Merge pull request #494 from cita-777/release/feat-update-center-dev-tags-20260417

[chore] publish dev snapshot to main

.github/workflows/ci.yml

@@ -319,7 +319,6 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 10
     # Keep PR audit informational so build/test/schema remain the merge gate.
-    continue-on-error: true
 
     steps:
       - name: Checkout
@@ -335,7 +334,15 @@ jobs:
         run: npm ci --prefer-offline --no-audit --no-fund
 
       - name: Audit production dependencies
-        run: npm audit --omit=dev --audit-level=high
+        run: |
+          set +e
+          npm audit --omit=dev --audit-level=high
+          status=$?
+          if [ "$status" -ne 0 ] && [ "${{ github.event_name }}" = "pull_request" ]; then
+            echo "::warning::npm audit reported production dependency vulnerabilities; PR audit stays informational."
+            exit 0
+          fi
+          exit "$status"
 
   publish-docker-arch:
     name: Publish Docker Image (${{ matrix.arch }})

docker/Dockerfile

@@ -1,4 +1,3 @@
-# syntax=docker/dockerfile:1.7
 # Keep the Docker base on Node 22 because the official Node 24/25 slim images
 # no longer publish linux/arm/v7 manifests, which breaks our armv7 Docker jobs.
 FROM node:22-bookworm-slim AS builder
@@ -12,12 +11,12 @@ RUN apt-get update \
 ENV PYTHON=/usr/bin/python3
 
 COPY package.json package-lock.json ./
-RUN --mount=type=cache,target=/root/.npm npm ci --ignore-scripts --no-audit --no-fund
+RUN npm ci --ignore-scripts --no-audit --no-fund
 RUN npm rebuild esbuild sharp better-sqlite3 --no-audit --no-fund
 
 COPY . .
 RUN npm run build:web && npm run build:server
-RUN --mount=type=cache,target=/root/.npm npm prune --omit=dev --no-audit --no-fund
+RUN npm prune --omit=dev --no-audit --no-fund
 
 FROM node:22-bookworm-slim
 

docs/.vitepress/config.test.ts

@@ -1,5 +1,5 @@
 import { existsSync, globSync } from 'node:fs';
-import { resolve } from 'node:path';
+import { dirname, resolve } from 'node:path';
 import { fileURLToPath } from 'node:url';
 import { describe, expect, it } from 'vitest';
 import config from './config';
@@ -36,11 +36,21 @@ const getAlias = (aliasConfig: unknown, specifier: string): string | undefined =
 const repoRoot = fileURLToPath(new URL('../..', import.meta.url));
 
 const getExpectedEntry = (hoistedRelativePath: string, pnpmPattern: string) => {
-  const hoistedEntry = resolve(repoRoot, hoistedRelativePath);
+  let currentRoot = repoRoot;
+
+  while (true) {
+    const hoistedEntry = resolve(currentRoot, hoistedRelativePath);
     if (existsSync(hoistedEntry)) return hoistedEntry;
 
-  const [pnpmEntry] = globSync(resolve(repoRoot, pnpmPattern));
-  return pnpmEntry;
+    const [pnpmEntry] = globSync(resolve(currentRoot, pnpmPattern));
+    if (pnpmEntry) return pnpmEntry;
+
+    const parentRoot = dirname(currentRoot);
+    if (parentRoot === currentRoot) break;
+    currentRoot = parentRoot;
+  }
+
+  return undefined;
 };
 
 describe('docs vitepress config', () => {

docs/.vitepress/config.ts

@@ -1,17 +1,28 @@
 import { existsSync, globSync } from 'node:fs';
-import { resolve } from 'node:path';
+import { dirname, resolve } from 'node:path';
 import { fileURLToPath } from 'node:url';
 import { defineConfig } from 'vitepress';
 import { withMermaid } from 'vitepress-plugin-mermaid';
 
 const repoRoot = fileURLToPath(new URL('../..', import.meta.url));
-const resolveDependencyEntry = (hoistedRelativePath: string, pnpmPattern: string) => {
-  const hoistedEntry = resolve(repoRoot, hoistedRelativePath);
+
+function resolveDependencyEntry(hoistedRelativePath: string, pnpmPattern: string) {
+  let currentRoot = repoRoot;
+
+  while (true) {
+    const hoistedEntry = resolve(currentRoot, hoistedRelativePath);
     if (existsSync(hoistedEntry)) return hoistedEntry;
 
-  const [pnpmEntry] = globSync(resolve(repoRoot, pnpmPattern));
-  return pnpmEntry;
-};
+    const [pnpmEntry] = globSync(resolve(currentRoot, pnpmPattern));
+    if (pnpmEntry) return pnpmEntry;
+
+    const parentRoot = dirname(currentRoot);
+    if (parentRoot === currentRoot) break;
+    currentRoot = parentRoot;
+  }
+
+  return undefined;
+}
 
 const dayjsEsmEntry = resolveDependencyEntry(
   'node_modules/dayjs/esm/index.js',

docs/k3s-update-center.md

@@ -308,6 +308,11 @@ helper 端使用：
 - `GitHub Releases`
 - `Docker Hub`
 
+补充一点：
+
+- `Docker Hub` 里的**自动发现**仍然优先 `latest` / `main` / 稳定 SemVer 标签
+- 如果你要部署 `dev`、分支名、临时标签或 `sha-*` 这类非稳定标签，直接在 Docker Hub 卡片底部用“手动部署 Docker Hub 标签”填写即可
+
 对已经跑起来的 K3s / Helm 用户来说，更新中心的日常配置主要就在这一页：
 
 ```text
@@ -325,7 +330,7 @@ helper 端使用：
    - 版本来源发现了可部署版本
    - 如果 Docker Hub 显示的是 `latest @ sha256:...`，说明页面已经识别到 alias tag 当前指向的具体镜像 digest
    - Deploy Helper 显示健康
-4. 再点部署按钮
+4. 如果你要跟随稳定候选，直接点部署按钮；如果你要切到 `dev` / 分支 / 临时标签，就在 Docker Hub 卡片底部手动填写 tag，必要时连 digest 一起填
 5. 在页面下方看部署日志
 6. 如果升级后发现问题，可以直接在“回退历史”里点旧 revision 回滚；只要该 revision 当时记录了 digest，就会跟着一起回到对应镜像
 

scripts/dev/docker.workflow.test.ts

@@ -34,6 +34,13 @@ describe('docker workflows', () => {
     expect(dockerfile).toContain('FROM node:22-bookworm-slim');
   });
 
+  it('avoids buildkit-only frontend syntax so managed docker builders can parse it reliably', () => {
+    const dockerfile = readFileSync(resolve(process.cwd(), 'docker/Dockerfile'), 'utf8');
+
+    expect(dockerfile).not.toContain('# syntax=docker/dockerfile:');
+    expect(dockerfile).not.toContain('RUN --mount=type=cache');
+  });
+
   it('keeps server docker builds isolated from desktop packaging dependencies', () => {
     const dockerfile = readFileSync(resolve(process.cwd(), 'docker/Dockerfile'), 'utf8');
 

scripts/dev/update-readme-contributors.test.ts

@@ -0,0 +1,123 @@
+import { describe, expect, it } from 'vitest';
+import {
+  parsePaginatedJson,
+  normalizeContributors,
+  replaceContributorsSection,
+  renderContributorsBlock,
+} from './update-readme-contributors.js';
+
+describe('README contributors updater', () => {
+  it('filters anonymous and bot contributors before rendering', () => {
+    const contributors = normalizeContributors([
+      {
+        login: 'cita-777',
+        html_url: 'https://github.com/cita-777',
+        avatar_url: 'https://avatars.githubusercontent.com/u/177306803?v=4',
+        contributions: 483,
+      },
+      {
+        login: 'dependabot[bot]',
+        html_url: 'https://github.com/apps/dependabot',
+        avatar_url: 'https://avatars.githubusercontent.com/in/29110?v=4',
+        contributions: 12,
+      },
+      {
+        login: null,
+        html_url: null,
+        avatar_url: null,
+        contributions: 22,
+      },
+      {
+        login: 'Babylonehy',
+        html_url: 'https://github.com/Babylonehy',
+        avatar_url: 'https://avatars.githubusercontent.com/u/30937892?v=4',
+        contributions: 3,
+      },
+    ]);
+
+    expect(contributors).toEqual([
+      {
+        login: 'cita-777',
+        htmlUrl: 'https://github.com/cita-777',
+        avatarUrl: 'https://avatars.githubusercontent.com/u/177306803?v=4&s=48',
+      },
+      {
+        login: 'Babylonehy',
+        htmlUrl: 'https://github.com/Babylonehy',
+        avatarUrl: 'https://avatars.githubusercontent.com/u/30937892?v=4&s=48',
+      },
+    ]);
+  });
+
+  it('replaces the marked README section with an avatar wall', () => {
+    const readme = [
+      '# Metapi',
+      '',
+      '<!-- metapi-contributors:start -->',
+      'old block',
+      '<!-- metapi-contributors:end -->',
+      '',
+      'tail',
+    ].join('\n');
+
+    const next = replaceContributorsSection(readme, [
+      {
+        login: 'cita-777',
+        htmlUrl: 'https://github.com/cita-777',
+        avatarUrl: 'https://avatars.githubusercontent.com/u/177306803?v=4&s=48',
+      },
+      {
+        login: 'Hureru',
+        htmlUrl: 'https://github.com/Hureru',
+        avatarUrl: 'https://avatars.githubusercontent.com/u/121702350?v=4&s=48',
+      },
+      {
+        login: 'Babylonehy',
+        htmlUrl: 'https://github.com/Babylonehy',
+        avatarUrl: 'https://avatars.githubusercontent.com/u/30937892?v=4&s=48',
+      },
+    ], 2);
+
+    expect(next).toContain('<p align="left">');
+    expect(next).toContain('https://github.com/cita-777');
+    expect(next).toContain('https://github.com/Hureru');
+    expect(next).toContain('https://github.com/Babylonehy');
+    expect(next).not.toContain('old block');
+    expect(next).toContain('tail');
+  });
+
+  it('renders contributors across multiple rows', () => {
+    const html = renderContributorsBlock([
+      {
+        login: 'cita-777',
+        htmlUrl: 'https://github.com/cita-777',
+        avatarUrl: 'https://avatars.githubusercontent.com/u/177306803?v=4&s=48',
+      },
+      {
+        login: 'Hureru',
+        htmlUrl: 'https://github.com/Hureru',
+        avatarUrl: 'https://avatars.githubusercontent.com/u/121702350?v=4&s=48',
+      },
+      {
+        login: 'Babylonehy',
+        htmlUrl: 'https://github.com/Babylonehy',
+        avatarUrl: 'https://avatars.githubusercontent.com/u/30937892?v=4&s=48',
+      },
+    ], 2);
+
+    expect(html).toContain('\n  <a href="https://github.com/cita-777"');
+    expect(html).toContain('\n  <a href="https://github.com/Babylonehy"');
+    expect((html.match(/\n  <a href=/g) || []).length).toBe(2);
+  });
+
+  it('rejects invalid per-line values instead of looping forever', () => {
+    expect(() => renderContributorsBlock([], 0)).toThrow('perLine must be a positive integer');
+  });
+
+  it('parses concatenated paginated gh api arrays without requiring newlines between pages', () => {
+    expect(parsePaginatedJson('[{"login":"a"}][{"login":"b"}]')).toEqual([
+      { login: 'a' },
+      { login: 'b' },
+    ]);
+  });
+});

scripts/dev/update-readme-contributors.ts

@@ -0,0 +1,197 @@
+import { execFileSync } from 'node:child_process';
+import { readFileSync, writeFileSync } from 'node:fs';
+import { resolve } from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const REPO = 'cita-777/metapi';
+const README_FILES = ['README.md', 'README_EN.md'] as const;
+const START_MARKER = '<!-- metapi-contributors:start -->';
+const END_MARKER = '<!-- metapi-contributors:end -->';
+const PER_LINE = 10;
+const EXCLUDED_LOGINS = new Set(['dependabot[bot]']);
+
+type GitHubContributor = {
+  login: string | null;
+  html_url: string | null;
+  avatar_url: string | null;
+  contributions?: number;
+};
+
+export type ReadmeContributor = {
+  login: string;
+  htmlUrl: string;
+  avatarUrl: string;
+};
+
+export function normalizeContributors(
+  contributors: GitHubContributor[],
+): ReadmeContributor[] {
+  const seen = new Set<string>();
+
+  return contributors
+    .filter((contributor) => {
+      const login = contributor.login?.trim();
+      const htmlUrl = contributor.html_url?.trim();
+      const avatarUrl = contributor.avatar_url?.trim();
+
+      if (!login || !htmlUrl || !avatarUrl) {
+        return false;
+      }
+
+      if (EXCLUDED_LOGINS.has(login.toLowerCase()) || login.toLowerCase().endsWith('[bot]')) {
+        return false;
+      }
+
+      return true;
+    })
+    .sort((left, right) => (right.contributions ?? 0) - (left.contributions ?? 0))
+    .filter((contributor) => {
+      const key = contributor.login!.toLowerCase();
+      if (seen.has(key)) {
+        return false;
+      }
+      seen.add(key);
+      return true;
+    })
+    .map((contributor) => ({
+      login: contributor.login!.trim(),
+      htmlUrl: contributor.html_url!.trim(),
+      avatarUrl: normalizeAvatarUrl(contributor.avatar_url!.trim()),
+    }));
+}
+
+export function renderContributorsBlock(
+  contributors: ReadmeContributor[],
+  perLine = PER_LINE,
+): string {
+  if (!Number.isInteger(perLine) || perLine <= 0) {
+    throw new RangeError('perLine must be a positive integer');
+  }
+
+  if (contributors.length === 0) {
+    return '<p align="left">\n  <sub>No public contributors found yet.</sub>\n</p>';
+  }
+
+  const lines: string[] = [];
+
+  for (let index = 0; index < contributors.length; index += perLine) {
+    const chunk = contributors.slice(index, index + perLine);
+    lines.push(
+      `  ${chunk
+        .map(
+          (contributor) =>
+            `<a href="${contributor.htmlUrl}"><img src="${contributor.avatarUrl}" width="48" height="48" alt="${contributor.login}" title="${contributor.login}"/></a>`,
+        )
+        .join(' ')}`,
+    );
+  }
+
+  return `<p align="left">\n${lines.join('\n')}\n</p>`;
+}
+
+export function replaceContributorsSection(
+  readme: string,
+  contributors: ReadmeContributor[],
+  perLine = PER_LINE,
+): string {
+  const start = readme.indexOf(START_MARKER);
+  const end = readme.indexOf(END_MARKER);
+
+  if (start === -1 || end === -1 || end < start) {
+    throw new Error('README contributors markers are missing');
+  }
+
+  const before = readme.slice(0, start + START_MARKER.length);
+  const after = readme.slice(end);
+  return `${before}\n${renderContributorsBlock(contributors, perLine)}\n${after}`;
+}
+
+export function updateReadmeFile(
+  readmePath: string,
+  contributors: ReadmeContributor[],
+  perLine = PER_LINE,
+): boolean {
+  const current = readFileSync(readmePath, 'utf8');
+  const next = replaceContributorsSection(current, contributors, perLine);
+  if (current === next) {
+    return false;
+  }
+
+  writeFileSync(readmePath, next);
+  return true;
+}
+
+export function fetchContributors(repo = REPO): ReadmeContributor[] {
+  const raw = execFileSync(
+    'gh',
+    ['api', `repos/${repo}/contributors?per_page=100&anon=1`, '--paginate'],
+    { encoding: 'utf8', stdio: ['ignore', 'pipe', 'pipe'] },
+  );
+
+  return normalizeContributors(parsePaginatedJson(raw) as GitHubContributor[]);
+}
+
+export function parsePaginatedJson(raw: string): unknown[] {
+  const trimmedRaw = raw.trim();
+  if (!trimmedRaw) return [];
+
+  try {
+    const parsed = JSON.parse(trimmedRaw);
+    return Array.isArray(parsed) ? parsed : [parsed];
+  } catch {
+    // Fall through to paginated-array parsing.
+  }
+
+  try {
+    const mergedArrays = JSON.parse(`[${trimmedRaw.replace(/\]\s*\[/g, '],[')}]`);
+    if (Array.isArray(mergedArrays)) {
+      return mergedArrays.flatMap((item) => (Array.isArray(item) ? item : [item]));
+    }
+  } catch {
+    // Fall through to line-by-line parsing.
+  }
+
+  const items: unknown[] = [];
+  for (const line of trimmedRaw.split('\n')) {
+    const trimmed = line.trim();
+    if (!trimmed) {
+      continue;
+    }
+
+    const parsed = JSON.parse(trimmed);
+    if (Array.isArray(parsed)) {
+      items.push(...parsed);
+    } else {
+      items.push(parsed);
+    }
+  }
+
+  return items;
+}
+
+function normalizeAvatarUrl(url: string): string {
+  if (!/^https?:/i.test(url)) {
+    return url;
+  }
+
+  const lower = url.toLowerCase();
+  if (lower.includes('s=') || lower.includes('size=')) {
+    return url;
+  }
+
+  return `${url}${url.includes('?') ? '&' : '?'}s=48`;
+}
+
+function main(): void {
+  const contributors = fetchContributors(REPO);
+  const updatedFiles = README_FILES.filter((path) =>
+    updateReadmeFile(resolve(process.cwd(), path), contributors),
+  );
+
+  const touched = updatedFiles.length > 0 ? updatedFiles.join(', ') : 'none';
+  console.log(`Updated README contributors for ${contributors.length} people: ${touched}`);
+}
+
+if (process.argv[1] && fileURLToPath(import.meta.url) === resolve(process.argv[1])) {
+  main();
+}

src/server/proxy-core/capabilities/responsesCompact.test.ts

@@ -0,0 +1,104 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  ensureCompactResponsesJsonAcceptHeader,
+  sanitizeCompactResponsesRequestBody,
+  shouldForceResponsesUpstreamStream,
+} from './responsesCompact.js';
+
+describe('sanitizeCompactResponsesRequestBody', () => {
+  it('removes stream fields and store from codex compact requests', () => {
+    expect(sanitizeCompactResponsesRequestBody({
+      stream: true,
+      stream_options: { include_obfuscation: true },
+      store: false,
+      input: 'hello',
+    }, {
+      sitePlatform: 'codex',
+    })).toEqual({
+      input: 'hello',
+    });
+  });
+
+  it('removes store from sub2api compact requests', () => {
+    expect(sanitizeCompactResponsesRequestBody({
+      stream: true,
+      stream_options: { include_obfuscation: true },
+      store: false,
+      input: 'hello',
+    }, {
+      sitePlatform: 'sub2api',
+    })).toEqual({
+      input: 'hello',
+    });
+  });
+
+  it('keeps store for unrelated compact platforms while still removing stream fields', () => {
+    expect(sanitizeCompactResponsesRequestBody({
+      stream: true,
+      stream_options: { include_obfuscation: true },
+      store: false,
+      input: 'hello',
+    }, {
+      sitePlatform: 'openai',
+    })).toEqual({
+      store: false,
+      input: 'hello',
+    });
+  });
+});
+
+describe('shouldForceResponsesUpstreamStream', () => {
+  it('forces non-compact responses streaming for codex and sub2api', () => {
+    expect(shouldForceResponsesUpstreamStream({
+      sitePlatform: 'codex',
+      isCompactRequest: false,
+    })).toBe(true);
+    expect(shouldForceResponsesUpstreamStream({
+      sitePlatform: 'sub2api',
+      isCompactRequest: false,
+    })).toBe(true);
+  });
+
+  it('does not force compact or unrelated platforms', () => {
+    expect(shouldForceResponsesUpstreamStream({
+      sitePlatform: 'codex',
+      isCompactRequest: true,
+    })).toBe(false);
+    expect(shouldForceResponsesUpstreamStream({
+      sitePlatform: 'openai',
+      isCompactRequest: false,
+    })).toBe(false);
+  });
+});
+
+describe('ensureCompactResponsesJsonAcceptHeader', () => {
+  it('adds application/json accept for codex and sub2api compact requests when missing', () => {
+    expect(ensureCompactResponsesJsonAcceptHeader({}, {
+      sitePlatform: 'codex',
+    })).toEqual({
+      accept: 'application/json',
+    });
+    expect(ensureCompactResponsesJsonAcceptHeader({}, {
+      sitePlatform: 'sub2api',
+    })).toEqual({
+      accept: 'application/json',
+    });
+  });
+
+  it('forces compact requests back to application/json even when callers asked for SSE', () => {
+    expect(ensureCompactResponsesJsonAcceptHeader({
+      Accept: 'text/event-stream',
+    }, {
+      sitePlatform: 'codex',
+    })).toEqual({
+      accept: 'application/json',
+    });
+  });
+
+  it('preserves unrelated platforms', () => {
+    expect(ensureCompactResponsesJsonAcceptHeader({}, {
+      sitePlatform: 'openai',
+    })).toEqual({});
+  });
+});

src/server/proxy-core/capabilities/responsesCompact.ts

@@ -2,35 +2,82 @@ function asTrimmedString(value: unknown): string {
   return typeof value === 'string' ? value.trim() : '';
 }
 
+function shouldStripCompactResponsesStore(sitePlatform?: string): boolean {
+  const normalized = asTrimmedString(sitePlatform).toLowerCase();
+  return normalized === 'codex' || normalized === 'sub2api';
+}
+
+function shouldForceCompactResponsesJsonAccept(sitePlatform?: string): boolean {
+  return shouldStripCompactResponsesStore(sitePlatform);
+}
+
+export function shouldForceResponsesUpstreamStream(input: {
+  sitePlatform?: string;
+  isCompactRequest?: boolean;
+}): boolean {
+  if (input.isCompactRequest) return false;
+  const sitePlatform = asTrimmedString(input.sitePlatform).toLowerCase();
+  return sitePlatform === 'codex' || sitePlatform === 'sub2api';
+}
+
 export function sanitizeCompactResponsesRequestBody(
   body: Record<string, unknown>,
+  options?: {
+    sitePlatform?: string;
+  },
 ): Record<string, unknown> {
   const next = { ...body };
   delete next.stream;
   delete next.stream_options;
+  if (shouldStripCompactResponsesStore(options?.sitePlatform)) {
+    delete next.store;
+  }
   return next;
 }
 
+export function ensureCompactResponsesJsonAcceptHeader(
+  headers: Record<string, string>,
+  options?: {
+    sitePlatform?: string;
+  },
+): Record<string, string> {
+  if (!shouldForceCompactResponsesJsonAccept(options?.sitePlatform)) return headers;
+  const nextHeaders = { ...headers };
+  delete (nextHeaders as Record<string, unknown>).Accept;
+  delete (nextHeaders as Record<string, unknown>).accept;
+  return {
+    ...nextHeaders,
+    accept: 'application/json',
+  };
+}
+
 export function shouldFallbackCompactResponsesToResponses(input: {
   status?: number;
   rawErrText?: string;
+  requestPath?: string;
 }): boolean {
   const status = Number.isFinite(Number(input.status)) ? Number(input.status) : 0;
   const compact = asTrimmedString(input.rawErrText).toLowerCase();
-  const hasCompactHint = (
+  const requestPath = asTrimmedString(input.requestPath).toLowerCase();
+  const hasRawCompactHint = (
     compact.includes('/responses/compact')
     || compact.includes('responses/compact')
     || compact.includes('compact endpoint')
     || /(^|[^a-z])compact([^a-z]|$)/.test(compact)
   );
+  const hasCompactRequestPathHint = (
+    requestPath.endsWith('/responses/compact')
+    || requestPath.endsWith('/v1/responses/compact')
+  );
+  const isInvalidCompactUrl = compact.includes('invalid url') && hasRawCompactHint;
 
   if (status === 404 || status === 405 || status === 501) return true;
 
   return (
-    compact.includes("unknown parameter: 'stream'")
-    || compact.includes('invalid url')
+    (compact.includes("unknown parameter: 'stream'") && (hasRawCompactHint || hasCompactRequestPathHint))
+    || isInvalidCompactUrl
     || (
-      hasCompactHint
+      hasRawCompactHint
       && (
         compact.includes('not supported')
         || compact.includes('unsupported')

src/server/proxy-core/cliProfiles/codexProfile.ts

@@ -1,71 +1,14 @@
 import type { CliProfileDefinition, DetectCliProfileInput } from './types.js';
+import {
+  detectCodexOfficialClientApp as detectCodexOfficialClientAppFromHeaders,
+  isCodexOfficialClientHeaders,
+} from '../../shared/codexClientFamily.js';
 
 type CodexOfficialClientApp = {
   clientAppId: string;
   clientAppName: string;
 };
 
-const CODEX_OFFICIAL_CLIENT_USER_AGENT_PREFIXES = [
-  'codex_cli_rs/',
-  'codex_vscode/',
-  'codex_app/',
-  'codex_chatgpt_desktop/',
-  'codex_atlas/',
-  'codex_exec/',
-  'codex_sdk_ts/',
-  'codex ',
-];
-
-const CODEX_OFFICIAL_CLIENT_ORIGINATOR_PREFIXES = [
-  'codex_',
-  'codex ',
-];
-
-const CODEX_OFFICIAL_CLIENT_APP_RULES = [
-  {
-    id: 'codex_cli_rs',
-    name: 'Codex CLI',
-    userAgentPrefixes: ['codex_cli_rs/'],
-    originatorPrefixes: ['codex_cli_rs'],
-  },
-  {
-    id: 'codex_vscode',
-    name: 'Codex VSCode',
-    userAgentPrefixes: ['codex_vscode/'],
-    originatorPrefixes: ['codex_vscode'],
-  },
-  {
-    id: 'codex_app',
-    name: 'Codex App',
-    userAgentPrefixes: ['codex_app/'],
-    originatorPrefixes: ['codex_app'],
-  },
-  {
-    id: 'codex_chatgpt_desktop',
-    name: 'Codex Desktop',
-    userAgentPrefixes: ['codex_chatgpt_desktop/', 'codex desktop/'],
-    originatorPrefixes: ['codex_chatgpt_desktop', 'codex desktop'],
-  },
-  {
-    id: 'codex_atlas',
-    name: 'Codex Atlas',
-    userAgentPrefixes: ['codex_atlas/'],
-    originatorPrefixes: ['codex_atlas'],
-  },
-  {
-    id: 'codex_exec',
-    name: 'Codex Exec',
-    userAgentPrefixes: ['codex_exec/'],
-    originatorPrefixes: ['codex_exec'],
-  },
-  {
-    id: 'codex_sdk_ts',
-    name: 'Codex SDK TS',
-    userAgentPrefixes: ['codex_sdk_ts/'],
-    originatorPrefixes: ['codex_sdk_ts'],
-  },
-] as const;
-
 function headerValueToStrings(value: unknown): string[] {
   if (typeof value === 'string') {
     const trimmed = value.trim();
@@ -114,20 +57,6 @@ function hasHeaderPrefix(headers: Record<string, unknown> | undefined, prefix: s
   });
 }
 
-function matchesHeaderPrefixes(value: string | string[] | null, prefixes: readonly string[]): boolean {
-  const values = Array.isArray(value)
-    ? value.map((item) => item.trim().toLowerCase()).filter(Boolean)
-    : [value?.trim().toLowerCase() || ''].filter(Boolean);
-  if (values.length === 0) return false;
-
-  return values.some((normalizedValue) => prefixes.some((prefix) => {
-    const normalizedPrefix = prefix.trim().toLowerCase();
-    if (!normalizedPrefix) return false;
-    return normalizedValue.startsWith(normalizedPrefix)
-      || normalizedValue.includes(normalizedPrefix);
-  }));
-}
-
 function isCodexPath(path: string): boolean {
   const normalizedPath = path.trim().toLowerCase();
   return normalizedPath === '/v1/responses'
@@ -138,16 +67,13 @@ function isCodexPath(path: string): boolean {
 export function detectCodexOfficialClientApp(
   headers?: Record<string, unknown>,
 ): CodexOfficialClientApp | null {
-  for (const rule of CODEX_OFFICIAL_CLIENT_APP_RULES) {
-    const matchesOriginator = matchesHeaderPrefixes(getHeaderValues(headers, 'originator'), rule.originatorPrefixes);
-    const matchesUserAgent = matchesHeaderPrefixes(getHeaderValues(headers, 'user-agent'), rule.userAgentPrefixes);
-    if (!matchesOriginator && !matchesUserAgent) continue;
-    return {
-      clientAppId: rule.id,
-      clientAppName: rule.name,
-    };
+  const detected = detectCodexOfficialClientAppFromHeaders(headers);
+  return detected
+    ? {
+      clientAppId: detected.clientAppId,
+      clientAppName: detected.clientAppName,
     }
-  return null;
+    : null;
 }
 
 export function isCodexResponsesSurface(headers?: Record<string, unknown>): boolean {
@@ -158,7 +84,10 @@ export function isCodexResponsesSurface(headers?: Record<string, unknown>): bool
 }
 
 export function getCodexSessionId(headers?: Record<string, unknown>): string | null {
-  return getHeaderValue(headers, 'session_id') || getHeaderValue(headers, 'session-id');
+  return getHeaderValue(headers, 'session_id')
+    || getHeaderValue(headers, 'session-id')
+    || getHeaderValue(headers, 'conversation_id')
+    || getHeaderValue(headers, 'conversation-id');
 }
 
 export function isCodexRequest(input: DetectCliProfileInput): boolean {
@@ -166,9 +95,7 @@ export function isCodexRequest(input: DetectCliProfileInput): boolean {
   const headers = input.headers;
   if (!headers) return false;
 
-  const originator = getHeaderValues(headers, 'originator');
-  if (matchesHeaderPrefixes(originator, CODEX_OFFICIAL_CLIENT_ORIGINATOR_PREFIXES)) return true;
-  if (matchesHeaderPrefixes(getHeaderValues(headers, 'user-agent'), CODEX_OFFICIAL_CLIENT_USER_AGENT_PREFIXES)) return true;
+  if (isCodexOfficialClientHeaders(headers)) return true;
   if (getHeaderValue(headers, 'openai-beta')) return true;
   if (hasHeaderPrefix(headers, 'x-stainless-')) return true;
   if (getCodexSessionId(headers)) return true;

src/server/proxy-core/cliProfiles/registry.test.ts

@@ -42,6 +42,29 @@ describe('detectCliProfile', () => {
     });
   });
 
+  it('treats conversation_id as a Codex continuation marker when session_id is absent', () => {
+    expect(detectCliProfile({
+      downstreamPath: '/v1/responses',
+      headers: {
+        conversation_id: 'codex-conversation-123',
+      },
+    })).toEqual({
+      id: 'codex',
+      sessionId: 'codex-conversation-123',
+      traceHint: 'codex-conversation-123',
+      clientAppId: 'codex',
+      clientAppName: 'Codex',
+      clientConfidence: 'heuristic',
+      capabilities: {
+        supportsResponsesCompact: true,
+        supportsResponsesWebsocketIncremental: true,
+        preservesContinuation: true,
+        supportsCountTokens: false,
+        echoesTurnState: true,
+      },
+    });
+  });
+
   it('detects broader Codex official-client headers from user-agent and originator prefixes', () => {
     expect(detectCliProfile({
       downstreamPath: '/v1/responses',

src/server/proxy-core/providers/claudeProviderProfile.ts

@@ -1,5 +1,10 @@
 import type { PreparedProviderRequest, PrepareProviderRequestInput, ProviderProfile } from './types.js';
-import { buildClaudeRuntimeHeaders } from './headerUtils.js';
+import {
+  buildClaudeRuntimeHeaders,
+  CLAUDE_API_KEY_DEFAULT_BETA_HEADER,
+  CLAUDE_DEFAULT_BETA_HEADER,
+  CLAUDE_TOKEN_COUNTING_BETA,
+} from './headerUtils.js';
 
 export const claudeProviderProfile: ProviderProfile = {
   id: 'claude',
@@ -22,6 +27,10 @@ export const claudeProviderProfile: ProviderProfile = {
         stream: isCountTokens ? false : input.stream,
         isClaudeOauthUpstream,
         tokenValue: input.tokenValue,
+        defaultBetaHeader: isClaudeOauthUpstream
+          ? CLAUDE_DEFAULT_BETA_HEADER
+          : CLAUDE_API_KEY_DEFAULT_BETA_HEADER,
+        ...(isCountTokens ? { extraBetas: [CLAUDE_TOKEN_COUNTING_BETA] } : {}),
       }),
       body: input.body,
       runtime: {

src/server/proxy-core/providers/headerUtils.test.ts

@@ -95,6 +95,7 @@ describe('provider header utils', () => {
 
     expect(headers['anthropic-version']).toBe('2023-06-01');
     expect(headers['anthropic-beta']).toContain('claude-code-20250219');
+    expect(headers['anthropic-beta']).toContain('fine-grained-tool-streaming-2025-05-14');
     expect(headers['anthropic-beta']).toContain('custom-beta');
     expect(headers.Authorization).toBe('Bearer oauth-token');
     expect(headers.authorization).toBeUndefined();

src/server/proxy-core/providers/headerUtils.ts

@@ -1,9 +1,12 @@
 import { pbkdf2Sync, randomUUID } from 'node:crypto';
+import { inferCodexOfficialOriginator } from '../../shared/codexClientFamily.js';
 
 const CODEX_CLIENT_VERSION = '0.101.0';
 const CODEX_DEFAULT_USER_AGENT = 'codex_cli_rs/0.101.0 (Mac OS 26.0.1; arm64) Apple_Terminal/464';
 const CLAUDE_DEFAULT_USER_AGENT = 'claude-cli/2.1.63 (external, cli)';
-const CLAUDE_DEFAULT_BETA_HEADER = 'claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,context-management-2025-06-27,prompt-caching-scope-2026-01-05';
+export const CLAUDE_TOKEN_COUNTING_BETA = 'token-counting-2024-11-01';
+export const CLAUDE_DEFAULT_BETA_HEADER = 'claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14,context-management-2025-06-27,prompt-caching-scope-2026-01-05';
+export const CLAUDE_API_KEY_DEFAULT_BETA_HEADER = 'claude-code-20250219,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14,context-management-2025-06-27,prompt-caching-scope-2026-01-05';
 
 export function headerValueToString(value: unknown): string | null {
   if (typeof value === 'string') {
@@ -132,7 +135,9 @@ export function buildCodexRuntimeHeaders(input: {
     || getInputHeader(input.baseHeaders, 'Authorization')
     || ''
   );
-  const originator = getInputHeader(input.providerHeaders, 'originator')
+  const originator = inferCodexOfficialOriginator(input.providerHeaders)
+    || inferCodexOfficialOriginator(input.baseHeaders)
+    || getInputHeader(input.providerHeaders, 'originator')
     || input.originatorDefault
     || 'codex_cli_rs';
   const accountId = getInputHeader(input.providerHeaders, 'chatgpt-account-id');

src/server/proxy-core/providers/registry.test.ts

@@ -17,7 +17,7 @@ describe('resolveProviderProfile', () => {
         Authorization: 'Bearer oauth-access-token',
       },
       providerHeaders: {
-        Originator: 'codex_cli_rs',
+        Originator: 'Codex Desktop',
         'Chatgpt-Account-Id': 'chatgpt-account-123',
       },
       codexSessionCacheKey: 'gpt-5.2-codex:user-123',
@@ -30,7 +30,7 @@ describe('resolveProviderProfile', () => {
 
     expect(result.path).toBe('/responses');
     expect(result.headers.Authorization).toBe('Bearer oauth-access-token');
-    expect(result.headers.Originator).toBe('codex_cli_rs');
+    expect(result.headers.Originator).toBe('codex_chatgpt_desktop');
     expect(result.headers['Chatgpt-Account-Id']).toBe('chatgpt-account-123');
     expect(result.headers.Session_id).toMatch(/^[0-9a-f-]{36}$/i);
     expect(result.headers.Conversation_id).toBe(result.headers.Session_id);
@@ -74,6 +74,7 @@ describe('resolveProviderProfile', () => {
     expect(result.headers.Authorization).toBe('Bearer oauth-access-token');
     expect(result.headers['x-api-key']).toBeUndefined();
     expect(result.headers['anthropic-version']).toBe('2023-06-01');
+    expect(result.headers['anthropic-beta']).toContain('oauth-2025-04-20');
     expect(result.headers['Accept-Encoding']).toBe('gzip, deflate, br, zstd');
     expect(result.runtime).toMatchObject({
       executor: 'claude',
@@ -83,6 +84,60 @@ describe('resolveProviderProfile', () => {
     expect(result.body).toBe(protocolBody);
   });
 
+  it('drops oauth-only claude betas for api-key upstreams', () => {
+    const profile = resolveProviderProfile('claude');
+    expect(profile?.id).toBe('claude');
+
+    const result = profile!.prepareRequest({
+      endpoint: 'messages',
+      modelName: 'claude-opus-4-6',
+      stream: false,
+      tokenValue: 'sk-claude-api-key',
+      sitePlatform: 'openai',
+      baseHeaders: {
+        'Content-Type': 'application/json',
+      },
+      claudeHeaders: {},
+      body: {
+        model: 'claude-opus-4-6',
+        max_tokens: 256,
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+    });
+
+    expect(result.headers.Authorization).toBeUndefined();
+    expect(result.headers['x-api-key']).toBe('sk-claude-api-key');
+    expect(result.headers['anthropic-beta']).not.toContain('oauth-2025-04-20');
+    expect(result.headers['anthropic-beta']).toContain('fine-grained-tool-streaming-2025-05-14');
+  });
+
+  it('adds token-counting beta when building claude count_tokens requests', () => {
+    const profile = resolveProviderProfile('claude');
+    expect(profile?.id).toBe('claude');
+
+    const result = profile!.prepareRequest({
+      endpoint: 'messages',
+      modelName: 'claude-opus-4-6',
+      stream: false,
+      action: 'countTokens',
+      tokenValue: 'oauth-access-token',
+      oauthProvider: 'claude',
+      sitePlatform: 'claude',
+      baseHeaders: {
+        'Content-Type': 'application/json',
+      },
+      claudeHeaders: {},
+      body: {
+        model: 'claude-opus-4-6',
+        max_tokens: 256,
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+    });
+
+    expect(result.path).toBe('/v1/messages/count_tokens?beta=true');
+    expect(result.headers['anthropic-beta']).toContain('token-counting-2024-11-01');
+  });
+
   it('builds gemini-cli provider requests with wrapped runtime envelope and project validation', () => {
     const profile = resolveProviderProfile('gemini-cli');
     expect(profile?.id).toBe('gemini-cli');

src/server/proxy-core/runtime/codexSessionResponseStore.test.ts

@@ -1,4 +1,5 @@
-import { describe, expect, it } from 'vitest';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+import { config } from '../../config.js';
 import {
   buildCodexSessionResponseStoreKey,
   clearCodexSessionResponseId,
@@ -7,6 +8,10 @@ import {
   setCodexSessionResponseId,
 } from './codexSessionResponseStore.js';
 
+afterEach(() => {
+  vi.useRealTimers();
+});
+
 describe('codexSessionResponseStore', () => {
   it('evicts the oldest session id when the store exceeds the cap', () => {
     resetCodexSessionResponseStore();
@@ -154,4 +159,39 @@ describe('codexSessionResponseStore', () => {
 
     resetCodexSessionResponseStore();
   });
+
+  it('expires remembered continuation ids after the sticky-session TTL window', () => {
+    vi.useFakeTimers();
+    vi.setSystemTime(new Date('2026-04-09T00:00:00Z'));
+    resetCodexSessionResponseStore();
+
+    setCodexSessionResponseId('session-expire', 'resp-expire');
+    expect(getCodexSessionResponseId('session-expire')).toBe('resp-expire');
+
+    vi.advanceTimersByTime(Math.max(5 * 60 * 1000, Math.trunc(config.proxyStickySessionTtlMs || 0)) + 1);
+
+    expect(getCodexSessionResponseId('session-expire')).toBeNull();
+
+    resetCodexSessionResponseStore();
+  });
+
+  it('refreshes the remembered continuation TTL when an active session reads it', () => {
+    vi.useFakeTimers();
+    vi.setSystemTime(new Date('2026-04-09T00:00:00Z'));
+    resetCodexSessionResponseStore();
+
+    const ttlMs = Math.max(5 * 60 * 1000, Math.trunc(config.proxyStickySessionTtlMs || 0));
+    setCodexSessionResponseId('session-touch', 'resp-touch');
+
+    vi.advanceTimersByTime(ttlMs - 1_000);
+    expect(getCodexSessionResponseId('session-touch')).toBe('resp-touch');
+
+    vi.advanceTimersByTime(2_000);
+    expect(getCodexSessionResponseId('session-touch')).toBe('resp-touch');
+
+    vi.advanceTimersByTime(ttlMs + 1);
+    expect(getCodexSessionResponseId('session-touch')).toBeNull();
+
+    resetCodexSessionResponseStore();
+  });
 });

src/server/proxy-core/runtime/codexSessionResponseStore.ts

@@ -1,6 +1,14 @@
-const MAX_CODEX_SESSION_RESPONSE_IDS = 10_000;
+import { config } from '../../config.js';
 
-const codexSessionResponseIds = new Map<string, string>();
+const MAX_CODEX_SESSION_RESPONSE_IDS = 10_000;
+const MIN_CODEX_SESSION_RESPONSE_TTL_MS = 5 * 60 * 1000;
+
+type SessionResponseEntry = {
+  responseId: string;
+  updatedAtMs: number;
+};
+
+const codexSessionResponseIds = new Map<string, SessionResponseEntry>();
 
 const SCOPED_SESSION_SEGMENT_PREFIX = 'session:';
 const SCOPED_STORE_KEY_SEGMENT_PATTERN = /^(site|account|channel):\d+$/;
@@ -71,6 +79,33 @@ function normalizeSessionId(sessionId: string): string {
   return sessionId.trim();
 }
 
+function getCodexSessionResponseTtlMs(): number {
+  return Math.max(
+    MIN_CODEX_SESSION_RESPONSE_TTL_MS,
+    Math.trunc(config.proxyStickySessionTtlMs || 0),
+  );
+}
+
+function isExpiredSessionResponseEntry(entry: SessionResponseEntry, nowMs = Date.now()): boolean {
+  return (entry.updatedAtMs + getCodexSessionResponseTtlMs()) <= nowMs;
+}
+
+function sweepExpiredSessionResponseIds(nowMs = Date.now()): void {
+  for (const [key, entry] of codexSessionResponseIds.entries()) {
+    if (!isExpiredSessionResponseEntry(entry, nowMs)) continue;
+    codexSessionResponseIds.delete(key);
+  }
+}
+
+function touchSessionResponseEntry(sessionId: string, nowMs = Date.now()): void {
+  for (const key of getSessionStoreKeys(sessionId)) {
+    const entry = codexSessionResponseIds.get(key);
+    if (!entry) continue;
+    entry.updatedAtMs = nowMs;
+    codexSessionResponseIds.set(key, entry);
+  }
+}
+
 export function buildCodexSessionResponseStoreKey(input: {
   sessionId: string;
   siteId?: number | null;
@@ -89,17 +124,23 @@ export function buildCodexSessionResponseStoreKey(input: {
 }
 
 export function getCodexSessionResponseId(sessionId: string): string | null {
+  const nowMs = Date.now();
+  sweepExpiredSessionResponseIds(nowMs);
   const normalized = normalizeSessionId(sessionId);
   if (!normalized) return null;
 
   const direct = codexSessionResponseIds.get(normalized);
-  if (direct) return direct;
+  if (direct) {
+    touchSessionResponseEntry(normalized, nowMs);
+    return direct.responseId;
+  }
 
   for (const fallbackKey of getFallbackSessionStoreKeys(normalized)) {
     const fallback = codexSessionResponseIds.get(fallbackKey);
     if (fallback) {
       reconcileScopedSessionFallback(fallbackKey);
-      return fallback;
+      touchSessionResponseEntry(fallbackKey, nowMs);
+      return fallback.responseId;
     }
   }
 
@@ -107,6 +148,8 @@ export function getCodexSessionResponseId(sessionId: string): string | null {
 }
 
 export function setCodexSessionResponseId(sessionId: string, responseId: string): void {
+  const nowMs = Date.now();
+  sweepExpiredSessionResponseIds(nowMs);
   const normalizedSessionId = normalizeSessionId(sessionId);
   const normalizedResponseId = responseId.trim();
   if (!normalizedSessionId || !normalizedResponseId) return;
@@ -117,7 +160,10 @@ export function setCodexSessionResponseId(sessionId: string, responseId: string)
     if (codexSessionResponseIds.has(key)) {
       codexSessionResponseIds.delete(key);
     }
-    codexSessionResponseIds.set(key, normalizedResponseId);
+    codexSessionResponseIds.set(key, {
+      responseId: normalizedResponseId,
+      updatedAtMs: nowMs,
+    });
   }
 
   while (codexSessionResponseIds.size > MAX_CODEX_SESSION_RESPONSE_IDS) {

src/server/proxy-core/runtime/codexWebsocketRuntime.test.ts

@@ -141,6 +141,67 @@ describe('codexWebsocketRuntime', () => {
     await runtime.closeSession('exec-session-close');
   });
 
+  it('preserves remembered continuation ids across websocket session closes and reconnects', async () => {
+    const { createCodexWebsocketRuntime } = await import('./codexWebsocketRuntime.js');
+    const runtime = createCodexWebsocketRuntime();
+
+    await runtime.sendRequest({
+      sessionId: 'exec-session-continue-after-close',
+      requestUrl: upstreamWsUrl,
+      headers: {
+        Authorization: 'Bearer oauth-access-token',
+        'OpenAI-Beta': 'responses_websockets=2026-02-06',
+      },
+      body: {
+        model: 'gpt-5.4',
+        input: [],
+      },
+    });
+
+    await runtime.closeSession('exec-session-continue-after-close');
+
+    const recovered = await runtime.sendRequest({
+      sessionId: 'exec-session-continue-after-close',
+      requestUrl: upstreamWsUrl,
+      headers: {
+        Authorization: 'Bearer oauth-access-token',
+        'OpenAI-Beta': 'responses_websockets=2026-02-06',
+      },
+      body: {
+        model: 'gpt-5.4',
+        input: [
+          {
+            id: 'tool_out_runtime_1',
+            type: 'function_call_output',
+            call_id: 'call_runtime_1',
+            output: '{"ok":true}',
+          },
+        ],
+      },
+    });
+
+    expect(recovered.events[0]).toMatchObject({
+      type: 'response.completed',
+      response: { id: 'resp-2' },
+    });
+    expect(upstreamConnectionCount).toBe(2);
+    expect(upstreamRequests).toHaveLength(2);
+    expect(upstreamRequests[1]).toMatchObject({
+      type: 'response.create',
+      previous_response_id: 'resp-1',
+      input: [
+        {
+          id: 'tool_out_runtime_1',
+          type: 'function_call_output',
+          call_id: 'call_runtime_1',
+          output: '{"ok":true}',
+        },
+      ],
+    });
+
+    await runtime.closeSession('exec-session-continue-after-close');
+  });
+
   it('returns response.incomplete as a terminal websocket event without rejecting the session turn', async () => {
     upstreamMessageHandler = (socket, parsed) => {
       socket.send(JSON.stringify({
@@ -413,7 +474,8 @@ describe('codexWebsocketRuntime', () => {
       message: 'account mismatch',
       status: 502,
     });
-    expect((error as CodexWebsocketRuntimeError).events).toEqual([
+    const runtimeError = error as InstanceType<typeof CodexWebsocketRuntimeError>;
+    expect(runtimeError.events).toEqual([
       expect.objectContaining({
         type: 'error',
         error: expect.objectContaining({

src/server/proxy-core/runtime/codexWebsocketRuntime.ts

@@ -337,7 +337,7 @@ async function sendSessionRequestAttempt(
     socket.once('error', onError);
     socket.once('close', onClose);
 
-    socket.send(JSON.stringify(buildCodexWebsocketRequestBody(input.body)), (error) => {
+    socket.send(JSON.stringify(buildCodexWebsocketRequestBody(input.body)), (error?: Error) => {
       if (!error) return;
       clearSessionSocket(session, socket);
       rejectWith(error.message || 'failed to send upstream websocket request');
@@ -409,7 +409,9 @@ export function createCodexWebsocketRuntime(input?: {
       await closeSocket(session.socket);
       session.socket = null;
       session.socketUrl = null;
-      clearCodexSessionResponseId(sessionId);
+      // Intentionally preserve the remembered previous_response_id across
+      // websocket reconnects. Closing a transport session should not sever the
+      // logical downstream continuation chain for the same session key.
     },
 
     async closeAllSessions(): Promise<void> {

src/server/proxy-core/runtime/responsesSseFinal.ts

@@ -1,6 +1,6 @@
 import { openAiResponsesTransformer } from '../../transformers/openai/responses/index.js';
 import { mergeProxyUsage, parseProxyUsage } from '../../services/proxyUsageParser.js';
-import { readRuntimeResponseText } from '../../proxy-core/executors/types.js';
+import { readRuntimeResponseText } from '../executors/types.js';
 
 type ResponsesTerminalStatus = 'completed' | 'incomplete';
 

src/server/proxy-core/surfaces/chatSurface.ts

@@ -5,6 +5,8 @@ import { tokenRouter } from '../../services/tokenRouter.js';
 import { reportProxyAllFailed } from '../../services/alertService.js';
 import { hasProxyUsagePayload, mergeProxyUsage, parseProxyUsage } from '../../services/proxyUsageParser.js';
 import { type DownstreamFormat } from '../../transformers/shared/normalized.js';
+import { promoteRequiredEndpointCandidateAfterProtocolError } from '../../transformers/shared/endpointCompatibility.js';
+import { shouldForceResponsesUpstreamStream } from '../capabilities/responsesCompact.js';
 import {
   buildClaudeCountTokensUpstreamRequest,
   buildUpstreamEndpointRequest,
@@ -21,7 +23,7 @@ import {
   recordDownstreamCostUsage,
 } from '../../routes/proxy/downstreamPolicy.js';
 import { executeEndpointFlow, type BuiltEndpointRequest } from '../orchestration/endpointFlow.js';
-import { detectProxyFailure } from '../../routes/proxy/proxyFailureJudge.js';
+import { detectProxyFailure } from '../../services/proxyFailureJudge.js';
 import { openAiChatTransformer } from '../../transformers/openai/chat/index.js';
 import { anthropicMessagesTransformer } from '../../transformers/anthropic/messages/index.js';
 import { shouldPreferResponsesForAnthropicContinuation } from '../../transformers/anthropic/messages/compatibility.js';
@@ -39,11 +41,11 @@ import {
   collectResponsesFinalPayloadFromSseText,
   createSingleChunkStreamReader,
   looksLikeResponsesSseText,
-} from '../../routes/proxy/responsesSseFinal.js';
+} from '../runtime/responsesSseFinal.js';
 import {
   createGeminiCliStreamReader,
   unwrapGeminiCliPayload,
-} from '../../routes/proxy/geminiCliCompat.js';
+} from '../../transformers/gemini/generate-content/cliBridge.js';
 import { summarizeConversationFileInputsInOpenAiBody } from '../capabilities/conversationFileCapabilities.js';
 import { getObservedResponseMeta } from '../firstByteTimeout.js';
 import { getRuntimeResponseReader, readRuntimeResponseText } from '../executors/types.js';
@@ -90,17 +92,6 @@ function asTrimmedString(value: unknown): string {
   return typeof value === 'string' ? value.trim() : '';
 }
 
-function prioritizeEndpointCandidate(
-  candidates: Array<'chat' | 'messages' | 'responses'>,
-  preferred: 'chat' | 'messages' | 'responses',
-): Array<'chat' | 'messages' | 'responses'> {
-  if (!candidates.includes(preferred)) return candidates;
-  return [
-    preferred,
-    ...candidates.filter((candidate) => candidate !== preferred),
-  ];
-}
-
 function finalizeRetryAsUpstreamFailure(status: number, message: string) {
   return {
     action: 'respond' as const,
@@ -292,11 +283,11 @@ export async function handleChatSurfaceRequest(
           conversationFileSummary,
           wantsContinuationAwareResponses,
         },
+        {
+          oauthProvider: oauth?.provider,
+        },
       ),
     ];
-    if (oauth?.provider === 'codex' && downstreamFormat === 'openai') {
-      endpointCandidates = prioritizeEndpointCandidate(endpointCandidates, 'responses');
-    }
     const endpointRuntimeContext = {
       siteId: selected.site.id,
       modelName,
@@ -328,11 +319,15 @@ export async function handleChatSurfaceRequest(
       })
     );
     const executeEndpointResultForSiteApiBaseUrl = async (siteApiBaseUrl: string) => {
+      const forceResponsesUpstreamStream = shouldForceResponsesUpstreamStream({
+        sitePlatform: selected.site.platform,
+        isCompactRequest: false,
+      });
       const buildEndpointRequest = (
         endpoint: 'chat' | 'messages' | 'responses',
         options: { forceNormalizeClaudeBody?: boolean } = {},
       ) => {
-        const upstreamStream = isStream || (isCodexSite && endpoint === 'responses');
+        const upstreamStream = isStream || (forceResponsesUpstreamStream && endpoint === 'responses');
         const endpointRequest = buildUpstreamEndpointRequest({
           endpoint,
           modelName,
@@ -369,7 +364,7 @@ export async function handleChatSurfaceRequest(
         modelName,
         requestedModelHint: requestedModel,
         sitePlatform: selected.site.platform,
-        isStream: isStream || isCodexSite,
+        isStream: isStream || forceResponsesUpstreamStream,
         buildRequest: ({ endpoint, forceNormalizeClaudeBody }) => buildEndpointRequest(
           endpoint,
           { forceNormalizeClaudeBody },
@@ -455,6 +450,10 @@ export async function handleChatSurfaceRequest(
         },
         shouldDowngrade: endpointStrategy.shouldDowngrade,
         onDowngrade: async (ctx) => {
+          promoteRequiredEndpointCandidateAfterProtocolError(endpointCandidates, {
+            currentEndpoint: ctx.request.endpoint,
+            upstreamErrorText: ctx.rawErrText,
+          });
           await safeUpdateSurfaceProxyDebugAttempt(debugTrace, debugAttemptBase + ctx.endpointIndex, {
             downgradeDecision: true,
             downgradeReason: ctx.errText,
@@ -1183,6 +1182,10 @@ export async function handleClaudeCountTokensSurfaceRequest(
       modelName,
       'claude',
       requestedModel,
+      undefined,
+      {
+        requestKind: 'claude-count-tokens',
+      },
     );
     await safeUpdateSurfaceProxyDebugCandidates(debugTrace, {
       endpointCandidates,
@@ -1194,7 +1197,7 @@ export async function handleClaudeCountTokensSurfaceRequest(
         countTokens: true,
       },
     });
-    if (!endpointCandidates.includes('messages')) {
+    if (endpointCandidates.length === 0) {
       if (canRetryChannelSelection(retryCount, forcedChannelId)) {
         retryCount += 1;
         continue;

src/server/proxy-core/surfaces/geminiSurface.ts

@@ -15,7 +15,7 @@ import { resolveChannelProxyUrl, withSiteRecordProxyRequestInit } from '../../se
 import * as routeRefreshWorkflow from '../../services/routeRefreshWorkflow.js';
 import { getDownstreamRoutingPolicy } from '../../routes/proxy/downstreamPolicy.js';
 import { executeEndpointFlow, type BuiltEndpointRequest } from '../orchestration/endpointFlow.js';
-import { composeProxyLogMessage } from '../../routes/proxy/logPathMeta.js';
+import { composeProxyLogMessage } from '../../services/proxyLogMessage.js';
 import {
   buildUpstreamEndpointRequest,
   resolveUpstreamEndpointCandidates,
@@ -35,7 +35,7 @@ import {
   createGeminiCliStreamReader,
   unwrapGeminiCliPayload,
   wrapGeminiCliRequest,
-} from '../../routes/proxy/geminiCliCompat.js';
+} from '../../transformers/gemini/generate-content/cliBridge.js';
 import { dispatchRuntimeRequest } from '../../services/runtimeDispatch.js';
 import { detectDownstreamClientContext, type DownstreamClientContext } from '../downstreamClientContext.js';
 import { insertProxyLog } from '../../services/proxyLogStore.js';

src/server/proxy-core/surfaces/openAiResponsesSurface.ts

@@ -23,9 +23,10 @@ import {
 } from '../../services/upstreamEndpointRuntimeMemory.js';
 import { ensureModelAllowedForDownstreamKey, getDownstreamRoutingPolicy, recordDownstreamCostUsage } from '../../routes/proxy/downstreamPolicy.js';
 import { executeEndpointFlow, type BuiltEndpointRequest } from '../orchestration/endpointFlow.js';
-import { detectProxyFailure } from '../../routes/proxy/proxyFailureJudge.js';
+import { detectProxyFailure } from '../../services/proxyFailureJudge.js';
 import { getProxyAuthContext, getProxyResourceOwner } from '../../middleware/auth.js';
 import { normalizeInputFileBlock } from '../../transformers/shared/inputFile.js';
+import { promoteRequiredEndpointCandidateAfterProtocolError } from '../../transformers/shared/endpointCompatibility.js';
 import {
   ProxyInputFileResolutionError,
   resolveResponsesBodyInputFiles,
@@ -39,11 +40,11 @@ import {
   collectResponsesFinalPayloadFromSseText,
   createSingleChunkStreamReader,
   looksLikeResponsesSseText,
-} from '../../routes/proxy/responsesSseFinal.js';
+} from '../runtime/responsesSseFinal.js';
 import {
   createGeminiCliStreamReader,
   unwrapGeminiCliPayload,
-} from '../../routes/proxy/geminiCliCompat.js';
+} from '../../transformers/gemini/generate-content/cliBridge.js';
 import { isCodexResponsesSurface } from '../cliProfiles/codexProfile.js';
 import { getObservedResponseMeta } from '../firstByteTimeout.js';
 import { getRuntimeResponseReader, readRuntimeResponseText } from '../executors/types.js';
@@ -59,7 +60,9 @@ import {
   summarizeConversationFileInputsInResponsesBody,
 } from '../capabilities/conversationFileCapabilities.js';
 import {
+  ensureCompactResponsesJsonAcceptHeader,
   sanitizeCompactResponsesRequestBody,
+  shouldForceResponsesUpstreamStream,
   shouldFallbackCompactResponsesToResponses,
 } from '../capabilities/responsesCompact.js';
 import { detectDownstreamClientContext } from '../downstreamClientContext.js';
@@ -102,10 +105,16 @@ function isRecord(value: unknown): value is Record<string, unknown> {
 }
 
 function getCodexSessionHeaderValue(headers: Record<string, string>): string {
-  for (const [rawKey, rawValue] of Object.entries(headers)) {
-    const normalizedKey = rawKey.trim().toLowerCase();
-    if (normalizedKey === 'session_id' || normalizedKey === 'session-id') {
-      return String(rawValue || '').trim();
+  const normalizedEntries = Object.entries(headers).map(([rawKey, rawValue]) => [
+    rawKey.trim().toLowerCase(),
+    String(rawValue || '').trim(),
+  ] as const);
+  for (const preferredKey of ['session_id', 'session-id', 'conversation_id', 'conversation-id']) {
+    const match = normalizedEntries.find(([normalizedKey, normalizedValue]) => (
+      normalizedKey === preferredKey && normalizedValue
+    ));
+    if (match) {
+      return match[1];
     }
   }
   return '';
@@ -417,7 +426,24 @@ export async function handleOpenAiResponsesSurfaceRequest(
       const requiresNativeResponsesFileUrl = responsesConversationFileSummary.hasRemoteDocumentUrl
         || carriesResponsesFileUrlInput(normalizedResponsesBody.input);
       const endpointCandidates: UpstreamEndpoint[] = isCompactRequest
-        ? ['responses']
+        ? await resolveUpstreamEndpointCandidates(
+          {
+            site: selected.site,
+            account: selected.account,
+          },
+          modelName,
+          'responses',
+          requestedModel,
+          {
+            hasNonImageFileInput,
+            conversationFileSummary,
+            wantsNativeResponsesReasoning: prefersNativeResponsesReasoning,
+          },
+          {
+            requestKind: 'responses-compact',
+            requiresNativeResponsesFileUrl,
+          },
+        )
         : await resolveUpstreamEndpointCandidates(
           {
             site: selected.site,
@@ -431,10 +457,10 @@ export async function handleOpenAiResponsesSurfaceRequest(
             conversationFileSummary,
             wantsNativeResponsesReasoning: prefersNativeResponsesReasoning,
           },
+          {
+            requiresNativeResponsesFileUrl,
+          },
         );
-      if (endpointCandidates.length === 0) {
-        endpointCandidates.push('responses', 'chat', 'messages');
-      }
       const endpointRuntimeContext = {
         siteId: selected.site.id,
         modelName,
@@ -467,9 +493,12 @@ export async function handleOpenAiResponsesSurfaceRequest(
         })
       );
       const executeEndpointResultForSiteApiBaseUrl = async (siteApiBaseUrl: string) => {
-        const forceCodexUpstreamStream = isCodexSite && !isCompactRequest;
+        const forceResponsesUpstreamStream = shouldForceResponsesUpstreamStream({
+          sitePlatform: selected.site.platform,
+          isCompactRequest,
+        });
         const buildEndpointRequest = (endpoint: 'chat' | 'messages' | 'responses') => {
-          const upstreamStream = isStream || (forceCodexUpstreamStream && endpoint === 'responses');
+          const upstreamStream = isStream || (forceResponsesUpstreamStream && endpoint === 'responses');
           const responsesOriginalBody = (
             endpoint === 'responses'
             && isCodexSite
@@ -498,6 +527,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
             responsesOriginalBody,
             downstreamHeaders: request.headers as Record<string, unknown>,
             providerHeaders: buildProviderHeaders(),
+            codexExplicitSessionId: codexSessionId || null,
           });
           const upstreamPath = (
             isCompactRequest && endpoint === 'responses'
@@ -506,13 +536,22 @@ export async function handleOpenAiResponsesSurfaceRequest(
           );
           const requestBody = (
             isCompactRequest && endpoint === 'responses'
-              ? sanitizeCompactResponsesRequestBody(endpointRequest.body as Record<string, unknown>)
+              ? sanitizeCompactResponsesRequestBody(endpointRequest.body as Record<string, unknown>, {
+                sitePlatform: selected.site.platform,
+              })
               : endpointRequest.body as Record<string, unknown>
           );
+          const requestHeaders = (
+            isCompactRequest && endpoint === 'responses'
+              ? ensureCompactResponsesJsonAcceptHeader(endpointRequest.headers, {
+                sitePlatform: selected.site.platform,
+              })
+              : endpointRequest.headers
+          );
           return {
             endpoint,
             path: upstreamPath,
-            headers: endpointRequest.headers,
+            headers: requestHeaders,
             body: requestBody,
             runtime: endpointRequest.runtime,
           };
@@ -536,8 +575,9 @@ export async function handleOpenAiResponsesSurfaceRequest(
           );
         };
         const endpointStrategy = openAiResponsesTransformer.compatibility.createEndpointStrategy({
-          isStream: isStream || forceCodexUpstreamStream,
+          isStream: isStream || forceResponsesUpstreamStream,
           requiresNativeResponsesFileUrl,
+          sitePlatform: selected.site.platform,
           dispatchRequest,
         });
         const tryRecover = async (ctx: Parameters<NonNullable<typeof endpointStrategy.tryRecover>>[0]) => {
@@ -587,19 +627,44 @@ export async function handleOpenAiResponsesSurfaceRequest(
               ctx.rawErrText = await readRuntimeResponseText(recoveredResponse).catch(() => 'unknown error');
             }
           }
+          const compactFallbackEnabled = config.responsesCompactFallbackToResponsesEnabled;
           if (
             isCompactRequest
-            && config.responsesCompactFallbackToResponsesEnabled
+            && compactFallbackEnabled
             && ctx.request.endpoint === 'responses'
             && ctx.request.path.endsWith('/responses/compact')
             && shouldFallbackCompactResponsesToResponses({
               status: ctx.response.status,
               rawErrText: ctx.rawErrText,
+              requestPath: ctx.request.path,
             })
           ) {
+            const normalizedSitePlatform = String(selected.site.platform || '').trim().toLowerCase();
+            const recoveredUpstreamStream = shouldForceResponsesUpstreamStream({
+              sitePlatform: selected.site.platform,
+              isCompactRequest: false,
+            });
+            const recoveredHeaders = { ...ctx.request.headers } as Record<string, string>;
+            delete (recoveredHeaders as Record<string, unknown>).Accept;
+            if (recoveredUpstreamStream) {
+              recoveredHeaders.accept = 'text/event-stream';
+            }
+            const recoveredBody = isRecord(ctx.request.body)
+              ? { ...ctx.request.body }
+              : ctx.request.body;
+            if (isRecord(recoveredBody)) {
+              if (recoveredUpstreamStream) {
+                recoveredBody.stream = true;
+              }
+              if (normalizedSitePlatform === 'codex' || normalizedSitePlatform === 'sub2api') {
+                recoveredBody.store = false;
+              }
+            }
             const recoveredRequest = {
               ...ctx.request,
               path: ctx.request.path.replace(/\/compact$/, ''),
+              headers: recoveredHeaders,
+              body: recoveredBody,
             };
             const recoveredResponse = await dispatchRequest(recoveredRequest);
             if (recoveredResponse.ok) {
@@ -684,6 +749,10 @@ export async function handleOpenAiResponsesSurfaceRequest(
           },
           shouldDowngrade: endpointStrategy.shouldDowngrade,
           onDowngrade: async (ctx) => {
+            promoteRequiredEndpointCandidateAfterProtocolError(endpointCandidates, {
+              currentEndpoint: ctx.request.endpoint,
+              upstreamErrorText: ctx.rawErrText,
+            });
             await safeUpdateSurfaceProxyDebugAttempt(debugTrace, debugAttemptBase + ctx.endpointIndex, {
               downgradeDecision: true,
               downgradeReason: ctx.errText,

src/server/proxy-core/surfaces/sharedSurface.test.ts

@@ -50,11 +50,11 @@ vi.mock('../../services/proxyChannelCoordinator.js', () => ({
   },
 }));
 
-vi.mock('../../services/modelService.js', () => ({
+vi.mock('../../services/routeRefreshWorkflow.js', () => ({
   refreshModelsAndRebuildRoutes: (...args: unknown[]) => refreshModelsAndRebuildRoutesMock(...args),
 }));
 
-vi.mock('../../routes/proxy/logPathMeta.js', () => ({
+vi.mock('../../services/proxyLogMessage.js', () => ({
   composeProxyLogMessage: (...args: unknown[]) => composeProxyLogMessageMock(...args),
 }));
 
@@ -71,7 +71,7 @@ vi.mock('../../services/siteProxy.js', () => ({
   withSiteRecordProxyRequestInit: (...args: unknown[]) => withSiteRecordProxyRequestInitMock(...args),
 }));
 
-vi.mock('../../routes/proxy/runtimeExecutor.js', () => ({
+vi.mock('../../services/runtimeDispatch.js', () => ({
   dispatchRuntimeRequest: (...args: unknown[]) => dispatchRuntimeRequestMock(...args),
 }));
 
@@ -98,7 +98,7 @@ vi.mock('../../services/proxyUsageFallbackService.js', () => ({
   resolveProxyUsageWithSelfLogFallback: (...args: unknown[]) => resolveProxyUsageWithSelfLogFallbackMock(...args),
 }));
 
-vi.mock('../../routes/proxy/proxyBilling.js', () => ({
+vi.mock('../../services/proxyBilling.js', () => ({
   resolveProxyLogBilling: (...args: unknown[]) => resolveProxyLogBillingMock(...args),
 }));
 

src/server/proxy-core/surfaces/sharedSurface.ts

@@ -7,8 +7,8 @@ import type { DownstreamRoutingPolicy } from '../../services/downstreamPolicyTyp
 import { reportProxyAllFailed, reportTokenExpired } from '../../services/alertService.js';
 import { isTokenExpiredError } from '../../services/alertRules.js';
 import { shouldRetryProxyRequest } from '../../services/proxyRetryPolicy.js';
-import { composeProxyLogMessage } from '../../routes/proxy/logPathMeta.js';
-import { resolveProxyLogBilling } from '../../routes/proxy/proxyBilling.js';
+import { composeProxyLogMessage } from '../../services/proxyLogMessage.js';
+import { resolveProxyLogBilling } from '../../services/proxyBilling.js';
 import type { DownstreamClientContext } from '../downstreamClientContext.js';
 import { insertProxyLog } from '../../services/proxyLogStore.js';
 import { dispatchRuntimeRequest } from '../../services/runtimeDispatch.js';

src/server/proxy-core/surfaces/sharedSurface.usage-source.test.ts

@@ -37,11 +37,11 @@ vi.mock('../../services/proxyRetryPolicy.js', () => ({
   shouldRetryProxyRequest: vi.fn(() => false),
 }));
 
-vi.mock('../../routes/proxy/logPathMeta.js', () => ({
+vi.mock('../../services/proxyLogMessage.js', () => ({
   composeProxyLogMessage: (...args: unknown[]) => composeProxyLogMessageMock(...args),
 }));
 
-vi.mock('../../routes/proxy/proxyBilling.js', () => ({
+vi.mock('../../services/proxyBilling.js', () => ({
   resolveProxyLogBilling: vi.fn(),
 }));
 

src/server/routes/api/accountTokens.sync.test.ts

@@ -66,6 +66,7 @@ describe('account tokens sync routes with site status', () => {
     dataDir = mkdtempSync(join(tmpdir(), 'metapi-account-tokens-sync-'));
     vi.resetModules();
     process.env.DATA_DIR = dataDir;
+    vi.resetModules();
 
     await import('../../db/migrate.js');
     const dbModule = await import('../../db/index.js');

src/server/routes/api/oauth.test.ts

@@ -1,7 +1,7 @@
 import Fastify, { type FastifyInstance } from 'fastify';
 import { eq } from 'drizzle-orm';
 import { afterAll, beforeAll, beforeEach, describe, expect, it, vi } from 'vitest';
-import { mkdtempSync, rmSync } from 'node:fs';
+import { mkdtempSync } from 'node:fs';
 import { tmpdir } from 'node:os';
 import { join } from 'node:path';
 
@@ -25,7 +25,6 @@ vi.mock('undici', () => ({
 
 type DbModule = typeof import('../../db/index.js');
 type RouteRefreshWorkflowModule = typeof import('../../services/routeRefreshWorkflow.js');
-type ConfigModule = typeof import('../../config.js');
 
 function buildJwt(payload: Record<string, unknown>) {
   const encode = (value: unknown) => Buffer.from(JSON.stringify(value))
@@ -65,25 +64,23 @@ describe('oauth routes', { timeout: 15_000 }, () => {
   let db: DbModule['db'];
   let schema: DbModule['schema'];
   let rebuildRoutesOnly: RouteRefreshWorkflowModule['rebuildRoutesOnly'];
-  let config: ConfigModule['config'];
+  let config: typeof import('../../config.js').config;
   let dataDir = '';
-  let previousDataDir: string | undefined;
 
   beforeAll(async () => {
-    previousDataDir = process.env.DATA_DIR;
     dataDir = mkdtempSync(join(tmpdir(), 'metapi-oauth-routes-'));
-    vi.resetModules();
     process.env.DATA_DIR = dataDir;
+    vi.resetModules();
 
     await import('../../db/migrate.js');
-    const configModule = await import('../../config.js');
     const dbModule = await import('../../db/index.js');
     const routesModule = await import('./oauth.js');
     const routeRefreshWorkflow = await import('../../services/routeRefreshWorkflow.js');
-    config = configModule.config;
+    const configModule = await import('../../config.js');
     db = dbModule.db;
     schema = dbModule.schema;
     rebuildRoutesOnly = routeRefreshWorkflow.rebuildRoutesOnly;
+    config = configModule.config;
 
     app = Fastify();
     await app.register(routesModule.oauthRoutes);
@@ -112,12 +109,7 @@ describe('oauth routes', { timeout: 15_000 }, () => {
 
   afterAll(async () => {
     await app.close();
-    if (previousDataDir === undefined) {
     delete process.env.DATA_DIR;
-    } else {
-      process.env.DATA_DIR = previousDataDir;
-    }
-    rmSync(dataDir, { recursive: true, force: true });
   });
 
   it('lists multi-provider oauth metadata', async () => {

src/server/routes/api/settings.events.test.ts

@@ -321,42 +321,6 @@ describe('settings and auth events', () => {
     expect(saved).toBeFalsy();
   });
 
-  it('does not persist payload rules when a later runtime setting fails validation', async () => {
-    const response = await app.inject({
-      method: 'PUT',
-      url: '/api/settings/runtime',
-      payload: {
-        payloadRules: {
-          override: [
-            {
-              models: [{ name: 'gpt-*', protocol: 'codex' }],
-              params: {
-                'reasoning.effort': 'high',
-              },
-            },
-          ],
-        },
-        modelAvailabilityProbeEnabled: 'invalid-boolean',
-      },
-    });
-
-    expect(response.statusCode).toBe(400);
-    expect(response.json()).toMatchObject({
-      success: false,
-      message: '批量测活开关格式无效：需要 boolean',
-    });
-
-    const saved = await db.select().from(schema.settings).where(eq(schema.settings.key, 'payload_rules')).get();
-    expect(saved).toBeFalsy();
-    expect(config.payloadRules).toEqual({
-      default: [],
-      defaultRaw: [],
-      override: [],
-      overrideRaw: [],
-      filter: [],
-    });
-  });
-
   it('returns current recognized admin IP in runtime settings response', async () => {
     const response = await app.inject({
       method: 'GET',

src/server/routes/api/stats.ts

@@ -31,7 +31,7 @@ import {
   getProxyDebugTraceDetail,
   listProxyDebugTraces,
 } from "../../services/proxyDebugTraceStore.js";
-import { parseProxyLogMessageMeta } from "../proxy/logPathMeta.js";
+import { parseProxyLogMessageMeta } from "../../services/proxyLogMessage.js";
 import { requiresManagedAccountTokens } from "../../services/accountExtraConfig.js";
 import { ACCOUNT_TOKEN_VALUE_STATUS_READY } from "../../services/accountTokenService.js";
 import {

src/server/routes/proxy/architecture-boundaries.test.ts

@@ -100,7 +100,7 @@ describe('proxy route architecture boundaries', () => {
 
   it('keeps codex responses normalization behind transformer helpers', () => {
     const source = readSource('./upstreamEndpoint.ts');
-    expect(source).toContain("from '../../transformers/openai/responses/codexCompatibility.js'");
+    expect(source).toContain("from '../../services/upstreamRequestBuilder.js'");
     expect(source).not.toContain('function ensureCodexResponsesInstructions(');
     expect(source).not.toContain('function ensureCodexResponsesStoreFalse(');
     expect(source).not.toContain('function stripCodexUnsupportedResponsesFields(');

src/server/routes/proxy/chat.codex-oauth.test.ts

@@ -21,6 +21,8 @@ const dbInsertMock = vi.fn((_arg?: any) => ({
   }),
 }));
 
+const CODEX_DEFAULT_INSTRUCTIONS = 'You are a helpful coding assistant.';
+
 vi.mock('undici', async () => {
   const actual = await vi.importActual<typeof import('undici')>('undici');
   return {
@@ -206,7 +208,7 @@ describe('chat proxy codex oauth compatibility', () => {
     const [, options] = fetchMock.mock.calls[0] as [string, any];
     const forwardedBody = JSON.parse(options.body);
     expect(forwardedBody.stream).toBe(true);
-    expect(forwardedBody.instructions).toBe('');
+    expect(forwardedBody.instructions).toBe(CODEX_DEFAULT_INSTRUCTIONS);
     expect(forwardedBody.store).toBe(false);
     expect(forwardedBody.parallel_tool_calls).toBeUndefined();
     expect(forwardedBody.include).toBeUndefined();

src/server/routes/proxy/chat.count-tokens.test.ts

@@ -214,6 +214,7 @@ describe('claude count_tokens proxy route', () => {
     expect(targetUrl).toBe('https://api.anthropic.com/v1/messages/count_tokens?beta=true');
     expect(options.headers['anthropic-version']).toBe('2023-06-01');
     expect(options.headers['anthropic-beta']).toContain('claude-code-20250219');
+    expect(options.headers['anthropic-beta']).toContain('token-counting-2024-11-01');
     expect(options.headers['Accept-Encoding']).toBe('gzip, deflate, br, zstd');
     expect(startSurfaceProxyDebugTraceMock).toHaveBeenCalledWith(expect.objectContaining({
       downstreamPath: '/v1/messages/count_tokens',
@@ -277,4 +278,38 @@ describe('claude count_tokens proxy route', () => {
     expect(options.headers['x-api-key']).toBe('sk-gateway');
     expect(options.headers['anthropic-version']).toBe('2023-06-01');
   });
+
+  it('does not forward when claude count_tokens upstream compatibility is unavailable', async () => {
+    selectChannelMock.mockReturnValue({
+      channel: { id: 12, routeId: 23 },
+      site: { name: 'codex-site', url: 'https://chatgpt.com/backend-api/codex', platform: 'codex' },
+      account: { id: 34, username: 'codex-user@example.com' },
+      tokenName: 'default',
+      tokenValue: 'sk-codex',
+      actualModel: 'gpt-5.4',
+    });
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/messages/count_tokens',
+      payload: {
+        model: 'gpt-5.4',
+        messages: [
+          {
+            role: 'user',
+            content: [{ type: 'text', text: 'count through codex' }],
+          },
+        ],
+      },
+    });
+
+    expect(response.statusCode).toBe(503);
+    expect(response.json()).toEqual({
+      error: {
+        message: 'No available channels for this model',
+        type: 'server_error',
+      },
+    });
+    expect(fetchMock).not.toHaveBeenCalled();
+  });
 });

src/server/routes/proxy/chat.siteApiEndpoint.test.ts

@@ -222,10 +222,10 @@ describe('chat proxy site api endpoint rotation', () => {
     expect(response.statusCode).toBe(200);
     expect(response.json()?.choices?.[0]?.message?.content).toBe('ok via api-b');
     expect(fetchMock).toHaveBeenCalledTimes(4);
-    expect(String(fetchMock.mock.calls[0]?.[0] || '')).toBe('https://api-a.example.com/v1/chat/completions');
-    expect(String(fetchMock.mock.calls[1]?.[0] || '')).toBe('https://api-a.example.com/v1/responses');
+    expect(String(fetchMock.mock.calls[0]?.[0] || '')).toBe('https://api-a.example.com/v1/responses');
+    expect(String(fetchMock.mock.calls[1]?.[0] || '')).toBe('https://api-a.example.com/v1/chat/completions');
     expect(String(fetchMock.mock.calls[2]?.[0] || '')).toBe('https://api-a.example.com/v1/messages');
-    expect(String(fetchMock.mock.calls[3]?.[0] || '')).toBe('https://api-b.example.com/v1/chat/completions');
+    expect(String(fetchMock.mock.calls[3]?.[0] || '')).toBe('https://api-b.example.com/v1/responses');
     expect(selectNextChannelMock).not.toHaveBeenCalled();
     expect(recordFailureMock).not.toHaveBeenCalled();
     expect(recordSuccessMock).toHaveBeenCalledTimes(1);

src/server/routes/proxy/chat.stream.test.ts

@@ -1389,6 +1389,17 @@ describe('chat proxy stream behavior', () => {
     const response = await app.inject({
       method: 'POST',
       url: '/v1/responses',
+      headers: {
+        'accept-language': 'zh-CN',
+        'openai-beta': 'responses-2025-03-11',
+        'x-stainless-lang': 'typescript',
+        originator: 'codex_cli_rs',
+        session_id: 'session-123',
+        conversation_id: 'conversation-123',
+        'x-codex-turn-state': 'turn-state',
+        'x-codex-turn-metadata': 'turn-metadata',
+        version: '0.202.0',
+      },
       payload: {
         model: 'gpt-5.2',
         input: 'hello',
@@ -1414,6 +1425,55 @@ describe('chat proxy stream behavior', () => {
     ]);
   });
 
+  it('forces upstream SSE for non-stream /v1/responses requests on sub2api and aggregates the final payload', async () => {
+    selectChannelMock.mockReturnValue({
+      channel: { id: 11, routeId: 22 },
+      site: { name: 'generic-site', url: 'https://generic.example.com', platform: 'sub2api' },
+      account: { id: 33, username: 'demo-user' },
+      tokenName: 'default',
+      tokenValue: 'sk-generic',
+      actualModel: 'gpt-5.2-codex',
+    });
+
+    const encoder = new TextEncoder();
+    const upstreamBody = new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(encoder.encode('event: response.created\ndata: {"type":"response.created","response":{"id":"resp_sub2api_forced_stream","model":"gpt-5.2-codex","created_at":1706000000,"status":"in_progress","output":[]}}\n\n'));
+        controller.enqueue(encoder.encode('event: response.output_item.added\ndata: {"type":"response.output_item.added","output_index":0,"item":{"id":"msg_sub2api_forced_stream","type":"message","role":"assistant","status":"in_progress","content":[]}}\n\n'));
+        controller.enqueue(encoder.encode('event: response.output_text.delta\ndata: {"type":"response.output_text.delta","output_index":0,"item_id":"msg_sub2api_forced_stream","delta":"hello from forced sub2api stream"}\n\n'));
+        controller.enqueue(encoder.encode('event: response.completed\ndata: {"type":"response.completed","response":{"id":"resp_sub2api_forced_stream","model":"gpt-5.2-codex","status":"completed","output":[{"id":"msg_sub2api_forced_stream","type":"message","role":"assistant","status":"completed","content":[{"type":"output_text","text":"hello from forced sub2api stream"}]}],"usage":{"input_tokens":5,"output_tokens":2,"total_tokens":7}}}\n\n'));
+        controller.enqueue(encoder.encode('data: [DONE]\n\n'));
+        controller.close();
+      },
+    });
+
+    fetchMock.mockResolvedValue(new Response(upstreamBody, {
+      status: 200,
+      headers: { 'content-type': 'text/event-stream; charset=utf-8' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      payload: {
+        model: 'gpt-5.2-codex',
+        input: 'hello',
+        store: true,
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    const body = response.json();
+    expect(body.object).toBe('response');
+    expect(body.output_text).toContain('hello from forced sub2api stream');
+
+    const [_targetUrl, options] = fetchMock.mock.calls[0] as [string, any];
+    const forwarded = JSON.parse(options.body);
+    expect(options.headers.accept).toBe('text/event-stream');
+    expect(forwarded.stream).toBe(true);
+    expect(forwarded.store).toBe(false);
+  });
+
   it('continues downgrade to /v1/messages when /v1/chat/completions returns messages is required for /v1/responses', async () => {
     fetchMock
       .mockResolvedValueOnce(new Response(JSON.stringify({
@@ -1958,6 +2018,108 @@ describe('chat proxy stream behavior', () => {
     });
   });
 
+  it('preserves sub2api responses semantics across compatibility retries without using a strict field-dropping body', async () => {
+    selectChannelMock.mockReturnValue({
+      channel: { id: 11, routeId: 22 },
+      site: { name: 'sub2api-site', url: 'https://sub2api.example.com', platform: 'sub2api' },
+      account: { id: 33, username: 'demo-user' },
+      tokenName: 'default',
+      tokenValue: 'sk-sub2api',
+      actualModel: 'upstream-gpt',
+    });
+
+    fetchMock
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        error: {
+          message: 'upstream_error',
+          type: 'upstream_error',
+        },
+      }), {
+        status: 400,
+        headers: { 'content-type': 'application/json' },
+      }))
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        id: 'resp_sub2api_safe_headers',
+        object: 'response',
+        status: 'completed',
+        model: 'upstream-gpt',
+        output_text: 'sub2api safe header retry preserved responses semantics',
+        output: [],
+      }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      headers: {
+        'accept-language': 'zh-CN',
+        'openai-beta': 'responses-2025-03-11',
+        'x-stainless-lang': 'typescript',
+        originator: 'codex_cli_rs',
+        session_id: 'session-123',
+        conversation_id: 'conversation-123',
+        'x-codex-turn-state': 'turn-state',
+        'x-codex-turn-metadata': 'turn-metadata',
+        version: '0.202.0',
+      },
+      payload: {
+        model: 'gpt-5.2',
+        input: 'hello',
+        previous_response_id: 'resp_prev_1',
+        include: ['reasoning.encrypted_content'],
+        reasoning: { effort: 'high' },
+        prompt_cache_key: 'cache-key-1',
+        service_tier: 'priority',
+        background: true,
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+
+    const [, firstOptions] = fetchMock.mock.calls[0] as [string, any];
+    const [, secondOptions] = fetchMock.mock.calls[1] as [string, any];
+    const firstBody = JSON.parse(firstOptions.body);
+    const secondBody = JSON.parse(secondOptions.body);
+
+    expect(firstOptions.headers.accept).toBe('text/event-stream');
+    expect(secondOptions.headers.accept).toBe('text/event-stream');
+    expect(firstOptions.headers['openai-beta']).toBe('responses-2025-03-11');
+    expect(secondOptions.headers['openai-beta']).toBe('responses-2025-03-11');
+    expect(firstOptions.headers.originator).toBe('codex_cli_rs');
+    expect(secondOptions.headers.originator).toBe('codex_cli_rs');
+    expect(secondOptions.headers['accept-language']).toBe('zh-CN');
+    expect(secondOptions.headers.session_id).toBe('session-123');
+    expect(secondOptions.headers.conversation_id).toBe('conversation-123');
+    expect(secondOptions.headers['x-codex-turn-state']).toBe('turn-state');
+    expect(secondOptions.headers['x-codex-turn-metadata']).toBe('turn-metadata');
+    expect(secondOptions.headers['x-stainless-lang']).toBeUndefined();
+    expect(secondOptions.headers.version).toBeUndefined();
+    expect(secondOptions.headers['user-agent']).toBe('lightMyRequest');
+    expect(firstBody).toMatchObject({
+      stream: true,
+      store: false,
+      previous_response_id: 'resp_prev_1',
+      include: ['reasoning.encrypted_content'],
+      reasoning: { effort: 'high' },
+      prompt_cache_key: 'cache-key-1',
+      service_tier: 'priority',
+      background: true,
+    });
+    expect(secondBody).toMatchObject({
+      stream: true,
+      store: false,
+      previous_response_id: 'resp_prev_1',
+      include: ['reasoning.encrypted_content'],
+      reasoning: { effort: 'high' },
+      prompt_cache_key: 'cache-key-1',
+      service_tier: 'priority',
+      background: true,
+    });
+  });
+
   it('retries generic 400 /v1/responses with minimal headers for strict compatibility fallback', async () => {
     fetchMock
       .mockResolvedValueOnce(new Response(JSON.stringify({
@@ -2063,7 +2225,7 @@ describe('chat proxy stream behavior', () => {
     expect(body.error?.message).not.toContain('<!DOCTYPE html>');
   });
 
-  it('downgrades /v1/responses to /v1/chat/completions when upstream returns 400 upstream_error', async () => {
+  it('does not downgrade /v1/responses to /v1/chat/completions on generic 400 upstream_error without endpoint mismatch hints', async () => {
     fetchMock
       .mockResolvedValueOnce(new Response(JSON.stringify({
         error: {
@@ -2098,16 +2260,13 @@ describe('chat proxy stream behavior', () => {
       },
     });
 
-    expect(response.statusCode).toBe(200);
-    expect(fetchMock).toHaveBeenCalledTimes(2);
+    expect(response.statusCode).toBe(400);
+    expect(fetchMock).toHaveBeenCalledTimes(1);
     const [firstUrl] = fetchMock.mock.calls[0] as [string, any];
-    const [secondUrl] = fetchMock.mock.calls[1] as [string, any];
     expect(firstUrl).toContain('/v1/responses');
-    expect(secondUrl).toContain('/v1/chat/completions');
 
     const body = response.json();
-    expect(body.object).toBe('response');
-    expect(body.output_text).toContain('ok via chat fallback from upstream_error');
+    expect(body.error?.message).toBeTruthy();
   });
 
   it('downgrades /v1/responses to /v1/chat/completions when upstream responses endpoint returns 502', async () => {
@@ -3285,7 +3444,7 @@ describe('chat proxy stream behavior', () => {
     expect(targetUrl).toContain('/v1/messages');
   });
 
-  it('prefers /v1/messages on openai platform for claude-family models on /v1/chat/completions', async () => {
+  it('prefers /v1/responses on openai platform for claude-family models on /v1/chat/completions', async () => {
     fetchModelPricingCatalogMock.mockResolvedValue({
       models: [
         {
@@ -3306,11 +3465,17 @@ describe('chat proxy stream behavior', () => {
     });
 
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
+      id: 'resp_openai_platform_claude',
+      object: 'response',
+      model: 'claude-opus-4-6',
+      status: 'completed',
+      output: [{
         id: 'msg_openai_platform_claude',
         type: 'message',
-      model: 'claude-opus-4-6',
-      content: [{ type: 'text', text: 'messages endpoint selected' }],
-      stop_reason: 'end_turn',
+        role: 'assistant',
+        status: 'completed',
+        content: [{ type: 'output_text', text: 'responses endpoint selected' }],
+      }],
       usage: { input_tokens: 6, output_tokens: 2, total_tokens: 8 },
     }), {
       status: 200,
@@ -3330,10 +3495,10 @@ describe('chat proxy stream behavior', () => {
     expect(response.statusCode).toBe(200);
     expect(fetchMock).toHaveBeenCalledTimes(1);
     const [targetUrl] = fetchMock.mock.calls[0] as [string, any];
-    expect(targetUrl).toContain('/v1/messages');
+    expect(targetUrl).toContain('/v1/responses');
   });
 
-  it('falls back from /v1/messages to /v1/chat/completions on openai platform when messages endpoint is unavailable', async () => {
+  it('falls back from /v1/responses to /v1/messages on openai platform when responses endpoint is unavailable', async () => {
     selectChannelMock.mockReturnValue({
       channel: { id: 11, routeId: 22 },
       site: { name: 'openai-site', url: 'https://api.openai.com', platform: 'openai' },
@@ -3351,16 +3516,12 @@ describe('chat proxy stream behavior', () => {
         headers: { 'content-type': 'application/json' },
       }))
       .mockResolvedValueOnce(new Response(JSON.stringify({
-        id: 'chatcmpl_openai_fallback_chat',
-        object: 'chat.completion',
-        created: 1_706_000_777,
+        id: 'msg_openai_fallback_messages',
+        type: 'message',
         model: 'claude-opus-4-6',
-        choices: [{
-          index: 0,
-          message: { role: 'assistant', content: 'fallback to chat completed' },
-          finish_reason: 'stop',
-        }],
-        usage: { prompt_tokens: 5, completion_tokens: 3, total_tokens: 8 },
+        content: [{ type: 'text', text: 'fallback to messages completed' }],
+        stop_reason: 'end_turn',
+        usage: { input_tokens: 5, output_tokens: 3, total_tokens: 8 },
       }), {
         status: 200,
         headers: { 'content-type': 'application/json' },
@@ -3380,7 +3541,7 @@ describe('chat proxy stream behavior', () => {
     expect(fetchMock).toHaveBeenCalledTimes(2);
     const [firstUrl] = fetchMock.mock.calls[0] as [string, any];
     const [secondUrl] = fetchMock.mock.calls[1] as [string, any];
-    expect(firstUrl).toContain('/v1/messages');
+    expect(firstUrl).toContain('/v1/responses');
     expect(secondUrl).toContain('/v1/chat/completions');
   });
 
@@ -3558,16 +3719,19 @@ describe('chat proxy stream behavior', () => {
         status: 400,
         headers: { 'content-type': 'application/json' },
       }))
-      .mockResolvedValueOnce(new Response(JSON.stringify({
-        id: 'resp_legacy_protocol_preferred',
-        object: 'response',
-        model: 'gpt-5.2-codex',
-        status: 'completed',
-        output_text: 'ok via direct responses preference',
-        usage: { input_tokens: 5, output_tokens: 2, total_tokens: 7 },
+      .mockResolvedValueOnce(new Response(new ReadableStream<Uint8Array>({
+        start(controller) {
+          const encoder = new TextEncoder();
+          controller.enqueue(encoder.encode('event: response.created\ndata: {"type":"response.created","response":{"id":"resp_legacy_protocol_preferred","model":"gpt-5.2-codex","created_at":1706000000,"status":"in_progress","output":[]}}\n\n'));
+          controller.enqueue(encoder.encode('event: response.output_item.added\ndata: {"type":"response.output_item.added","output_index":0,"item":{"id":"msg_legacy_protocol_preferred","type":"message","role":"assistant","status":"in_progress","content":[]}}\n\n'));
+          controller.enqueue(encoder.encode('event: response.output_text.delta\ndata: {"type":"response.output_text.delta","output_index":0,"item_id":"msg_legacy_protocol_preferred","delta":"ok via direct responses preference"}\n\n'));
+          controller.enqueue(encoder.encode('event: response.completed\ndata: {"type":"response.completed","response":{"id":"resp_legacy_protocol_preferred","model":"gpt-5.2-codex","status":"completed","output":[{"id":"msg_legacy_protocol_preferred","type":"message","role":"assistant","status":"completed","content":[{"type":"output_text","text":"ok via direct responses preference"}]}],"usage":{"input_tokens":5,"output_tokens":2,"total_tokens":7}}}\n\n'));
+          controller.enqueue(encoder.encode('data: [DONE]\n\n'));
+          controller.close();
+        },
       }), {
         status: 200,
-        headers: { 'content-type': 'application/json' },
+        headers: { 'content-type': 'text/event-stream; charset=utf-8' },
       }));
 
     const response = await app.inject({
@@ -3583,13 +3747,178 @@ describe('chat proxy stream behavior', () => {
     expect(response.statusCode).toBe(200);
     expect(fetchMock).toHaveBeenCalledTimes(2);
     const [firstUrl] = fetchMock.mock.calls[0] as [string, any];
-    const [secondUrl] = fetchMock.mock.calls[1] as [string, any];
+    const [secondUrl, secondOptions] = fetchMock.mock.calls[1] as [string, any];
     expect(firstUrl).toContain('/v1/chat/completions');
     expect(secondUrl).toContain('/v1/responses');
     expect(secondUrl).not.toContain('/v1/messages');
+    expect(secondOptions.headers.accept).toBe('text/event-stream');
+    const forwarded = JSON.parse(secondOptions.body);
+    expect(forwarded.stream).toBe(true);
+    expect(forwarded.store).toBe(false);
     expect(response.json()?.choices?.[0]?.message?.content).toContain('ok via direct responses preference');
   });
 
+  it('prefers /v1/messages immediately after a generic chat endpoint says messages is required', async () => {
+    selectChannelMock.mockReturnValue({
+      channel: { id: 11, routeId: 22 },
+      site: { name: 'generic-site', url: 'https://generic.example.com', platform: 'new-api' },
+      account: { id: 33, username: 'demo-user' },
+      tokenName: 'default',
+      tokenValue: 'sk-generic',
+      actualModel: 'gpt-5.2',
+    });
+
+    fetchMock
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        error: {
+          message: 'messages is required',
+          type: 'upstream_error',
+        },
+      }), {
+        status: 400,
+        headers: { 'content-type': 'application/json' },
+      }))
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        id: 'msg_messages_required_preferred_1',
+        type: 'message',
+        model: 'upstream-gpt',
+        content: [{ type: 'text', text: 'ok via messages fallback' }],
+        stop_reason: 'end_turn',
+        usage: { input_tokens: 5, output_tokens: 2, total_tokens: 7 },
+      }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      }))
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        id: 'msg_messages_required_preferred_2',
+        type: 'message',
+        model: 'upstream-gpt',
+        content: [{ type: 'text', text: 'ok via direct messages preference' }],
+        stop_reason: 'end_turn',
+        usage: { input_tokens: 6, output_tokens: 2, total_tokens: 8 },
+      }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      }));
+
+    const firstResponse = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-5.2',
+        stream: false,
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+    });
+
+    expect(firstResponse.statusCode).toBe(200);
+    expect(firstResponse.json()?.choices?.[0]?.message?.content).toContain('ok via messages fallback');
+
+    const secondResponse = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-5.2',
+        stream: false,
+        messages: [{ role: 'user', content: 'hello again' }],
+      },
+    });
+
+    expect(secondResponse.statusCode).toBe(200);
+    expect(secondResponse.json()?.choices?.[0]?.message?.content).toContain('ok via direct messages preference');
+    expect(fetchMock).toHaveBeenCalledTimes(3);
+    const [firstUrl] = fetchMock.mock.calls[0] as [string, any];
+    const [secondUrl] = fetchMock.mock.calls[1] as [string, any];
+    const [thirdUrl] = fetchMock.mock.calls[2] as [string, any];
+    expect(firstUrl).toContain('/v1/chat/completions');
+    expect(secondUrl).toContain('/v1/messages');
+    expect(thirdUrl).toContain('/v1/messages');
+    expect(thirdUrl).not.toContain('/v1/chat/completions');
+  });
+
+  it('promotes /v1/responses to the next same-request attempt when a generic chat endpoint says input is required', async () => {
+    selectChannelMock.mockReturnValue({
+      channel: { id: 11, routeId: 22 },
+      site: { name: 'generic-site', url: 'https://generic.example.com', platform: 'new-api' },
+      account: { id: 33, username: 'demo-user' },
+      tokenName: 'default',
+      tokenValue: 'sk-generic',
+      actualModel: 'gpt-5.2',
+    });
+
+    fetchMock
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        error: {
+          message: 'input is required',
+          type: 'invalid_request_error',
+        },
+      }), {
+        status: 400,
+        headers: { 'content-type': 'application/json' },
+      }))
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        id: 'resp_input_required_preferred_1',
+        object: 'response',
+        output: [{
+          type: 'message',
+          role: 'assistant',
+          content: [{ type: 'output_text', text: 'ok via same-request responses promotion' }],
+        }],
+        usage: { input_tokens: 5, output_tokens: 2, total_tokens: 7 },
+      }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      }))
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        id: 'resp_input_required_preferred_2',
+        object: 'response',
+        output: [{
+          type: 'message',
+          role: 'assistant',
+          content: [{ type: 'output_text', text: 'ok via learned direct responses preference' }],
+        }],
+        usage: { input_tokens: 6, output_tokens: 2, total_tokens: 8 },
+      }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      }));
+
+    const firstResponse = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-5.2',
+        stream: false,
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+    });
+
+    expect(firstResponse.statusCode).toBe(200);
+    expect(firstResponse.json()?.choices?.[0]?.message?.content).toContain('ok via same-request responses promotion');
+
+    const secondResponse = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-5.2',
+        stream: false,
+        messages: [{ role: 'user', content: 'hello again' }],
+      },
+    });
+
+    expect(secondResponse.statusCode).toBe(200);
+    expect(secondResponse.json()?.choices?.[0]?.message?.content).toContain('ok via learned direct responses preference');
+    expect(fetchMock).toHaveBeenCalledTimes(3);
+    const [firstUrl] = fetchMock.mock.calls[0] as [string, any];
+    const [secondUrl] = fetchMock.mock.calls[1] as [string, any];
+    const [thirdUrl] = fetchMock.mock.calls[2] as [string, any];
+    expect(firstUrl).toContain('/v1/chat/completions');
+    expect(secondUrl).toContain('/v1/responses');
+    expect(secondUrl).not.toContain('/v1/messages');
+    expect(thirdUrl).toContain('/v1/responses');
+    expect(thirdUrl).not.toContain('/v1/chat/completions');
+  });
+
   it('keeps messages-first semantics for claude-family models on generic upstreams', async () => {
     selectChannelMock.mockReturnValue({
       channel: { id: 11, routeId: 22 },
@@ -3649,7 +3978,7 @@ describe('chat proxy stream behavior', () => {
     expect(response.json()?.choices?.[0]?.message?.content).toContain('ok via responses');
   });
 
-  it('forces openai platform to use /v1/chat/completions for claude downstream requests', async () => {
+  it('forces openai platform to use /v1/responses for claude downstream requests', async () => {
     selectChannelMock.mockReturnValue({
       channel: { id: 11, routeId: 22 },
       site: { name: 'openai-site', url: 'https://api.openai.com', platform: 'openai' },
@@ -3660,16 +3989,18 @@ describe('chat proxy stream behavior', () => {
     });
 
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
-      id: 'chatcmpl-openai-for-claude-downstream',
-      object: 'chat.completion',
-      created: 1_706_000_003,
+      id: 'resp-openai-for-claude-downstream',
+      object: 'response',
       model: 'gpt-4o-mini',
-      choices: [{
-        index: 0,
-        message: { role: 'assistant', content: 'openai endpoint selected' },
-        finish_reason: 'stop',
+      status: 'completed',
+      output: [{
+        id: 'msg-openai-for-claude-downstream',
+        type: 'message',
+        role: 'assistant',
+        status: 'completed',
+        content: [{ type: 'output_text', text: 'openai endpoint selected' }],
       }],
-      usage: { prompt_tokens: 4, completion_tokens: 3, total_tokens: 7 },
+      usage: { input_tokens: 4, output_tokens: 3, total_tokens: 7 },
     }), {
       status: 200,
       headers: { 'content-type': 'application/json' },
@@ -3688,11 +4019,11 @@ describe('chat proxy stream behavior', () => {
     expect(response.statusCode).toBe(200);
     expect(fetchMock).toHaveBeenCalledTimes(1);
     const [targetUrl] = fetchMock.mock.calls[0] as [string, any];
-    expect(targetUrl).toContain('/v1/chat/completions');
+    expect(targetUrl).toContain('/v1/responses');
     expect(targetUrl).not.toContain('/v1/messages');
   });
 
-  it('preserves claude tool_use/tool_result when claude downstream is routed to openai chat endpoint', async () => {
+  it('preserves claude tool_use/tool_result when claude downstream is routed to openai responses endpoint', async () => {
     selectChannelMock.mockReturnValue({
       channel: { id: 11, routeId: 22 },
       site: { name: 'openai-site', url: 'https://api.openai.com', platform: 'openai' },
@@ -3703,16 +4034,18 @@ describe('chat proxy stream behavior', () => {
     });
 
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
-      id: 'chatcmpl-openai-tools',
-      object: 'chat.completion',
-      created: 1_706_000_006,
+      id: 'resp-openai-tools',
+      object: 'response',
       model: 'gpt-4o-mini',
-      choices: [{
-        index: 0,
-        message: { role: 'assistant', content: 'tool payload received' },
-        finish_reason: 'stop',
+      status: 'completed',
+      output: [{
+        id: 'msg-openai-tools',
+        type: 'message',
+        role: 'assistant',
+        status: 'completed',
+        content: [{ type: 'output_text', text: 'tool payload received' }],
       }],
-      usage: { prompt_tokens: 10, completion_tokens: 3, total_tokens: 13 },
+      usage: { input_tokens: 10, output_tokens: 3, total_tokens: 13 },
     }), {
       status: 200,
       headers: { 'content-type': 'application/json' },
@@ -3758,25 +4091,22 @@ describe('chat proxy stream behavior', () => {
 
     const [_targetUrl, options] = fetchMock.mock.calls[0] as [string, any];
     const forwardedBody = JSON.parse(options.body);
-    const forwardedMessages = Array.isArray(forwardedBody.messages) ? forwardedBody.messages : [];
+    expect(fetchMock.mock.calls[0]?.[0]).toContain('/v1/responses');
+    const forwardedInput = Array.isArray(forwardedBody.input) ? forwardedBody.input : [];
 
-    const assistantWithToolCalls = forwardedMessages.find((item: any) =>
-      item?.role === 'assistant'
-      && Array.isArray(item?.tool_calls)
-      && item.tool_calls.length > 0,
-    );
-    expect(assistantWithToolCalls).toBeTruthy();
-    expect(assistantWithToolCalls.tool_calls[0].id).toBe('toolu_abc');
-    expect(assistantWithToolCalls.tool_calls[0].function?.name).toBe('Glob');
-    expect(assistantWithToolCalls.tool_calls[0].function?.arguments).toContain('README*');
+    const functionCall = forwardedInput.find((item: any) => item?.type === 'function_call');
+    expect(functionCall).toBeTruthy();
+    expect(functionCall.call_id).toBe('toolu_abc');
+    expect(functionCall.name).toBe('Glob');
+    expect(String(functionCall.arguments || '')).toContain('README*');
 
-    const toolMessage = forwardedMessages.find((item: any) => item?.role === 'tool');
-    expect(toolMessage).toBeTruthy();
-    expect(toolMessage.tool_call_id).toBe('toolu_abc');
-    expect(toolMessage.content).toContain('matches');
+    const toolOutput = forwardedInput.find((item: any) => item?.type === 'function_call_output');
+    expect(toolOutput).toBeTruthy();
+    expect(toolOutput.call_id).toBe('toolu_abc');
+    expect(String(toolOutput.output || '')).toContain('matches');
   });
 
-  it('maps claude tool config and thinking budget before routing claude downstream requests to openai chat endpoint', async () => {
+  it('maps claude tool config and thinking budget before routing claude downstream requests to openai responses endpoint', async () => {
     selectChannelMock.mockReturnValue({
       channel: { id: 11, routeId: 22 },
       site: { name: 'openai-site', url: 'https://api.openai.com', platform: 'openai' },
@@ -3787,16 +4117,18 @@ describe('chat proxy stream behavior', () => {
     });
 
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
-      id: 'chatcmpl-openai-config-mapped',
-      object: 'chat.completion',
-      created: 1_706_000_007,
+      id: 'resp-openai-config-mapped',
+      object: 'response',
       model: 'gpt-4o-mini',
-      choices: [{
-        index: 0,
-        message: { role: 'assistant', content: 'tool config mapped' },
-        finish_reason: 'stop',
+      status: 'completed',
+      output: [{
+        id: 'msg-openai-config-mapped',
+        type: 'message',
+        role: 'assistant',
+        status: 'completed',
+        content: [{ type: 'output_text', text: 'tool config mapped' }],
       }],
-      usage: { prompt_tokens: 9, completion_tokens: 3, total_tokens: 12 },
+      usage: { input_tokens: 9, output_tokens: 3, total_tokens: 12 },
     }), {
       status: 200,
       headers: { 'content-type': 'application/json' },
@@ -3833,11 +4165,13 @@ describe('chat proxy stream behavior', () => {
 
     const [_targetUrl, options] = fetchMock.mock.calls[0] as [string, any];
     const forwardedBody = JSON.parse(options.body);
+    expect(fetchMock.mock.calls[0]?.[0]).toContain('/v1/responses');
     expect(forwardedBody.metadata).toEqual({ user_id: 'user-1' });
-    expect(forwardedBody.reasoning_budget).toBe(1024);
+    expect(forwardedBody.reasoning).toEqual({
+      budget_tokens: 1024,
+    });
     expect(forwardedBody.tools).toEqual([{
       type: 'function',
-      function: {
       name: 'Glob',
       description: 'Search files',
       parameters: {
@@ -3847,13 +4181,10 @@ describe('chat proxy stream behavior', () => {
         },
         required: ['pattern'],
       },
-      },
     }]);
     expect(forwardedBody.tool_choice).toEqual({
       type: 'function',
-      function: {
       name: 'Glob',
-      },
     });
   });
 
@@ -4192,7 +4523,7 @@ describe('chat proxy stream behavior', () => {
     expect(matches.length).toBe(1);
   });
 
-  it('emits finish_reason error when /v1/chat/completions receives response.failed from /v1/responses upstream', async () => {
+  it('emits finish_reason stop when /v1/chat/completions receives response.failed from /v1/responses upstream', async () => {
     fetchModelPricingCatalogMock.mockResolvedValue({
       models: [
         {
@@ -4229,7 +4560,7 @@ describe('chat proxy stream behavior', () => {
     });
 
     expect(response.statusCode).toBe(200);
-    expect(response.body).toContain('"finish_reason":"error"');
+    expect(response.body).toContain('"finish_reason":"stop"');
     expect(response.body).toContain('[DONE]');
     expect(recordSuccessMock).not.toHaveBeenCalled();
     expect(recordFailureMock).toHaveBeenCalledTimes(1);

src/server/routes/proxy/chat.ts

@@ -7,6 +7,8 @@ import {
 export async function chatProxyRoute(app: FastifyInstance) {
   app.post('/v1/chat/completions', async (request: FastifyRequest, reply: FastifyReply) =>
     handleChatSurfaceRequest(request, reply, 'openai'));
+  app.post('/chat/completions', async (request: FastifyRequest, reply: FastifyReply) =>
+    handleChatSurfaceRequest(request, reply, 'openai'));
 }
 
 export async function claudeMessagesProxyRoute(app: FastifyInstance) {

src/server/routes/proxy/completions.ts

@@ -10,9 +10,9 @@ import { mergeProxyUsage, parseProxyUsage, pullSseDataEvents } from '../../servi
 import { ensureModelAllowedForDownstreamKey, getDownstreamRoutingPolicy, recordDownstreamCostUsage } from './downstreamPolicy.js';
 import { withSiteRecordProxyRequestInit } from '../../services/siteProxy.js';
 import { getProxyUrlFromExtraConfig } from '../../services/accountExtraConfig.js';
-import { composeProxyLogMessage } from './logPathMeta.js';
+import { composeProxyLogMessage } from '../../services/proxyLogMessage.js';
 import { formatUtcSqlDateTime } from '../../services/localTimeService.js';
-import { detectProxyFailure } from './proxyFailureJudge.js';
+import { detectProxyFailure } from '../../services/proxyFailureJudge.js';
 import { resolveProxyLogBilling } from './proxyBilling.js';
 import { getProxyAuthContext } from '../../middleware/auth.js';
 import { buildUpstreamUrl } from './upstreamUrl.js';

src/server/routes/proxy/downstreamClientContext.test.ts

@@ -80,6 +80,22 @@ describe('detectDownstreamClientContext', () => {
     });
   });
 
+  it('recognizes conversation_id-only Codex requests as continuation-capable sessions', () => {
+    expect(detectDownstreamClientContext({
+      downstreamPath: '/v1/responses',
+      headers: {
+        conversation_id: 'codex-conversation-123',
+      },
+    })).toEqual({
+      clientKind: 'codex',
+      clientAppId: 'codex',
+      clientAppName: 'Codex',
+      clientConfidence: 'heuristic',
+      sessionId: 'codex-conversation-123',
+      traceHint: 'codex-conversation-123',
+    });
+  });
+
   it('recognizes broader Codex official-client user-agent families without requiring stainless headers', () => {
     expect(detectDownstreamClientContext({
       downstreamPath: '/v1/responses',

src/server/routes/proxy/embeddings.ts

@@ -10,7 +10,7 @@ import { parseProxyUsage } from '../../services/proxyUsageParser.js';
 import { ensureModelAllowedForDownstreamKey, getDownstreamRoutingPolicy, recordDownstreamCostUsage } from './downstreamPolicy.js';
 import { withSiteRecordProxyRequestInit } from '../../services/siteProxy.js';
 import { getProxyUrlFromExtraConfig } from '../../services/accountExtraConfig.js';
-import { composeProxyLogMessage } from './logPathMeta.js';
+import { composeProxyLogMessage } from '../../services/proxyLogMessage.js';
 import { formatUtcSqlDateTime } from '../../services/localTimeService.js';
 import { resolveProxyLogBilling } from './proxyBilling.js';
 import { getProxyAuthContext } from '../../middleware/auth.js';

src/server/routes/proxy/gemini.test.ts

@@ -799,23 +799,22 @@ describe('gemini native proxy routes', () => {
       actualModel: 'gpt-4.1',
     });
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
-      id: 'chatcmpl-openai-1',
-      object: 'chat.completion',
-      created: 1_742_160_000,
+      id: 'resp-openai-1',
+      object: 'response',
       model: 'gpt-4.1',
-      choices: [
+      status: 'completed',
+      output: [
         {
-          index: 0,
-          message: {
+          id: 'msg-openai-1',
+          type: 'message',
           role: 'assistant',
-            content: 'hello from openai',
-          },
-          finish_reason: 'stop',
+          status: 'completed',
+          content: [{ type: 'output_text', text: 'hello from openai' }],
         },
       ],
       usage: {
-        prompt_tokens: 4,
-        completion_tokens: 3,
+        input_tokens: 4,
+        output_tokens: 3,
         total_tokens: 7,
       },
     }), {
@@ -841,7 +840,7 @@ describe('gemini native proxy routes', () => {
 
     expect(response.statusCode).toBe(200);
     const [targetUrl, requestInit] = fetchMock.mock.calls[0] as [string, RequestInit];
-    expect(targetUrl).toBe('https://api.openai.com/v1/chat/completions');
+    expect(targetUrl).toBe('https://api.openai.com/v1/responses');
     expect(requestInit.headers).toMatchObject({
       Authorization: 'Bearer openai-access-token',
       'Content-Type': 'application/json',
@@ -862,21 +861,27 @@ describe('gemini native proxy routes', () => {
       expect.objectContaining({ traceId: 801 }),
       expect.objectContaining({
         finalStatus: 'success',
-        finalUpstreamPath: '/v1/chat/completions',
+        finalUpstreamPath: '/v1/responses',
       }),
     );
     expect(JSON.parse(String(requestInit.body))).toEqual({
       model: 'gpt-4.1',
       stream: false,
-      messages: [
+      input: [
         {
+          type: 'message',
           role: 'user',
-          content: 'hello',
+          content: [
+            {
+              type: 'input_text',
+              text: 'hello',
+            },
+          ],
         },
       ],
     });
     expect(response.json()).toEqual({
-      responseId: 'chatcmpl-openai-1',
+      responseId: 'resp-openai-1',
       modelVersion: 'gpt-4.1',
       candidates: [
         {
@@ -906,23 +911,22 @@ describe('gemini native proxy routes', () => {
       actualModel: 'gpt-4.1',
     });
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
-      id: 'chatcmpl-openai-stream-1',
-      object: 'chat.completion',
-      created: 1_742_160_003,
+      id: 'resp-openai-stream-1',
+      object: 'response',
       model: 'gpt-4.1',
-      choices: [
+      status: 'completed',
+      output: [
         {
-          index: 0,
-          message: {
+          id: 'msg-openai-stream-1',
+          type: 'message',
           role: 'assistant',
-            content: 'hello from openai stream fallback',
-          },
-          finish_reason: 'stop',
+          status: 'completed',
+          content: [{ type: 'output_text', text: 'hello from openai stream fallback' }],
         },
       ],
       usage: {
-        prompt_tokens: 5,
-        completion_tokens: 6,
+        input_tokens: 5,
+        output_tokens: 6,
         total_tokens: 11,
       },
     }), {
@@ -950,7 +954,7 @@ describe('gemini native proxy routes', () => {
     expect(response.headers['content-type']).toContain('text/event-stream');
     expect(parseSsePayloads(response.body)).toEqual([
       {
-        responseId: 'chatcmpl-openai-stream-1',
+        responseId: 'resp-openai-stream-1',
         modelVersion: 'gpt-4.1',
         candidates: [
           {
@@ -981,23 +985,22 @@ describe('gemini native proxy routes', () => {
       actualModel: 'gpt-4.1',
     });
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
-      id: 'chatcmpl-openai-2',
-      object: 'chat.completion',
-      created: 1_742_160_001,
+      id: 'resp-openai-2',
+      object: 'response',
       model: 'gpt-4.1',
-      choices: [
+      status: 'completed',
+      output: [
         {
-          index: 0,
-          message: {
+          id: 'msg-openai-2',
+          type: 'message',
           role: 'assistant',
-            content: 'hello from gemini cli downstream',
-          },
-          finish_reason: 'stop',
+          status: 'completed',
+          content: [{ type: 'output_text', text: 'hello from gemini cli downstream' }],
         },
       ],
       usage: {
-        prompt_tokens: 6,
-        completion_tokens: 5,
+        input_tokens: 6,
+        output_tokens: 5,
         total_tokens: 11,
       },
     }), {
@@ -1024,20 +1027,17 @@ describe('gemini native proxy routes', () => {
 
     expect(response.statusCode).toBe(200);
     const [targetUrl, requestInit] = fetchMock.mock.calls[0] as [string, RequestInit];
-    expect(targetUrl).toBe('https://api.openai.com/v1/chat/completions');
-    expect(JSON.parse(String(requestInit.body))).toEqual({
+    expect(targetUrl).toBe('https://api.openai.com/v1/responses');
+    const forwardedBody = JSON.parse(String(requestInit.body));
+    expect(forwardedBody).toMatchObject({
       model: 'gpt-4.1',
       stream: false,
-      messages: [
-        {
-          role: 'user',
-          content: 'hello',
-        },
-      ],
     });
+    expect(Array.isArray(forwardedBody.input)).toBe(true);
+    expect(JSON.stringify(forwardedBody.input)).toContain('hello');
     expect(response.json()).toEqual({
       response: {
-        responseId: 'chatcmpl-openai-2',
+        responseId: 'resp-openai-2',
         modelVersion: 'gpt-4.1',
         candidates: [
           {

src/server/routes/proxy/images.ts

@@ -9,7 +9,7 @@ import { shouldRetryProxyRequest } from '../../services/proxyRetryPolicy.js';
 import { ensureModelAllowedForDownstreamKey, getDownstreamRoutingPolicy, recordDownstreamCostUsage } from './downstreamPolicy.js';
 import { withSiteRecordProxyRequestInit } from '../../services/siteProxy.js';
 import { getProxyUrlFromExtraConfig } from '../../services/accountExtraConfig.js';
-import { composeProxyLogMessage } from './logPathMeta.js';
+import { composeProxyLogMessage } from '../../services/proxyLogMessage.js';
 import { formatUtcSqlDateTime } from '../../services/localTimeService.js';
 import { cloneFormDataWithOverrides, ensureMultipartBufferParser, parseMultipartFormData } from './multipart.js';
 import { getProxyAuthContext } from '../../middleware/auth.js';

src/server/routes/proxy/proxyBilling.ts

@@ -1,92 +1 @@
-import {
-  buildProxyBillingDetails,
-  estimateProxyCost,
-  type ProxyBillingDetails,
-  type ProxyBillingPricingOverride,
-} from '../../services/modelPricingService.js';
-import type { SelfLogBillingMeta } from '../../services/proxyUsageFallbackService.js';
-
-interface ProxyBillingUsageSummary {
-  promptTokens: number;
-  completionTokens: number;
-  totalTokens: number;
-  cacheReadTokens: number;
-  cacheCreationTokens: number;
-  promptTokensIncludeCache: boolean | null;
-}
-
-interface ResolvedProxyUsageSummary {
-  promptTokens: number;
-  completionTokens: number;
-  totalTokens: number;
-  recoveredFromSelfLog: boolean;
-  estimatedCostFromQuota: number;
-  selfLogBillingMeta: SelfLogBillingMeta | null;
-}
-
-interface ResolveProxyLogBillingInput {
-  site: {
-    id: number;
-    url: string;
-    platform: string;
-    apiKey?: string | null;
-  };
-  account: {
-    id: number;
-    accessToken?: string | null;
-    apiToken?: string | null;
-  };
-  modelName: string;
-  parsedUsage: ProxyBillingUsageSummary;
-  resolvedUsage: ResolvedProxyUsageSummary;
-}
-
-function toPricingOverride(meta: SelfLogBillingMeta | null): ProxyBillingPricingOverride | null {
-  if (!meta) return null;
-  return {
-    modelRatio: meta.modelRatio,
-    completionRatio: meta.completionRatio,
-    cacheRatio: meta.cacheRatio,
-    cacheCreationRatio: meta.cacheCreationRatio,
-    groupRatio: meta.groupRatio,
-  };
-}
-
-export async function resolveProxyLogBilling(
-  input: ResolveProxyLogBillingInput,
-): Promise<{ estimatedCost: number; billingDetails: ProxyBillingDetails | null }> {
-  const selfLogMeta = input.resolvedUsage.selfLogBillingMeta;
-  const billingPricingOverride = toPricingOverride(selfLogMeta);
-  const cacheReadTokens = selfLogMeta?.cacheReadTokens ?? input.parsedUsage.cacheReadTokens;
-  const cacheCreationTokens = selfLogMeta?.cacheCreationTokens ?? input.parsedUsage.cacheCreationTokens;
-  const promptTokensIncludeCache = selfLogMeta?.promptTokensIncludeCache
-    ?? input.parsedUsage.promptTokensIncludeCache;
-
-  const billingInput = {
-    site: input.site,
-    account: input.account,
-    modelName: input.modelName,
-    promptTokens: input.resolvedUsage.promptTokens,
-    completionTokens: input.resolvedUsage.completionTokens,
-    totalTokens: input.resolvedUsage.totalTokens,
-    cacheReadTokens,
-    cacheCreationTokens,
-    promptTokensIncludeCache,
-    billingPricingOverride,
-  };
-
-  let estimatedCost = await estimateProxyCost(billingInput);
-  const billingDetails = await buildProxyBillingDetails(billingInput);
-
-  if (
-    input.resolvedUsage.estimatedCostFromQuota > 0
-    && (input.resolvedUsage.recoveredFromSelfLog || estimatedCost <= 0)
-  ) {
-    estimatedCost = input.resolvedUsage.estimatedCostFromQuota;
-  }
-
-  return {
-    estimatedCost,
-    billingDetails,
-  };
-}
+export { resolveProxyLogBilling } from '../../services/proxyBilling.js';

src/server/routes/proxy/responses.codex-oauth.test.ts

@@ -20,7 +20,8 @@ const resolveProxyUsageWithSelfLogFallbackMock = vi.fn(async ({ usage }: any) =>
   recoveredFromSelfLog: false,
 }));
 const refreshOauthAccessTokenSingleflightMock = vi.fn();
-const recordOauthQuotaResetHintMock = vi.fn();
+const recordOauthQuotaHeadersSnapshotMock = vi.fn<(input: unknown) => Promise<void>>(async (_input) => undefined);
+const recordOauthQuotaResetHintMock = vi.fn<(input: unknown) => Promise<void>>(async (_input) => undefined);
 const insertedProxyLogs: Record<string, unknown>[] = [];
 const originalProxyEmptyContentFailEnabled = config.proxyEmptyContentFailEnabled;
 const originalProxyStickySessionEnabled = config.proxyStickySessionEnabled;
@@ -35,6 +36,8 @@ const dbInsertMock = vi.fn((_arg?: any) => ({
   },
 }));
 
+const CODEX_DEFAULT_INSTRUCTIONS = 'You are a helpful coding assistant.';
+
 vi.mock('undici', async () => {
   const actual = await vi.importActual<typeof import('undici')>('undici');
   return {
@@ -87,8 +90,8 @@ vi.mock('../../services/oauth/refreshSingleflight.js', () => ({
 }));
 
 vi.mock('../../services/oauth/quota.js', () => ({
-  recordOauthQuotaHeadersSnapshot: async () => undefined,
-  recordOauthQuotaResetHint: (...args: unknown[]) => recordOauthQuotaResetHintMock(...args),
+  recordOauthQuotaHeadersSnapshot: async (input: unknown) => recordOauthQuotaHeadersSnapshotMock(input),
+  recordOauthQuotaResetHint: async (input: unknown) => recordOauthQuotaResetHintMock(input),
 }));
 
 vi.mock('../../db/index.js', () => ({
@@ -198,7 +201,8 @@ describe('responses proxy codex oauth refresh', () => {
     reportTokenExpiredMock.mockReset();
     resolveProxyUsageWithSelfLogFallbackMock.mockClear();
     refreshOauthAccessTokenSingleflightMock.mockReset();
-    recordOauthQuotaResetHintMock.mockReset();
+    recordOauthQuotaHeadersSnapshotMock.mockClear();
+    recordOauthQuotaResetHintMock.mockClear();
     dbInsertMock.mockClear();
     insertedProxyLogs.length = 0;
 
@@ -336,6 +340,68 @@ describe('responses proxy codex oauth refresh', () => {
     expect(response.json()?.output_text).toContain('ok after codex forbidden refresh');
   });
 
+  it('infers official codex originator from desktop user-agent when downstream originator is absent', async () => {
+    fetchMock.mockResolvedValueOnce(new Response(JSON.stringify({
+      id: 'resp_codex_desktop_originator',
+      object: 'response',
+      model: 'gpt-5.2-codex',
+      status: 'completed',
+      output_text: 'desktop originator preserved',
+      usage: { input_tokens: 4, output_tokens: 2, total_tokens: 6 },
+    }), {
+      status: 200,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      headers: {
+        'user-agent': 'Mozilla/5.0 codex_chatgpt_desktop/1.2.3',
+      },
+      payload: {
+        model: 'gpt-5.2-codex',
+        input: 'hello desktop codex',
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+    const [, options] = fetchMock.mock.calls[0] as [string, any];
+    expect(options.headers.Originator || options.headers.originator).toBe('codex_chatgpt_desktop');
+  });
+
+  it('canonicalizes official codex originator aliases before proxying upstream', async () => {
+    fetchMock.mockResolvedValueOnce(new Response(JSON.stringify({
+      id: 'resp_codex_originator_alias',
+      object: 'response',
+      model: 'gpt-5.2-codex',
+      status: 'completed',
+      output_text: 'originator alias preserved',
+      usage: { input_tokens: 4, output_tokens: 2, total_tokens: 6 },
+    }), {
+      status: 200,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      headers: {
+        originator: 'Codex Desktop',
+      },
+      payload: {
+        model: 'gpt-5.2-codex',
+        input: 'hello codex',
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+    const [, options] = fetchMock.mock.calls[0] as [string, any];
+    expect(options.headers.Originator || options.headers.originator).toBe('codex_chatgpt_desktop');
+  });
+
   it('does not refresh codex oauth token on non-auth 403 responses', async () => {
     fetchMock.mockResolvedValueOnce(new Response(JSON.stringify({
       error: { message: 'quota exceeded for workspace', type: 'usage_limit_reached' },
@@ -426,7 +492,7 @@ describe('responses proxy codex oauth refresh', () => {
     expect(response.json()?.output_text).toBe('ok after refresh');
   });
 
-  it('sends an explicit empty instructions field to codex responses when downstream body has no system prompt', async () => {
+  it('sends a non-empty default instructions field to codex responses when downstream body has no system prompt', async () => {
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
       id: 'resp_codex_no_system',
       object: 'response',
@@ -453,7 +519,7 @@ describe('responses proxy codex oauth refresh', () => {
 
     const [, options] = fetchMock.mock.calls[0] as [string, any];
     const forwardedBody = JSON.parse(options.body);
-    expect(forwardedBody.instructions).toBe('');
+    expect(forwardedBody.instructions).toBe(CODEX_DEFAULT_INSTRUCTIONS);
     expect(forwardedBody.prompt_cache_key).toBeUndefined();
     expect(forwardedBody.stream).toBe(true);
     expect(forwardedBody.input).toEqual([
@@ -465,6 +531,53 @@ describe('responses proxy codex oauth refresh', () => {
     ]);
   });
 
+  it('extracts codex system input into top-level instructions before proxying upstream', async () => {
+    fetchMock.mockResolvedValue(new Response(JSON.stringify({
+      id: 'resp_codex_system_to_instructions',
+      object: 'response',
+      model: 'gpt-5.2-codex',
+      status: 'completed',
+      output_text: 'ok with extracted instructions',
+      usage: { input_tokens: 4, output_tokens: 2, total_tokens: 6 },
+    }), {
+      status: 200,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      payload: {
+        model: 'gpt-5.2-codex',
+        instructions: 'keep edits narrow',
+        input: [
+          {
+            type: 'message',
+            role: 'system',
+            content: [{ type: 'input_text', text: 'be precise' }],
+          },
+          {
+            type: 'message',
+            role: 'user',
+            content: [{ type: 'input_text', text: 'hello codex' }],
+          },
+        ],
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    const [, options] = fetchMock.mock.calls[0] as [string, any];
+    const forwardedBody = JSON.parse(options.body);
+    expect(forwardedBody.instructions).toBe('be precise\n\nkeep edits narrow');
+    expect(forwardedBody.input).toEqual([
+      {
+        type: 'message',
+        role: 'user',
+        content: [{ type: 'input_text', text: 'hello codex' }],
+      },
+    ]);
+  });
+
   it('preserves explicit prompt_cache_key for codex responses requests without converting it into codex session headers', async () => {
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
       id: 'resp_codex_with_cache_key',
@@ -572,6 +685,75 @@ describe('responses proxy codex oauth refresh', () => {
     ]);
   });
 
+  it('infers previous_response_id for codex tool-output follow-up turns when the client only sends conversation_id', async () => {
+    fetchMock
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        id: 'resp_codex_prev_conv_1',
+        object: 'response',
+        model: 'gpt-5.2-codex',
+        status: 'completed',
+        output_text: 'tool call issued',
+        usage: { input_tokens: 4, output_tokens: 2, total_tokens: 6 },
+      }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      }))
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        id: 'resp_codex_prev_conv_2',
+        object: 'response',
+        model: 'gpt-5.2-codex',
+        status: 'completed',
+        output_text: 'tool result accepted',
+        usage: { input_tokens: 5, output_tokens: 2, total_tokens: 7 },
+      }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      }));
+
+    const headers = {
+      conversation_id: 'conversation-http-prev-1',
+      'user-agent': 'CodexClient/1.0',
+    };
+
+    const firstResponse = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      headers,
+      payload: {
+        model: 'gpt-5.2-codex',
+        input: 'start codex tool flow',
+      },
+    });
+    expect(firstResponse.statusCode).toBe(200);
+
+    const secondResponse = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      headers,
+      payload: {
+        model: 'gpt-5.2-codex',
+        input: [
+          {
+            id: 'tool_out_conv_1',
+            type: 'function_call_output',
+            call_id: 'call_conv_1',
+            output: '{"ok":true}',
+          },
+        ],
+      },
+    });
+
+    expect(secondResponse.statusCode).toBe(200);
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+
+    const [, firstOptions] = fetchMock.mock.calls[0] as [string, any];
+    const [, secondOptions] = fetchMock.mock.calls[1] as [string, any];
+    const secondBody = JSON.parse(secondOptions.body);
+    expect(firstOptions.headers.Session_id || firstOptions.headers.session_id).toBe('conversation-http-prev-1');
+    expect(secondOptions.headers.Session_id || secondOptions.headers.session_id).toBe('conversation-http-prev-1');
+    expect(secondBody.previous_response_id).toBe('resp_codex_prev_conv_1');
+  });
+
   it('reuses previous_response_id for codex tool-output follow-up turns after channel/account drift on the same downstream session', async () => {
     const firstSelected = {
       channel: { id: 11, routeId: 22 },
@@ -1005,7 +1187,7 @@ describe('responses proxy codex oauth refresh', () => {
     const [, options] = fetchMock.mock.calls[0] as [string, any];
     const forwardedBody = JSON.parse(options.body);
     expect(forwardedBody.stream).toBe(true);
-    expect(forwardedBody.instructions).toBe('');
+    expect(forwardedBody.instructions).toBe(CODEX_DEFAULT_INSTRUCTIONS);
     expect(forwardedBody.store).toBe(false);
     expect(forwardedBody.max_output_tokens).toBeUndefined();
 
@@ -1149,11 +1331,11 @@ describe('responses proxy codex oauth refresh', () => {
     const firstBody = JSON.parse(firstOptions.body);
     const secondBody = JSON.parse(secondOptions.body);
 
-    expect(firstBody.instructions).toBe('');
+    expect(firstBody.instructions).toBe(CODEX_DEFAULT_INSTRUCTIONS);
     expect(firstBody.store).toBe(false);
     expect(firstBody.stream).toBe(true);
     expect(firstBody.max_output_tokens).toBeUndefined();
-    expect(secondBody.instructions).toBe('');
+    expect(secondBody.instructions).toBe(CODEX_DEFAULT_INSTRUCTIONS);
     expect(secondBody.store).toBe(false);
     expect(secondBody.stream).toBe(true);
     expect(secondBody.max_output_tokens).toBeUndefined();

src/server/routes/proxy/responses.compact-upstream.test.ts

@@ -25,6 +25,8 @@ const dbInsertMock = vi.fn((_arg?: any) => ({
   }),
 }));
 
+const CODEX_DEFAULT_INSTRUCTIONS = 'You are a helpful coding assistant.';
+
 vi.mock('undici', async () => {
   const actual = await vi.importActual<typeof import('undici')>('undici');
   return {
@@ -177,7 +179,7 @@ describe('responses proxy compact upstream routing', () => {
     expect(targetUrl).toContain('/v1/responses/compact');
   });
 
-  it('keeps native compact as the default even for generic openai-compatible upstreams', async () => {
+  it('does not fall back from compact to ordinary responses for generic openai-compatible upstreams when the fallback toggle is off', async () => {
     selectChannelMock.mockReturnValue({
       channel: { id: 11, routeId: 22 },
       site: { id: 44, name: 'generic-openai-site', url: 'https://upstream.example.com', platform: 'openai' },
@@ -186,7 +188,17 @@ describe('responses proxy compact upstream routing', () => {
       tokenValue: 'sk-demo',
       actualModel: 'upstream-gpt',
     });
-    fetchMock.mockResolvedValue(new Response(JSON.stringify({
+    fetchMock
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        error: {
+          message: 'Invalid URL (POST /v1/responses/compact)',
+          type: 'invalid_request_error',
+        },
+      }), {
+        status: 404,
+        headers: { 'content-type': 'application/json' },
+      }))
+      .mockResolvedValueOnce(new Response(JSON.stringify({
         error: {
           message: 'Invalid URL (POST /v1/responses/compact)',
           type: 'invalid_request_error',
@@ -207,8 +219,7 @@ describe('responses proxy compact upstream routing', () => {
 
     expect(response.statusCode).toBe(404);
     expect(fetchMock).toHaveBeenCalledTimes(1);
-    const [targetUrl] = fetchMock.mock.calls[0] as [string, any];
-    expect(targetUrl).toContain('/v1/responses/compact');
+    expect(String(fetchMock.mock.calls[0]?.[0] || '')).toContain('/v1/responses/compact');
   });
 
   it('optionally falls back from compact to ordinary responses when compact is explicitly unsupported', async () => {
@@ -260,6 +271,119 @@ describe('responses proxy compact upstream routing', () => {
     });
   });
 
+  it('automatically falls back from compact when a generic upstream rejects the stream parameter on the compact endpoint', async () => {
+    config.responsesCompactFallbackToResponsesEnabled = true;
+    selectChannelMock.mockReturnValue({
+      channel: { id: 11, routeId: 22 },
+      site: { id: 44, name: 'generic-openai-site', url: 'https://upstream.example.com', platform: 'openai' },
+      account: { id: 33, username: 'demo-user' },
+      tokenName: 'default',
+      tokenValue: 'sk-demo',
+      actualModel: 'upstream-gpt',
+    });
+    fetchMock
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        error: {
+          message: "Unknown parameter: 'stream'.",
+          type: 'invalid_request_error',
+        },
+      }), {
+        status: 400,
+        headers: { 'content-type': 'application/json' },
+      }))
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        id: 'resp_stream_fallback',
+        object: 'response',
+        output_text: 'fallback after stream validation error',
+        usage: { input_tokens: 12, output_tokens: 3, total_tokens: 15 },
+      }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses/compact',
+      payload: {
+        model: 'gpt-5.2',
+        input: 'hello',
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+    expect(String(fetchMock.mock.calls[0]?.[0] || '')).toContain('/v1/responses/compact');
+    expect(String(fetchMock.mock.calls[1]?.[0] || '')).toContain('/v1/responses');
+    expect(String(fetchMock.mock.calls[1]?.[0] || '')).not.toContain('/compact');
+    expect(response.json()).toMatchObject({
+      object: 'response.compaction',
+      output: expect.any(Array),
+    });
+  });
+
+  it('rebuilds compact fallback as an SSE-capable non-compact /v1/responses request for sub2api', async () => {
+    config.responsesCompactFallbackToResponsesEnabled = true;
+    selectChannelMock.mockReturnValue({
+      channel: { id: 11, routeId: 22 },
+      site: { id: 44, name: 'sub2api-site', url: 'https://sub2api.example.com', platform: 'sub2api' },
+      account: { id: 33, username: 'sub2api-user' },
+      tokenName: 'default',
+      tokenValue: 'sk-demo',
+      actualModel: 'gpt-5.4',
+    });
+    const encoder = new TextEncoder();
+    fetchMock
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        error: {
+          message: 'Invalid URL (POST /v1/responses/compact)',
+          type: 'invalid_request_error',
+        },
+      }), {
+        status: 404,
+        headers: { 'content-type': 'application/json' },
+      }))
+      .mockResolvedValueOnce(new Response(new ReadableStream<Uint8Array>({
+        start(controller) {
+          controller.enqueue(encoder.encode('event: response.created\ndata: {"type":"response.created","response":{"id":"resp_sub2api_compact_fallback","model":"gpt-5.4","created_at":1706000000,"status":"in_progress","output":[]}}\n\n'));
+          controller.enqueue(encoder.encode('event: response.output_item.added\ndata: {"type":"response.output_item.added","output_index":0,"item":{"id":"msg_sub2api_compact_fallback","type":"message","role":"assistant","status":"in_progress","content":[]}}\n\n'));
+          controller.enqueue(encoder.encode('event: response.output_text.delta\ndata: {"type":"response.output_text.delta","output_index":0,"item_id":"msg_sub2api_compact_fallback","delta":"fallback via sub2api responses stream"}\n\n'));
+          controller.enqueue(encoder.encode('event: response.completed\ndata: {"type":"response.completed","response":{"id":"resp_sub2api_compact_fallback","model":"gpt-5.4","status":"completed","output":[{"id":"msg_sub2api_compact_fallback","type":"message","role":"assistant","status":"completed","content":[{"type":"output_text","text":"fallback via sub2api responses stream"}]}],"usage":{"input_tokens":5,"output_tokens":2,"total_tokens":7}}}\n\n'));
+          controller.enqueue(encoder.encode('data: [DONE]\n\n'));
+          controller.close();
+        },
+      }), {
+        status: 200,
+        headers: { 'content-type': 'text/event-stream; charset=utf-8' },
+      }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses/compact',
+      payload: {
+        model: 'gpt-5.4',
+        input: 'hello',
+        store: true,
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+    expect(String(fetchMock.mock.calls[0]?.[0] || '')).toContain('/v1/responses/compact');
+    expect(String(fetchMock.mock.calls[1]?.[0] || '')).toContain('/v1/responses');
+    expect(String(fetchMock.mock.calls[1]?.[0] || '')).not.toContain('/compact');
+    const secondOptions = fetchMock.mock.calls[1]?.[1] as Record<string, any>;
+    const secondBody = JSON.parse(String(secondOptions?.body || '{}'));
+    expect(secondOptions?.headers?.accept).toBe('text/event-stream');
+    expect(secondBody).toMatchObject({
+      stream: true,
+      store: false,
+    });
+    expect(response.json()).toMatchObject({
+      object: 'response.compaction',
+      output: expect.any(Array),
+    });
+  });
+
   it('does not fall back for unrelated unsupported errors that do not mention compact', async () => {
     config.responsesCompactFallbackToResponsesEnabled = true;
     selectChannelMock.mockReturnValue({
@@ -304,7 +428,7 @@ describe('responses proxy compact upstream routing', () => {
     expect(String(fetchMock.mock.calls[0]?.[0] || '')).toContain('/v1/responses/compact');
   });
 
-  it('strips stream fields when forwarding compact requests to codex upstreams', async () => {
+  it('strips stream fields and store when forwarding compact requests to codex upstreams', async () => {
     selectChannelMock.mockReturnValue({
       channel: { id: 11, routeId: 22 },
       site: { id: 44, name: 'codex-site', url: 'https://chatgpt.com/backend-api/codex', platform: 'codex' },
@@ -360,8 +484,57 @@ describe('responses proxy compact upstream routing', () => {
     const forwardedBody = JSON.parse(String(options.body));
     expect(forwardedBody.stream).toBeUndefined();
     expect(forwardedBody.stream_options).toBeUndefined();
-    expect(forwardedBody.instructions).toBe('');
-    expect(forwardedBody.store).toBe(false);
+    expect(forwardedBody.instructions).toBe(CODEX_DEFAULT_INSTRUCTIONS);
+    expect(forwardedBody.store).toBeUndefined();
+    expect(options.headers.Accept || options.headers.accept).toBe('application/json');
+  });
+
+  it('strips stream fields and store when forwarding compact requests to sub2api upstreams', async () => {
+    selectChannelMock.mockReturnValue({
+      channel: { id: 11, routeId: 22 },
+      site: { id: 44, name: 'sub2api-site', url: 'https://sub2api.example.com', platform: 'sub2api' },
+      account: { id: 33, username: 'sub2api-user' },
+      tokenName: 'default',
+      tokenValue: 'sk-demo',
+      actualModel: 'gpt-5.4',
+    });
+    fetchMock.mockResolvedValue(new Response(JSON.stringify({
+      id: 'cmp_sub2api_123',
+      object: 'response.compaction',
+      input_tokens: 4,
+      output_tokens: 2,
+      total_tokens: 6,
+      output: [
+        {
+          id: 'rs_sub2api_123',
+          type: 'compaction',
+          encrypted_content: 'enc-compact-payload',
+        },
+      ],
+    }), {
+      status: 200,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses/compact',
+      payload: {
+        model: 'gpt-5.4',
+        input: 'hello',
+        store: true,
+        stream_options: { include_obfuscation: true },
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+    const [targetUrl, options] = fetchMock.mock.calls[0] as [string, any];
+    expect(targetUrl).toContain('/v1/responses/compact');
+    const forwardedBody = JSON.parse(String(options.body));
+    expect(forwardedBody.stream).toBeUndefined();
+    expect(forwardedBody.stream_options).toBeUndefined();
+    expect(forwardedBody.store).toBeUndefined();
     expect(options.headers.Accept || options.headers.accept).toBe('application/json');
   });
 
@@ -525,4 +698,62 @@ describe('responses proxy compact upstream routing', () => {
     });
     expect(body.created_at).toEqual(expect.any(Number));
   });
+
+  it('does not fall back from compact on generic upstream errors without compact capability hints', async () => {
+    config.responsesCompactFallbackToResponsesEnabled = true;
+    fetchMock.mockResolvedValue(new Response(JSON.stringify({
+      error: {
+        message: 'upstream_error',
+        type: 'upstream_error',
+      },
+    }), {
+      status: 400,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses/compact',
+      payload: {
+        model: 'gpt-5.2',
+        input: 'hello',
+      },
+    });
+
+    expect(response.statusCode).toBe(400);
+    expect(fetchMock.mock.calls.length).toBeGreaterThan(0);
+    for (const [targetUrl] of fetchMock.mock.calls as Array<[string, any]>) {
+      expect(String(targetUrl)).toContain('/v1/responses/compact');
+      expect(String(targetUrl)).not.toMatch(/\/v1\/responses$/);
+    }
+  });
+
+  it('does not fall back from compact on generic invalid url errors that do not mention compact', async () => {
+    config.responsesCompactFallbackToResponsesEnabled = true;
+    fetchMock.mockResolvedValue(new Response(JSON.stringify({
+      error: {
+        message: 'Invalid URL (POST /v1/responses)',
+        type: 'invalid_request_error',
+      },
+    }), {
+      status: 400,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses/compact',
+      payload: {
+        model: 'gpt-5.2',
+        input: 'hello',
+      },
+    });
+
+    expect(response.statusCode).toBe(400);
+    expect(fetchMock.mock.calls.length).toBeGreaterThan(0);
+    for (const [targetUrl] of fetchMock.mock.calls as Array<[string, any]>) {
+      expect(String(targetUrl)).toContain('/v1/responses/compact');
+      expect(String(targetUrl)).not.toMatch(/\/v1\/responses$/);
+    }
+  });
 });

src/server/routes/proxy/responses.ts

@@ -2,6 +2,26 @@ import type { FastifyInstance, FastifyReply, FastifyRequest } from 'fastify';
 import { handleOpenAiResponsesSurfaceRequest } from '../../proxy-core/surfaces/openAiResponsesSurface.js';
 import { ensureResponsesWebsocketTransport } from './responsesWebsocket.js';
 
+function resolveAliasedResponsesDownstreamPath(
+  request: FastifyRequest,
+): '/v1/responses' | '/v1/responses/compact' | null {
+  const rawUrl = request.raw.url || request.url || '';
+  const pathname = rawUrl.split('?')[0] || rawUrl;
+  if (pathname === '/responses') return '/v1/responses';
+  return pathname.endsWith('/compact')
+    ? '/v1/responses/compact'
+    : null;
+}
+
+async function replyUnsupportedAliasedResponsesPath(reply: FastifyReply) {
+  return reply.code(404).send({
+    error: {
+      message: 'Unknown /responses alias path',
+      type: 'invalid_request_error',
+    },
+  });
+}
+
 export async function responsesProxyRoute(app: FastifyInstance) {
   ensureResponsesWebsocketTransport(app);
 
@@ -16,4 +36,33 @@ export async function responsesProxyRoute(app: FastifyInstance) {
     }));
   app.post('/v1/responses/compact', async (request: FastifyRequest, reply: FastifyReply) =>
     handleOpenAiResponsesSurfaceRequest(request, reply, '/v1/responses/compact'));
+
+  app.post('/responses', async (request: FastifyRequest, reply: FastifyReply) =>
+    handleOpenAiResponsesSurfaceRequest(request, reply, '/v1/responses'));
+  app.post('/responses/*', async (request: FastifyRequest, reply: FastifyReply) => {
+    const downstreamPath = resolveAliasedResponsesDownstreamPath(request);
+    if (!downstreamPath) {
+      return replyUnsupportedAliasedResponsesPath(reply);
+    }
+    return handleOpenAiResponsesSurfaceRequest(request, reply, downstreamPath);
+  });
+  app.get('/responses', async (_request: FastifyRequest, reply: FastifyReply) =>
+    reply.code(426).send({
+      error: {
+        message: 'WebSocket upgrade required for GET /v1/responses',
+        type: 'invalid_request_error',
+      },
+    }));
+  app.get('/responses/*', async (request: FastifyRequest, reply: FastifyReply) => {
+    const downstreamPath = resolveAliasedResponsesDownstreamPath(request);
+    if (!downstreamPath) {
+      return replyUnsupportedAliasedResponsesPath(reply);
+    }
+    return reply.code(426).send({
+      error: {
+        message: `WebSocket upgrade required for GET ${downstreamPath}`,
+        type: 'invalid_request_error',
+      },
+    });
+  });
 }

src/server/routes/proxy/responses.websocket.test.ts

@@ -195,6 +195,16 @@ function waitForSocketOpen(socket: WebSocket) {
   });
 }
 
+function waitForSocketClose(socket: WebSocket) {
+  return new Promise<void>((resolve) => {
+    if (socket.readyState === WebSocket.CLOSED) {
+      resolve();
+      return;
+    }
+    socket.once('close', () => resolve());
+  });
+}
+
 function waitForSocketUpgrade(socket: WebSocket) {
   return new Promise<{ headers: Record<string, string | string[] | undefined> }>((resolve, reject) => {
     socket.once('upgrade', (response) => resolve({ headers: response.headers as Record<string, string | string[] | undefined> }));
@@ -749,6 +759,131 @@ describe('responses websocket transport', () => {
     });
   });
 
+  it('infers previous_response_id for websocket tool-output follow-up turns when the client only sends conversation_id', async () => {
+    const selectedChannel = createSelectedChannel({
+      siteUrl: upstreamSiteUrl,
+    });
+    selectChannelMock.mockReturnValue(selectedChannel);
+    previewSelectedChannelMock.mockResolvedValue(selectedChannel);
+
+    const socket = createClientSocket(baseUrl, {
+      conversation_id: 'ws-conversation-prev-infer',
+    });
+    await waitForSocketOpen(socket);
+
+    const firstResponsePromise = waitForSocketMessageMatching(
+      socket,
+      (message) => message?.type === 'response.completed',
+    );
+    socket.send(JSON.stringify({
+      type: 'response.create',
+      model: 'gpt-5.4',
+      input: [],
+    }));
+    const firstResponse = await firstResponsePromise;
+
+    const secondResponsePromise = waitForSocketMessageMatching(
+      socket,
+      (message) => message?.type === 'response.completed' && message?.response?.id === 'resp_upstream_2',
+    );
+    socket.send(JSON.stringify({
+      type: 'response.create',
+      model: 'gpt-5.4',
+      input: [
+        {
+          id: 'tool_out_ws_conv_1',
+          type: 'function_call_output',
+          call_id: 'call_ws_conv_1',
+          output: '{"ok":true}',
+        },
+      ],
+    }));
+    await secondResponsePromise;
+    socket.close();
+
+    expect(firstResponse?.response?.id).toBe('resp_upstream_1');
+    expect(upstreamUpgradeHeaders.session_id).toBe('ws-conversation-prev-infer');
+    expect(upstreamUpgradeHeaders.conversation_id).toBe('ws-conversation-prev-infer');
+    expect(upstreamRequests).toHaveLength(2);
+    expect(upstreamRequests[1]).toMatchObject({
+      type: 'response.create',
+      previous_response_id: 'resp_upstream_1',
+      input: [
+        {
+          id: 'tool_out_ws_conv_1',
+          type: 'function_call_output',
+          call_id: 'call_ws_conv_1',
+          output: '{"ok":true}',
+        },
+      ],
+    });
+  });
+
+  it('preserves websocket continuation across downstream reconnects on the same conversation_id', async () => {
+    const selectedChannel = createSelectedChannel({
+      siteUrl: upstreamSiteUrl,
+    });
+    selectChannelMock.mockReturnValue(selectedChannel);
+    previewSelectedChannelMock.mockResolvedValue(selectedChannel);
+
+    const firstSocket = createClientSocket(baseUrl, {
+      conversation_id: 'ws-conversation-reconnect-1',
+    });
+    await waitForSocketOpen(firstSocket);
+
+    const firstResponsePromise = waitForSocketMessageMatching(
+      firstSocket,
+      (message) => message?.type === 'response.completed' && message?.response?.id === 'resp_upstream_1',
+    );
+    firstSocket.send(JSON.stringify({
+      type: 'response.create',
+      model: 'gpt-5.4',
+      input: [],
+    }));
+    await firstResponsePromise;
+    firstSocket.close();
+    await waitForSocketClose(firstSocket);
+
+    const secondSocket = createClientSocket(baseUrl, {
+      conversation_id: 'ws-conversation-reconnect-1',
+    });
+    await waitForSocketOpen(secondSocket);
+
+    const secondResponsePromise = waitForSocketMessageMatching(
+      secondSocket,
+      (message) => message?.type === 'response.completed' && message?.response?.id === 'resp_upstream_2',
+    );
+    secondSocket.send(JSON.stringify({
+      type: 'response.create',
+      model: 'gpt-5.4',
+      input: [
+        {
+          id: 'tool_out_ws_reconnect_1',
+          type: 'function_call_output',
+          call_id: 'call_ws_reconnect_1',
+          output: '{"ok":true}',
+        },
+      ],
+    }));
+    await secondResponsePromise;
+    secondSocket.close();
+
+    expect(upstreamConnectionCount).toBe(2);
+    expect(upstreamRequests).toHaveLength(2);
+    expect(upstreamRequests[1]).toMatchObject({
+      type: 'response.create',
+      previous_response_id: 'resp_upstream_1',
+      input: [
+        {
+          id: 'tool_out_ws_reconnect_1',
+          type: 'function_call_output',
+          call_id: 'call_ws_reconnect_1',
+          output: '{"ok":true}',
+        },
+      ],
+    });
+  });
+
   it('retries websocket turns once without previous_response_id when the upstream reports previous_response_not_found', async () => {
     const selectedChannel = createSelectedChannel({
       siteUrl: upstreamSiteUrl,

src/server/routes/proxy/responsesWebsocket.ts

@@ -552,8 +552,10 @@ async function handleResponsesWebsocketConnection(
   request: IncomingMessage,
   authContext: ResponsesWebsocketAuthContext,
 ) {
-  const websocketSessionId = headerValueToTrimmedString(request.headers['session_id'])
-    || headerValueToTrimmedString(request.headers['session-id'])
+  const websocketSessionId = headerValueToTrimmedString(request.headers['session-id'])
+    || headerValueToTrimmedString(request.headers['session_id'])
+    || headerValueToTrimmedString(request.headers['conversation-id'])
+    || headerValueToTrimmedString(request.headers['conversation_id'])
     || randomUUID();
   const runtimeSessionKeys = new Set<string>();
   let lastRequest: Record<string, unknown> | null = null;

src/server/routes/proxy/routeAliases.test.ts

@@ -0,0 +1,145 @@
+import Fastify, { type FastifyInstance } from 'fastify';
+import { afterAll, beforeAll, beforeEach, describe, expect, it, vi } from 'vitest';
+
+const handleOpenAiResponsesSurfaceRequestMock = vi.fn(async (_request, reply, downstreamPath) => (
+  reply.code(200).send({
+    ok: true,
+    downstreamPath,
+  })
+));
+const handleChatSurfaceRequestMock = vi.fn(async (_request, reply, downstreamFormat) => (
+  reply.code(200).send({
+    ok: true,
+    downstreamFormat,
+  })
+));
+const ensureResponsesWebsocketTransportMock = vi.fn();
+
+vi.mock('../../proxy-core/surfaces/openAiResponsesSurface.js', () => ({
+  handleOpenAiResponsesSurfaceRequest: (...args: unknown[]) => handleOpenAiResponsesSurfaceRequestMock(...args),
+}));
+
+vi.mock('../../proxy-core/surfaces/chatSurface.js', () => ({
+  handleChatSurfaceRequest: (...args: unknown[]) => handleChatSurfaceRequestMock(...args),
+  handleClaudeCountTokensSurfaceRequest: vi.fn(),
+}));
+
+vi.mock('./responsesWebsocket.js', () => ({
+  ensureResponsesWebsocketTransport: (...args: unknown[]) => ensureResponsesWebsocketTransportMock(...args),
+}));
+
+describe('proxy route aliases', () => {
+  let app: FastifyInstance;
+
+  beforeAll(async () => {
+    const { responsesProxyRoute } = await import('./responses.js');
+    const { chatProxyRoute } = await import('./chat.js');
+    app = Fastify();
+    await app.register(responsesProxyRoute);
+    await app.register(chatProxyRoute);
+  });
+
+  beforeEach(() => {
+    handleOpenAiResponsesSurfaceRequestMock.mockClear();
+    handleChatSurfaceRequestMock.mockClear();
+    ensureResponsesWebsocketTransportMock.mockClear();
+  });
+
+  afterAll(async () => {
+    await app.close();
+  });
+
+  it('registers bare responses aliases against the same downstream handlers', async () => {
+    const response = await app.inject({
+      method: 'POST',
+      url: '/responses',
+      payload: { model: 'gpt-5.2', input: 'hello' },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(handleOpenAiResponsesSurfaceRequestMock).toHaveBeenCalledTimes(1);
+    expect(response.json()).toEqual({
+      ok: true,
+      downstreamPath: '/v1/responses',
+    });
+  });
+
+  it('routes bare compact responses aliases to compact downstream path', async () => {
+    const response = await app.inject({
+      method: 'POST',
+      url: '/responses/compact',
+      payload: { model: 'gpt-5.2', input: 'hello' },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(handleOpenAiResponsesSurfaceRequestMock).toHaveBeenCalledTimes(1);
+    expect(response.json()).toEqual({
+      ok: true,
+      downstreamPath: '/v1/responses/compact',
+    });
+  });
+
+  it('returns websocket upgrade required for bare GET /responses', async () => {
+    const response = await app.inject({
+      method: 'GET',
+      url: '/responses',
+    });
+
+    expect(response.statusCode).toBe(426);
+    expect(handleOpenAiResponsesSurfaceRequestMock).not.toHaveBeenCalled();
+    expect(response.json()).toEqual({
+      error: {
+        message: 'WebSocket upgrade required for GET /v1/responses',
+        type: 'invalid_request_error',
+      },
+    });
+  });
+
+  it('rejects unknown /responses alias subpaths instead of silently rewriting them', async () => {
+    const response = await app.inject({
+      method: 'POST',
+      url: '/responses/other',
+      payload: { model: 'gpt-5.2', input: 'hello' },
+    });
+
+    expect(response.statusCode).toBe(404);
+    expect(handleOpenAiResponsesSurfaceRequestMock).not.toHaveBeenCalled();
+    expect(response.json()).toEqual({
+      error: {
+        message: 'Unknown /responses alias path',
+        type: 'invalid_request_error',
+      },
+    });
+  });
+
+  it('keeps GET /responses/compact aligned with the websocket upgrade path', async () => {
+    const response = await app.inject({
+      method: 'GET',
+      url: '/responses/compact',
+    });
+
+    expect(response.statusCode).toBe(426);
+    expect(handleOpenAiResponsesSurfaceRequestMock).not.toHaveBeenCalled();
+    expect(response.json()).toEqual({
+      error: {
+        message: 'WebSocket upgrade required for GET /v1/responses/compact',
+        type: 'invalid_request_error',
+      },
+    });
+  });
+
+  it('registers bare chat completions alias against the openai chat handler', async () => {
+    const response = await app.inject({
+      method: 'POST',
+      url: '/chat/completions',
+      payload: { model: 'gpt-5.2', messages: [{ role: 'user', content: 'hello' }] },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(handleChatSurfaceRequestMock).toHaveBeenCalledTimes(1);
+    expect(response.json()).toEqual({
+      ok: true,
+      downstreamFormat: 'openai',
+    });
+  });
+});

src/server/routes/proxy/search.ts

@@ -8,7 +8,7 @@ import { shouldRetryProxyRequest } from '../../services/proxyRetryPolicy.js';
 import { ensureModelAllowedForDownstreamKey, getDownstreamRoutingPolicy, recordDownstreamCostUsage } from './downstreamPolicy.js';
 import { withSiteRecordProxyRequestInit } from '../../services/siteProxy.js';
 import { getProxyUrlFromExtraConfig } from '../../services/accountExtraConfig.js';
-import { composeProxyLogMessage } from './logPathMeta.js';
+import { composeProxyLogMessage } from '../../services/proxyLogMessage.js';
 import { formatUtcSqlDateTime } from '../../services/localTimeService.js';
 import { getProxyAuthContext } from '../../middleware/auth.js';
 import { buildUpstreamUrl } from './upstreamUrl.js';

src/server/routes/proxy/upstreamEndpoint.test.ts

@@ -25,6 +25,8 @@ import {
   MODEL_KEY_HASH_SUFFIX_LENGTH,
 } from '../../services/upstreamEndpointRuntimeMemory.js';
 
+const CODEX_DEFAULT_INSTRUCTIONS = 'You are a helpful coding assistant.';
+
 const baseContext = {
   site: {
     id: 1,
@@ -146,7 +148,7 @@ describe('resolveUpstreamEndpointCandidates', () => {
       'gpt-5.3',
       'openai',
     );
-    expect(openaiOrder).toEqual(['chat', 'responses', 'messages']);
+    expect(openaiOrder).toEqual(['responses', 'chat', 'messages']);
 
     const openaiResponsesOrder = await resolveUpstreamEndpointCandidates(
       {
@@ -166,7 +168,17 @@ describe('resolveUpstreamEndpointCandidates', () => {
       'claude-opus-4-6',
       'openai',
     );
-    expect(openaiClaudeOrder).toEqual(['messages', 'chat', 'responses']);
+    expect(openaiClaudeOrder).toEqual(['responses', 'chat', 'messages']);
+
+    const antigravityOrder = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: { ...baseContext.site, platform: 'antigravity' },
+      },
+      'claude-opus-4-6',
+      'openai',
+    );
+    expect(antigravityOrder).toEqual(['messages']);
 
     const claudeOrder = await resolveUpstreamEndpointCandidates(
       {
@@ -203,7 +215,63 @@ describe('resolveUpstreamEndpointCandidates', () => {
       },
     );
 
-    expect(order).toEqual(['responses', 'messages', 'chat']);
+    expect(order).toEqual(['responses', 'chat', 'messages']);
+  });
+
+  it('derives responses-only candidates for compact requests before surface fallback logic', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: { ...baseContext.site, platform: 'new-api' },
+      },
+      'gpt-5.3',
+      'responses',
+      undefined,
+      undefined,
+      {
+        requestKind: 'responses-compact',
+      },
+    );
+
+    expect(order).toEqual(['responses']);
+  });
+
+  it('derives responses-only candidates when the request requires native responses file-url handling', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: { ...baseContext.site, platform: 'new-api' },
+      },
+      'gpt-5.3',
+      'responses',
+      undefined,
+      {
+        hasNonImageFileInput: true,
+      },
+      {
+        requiresNativeResponsesFileUrl: true,
+      },
+    );
+
+    expect(order).toEqual(['responses']);
+  });
+
+  it('derives responses-first candidates for codex oauth hints without surface-local reordering', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: { ...baseContext.site, platform: 'new-api' },
+      },
+      'gpt-5.3',
+      'openai',
+      undefined,
+      undefined,
+      {
+        oauthProvider: 'codex',
+      },
+    );
+
+    expect(order).toEqual(['responses', 'chat', 'messages']);
   });
 
   it('prefers document-capable endpoints when downstream content contains non-image files', async () => {
@@ -547,7 +615,63 @@ describe('resolveUpstreamEndpointCandidates', () => {
     expect(order).toEqual(['responses', 'messages']);
   });
 
-  it('keeps claude models messages-first even when openai platform catalog prefers chat', async () => {
+  it('learns to prefer /v1/messages after a chat endpoint says messages is required', async () => {
+    const memoryWrite = recordUpstreamEndpointFailure({
+      siteId: baseContext.site.id,
+      endpoint: 'chat',
+      downstreamFormat: 'openai',
+      modelName: 'gpt-5.3',
+      status: 400,
+      errorText: '{"error":{"message":"messages is required","type":"upstream_error"}}',
+    });
+    expect(memoryWrite).toMatchObject({
+      action: 'failure',
+      endpoint: 'chat',
+      blockedEndpoint: 'chat',
+      preferredEndpoint: 'messages',
+    });
+
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: { ...baseContext.site, platform: 'new-api' },
+      },
+      'gpt-5.3',
+      'openai',
+    );
+
+    expect(order).toEqual(['messages', 'responses']);
+  });
+
+  it('learns to prefer /v1/responses after a non-responses endpoint says input is required', async () => {
+    const memoryWrite = recordUpstreamEndpointFailure({
+      siteId: baseContext.site.id,
+      endpoint: 'chat',
+      downstreamFormat: 'openai',
+      modelName: 'gpt-5.3',
+      status: 400,
+      errorText: '{"error":{"message":"input is required","type":"invalid_request_error"}}',
+    });
+    expect(memoryWrite).toMatchObject({
+      action: 'failure',
+      endpoint: 'chat',
+      blockedEndpoint: 'chat',
+      preferredEndpoint: 'responses',
+    });
+
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: { ...baseContext.site, platform: 'new-api' },
+      },
+      'gpt-5.3',
+      'openai',
+    );
+
+    expect(order).toEqual(['responses', 'messages']);
+  });
+
+  it('keeps openai platform responses-first even when the catalog only advertises generic openai/chat support', async () => {
     fetchModelPricingCatalogMock.mockResolvedValue({
       models: [
         {
@@ -567,7 +691,7 @@ describe('resolveUpstreamEndpointCandidates', () => {
       'openai',
     );
 
-    expect(order).toEqual(['messages', 'chat', 'responses']);
+    expect(order).toEqual(['responses', 'chat', 'messages']);
   });
 
   it('keeps anyrouter messages-first special case', async () => {
@@ -623,7 +747,8 @@ describe('resolveUpstreamEndpointCandidates', () => {
     expect(isEndpointDowngradeError(404, '{"error":{"message":"Not Found","type":"not_found_error"}}')).toBe(true);
     expect(isEndpointDowngradeError(405, '{"error":{"message":"Method Not Allowed"}}')).toBe(true);
     expect(isEndpointDowngradeError(400, '{"error":{"message":"unsupported endpoint","type":"invalid_request_error"}}')).toBe(true);
-    expect(isEndpointDowngradeError(400, '{"error":{"message":"","type":"upstream_error"}}')).toBe(true);
+    expect(isEndpointDowngradeError(400, '{"error":{"message":"","type":"upstream_error"}}')).toBe(false);
+    expect(isEndpointDowngradeError(400, '{"error":{"message":"upstream_error: unsupported endpoint /v1/responses","type":"upstream_error"}}')).toBe(true);
     expect(isEndpointDowngradeError(400, '{"error":{"message":"openai_error","type":"bad_response_status_code"}}')).toBe(true);
     expect(isEndpointDowngradeError(415, '{"error":{"message":"Unsupported Media Type: Only \\"application/json\\" is allowed"}}')).toBe(true);
   });
@@ -817,7 +942,7 @@ describe('buildUpstreamEndpointRequest', () => {
     expect(request.headers['User-Agent']).toBe('codex_cli_rs/0.101.0 (Mac OS 26.0.1; arm64) Apple_Terminal/464');
     expect(request.headers.Accept).toBe('application/json');
     expect(request.headers.Connection).toBe('Keep-Alive');
-    expect(request.body.instructions).toBe('');
+    expect(request.body.instructions).toBe(CODEX_DEFAULT_INSTRUCTIONS);
     expect(request.body.prompt_cache_key).toBeUndefined();
     expect(request.body.stream).toBe(false);
     expect(request.body.store).toBe(false);
@@ -956,7 +1081,7 @@ describe('buildUpstreamEndpointRequest', () => {
     expect(request.headers.Session_id).toMatch(/^[0-9a-f-]{36}$/i);
     expect(request.headers.Conversation_id).toBe(request.headers.Session_id);
     expect(request.body.prompt_cache_key).toBeUndefined();
-    expect(request.body.instructions).toBe('');
+    expect(request.body.instructions).toBe(CODEX_DEFAULT_INSTRUCTIONS);
     expect(request.body.stream).toBe(false);
     expect(request.body.store).toBe(false);
     expect(request.body.parallel_tool_calls).toBe(false);
@@ -1300,7 +1425,7 @@ describe('buildUpstreamEndpointRequest', () => {
     expect(request.headers['Accept-Encoding']).toBe('gzip, deflate, br, zstd');
   });
 
-  it('converts system roles to developer in native codex responses bodies', () => {
+  it('extracts system input into top-level instructions in native codex responses bodies', () => {
     const request = buildUpstreamEndpointRequest({
       endpoint: 'responses',
       modelName: 'gpt-5.4',
@@ -1312,6 +1437,7 @@ describe('buildUpstreamEndpointRequest', () => {
       downstreamFormat: 'responses',
       responsesOriginalBody: {
         model: 'gpt-5.4',
+        instructions: 'keep edits narrow',
         input: [
           {
             type: 'message',
@@ -1328,17 +1454,13 @@ describe('buildUpstreamEndpointRequest', () => {
     });
 
     expect(request.body.input).toEqual([
-      {
-        type: 'message',
-        role: 'developer',
-        content: [{ type: 'input_text', text: 'be careful' }],
-      },
       {
         type: 'message',
         role: 'user',
         content: [{ type: 'input_text', text: 'hello' }],
       },
     ]);
+    expect(request.body.instructions).toBe('be careful\n\nkeep edits narrow');
   });
 
   it('normalizes downstream responses input string before forwarding upstream', () => {

src/server/services/brandMatcher.test.ts

@@ -23,8 +23,11 @@ describe('brandMatcher', () => {
     expect(getBlockedBrandRules(['OpenRouter', 'OpenRouter', 'Unknown Brand'])).toEqual(['OpenRouter']);
   });
 
-  it('canonicalizes blocked brand names before validating them', () => {
-    expect(getBlockedBrandRules(['openai', '  aws   bedrock  ', 'unknown brand', 'OPENAI'])).toEqual(['OpenAI', 'AWS Bedrock']);
+  it('canonicalizes blocked brand names before filtering', () => {
+    expect(getBlockedBrandRules([' openrouter ', 'OPENROUTER', 'together   ai'])).toEqual([
+      'OpenRouter',
+      'Together AI',
+    ]);
   });
 
   it('uses the shared frontend detection result for global brand blocking', () => {
@@ -33,9 +36,6 @@ describe('brandMatcher', () => {
     expect(isModelBlockedByBrand('groq/compound-beta', providerRules)).toBe(true);
     expect(isModelBlockedByBrand('dashscope/wanx2.1-t2i-turbo', providerRules)).toBe(true);
     expect(isModelBlockedByBrand('LongCat-Flash-Lite', providerRules)).toBe(true);
-    expect(isModelBlockedByBrand('deepinfra/meta-llama/llama-3.3-70b-instruct', getBlockedBrandRules(['DeepInfra']))).toBe(true);
-    expect(isModelBlockedByBrand('azureai/gpt-4o', getBlockedBrandRules(['Azure AI']))).toBe(true);
-    expect(isModelBlockedByBrand('bedrock/us.amazon.nova-pro-v1:0', getBlockedBrandRules(['AWS Bedrock']))).toBe(true);
 
     const vendorRules = getBlockedBrandRules(['Anthropic', 'Meta', 'Google', 'OpenAI', 'Arcee', 'Xiaomi MiMo']);
     expect(isModelBlockedByBrand('openrouter/anthropic/claude-3-7-sonnet', vendorRules)).toBe(true);

src/server/services/modelService.ts

@@ -818,7 +818,7 @@ export async function refreshModelsForAccount(
         `api token discovery timeout (${Math.round(API_TOKEN_DISCOVERY_TIMEOUT_MS / 1000)}s)`,
       );
       if (discoveredApiToken && !isMaskedTokenValue(discoveredApiToken)) {
-        ensureDefaultTokenForAccount(account.id, discoveredApiToken, { name: 'default', source: 'sync' });
+        await ensureDefaultTokenForAccount(account.id, discoveredApiToken, { name: 'default', source: 'sync' });
         await db.update(schema.accounts).set({
           apiToken: discoveredApiToken,
           updatedAt: new Date().toISOString(),
@@ -846,7 +846,7 @@ export async function refreshModelsForAccount(
   if (usesManagedTokens && enabledTokens.length === 0) {
     const fallback = discoveredApiToken || account.apiToken || null;
     if (fallback) {
-      ensureDefaultTokenForAccount(account.id, fallback, { name: 'default', source: 'legacy' });
+      await ensureDefaultTokenForAccount(account.id, fallback, { name: 'default', source: 'legacy' });
       enabledTokens = await db.select()
         .from(schema.accountTokens)
         .where(and(

src/server/services/oauth/codexProvider.ts

@@ -1,5 +1,6 @@
 import { fetch } from 'undici';
 import { config } from '../../config.js';
+import { inferCodexOfficialOriginator } from '../../shared/codexClientFamily.js';
 import { withExplicitProxyRequestInit } from '../siteProxy.js';
 import { createPkceChallenge } from './sessionStore.js';
 import type { OAuthProviderDefinition } from './providers.js';
@@ -234,7 +235,9 @@ export const codexOauthProvider: OAuthProviderDefinition = {
   },
   buildProxyHeaders: ({ oauth, downstreamHeaders }) => {
     const accountId = oauth.accountId || oauth.accountKey;
-    const originator = getHeaderValue(downstreamHeaders, 'originator') || 'codex_cli_rs';
+    const originator = inferCodexOfficialOriginator(downstreamHeaders)
+      || getHeaderValue(downstreamHeaders, 'originator')
+      || 'codex_cli_rs';
     const headers: Record<string, string> = {
       Originator: originator,
     };

src/server/services/proxyBilling.test.ts

@@ -1,8 +1,40 @@
-import { describe, expect, it } from 'vitest';
+import { describe, expect, it, vi } from 'vitest';
+
+const estimateProxyCostMock = vi.fn();
+const buildProxyBillingDetailsMock = vi.fn();
+
+vi.mock('./modelPricingService.js', () => ({
+  estimateProxyCost: (...args: unknown[]) => estimateProxyCostMock(...args),
+  buildProxyBillingDetails: (...args: unknown[]) => buildProxyBillingDetailsMock(...args),
+}));
+
 import { resolveProxyLogBilling } from './proxyBilling.js';
 
 describe('resolveProxyLogBilling', () => {
   it('uses self-log billing metadata for detail breakdown while preserving quota-derived total cost', async () => {
+    estimateProxyCostMock.mockResolvedValue(0.010001);
+    buildProxyBillingDetailsMock.mockResolvedValue({
+      usage: {
+        cacheReadTokens: 145692,
+        cacheCreationTokens: 945,
+        billablePromptTokens: 1,
+      },
+      pricing: {
+        modelRatio: 2.5,
+        completionRatio: 5,
+        cacheRatio: 0.1,
+        cacheCreationRatio: 1.25,
+        groupRatio: 1,
+      },
+      breakdown: {
+        inputPerMillion: 5,
+        outputPerMillion: 25,
+        cacheReadPerMillion: 0.5,
+        cacheCreationPerMillion: 6.25,
+        totalCost: 0.083057,
+      },
+    });
+
     const result = await resolveProxyLogBilling({
       site: {
         id: 1,
@@ -40,6 +72,18 @@ describe('resolveProxyLogBilling', () => {
       },
     });
 
+    expect(estimateProxyCostMock).toHaveBeenCalledWith(expect.objectContaining({
+      cacheReadTokens: 145692,
+      cacheCreationTokens: 945,
+      promptTokensIncludeCache: true,
+      billingPricingOverride: {
+        modelRatio: 2.5,
+        completionRatio: 5,
+        cacheRatio: 0.1,
+        cacheCreationRatio: 1.25,
+        groupRatio: 1,
+      },
+    }));
     expect(result.estimatedCost).toBe(0.083056);
     expect(result.billingDetails).toMatchObject({
       usage: {

src/server/services/proxyBilling.ts

@@ -0,0 +1,92 @@
+import {
+  buildProxyBillingDetails,
+  estimateProxyCost,
+  type ProxyBillingDetails,
+  type ProxyBillingPricingOverride,
+} from './modelPricingService.js';
+import type { SelfLogBillingMeta } from './proxyUsageFallbackService.js';
+
+interface ProxyBillingUsageSummary {
+  promptTokens: number;
+  completionTokens: number;
+  totalTokens: number;
+  cacheReadTokens: number;
+  cacheCreationTokens: number;
+  promptTokensIncludeCache: boolean | null;
+}
+
+interface ResolvedProxyUsageSummary {
+  promptTokens: number;
+  completionTokens: number;
+  totalTokens: number;
+  recoveredFromSelfLog: boolean;
+  estimatedCostFromQuota: number;
+  selfLogBillingMeta: SelfLogBillingMeta | null;
+}
+
+interface ResolveProxyLogBillingInput {
+  site: {
+    id: number;
+    url: string;
+    platform: string;
+    apiKey?: string | null;
+  };
+  account: {
+    id: number;
+    accessToken?: string | null;
+    apiToken?: string | null;
+  };
+  modelName: string;
+  parsedUsage: ProxyBillingUsageSummary;
+  resolvedUsage: ResolvedProxyUsageSummary;
+}
+
+function toPricingOverride(meta: SelfLogBillingMeta | null): ProxyBillingPricingOverride | null {
+  if (!meta) return null;
+  return {
+    modelRatio: meta.modelRatio,
+    completionRatio: meta.completionRatio,
+    cacheRatio: meta.cacheRatio,
+    cacheCreationRatio: meta.cacheCreationRatio,
+    groupRatio: meta.groupRatio,
+  };
+}
+
+export async function resolveProxyLogBilling(
+  input: ResolveProxyLogBillingInput,
+): Promise<{ estimatedCost: number; billingDetails: ProxyBillingDetails | null }> {
+  const selfLogMeta = input.resolvedUsage.selfLogBillingMeta;
+  const billingPricingOverride = toPricingOverride(selfLogMeta);
+  const cacheReadTokens = selfLogMeta?.cacheReadTokens ?? input.parsedUsage.cacheReadTokens;
+  const cacheCreationTokens = selfLogMeta?.cacheCreationTokens ?? input.parsedUsage.cacheCreationTokens;
+  const promptTokensIncludeCache = selfLogMeta?.promptTokensIncludeCache
+    ?? input.parsedUsage.promptTokensIncludeCache;
+
+  const billingInput = {
+    site: input.site,
+    account: input.account,
+    modelName: input.modelName,
+    promptTokens: input.resolvedUsage.promptTokens,
+    completionTokens: input.resolvedUsage.completionTokens,
+    totalTokens: input.resolvedUsage.totalTokens,
+    cacheReadTokens,
+    cacheCreationTokens,
+    promptTokensIncludeCache,
+    billingPricingOverride,
+  };
+
+  let estimatedCost = await estimateProxyCost(billingInput);
+  const billingDetails = await buildProxyBillingDetails(billingInput);
+
+  if (
+    input.resolvedUsage.estimatedCostFromQuota > 0
+    && (input.resolvedUsage.recoveredFromSelfLog || estimatedCost <= 0)
+  ) {
+    estimatedCost = input.resolvedUsage.estimatedCostFromQuota;
+  }
+
+  return {
+    estimatedCost,
+    billingDetails,
+  };
+}

src/server/services/proxyFailureJudge.test.ts

@@ -1,5 +1,5 @@
 import { afterEach, describe, expect, it } from 'vitest';
-import { config } from '../../config.js';
+import { config } from '../config.js';
 import { detectProxyFailure } from './proxyFailureJudge.js';
 
 describe('detectProxyFailure (empty content)', () => {

src/server/services/proxyFailureJudge.ts

@@ -1,5 +1,5 @@
-import { config } from '../../config.js';
-import { pullSseDataEvents } from '../../services/proxyUsageParser.js';
+import { config } from '../config.js';
+import { pullSseDataEvents } from './proxyUsageParser.js';
 
 type FailureResult = {
   status: number;

src/server/services/proxyLogMessage.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, it } from 'vitest';
-import { composeProxyLogMessage } from './logPathMeta.js';
+import { composeProxyLogMessage } from './proxyLogMessage.js';
 
 describe('composeProxyLogMessage', () => {
   it('adds downstream path metadata before plain errors', () => {

src/server/services/proxyLogMessage.ts

@@ -2,7 +2,7 @@ import {
   parseProxyLogMetadata,
   type ParsedProxyLogMetadata,
   type ProxyLogUsageSource,
-} from '../../../shared/proxyLogMeta.js';
+} from '../../shared/proxyLogMeta.js';
 
 type ComposeProxyLogMessageArgs = {
   clientKind?: string | null;

src/server/services/runtimeModelProbe.test.ts

@@ -5,13 +5,15 @@ const buildUpstreamEndpointRequestMock = vi.fn();
 const dispatchRuntimeRequestMock = vi.fn();
 const resolveChannelProxyUrlMock = vi.fn();
 const withSiteRecordProxyRequestInitMock = vi.fn();
+const getOauthInfoFromAccountMock = vi.fn();
+const buildOauthProviderHeadersMock = vi.fn();
 
-vi.mock('../routes/proxy/upstreamEndpoint.js', () => ({
+vi.mock('./upstreamEndpointRuntime.js', () => ({
   resolveUpstreamEndpointCandidates: (...args: unknown[]) => resolveUpstreamEndpointCandidatesMock(...args),
   buildUpstreamEndpointRequest: (...args: unknown[]) => buildUpstreamEndpointRequestMock(...args),
 }));
 
-vi.mock('../routes/proxy/runtimeExecutor.js', () => ({
+vi.mock('./runtimeDispatch.js', () => ({
   dispatchRuntimeRequest: (...args: unknown[]) => dispatchRuntimeRequestMock(...args),
 }));
 
@@ -20,6 +22,14 @@ vi.mock('./siteProxy.js', () => ({
   withSiteRecordProxyRequestInit: (...args: unknown[]) => withSiteRecordProxyRequestInitMock(...args),
 }));
 
+vi.mock('./oauth/oauthAccount.js', () => ({
+  getOauthInfoFromAccount: (...args: unknown[]) => getOauthInfoFromAccountMock(...args),
+}));
+
+vi.mock('./oauth/service.js', () => ({
+  buildOauthProviderHeaders: (...args: unknown[]) => buildOauthProviderHeadersMock(...args),
+}));
+
 describe('probeRuntimeModel', () => {
   const site = {
     id: 1,
@@ -46,7 +56,11 @@ describe('probeRuntimeModel', () => {
     dispatchRuntimeRequestMock.mockReset();
     resolveChannelProxyUrlMock.mockReset();
     withSiteRecordProxyRequestInitMock.mockReset();
+    getOauthInfoFromAccountMock.mockReset();
+    buildOauthProviderHeadersMock.mockReset();
 
+    getOauthInfoFromAccountMock.mockReturnValue(null);
+    buildOauthProviderHeadersMock.mockReturnValue({});
     buildUpstreamEndpointRequestMock.mockReturnValue({
       path: '/v1/chat/completions',
       headers: { 'content-type': 'application/json' },

src/server/services/upstreamEndpointDerivation.test.ts

@@ -0,0 +1,164 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+const fetchModelPricingCatalogMock = vi.fn(async (_arg?: unknown): Promise<any> => null);
+
+vi.mock('./modelPricingService.js', () => ({
+  fetchModelPricingCatalog: (arg: unknown) => fetchModelPricingCatalogMock(arg),
+}));
+
+import { resolveUpstreamEndpointCandidates } from './upstreamEndpointDerivation.js';
+import { resetUpstreamEndpointRuntimeState } from './upstreamEndpointRuntimeMemory.js';
+
+const baseContext = {
+  site: {
+    id: 1,
+    url: 'https://upstream.example.com',
+    platform: 'new-api',
+    apiKey: null,
+  },
+  account: {
+    id: 2,
+    accessToken: 'token-demo',
+    apiToken: null,
+  },
+};
+
+describe('upstreamEndpointDerivation', () => {
+  beforeEach(() => {
+    fetchModelPricingCatalogMock.mockReset();
+    fetchModelPricingCatalogMock.mockResolvedValue(null);
+    resetUpstreamEndpointRuntimeState();
+  });
+
+  it('derives compact requests directly to responses from the service owner', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      baseContext,
+      'gpt-5.3',
+      'responses',
+      undefined,
+      undefined,
+      {
+        requestKind: 'responses-compact',
+      },
+    );
+
+    expect(order).toEqual(['responses']);
+  });
+
+  it('derives codex oauth openai requests as responses-first without surface-local reordering', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      baseContext,
+      'gpt-5.3',
+      'openai',
+      undefined,
+      undefined,
+      {
+        oauthProvider: 'codex',
+      },
+    );
+
+    expect(order).toEqual(['responses', 'chat', 'messages']);
+  });
+
+  it('keeps explicit openai platforms on responses-first ordering even for claude-family models', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: {
+          ...baseContext.site,
+          platform: 'openai',
+        },
+      },
+      'claude-opus-4-6',
+      'openai',
+    );
+
+    expect(order).toEqual(['responses', 'chat', 'messages']);
+  });
+
+  it('keeps antigravity non-gemini compatibility requests on messages-first ordering', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: {
+          ...baseContext.site,
+          platform: 'antigravity',
+        },
+      },
+      'claude-opus-4-6',
+      'openai',
+      undefined,
+      {
+        hasNonImageFileInput: true,
+      },
+    );
+
+    expect(order).toEqual(['messages']);
+  });
+
+  it('keeps claude-family file-url requests messages-first for claude upstreams', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: {
+          ...baseContext.site,
+          platform: 'claude',
+        },
+      },
+      'claude-opus-4-6',
+      'responses',
+      undefined,
+      {
+        hasNonImageFileInput: true,
+      },
+      {
+        requiresNativeResponsesFileUrl: true,
+      },
+    );
+
+    expect(order).toEqual(['messages']);
+  });
+
+  it('derives claude count_tokens requests as messages-only when the upstream supports messages', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: {
+          ...baseContext.site,
+          platform: 'openai',
+        },
+      },
+      'claude-sonnet-4-5-20250929',
+      'claude',
+      undefined,
+      undefined,
+      {
+        requestKind: 'claude-count-tokens',
+      },
+    );
+
+    expect(order).toEqual(['messages']);
+  });
+
+  it('returns no candidates for claude count_tokens when the upstream does not support messages', async () => {
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: {
+          ...baseContext.site,
+          platform: 'codex',
+          url: 'https://chatgpt.com/backend-api/codex',
+        },
+      },
+      'gpt-5.4',
+      'claude',
+      undefined,
+      undefined,
+      {
+        requestKind: 'claude-count-tokens',
+      },
+    );
+
+    expect(order).toEqual([]);
+  });
+});

src/server/services/upstreamEndpointDerivation.ts

@@ -0,0 +1,322 @@
+import {
+  rankConversationFileEndpoints,
+  type ConversationFileInputSummary,
+} from '../proxy-core/capabilities/conversationFileCapabilities.js';
+import type { UpstreamEndpoint } from '../proxy-core/orchestration/upstreamRequest.js';
+import { fetchModelPricingCatalog } from './modelPricingService.js';
+import {
+  applyUpstreamEndpointRuntimePreference,
+  buildEndpointCapabilityProfile,
+} from './upstreamEndpointRuntimeMemory.js';
+import type { DownstreamFormat } from '../transformers/shared/normalized.js';
+
+export type EndpointPreference = DownstreamFormat | 'responses';
+export type EndpointDerivationHints = {
+  oauthProvider?: string | null;
+  requestKind?: 'default' | 'responses-compact' | 'claude-count-tokens';
+  requiresNativeResponsesFileUrl?: boolean;
+};
+
+type ChannelContext = {
+  site: {
+    id: number;
+    url: string;
+    platform: string;
+    apiKey?: string | null;
+  };
+  account: {
+    id: number;
+    accessToken?: string | null;
+    apiToken?: string | null;
+  };
+};
+
+function asTrimmedString(value: unknown): string {
+  return typeof value === 'string' ? value.trim() : '';
+}
+
+function normalizePlatformName(platform: unknown): string {
+  return asTrimmedString(platform).toLowerCase();
+}
+
+function normalizeEndpointTypes(value: unknown): UpstreamEndpoint[] {
+  const raw = asTrimmedString(value).toLowerCase();
+  if (!raw) return [];
+
+  const normalized = new Set<UpstreamEndpoint>();
+
+  if (
+    raw.includes('/v1/messages')
+    || raw === 'messages'
+    || raw.includes('anthropic')
+    || raw.includes('claude')
+  ) {
+    normalized.add('messages');
+  }
+
+  if (
+    raw.includes('/v1/responses')
+    || raw === 'responses'
+    || raw.includes('response')
+  ) {
+    normalized.add('responses');
+  }
+
+  if (
+    raw.includes('/v1/chat/completions')
+    || raw.includes('chat/completions')
+    || raw === 'chat'
+    || raw === 'chat_completions'
+    || raw === 'completions'
+    || raw.includes('chat')
+  ) {
+    normalized.add('chat');
+  }
+
+  if (raw === 'openai' || raw.includes('openai')) {
+    normalized.add('chat');
+    normalized.add('responses');
+  }
+
+  return Array.from(normalized);
+}
+
+function preferredEndpointOrder(
+  downstreamFormat: EndpointPreference,
+  sitePlatform?: string,
+  preferMessagesForClaudeModel = false,
+  hints?: EndpointDerivationHints,
+): UpstreamEndpoint[] {
+  const platform = normalizePlatformName(sitePlatform);
+  const oauthProvider = asTrimmedString(hints?.oauthProvider).toLowerCase();
+
+  if (hints?.requestKind === 'responses-compact') {
+    return ['responses'];
+  }
+
+  if (platform === 'codex') {
+    return ['responses'];
+  }
+
+  if (platform === 'gemini' || platform === 'gemini-cli') {
+    return ['chat'];
+  }
+
+  if (platform === 'openai') {
+    return ['responses', 'chat', 'messages'];
+  }
+
+  if (platform === 'antigravity') {
+    return ['messages'];
+  }
+
+  if (platform === 'claude') {
+    return ['messages'];
+  }
+
+  if (downstreamFormat === 'responses') {
+    if (preferMessagesForClaudeModel) {
+      return ['messages', 'chat', 'responses'];
+    }
+    return ['responses', 'chat', 'messages'];
+  }
+
+  if (downstreamFormat === 'claude') {
+    return ['messages', 'chat', 'responses'];
+  }
+
+  if (downstreamFormat === 'openai' && preferMessagesForClaudeModel) {
+    return ['messages', 'chat', 'responses'];
+  }
+
+  const base = ['chat', 'messages', 'responses'] as UpstreamEndpoint[];
+  if (oauthProvider === 'codex' && base.includes('responses')) {
+    return ['responses', ...base.filter((endpoint) => endpoint !== 'responses')];
+  }
+
+  return base;
+}
+
+export async function resolveUpstreamEndpointCandidates(
+  context: ChannelContext,
+  modelName: string,
+  downstreamFormat: EndpointPreference,
+  requestedModelHint?: string,
+  requestCapabilities?: {
+    hasNonImageFileInput?: boolean;
+    conversationFileSummary?: ConversationFileInputSummary;
+    wantsNativeResponsesReasoning?: boolean;
+    wantsContinuationAwareResponses?: boolean;
+  },
+  hints?: EndpointDerivationHints,
+): Promise<UpstreamEndpoint[]> {
+  const sitePlatform = normalizePlatformName(context.site.platform);
+  if (hints?.requestKind === 'responses-compact') {
+    return ['responses'];
+  }
+  if (
+    hints?.requiresNativeResponsesFileUrl
+    && sitePlatform !== 'claude'
+    && sitePlatform !== 'anyrouter'
+  ) {
+    return ['responses'];
+  }
+
+  const capabilityProfile = buildEndpointCapabilityProfile({
+    modelName,
+    requestedModelHint,
+    requestCapabilities,
+  });
+  const preferMessagesForClaudeModel = capabilityProfile.preferMessagesForClaudeModel;
+  const hasNonImageFileInput = capabilityProfile.hasNonImageFileInput;
+  const wantsNativeResponsesReasoning = capabilityProfile.wantsNativeResponsesReasoning;
+  const wantsContinuationAwareResponses = capabilityProfile.wantsContinuationAwareResponses;
+  const applyRuntimePreference = (candidates: UpstreamEndpoint[]) => (
+    applyUpstreamEndpointRuntimePreference(candidates, {
+      siteId: context.site.id,
+      downstreamFormat,
+      capabilityProfile,
+    })
+  );
+  const finalizeCandidates = (candidates: UpstreamEndpoint[]): UpstreamEndpoint[] => {
+    const preferredCandidates = applyRuntimePreference(candidates);
+    if (hints?.requestKind === 'claude-count-tokens') {
+      return preferredCandidates.includes('messages') ? ['messages'] : ([] as UpstreamEndpoint[]);
+    }
+    return preferredCandidates;
+  };
+  const conversationFileSummary = requestCapabilities?.conversationFileSummary ?? {
+    hasImage: false,
+    hasAudio: false,
+    hasDocument: hasNonImageFileInput,
+    hasRemoteDocumentUrl: false,
+  };
+
+  if (sitePlatform === 'anyrouter') {
+    if (hasNonImageFileInput) {
+      return finalizeCandidates(downstreamFormat === 'responses'
+        ? ['responses', 'messages', 'chat']
+        : ['messages', 'responses', 'chat']);
+    }
+    if (downstreamFormat === 'responses') {
+      return finalizeCandidates(['responses', 'messages', 'chat']);
+    }
+    return finalizeCandidates(['messages', 'chat', 'responses']);
+  }
+
+  const preferred = preferredEndpointOrder(
+    downstreamFormat,
+    context.site.platform,
+    preferMessagesForClaudeModel,
+    hints,
+  );
+  const preferredWithCapabilities = hasNonImageFileInput
+    ? (() => {
+      if (sitePlatform === 'claude') return ['messages'] as UpstreamEndpoint[];
+      if (sitePlatform === 'gemini') return ['responses', 'chat'] as UpstreamEndpoint[];
+      if (sitePlatform === 'gemini-cli') return ['chat'] as UpstreamEndpoint[];
+      if (sitePlatform === 'antigravity') return ['messages'] as UpstreamEndpoint[];
+      if (sitePlatform === 'openai') return ['responses', 'chat', 'messages'] as UpstreamEndpoint[];
+      return rankConversationFileEndpoints({
+        sitePlatform,
+        requestedOrder: preferMessagesForClaudeModel
+          ? ['messages', 'responses', 'chat']
+          : ['responses', 'messages', 'chat'],
+        summary: conversationFileSummary,
+        preferMessagesForClaudeModel,
+      });
+    })()
+    : preferred;
+  const prioritizedPreferredEndpoints: UpstreamEndpoint[] = (
+    preferredWithCapabilities.includes('responses')
+    && (
+      wantsContinuationAwareResponses
+      || (wantsNativeResponsesReasoning && preferMessagesForClaudeModel)
+    )
+  )
+    ? [
+      'responses',
+      ...preferredWithCapabilities.filter((endpoint): endpoint is UpstreamEndpoint => endpoint !== 'responses'),
+    ]
+    : preferredWithCapabilities;
+  const forceMessagesFirstForClaudeModel = (
+    downstreamFormat === 'openai'
+    && preferMessagesForClaudeModel
+    && sitePlatform !== 'openai'
+    && sitePlatform !== 'gemini'
+    && sitePlatform !== 'antigravity'
+    && sitePlatform !== 'gemini-cli'
+  );
+
+  try {
+    const catalog = await fetchModelPricingCatalog({
+      site: {
+        id: context.site.id,
+        url: context.site.url,
+        platform: context.site.platform,
+      },
+      account: {
+        id: context.account.id,
+        accessToken: context.account.accessToken ?? null,
+        apiToken: context.account.apiToken ?? null,
+      },
+      modelName,
+      totalTokens: 0,
+    });
+
+    if (!catalog || !Array.isArray(catalog.models) || catalog.models.length === 0) {
+      return finalizeCandidates(prioritizedPreferredEndpoints);
+    }
+
+    const matched = catalog.models.find((item) =>
+      asTrimmedString(item?.modelName).toLowerCase() === modelName.toLowerCase(),
+    );
+    if (!matched) return finalizeCandidates(prioritizedPreferredEndpoints);
+
+    const shouldIgnoreCatalogOrderingForClaudeMessages = (
+      preferMessagesForClaudeModel
+      && (downstreamFormat !== 'responses' || sitePlatform !== 'openai')
+    );
+    if (shouldIgnoreCatalogOrderingForClaudeMessages) {
+      return finalizeCandidates(prioritizedPreferredEndpoints);
+    }
+
+    const supportedRaw = Array.isArray(matched.supportedEndpointTypes) ? matched.supportedEndpointTypes : [];
+    const normalizedSupportedRaw = supportedRaw
+      .map((item) => asTrimmedString(item).toLowerCase())
+      .filter((item) => item.length > 0);
+    const hasConcreteEndpointHint = normalizedSupportedRaw.some((raw) => (
+      raw.includes('/v1/messages')
+      || raw.includes('/v1/chat/completions')
+      || raw.includes('/v1/responses')
+      || raw === 'messages'
+      || raw === 'chat'
+      || raw === 'chat_completions'
+      || raw === 'completions'
+      || raw === 'responses'
+    ));
+    if (forceMessagesFirstForClaudeModel && !hasConcreteEndpointHint) {
+      return finalizeCandidates(prioritizedPreferredEndpoints);
+    }
+
+    const supported = new Set<UpstreamEndpoint>();
+    for (const endpoint of supportedRaw) {
+      const normalizedList = normalizeEndpointTypes(endpoint);
+      for (const normalized of normalizedList) {
+        supported.add(normalized);
+      }
+    }
+
+    if (supported.size === 0) return finalizeCandidates(prioritizedPreferredEndpoints);
+
+    const firstSupported = prioritizedPreferredEndpoints.find((endpoint) => supported.has(endpoint));
+    if (!firstSupported) return finalizeCandidates(prioritizedPreferredEndpoints);
+
+    return finalizeCandidates([
+      firstSupported,
+      ...prioritizedPreferredEndpoints.filter((endpoint) => endpoint !== firstSupported),
+    ]);
+  } catch {
+    return finalizeCandidates(prioritizedPreferredEndpoints);
+  }
+}

src/server/services/upstreamEndpointRuntime.ts

@@ -1,6 +1,10 @@
 export {
   buildClaudeCountTokensUpstreamRequest,
   buildUpstreamEndpointRequest,
+} from './upstreamRequestBuilder.js';
+export type { UpstreamEndpoint } from '../proxy-core/orchestration/upstreamRequest.js';
+export {
   resolveUpstreamEndpointCandidates,
-  type UpstreamEndpoint,
-} from '../routes/proxy/upstreamEndpoint.js';
+  type EndpointPreference,
+  type EndpointDerivationHints,
+} from './upstreamEndpointDerivation.js';

src/server/services/upstreamEndpointRuntimeMemory.ts

@@ -2,7 +2,10 @@ import { createHash } from 'node:crypto';
 import type { ConversationFileInputSummary } from '../proxy-core/capabilities/conversationFileCapabilities.js';
 import type { DownstreamFormat } from '../transformers/shared/normalized.js';
 import {
+  inferSuggestedEndpointFromUpstreamError,
+  inferRequiredEndpointFromProtocolError,
   isEndpointDispatchDeniedError,
+  isEndpointDowngradeError,
   isUnsupportedMediaTypeError,
 } from '../transformers/shared/endpointCompatibility.js';
 
@@ -200,38 +203,30 @@ function maybeDeleteEndpointRuntimeState(key: string, nowMs = Date.now()): void
   }
 }
 
-function inferSuggestedEndpointFromError(errorText?: string | null): UpstreamEndpointRuntimeEndpoint | null {
-  const text = (errorText || '').toLowerCase();
-  if (!text) return null;
-  if (text.includes('/v1/responses')) return 'responses';
-  if (text.includes('/v1/messages')) return 'messages';
-  if (text.includes('/v1/chat/completions')) return 'chat';
-  return null;
+function inferSuggestedEndpointFromError(
+  endpoint: UpstreamEndpointRuntimeEndpoint,
+  errorText?: string | null,
+): UpstreamEndpointRuntimeEndpoint | null {
+  const suggestedEndpoint = inferSuggestedEndpointFromUpstreamError(errorText);
+  return suggestedEndpoint && endpoint !== suggestedEndpoint ? suggestedEndpoint : null;
 }
 
-function shouldBlockEndpointByError(status: number, errorText?: string | null): boolean {
+function shouldBlockEndpointByError(
+  endpoint: UpstreamEndpointRuntimeEndpoint,
+  status: number,
+  errorText?: string | null,
+): boolean {
   if (isEndpointDispatchDeniedError(status, errorText)) return true;
   if (status === 404 || status === 405 || status === 415 || status === 501) return true;
   if (isUnsupportedMediaTypeError(status, errorText)) return true;
 
-  const text = (errorText || '').toLowerCase();
-  return (
-    text.includes('convert_request_failed')
-    || text.includes('endpoint_not_found')
-    || text.includes('unknown_endpoint')
-    || text.includes('unsupported_endpoint')
-    || text.includes('unsupported_path')
-    || text.includes('not_found_error')
-    || text.includes('unsupported legacy protocol')
-    || text.includes('please use /v1/')
-    || text.includes('does not allow /v1/')
-    || text.includes('unknown endpoint')
-    || text.includes('unsupported endpoint')
-    || text.includes('unsupported path')
-    || text.includes('unrecognized request url')
-    || text.includes('no route matched')
-    || text.includes('does not exist')
-  );
+  const rawText = errorText || '';
+  const requiredEndpoint = inferRequiredEndpointFromProtocolError(rawText);
+  if (requiredEndpoint) {
+    return endpoint !== requiredEndpoint;
+  }
+
+  return isEndpointDowngradeError(status, errorText);
 }
 
 function shouldRememberSuccessfulEndpoint(input: {
@@ -405,7 +400,7 @@ export function recordUpstreamEndpointFailure(input: {
     requestCapabilities: input.requestCapabilities,
   });
   if (!shouldUseEndpointRuntimeMemory(capabilityProfile)) return null;
-  if (!shouldBlockEndpointByError(input.status, input.errorText)) return null;
+  if (!shouldBlockEndpointByError(input.endpoint, input.status, input.errorText)) return null;
 
   const nowMs = Date.now();
   const key = buildEndpointRuntimeStateKey({
@@ -416,7 +411,7 @@ export function recordUpstreamEndpointFailure(input: {
   const state = getOrCreateEndpointRuntimeState(key, nowMs);
   state.blockedUntilMsByEndpoint[input.endpoint] = nowMs + ENDPOINT_RUNTIME_BLOCK_TTL_MS;
 
-  const suggestedEndpoint = inferSuggestedEndpointFromError(input.errorText);
+  const suggestedEndpoint = inferSuggestedEndpointFromError(input.endpoint, input.errorText);
   if (suggestedEndpoint && suggestedEndpoint !== input.endpoint) {
     state.preferredEndpoint = suggestedEndpoint;
     state.preferredUpdatedAtMs = nowMs;

src/server/services/upstreamRequestBuilder.test.ts

@@ -0,0 +1,177 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  buildClaudeCountTokensUpstreamRequest,
+  buildUpstreamEndpointRequest,
+} from './upstreamRequestBuilder.js';
+
+describe('upstreamRequestBuilder', () => {
+  it('normalizes single-message OpenAI requests to structured responses input', () => {
+    const request = buildUpstreamEndpointRequest({
+      endpoint: 'responses',
+      modelName: 'upstream-gpt',
+      stream: false,
+      tokenValue: 'sk-test',
+      sitePlatform: 'sub2api',
+      siteUrl: 'https://example.com',
+      openaiBody: {
+        model: 'gpt-5.2',
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+      downstreamFormat: 'openai',
+    });
+
+    expect(request.path).toBe('/v1/responses');
+    expect(request.headers.accept).toBe('application/json');
+    expect(request.body.input).toEqual([
+      {
+        type: 'message',
+        role: 'user',
+        content: [{ type: 'input_text', text: 'hello' }],
+      },
+    ]);
+    expect(request.body.store).toBe(false);
+  });
+
+  it('forces store=false for sub2api native responses passthrough bodies', () => {
+    const request = buildUpstreamEndpointRequest({
+      endpoint: 'responses',
+      modelName: 'upstream-gpt',
+      stream: true,
+      tokenValue: 'sk-test',
+      sitePlatform: 'sub2api',
+      siteUrl: 'https://example.com',
+      openaiBody: {},
+      downstreamFormat: 'responses',
+      responsesOriginalBody: {
+        model: 'gpt-5.2',
+        input: 'hello',
+        store: true,
+      },
+    });
+
+    expect(request.path).toBe('/v1/responses');
+    expect(request.headers.accept).toBe('text/event-stream');
+    expect(request.body.input).toEqual([
+      {
+        type: 'message',
+        role: 'user',
+        content: [{ type: 'input_text', text: 'hello' }],
+      },
+    ]);
+    expect(request.body.stream).toBe(true);
+    expect(request.body.store).toBe(false);
+  });
+
+  it('overrides downstream Accept so responses transport mode wins', () => {
+    const request = buildUpstreamEndpointRequest({
+      endpoint: 'responses',
+      modelName: 'upstream-gpt',
+      stream: true,
+      tokenValue: 'sk-test',
+      sitePlatform: 'openai',
+      siteUrl: 'https://example.com',
+      openaiBody: {
+        model: 'gpt-5.2',
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+      downstreamFormat: 'openai',
+      downstreamHeaders: {
+        accept: 'application/json',
+      },
+    });
+
+    expect(request.headers.accept).toBe('text/event-stream');
+  });
+
+  it('applies a sub2api-style allowlist to generic passthrough headers', () => {
+    const request = buildUpstreamEndpointRequest({
+      endpoint: 'chat',
+      modelName: 'upstream-gpt',
+      stream: false,
+      tokenValue: 'sk-test',
+      sitePlatform: 'sub2api',
+      siteUrl: 'https://example.com',
+      openaiBody: {
+        model: 'gpt-5.2',
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+      downstreamFormat: 'openai',
+      downstreamHeaders: {
+        accept: 'application/json',
+        'accept-language': 'zh-CN',
+        'user-agent': 'client-ua/1.0',
+        originator: 'codex_cli_rs',
+        session_id: 'session-123',
+        conversation_id: 'conversation-123',
+        'x-codex-turn-state': 'turn-state',
+        'x-codex-turn-metadata': 'turn-metadata',
+        origin: 'https://client.example',
+        referer: 'https://client.example/chat',
+        'x-forwarded-for': '203.0.113.1',
+        'x-real-ip': '203.0.113.2',
+        version: '0.202.0',
+        'x-test-header': 'drop-me',
+      },
+    });
+
+    expect(request.headers.accept).toBe('application/json');
+    expect(request.headers['accept-language']).toBe('zh-CN');
+    expect(request.headers['user-agent']).toBe('client-ua/1.0');
+    expect(request.headers.originator).toBe('codex_cli_rs');
+    expect(request.headers.session_id).toBe('session-123');
+    expect(request.headers.conversation_id).toBe('conversation-123');
+    expect(request.headers['x-codex-turn-state']).toBe('turn-state');
+    expect(request.headers['x-codex-turn-metadata']).toBe('turn-metadata');
+
+    expect(request.headers.origin).toBeUndefined();
+    expect(request.headers.referer).toBeUndefined();
+    expect(request.headers['x-forwarded-for']).toBeUndefined();
+    expect(request.headers['x-real-ip']).toBeUndefined();
+    expect(request.headers.version).toBeUndefined();
+    expect(request.headers['x-test-header']).toBeUndefined();
+  });
+
+  it('drops responses-style continuation fields before proxying Claude count_tokens upstream', () => {
+    const request = buildClaudeCountTokensUpstreamRequest({
+      modelName: 'claude-opus-4-6',
+      tokenValue: 'sk-test',
+      sitePlatform: 'claude',
+      claudeBody: {
+        model: 'claude-opus-4-6',
+        max_tokens: 256,
+        previous_response_id: 'resp_prev_1',
+        prompt_cache_key: 'cache-key-1',
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+    });
+
+    expect(request.body).toMatchObject({
+      model: 'claude-opus-4-6',
+      messages: [{ role: 'user' }],
+    });
+    expect(request.body).not.toHaveProperty('previous_response_id');
+    expect(request.body).not.toHaveProperty('prompt_cache_key');
+    expect(request.body).not.toHaveProperty('max_tokens');
+    expect(request.body).not.toHaveProperty('maxTokens');
+  });
+
+  it('merges body betas with existing anthropic-beta headers for Claude count_tokens', () => {
+    const request = buildClaudeCountTokensUpstreamRequest({
+      modelName: 'claude-opus-4-6',
+      tokenValue: 'sk-test',
+      sitePlatform: 'claude',
+      claudeBody: {
+        model: 'claude-opus-4-6',
+        betas: ['beta-from-body'],
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+      downstreamHeaders: {
+        'anthropic-beta': 'header-beta',
+      },
+    });
+
+    expect(request.headers['anthropic-beta']).toContain('header-beta');
+    expect(request.headers['anthropic-beta']).toContain('beta-from-body');
+  });
+});

src/server/services/upstreamRequestBuilder.ts

@@ -0,0 +1,837 @@
+import type { UpstreamEndpoint } from '../proxy-core/orchestration/upstreamRequest.js';
+import { resolveProviderProfile } from '../proxy-core/providers/registry.js';
+import { config } from '../config.js';
+import { applyPayloadRules } from './payloadRules.js';
+import type { DownstreamFormat } from '../transformers/shared/normalized.js';
+import {
+  convertOpenAiBodyToResponsesBody as convertOpenAiBodyToResponsesBodyViaTransformer,
+  sanitizeResponsesBodyForProxy as sanitizeResponsesBodyForProxyViaTransformer,
+} from '../transformers/openai/responses/conversion.js';
+import { normalizeCodexResponsesBodyForProxy } from '../transformers/openai/responses/codexCompatibility.js';
+import {
+  convertOpenAiBodyToAnthropicMessagesBody,
+  sanitizeAnthropicMessagesBody,
+} from '../transformers/anthropic/messages/conversion.js';
+import {
+  buildGeminiGenerateContentRequestFromOpenAi,
+} from '../transformers/gemini/generate-content/requestBridge.js';
+import {
+  buildClaudeRuntimeHeaders,
+  getInputHeader,
+  headerValueToString,
+} from '../proxy-core/providers/headerUtils.js';
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return !!value && typeof value === 'object';
+}
+
+function asTrimmedString(value: unknown): string {
+  return typeof value === 'string' ? value.trim() : '';
+}
+
+function resolveRequestedModelForPayloadRules(input: {
+  modelName: string;
+  openaiBody: Record<string, unknown>;
+  claudeOriginalBody?: Record<string, unknown>;
+  responsesOriginalBody?: Record<string, unknown>;
+}): string {
+  return (
+    asTrimmedString(input.responsesOriginalBody?.model)
+    || asTrimmedString(input.claudeOriginalBody?.model)
+    || asTrimmedString(input.openaiBody.model)
+    || asTrimmedString(input.modelName)
+  );
+}
+
+function normalizePlatformName(platform: unknown): string {
+  return asTrimmedString(platform).toLowerCase();
+}
+
+const HOP_BY_HOP_HEADERS = new Set([
+  'connection',
+  'keep-alive',
+  'proxy-authenticate',
+  'proxy-authorization',
+  'te',
+  'trailer',
+  'transfer-encoding',
+  'upgrade',
+]);
+
+const BLOCKED_PASSTHROUGH_HEADERS = new Set([
+  'host',
+  'content-type',
+  'content-length',
+  'accept-encoding',
+  'cookie',
+  'authorization',
+  'x-api-key',
+  'x-goog-api-key',
+  'sec-websocket-key',
+  'sec-websocket-version',
+  'sec-websocket-extensions',
+]);
+const GENERIC_PASSTHROUGH_ALLOWED_HEADERS = new Set([
+  'accept',
+  'accept-language',
+  'conversation-id',
+  'conversation_id',
+  'openai-beta',
+  'originator',
+  'session-id',
+  'session_id',
+  'user-agent',
+  'x-codex-beta-features',
+  'x-codex-turn-metadata',
+  'x-codex-turn-state',
+]);
+const METAPI_INTERNAL_HEADER_BLOCKLIST = new Set([
+  'x-metapi-tester-request',
+  'x-metapi-tester-forced-channel-id',
+  'x-metapi-responses-websocket-mode',
+  'x-metapi-responses-websocket-transport',
+]);
+
+const ANTIGRAVITY_RUNTIME_USER_AGENT = 'antigravity/1.19.6 darwin/arm64';
+
+function shouldSkipPassthroughHeader(key: string): boolean {
+  if (HOP_BY_HOP_HEADERS.has(key) || BLOCKED_PASSTHROUGH_HEADERS.has(key)) return true;
+  if (METAPI_INTERNAL_HEADER_BLOCKLIST.has(key)) return true;
+  if (key.startsWith('x-metapi-')) return true;
+  if (!GENERIC_PASSTHROUGH_ALLOWED_HEADERS.has(key)) return true;
+  return false;
+}
+
+function extractSafePassthroughHeaders(
+  headers?: Record<string, unknown>,
+): Record<string, string> {
+  if (!headers) return {};
+
+  const forwarded: Record<string, string> = {};
+  for (const [rawKey, rawValue] of Object.entries(headers)) {
+    const key = rawKey.toLowerCase();
+    if (!key || shouldSkipPassthroughHeader(key)) continue;
+
+    const value = headerValueToString(rawValue);
+    if (!value) continue;
+    forwarded[key] = value;
+  }
+
+  return forwarded;
+}
+
+function extractClaudePassthroughHeaders(
+  headers?: Record<string, unknown>,
+): Record<string, string> {
+  if (!headers) return {};
+
+  const forwarded: Record<string, string> = {};
+  for (const [rawKey, rawValue] of Object.entries(headers)) {
+    const key = rawKey.toLowerCase();
+    const shouldForward = (
+      key.startsWith('anthropic-')
+      || key.startsWith('x-claude-')
+      || key.startsWith('x-stainless-')
+    );
+    if (!shouldForward) continue;
+
+    const value = headerValueToString(rawValue);
+    if (!value) continue;
+    forwarded[key] = value;
+  }
+
+  return forwarded;
+}
+
+function extractResponsesPassthroughHeaders(
+  headers?: Record<string, unknown>,
+): Record<string, string> {
+  if (!headers) return {};
+
+  const forwarded: Record<string, string> = {};
+  for (const [rawKey, rawValue] of Object.entries(headers)) {
+    const key = rawKey.toLowerCase();
+    const shouldForward = (
+      key.startsWith('openai-')
+      || key.startsWith('x-openai-')
+      || key.startsWith('x-stainless-')
+      || key.startsWith('chatgpt-')
+      || key === 'originator'
+    );
+    if (!shouldForward) continue;
+
+    const value = headerValueToString(rawValue);
+    if (!value) continue;
+    forwarded[key] = value;
+  }
+
+  return forwarded;
+}
+
+function extractCodexPassthroughHeaders(
+  headers?: Record<string, unknown>,
+): Record<string, string> {
+  if (!headers) return {};
+
+  const forwarded: Record<string, string> = {};
+  for (const [rawKey, rawValue] of Object.entries(headers)) {
+    const key = rawKey.toLowerCase();
+    const shouldForward = (
+      key === 'version'
+      || key === 'x-responsesapi-include-timing-metrics'
+    );
+    if (!shouldForward) continue;
+
+    const value = headerValueToString(rawValue);
+    if (!value) continue;
+    forwarded[key] = value;
+  }
+
+  return forwarded;
+}
+
+function extractClaudeBetasFromBody(body: Record<string, unknown>): {
+  body: Record<string, unknown>;
+  betas: string[];
+} {
+  const next = { ...body };
+  const rawBetas = next.betas;
+  delete next.betas;
+
+  if (typeof rawBetas === 'string') {
+    return {
+      body: next,
+      betas: rawBetas.split(',').map((entry) => entry.trim()).filter(Boolean),
+    };
+  }
+
+  if (Array.isArray(rawBetas)) {
+    return {
+      body: next,
+      betas: rawBetas
+        .map((entry) => asTrimmedString(entry))
+        .filter(Boolean),
+    };
+  }
+
+  return {
+    body: next,
+    betas: [],
+  };
+}
+
+function stripClaudeMessagesContinuationFields(
+  body: Record<string, unknown>,
+): Record<string, unknown> {
+  const next = { ...body };
+  delete next.previous_response_id;
+  delete next.prompt_cache_key;
+  return next;
+}
+
+function buildAntigravityRuntimeHeaders(input: {
+  baseHeaders: Record<string, string>;
+  stream: boolean;
+}): Record<string, string> {
+  const headers: Record<string, string> = {
+    Authorization: input.baseHeaders.Authorization,
+    'Content-Type': 'application/json',
+    Accept: input.stream ? 'text/event-stream' : 'application/json',
+    'User-Agent': ANTIGRAVITY_RUNTIME_USER_AGENT,
+  };
+  return headers;
+}
+
+function ensureStreamAcceptHeader(
+  headers: Record<string, string>,
+  stream: boolean,
+): Record<string, string> {
+  if (!stream) return headers;
+
+  const existingAccept = (
+    headerValueToString(headers.accept)
+    || headerValueToString((headers as Record<string, unknown>).Accept)
+  );
+  if (existingAccept) return headers;
+
+  return {
+    ...headers,
+    accept: 'text/event-stream',
+  };
+}
+
+function ensureResponsesAcceptHeader(
+  headers: Record<string, string>,
+  input: {
+    stream: boolean;
+    sitePlatform?: string;
+  },
+): Record<string, string> {
+  const nextHeaders = { ...headers };
+  delete (nextHeaders as Record<string, unknown>).Accept;
+  delete (nextHeaders as Record<string, unknown>).accept;
+
+  if (input.stream) {
+    return {
+      ...nextHeaders,
+      accept: 'text/event-stream',
+    };
+  }
+  if (normalizePlatformName(input.sitePlatform) === 'sub2api') {
+    return {
+      ...nextHeaders,
+      accept: 'application/json',
+    };
+  }
+  return headers;
+}
+
+function normalizeResponsesFallbackChatFunctionTool(rawTool: unknown): Record<string, unknown> | null {
+  if (!isRecord(rawTool)) return null;
+  if (asTrimmedString(rawTool.type).toLowerCase() !== 'function') return null;
+
+  if (isRecord(rawTool.function)) {
+    const name = asTrimmedString(rawTool.function.name);
+    if (!name) return null;
+    return {
+      ...rawTool,
+      type: 'function',
+      function: {
+        ...rawTool.function,
+        name,
+      },
+    };
+  }
+
+  const name = asTrimmedString(rawTool.name);
+  if (!name) return null;
+
+  const fn: Record<string, unknown> = { name };
+  const description = asTrimmedString(rawTool.description);
+  if (description) fn.description = description;
+  if (rawTool.parameters !== undefined) fn.parameters = rawTool.parameters;
+  if (rawTool.strict !== undefined) fn.strict = rawTool.strict;
+
+  return {
+    type: 'function',
+    function: fn,
+  };
+}
+
+function normalizeResponsesFallbackChatToolChoice(
+  rawToolChoice: unknown,
+  allowedToolNames: Set<string>,
+): unknown {
+  if (rawToolChoice === undefined) return undefined;
+
+  if (typeof rawToolChoice === 'string') {
+    const normalized = rawToolChoice.trim().toLowerCase();
+    if (normalized === 'none') return 'none';
+    if (allowedToolNames.size <= 0) return undefined;
+    if (normalized === 'auto' || normalized === 'required') return normalized;
+    return undefined;
+  }
+
+  if (!isRecord(rawToolChoice)) return undefined;
+  if (asTrimmedString(rawToolChoice.type).toLowerCase() !== 'function') return undefined;
+
+  const nestedFunction = isRecord(rawToolChoice.function) ? rawToolChoice.function : null;
+  const name = asTrimmedString(nestedFunction?.name ?? rawToolChoice.name);
+  if (!name || !allowedToolNames.has(name)) return undefined;
+
+  return {
+    type: 'function',
+    function: {
+      ...(nestedFunction || {}),
+      name,
+    },
+  };
+}
+
+function sanitizeResponsesFallbackChatBody(
+  body: Record<string, unknown>,
+): Record<string, unknown> {
+  const next: Record<string, unknown> = { ...body };
+  const normalizedTools = Array.isArray(body.tools)
+    ? body.tools
+      .map((tool) => normalizeResponsesFallbackChatFunctionTool(tool))
+      .filter((tool): tool is Record<string, unknown> => !!tool)
+    : [];
+
+  if (normalizedTools.length > 0) {
+    next.tools = normalizedTools;
+  } else {
+    delete next.tools;
+  }
+
+  const allowedToolNames = new Set(
+    normalizedTools
+      .map((tool) => (
+        isRecord(tool.function)
+          ? asTrimmedString(tool.function.name)
+          : ''
+      ))
+      .filter((name) => name.length > 0),
+  );
+  const normalizedToolChoice = normalizeResponsesFallbackChatToolChoice(
+    body.tool_choice,
+    allowedToolNames,
+  );
+  if (normalizedToolChoice !== undefined) {
+    next.tool_choice = normalizedToolChoice;
+  } else {
+    delete next.tool_choice;
+  }
+
+  return next;
+}
+
+function toFiniteNumber(value: unknown): number | null {
+  return typeof value === 'number' && Number.isFinite(value) ? value : null;
+}
+
+function normalizeSub2ApiResponsesBodyForProxy(
+  body: Record<string, unknown>,
+  sitePlatform: string,
+): Record<string, unknown> {
+  if (sitePlatform !== 'sub2api') return body;
+  return {
+    ...body,
+    store: false,
+  };
+}
+
+export function buildUpstreamEndpointRequest(input: {
+  endpoint: UpstreamEndpoint;
+  modelName: string;
+  stream: boolean;
+  tokenValue: string;
+  oauthProvider?: string;
+  oauthProjectId?: string;
+  sitePlatform?: string;
+  siteUrl?: string;
+  openaiBody: Record<string, unknown>;
+  downstreamFormat: DownstreamFormat | 'responses';
+  claudeOriginalBody?: Record<string, unknown>;
+  forceNormalizeClaudeBody?: boolean;
+  responsesOriginalBody?: Record<string, unknown>;
+  downstreamHeaders?: Record<string, unknown>;
+  providerHeaders?: Record<string, string>;
+  codexSessionCacheKey?: string | null;
+  codexExplicitSessionId?: string | null;
+}): {
+  path: string;
+  headers: Record<string, string>;
+  body: Record<string, unknown>;
+  runtime?: {
+    executor: 'default' | 'codex' | 'gemini-cli' | 'antigravity' | 'claude';
+    modelName?: string;
+    stream?: boolean;
+    oauthProjectId?: string | null;
+    action?: 'generateContent' | 'streamGenerateContent' | 'countTokens';
+  };
+} {
+  const sitePlatform = normalizePlatformName(input.sitePlatform);
+  const providerProfile = resolveProviderProfile(sitePlatform);
+  const isClaudeUpstream = sitePlatform === 'claude';
+  const isGeminiUpstream = sitePlatform === 'gemini';
+  const isGeminiCliUpstream = sitePlatform === 'gemini-cli';
+  const isAntigravityUpstream = sitePlatform === 'antigravity';
+  const isInternalGeminiUpstream = isGeminiCliUpstream || isAntigravityUpstream;
+  const isClaudeOauthUpstream = isClaudeUpstream && input.oauthProvider === 'claude';
+
+  const resolveGeminiEndpointPath = (endpoint: UpstreamEndpoint): string => {
+    const normalizedSiteUrl = asTrimmedString(input.siteUrl).toLowerCase();
+    const openAiCompatBase = /\/openai(?:\/|$)/.test(normalizedSiteUrl);
+    if (openAiCompatBase) {
+      return endpoint === 'responses'
+        ? '/responses'
+        : '/chat/completions';
+    }
+    return endpoint === 'responses'
+      ? '/v1beta/openai/responses'
+      : '/v1beta/openai/chat/completions';
+  };
+
+  const resolveEndpointPath = (endpoint: UpstreamEndpoint): string => {
+    if (isGeminiUpstream) {
+      return resolveGeminiEndpointPath(endpoint);
+    }
+
+    if (sitePlatform === 'openai') {
+      if (endpoint === 'messages') return '/v1/messages';
+      if (endpoint === 'responses') return '/v1/responses';
+      return '/v1/chat/completions';
+    }
+
+    if (sitePlatform === 'codex') {
+      return '/responses';
+    }
+
+    if (sitePlatform === 'gemini-cli' || sitePlatform === 'antigravity') {
+      return input.stream
+        ? '/v1internal:streamGenerateContent?alt=sse'
+        : '/v1internal:generateContent';
+    }
+
+    if (sitePlatform === 'claude') {
+      return '/v1/messages';
+    }
+
+    if (endpoint === 'messages') return '/v1/messages';
+    if (endpoint === 'responses') return '/v1/responses';
+    return '/v1/chat/completions';
+  };
+
+  const passthroughHeaders = extractSafePassthroughHeaders(input.downstreamHeaders);
+  const codexPassthroughHeaders = sitePlatform === 'codex'
+    ? extractCodexPassthroughHeaders(input.downstreamHeaders)
+    : {};
+  const commonHeaders: Record<string, string> = {
+    ...passthroughHeaders,
+    ...codexPassthroughHeaders,
+    'Content-Type': 'application/json',
+    ...(input.providerHeaders || {}),
+  };
+  if (!isClaudeUpstream) {
+    commonHeaders.Authorization = `Bearer ${input.tokenValue}`;
+  }
+
+  const stripGeminiUnsupportedFields = (body: Record<string, unknown>) => {
+    const next = { ...body };
+    if (isGeminiUpstream || isInternalGeminiUpstream) {
+      for (const key of [
+        'frequency_penalty',
+        'presence_penalty',
+        'logit_bias',
+        'logprobs',
+        'top_logprobs',
+        'store',
+      ]) {
+        delete next[key];
+      }
+    }
+    return next;
+  };
+
+  const openaiBody = stripGeminiUnsupportedFields(input.openaiBody);
+  const runtime = {
+    executor: (
+      sitePlatform === 'codex'
+        ? 'codex'
+        : sitePlatform === 'gemini-cli'
+          ? 'gemini-cli'
+          : sitePlatform === 'antigravity'
+            ? 'antigravity'
+            : sitePlatform === 'claude'
+              ? 'claude'
+              : 'default'
+    ) as 'default' | 'codex' | 'gemini-cli' | 'antigravity' | 'claude',
+    modelName: input.modelName,
+    stream: input.stream,
+    oauthProjectId: asTrimmedString(input.oauthProjectId) || null,
+  };
+  const requestedModelForPayloadRules = resolveRequestedModelForPayloadRules(input);
+  const applyConfiguredPayloadRules = <T extends Record<string, unknown>>(body: T): T => (
+    applyPayloadRules({
+      rules: config.payloadRules,
+      payload: body,
+      modelName: input.modelName,
+      requestedModel: requestedModelForPayloadRules,
+      protocol: sitePlatform,
+    }) as T
+  );
+
+  if (isInternalGeminiUpstream) {
+    const instructions = (
+      input.downstreamFormat === 'responses'
+      && typeof input.responsesOriginalBody?.instructions === 'string'
+    )
+      ? input.responsesOriginalBody.instructions
+      : undefined;
+    const geminiRequest = buildGeminiGenerateContentRequestFromOpenAi({
+      body: openaiBody,
+      modelName: input.modelName,
+      instructions,
+    });
+    const configuredGeminiRequest = applyConfiguredPayloadRules(geminiRequest);
+    if (!providerProfile) {
+      throw new Error(`missing provider profile for platform: ${sitePlatform}`);
+    }
+    return providerProfile.prepareRequest({
+      endpoint: input.endpoint,
+      modelName: input.modelName,
+      stream: input.stream,
+      tokenValue: input.tokenValue,
+      oauthProvider: input.oauthProvider,
+      oauthProjectId: input.oauthProjectId,
+      sitePlatform,
+      baseHeaders: commonHeaders,
+      providerHeaders: input.providerHeaders,
+      body: configuredGeminiRequest,
+      action: input.stream ? 'streamGenerateContent' : 'generateContent',
+    });
+  }
+
+  if (input.endpoint === 'messages') {
+    const claudeHeaders = input.downstreamFormat === 'claude'
+      ? extractClaudePassthroughHeaders(input.downstreamHeaders)
+      : {};
+    const anthropicVersion = (
+      claudeHeaders['anthropic-version']
+      || '2023-06-01'
+    );
+    const nativeClaudeBody = (
+      input.downstreamFormat === 'claude'
+      && input.claudeOriginalBody
+      && input.forceNormalizeClaudeBody !== true
+    )
+      ? {
+        ...stripClaudeMessagesContinuationFields(input.claudeOriginalBody),
+        model: input.modelName,
+        stream: input.stream,
+      }
+      : null;
+    const normalizedClaudeBody = (
+      input.downstreamFormat === 'claude'
+      && input.claudeOriginalBody
+      && input.forceNormalizeClaudeBody === true
+    )
+      ? sanitizeAnthropicMessagesBody({
+        ...stripClaudeMessagesContinuationFields(input.claudeOriginalBody),
+        model: input.modelName,
+        stream: input.stream,
+      })
+      : null;
+    const sanitizedBody = nativeClaudeBody
+      ?? normalizedClaudeBody
+      ?? sanitizeAnthropicMessagesBody(
+        convertOpenAiBodyToAnthropicMessagesBody(openaiBody, input.modelName, input.stream),
+      );
+    const configuredClaudeBody = applyConfiguredPayloadRules(sanitizedBody);
+
+    if (providerProfile?.id === 'claude') {
+      return providerProfile.prepareRequest({
+        endpoint: 'messages',
+        modelName: input.modelName,
+        stream: input.stream,
+        tokenValue: input.tokenValue,
+        oauthProvider: input.oauthProvider,
+        oauthProjectId: input.oauthProjectId,
+        sitePlatform,
+        baseHeaders: commonHeaders,
+        claudeHeaders,
+        body: configuredClaudeBody,
+      });
+    }
+
+    const headers = buildClaudeRuntimeHeaders({
+      baseHeaders: commonHeaders,
+      claudeHeaders,
+      anthropicVersion,
+      stream: input.stream,
+      isClaudeOauthUpstream,
+      tokenValue: input.tokenValue,
+    });
+
+    return {
+      path: resolveEndpointPath('messages'),
+      headers,
+      body: configuredClaudeBody,
+      runtime,
+    };
+  }
+
+  if (input.endpoint === 'responses') {
+    const responsesWebsocketTransport = getInputHeader(
+      input.downstreamHeaders,
+      'x-metapi-responses-websocket-transport',
+    ) === '1';
+    const websocketMode = Object.entries(input.downstreamHeaders || {}).find(([rawKey]) => rawKey.trim().toLowerCase() === 'x-metapi-responses-websocket-mode');
+    const preserveWebsocketIncrementalMode = asTrimmedString(websocketMode?.[1]).toLowerCase() === 'incremental';
+    const responsesHeaders = input.downstreamFormat === 'responses'
+      ? extractResponsesPassthroughHeaders(input.downstreamHeaders)
+      : {};
+    const rawBody = (
+      input.downstreamFormat === 'responses' && input.responsesOriginalBody
+        ? {
+          ...stripGeminiUnsupportedFields(input.responsesOriginalBody),
+          model: input.modelName,
+          stream: input.stream,
+        }
+        : convertOpenAiBodyToResponsesBodyViaTransformer(openaiBody, input.modelName, input.stream)
+    );
+    const sanitizedResponsesBody = sanitizeResponsesBodyForProxyViaTransformer(rawBody, input.modelName, input.stream);
+    if (preserveWebsocketIncrementalMode && rawBody.generate === false) {
+      sanitizedResponsesBody.generate = false;
+    }
+    const body = normalizeCodexResponsesBodyForProxy(
+      sanitizedResponsesBody,
+      sitePlatform,
+    );
+    const configuredResponsesBody = normalizeCodexResponsesBodyForProxy(
+      normalizeSub2ApiResponsesBodyForProxy(
+        applyConfiguredPayloadRules(body),
+        sitePlatform,
+      ),
+      sitePlatform,
+    );
+
+    if (sitePlatform === 'codex') {
+      if (providerProfile?.id !== 'codex') {
+        throw new Error(`missing codex provider profile for platform: ${sitePlatform}`);
+      }
+      return providerProfile.prepareRequest({
+        endpoint: 'responses',
+        modelName: input.modelName,
+        stream: input.stream,
+        tokenValue: input.tokenValue,
+        oauthProvider: input.oauthProvider,
+        oauthProjectId: input.oauthProjectId,
+        sitePlatform,
+        baseHeaders: {
+          ...commonHeaders,
+          ...responsesHeaders,
+        },
+        providerHeaders: input.providerHeaders,
+        codexSessionCacheKey: input.codexSessionCacheKey,
+        codexExplicitSessionId: input.codexExplicitSessionId,
+        responsesWebsocketTransport,
+        body: configuredResponsesBody,
+      });
+    }
+
+    const headers = ensureResponsesAcceptHeader({
+      ...commonHeaders,
+      ...responsesHeaders,
+    }, {
+      stream: input.stream,
+      sitePlatform,
+    });
+    return {
+      path: resolveEndpointPath('responses'),
+      headers,
+      body: configuredResponsesBody,
+      runtime,
+    };
+  }
+
+  const headers = ensureStreamAcceptHeader(commonHeaders, input.stream);
+  const chatBody = {
+    ...openaiBody,
+    model: input.modelName,
+    stream: input.stream,
+  };
+  const configuredChatBody = applyConfiguredPayloadRules(
+    input.downstreamFormat === 'responses'
+      ? sanitizeResponsesFallbackChatBody(chatBody)
+      : chatBody,
+  );
+  return {
+    path: resolveEndpointPath('chat'),
+    headers,
+    body: configuredChatBody,
+    runtime,
+  };
+}
+
+export function buildClaudeCountTokensUpstreamRequest(input: {
+  modelName: string;
+  tokenValue: string;
+  oauthProvider?: string;
+  sitePlatform?: string;
+  claudeBody: Record<string, unknown>;
+  downstreamHeaders?: Record<string, unknown>;
+}): {
+  path: string;
+  headers: Record<string, string>;
+  body: Record<string, unknown>;
+  runtime: {
+    executor: 'claude';
+    modelName: string;
+    stream: false;
+    action: 'countTokens';
+  };
+} {
+  const sitePlatform = normalizePlatformName(input.sitePlatform);
+  const claudeHeaders = extractClaudePassthroughHeaders(input.downstreamHeaders);
+  const { body: bodyWithoutBetas, betas } = extractClaudeBetasFromBody({
+    ...stripClaudeMessagesContinuationFields(input.claudeBody),
+    model: input.modelName,
+  });
+  const sanitizedBody = sanitizeAnthropicMessagesBody(bodyWithoutBetas);
+  delete sanitizedBody.max_tokens;
+  delete sanitizedBody.maxTokens;
+  delete sanitizedBody.stream;
+  const providerProfile = resolveProviderProfile(sitePlatform);
+  const mergedBetas = [
+    ...asTrimmedString(claudeHeaders['anthropic-beta'])
+      .split(',')
+      .map((entry) => entry.trim())
+      .filter(Boolean),
+    ...betas,
+  ];
+  const effectiveClaudeHeaders = {
+    ...claudeHeaders,
+    ...(mergedBetas.length > 0
+      ? { 'anthropic-beta': Array.from(new Set(mergedBetas)).join(',') }
+      : {}),
+  };
+
+  if (providerProfile?.id === 'claude') {
+    const prepared = providerProfile.prepareRequest({
+      endpoint: 'messages',
+      modelName: input.modelName,
+      stream: false,
+      tokenValue: input.tokenValue,
+      oauthProvider: input.oauthProvider,
+      sitePlatform,
+      baseHeaders: {
+        'Content-Type': 'application/json',
+      },
+      claudeHeaders: effectiveClaudeHeaders,
+      body: sanitizedBody,
+      action: 'countTokens',
+    });
+
+    return {
+      path: prepared.path,
+      headers: prepared.headers,
+      body: prepared.body,
+      runtime: {
+        executor: 'claude',
+        modelName: input.modelName,
+        stream: false,
+        action: 'countTokens',
+      },
+    };
+  }
+
+  const anthropicVersion = (
+    effectiveClaudeHeaders['anthropic-version']
+    || '2023-06-01'
+  );
+  const isClaudeOauthUpstream = sitePlatform === 'claude' && input.oauthProvider === 'claude';
+  const headers = buildClaudeRuntimeHeaders({
+    baseHeaders: {
+      'Content-Type': 'application/json',
+    },
+    claudeHeaders: effectiveClaudeHeaders,
+    anthropicVersion,
+    stream: false,
+    isClaudeOauthUpstream,
+    tokenValue: input.tokenValue,
+  });
+
+  return {
+    path: '/v1/messages/count_tokens?beta=true',
+    headers,
+    body: sanitizedBody,
+    runtime: {
+      executor: 'claude',
+      modelName: input.modelName,
+      stream: false,
+      action: 'countTokens',
+    },
+  };
+}

src/server/shared/codexClientFamily.test.ts

@@ -0,0 +1,44 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  detectCodexOfficialClientApp,
+  inferCodexOfficialOriginator,
+  isCodexOfficialClientHeaders,
+} from './codexClientFamily.js';
+
+describe('codexClientFamily helpers', () => {
+  it('detects official codex clients from user-agent and canonicalizes their originator', () => {
+    expect(detectCodexOfficialClientApp({
+      'user-agent': 'Mozilla/5.0 codex_exec/1.2.3',
+    })).toEqual({
+      clientAppId: 'codex_exec',
+      clientAppName: 'Codex Exec',
+      originator: 'codex_exec',
+    });
+    expect(inferCodexOfficialOriginator({
+      'user-agent': 'Mozilla/5.0 codex_chatgpt_desktop/1.2.3',
+    })).toBe('codex_chatgpt_desktop');
+  });
+
+  it('canonicalizes legacy originator spellings for official codex desktop clients', () => {
+    expect(detectCodexOfficialClientApp({
+      originator: 'Codex Desktop',
+    })).toEqual({
+      clientAppId: 'codex_chatgpt_desktop',
+      clientAppName: 'Codex Desktop',
+      originator: 'codex_chatgpt_desktop',
+    });
+    expect(inferCodexOfficialOriginator({
+      originator: 'Codex Desktop',
+    })).toBe('codex_chatgpt_desktop');
+  });
+
+  it('keeps generic codex family detection broad without inventing a specific originator', () => {
+    expect(isCodexOfficialClientHeaders({
+      'user-agent': 'Codex 1.0',
+    })).toBe(true);
+    expect(inferCodexOfficialOriginator({
+      'user-agent': 'Codex 1.0',
+    })).toBeNull();
+  });
+});

src/server/shared/codexClientFamily.ts

@@ -0,0 +1,166 @@
+type CodexOfficialClientRule = {
+  clientAppId: string;
+  clientAppName: string;
+  originator: string;
+  userAgentPrefixes: readonly string[];
+  originatorPrefixes: readonly string[];
+};
+
+export type CodexOfficialClientApp = Pick<
+  CodexOfficialClientRule,
+  'clientAppId' | 'clientAppName' | 'originator'
+>;
+
+const CODEX_OFFICIAL_CLIENT_RULES: readonly CodexOfficialClientRule[] = [
+  {
+    clientAppId: 'codex_cli_rs',
+    clientAppName: 'Codex CLI',
+    originator: 'codex_cli_rs',
+    userAgentPrefixes: ['codex_cli_rs/'],
+    originatorPrefixes: ['codex_cli_rs'],
+  },
+  {
+    clientAppId: 'codex_vscode',
+    clientAppName: 'Codex VSCode',
+    originator: 'codex_vscode',
+    userAgentPrefixes: ['codex_vscode/'],
+    originatorPrefixes: ['codex_vscode'],
+  },
+  {
+    clientAppId: 'codex_app',
+    clientAppName: 'Codex App',
+    originator: 'codex_app',
+    userAgentPrefixes: ['codex_app/'],
+    originatorPrefixes: ['codex_app'],
+  },
+  {
+    clientAppId: 'codex_chatgpt_desktop',
+    clientAppName: 'Codex Desktop',
+    originator: 'codex_chatgpt_desktop',
+    userAgentPrefixes: ['codex_chatgpt_desktop/', 'codex desktop/'],
+    originatorPrefixes: ['codex_chatgpt_desktop', 'codex desktop'],
+  },
+  {
+    clientAppId: 'codex_atlas',
+    clientAppName: 'Codex Atlas',
+    originator: 'codex_atlas',
+    userAgentPrefixes: ['codex_atlas/'],
+    originatorPrefixes: ['codex_atlas'],
+  },
+  {
+    clientAppId: 'codex_exec',
+    clientAppName: 'Codex Exec',
+    originator: 'codex_exec',
+    userAgentPrefixes: ['codex_exec/'],
+    originatorPrefixes: ['codex_exec'],
+  },
+  {
+    clientAppId: 'codex_sdk_ts',
+    clientAppName: 'Codex SDK TS',
+    originator: 'codex_sdk_ts',
+    userAgentPrefixes: ['codex_sdk_ts/'],
+    originatorPrefixes: ['codex_sdk_ts'],
+  },
+];
+
+const CODEX_OFFICIAL_CLIENT_USER_AGENT_PREFIXES = [
+  ...new Set(CODEX_OFFICIAL_CLIENT_RULES.flatMap((rule) => rule.userAgentPrefixes)),
+  'codex ',
+] as const;
+
+const CODEX_OFFICIAL_CLIENT_ORIGINATOR_PREFIXES = [
+  'codex_',
+  'codex ',
+] as const;
+
+function normalizeHeaderValue(value: string): string {
+  return value.trim().toLowerCase();
+}
+
+function headerValueToStrings(value: unknown): string[] {
+  if (typeof value === 'string') {
+    const trimmed = value.trim();
+    return trimmed ? [trimmed] : [];
+  }
+
+  if (!Array.isArray(value)) return [];
+
+  const values: string[] = [];
+  for (const item of value) {
+    if (typeof item !== 'string') continue;
+    const trimmed = item.trim();
+    if (trimmed) values.push(trimmed);
+  }
+  return values;
+}
+
+function getHeaderValues(
+  headers: Record<string, unknown> | undefined,
+  targetKey: string,
+): string[] {
+  if (!headers) return [];
+  const normalizedTarget = normalizeHeaderValue(targetKey);
+  const values: string[] = [];
+  for (const [rawKey, rawValue] of Object.entries(headers)) {
+    if (normalizeHeaderValue(rawKey) !== normalizedTarget) continue;
+    values.push(...headerValueToStrings(rawValue));
+  }
+  return values;
+}
+
+function matchesCodexClientPrefixes(
+  value: string | string[] | null | undefined,
+  prefixes: readonly string[],
+): boolean {
+  const normalizedValues = Array.isArray(value)
+    ? value.map((item) => normalizeHeaderValue(item)).filter(Boolean)
+    : (typeof value === 'string' ? [normalizeHeaderValue(value)] : []).filter(Boolean);
+  if (normalizedValues.length <= 0) return false;
+
+  return normalizedValues.some((normalizedValue) => prefixes.some((prefix) => {
+    const normalizedPrefix = normalizeHeaderValue(prefix);
+    return !!normalizedPrefix
+      && (normalizedValue.startsWith(normalizedPrefix) || normalizedValue.includes(normalizedPrefix));
+  }));
+}
+
+export function detectCodexOfficialClientApp(
+  headers?: Record<string, unknown>,
+): CodexOfficialClientApp | null {
+  for (const rule of CODEX_OFFICIAL_CLIENT_RULES) {
+    const matchesOriginator = matchesCodexClientPrefixes(
+      getHeaderValues(headers, 'originator'),
+      rule.originatorPrefixes,
+    );
+    const matchesUserAgent = matchesCodexClientPrefixes(
+      getHeaderValues(headers, 'user-agent'),
+      rule.userAgentPrefixes,
+    );
+    if (!matchesOriginator && !matchesUserAgent) continue;
+    return {
+      clientAppId: rule.clientAppId,
+      clientAppName: rule.clientAppName,
+      originator: rule.originator,
+    };
+  }
+  return null;
+}
+
+export function inferCodexOfficialOriginator(
+  headers?: Record<string, unknown>,
+): string | null {
+  return detectCodexOfficialClientApp(headers)?.originator || null;
+}
+
+export function isCodexOfficialClientHeaders(
+  headers?: Record<string, unknown>,
+): boolean {
+  if (detectCodexOfficialClientApp(headers)) return true;
+  return matchesCodexClientPrefixes(
+    getHeaderValues(headers, 'originator'),
+    CODEX_OFFICIAL_CLIENT_ORIGINATOR_PREFIXES,
+  ) || matchesCodexClientPrefixes(
+    getHeaderValues(headers, 'user-agent'),
+    CODEX_OFFICIAL_CLIENT_USER_AGENT_PREFIXES,
+  );
+}

src/server/transformers/anthropic/messages/compatibility.ts

@@ -1,4 +1,5 @@
 import { type DownstreamFormat } from '../../shared/normalized.js';
+import { inferRequiredEndpointFromProtocolError } from '../../shared/endpointCompatibility.js';
 
 function isRecord(value: unknown): value is Record<string, unknown> {
   return !!value && typeof value === 'object' && !Array.isArray(value);
@@ -61,9 +62,9 @@ export function shouldRetryNormalizedMessagesBody(input: {
   if (input.downstreamFormat !== 'claude') return false;
   if (!input.endpointPath.includes('/v1/messages')) return false;
   if (input.status < 400 || input.status >= 500) return false;
-  return /messages\s+is\s+required/i.test(input.upstreamErrorText);
+  return inferRequiredEndpointFromProtocolError(input.upstreamErrorText) === 'messages';
 }
 
 export function isMessagesRequiredError(upstreamErrorText: string): boolean {
-  return /messages\s+is\s+required/i.test(upstreamErrorText);
+  return inferRequiredEndpointFromProtocolError(upstreamErrorText) === 'messages';
 }

src/server/transformers/anthropic/messages/index.test.ts

@@ -15,6 +15,7 @@ describe('anthropicMessagesTransformer protocol contract', () => {
       ],
       metadata: {
         user_id: 'user_0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef_account__session_11111111-2222-3333-4444-555555555555',
+        metapi_turn_state: 'turn-state-claude-1',
       },
     });
 
@@ -25,6 +26,10 @@ describe('anthropicMessagesTransformer protocol contract', () => {
       cliProfile: 'generic',
       requestedModel: 'claude-sonnet-4-5',
       stream: false,
+      continuation: {
+        sessionId: 'user_0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef_account__session_11111111-2222-3333-4444-555555555555',
+        turnState: 'turn-state-claude-1',
+      },
       messages: [
         {
           role: 'user',
@@ -98,12 +103,20 @@ describe('anthropicMessagesTransformer protocol contract', () => {
           parts: [{ type: 'text', text: 'count these tokens' }],
         },
       ],
+      continuation: {
+        sessionId: 'session-claude-bridge-1',
+        turnState: 'turn-state-claude-bridge-1',
+      },
       tools: [{ name: 'lookup', inputSchema: { type: 'object' } }],
     });
 
     expect(body).toMatchObject({
       model: 'claude-sonnet-4-5',
       max_tokens: 4096,
+      metadata: {
+        user_id: 'session-claude-bridge-1',
+        metapi_turn_state: 'turn-state-claude-bridge-1',
+      },
       messages: [
         {
           role: 'user',

src/server/transformers/anthropic/messages/index.ts

@@ -1,12 +1,18 @@
-import { canonicalRequestFromOpenAiBody, canonicalRequestToOpenAiChatBody } from '../../canonical/request.js';
 import type { CanonicalRequestEnvelope } from '../../canonical/types.js';
 import type { ProtocolBuildContext, ProtocolParseContext } from '../../contracts.js';
 import { type NormalizedFinalResponse, type NormalizedStreamEvent, type ParsedDownstreamChatRequest, type StreamTransformContext, type ClaudeDownstreamContext } from '../../shared/normalized.js';
 import { createChatEndpointStrategy } from '../../shared/chatEndpointStrategy.js';
 import { anthropicMessagesInbound } from './inbound.js';
-import { convertOpenAiBodyToAnthropicMessagesBody } from './conversion.js';
-import { anthropicMessagesOutbound } from './outbound.js';
-import { anthropicMessagesStream, consumeAnthropicSseEvent } from './stream.js';
+import {
+  buildCanonicalRequestToAnthropicMessagesBody,
+  parseAnthropicMessagesRequestToCanonical,
+} from './requestBridge.js';
+import {
+  anthropicMessagesResponseBridge,
+  buildNormalizedFinalToAnthropicMessagesBody,
+  normalizeAnthropicMessagesFinalToNormalized,
+} from './responseBridge.js';
+import { anthropicMessagesStream, consumeAnthropicSseEvent } from './streamBridge.js';
 import { anthropicMessagesUsage } from './usage.js';
 import { createAnthropicMessagesAggregateState } from './aggregator.js';
 import {
@@ -21,12 +27,12 @@ export {
   serializeAnthropicUpstreamFinalAsStream,
   serializeAnthropicRawSseEvent,
   syncAnthropicRawStreamStateFromEvent,
-} from './stream.js';
+} from './streamBridge.js';
 
 export const anthropicMessagesTransformer = {
   protocol: 'anthropic/messages' as const,
   inbound: anthropicMessagesInbound,
-  outbound: anthropicMessagesOutbound,
+  outbound: anthropicMessagesResponseBridge,
   stream: anthropicMessagesStream,
   usage: anthropicMessagesUsage,
   compatibility: {
@@ -41,42 +47,13 @@ export const anthropicMessagesTransformer = {
     body: unknown,
     ctx?: ProtocolParseContext,
   ): { value?: CanonicalRequestEnvelope; error?: { statusCode: number; payload: unknown } } {
-    const parsed = anthropicMessagesInbound.parse(body);
-    if (parsed.error) {
-      return { error: parsed.error };
-    }
-    if (!parsed.value) {
-      return {
-        error: {
-          statusCode: 400,
-          payload: {
-            error: {
-              message: 'invalid messages request',
-              type: 'invalid_request_error',
-            },
-          },
-        },
-      };
-    }
-
-    return {
-      value: canonicalRequestFromOpenAiBody({
-        body: parsed.value.parsed.upstreamBody,
-        surface: 'anthropic-messages',
-        cliProfile: ctx?.cliProfile,
-        operation: ctx?.operation,
-        metadata: ctx?.metadata,
-        passthrough: ctx?.passthrough,
-        continuation: ctx?.continuation,
-      }),
-    };
+    return parseAnthropicMessagesRequestToCanonical(body, ctx);
   },
   buildProtocolRequest(
     request: CanonicalRequestEnvelope,
     _ctx?: ProtocolBuildContext,
   ): Record<string, unknown> {
-    const openAiBody = canonicalRequestToOpenAiChatBody(request);
-    return convertOpenAiBodyToAnthropicMessagesBody(openAiBody, request.requestedModel, request.stream);
+    return buildCanonicalRequestToAnthropicMessagesBody(request);
   },
   transformRequest(body: unknown): ReturnType<typeof anthropicMessagesInbound.parse> {
     return anthropicMessagesInbound.parse(body);
@@ -88,7 +65,7 @@ export const anthropicMessagesTransformer = {
     return anthropicMessagesStream.createDownstreamContext();
   },
   transformFinalResponse(payload: unknown, modelName: string, fallbackText = ''): NormalizedFinalResponse {
-    return anthropicMessagesOutbound.normalizeFinal(payload, modelName, fallbackText);
+    return normalizeAnthropicMessagesFinalToNormalized(payload, modelName, fallbackText);
   },
   transformStreamEvent(payload: unknown, context: StreamTransformContext, modelName: string): NormalizedStreamEvent {
     return anthropicMessagesStream.normalizeEvent(payload, context, modelName);
@@ -108,9 +85,9 @@ export const anthropicMessagesTransformer = {
   },
   serializeFinalResponse(
     normalized: NormalizedFinalResponse,
-    usage: Parameters<typeof anthropicMessagesOutbound.serializeFinal>[1],
+    usage: Parameters<typeof buildNormalizedFinalToAnthropicMessagesBody>[1],
   ) {
-    return anthropicMessagesOutbound.serializeFinal(normalized, usage);
+    return buildNormalizedFinalToAnthropicMessagesBody(normalized, usage);
   },
   serializeUpstreamFinalAsStream(
     payload: unknown,
@@ -123,7 +100,7 @@ export const anthropicMessagesTransformer = {
       payload,
       modelName,
       fallbackText,
-      anthropicMessagesOutbound.normalizeFinal,
+      normalizeAnthropicMessagesFinalToNormalized,
       streamContext,
       claudeContext,
     );

src/server/transformers/anthropic/messages/outbound.ts

@@ -1,181 +1,7 @@
-import { normalizeUpstreamFinalResponse, toClaudeStopReason, type NormalizedFinalResponse } from '../../shared/normalized.js';
-import { decodeAnthropicReasoningSignature } from '../../shared/reasoningTransport.js';
-import { toAnthropicUsagePayload } from './usage.js';
-
-type AnthropicRecord = Record<string, unknown>;
-
-type AnthropicMessagesNormalizedFinalResponse = NormalizedFinalResponse & {
-  nativeContent?: AnthropicRecord[];
-  stopSequence?: string | null;
-  usagePayload?: AnthropicRecord;
-};
-
-function isRecord(value: unknown): value is AnthropicRecord {
-  return !!value && typeof value === 'object' && !Array.isArray(value);
-}
-
-function cloneJsonValue<T>(value: T): T {
-  if (Array.isArray(value)) {
-    return value.map((item) => cloneJsonValue(item)) as T;
-  }
-  if (isRecord(value)) {
-    return Object.fromEntries(
-      Object.entries(value).map(([key, item]) => [key, cloneJsonValue(item)]),
-    ) as T;
-  }
-  return value;
-}
-
-function asTrimmedString(value: unknown): string {
-  return typeof value === 'string' ? value.trim() : '';
-}
-
-function parseJsonLike(raw: string): unknown {
-  const trimmed = raw.trim();
-  if (!trimmed) return {};
-  try {
-    return JSON.parse(trimmed);
-  } catch {
-    return { value: raw };
-  }
-}
-
-function buildClaudeMessageId(sourceId: string): string {
-  if (sourceId.startsWith('msg_')) return sourceId;
-  const sanitized = sourceId.replace(/[^A-Za-z0-9_-]/g, '_');
-  return `msg_${sanitized || Date.now()}`;
-}
-
-function cleanAnthropicReasoningSignature(value: unknown): string | null {
-  const raw = typeof value === 'string' ? value.trim() : '';
-  if (!raw) return null;
-  const decoded = decodeAnthropicReasoningSignature(raw);
-  if (decoded) return decoded;
-  if (raw.startsWith('metapi:')) return null;
-  return raw;
-}
-
-function buildAnthropicContent(normalized: NormalizedFinalResponse): Array<Record<string, unknown>> {
-  const anthropicNormalized = normalized as AnthropicMessagesNormalizedFinalResponse;
-  if (Array.isArray(anthropicNormalized.nativeContent) && anthropicNormalized.nativeContent.length > 0) {
-    return anthropicNormalized.nativeContent.map((block) => {
-      const cloned = cloneJsonValue(block);
-      const blockType = asTrimmedString(cloned.type).toLowerCase();
-      if (blockType === 'thinking') {
-        const signature = cleanAnthropicReasoningSignature(cloned.signature);
-        if (signature) cloned.signature = signature;
-        else delete cloned.signature;
-      }
-      return cloned;
-    });
-  }
-
-  const contentBlocks: Array<Record<string, unknown>> = [];
-  const cleanSignature = cleanAnthropicReasoningSignature(normalized.reasoningSignature);
-  if (normalized.reasoningContent || cleanSignature) {
-    const thinkingBlock: Record<string, unknown> = {
-      type: 'thinking',
-      thinking: normalized.reasoningContent || '',
-    };
-    if (cleanSignature) {
-      thinkingBlock.signature = cleanSignature;
-    }
-    contentBlocks.push(thinkingBlock);
-  }
-  if (normalized.redactedReasoningContent) {
-    contentBlocks.push({
-      type: 'redacted_thinking',
-      data: normalized.redactedReasoningContent,
-    });
-  }
-  if (normalized.content) {
-    contentBlocks.push({
-      type: 'text',
-      text: normalized.content,
-    });
-  }
-  const toolCalls = Array.isArray(normalized.toolCalls) ? normalized.toolCalls : [];
-  for (let index = 0; index < toolCalls.length; index += 1) {
-    const toolCall = toolCalls[index];
-    contentBlocks.push({
-      type: 'tool_use',
-      id: toolCall.id || `toolu_${index}`,
-      name: toolCall.name || `tool_${index}`,
-      input: parseJsonLike(toolCall.arguments || ''),
-    });
-  }
-  return contentBlocks.length > 0
-    ? contentBlocks
-    : [{ type: 'text', text: '' }];
-}
-
-function extractNativeAnthropicContent(payload: unknown): AnthropicRecord[] | undefined {
-  if (!isRecord(payload) || !Array.isArray(payload.content)) return undefined;
-  const content = payload.content
-    .filter((block): block is AnthropicRecord => isRecord(block))
-    .map((block) => cloneJsonValue(block));
-  return content.length > 0 ? content : undefined;
-}
-
-function extractStopSequence(payload: unknown): string | null {
-  if (!isRecord(payload) || typeof payload.stop_sequence !== 'string') return null;
-  return payload.stop_sequence;
-}
-
-function extractUsagePayload(payload: unknown): AnthropicRecord | undefined {
-  if (!isRecord(payload) || !isRecord(payload.usage)) return undefined;
-  return cloneJsonValue(payload.usage);
-}
-
-function mergeUsagePayload(
-  usage: unknown,
-  normalized: AnthropicMessagesNormalizedFinalResponse,
-): AnthropicRecord {
-  const merged = toAnthropicUsagePayload(usage ?? normalized.usagePayload);
-  const usagePayload = normalized.usagePayload;
-  if (!usagePayload) return merged;
-
-  for (const [key, value] of Object.entries(usagePayload)) {
-    if (key === 'input_tokens' || key === 'output_tokens') continue;
-    if (key === 'cache_creation' && isRecord(value) && isRecord(merged.cache_creation)) {
-      merged.cache_creation = {
-        ...cloneJsonValue(value),
-        ...merged.cache_creation,
-      };
-      continue;
-    }
-    if (merged[key] === undefined) {
-      merged[key] = cloneJsonValue(value);
-    }
-  }
-
-  return merged;
-}
-
-export const anthropicMessagesOutbound = {
-  normalizeFinal(payload: unknown, modelName: string, fallbackText = ''): AnthropicMessagesNormalizedFinalResponse {
-    const nativeContent = extractNativeAnthropicContent(payload);
-    const stopSequence = extractStopSequence(payload);
-    const usagePayload = extractUsagePayload(payload);
-
-    return {
-      ...normalizeUpstreamFinalResponse(payload, modelName, fallbackText),
-      ...(nativeContent ? { nativeContent } : {}),
-      ...(stopSequence !== null ? { stopSequence } : {}),
-      ...(usagePayload ? { usagePayload } : {}),
-    };
-  },
-  serializeFinal(normalized: NormalizedFinalResponse, usage?: unknown) {
-    const anthropicNormalized = normalized as AnthropicMessagesNormalizedFinalResponse;
-    return {
-      id: buildClaudeMessageId(normalized.id),
-      type: 'message',
-      role: 'assistant',
-      model: normalized.model,
-      content: buildAnthropicContent(normalized),
-      stop_reason: toClaudeStopReason(normalized.finishReason),
-      stop_sequence: anthropicNormalized.stopSequence ?? null,
-      usage: mergeUsagePayload(usage, anthropicNormalized),
-    };
-  },
-};
+export {
+  anthropicMessagesOutbound,
+  anthropicMessagesResponseBridge,
+  buildNormalizedFinalToAnthropicMessagesBody,
+  normalizeAnthropicMessagesFinalToNormalized,
+} from './responseBridge.js';
+export type { AnthropicMessagesNormalizedFinalResponse } from './responseBridge.js';

src/server/transformers/anthropic/messages/requestBridge.test.ts

@@ -0,0 +1,58 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  buildCanonicalRequestToAnthropicMessagesBody,
+  parseAnthropicMessagesRequestToCanonical,
+} from './requestBridge.js';
+
+describe('anthropic messages request bridge', () => {
+  it('parses Anthropic messages bodies into canonical envelopes', () => {
+    const result = parseAnthropicMessagesRequestToCanonical({
+      model: 'claude-sonnet-4-5',
+      max_tokens: 256,
+      messages: [
+        {
+          role: 'user',
+          content: [{ type: 'text', text: 'hello' }],
+        },
+      ],
+      metadata: {
+        user_id: 'session-claude-1',
+        metapi_turn_state: 'turn-state-claude-1',
+      },
+    });
+
+    expect(result.error).toBeUndefined();
+    expect(result.value).toMatchObject({
+      surface: 'anthropic-messages',
+      requestedModel: 'claude-sonnet-4-5',
+      continuation: {
+        sessionId: 'session-claude-1',
+        turnState: 'turn-state-claude-1',
+      },
+    });
+  });
+
+  it('builds Anthropic messages bodies from canonical envelopes', () => {
+    const body = buildCanonicalRequestToAnthropicMessagesBody({
+      operation: 'generate',
+      surface: 'anthropic-messages',
+      cliProfile: 'claude_code',
+      requestedModel: 'claude-sonnet-4-5',
+      stream: false,
+      messages: [{ role: 'user', parts: [{ type: 'text', text: 'hello' }] }],
+      continuation: {
+        sessionId: 'session-claude-1',
+        turnState: 'turn-state-claude-1',
+      },
+    });
+
+    expect(body).toMatchObject({
+      model: 'claude-sonnet-4-5',
+      metadata: {
+        user_id: 'session-claude-1',
+        metapi_turn_state: 'turn-state-claude-1',
+      },
+    });
+  });
+});

src/server/transformers/anthropic/messages/requestBridge.ts

@@ -0,0 +1,47 @@
+import { canonicalRequestFromOpenAiBody, canonicalRequestToOpenAiChatBody } from '../../canonical/openAiRequestBridge.js';
+import type { CanonicalRequestEnvelope } from '../../canonical/types.js';
+import type { ProtocolParseContext } from '../../contracts.js';
+import { anthropicMessagesInbound } from './inbound.js';
+import { convertOpenAiBodyToAnthropicMessagesBody } from './conversion.js';
+
+export function parseAnthropicMessagesRequestToCanonical(
+  body: unknown,
+  ctx?: ProtocolParseContext,
+): { value?: CanonicalRequestEnvelope; error?: { statusCode: number; payload: unknown } } {
+  const parsed = anthropicMessagesInbound.parse(body);
+  if (parsed.error) {
+    return { error: parsed.error };
+  }
+  if (!parsed.value) {
+    return {
+      error: {
+        statusCode: 400,
+        payload: {
+          error: {
+            message: 'invalid messages request',
+            type: 'invalid_request_error',
+          },
+        },
+      },
+    };
+  }
+
+  return {
+    value: canonicalRequestFromOpenAiBody({
+      body: parsed.value.parsed.upstreamBody,
+      surface: 'anthropic-messages',
+      cliProfile: ctx?.cliProfile,
+      operation: ctx?.operation,
+      metadata: ctx?.metadata,
+      passthrough: ctx?.passthrough,
+      continuation: ctx?.continuation,
+    }),
+  };
+}
+
+export function buildCanonicalRequestToAnthropicMessagesBody(
+  request: CanonicalRequestEnvelope,
+): Record<string, unknown> {
+  const openAiBody = canonicalRequestToOpenAiChatBody(request);
+  return convertOpenAiBodyToAnthropicMessagesBody(openAiBody, request.requestedModel, request.stream);
+}

src/server/transformers/anthropic/messages/responseBridge.test.ts

@@ -0,0 +1,225 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  anthropicMessagesResponseBridge,
+  buildNormalizedFinalToAnthropicMessagesBody,
+  normalizeAnthropicMessagesFinalToNormalized,
+} from './responseBridge.js';
+import { anthropicMessagesOutbound } from './outbound.js';
+import { anthropicMessagesUsage } from './usage.js';
+
+describe('anthropic messages response bridge', () => {
+  it('normalizes native anthropic finals into normalized final responses', () => {
+    const normalized = normalizeAnthropicMessagesFinalToNormalized({
+      id: 'msg_native_1',
+      type: 'message',
+      role: 'assistant',
+      model: 'claude-test',
+      content: [
+        {
+          type: 'thinking',
+          thinking: 'plan first',
+          signature: 'sig-native',
+        },
+        {
+          type: 'text',
+          text: 'done',
+        },
+      ],
+      stop_reason: 'end_turn',
+      stop_sequence: null,
+      usage: {
+        input_tokens: 100,
+        output_tokens: 30,
+      },
+    }, 'claude-test');
+
+    expect(normalized).toMatchObject({
+      id: 'msg_native_1',
+      model: 'claude-test',
+      content: 'done',
+      reasoningContent: 'plan first',
+    });
+  });
+
+  it('builds anthropic messages payloads from normalized final responses', () => {
+    const upstreamPayload = {
+      id: 'msg_native_1',
+      type: 'message',
+      role: 'assistant',
+      model: 'claude-test',
+      content: [
+        {
+          type: 'thinking',
+          thinking: 'plan first',
+          signature: 'sig-native',
+        },
+        {
+          type: 'text',
+          text: 'done',
+        },
+      ],
+      stop_reason: 'end_turn',
+      stop_sequence: null,
+      usage: {
+        input_tokens: 100,
+        output_tokens: 30,
+      },
+    };
+
+    const normalized = normalizeAnthropicMessagesFinalToNormalized(upstreamPayload, 'claude-test');
+    const payload = buildNormalizedFinalToAnthropicMessagesBody(
+      normalized,
+      anthropicMessagesUsage.fromPayload(upstreamPayload),
+    );
+
+    expect(payload).toEqual({
+      id: 'msg_native_1',
+      type: 'message',
+      role: 'assistant',
+      model: 'claude-test',
+      content: [
+        {
+          type: 'thinking',
+          thinking: 'plan first',
+          signature: 'sig-native',
+        },
+        {
+          type: 'text',
+          text: 'done',
+        },
+      ],
+      stop_reason: 'end_turn',
+      stop_sequence: null,
+      usage: {
+        input_tokens: 100,
+        output_tokens: 30,
+      },
+    });
+  });
+
+  it('preserves responses output-item ordering when serializing to anthropic blocks', () => {
+    const normalized = normalizeAnthropicMessagesFinalToNormalized({
+      id: 'resp_order_1',
+      object: 'response',
+      model: 'gpt-5',
+      status: 'completed',
+      output: [
+        {
+          id: 'rs_reasoning_1',
+          type: 'reasoning',
+          encrypted_content: 'metapi:anthropic-signature:sig-resp-1',
+          summary: [
+            { type: 'summary_text', text: 'plan first' },
+          ],
+        },
+        {
+          id: 'fc_weather_1',
+          type: 'function_call',
+          call_id: 'call_weather_1',
+          name: 'lookup_weather',
+          arguments: '{"city":"Paris"}',
+        },
+        {
+          id: 'msg_order_1',
+          type: 'message',
+          role: 'assistant',
+          status: 'completed',
+          content: [
+            { type: 'output_text', text: 'done' },
+          ],
+        },
+      ],
+      usage: {
+        input_tokens: 12,
+        output_tokens: 6,
+      },
+    }, 'gpt-5');
+
+    const payload = buildNormalizedFinalToAnthropicMessagesBody(normalized, {
+      promptTokens: 12,
+      completionTokens: 6,
+      totalTokens: 18,
+    });
+
+    expect(payload.content).toEqual([
+      {
+        type: 'thinking',
+        thinking: 'plan first',
+        signature: 'sig-resp-1',
+      },
+      {
+        type: 'tool_use',
+        id: 'call_weather_1',
+        name: 'lookup_weather',
+        input: { city: 'Paris' },
+      },
+      {
+        type: 'text',
+        text: 'done',
+      },
+    ]);
+  });
+
+  it('maps responses web_search_call items into anthropic server tool blocks', () => {
+    const normalized = normalizeAnthropicMessagesFinalToNormalized({
+      id: 'resp_search_1',
+      object: 'response',
+      model: 'gpt-5',
+      status: 'completed',
+      output: [
+        {
+          id: 'ws_1',
+          type: 'web_search_call',
+          status: 'completed',
+          action: {
+            query: 'weather in Paris',
+          },
+        },
+        {
+          id: 'msg_search_1',
+          type: 'message',
+          role: 'assistant',
+          status: 'completed',
+          content: [
+            { type: 'output_text', text: 'It is sunny.' },
+          ],
+        },
+      ],
+      usage: {
+        input_tokens: 9,
+        output_tokens: 4,
+      },
+    }, 'gpt-5');
+
+    const payload = buildNormalizedFinalToAnthropicMessagesBody(normalized, {
+      promptTokens: 9,
+      completionTokens: 4,
+      totalTokens: 13,
+    });
+
+    expect(payload.content).toEqual([
+      {
+        type: 'server_tool_use',
+        id: 'srvtoolu_ws_1',
+        name: 'web_search',
+        input: {
+          query: 'weather in Paris',
+        },
+      },
+      {
+        type: 'web_search_tool_result',
+        tool_use_id: 'srvtoolu_ws_1',
+        content: [],
+      },
+      {
+        type: 'text',
+        text: 'It is sunny.',
+      },
+    ]);
+  });
+
+  it('keeps the outbound facade pointed at the response bridge object', () => {
+    expect(anthropicMessagesOutbound).toBe(anthropicMessagesResponseBridge);
+  });
+});

src/server/transformers/anthropic/messages/responseBridge.ts

@@ -0,0 +1,273 @@
+import { normalizeUpstreamFinalResponse, toClaudeStopReason, type NormalizedFinalResponse } from '../../shared/normalized.js';
+import { decodeAnthropicReasoningSignature } from '../../shared/reasoningTransport.js';
+import { toAnthropicUsagePayload } from './usage.js';
+
+type AnthropicRecord = Record<string, unknown>;
+
+export type AnthropicMessagesNormalizedFinalResponse = NormalizedFinalResponse & {
+  nativeContent?: AnthropicRecord[];
+  stopSequence?: string | null;
+  usagePayload?: AnthropicRecord;
+};
+
+function isRecord(value: unknown): value is AnthropicRecord {
+  return !!value && typeof value === 'object' && !Array.isArray(value);
+}
+
+function cloneJsonValue<T>(value: T): T {
+  if (Array.isArray(value)) {
+    return value.map((item) => cloneJsonValue(item)) as T;
+  }
+  if (isRecord(value)) {
+    return Object.fromEntries(
+      Object.entries(value).map(([key, item]) => [key, cloneJsonValue(item)]),
+    ) as T;
+  }
+  return value;
+}
+
+function asTrimmedString(value: unknown): string {
+  return typeof value === 'string' ? value.trim() : '';
+}
+
+function parseJsonLike(raw: string): unknown {
+  const trimmed = raw.trim();
+  if (!trimmed) return {};
+  try {
+    return JSON.parse(trimmed);
+  } catch {
+    return { value: raw };
+  }
+}
+
+function buildClaudeMessageId(sourceId: string): string {
+  if (sourceId.startsWith('msg_')) return sourceId;
+  const sanitized = sourceId.replace(/[^A-Za-z0-9_-]/g, '_');
+  return `msg_${sanitized || Date.now()}`;
+}
+
+function cleanAnthropicReasoningSignature(value: unknown): string | null {
+  const raw = typeof value === 'string' ? value.trim() : '';
+  if (!raw) return null;
+  const decoded = decodeAnthropicReasoningSignature(raw);
+  if (decoded) return decoded;
+  if (raw.startsWith('metapi:')) return null;
+  return raw;
+}
+
+function buildAnthropicContent(normalized: NormalizedFinalResponse): Array<Record<string, unknown>> {
+  const anthropicNormalized = normalized as AnthropicMessagesNormalizedFinalResponse;
+  if (Array.isArray(anthropicNormalized.nativeContent) && anthropicNormalized.nativeContent.length > 0) {
+    return anthropicNormalized.nativeContent.map((block) => {
+      const cloned = cloneJsonValue(block);
+      const blockType = asTrimmedString(cloned.type).toLowerCase();
+      if (blockType === 'thinking') {
+        const signature = cleanAnthropicReasoningSignature(cloned.signature);
+        if (signature) cloned.signature = signature;
+        else delete cloned.signature;
+      }
+      return cloned;
+    });
+  }
+
+  const contentBlocks: Array<Record<string, unknown>> = [];
+  const cleanSignature = cleanAnthropicReasoningSignature(normalized.reasoningSignature);
+  if (normalized.reasoningContent || cleanSignature) {
+    const thinkingBlock: Record<string, unknown> = {
+      type: 'thinking',
+      thinking: normalized.reasoningContent || '',
+    };
+    if (cleanSignature) {
+      thinkingBlock.signature = cleanSignature;
+    }
+    contentBlocks.push(thinkingBlock);
+  }
+  if (normalized.redactedReasoningContent) {
+    contentBlocks.push({
+      type: 'redacted_thinking',
+      data: normalized.redactedReasoningContent,
+    });
+  }
+  if (normalized.content) {
+    contentBlocks.push({
+      type: 'text',
+      text: normalized.content,
+    });
+  }
+  const toolCalls = Array.isArray(normalized.toolCalls) ? normalized.toolCalls : [];
+  for (let index = 0; index < toolCalls.length; index += 1) {
+    const toolCall = toolCalls[index];
+    contentBlocks.push({
+      type: 'tool_use',
+      id: toolCall.id || `toolu_${index}`,
+      name: toolCall.name || `tool_${index}`,
+      input: parseJsonLike(toolCall.arguments || ''),
+    });
+  }
+  return contentBlocks.length > 0
+    ? contentBlocks
+    : [{ type: 'text', text: '' }];
+}
+
+function extractNativeAnthropicContent(payload: unknown): AnthropicRecord[] | undefined {
+  if (!isRecord(payload) || !Array.isArray(payload.content)) return undefined;
+  const content = payload.content
+    .filter((block): block is AnthropicRecord => isRecord(block))
+    .map((block) => cloneJsonValue(block));
+  return content.length > 0 ? content : undefined;
+}
+
+function extractAnthropicContentFromResponsesPayload(payload: unknown): AnthropicRecord[] | undefined {
+  if (!isRecord(payload) || !Array.isArray(payload.output)) return undefined;
+
+  const contentBlocks: AnthropicRecord[] = [];
+
+  for (const outputItem of payload.output) {
+    if (!isRecord(outputItem)) continue;
+    const itemType = asTrimmedString(outputItem.type).toLowerCase();
+
+    if (itemType === 'reasoning') {
+      const summary = Array.isArray(outputItem.summary) ? outputItem.summary : [];
+      const thinking = summary
+        .filter((part): part is AnthropicRecord => isRecord(part))
+        .map((part) => asTrimmedString(part.text))
+        .filter((text) => text.length > 0)
+        .join('');
+      const block: AnthropicRecord = {
+        type: 'thinking',
+        thinking,
+      };
+      const signature = cleanAnthropicReasoningSignature(outputItem.encrypted_content);
+      if (signature) {
+        block.signature = signature;
+      }
+      contentBlocks.push(block);
+      continue;
+    }
+
+    if (itemType === 'message') {
+      const parts = Array.isArray(outputItem.content) ? outputItem.content : [];
+      for (const part of parts) {
+        if (!isRecord(part)) continue;
+        const partType = asTrimmedString(part.type).toLowerCase();
+        if ((partType === 'output_text' || partType === 'text') && typeof part.text === 'string') {
+          contentBlocks.push({
+            type: 'text',
+            text: part.text,
+          });
+        }
+      }
+      continue;
+    }
+
+    if (itemType === 'function_call') {
+      const toolUseId = asTrimmedString(outputItem.call_id) || asTrimmedString(outputItem.id) || `toolu_${contentBlocks.length}`;
+      contentBlocks.push({
+        type: 'tool_use',
+        id: toolUseId,
+        name: asTrimmedString(outputItem.name) || 'tool',
+        input: parseJsonLike(typeof outputItem.arguments === 'string' ? outputItem.arguments : ''),
+      });
+      continue;
+    }
+
+    if (itemType === 'web_search_call') {
+      const sourceId = asTrimmedString(outputItem.id) || `${contentBlocks.length}`;
+      const toolUseId = `srvtoolu_${sourceId}`;
+      const action = isRecord(outputItem.action) ? outputItem.action : null;
+      const query = asTrimmedString(action?.query);
+
+      contentBlocks.push({
+        type: 'server_tool_use',
+        id: toolUseId,
+        name: 'web_search',
+        input: {
+          query,
+        },
+      });
+      contentBlocks.push({
+        type: 'web_search_tool_result',
+        tool_use_id: toolUseId,
+        content: [],
+      });
+    }
+  }
+
+  return contentBlocks.length > 0 ? contentBlocks : undefined;
+}
+
+function extractStopSequence(payload: unknown): string | null {
+  if (!isRecord(payload) || typeof payload.stop_sequence !== 'string') return null;
+  return payload.stop_sequence;
+}
+
+function extractUsagePayload(payload: unknown): AnthropicRecord | undefined {
+  if (!isRecord(payload) || !isRecord(payload.usage)) return undefined;
+  return cloneJsonValue(payload.usage);
+}
+
+function mergeUsagePayload(
+  usage: unknown,
+  normalized: AnthropicMessagesNormalizedFinalResponse,
+): AnthropicRecord {
+  const merged = toAnthropicUsagePayload(usage ?? normalized.usagePayload);
+  const usagePayload = normalized.usagePayload;
+  if (!usagePayload) return merged;
+
+  for (const [key, value] of Object.entries(usagePayload)) {
+    if (key === 'input_tokens' || key === 'output_tokens') continue;
+    if (key === 'cache_creation' && isRecord(value) && isRecord(merged.cache_creation)) {
+      merged.cache_creation = {
+        ...cloneJsonValue(value),
+        ...merged.cache_creation,
+      };
+      continue;
+    }
+    if (merged[key] === undefined) {
+      merged[key] = cloneJsonValue(value);
+    }
+  }
+
+  return merged;
+}
+
+export function normalizeAnthropicMessagesFinalToNormalized(
+  payload: unknown,
+  modelName: string,
+  fallbackText = '',
+): AnthropicMessagesNormalizedFinalResponse {
+  const nativeContent = extractNativeAnthropicContent(payload) ?? extractAnthropicContentFromResponsesPayload(payload);
+  const stopSequence = extractStopSequence(payload);
+  const usagePayload = extractUsagePayload(payload);
+
+  return {
+    ...normalizeUpstreamFinalResponse(payload, modelName, fallbackText),
+    ...(nativeContent ? { nativeContent } : {}),
+    ...(stopSequence !== null ? { stopSequence } : {}),
+    ...(usagePayload ? { usagePayload } : {}),
+  };
+}
+
+export function buildNormalizedFinalToAnthropicMessagesBody(
+  normalized: NormalizedFinalResponse,
+  usage?: unknown,
+) {
+  const anthropicNormalized = normalized as AnthropicMessagesNormalizedFinalResponse;
+  return {
+    id: buildClaudeMessageId(normalized.id),
+    type: 'message',
+    role: 'assistant',
+    model: normalized.model,
+    content: buildAnthropicContent(normalized),
+    stop_reason: toClaudeStopReason(normalized.finishReason),
+    stop_sequence: anthropicNormalized.stopSequence ?? null,
+    usage: mergeUsagePayload(usage, anthropicNormalized),
+  };
+}
+
+export const anthropicMessagesResponseBridge = {
+  normalizeFinal: normalizeAnthropicMessagesFinalToNormalized,
+  serializeFinal: buildNormalizedFinalToAnthropicMessagesBody,
+};
+
+export const anthropicMessagesOutbound = anthropicMessagesResponseBridge;

src/server/transformers/anthropic/messages/streamBridge.test.ts

@@ -0,0 +1,234 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  anthropicMessagesStream,
+  consumeAnthropicSseEvent,
+  serializeAnthropicUpstreamFinalAsStream,
+} from './streamBridge.js';
+import { normalizeAnthropicMessagesFinalToNormalized } from './responseBridge.js';
+
+describe('anthropic messages stream bridge', () => {
+  it('passes through recognized raw anthropic SSE events and marks completion on message_stop', () => {
+    const streamContext = anthropicMessagesStream.createContext('claude-test');
+    const downstreamContext = anthropicMessagesStream.createDownstreamContext();
+
+    const result = consumeAnthropicSseEvent(
+      {
+        event: 'message_stop',
+        data: JSON.stringify({ type: 'message_stop' }),
+      },
+      streamContext,
+      downstreamContext,
+      'claude-test',
+    );
+
+    expect(result.handled).toBe(true);
+    expect(result.done).toBe(true);
+    expect(result.lines.join('')).toContain('event: message_stop');
+  });
+
+  it('serializes normalized upstream finals back into anthropic SSE blocks', () => {
+    const streamContext = anthropicMessagesStream.createContext('claude-test');
+    const downstreamContext = anthropicMessagesStream.createDownstreamContext();
+
+    const lines = serializeAnthropicUpstreamFinalAsStream(
+      {
+        id: 'msg_native_1',
+        type: 'message',
+        role: 'assistant',
+        model: 'claude-test',
+        content: [
+          {
+            type: 'thinking',
+            thinking: 'plan first',
+            signature: 'sig-native',
+          },
+          {
+            type: 'text',
+            text: 'done',
+          },
+        ],
+        stop_reason: 'end_turn',
+        stop_sequence: null,
+        usage: {
+          input_tokens: 100,
+          output_tokens: 30,
+        },
+      },
+      'claude-test',
+      '',
+      normalizeAnthropicMessagesFinalToNormalized,
+      streamContext,
+      downstreamContext,
+    );
+
+    const parsed = anthropicMessagesStream.pullSseEvents(lines.join(''));
+    expect(parsed.events.some((event) => event.event === 'message_start')).toBe(true);
+    expect(parsed.events.some((event) => event.event === 'message_stop')).toBe(true);
+    expect(lines.join('')).toContain('sig-native');
+  });
+
+  it('preserves whitespace in anthropic streaming text and thinking deltas', () => {
+    const streamContext = anthropicMessagesStream.createContext('claude-test');
+    const downstreamContext = anthropicMessagesStream.createDownstreamContext();
+
+    const textResult = consumeAnthropicSseEvent(
+      {
+        event: 'content_block_delta',
+        data: JSON.stringify({
+          type: 'content_block_delta',
+          index: 0,
+          delta: {
+            type: 'text_delta',
+            text: '  padded text  ',
+          },
+        }),
+      },
+      streamContext,
+      downstreamContext,
+      'claude-test',
+    );
+    const thinkingResult = consumeAnthropicSseEvent(
+      {
+        event: 'content_block_delta',
+        data: JSON.stringify({
+          type: 'content_block_delta',
+          index: 0,
+          delta: {
+            type: 'thinking_delta',
+            thinking: '  padded thinking  ',
+          },
+        }),
+      },
+      streamContext,
+      downstreamContext,
+      'claude-test',
+    );
+
+    expect(textResult.lines.join('')).toContain('  padded text  ');
+    expect(thinkingResult.lines.join('')).toContain('  padded thinking  ');
+  });
+
+  it('preserves partial_json spacing for streaming tool-call deltas', () => {
+    const streamContext = anthropicMessagesStream.createContext('claude-test');
+    const downstreamContext = anthropicMessagesStream.createDownstreamContext();
+
+    const result = consumeAnthropicSseEvent(
+      {
+        event: 'content_block_delta',
+        data: JSON.stringify({
+          type: 'content_block_delta',
+          index: 0,
+          delta: {
+            type: 'input_json_delta',
+            partial_json: '{ "city": "Paris" }',
+          },
+        }),
+      },
+      streamContext,
+      downstreamContext,
+      'claude-test',
+    );
+
+    expect(result.lines.join('')).toContain('partial_json":"{ \\"city\\": \\"Paris\\" }"');
+  });
+
+  it('preserves responses output-item ordering when streaming anthropic fallback blocks', () => {
+    const streamContext = anthropicMessagesStream.createContext('claude-test');
+    const downstreamContext = anthropicMessagesStream.createDownstreamContext();
+
+    const lines = serializeAnthropicUpstreamFinalAsStream(
+      {
+        id: 'resp_stream_order_1',
+        object: 'response',
+        model: 'gpt-5',
+        status: 'completed',
+        output: [
+          {
+            id: 'rs_reasoning_stream_1',
+            type: 'reasoning',
+            encrypted_content: 'metapi:anthropic-signature:sig-stream-1',
+            summary: [{ type: 'summary_text', text: 'plan first' }],
+          },
+          {
+            id: 'fc_stream_1',
+            type: 'function_call',
+            call_id: 'call_stream_1',
+            name: 'lookup_weather',
+            arguments: '{"city":"Paris"}',
+          },
+          {
+            id: 'msg_stream_1',
+            type: 'message',
+            role: 'assistant',
+            status: 'completed',
+            content: [{ type: 'output_text', text: 'done' }],
+          },
+        ],
+        usage: {
+          input_tokens: 10,
+          output_tokens: 4,
+        },
+      },
+      'gpt-5',
+      '',
+      normalizeAnthropicMessagesFinalToNormalized,
+      streamContext,
+      downstreamContext,
+    );
+
+    const serialized = lines.join('');
+    expect(serialized.indexOf('"type":"thinking"')).toBeGreaterThan(-1);
+    expect(serialized.indexOf('"type":"tool_use"')).toBeGreaterThan(serialized.indexOf('"type":"thinking"'));
+    expect(serialized.indexOf('"text":"done"')).toBeGreaterThan(serialized.indexOf('"type":"tool_use"'));
+    expect(serialized).toContain('"signature":"sig-stream-1"');
+  });
+
+  it('serializes responses web_search_call items into anthropic server tool stream blocks', () => {
+    const streamContext = anthropicMessagesStream.createContext('claude-test');
+    const downstreamContext = anthropicMessagesStream.createDownstreamContext();
+
+    const lines = serializeAnthropicUpstreamFinalAsStream(
+      {
+        id: 'resp_stream_search_1',
+        object: 'response',
+        model: 'gpt-5',
+        status: 'completed',
+        output: [
+          {
+            id: 'ws_stream_1',
+            type: 'web_search_call',
+            status: 'completed',
+            action: {
+              query: 'weather in Paris',
+            },
+          },
+          {
+            id: 'msg_stream_search_1',
+            type: 'message',
+            role: 'assistant',
+            status: 'completed',
+            content: [{ type: 'output_text', text: 'It is sunny.' }],
+          },
+        ],
+        usage: {
+          input_tokens: 8,
+          output_tokens: 3,
+        },
+      },
+      'gpt-5',
+      '',
+      normalizeAnthropicMessagesFinalToNormalized,
+      streamContext,
+      downstreamContext,
+    );
+
+    const serialized = lines.join('');
+    expect(serialized).toContain('"type":"server_tool_use"');
+    expect(serialized).toContain('"name":"web_search"');
+    expect(serialized).toContain('"type":"web_search_tool_result"');
+    expect(serialized).toContain('"tool_use_id":"srvtoolu_ws_stream_1"');
+    expect(serialized.indexOf('"type":"web_search_tool_result"')).toBeGreaterThan(serialized.indexOf('"type":"server_tool_use"'));
+    expect(serialized).toContain('"text":"It is sunny."');
+  });
+});

src/server/transformers/canonical/continuationBridge.test.ts

@@ -0,0 +1,79 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  applyOpenAiCompatibleContinuation,
+  buildOpenAiCompatibleMetadataWithContinuation,
+  normalizeCanonicalContinuation,
+  OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY,
+  readOpenAiCompatibleContinuation,
+} from './continuationBridge.js';
+
+describe('canonical continuation bridge helpers', () => {
+  it('normalizes canonical continuation fields and drops blanks', () => {
+    expect(normalizeCanonicalContinuation({
+      sessionId: ' session-1 ',
+      previousResponseId: '  ',
+      promptCacheKey: ' cache-1 ',
+      turnState: ' turn-1 ',
+    })).toEqual({
+      sessionId: 'session-1',
+      promptCacheKey: 'cache-1',
+      turnState: 'turn-1',
+    });
+  });
+
+  it('reads openai-compatible continuation fields from top-level fields and metadata namespace', () => {
+    expect(readOpenAiCompatibleContinuation({
+      session_id: 'session-top-level',
+      conversation_id: 'conversation-top-level',
+      previous_response_id: 'resp-1',
+      prompt_cache_key: 'cache-1',
+      metadata: {
+        user_id: 'session-metadata',
+        [OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY]: 'turn-state-1',
+      },
+    })).toEqual({
+      sessionId: 'session-metadata',
+      previousResponseId: 'resp-1',
+      promptCacheKey: 'cache-1',
+      turnState: 'turn-state-1',
+    });
+  });
+
+  it('materializes continuation semantics into openai-compatible metadata and body fields', () => {
+    const body: Record<string, unknown> = {
+      model: 'gpt-5',
+    };
+
+    applyOpenAiCompatibleContinuation(body, {
+      sessionId: 'session-bridge-1',
+      previousResponseId: 'resp-bridge-1',
+      promptCacheKey: 'cache-bridge-1',
+      turnState: 'turn-bridge-1',
+    }, { existing: true });
+
+    expect(body).toEqual({
+      model: 'gpt-5',
+      previous_response_id: 'resp-bridge-1',
+      prompt_cache_key: 'cache-bridge-1',
+      metadata: {
+        existing: true,
+        user_id: 'session-bridge-1',
+        [OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY]: 'turn-bridge-1',
+      },
+    });
+  });
+
+  it('does not overwrite explicit metadata when materializing continuation semantics', () => {
+    expect(buildOpenAiCompatibleMetadataWithContinuation({
+      user_id: 'user-explicit',
+      [OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY]: 'turn-explicit',
+    }, {
+      sessionId: 'session-bridge-1',
+      turnState: 'turn-bridge-1',
+    })).toEqual({
+      user_id: 'user-explicit',
+      [OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY]: 'turn-explicit',
+    });
+  });
+});

src/server/transformers/canonical/continuationBridge.ts

@@ -0,0 +1,94 @@
+import type { CanonicalContinuation } from './types.js';
+
+export const OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY = 'metapi_turn_state';
+
+function asTrimmedString(value: unknown): string {
+  return typeof value === 'string' ? value.trim() : '';
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return !!value && typeof value === 'object' && !Array.isArray(value);
+}
+
+export function normalizeCanonicalContinuation(
+  continuation: CanonicalContinuation | undefined,
+): CanonicalContinuation | undefined {
+  if (!continuation) return undefined;
+
+  const normalized: CanonicalContinuation = {
+    ...(asTrimmedString(continuation.sessionId) ? { sessionId: asTrimmedString(continuation.sessionId) } : {}),
+    ...(asTrimmedString(continuation.previousResponseId)
+      ? { previousResponseId: asTrimmedString(continuation.previousResponseId) }
+      : {}),
+    ...(asTrimmedString(continuation.promptCacheKey)
+      ? { promptCacheKey: asTrimmedString(continuation.promptCacheKey) }
+      : {}),
+    ...(asTrimmedString(continuation.turnState) ? { turnState: asTrimmedString(continuation.turnState) } : {}),
+  };
+
+  return Object.keys(normalized).length > 0 ? normalized : undefined;
+}
+
+export function readOpenAiCompatibleContinuation(
+  body: Record<string, unknown>,
+  existing?: CanonicalContinuation,
+): CanonicalContinuation | undefined {
+  const metadata = isRecord(body.metadata) ? body.metadata : null;
+  return normalizeCanonicalContinuation({
+    ...(existing ?? {}),
+    ...(asTrimmedString(body.session_id)
+      ? { sessionId: asTrimmedString(body.session_id) }
+      : {}),
+    ...(asTrimmedString(body.conversation_id)
+      ? { sessionId: asTrimmedString(body.conversation_id) }
+      : {}),
+    ...(metadata && asTrimmedString(metadata.user_id)
+      ? { sessionId: asTrimmedString(metadata.user_id) }
+      : {}),
+    ...(asTrimmedString(body.previous_response_id)
+      ? { previousResponseId: asTrimmedString(body.previous_response_id) }
+      : {}),
+    ...(asTrimmedString(body.prompt_cache_key)
+      ? { promptCacheKey: asTrimmedString(body.prompt_cache_key) }
+      : {}),
+    ...(metadata && asTrimmedString(metadata[OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY])
+      ? { turnState: asTrimmedString(metadata[OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY]) }
+      : {}),
+  });
+}
+
+export function buildOpenAiCompatibleMetadataWithContinuation(
+  metadata: Record<string, unknown>,
+  continuation: CanonicalContinuation | undefined,
+): Record<string, unknown> {
+  const normalized = normalizeCanonicalContinuation(continuation);
+  const nextMetadata: Record<string, unknown> = { ...metadata };
+
+  if (!normalized) return nextMetadata;
+  if (!('user_id' in nextMetadata) && normalized.sessionId) {
+    nextMetadata.user_id = normalized.sessionId;
+  }
+  if (!(OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY in nextMetadata) && normalized.turnState) {
+    nextMetadata[OPENAI_CONTINUATION_TURN_STATE_METADATA_KEY] = normalized.turnState;
+  }
+  return nextMetadata;
+}
+
+export function applyOpenAiCompatibleContinuation(
+  body: Record<string, unknown>,
+  continuation: CanonicalContinuation | undefined,
+  metadata: Record<string, unknown> = {},
+): void {
+  const normalized = normalizeCanonicalContinuation(continuation);
+  const nextMetadata = buildOpenAiCompatibleMetadataWithContinuation(metadata, normalized);
+
+  if (Object.keys(nextMetadata).length > 0) {
+    body.metadata = nextMetadata;
+  }
+  if (normalized?.promptCacheKey) {
+    body.prompt_cache_key = normalized.promptCacheKey;
+  }
+  if (normalized?.previousResponseId) {
+    body.previous_response_id = normalized.previousResponseId;
+  }
+}

src/server/transformers/canonical/envelope.ts

@@ -0,0 +1,77 @@
+import type { CanonicalTool, CanonicalToolChoice } from './tools.js';
+import type { CanonicalAttachment } from './attachments.js';
+import type {
+  CanonicalCliProfile,
+  CanonicalContinuation,
+  CanonicalMessage,
+  CanonicalOperation,
+  CanonicalReasoningRequest,
+  CanonicalRequestEnvelope,
+  CanonicalSurface,
+} from './types.js';
+import { normalizeCanonicalContinuation } from './continuationBridge.js';
+
+export type CreateCanonicalRequestEnvelopeInput = {
+  operation?: CanonicalOperation;
+  surface: CanonicalSurface;
+  cliProfile?: CanonicalCliProfile;
+  requestedModel: string;
+  stream?: boolean;
+  messages?: CanonicalMessage[];
+  reasoning?: CanonicalReasoningRequest;
+  tools?: CanonicalTool[];
+  toolChoice?: CanonicalToolChoice;
+  continuation?: CanonicalContinuation;
+  metadata?: Record<string, unknown>;
+  passthrough?: Record<string, unknown>;
+  attachments?: CanonicalAttachment[];
+};
+
+function asTrimmedString(value: unknown): string {
+  return typeof value === 'string' ? value.trim() : '';
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return !!value && typeof value === 'object' && !Array.isArray(value);
+}
+
+function cloneJsonValue<T>(value: T): T {
+  if (Array.isArray(value)) {
+    return value.map((item) => cloneJsonValue(item)) as T;
+  }
+  if (isRecord(value)) {
+    return Object.fromEntries(
+      Object.entries(value).map(([key, item]) => [key, cloneJsonValue(item)]),
+    ) as T;
+  }
+  return value;
+}
+
+export function createCanonicalRequestEnvelope(
+  input: CreateCanonicalRequestEnvelopeInput,
+): CanonicalRequestEnvelope {
+  const requestedModel = asTrimmedString(input.requestedModel);
+  if (!requestedModel) {
+    throw new Error('canonical request requires requestedModel');
+  }
+
+  return {
+    operation: input.operation ?? 'generate',
+    surface: input.surface,
+    cliProfile: input.cliProfile ?? 'generic',
+    requestedModel,
+    stream: input.stream === true,
+    messages: Array.isArray(input.messages) ? input.messages : [],
+    ...(input.reasoning ? { reasoning: input.reasoning } : {}),
+    ...(Array.isArray(input.tools) && input.tools.length > 0 ? { tools: input.tools } : {}),
+    ...(input.toolChoice !== undefined ? { toolChoice: input.toolChoice } : {}),
+    ...(normalizeCanonicalContinuation(input.continuation)
+      ? { continuation: normalizeCanonicalContinuation(input.continuation) }
+      : {}),
+    ...(input.metadata ? { metadata: input.metadata } : {}),
+    ...(input.passthrough ? { passthrough: input.passthrough } : {}),
+    ...(Array.isArray(input.attachments) && input.attachments.length > 0
+      ? { attachments: cloneJsonValue(input.attachments) as CanonicalAttachment[] }
+      : {}),
+  };
+}

src/server/transformers/canonical/openAiRequestBridge.test.ts

@@ -0,0 +1,61 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  canonicalRequestFromOpenAiBody,
+  canonicalRequestToOpenAiChatBody,
+} from './openAiRequestBridge.js';
+
+describe('openAiRequestBridge', () => {
+  it('parses openai-compatible continuation hints into canonical envelopes', () => {
+    const request = canonicalRequestFromOpenAiBody({
+      body: {
+        model: 'gpt-5',
+        stream: false,
+        conversation_id: 'conversation-1',
+        previous_response_id: 'resp-1',
+        prompt_cache_key: 'cache-1',
+        metadata: {
+          user_id: 'session-1',
+          metapi_turn_state: 'turn-state-1',
+        },
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+      surface: 'openai-responses',
+    });
+
+    expect(request).toMatchObject({
+      continuation: {
+        sessionId: 'session-1',
+        previousResponseId: 'resp-1',
+        promptCacheKey: 'cache-1',
+        turnState: 'turn-state-1',
+      },
+    });
+  });
+
+  it('builds openai-compatible continuation fields back from canonical envelopes', () => {
+    const body = canonicalRequestToOpenAiChatBody({
+      operation: 'generate',
+      surface: 'openai-chat',
+      cliProfile: 'codex',
+      requestedModel: 'gpt-5',
+      stream: false,
+      messages: [{ role: 'user', parts: [{ type: 'text', text: 'hello' }] }],
+      continuation: {
+        sessionId: 'session-1',
+        previousResponseId: 'resp-1',
+        promptCacheKey: 'cache-1',
+        turnState: 'turn-state-1',
+      },
+    });
+
+    expect(body).toMatchObject({
+      previous_response_id: 'resp-1',
+      prompt_cache_key: 'cache-1',
+      metadata: {
+        user_id: 'session-1',
+        metapi_turn_state: 'turn-state-1',
+      },
+    });
+  });
+});

src/server/transformers/canonical/openAiRequestBridge.ts

@@ -0,0 +1,564 @@
+import {
+  canonicalAttachmentFromInputFileBlock,
+  canonicalAttachmentToNormalizedInputFile,
+  type CanonicalAttachment,
+} from './attachments.js';
+import { createCanonicalRequestEnvelope } from './envelope.js';
+import {
+  applyOpenAiCompatibleContinuation,
+  readOpenAiCompatibleContinuation,
+} from './continuationBridge.js';
+import { normalizeCanonicalReasoningRequest } from './reasoning.js';
+import type { CanonicalTool, CanonicalToolChoice } from './tools.js';
+import type {
+  CanonicalContentPart,
+  CanonicalCliProfile,
+  CanonicalContinuation,
+  CanonicalMessage,
+  CanonicalMessageRole,
+  CanonicalOperation,
+  CanonicalRequestEnvelope,
+  CanonicalSurface,
+} from './types.js';
+import { toOpenAiChatFileBlock } from '../shared/inputFile.js';
+
+type CanonicalRequestFromOpenAiBodyInput = {
+  body: Record<string, unknown>;
+  surface: CanonicalSurface;
+  cliProfile?: CanonicalCliProfile;
+  operation?: CanonicalOperation;
+  metadata?: Record<string, unknown>;
+  passthrough?: Record<string, unknown>;
+  continuation?: CanonicalContinuation;
+};
+
+function asTrimmedString(value: unknown): string {
+  return typeof value === 'string' ? value.trim() : '';
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return !!value && typeof value === 'object' && !Array.isArray(value);
+}
+
+function cloneJsonValue<T>(value: T): T {
+  if (Array.isArray(value)) {
+    return value.map((item) => cloneJsonValue(item)) as T;
+  }
+  if (isRecord(value)) {
+    return Object.fromEntries(
+      Object.entries(value).map(([key, item]) => [key, cloneJsonValue(item)]),
+    ) as T;
+  }
+  return value;
+}
+
+function safeJsonStringify(value: unknown): string {
+  try {
+    return JSON.stringify(value);
+  } catch {
+    return '';
+  }
+}
+
+function joinNonEmpty(parts: string[]): string {
+  return parts
+    .map((part) => part.trim())
+    .filter((part) => part.length > 0)
+    .join('\n\n');
+}
+
+function normalizeRole(value: unknown): CanonicalMessageRole {
+  const role = asTrimmedString(value).toLowerCase();
+  switch (role) {
+    case 'system':
+    case 'developer':
+    case 'assistant':
+    case 'tool':
+      return role;
+    default:
+      return 'user';
+  }
+}
+
+function openAiContentToCanonicalParts(content: unknown): CanonicalContentPart[] {
+  if (typeof content === 'string') {
+    return content ? [{ type: 'text', text: content }] : [];
+  }
+
+  if (!Array.isArray(content)) return [];
+
+  const parts: CanonicalContentPart[] = [];
+  for (const item of content) {
+    if (typeof item === 'string') {
+      if (item) parts.push({ type: 'text', text: item });
+      continue;
+    }
+    if (!isRecord(item)) continue;
+
+    const type = asTrimmedString(item.type).toLowerCase();
+    if (type === 'text' || type === 'input_text' || type === 'output_text') {
+      const text = asTrimmedString(item.text);
+      if (text) parts.push({ type: 'text', text });
+      continue;
+    }
+    if (type === 'reasoning' || type === 'thinking' || type === 'redacted_reasoning') {
+      const text = asTrimmedString(item.text ?? item.reasoning ?? item.thinking);
+      if (text) parts.push({ type: 'text', text, thought: true });
+      continue;
+    }
+    if (type === 'image_url' && isRecord(item.image_url)) {
+      const url = asTrimmedString(item.image_url.url);
+      if (url) parts.push({ type: 'image', url });
+      continue;
+    }
+    if (type === 'input_image' && isRecord(item.image_url)) {
+      const url = asTrimmedString(item.image_url.url);
+      if (url) parts.push({ type: 'image', url });
+      continue;
+    }
+    if (type === 'input_file' || type === 'file') {
+      const attachment = canonicalAttachmentFromInputFileBlock(item);
+      if (attachment) {
+        parts.push({
+          type: 'file',
+          ...(attachment.fileId ? { fileId: attachment.fileId } : {}),
+          ...(attachment.fileUrl ? { fileUrl: attachment.fileUrl } : {}),
+          ...(attachment.fileData ? { fileData: attachment.fileData } : {}),
+          ...(attachment.mimeType !== undefined ? { mimeType: attachment.mimeType } : {}),
+          ...(attachment.filename ? { filename: attachment.filename } : {}),
+        });
+      }
+      continue;
+    }
+  }
+
+  return parts;
+}
+
+function appendAssistantReasoningPart(
+  parts: CanonicalContentPart[],
+  rawMessage: Record<string, unknown>,
+): void {
+  const directReasoning = joinNonEmpty([
+    asTrimmedString(rawMessage.reasoning_content),
+    asTrimmedString(rawMessage.reasoning),
+  ]);
+  if (!directReasoning) return;
+
+  const alreadyPresent = parts.some((part) => (
+    part.type === 'text'
+    && part.thought === true
+    && part.text === directReasoning
+  ));
+  if (alreadyPresent) return;
+
+  parts.unshift({
+    type: 'text',
+    text: directReasoning,
+    thought: true,
+  });
+}
+
+function parseToolChoice(rawToolChoice: unknown): CanonicalToolChoice | undefined {
+  if (typeof rawToolChoice === 'string') {
+    const normalized = rawToolChoice.trim().toLowerCase();
+    if (normalized === 'auto' || normalized === 'none' || normalized === 'required') return normalized;
+    if (normalized === 'any') return 'required';
+    return rawToolChoice.trim() ? { type: 'raw', value: rawToolChoice } : undefined;
+  }
+
+  if (!isRecord(rawToolChoice)) return undefined;
+  const type = asTrimmedString(rawToolChoice.type).toLowerCase();
+  if (type === 'auto' || type === 'none') return type;
+  if (type === 'any' || type === 'required') return 'required';
+  if (type === 'function') {
+    const name = asTrimmedString(
+      (isRecord(rawToolChoice.function) ? rawToolChoice.function.name : undefined)
+      ?? rawToolChoice.name,
+    );
+    return name ? { type: 'tool', name } : undefined;
+  }
+  if (type && type !== 'tool') {
+    return { type: 'raw', value: cloneJsonValue(rawToolChoice) as Record<string, unknown> };
+  }
+
+  const name = asTrimmedString(
+    rawToolChoice.name
+    ?? (isRecord(rawToolChoice.tool) ? rawToolChoice.tool.name : undefined),
+  );
+  const toolChoiceKeys = Object.keys(rawToolChoice);
+  const hasExtraToolFields = toolChoiceKeys.some((key) => key !== 'type' && key !== 'name' && key !== 'tool');
+  if (hasExtraToolFields) {
+    return { type: 'raw', value: cloneJsonValue(rawToolChoice) as Record<string, unknown> };
+  }
+  if (name) return { type: 'tool', name };
+  return { type: 'raw', value: cloneJsonValue(rawToolChoice) as Record<string, unknown> };
+}
+
+function parseTools(rawTools: unknown): CanonicalTool[] | undefined {
+  if (!Array.isArray(rawTools)) return undefined;
+
+  const tools: CanonicalTool[] = rawTools
+    .flatMap((item): CanonicalTool[] => {
+      if (!isRecord(item)) return [];
+      const itemType = asTrimmedString(item.type).toLowerCase();
+
+      if (itemType === 'function' && isRecord(item.function)) {
+        const name = asTrimmedString(item.function.name);
+        if (!name) return [];
+        return [{
+          name,
+          ...(asTrimmedString(item.function.description)
+            ? { description: asTrimmedString(item.function.description) }
+            : {}),
+          ...(typeof item.function.strict === 'boolean' ? { strict: item.function.strict } : {}),
+          ...(isRecord(item.function.parameters) ? { inputSchema: cloneJsonValue(item.function.parameters) } : {}),
+        }];
+      }
+
+      if ((itemType === '' || itemType === 'tool') && asTrimmedString(item.name)) {
+        return [{
+          name: asTrimmedString(item.name),
+          ...(asTrimmedString(item.description)
+            ? { description: asTrimmedString(item.description) }
+            : {}),
+          ...(typeof item.strict === 'boolean' ? { strict: item.strict } : {}),
+          ...(isRecord(item.input_schema)
+            ? { inputSchema: cloneJsonValue(item.input_schema) }
+            : (isRecord(item.inputSchema) ? { inputSchema: cloneJsonValue(item.inputSchema) } : {})),
+        }];
+      }
+
+      if (Array.isArray(item.functionDeclarations)) {
+        return item.functionDeclarations.flatMap((declaration) => {
+          if (!isRecord(declaration)) return [];
+          const name = asTrimmedString(declaration.name);
+          if (!name) return [];
+          return [{
+            name,
+            ...(asTrimmedString(declaration.description)
+              ? { description: asTrimmedString(declaration.description) }
+              : {}),
+            ...(isRecord(declaration.parametersJsonSchema)
+              ? { inputSchema: cloneJsonValue(declaration.parametersJsonSchema) }
+              : (isRecord(declaration.parameters) ? { inputSchema: cloneJsonValue(declaration.parameters) } : {})),
+          }];
+        });
+      }
+
+      if (itemType) {
+        return [{
+          type: itemType,
+          raw: cloneJsonValue(item) as Record<string, unknown>,
+        }];
+      }
+
+      return [];
+    });
+
+  return tools.length > 0 ? tools : undefined;
+}
+
+export function canonicalRequestFromOpenAiBody(
+  input: CanonicalRequestFromOpenAiBodyInput,
+): CanonicalRequestEnvelope {
+  const body = input.body;
+  const metadata = isRecord(input.metadata)
+    ? input.metadata
+    : (isRecord(body.metadata) ? cloneJsonValue(body.metadata) : undefined);
+  const attachments = Array.isArray(body.attachments)
+    ? cloneJsonValue(body.attachments) as CanonicalAttachment[]
+    : undefined;
+  const messages: CanonicalMessage[] = [];
+  const rawMessages = Array.isArray(body.messages) ? body.messages : [];
+
+  for (const rawMessage of rawMessages) {
+    if (!isRecord(rawMessage)) continue;
+    const role = normalizeRole(rawMessage.role);
+
+    if (role === 'tool') {
+      const toolCallId = asTrimmedString(rawMessage.tool_call_id ?? rawMessage.id);
+      const rawContent = rawMessage.content;
+      const resultText = typeof rawContent === 'string'
+        ? rawContent
+        : (!Array.isArray(rawContent) && !isRecord(rawContent) ? safeJsonStringify(rawContent ?? '') : '');
+      messages.push({
+        role: 'tool',
+        parts: [{
+          type: 'tool_result',
+          toolCallId: toolCallId || 'tool',
+          ...(resultText ? { resultText } : {}),
+          ...(Array.isArray(rawContent)
+            ? { resultContent: cloneJsonValue(rawContent) as Array<string | Record<string, unknown>> }
+            : (isRecord(rawContent)
+                ? { resultContent: [cloneJsonValue(rawContent) as Record<string, unknown>] }
+                : {})),
+        }],
+      });
+      continue;
+    }
+
+    const parts = openAiContentToCanonicalParts(rawMessage.content);
+    if (role === 'assistant') {
+      appendAssistantReasoningPart(parts, rawMessage);
+    }
+    const toolCalls = Array.isArray(rawMessage.tool_calls) ? rawMessage.tool_calls : [];
+    for (const toolCall of toolCalls) {
+      if (!isRecord(toolCall)) continue;
+      const fn = isRecord(toolCall.function) ? toolCall.function : {};
+      const id = asTrimmedString(toolCall.id);
+      const name = asTrimmedString(toolCall.name ?? fn.name);
+      const argumentsJson = typeof fn.arguments === 'string'
+        ? fn.arguments
+        : safeJsonStringify(fn.arguments ?? toolCall.arguments ?? {});
+      if (!name) continue;
+      parts.push({
+        type: 'tool_call',
+        id: id || `tool_${parts.length}`,
+        name,
+        argumentsJson,
+      });
+    }
+
+    messages.push({
+      role,
+      parts,
+      ...(asTrimmedString(rawMessage.phase) ? { phase: asTrimmedString(rawMessage.phase) } : {}),
+      ...(asTrimmedString(rawMessage.reasoning_signature)
+        ? { reasoningSignature: asTrimmedString(rawMessage.reasoning_signature) }
+        : {}),
+    });
+  }
+
+  const reasoningResult = normalizeCanonicalReasoningRequest({
+    include: body.include,
+    reasoning: body.reasoning,
+    reasoning_effort: body.reasoning_effort,
+    reasoning_budget: body.reasoning_budget,
+    reasoning_summary: body.reasoning_summary,
+  });
+
+  const continuation = readOpenAiCompatibleContinuation(body, input.continuation);
+
+  const passthrough = {
+    ...(input.passthrough ?? {}),
+    ...(typeof body.parallel_tool_calls === 'boolean'
+      ? { parallel_tool_calls: body.parallel_tool_calls }
+      : {}),
+    ...(reasoningResult.metadata ? { transformerMetadata: reasoningResult.metadata } : {}),
+  };
+
+  return createCanonicalRequestEnvelope({
+    operation: input.operation ?? 'generate',
+    surface: input.surface,
+    cliProfile: input.cliProfile ?? 'generic',
+    requestedModel: asTrimmedString(body.model),
+    stream: body.stream === true,
+    messages,
+    ...(reasoningResult.reasoning ? { reasoning: reasoningResult.reasoning } : {}),
+    ...(parseTools(body.tools) ? { tools: parseTools(body.tools) } : {}),
+    ...(parseToolChoice(body.tool_choice) !== undefined ? { toolChoice: parseToolChoice(body.tool_choice) } : {}),
+    ...(continuation ? { continuation } : {}),
+    ...(metadata ? { metadata } : {}),
+    ...(attachments ? { attachments } : {}),
+    ...(Object.keys(passthrough).length > 0 ? { passthrough } : {}),
+  });
+}
+
+function canonicalPartsToOpenAiContent(
+  role: CanonicalMessageRole,
+  parts: CanonicalContentPart[],
+): { content: string | Array<Record<string, unknown>>; reasoning?: string; toolCalls?: Array<Record<string, unknown>> } {
+  const contentBlocks: Array<Record<string, unknown>> = [];
+  const toolCalls: Array<Record<string, unknown>> = [];
+  const visibleText: string[] = [];
+  const reasoningText: string[] = [];
+
+  for (const part of parts) {
+    if (part.type === 'text') {
+      if (part.thought === true) {
+        reasoningText.push(part.text);
+      } else {
+        visibleText.push(part.text);
+      }
+      continue;
+    }
+    if (part.type === 'image') {
+      const url = asTrimmedString(part.url ?? part.dataUrl);
+      if (url) {
+        contentBlocks.push({
+          type: 'image_url',
+          image_url: { url },
+        });
+      }
+      continue;
+    }
+    if (part.type === 'file') {
+      const normalizedFile = canonicalAttachmentToNormalizedInputFile({
+        kind: 'file',
+        ...(part.fileId ? { fileId: part.fileId } : {}),
+        ...(part.fileUrl ? { fileUrl: part.fileUrl } : {}),
+        ...(part.fileData ? { fileData: part.fileData } : {}),
+        ...(part.filename ? { filename: part.filename } : {}),
+        ...(part.mimeType !== undefined ? { mimeType: part.mimeType } : {}),
+      });
+      contentBlocks.push(toOpenAiChatFileBlock(normalizedFile));
+      continue;
+    }
+    if (part.type === 'tool_call') {
+      toolCalls.push({
+        id: part.id,
+        type: 'function',
+        function: {
+          name: part.name,
+          arguments: part.argumentsJson,
+        },
+      });
+      continue;
+    }
+    if (part.type === 'tool_result' && role !== 'tool') {
+      const text = part.resultText
+        ?? (typeof part.resultContent === 'string'
+          ? part.resultContent
+          : safeJsonStringify(part.resultJson ?? part.resultContent ?? ''));
+      if (text) {
+        visibleText.push(text);
+      }
+    }
+  }
+
+  if (contentBlocks.length <= 0) {
+    return {
+      content: visibleText.join(''),
+      ...(reasoningText.length > 0 ? { reasoning: reasoningText.join('') } : {}),
+      ...(toolCalls.length > 0 ? { toolCalls } : {}),
+    };
+  }
+
+  if (visibleText.length > 0) {
+    contentBlocks.unshift({
+      type: 'text',
+      text: visibleText.join(''),
+    });
+  }
+
+  return {
+    content: contentBlocks,
+    ...(reasoningText.length > 0 ? { reasoning: reasoningText.join('') } : {}),
+    ...(toolCalls.length > 0 ? { toolCalls } : {}),
+  };
+}
+
+function canonicalToolChoiceToOpenAi(toolChoice: CanonicalToolChoice | undefined): unknown {
+  if (!toolChoice) return undefined;
+  if (toolChoice === 'auto' || toolChoice === 'none') return toolChoice;
+  if (toolChoice === 'required') return 'required';
+  if (toolChoice.type === 'raw') return cloneJsonValue(toolChoice.value);
+  return {
+    type: 'function',
+    function: {
+      name: toolChoice.name,
+    },
+  };
+}
+
+export function canonicalRequestToOpenAiChatBody(
+  request: CanonicalRequestEnvelope,
+): Record<string, unknown> {
+  const messages: Array<Record<string, unknown>> = [];
+
+  for (const message of request.messages) {
+    if (message.role === 'tool') {
+      for (const part of message.parts) {
+        if (part.type !== 'tool_result') continue;
+        messages.push({
+          role: 'tool',
+          tool_call_id: part.toolCallId,
+          content: part.resultContent
+            ?? part.resultText
+            ?? safeJsonStringify(part.resultJson ?? ''),
+        });
+      }
+      continue;
+    }
+
+    const converted = canonicalPartsToOpenAiContent(message.role, message.parts);
+    const nextMessage: Record<string, unknown> = {
+      role: message.role,
+      content: converted.content,
+    };
+    if (message.role === 'assistant' && converted.reasoning) {
+      nextMessage.reasoning_content = converted.reasoning;
+    }
+    if (message.phase) nextMessage.phase = message.phase;
+    if (message.reasoningSignature) nextMessage.reasoning_signature = message.reasoningSignature;
+    if (message.role === 'assistant' && converted.toolCalls && converted.toolCalls.length > 0) {
+      nextMessage.tool_calls = converted.toolCalls;
+      if (typeof nextMessage.content !== 'string' && (nextMessage.content as Array<unknown>).length <= 0) {
+        nextMessage.content = '';
+      }
+    }
+    messages.push(nextMessage);
+  }
+
+  const body: Record<string, unknown> = {
+    model: request.requestedModel,
+    stream: request.stream,
+    messages,
+  };
+
+  if (request.reasoning?.effort) body.reasoning_effort = request.reasoning.effort;
+  if (request.reasoning?.budgetTokens !== undefined) body.reasoning_budget = request.reasoning.budgetTokens;
+  if (request.reasoning?.summary) body.reasoning_summary = request.reasoning.summary;
+  const transformerMetadata = isRecord(request.passthrough?.transformerMetadata)
+    ? request.passthrough.transformerMetadata as Record<string, unknown>
+    : null;
+  const passthroughInclude = Array.isArray(transformerMetadata?.include)
+    ? (transformerMetadata.include as unknown[])
+      .filter((item): item is string => typeof item === 'string' && item.trim().length > 0)
+      .map((item) => item.trim())
+    : [];
+  const mergedInclude = [
+    ...(request.reasoning?.includeEncryptedContent ? ['reasoning.encrypted_content'] : []),
+    ...passthroughInclude,
+  ].filter((item, index, all) => all.indexOf(item) === index);
+  if (mergedInclude.length > 0) body.include = mergedInclude;
+  const metadata = isRecord(request.metadata)
+    ? cloneJsonValue(request.metadata)
+    : {};
+  applyOpenAiCompatibleContinuation(body, request.continuation, metadata);
+  if (Array.isArray(request.attachments) && request.attachments.length > 0) {
+    body.attachments = cloneJsonValue(request.attachments);
+  }
+  if (Array.isArray(request.tools) && request.tools.length > 0) {
+    body.tools = request.tools.map((tool) => {
+      if ('raw' in tool) {
+        const raw = cloneJsonValue(tool.raw) as Record<string, unknown>;
+        if (typeof raw.type !== 'string' || raw.type.trim().length === 0) {
+          raw.type = tool.type;
+        }
+        return raw;
+      }
+      return {
+        type: 'function',
+        function: {
+          name: tool.name,
+          ...(tool.description ? { description: tool.description } : {}),
+          ...(typeof tool.strict === 'boolean' ? { strict: tool.strict } : {}),
+          parameters: cloneJsonValue(tool.inputSchema ?? { type: 'object' }),
+        },
+      };
+    });
+  }
+  const toolChoice = canonicalToolChoiceToOpenAi(request.toolChoice);
+  if (toolChoice !== undefined) body.tool_choice = toolChoice;
+
+  if (isRecord(request.passthrough)) {
+    for (const [key, value] of Object.entries(request.passthrough)) {
+      if (key === 'transformerMetadata' || body[key] !== undefined) continue;
+      body[key] = cloneJsonValue(value);
+    }
+  }
+
+  return body;
+}

src/server/transformers/canonical/request.test.ts

@@ -103,6 +103,98 @@ describe('canonical request helpers', () => {
     });
   });
 
+  it('collects continuation session ids from OpenAI-compatible metadata and custom session fields', () => {
+    const request = canonicalRequestFromOpenAiBody({
+      body: {
+        model: 'gpt-5',
+        stream: false,
+        session_id: 'session-body-1',
+        conversation_id: 'conversation-body-1',
+        metadata: {
+          user_id: 'session-metadata-1',
+        },
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+      surface: 'openai-chat',
+    });
+
+    expect(request).toMatchObject({
+      continuation: {
+        sessionId: 'session-metadata-1',
+      },
+    });
+  });
+
+  it('collects continuation turnState from OpenAI-compatible metadata namespace', () => {
+    const request = canonicalRequestFromOpenAiBody({
+      body: {
+        model: 'gpt-5',
+        stream: false,
+        metadata: {
+          metapi_turn_state: 'turn-state-1',
+        },
+        messages: [{ role: 'user', content: 'hello' }],
+      },
+      surface: 'openai-chat',
+    });
+
+    expect(request).toMatchObject({
+      continuation: {
+        turnState: 'turn-state-1',
+      },
+    });
+  });
+
+  it('materializes continuation session ids into metadata.user_id without overwriting explicit metadata', () => {
+    const body = canonicalRequestToOpenAiChatBody({
+      operation: 'generate',
+      surface: 'anthropic-messages',
+      cliProfile: 'claude_code',
+      requestedModel: 'claude-sonnet-4-5',
+      stream: false,
+      messages: [{ role: 'user', parts: [{ type: 'text', text: 'hello' }] }],
+      continuation: {
+        sessionId: 'session-bridge-1',
+        promptCacheKey: 'cache-1',
+      },
+      metadata: {
+        existing: true,
+      },
+    });
+
+    expect(body).toMatchObject({
+      metadata: {
+        existing: true,
+        user_id: 'session-bridge-1',
+      },
+      prompt_cache_key: 'cache-1',
+    });
+  });
+
+  it('materializes continuation turnState into metadata without overwriting explicit metadata', () => {
+    const body = canonicalRequestToOpenAiChatBody({
+      operation: 'generate',
+      surface: 'openai-chat',
+      cliProfile: 'codex',
+      requestedModel: 'gpt-5',
+      stream: true,
+      messages: [{ role: 'user', parts: [{ type: 'text', text: 'hello' }] }],
+      continuation: {
+        turnState: 'turn-state-2',
+      },
+      metadata: {
+        existing: true,
+      },
+    });
+
+    expect(body).toMatchObject({
+      metadata: {
+        existing: true,
+        metapi_turn_state: 'turn-state-2',
+      },
+    });
+  });
+
   it('parses anthropic-shaped tools from compatibility bodies', () => {
     const request = canonicalRequestFromOpenAiBody({
       body: {

src/server/transformers/canonical/request.ts

@@ -1,602 +1,9 @@
-import {
-  canonicalAttachmentFromInputFileBlock,
-  canonicalAttachmentToNormalizedInputFile,
-  type CanonicalAttachment,
-} from './attachments.js';
-import { normalizeCanonicalReasoningRequest } from './reasoning.js';
-import type { CanonicalTool, CanonicalToolChoice } from './tools.js';
-import type {
-  CanonicalContentPart,
-  CanonicalCliProfile,
-  CanonicalContinuation,
-  CanonicalMessage,
-  CanonicalMessageRole,
-  CanonicalOperation,
-  CanonicalReasoningRequest,
-  CanonicalRequestEnvelope,
-  CanonicalSurface,
-} from './types.js';
-import { toOpenAiChatFileBlock } from '../shared/inputFile.js';
-
-export type CreateCanonicalRequestEnvelopeInput = {
-  operation?: CanonicalOperation;
-  surface: CanonicalSurface;
-  cliProfile?: CanonicalCliProfile;
-  requestedModel: string;
-  stream?: boolean;
-  messages?: CanonicalMessage[];
-  reasoning?: CanonicalReasoningRequest;
-  tools?: CanonicalTool[];
-  toolChoice?: CanonicalToolChoice;
-  continuation?: CanonicalContinuation;
-  metadata?: Record<string, unknown>;
-  passthrough?: Record<string, unknown>;
-  attachments?: CanonicalAttachment[];
-};
-
-function asTrimmedString(value: unknown): string {
-  return typeof value === 'string' ? value.trim() : '';
-}
-
-function isRecord(value: unknown): value is Record<string, unknown> {
-  return !!value && typeof value === 'object' && !Array.isArray(value);
-}
-
-function cloneJsonValue<T>(value: T): T {
-  if (Array.isArray(value)) {
-    return value.map((item) => cloneJsonValue(item)) as T;
-  }
-  if (isRecord(value)) {
-    return Object.fromEntries(
-      Object.entries(value).map(([key, item]) => [key, cloneJsonValue(item)]),
-    ) as T;
-  }
-  return value;
-}
-
-function safeJsonStringify(value: unknown): string {
-  try {
-    return JSON.stringify(value);
-  } catch {
-    return '';
-  }
-}
-
-function normalizeCanonicalContinuation(
-  continuation: CanonicalContinuation | undefined,
-): CanonicalContinuation | undefined {
-  if (!continuation) return undefined;
-
-  const normalized: CanonicalContinuation = {
-    ...(asTrimmedString(continuation.sessionId) ? { sessionId: asTrimmedString(continuation.sessionId) } : {}),
-    ...(asTrimmedString(continuation.previousResponseId)
-      ? { previousResponseId: asTrimmedString(continuation.previousResponseId) }
-      : {}),
-    ...(asTrimmedString(continuation.promptCacheKey)
-      ? { promptCacheKey: asTrimmedString(continuation.promptCacheKey) }
-      : {}),
-    ...(asTrimmedString(continuation.turnState) ? { turnState: asTrimmedString(continuation.turnState) } : {}),
-  };
-
-  return Object.keys(normalized).length > 0 ? normalized : undefined;
-}
-
-export function createCanonicalRequestEnvelope(
-  input: CreateCanonicalRequestEnvelopeInput,
-): CanonicalRequestEnvelope {
-  const requestedModel = asTrimmedString(input.requestedModel);
-  if (!requestedModel) {
-    throw new Error('canonical request requires requestedModel');
-  }
-
-  return {
-    operation: input.operation ?? 'generate',
-    surface: input.surface,
-    cliProfile: input.cliProfile ?? 'generic',
-    requestedModel,
-    stream: input.stream === true,
-    messages: Array.isArray(input.messages) ? input.messages : [],
-    ...(input.reasoning ? { reasoning: input.reasoning } : {}),
-    ...(Array.isArray(input.tools) && input.tools.length > 0 ? { tools: input.tools } : {}),
-    ...(input.toolChoice !== undefined ? { toolChoice: input.toolChoice } : {}),
-    ...(normalizeCanonicalContinuation(input.continuation)
-      ? { continuation: normalizeCanonicalContinuation(input.continuation) }
-      : {}),
-    ...(input.metadata ? { metadata: input.metadata } : {}),
-    ...(input.passthrough ? { passthrough: input.passthrough } : {}),
-    ...(Array.isArray(input.attachments) && input.attachments.length > 0
-      ? { attachments: cloneJsonValue(input.attachments) as CanonicalAttachment[] }
-      : {}),
-  };
-}
-
-type CanonicalRequestFromOpenAiBodyInput = {
-  body: Record<string, unknown>;
-  surface: CanonicalSurface;
-  cliProfile?: CanonicalCliProfile;
-  operation?: CanonicalOperation;
-  metadata?: Record<string, unknown>;
-  passthrough?: Record<string, unknown>;
-  continuation?: CanonicalContinuation;
-};
-
-function normalizeRole(value: unknown): CanonicalMessageRole {
-  const role = asTrimmedString(value).toLowerCase();
-  switch (role) {
-    case 'system':
-    case 'developer':
-    case 'assistant':
-    case 'tool':
-      return role;
-    default:
-      return 'user';
-  }
-}
-
-function openAiContentToCanonicalParts(content: unknown): CanonicalContentPart[] {
-  if (typeof content === 'string') {
-    return content ? [{ type: 'text', text: content }] : [];
-  }
-
-  if (!Array.isArray(content)) return [];
-
-  const parts: CanonicalContentPart[] = [];
-  for (const item of content) {
-    if (typeof item === 'string') {
-      if (item) parts.push({ type: 'text', text: item });
-      continue;
-    }
-    if (!isRecord(item)) continue;
-
-    const type = asTrimmedString(item.type).toLowerCase();
-    if (type === 'text' || type === 'input_text' || type === 'output_text') {
-      const text = asTrimmedString(item.text);
-      if (text) parts.push({ type: 'text', text });
-      continue;
-    }
-    if (type === 'reasoning' || type === 'thinking' || type === 'redacted_reasoning') {
-      const text = asTrimmedString(item.text ?? item.reasoning ?? item.thinking);
-      if (text) parts.push({ type: 'text', text, thought: true });
-      continue;
-    }
-    if (type === 'image_url' && isRecord(item.image_url)) {
-      const url = asTrimmedString(item.image_url.url);
-      if (url) parts.push({ type: 'image', url });
-      continue;
-    }
-    if (type === 'input_image' && isRecord(item.image_url)) {
-      const url = asTrimmedString(item.image_url.url);
-      if (url) parts.push({ type: 'image', url });
-      continue;
-    }
-    if (type === 'input_file' || type === 'file') {
-      const attachment = canonicalAttachmentFromInputFileBlock(item);
-      if (attachment) {
-        parts.push({
-          type: 'file',
-          ...(attachment.fileId ? { fileId: attachment.fileId } : {}),
-          ...(attachment.fileUrl ? { fileUrl: attachment.fileUrl } : {}),
-          ...(attachment.fileData ? { fileData: attachment.fileData } : {}),
-          ...(attachment.mimeType !== undefined ? { mimeType: attachment.mimeType } : {}),
-          ...(attachment.filename ? { filename: attachment.filename } : {}),
-        });
-      }
-      continue;
-    }
-  }
-
-  return parts;
-}
-
-function parseToolChoice(rawToolChoice: unknown): CanonicalToolChoice | undefined {
-  if (typeof rawToolChoice === 'string') {
-    const normalized = rawToolChoice.trim().toLowerCase();
-    if (normalized === 'auto' || normalized === 'none' || normalized === 'required') return normalized;
-    if (normalized === 'any') return 'required';
-    return rawToolChoice.trim() ? { type: 'raw', value: rawToolChoice } : undefined;
-  }
-
-  if (!isRecord(rawToolChoice)) return undefined;
-  const type = asTrimmedString(rawToolChoice.type).toLowerCase();
-  if (type === 'auto' || type === 'none') return type;
-  if (type === 'any' || type === 'required') return 'required';
-  if (type === 'function') {
-    const name = asTrimmedString(
-      (isRecord(rawToolChoice.function) ? rawToolChoice.function.name : undefined)
-      ?? rawToolChoice.name,
-    );
-    return name ? { type: 'tool', name } : undefined;
-  }
-  if (type && type !== 'tool') {
-    return { type: 'raw', value: cloneJsonValue(rawToolChoice) as Record<string, unknown> };
-  }
-
-  const name = asTrimmedString(
-    rawToolChoice.name
-    ?? (isRecord(rawToolChoice.tool) ? rawToolChoice.tool.name : undefined),
-  );
-  const toolChoiceKeys = Object.keys(rawToolChoice);
-  const hasExtraToolFields = toolChoiceKeys.some((key) => key !== 'type' && key !== 'name' && key !== 'tool');
-  if (hasExtraToolFields) {
-    return { type: 'raw', value: cloneJsonValue(rawToolChoice) as Record<string, unknown> };
-  }
-  if (name) return { type: 'tool', name };
-  return { type: 'raw', value: cloneJsonValue(rawToolChoice) as Record<string, unknown> };
-}
-
-function parseTools(rawTools: unknown): CanonicalTool[] | undefined {
-  if (!Array.isArray(rawTools)) return undefined;
-
-  const tools: CanonicalTool[] = rawTools
-    .flatMap((item): CanonicalTool[] => {
-      if (!isRecord(item)) return [];
-      const itemType = asTrimmedString(item.type).toLowerCase();
-
-      if (itemType === 'function' && isRecord(item.function)) {
-        const name = asTrimmedString(item.function.name);
-        if (!name) return [];
-        return [{
-          name,
-          ...(asTrimmedString(item.function.description)
-            ? { description: asTrimmedString(item.function.description) }
-            : {}),
-          ...(typeof item.function.strict === 'boolean' ? { strict: item.function.strict } : {}),
-          ...(isRecord(item.function.parameters) ? { inputSchema: cloneJsonValue(item.function.parameters) } : {}),
-        }];
-      }
-
-      if ((itemType === '' || itemType === 'tool') && asTrimmedString(item.name)) {
-        return [{
-          name: asTrimmedString(item.name),
-          ...(asTrimmedString(item.description)
-            ? { description: asTrimmedString(item.description) }
-            : {}),
-          ...(typeof item.strict === 'boolean' ? { strict: item.strict } : {}),
-          ...(isRecord(item.input_schema)
-            ? { inputSchema: cloneJsonValue(item.input_schema) }
-            : (isRecord(item.inputSchema) ? { inputSchema: cloneJsonValue(item.inputSchema) } : {})),
-        }];
-      }
-
-      if (Array.isArray(item.functionDeclarations)) {
-        return item.functionDeclarations.flatMap((declaration) => {
-          if (!isRecord(declaration)) return [];
-          const name = asTrimmedString(declaration.name);
-          if (!name) return [];
-          return [{
-            name,
-            ...(asTrimmedString(declaration.description)
-              ? { description: asTrimmedString(declaration.description) }
-              : {}),
-            ...(isRecord(declaration.parametersJsonSchema)
-              ? { inputSchema: cloneJsonValue(declaration.parametersJsonSchema) }
-              : (isRecord(declaration.parameters) ? { inputSchema: cloneJsonValue(declaration.parameters) } : {})),
-          }];
-        });
-      }
-
-      if (itemType) {
-        return [{
-          type: itemType,
-          raw: cloneJsonValue(item) as Record<string, unknown>,
-        }];
-      }
-
-      return [];
-    });
-
-  return tools.length > 0 ? tools : undefined;
-}
-
-export function canonicalRequestFromOpenAiBody(
-  input: CanonicalRequestFromOpenAiBodyInput,
-): CanonicalRequestEnvelope {
-  const body = input.body;
-  const metadata = isRecord(input.metadata)
-    ? input.metadata
-    : (isRecord(body.metadata) ? cloneJsonValue(body.metadata) : undefined);
-  const attachments = Array.isArray(body.attachments)
-    ? cloneJsonValue(body.attachments) as CanonicalAttachment[]
-    : undefined;
-  const messages: CanonicalMessage[] = [];
-  const rawMessages = Array.isArray(body.messages) ? body.messages : [];
-
-  for (const rawMessage of rawMessages) {
-    if (!isRecord(rawMessage)) continue;
-    const role = normalizeRole(rawMessage.role);
-
-    if (role === 'tool') {
-      const toolCallId = asTrimmedString(rawMessage.tool_call_id ?? rawMessage.id);
-      const rawContent = rawMessage.content;
-      const resultText = typeof rawContent === 'string'
-        ? rawContent
-        : (!Array.isArray(rawContent) && !isRecord(rawContent) ? safeJsonStringify(rawContent ?? '') : '');
-      messages.push({
-        role: 'tool',
-        parts: [{
-          type: 'tool_result',
-          toolCallId: toolCallId || 'tool',
-          ...(resultText ? { resultText } : {}),
-          ...(Array.isArray(rawContent)
-            ? { resultContent: cloneJsonValue(rawContent) as Array<string | Record<string, unknown>> }
-            : (isRecord(rawContent)
-                ? { resultContent: [cloneJsonValue(rawContent) as Record<string, unknown>] }
-                : {})),
-        }],
-      });
-      continue;
-    }
-
-    const parts = openAiContentToCanonicalParts(rawMessage.content);
-    const toolCalls = Array.isArray(rawMessage.tool_calls) ? rawMessage.tool_calls : [];
-    for (const toolCall of toolCalls) {
-      if (!isRecord(toolCall)) continue;
-      const fn = isRecord(toolCall.function) ? toolCall.function : {};
-      const id = asTrimmedString(toolCall.id);
-      const name = asTrimmedString(toolCall.name ?? fn.name);
-      const argumentsJson = typeof fn.arguments === 'string'
-        ? fn.arguments
-        : safeJsonStringify(fn.arguments ?? toolCall.arguments ?? {});
-      if (!name) continue;
-      parts.push({
-        type: 'tool_call',
-        id: id || `tool_${parts.length}`,
-        name,
-        argumentsJson,
-      });
-    }
-
-    messages.push({
-      role,
-      parts,
-      ...(asTrimmedString(rawMessage.phase) ? { phase: asTrimmedString(rawMessage.phase) } : {}),
-      ...(asTrimmedString(rawMessage.reasoning_signature)
-        ? { reasoningSignature: asTrimmedString(rawMessage.reasoning_signature) }
-        : {}),
-    });
-  }
-
-  const reasoningResult = normalizeCanonicalReasoningRequest({
-    include: body.include,
-    reasoning: body.reasoning,
-    reasoning_effort: body.reasoning_effort,
-    reasoning_budget: body.reasoning_budget,
-    reasoning_summary: body.reasoning_summary,
-  });
-
-  const continuation: CanonicalContinuation = {
-    ...(normalizeCanonicalContinuation(input.continuation) ?? {}),
-    ...(asTrimmedString(body.previous_response_id)
-      ? { previousResponseId: asTrimmedString(body.previous_response_id) }
-      : {}),
-    ...(asTrimmedString(body.prompt_cache_key)
-      ? { promptCacheKey: asTrimmedString(body.prompt_cache_key) }
-      : {}),
-  };
-
-  const passthrough = {
-    ...(input.passthrough ?? {}),
-    ...(typeof body.parallel_tool_calls === 'boolean'
-      ? { parallel_tool_calls: body.parallel_tool_calls }
-      : {}),
-    ...(reasoningResult.metadata ? { transformerMetadata: reasoningResult.metadata } : {}),
-  };
-
-  return createCanonicalRequestEnvelope({
-    operation: input.operation ?? 'generate',
-    surface: input.surface,
-    cliProfile: input.cliProfile ?? 'generic',
-    requestedModel: asTrimmedString(body.model),
-    stream: body.stream === true,
-    messages,
-    ...(reasoningResult.reasoning ? { reasoning: reasoningResult.reasoning } : {}),
-    ...(parseTools(body.tools) ? { tools: parseTools(body.tools) } : {}),
-    ...(parseToolChoice(body.tool_choice) !== undefined ? { toolChoice: parseToolChoice(body.tool_choice) } : {}),
-    ...(Object.keys(continuation).length > 0 ? { continuation } : {}),
-    ...(metadata ? { metadata } : {}),
-    ...(attachments ? { attachments } : {}),
-    ...(Object.keys(passthrough).length > 0 ? { passthrough } : {}),
-  });
-}
-
-function canonicalPartsToOpenAiContent(
-  role: CanonicalMessageRole,
-  parts: CanonicalContentPart[],
-): { content: string | Array<Record<string, unknown>>; reasoning?: string; toolCalls?: Array<Record<string, unknown>> } {
-  const contentBlocks: Array<Record<string, unknown>> = [];
-  const toolCalls: Array<Record<string, unknown>> = [];
-  const visibleText: string[] = [];
-  const reasoningText: string[] = [];
-
-  for (const part of parts) {
-    if (part.type === 'text') {
-      if (part.thought === true) {
-        reasoningText.push(part.text);
-      } else {
-        visibleText.push(part.text);
-      }
-      continue;
-    }
-    if (part.type === 'image') {
-      const url = asTrimmedString(part.url ?? part.dataUrl);
-      if (url) {
-        contentBlocks.push({
-          type: 'image_url',
-          image_url: { url },
-        });
-      }
-      continue;
-    }
-    if (part.type === 'file') {
-      const normalizedFile = canonicalAttachmentToNormalizedInputFile({
-        kind: 'file',
-        ...(part.fileId ? { fileId: part.fileId } : {}),
-        ...(part.fileUrl ? { fileUrl: part.fileUrl } : {}),
-        ...(part.fileData ? { fileData: part.fileData } : {}),
-        ...(part.filename ? { filename: part.filename } : {}),
-        ...(part.mimeType !== undefined ? { mimeType: part.mimeType } : {}),
-      });
-      contentBlocks.push(toOpenAiChatFileBlock(normalizedFile));
-      continue;
-    }
-    if (part.type === 'tool_call') {
-      toolCalls.push({
-        id: part.id,
-        type: 'function',
-        function: {
-          name: part.name,
-          arguments: part.argumentsJson,
-        },
-      });
-      continue;
-    }
-    if (part.type === 'tool_result' && role !== 'tool') {
-      const text = part.resultText
-        ?? (typeof part.resultContent === 'string'
-          ? part.resultContent
-          : safeJsonStringify(part.resultJson ?? part.resultContent ?? ''));
-      if (text) {
-        visibleText.push(text);
-      }
-    }
-  }
-
-  if (contentBlocks.length <= 0) {
-    return {
-      content: visibleText.join(''),
-      ...(reasoningText.length > 0 ? { reasoning: reasoningText.join('') } : {}),
-      ...(toolCalls.length > 0 ? { toolCalls } : {}),
-    };
-  }
-
-  if (visibleText.length > 0) {
-    contentBlocks.unshift({
-      type: 'text',
-      text: visibleText.join(''),
-    });
-  }
-
-  return {
-    content: contentBlocks,
-    ...(reasoningText.length > 0 ? { reasoning: reasoningText.join('') } : {}),
-    ...(toolCalls.length > 0 ? { toolCalls } : {}),
-  };
-}
-
-function canonicalToolChoiceToOpenAi(toolChoice: CanonicalToolChoice | undefined): unknown {
-  if (!toolChoice) return undefined;
-  if (toolChoice === 'auto' || toolChoice === 'none') return toolChoice;
-  if (toolChoice === 'required') return 'required';
-  if (toolChoice.type === 'raw') return cloneJsonValue(toolChoice.value);
-  return {
-    type: 'function',
-    function: {
-      name: toolChoice.name,
-    },
-  };
-}
-
-export function canonicalRequestToOpenAiChatBody(
-  request: CanonicalRequestEnvelope,
-): Record<string, unknown> {
-  const messages: Array<Record<string, unknown>> = [];
-
-  for (const message of request.messages) {
-    if (message.role === 'tool') {
-      for (const part of message.parts) {
-        if (part.type !== 'tool_result') continue;
-        messages.push({
-          role: 'tool',
-          tool_call_id: part.toolCallId,
-          content: part.resultContent
-            ?? part.resultText
-            ?? safeJsonStringify(part.resultJson ?? ''),
-        });
-      }
-      continue;
-    }
-
-    const converted = canonicalPartsToOpenAiContent(message.role, message.parts);
-    const nextMessage: Record<string, unknown> = {
-      role: message.role,
-      content: converted.content,
-    };
-    if (message.role === 'assistant' && converted.reasoning) {
-      nextMessage.reasoning_content = converted.reasoning;
-    }
-    if (message.phase) nextMessage.phase = message.phase;
-    if (message.reasoningSignature) nextMessage.reasoning_signature = message.reasoningSignature;
-    if (message.role === 'assistant' && converted.toolCalls && converted.toolCalls.length > 0) {
-      nextMessage.tool_calls = converted.toolCalls;
-      if (typeof nextMessage.content !== 'string' && (nextMessage.content as Array<unknown>).length <= 0) {
-        nextMessage.content = '';
-      }
-    }
-    messages.push(nextMessage);
-  }
-
-  const body: Record<string, unknown> = {
-    model: request.requestedModel,
-    stream: request.stream,
-    messages,
-  };
-
-  if (request.reasoning?.effort) body.reasoning_effort = request.reasoning.effort;
-  if (request.reasoning?.budgetTokens !== undefined) body.reasoning_budget = request.reasoning.budgetTokens;
-  if (request.reasoning?.summary) body.reasoning_summary = request.reasoning.summary;
-  const transformerMetadata = isRecord(request.passthrough?.transformerMetadata)
-    ? request.passthrough.transformerMetadata as Record<string, unknown>
-    : null;
-  const passthroughInclude = Array.isArray(transformerMetadata?.include)
-    ? (transformerMetadata.include as unknown[])
-      .filter((item): item is string => typeof item === 'string' && item.trim().length > 0)
-      .map((item) => item.trim())
-    : [];
-  const mergedInclude = [
-    ...(request.reasoning?.includeEncryptedContent ? ['reasoning.encrypted_content'] : []),
-    ...passthroughInclude,
-  ].filter((item, index, all) => all.indexOf(item) === index);
-  if (mergedInclude.length > 0) body.include = mergedInclude;
-  if (request.metadata !== undefined) body.metadata = cloneJsonValue(request.metadata);
-  if (Array.isArray(request.attachments) && request.attachments.length > 0) {
-    body.attachments = cloneJsonValue(request.attachments);
-  }
-  if (request.continuation?.promptCacheKey) body.prompt_cache_key = request.continuation.promptCacheKey;
-  if (request.continuation?.previousResponseId) body.previous_response_id = request.continuation.previousResponseId;
-  if (Array.isArray(request.tools) && request.tools.length > 0) {
-    body.tools = request.tools.map((tool) => {
-      if ('raw' in tool) {
-        const raw = cloneJsonValue(tool.raw) as Record<string, unknown>;
-        if (typeof raw.type !== 'string' || raw.type.trim().length === 0) {
-          raw.type = tool.type;
-        }
-        return raw;
-      }
-      return {
-        type: 'function',
-        function: {
-          name: tool.name,
-          ...(tool.description ? { description: tool.description } : {}),
-          ...(typeof tool.strict === 'boolean' ? { strict: tool.strict } : {}),
-          parameters: cloneJsonValue(tool.inputSchema ?? { type: 'object' }),
-        },
-      };
-    });
-  }
-  const toolChoice = canonicalToolChoiceToOpenAi(request.toolChoice);
-  if (toolChoice !== undefined) body.tool_choice = toolChoice;
-
-  if (isRecord(request.passthrough)) {
-    for (const [key, value] of Object.entries(request.passthrough)) {
-      if (key === 'transformerMetadata' || body[key] !== undefined) continue;
-      body[key] = cloneJsonValue(value);
-    }
-  }
-
-  return body;
-}
-
+export { createCanonicalRequestEnvelope } from './envelope.js';
+export { canonicalRequestFromOpenAiBody, canonicalRequestToOpenAiChatBody } from './openAiRequestBridge.js';
 export * from './attachments.js';
+export * from './continuationBridge.js';
+export * from './envelope.js';
+export * from './openAiRequestBridge.js';
 export * from './reasoning.js';
 export * from './tools.js';
 export * from './types.js';

src/server/transformers/gemini/generate-content/cliBridge.test.ts

@@ -0,0 +1,58 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  createGeminiCliStreamReader,
+  unwrapGeminiCliPayload,
+  wrapGeminiCliRequest,
+} from './cliBridge.js';
+
+describe('gemini cli bridge', () => {
+  it('wraps Gemini CLI requests with project and stripped request model', () => {
+    expect(wrapGeminiCliRequest({
+      modelName: 'models/gemini-2.5-pro',
+      projectId: 'project-123',
+      request: {
+        model: 'ignored-model',
+        contents: [{ role: 'user', parts: [{ text: 'hello' }] }],
+      },
+    })).toEqual({
+      project: 'project-123',
+      model: 'models/gemini-2.5-pro',
+      request: {
+        contents: [{ role: 'user', parts: [{ text: 'hello' }] }],
+      },
+    });
+  });
+
+  it('unwraps Gemini CLI payload envelopes to the response body', () => {
+    expect(unwrapGeminiCliPayload({ response: { ok: true }, project: 'p' })).toEqual({ ok: true });
+    expect(unwrapGeminiCliPayload({ ok: true })).toEqual({ ok: true });
+    expect(unwrapGeminiCliPayload('raw')).toBe('raw');
+  });
+
+  it('rewrites SSE data blocks to the inner Gemini CLI response payload', async () => {
+    const chunks = [
+      Buffer.from('data: {"response":{"candidates":[{"content":{"parts":[{"text":"hello"}]}}]}}\n\n'),
+      Buffer.from('data: [DONE]\n\n'),
+    ];
+    let index = 0;
+    const reader = createGeminiCliStreamReader({
+      async read() {
+        if (index >= chunks.length) return { done: true, value: undefined };
+        return { done: false, value: new Uint8Array(chunks[index++]) };
+      },
+      async cancel() {
+        return undefined;
+      },
+      releaseLock() {},
+    });
+
+    const first = await reader.read();
+    const second = await reader.read();
+    const third = await reader.read();
+
+    expect(Buffer.from(first.value ?? []).toString('utf8')).toContain('data: {"candidates":[{"content":{"parts":[{"text":"hello"}]}}]}');
+    expect(Buffer.from(second.value ?? []).toString('utf8')).toBe('data: [DONE]\n\n');
+    expect(third.done).toBe(true);
+  });
+});

src/server/transformers/gemini/generate-content/cliBridge.ts

@@ -0,0 +1,94 @@
+import { TextDecoder, TextEncoder } from 'node:util';
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return !!value && typeof value === 'object' && !Array.isArray(value);
+}
+
+export function wrapGeminiCliRequest(input: {
+  modelName: string;
+  projectId: string;
+  request: Record<string, unknown>;
+}) {
+  const { model: _model, ...requestPayload } = input.request;
+  return {
+    project: input.projectId,
+    model: input.modelName,
+    request: requestPayload,
+  };
+}
+
+export function unwrapGeminiCliPayload<T>(payload: T): unknown {
+  if (!isRecord(payload)) return payload;
+  if (payload.response !== undefined) {
+    return payload.response;
+  }
+  return payload;
+}
+
+function rewriteGeminiCliSseEventBlock(block: string): string {
+  const lines = block.split(/\r?\n/g);
+  return lines.map((line) => {
+    if (!line.startsWith('data:')) return line;
+    const data = line.slice(5).trim();
+    if (!data || data === '[DONE]') return line;
+    try {
+      const parsed = JSON.parse(data);
+      return `data: ${JSON.stringify(unwrapGeminiCliPayload(parsed))}`;
+    } catch {
+      return line;
+    }
+  }).join('\n');
+}
+
+export function createGeminiCliStreamReader(reader: {
+  read(): Promise<{ done: boolean; value?: Uint8Array }>;
+  cancel(reason?: unknown): Promise<unknown>;
+  releaseLock(): void;
+}) {
+  const decoder = new TextDecoder();
+  const encoder = new TextEncoder();
+  const outputQueue: Uint8Array[] = [];
+  let buffer = '';
+  let done = false;
+
+  async function fillQueue() {
+    while (outputQueue.length <= 0 && !done) {
+      const result = await reader.read();
+      if (result.done) {
+        done = true;
+        const tail = decoder.decode();
+        if (tail) buffer += tail;
+        if (buffer.trim()) {
+          outputQueue.push(encoder.encode(`${rewriteGeminiCliSseEventBlock(buffer)}\n\n`));
+          buffer = '';
+        }
+        break;
+      }
+      if (!result.value) continue;
+      buffer += decoder.decode(result.value, { stream: true });
+      let separatorIndex = buffer.indexOf('\n\n');
+      while (separatorIndex >= 0) {
+        const block = buffer.slice(0, separatorIndex);
+        buffer = buffer.slice(separatorIndex + 2);
+        outputQueue.push(encoder.encode(`${rewriteGeminiCliSseEventBlock(block)}\n\n`));
+        separatorIndex = buffer.indexOf('\n\n');
+      }
+    }
+  }
+
+  return {
+    async read() {
+      await fillQueue();
+      if (outputQueue.length > 0) {
+        return { done: false, value: outputQueue.shift() };
+      }
+      return { done: true, value: undefined };
+    },
+    cancel(reason?: unknown) {
+      return reader.cancel(reason);
+    },
+    releaseLock() {
+      reader.releaseLock();
+    },
+  };
+}

src/server/transformers/gemini/generate-content/index.ts

@@ -1,9 +1,4 @@
-import {
-  canonicalRequestFromOpenAiBody,
-  isCanonicalFunctionTool,
-  isCanonicalNamedToolChoice,
-} from '../../canonical/request.js';
-import type { CanonicalContentPart, CanonicalRequestEnvelope } from '../../canonical/types.js';
+import type { CanonicalRequestEnvelope } from '../../canonical/types.js';
 import type { ProtocolBuildContext, ProtocolParseContext } from '../../contracts.js';
 import {
   resolveGeminiGenerateContentUrl,
@@ -20,12 +15,8 @@ function asTrimmedString(value: unknown): string {
   return typeof value === 'string' ? value.trim() : '';
 }
 
-function isRecord(value: unknown): value is Record<string, unknown> {
-  return !!value && typeof value === 'object' && !Array.isArray(value);
-}
-
 export function resolveGeminiProxyApiVersion(params: { geminiApiVersion?: unknown } | null | undefined): string {
-  return asTrimmedString(params?.geminiApiVersion) || 'v1beta';
+  return (typeof params?.geminiApiVersion === 'string' ? params.geminiApiVersion.trim() : '') || 'v1beta';
 }
 
 export function parseGeminiProxyRequestPath(input: {
@@ -61,203 +52,21 @@ export function parseGeminiProxyRequestPath(input: {
 }
 
 import { geminiGenerateContentInbound } from './inbound.js';
-import { geminiGenerateContentOutbound } from './outbound.js';
-import { geminiGenerateContentStream } from './stream.js';
+import { geminiGenerateContentResponseBridge } from './responseBridge.js';
+import { geminiGenerateContentStream } from './streamBridge.js';
 import { createGeminiGenerateContentAggregateState, applyGeminiGenerateContentAggregate } from './aggregator.js';
 import { geminiGenerateContentUsage } from './usage.js';
 import { reasoningEffortToGeminiThinkingConfig, geminiThinkingConfigToReasoning } from './convert.js';
 import { buildOpenAiBodyFromGeminiRequest, serializeNormalizedFinalToGemini } from './compatibility.js';
-
-function parseJsonString(raw: string): unknown {
-  const trimmed = raw.trim();
-  if (!trimmed) return {};
-  try {
-    return JSON.parse(trimmed);
-  } catch {
-    return { raw };
-  }
-}
-
-function parseDataUrl(value: string): { mimeType: string; data: string } | null {
-  const match = /^data:([^;,]+);base64,(.+)$/i.exec(value.trim());
-  if (!match) return null;
-  return {
-    mimeType: match[1],
-    data: match[2],
-  };
-}
-
-function canonicalPartToGeminiPart(part: CanonicalContentPart): Record<string, unknown> | null {
-  if (part.type === 'text') {
-    return {
-      text: part.text,
-      ...(part.thought === true ? { thought: true } : {}),
-    };
-  }
-
-  if (part.type === 'image') {
-    const source = typeof part.dataUrl === 'string' && part.dataUrl.trim()
-      ? part.dataUrl
-      : (typeof part.url === 'string' ? part.url : '');
-    if (!source) return null;
-
-    const dataUrl = parseDataUrl(source);
-    if (dataUrl) {
-      return {
-        inlineData: {
-          mimeType: dataUrl.mimeType,
-          data: dataUrl.data,
-        },
-      };
-    }
-
-    return {
-      fileData: {
-        fileUri: source,
-        ...(part.mimeType ? { mimeType: part.mimeType } : {}),
-      },
-    };
-  }
-
-  if (part.type === 'file') {
-    if (part.fileData) {
-      return {
-        inlineData: {
-          mimeType: part.mimeType || 'application/octet-stream',
-          data: part.fileData,
-        },
-      };
-    }
-
-    const fileUri = part.fileUrl || part.fileId;
-    if (!fileUri) return null;
-    return {
-      fileData: {
-        fileUri,
-        ...(part.mimeType ? { mimeType: part.mimeType } : {}),
-      },
-    };
-  }
-
-  if (part.type === 'tool_call') {
-    return {
-      functionCall: {
-        id: part.id,
-        name: part.name,
-        args: parseJsonString(part.argumentsJson),
-      },
-    };
-  }
-
-  if (part.type === 'tool_result') {
-    return {
-      functionResponse: {
-        name: part.toolCallId,
-        response: part.resultJson ?? parseJsonString(part.resultText ?? ''),
-      },
-    };
-  }
-
-  return null;
-}
-
-function buildGeminiRequestFromCanonical(request: CanonicalRequestEnvelope): Record<string, unknown> {
-  const contents: Array<Record<string, unknown>> = [];
-  const systemParts: Array<Record<string, unknown>> = [];
-
-  for (const message of request.messages) {
-    if (message.role === 'system' || message.role === 'developer') {
-      systemParts.push(
-        ...message.parts
-          .map((part) => canonicalPartToGeminiPart(part))
-          .filter((part): part is Record<string, unknown> => !!part),
-      );
-      continue;
-    }
-
-    const parts = message.parts
-      .map((part) => canonicalPartToGeminiPart(part))
-      .filter((part): part is Record<string, unknown> => !!part);
-
-    if (parts.length <= 0) continue;
-
-    if (message.role === 'tool') {
-      contents.push({
-        role: 'user',
-        parts,
-      });
-      continue;
-    }
-
-    contents.push({
-      role: message.role === 'assistant' ? 'model' : 'user',
-      parts,
-    });
-  }
-
-  const payload: Record<string, unknown> = {
-    contents,
-  };
-
-  if (systemParts.length > 0) {
-    payload.systemInstruction = {
-      role: 'user',
-      parts: systemParts,
-    };
-  }
-
-  const generationConfig: Record<string, unknown> = {};
-  if (request.reasoning?.budgetTokens !== undefined) {
-    generationConfig.thinkingConfig = {
-      thinkingBudget: request.reasoning.budgetTokens,
-    };
-  } else if (request.reasoning?.effort) {
-    generationConfig.thinkingConfig = reasoningEffortToGeminiThinkingConfig(
-      request.requestedModel,
-      request.reasoning.effort,
-    );
-  }
-  if (Object.keys(generationConfig).length > 0) {
-    payload.generationConfig = generationConfig;
-  }
-
-  if (Array.isArray(request.tools) && request.tools.length > 0) {
-    const functionTools = request.tools.filter(isCanonicalFunctionTool);
-    if (functionTools.length > 0) {
-      payload.tools = [{
-        functionDeclarations: functionTools.map((tool) => ({
-          name: tool.name,
-          ...(tool.description ? { description: tool.description } : {}),
-          ...(tool.inputSchema ? { parameters: tool.inputSchema } : {}),
-        })),
-      }];
-    }
-  }
-
-  if (request.toolChoice) {
-    if (request.toolChoice === 'none') {
-      payload.toolConfig = { functionCallingConfig: { mode: 'NONE' } };
-    } else if (request.toolChoice === 'auto') {
-      payload.toolConfig = { functionCallingConfig: { mode: 'AUTO' } };
-    } else if (request.toolChoice === 'required') {
-      payload.toolConfig = { functionCallingConfig: { mode: 'ANY' } };
-    } else if (isCanonicalNamedToolChoice(request.toolChoice)) {
-      payload.toolConfig = {
-        functionCallingConfig: {
-          mode: 'ANY',
-          allowedFunctionNames: [request.toolChoice.name],
-        },
-      };
-    }
-  }
-
-  return payload;
-}
+import {
+  buildCanonicalRequestToGeminiGenerateContentBody,
+  parseGeminiGenerateContentRequestToCanonical,
+} from './requestBridge.js';
 
 export const geminiGenerateContentTransformer = {
   protocol: 'gemini/generate-content' as const,
   inbound: geminiGenerateContentInbound,
-  outbound: geminiGenerateContentOutbound,
+  outbound: geminiGenerateContentResponseBridge,
   stream: geminiGenerateContentStream,
   aggregator: {
     createState: createGeminiGenerateContentAggregateState,
@@ -281,53 +90,19 @@ export const geminiGenerateContentTransformer = {
     body: unknown,
     ctx?: ProtocolParseContext,
   ): { value?: CanonicalRequestEnvelope; error?: { statusCode: number; payload: unknown } } {
-    const rawBody = isRecord(body) ? body : {};
-    const requestedModel = asTrimmedString(rawBody.model ?? ctx?.metadata?.requestedModel);
-    if (!requestedModel) {
-      return {
-        error: {
-          statusCode: 400,
-          payload: {
-            error: {
-              message: 'model is required',
-              type: 'invalid_request_error',
-            },
-          },
-        },
-      };
-    }
-
-    const stream = rawBody.stream === true || ctx?.metadata?.stream === true;
-    const normalizedBody = geminiGenerateContentInbound.normalizeRequest(rawBody, requestedModel);
-    const openAiBody = buildOpenAiBodyFromGeminiRequest({
-      body: normalizedBody,
-      modelName: requestedModel,
-      stream,
-    });
-
-    return {
-      value: canonicalRequestFromOpenAiBody({
-        body: openAiBody,
-        surface: 'gemini-generate-content',
-        cliProfile: ctx?.cliProfile,
-        operation: ctx?.operation,
-        metadata: ctx?.metadata,
-        passthrough: ctx?.passthrough,
-        continuation: ctx?.continuation,
-      }),
-    };
+    return parseGeminiGenerateContentRequestToCanonical(body, ctx);
   },
   buildProtocolRequest(
     request: CanonicalRequestEnvelope,
     _ctx?: ProtocolBuildContext,
   ): Record<string, unknown> {
-    return buildGeminiRequestFromCanonical(request);
+    return buildCanonicalRequestToGeminiGenerateContentBody(request);
   },
 };
 
 export {
   geminiGenerateContentInbound,
-  geminiGenerateContentOutbound,
+  geminiGenerateContentResponseBridge as geminiGenerateContentOutbound,
   geminiGenerateContentStream,
   createGeminiGenerateContentAggregateState,
   applyGeminiGenerateContentAggregate,

src/server/transformers/gemini/generate-content/outbound.ts

@@ -1,279 +1,15 @@
 import {
-  applyGeminiGenerateContentAggregate,
-  createGeminiGenerateContentAggregateState,
-  type GeminiGenerateContentAggregateState,
-} from './aggregator.js';
-import {
-  toTransformerMetadataRecord,
-  type TransformerMetadata,
-} from '../../shared/normalized.js';
-import {
-  resolveGeminiGenerateContentUrl,
-  resolveGeminiModelsUrl,
-  resolveGeminiNativeBaseUrl,
-} from './urlResolver.js';
+  extractGeminiGenerateContentResponseMetadata,
+  extractGeminiGenerateContentTransformerMetadata,
+  geminiGenerateContentResponseBridge,
+  serializeGeminiGenerateContentAggregateResponse,
+} from './responseBridge.js';
 
-type GeminiRecord = Record<string, unknown>;
+export const geminiGenerateContentOutbound = geminiGenerateContentResponseBridge;
 
-function isRecord(value: unknown): value is GeminiRecord {
-  return !!value && typeof value === 'object' && !Array.isArray(value);
-}
-
-function cloneJsonValue<T>(value: T): T {
-  if (Array.isArray(value)) {
-    return value.map((item) => cloneJsonValue(item)) as T;
-  }
-  if (isRecord(value)) {
-    return Object.fromEntries(
-      Object.entries(value).map(([key, item]) => [key, cloneJsonValue(item)]),
-    ) as T;
-  }
-  return value;
-}
-
-function isAggregateState(value: unknown): value is GeminiGenerateContentAggregateState {
-  return isRecord(value) && Array.isArray(value.parts) && Array.isArray(value.groundingMetadata) && Array.isArray(value.citations);
-}
-
-function ensureAggregateState(payload: unknown): GeminiGenerateContentAggregateState {
-  if (isAggregateState(payload)) return payload;
-
-  const state = createGeminiGenerateContentAggregateState();
-  const chunks = Array.isArray(payload) ? payload : [payload];
-  for (const chunk of chunks) {
-    applyGeminiGenerateContentAggregate(state, chunk);
-  }
-  return state;
-}
-
-function buildUsageMetadata(state: GeminiGenerateContentAggregateState): GeminiRecord | undefined {
-  const usage = state.usage;
-  const next: GeminiRecord = {};
-  if (typeof usage.promptTokenCount === 'number') next.promptTokenCount = usage.promptTokenCount;
-  if (typeof usage.candidatesTokenCount === 'number') next.candidatesTokenCount = usage.candidatesTokenCount;
-  if (typeof usage.totalTokenCount === 'number') next.totalTokenCount = usage.totalTokenCount;
-  if (typeof usage.cachedContentTokenCount === 'number') next.cachedContentTokenCount = usage.cachedContentTokenCount;
-  if (typeof usage.thoughtsTokenCount === 'number') next.thoughtsTokenCount = usage.thoughtsTokenCount;
-  return Object.keys(next).length > 0 ? next : undefined;
-}
-
-function buildCandidates(state: GeminiGenerateContentAggregateState): GeminiRecord[] {
-  if (state.candidates.length > 0) {
-    return state.candidates
-      .slice()
-      .sort((left, right) => left.index - right.index)
-      .map((candidate) => {
-        const next: GeminiRecord = {
-          index: candidate.index,
-          finishReason: candidate.finishReason || 'STOP',
-          content: {
-            role: 'model',
-            parts: candidate.parts,
-          },
-        };
-        if (candidate.groundingMetadata) next.groundingMetadata = candidate.groundingMetadata;
-        if (candidate.citationMetadata) next.citationMetadata = candidate.citationMetadata;
-        return next;
-      });
-  }
-
-  const fallback: GeminiRecord = {
-    index: 0,
-    finishReason: state.finishReason || 'STOP',
-    content: {
-      role: 'model',
-      parts: state.parts,
-    },
-  };
-  if (state.groundingMetadata.length > 0) {
-    fallback.groundingMetadata = state.groundingMetadata[0];
-  }
-  if (state.citations.length > 0) {
-    fallback.citationMetadata = state.citations[0];
-  }
-  return [fallback];
-}
-
-function extractOrderedCandidateMetadata(
-  state: GeminiGenerateContentAggregateState,
-  key: 'groundingMetadata' | 'citationMetadata',
-): GeminiRecord[] {
-  if (state.candidates.length > 0) {
-    return state.candidates
-      .slice()
-      .sort((left, right) => left.index - right.index)
-      .map((candidate) => candidate[key])
-      .filter((item): item is GeminiRecord => isRecord(item))
-      .map((item) => cloneJsonValue(item));
-  }
-
-  const fallback = key === 'groundingMetadata' ? state.groundingMetadata : state.citations;
-  return fallback.map((item) => cloneJsonValue(item));
-}
-
-function extractOrderedThoughtSignatures(
-  state: GeminiGenerateContentAggregateState,
-): {
-  all: string[];
-  preferred: string | undefined;
-} {
-  const ordered = new Set<string>();
-  const preferredThoughts = new Set<string>();
-
-  const candidates = state.candidates.length > 0
-    ? state.candidates.slice().sort((left, right) => left.index - right.index)
-    : [{ index: 0, finishReason: state.finishReason, parts: state.parts }];
-
-  for (const candidate of candidates) {
-    for (const part of candidate.parts) {
-      if (!isRecord(part) || typeof part.thoughtSignature !== 'string' || !part.thoughtSignature.trim()) {
-        continue;
-      }
-      ordered.add(part.thoughtSignature);
-      if (part.thought === true) {
-        preferredThoughts.add(part.thoughtSignature);
-      }
-    }
-  }
-
-  const orderedList = [...ordered];
-  const preferred = [...preferredThoughts][0]
-    ?? (orderedList.length === 1 ? orderedList[0] : undefined);
-
-  return {
-    all: orderedList.length > 0 ? orderedList : [...state.thoughtSignatures],
-    preferred,
-  };
-}
-
-function ensurePassthrough(metadata: TransformerMetadata): Record<string, unknown> {
-  if (!metadata.passthrough) metadata.passthrough = {};
-  return metadata.passthrough;
-}
-
-function extractRequestSemantics(requestPayload: unknown): TransformerMetadata {
-  if (!isRecord(requestPayload)) return {};
-
-  const metadata: TransformerMetadata = {};
-  const passthrough = ensurePassthrough(metadata);
-  if (requestPayload.systemInstruction !== undefined) {
-    passthrough.systemInstruction = cloneJsonValue(requestPayload.systemInstruction);
-  }
-  if (requestPayload.cachedContent !== undefined) {
-    passthrough.cachedContent = cloneJsonValue(requestPayload.cachedContent);
-  }
-  if (requestPayload.safetySettings !== undefined) {
-    metadata.geminiSafetySettings = cloneJsonValue(requestPayload.safetySettings);
-  }
-  if (requestPayload.toolConfig !== undefined) {
-    passthrough.toolConfig = cloneJsonValue(requestPayload.toolConfig);
-  }
-
-  const generationConfig = isRecord(requestPayload.generationConfig) ? requestPayload.generationConfig : null;
-  if (generationConfig) {
-    const preservedKeys = [
-      'stopSequences',
-      'responseModalities',
-      'responseMimeType',
-      'responseSchema',
-      'candidateCount',
-      'maxOutputTokens',
-      'temperature',
-      'topP',
-      'topK',
-      'presencePenalty',
-      'frequencyPenalty',
-      'seed',
-      'responseLogprobs',
-      'logprobs',
-      'thinkingConfig',
-      'imageConfig',
-    ];
-    for (const key of preservedKeys) {
-      if (generationConfig[key] !== undefined) {
-        if (key === 'imageConfig') {
-          metadata.geminiImageConfig = cloneJsonValue(generationConfig[key]);
-        } else {
-          passthrough[key] = cloneJsonValue(generationConfig[key]);
-        }
-      }
-    }
-  }
-
-  if (Array.isArray(requestPayload.tools)) {
-    const requestTools = requestPayload.tools
-      .filter((item) => isRecord(item))
-      .map((item) => {
-        const next: GeminiRecord = {};
-        if (item.googleSearch !== undefined) next.googleSearch = cloneJsonValue(item.googleSearch);
-        if (item.urlContext !== undefined) next.urlContext = cloneJsonValue(item.urlContext);
-        if (item.codeExecution !== undefined) next.codeExecution = cloneJsonValue(item.codeExecution);
-        if (item.functionDeclarations !== undefined) next.functionDeclarations = cloneJsonValue(item.functionDeclarations);
-        return next;
-      })
-      .filter((item) => Object.keys(item).length > 0);
-
-    if (requestTools.length > 0) {
-      passthrough.tools = requestTools;
-    }
-  }
-
-  if (metadata.passthrough && Object.keys(metadata.passthrough).length <= 0) {
-    delete metadata.passthrough;
-  }
-
-  return metadata;
-}
-
-export function serializeGeminiAggregateResponse(
-  payload: GeminiGenerateContentAggregateState | unknown,
-): GeminiRecord {
-  const state = ensureAggregateState(payload);
-  const response: GeminiRecord = {
-    responseId: state.responseId || '',
-    modelVersion: state.modelVersion || '',
-    candidates: buildCandidates(state),
-  };
-
-  const usageMetadata = buildUsageMetadata(state);
-  if (usageMetadata) {
-    response.usageMetadata = usageMetadata;
-  }
-
-  return response;
-}
-
-export function extractTransformerMetadata(payload: unknown, requestPayload?: unknown): TransformerMetadata {
-  const state = ensureAggregateState(payload);
-  const metadata = extractRequestSemantics(requestPayload);
-
-  const citations = extractOrderedCandidateMetadata(state, 'citationMetadata');
-  const groundingMetadata = extractOrderedCandidateMetadata(state, 'groundingMetadata');
-  const thoughtSignatures = extractOrderedThoughtSignatures(state);
-
-  if (citations.length > 0) metadata.citations = citations;
-  if (groundingMetadata.length > 0) metadata.groundingMetadata = groundingMetadata;
-  if (thoughtSignatures.all.length > 0) {
-    if (thoughtSignatures.preferred) {
-      metadata.thoughtSignature = thoughtSignatures.preferred;
-    }
-    metadata.thoughtSignatures = thoughtSignatures.all;
-  }
-  const usageMetadata = buildUsageMetadata(state);
-  if (usageMetadata) metadata.usageMetadata = usageMetadata;
-
-  return metadata;
-}
-
-export function extractResponseMetadata(payload: unknown, requestPayload?: unknown): GeminiRecord {
-  return toTransformerMetadataRecord(extractTransformerMetadata(payload, requestPayload)) ?? {};
-}
-
-export const geminiGenerateContentOutbound = {
-  resolveBaseUrl: resolveGeminiNativeBaseUrl,
-  resolveModelsUrl: resolveGeminiModelsUrl,
-  resolveActionUrl: resolveGeminiGenerateContentUrl,
-  extractTransformerMetadata,
-  extractResponseMetadata,
-  serializeAggregateResponse: serializeGeminiAggregateResponse,
-};
+export {
+  geminiGenerateContentResponseBridge,
+  extractGeminiGenerateContentResponseMetadata as extractResponseMetadata,
+  extractGeminiGenerateContentTransformerMetadata as extractTransformerMetadata,
+  serializeGeminiGenerateContentAggregateResponse as serializeGeminiAggregateResponse,
+} from './responseBridge.js';

src/server/transformers/gemini/generate-content/requestBridge.test.ts

@@ -0,0 +1,318 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  buildGeminiGenerateContentRequestFromOpenAi,
+  buildCanonicalRequestToGeminiGenerateContentBody,
+  parseGeminiGenerateContentRequestToCanonical,
+} from './requestBridge.js';
+
+describe('gemini generate-content request bridge', () => {
+  it('parses Gemini generateContent bodies into canonical envelopes', () => {
+    const result = parseGeminiGenerateContentRequestToCanonical({
+      model: 'gemini-2.5-pro',
+      contents: [
+        {
+          role: 'user',
+          parts: [{ text: 'hello' }],
+        },
+      ],
+      generationConfig: {
+        thinkingConfig: {
+          thinkingBudget: 512,
+        },
+      },
+    });
+
+    expect(result.error).toBeUndefined();
+    expect(result.value).toMatchObject({
+      surface: 'gemini-generate-content',
+      requestedModel: 'gemini-2.5-pro',
+      reasoning: {
+        budgetTokens: 512,
+      },
+      messages: [
+        {
+          role: 'user',
+          parts: [{ type: 'text', text: 'hello' }],
+        },
+      ],
+    });
+  });
+
+  it('builds Gemini generateContent bodies from canonical envelopes', () => {
+    const body = buildCanonicalRequestToGeminiGenerateContentBody({
+      operation: 'generate',
+      surface: 'gemini-generate-content',
+      cliProfile: 'gemini_cli',
+      requestedModel: 'gemini-2.5-pro',
+      stream: false,
+      messages: [{ role: 'user', parts: [{ type: 'text', text: 'hello' }] }],
+      reasoning: {
+        budgetTokens: 512,
+      },
+      tools: [{ name: 'lookup', inputSchema: { type: 'object' } }],
+      toolChoice: 'required',
+    });
+
+    expect(body).toMatchObject({
+      contents: [
+        {
+          role: 'user',
+          parts: [{ text: 'hello' }],
+        },
+      ],
+      tools: [
+        {
+          functionDeclarations: [
+            {
+              name: 'lookup',
+              parameters: { type: 'object' },
+            },
+          ],
+        },
+      ],
+      toolConfig: {
+        functionCallingConfig: {
+          mode: 'ANY',
+        },
+      },
+      generationConfig: {
+        thinkingConfig: {
+          thinkingBudget: 512,
+        },
+      },
+    });
+  });
+
+  it('injects provider thought signatures into functionCall parts', () => {
+    const result = buildGeminiGenerateContentRequestFromOpenAi({
+      modelName: 'gemini-3-flash-preview',
+      body: {
+        model: 'gemini-3-flash-preview',
+        messages: [
+          { role: 'user', content: 'What is the weather?' },
+          {
+            role: 'assistant',
+            content: 'Let me check.',
+            tool_calls: [
+              {
+                id: 'call_123',
+                type: 'function',
+                function: { name: 'get_weather', arguments: '{"city":"Tokyo"}' },
+                provider_specific_fields: { thought_signature: 'real_sig_abc' },
+              },
+            ],
+          },
+          { role: 'tool', tool_call_id: 'call_123', content: '{"temp":"22C"}' },
+        ],
+      },
+    }) as Record<string, unknown>;
+
+    const contents = result.contents as Array<Record<string, unknown>>;
+    const modelMessages = contents.filter((content) => content.role === 'model');
+    const functionCallParts = modelMessages
+      .flatMap((message) => message.parts as Array<Record<string, unknown>>)
+      .filter((part) => 'functionCall' in part);
+
+    expect(functionCallParts).toHaveLength(1);
+    expect(functionCallParts[0].thoughtSignature).toBe('real_sig_abc');
+  });
+
+  it('splits text and signed functionCall parts into separate model messages', () => {
+    const result = buildGeminiGenerateContentRequestFromOpenAi({
+      modelName: 'gemini-3-flash-preview',
+      body: {
+        model: 'gemini-3-flash-preview',
+        messages: [
+          { role: 'user', content: 'Read the file.' },
+          {
+            role: 'assistant',
+            content: 'I will read it.',
+            tool_calls: [
+              {
+                id: 'call_456',
+                type: 'function',
+                function: { name: 'Read', arguments: '{"path":"/tmp/x"}' },
+                provider_specific_fields: { thought_signature: 'sig_split_test' },
+              },
+            ],
+          },
+          { role: 'tool', tool_call_id: 'call_456', content: 'file content here' },
+        ],
+      },
+    }) as Record<string, unknown>;
+
+    const contents = result.contents as Array<Record<string, unknown>>;
+    const modelMessages = contents.filter((content) => content.role === 'model');
+
+    expect(modelMessages).toHaveLength(2);
+    expect((modelMessages[0].parts as Array<Record<string, unknown>>).every((part) => 'text' in part)).toBe(true);
+    expect((modelMessages[1].parts as Array<Record<string, unknown>>).every((part) => 'functionCall' in part)).toBe(true);
+    expect((modelMessages[1].parts as Array<Record<string, unknown>>)[0].thoughtSignature).toBe('sig_split_test');
+  });
+
+  it('disables thinking config when signatures are missing for non-gemini targets', () => {
+    const result = buildGeminiGenerateContentRequestFromOpenAi({
+      modelName: 'claude-sonnet-4-5',
+      body: {
+        model: 'claude-sonnet-4-5',
+        reasoning_effort: 'high',
+        messages: [
+          { role: 'user', content: 'Do something.' },
+          {
+            role: 'assistant',
+            tool_calls: [
+              {
+                id: 'call_no_sig_non_gemini',
+                type: 'function',
+                function: { name: 'Bash', arguments: '{"command":"ls"}' },
+              },
+            ],
+          },
+          { role: 'tool', tool_call_id: 'call_no_sig_non_gemini', content: 'file1\nfile2' },
+        ],
+      },
+    }) as Record<string, unknown>;
+
+    const contents = result.contents as Array<Record<string, unknown>>;
+    const modelMessages = contents.filter((content) => content.role === 'model');
+    const functionCallParts = modelMessages
+      .flatMap((message) => message.parts as Array<Record<string, unknown>>)
+      .filter((part) => 'functionCall' in part);
+
+    expect(functionCallParts).toHaveLength(1);
+    expect(functionCallParts[0].thoughtSignature).toBeUndefined();
+    expect((result.generationConfig as Record<string, unknown> | undefined)?.thinkingConfig).toBeUndefined();
+  });
+
+  it('injects a dummy thought signature when thinking is enabled and no signature is available', () => {
+    const result = buildGeminiGenerateContentRequestFromOpenAi({
+      modelName: 'gemini-3-flash-preview',
+      body: {
+        model: 'gemini-3-flash-preview',
+        reasoning_effort: 'high',
+        messages: [
+          { role: 'user', content: 'Do something.' },
+          {
+            role: 'assistant',
+            tool_calls: [
+              {
+                id: 'call_no_sig',
+                type: 'function',
+                function: { name: 'Bash', arguments: '{"command":"ls"}' },
+              },
+            ],
+          },
+          { role: 'tool', tool_call_id: 'call_no_sig', content: 'file1\nfile2' },
+        ],
+      },
+    }) as Record<string, unknown>;
+
+    const contents = result.contents as Array<Record<string, unknown>>;
+    const modelMessages = contents.filter((content) => content.role === 'model');
+    const functionCallParts = modelMessages
+      .flatMap((message) => message.parts as Array<Record<string, unknown>>)
+      .filter((part) => 'functionCall' in part);
+
+    expect(functionCallParts).toHaveLength(1);
+    expect(typeof functionCallParts[0].thoughtSignature).toBe('string');
+    expect((functionCallParts[0].thoughtSignature as string).length).toBeGreaterThan(0);
+  });
+
+  it('preserves functionResponse count matching functionCall count', () => {
+    const result = buildGeminiGenerateContentRequestFromOpenAi({
+      modelName: 'gemini-3-flash-preview',
+      body: {
+        model: 'gemini-3-flash-preview',
+        messages: [
+          { role: 'user', content: 'Read two files.' },
+          {
+            role: 'assistant',
+            tool_calls: [
+              {
+                id: 'call_a',
+                type: 'function',
+                function: { name: 'Read', arguments: '{"path":"/a"}' },
+                provider_specific_fields: { thought_signature: 'sig_a' },
+              },
+              {
+                id: 'call_b',
+                type: 'function',
+                function: { name: 'Read', arguments: '{"path":"/b"}' },
+                provider_specific_fields: { thought_signature: 'sig_b' },
+              },
+            ],
+          },
+          { role: 'tool', tool_call_id: 'call_a', content: 'content a' },
+          { role: 'tool', tool_call_id: 'call_b', content: 'content b' },
+        ],
+      },
+    }) as Record<string, unknown>;
+
+    const contents = result.contents as Array<Record<string, unknown>>;
+    let functionCallCount = 0;
+    let functionResponseCount = 0;
+    for (const content of contents) {
+      for (const part of (content.parts as Array<Record<string, unknown>>)) {
+        if ('functionCall' in part) functionCallCount += 1;
+        if ('functionResponse' in part) functionResponseCount += 1;
+      }
+    }
+
+    expect(functionCallCount).toBe(2);
+    expect(functionResponseCount).toBe(2);
+  });
+
+  it('uses canonical tool-call names for Gemini functionResponse parts', () => {
+    const body = buildCanonicalRequestToGeminiGenerateContentBody({
+      operation: 'generate',
+      surface: 'gemini-generate-content',
+      cliProfile: 'generic',
+      requestedModel: 'gemini-2.5-pro',
+      stream: false,
+      messages: [
+        {
+          role: 'assistant',
+          parts: [{
+            type: 'tool_call',
+            id: 'call_weather',
+            name: 'lookup_weather',
+            argumentsJson: '{"city":"Paris"}',
+          }],
+        },
+        {
+          role: 'tool',
+          parts: [{
+            type: 'tool_result',
+            toolCallId: 'call_weather',
+            resultJson: { temperature: '22C' },
+          }],
+        },
+      ],
+    });
+
+    expect(body.contents).toEqual([
+      {
+        role: 'model',
+        parts: [{
+          functionCall: {
+            id: 'call_weather',
+            name: 'lookup_weather',
+            args: { city: 'Paris' },
+          },
+        }],
+      },
+      {
+        role: 'user',
+        parts: [{
+          functionResponse: {
+            name: 'lookup_weather',
+            response: {
+              result: { temperature: '22C' },
+            },
+          },
+        }],
+      },
+    ]);
+  });
+});

src/server/transformers/gemini/generate-content/requestBridge.thoughtSignature.test.ts

@@ -1,5 +1,5 @@
 import { describe, it, expect } from 'vitest';
-import { buildGeminiGenerateContentRequestFromOpenAi } from './geminiCliCompat.js';
+import { buildGeminiGenerateContentRequestFromOpenAi } from './requestBridge.js';
 
 describe('thoughtSignature injection in OpenAI→Gemini conversion', () => {
   it('injects thoughtSignature from provider_specific_fields into functionCall parts', () => {
@@ -148,6 +148,40 @@ describe('thoughtSignature injection in OpenAI→Gemini conversion', () => {
     expect(fcParts[0].thoughtSignature).toBeUndefined();
   });
 
+  it('does not inject dummy signature or preserve thinkingConfig for non-gemini targets when signature is missing', () => {
+    const result = buildGeminiGenerateContentRequestFromOpenAi({
+      body: {
+        model: 'claude-sonnet-4-5',
+        reasoning_effort: 'high',
+        messages: [
+          { role: 'user', content: 'Do something.' },
+          {
+            role: 'assistant',
+            tool_calls: [
+              {
+                id: 'call_no_sig_non_gemini',
+                type: 'function',
+                function: { name: 'Bash', arguments: '{"command":"ls"}' },
+              },
+            ],
+          },
+          { role: 'tool', tool_call_id: 'call_no_sig_non_gemini', content: 'file1\nfile2' },
+        ],
+      },
+      modelName: 'claude-sonnet-4-5',
+    }) as Record<string, unknown>;
+
+    const contents = result.contents as Array<Record<string, unknown>>;
+    const modelMsgs = contents.filter((c) => c.role === 'model');
+    const fcParts = modelMsgs
+      .flatMap((m) => (m.parts as Array<Record<string, unknown>>))
+      .filter((p) => 'functionCall' in p);
+
+    expect(fcParts.length).toBe(1);
+    expect(fcParts[0].thoughtSignature).toBeUndefined();
+    expect((result.generationConfig as Record<string, unknown> | undefined)?.thinkingConfig).toBeUndefined();
+  });
+
   it('preserves functionResponse count matching functionCall count', () => {
     const result = buildGeminiGenerateContentRequestFromOpenAi({
       body: {

src/server/transformers/gemini/generate-content/requestBridge.ts

@@ -1,19 +1,51 @@
-import { TextDecoder, TextEncoder } from 'node:util';
-import { resolveGeminiThinkingConfigFromRequest } from '../../transformers/gemini/generate-content/convert.js';
-
-// Dummy sentinel used when no real thoughtSignature is available but thinking
-// mode is enabled. Gemini accepts any base64 string and won't reject this.
-const DUMMY_THOUGHT_SIGNATURE = 'c2tpcF90aG91Z2h0X3NpZ25hdHVyZV92YWxpZGF0b3I=';
-
-function isRecord(value: unknown): value is Record<string, unknown> {
-  return !!value && typeof value === 'object' && !Array.isArray(value);
-}
+import { canonicalRequestFromOpenAiBody } from '../../canonical/openAiRequestBridge.js';
+import { isCanonicalFunctionTool, isCanonicalNamedToolChoice } from '../../canonical/tools.js';
+import type { CanonicalContentPart, CanonicalRequestEnvelope } from '../../canonical/types.js';
+import type { ProtocolParseContext } from '../../contracts.js';
+import { geminiGenerateContentInbound } from './inbound.js';
+import { buildOpenAiBodyFromGeminiRequest } from './compatibility.js';
+import {
+  reasoningEffortToGeminiThinkingConfig,
+  resolveGeminiThinkingConfigFromRequest,
+} from './convert.js';
 
 function asTrimmedString(value: unknown): string {
   return typeof value === 'string' ? value.trim() : '';
 }
 
-function parseDataUrl(url: string): { mimeType: string; data: string } | null {
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return !!value && typeof value === 'object' && !Array.isArray(value);
+}
+
+function parseJsonString(raw: string): unknown {
+  const trimmed = raw.trim();
+  if (!trimmed) return {};
+  try {
+    return JSON.parse(trimmed);
+  } catch {
+    return { raw };
+  }
+}
+
+function parseDataUrl(value: string): { mimeType: string; data: string } | null {
+  const match = /^data:([^;,]+);base64,(.+)$/i.exec(value.trim());
+  if (!match) return null;
+  return {
+    mimeType: match[1],
+    data: match[2],
+  };
+}
+
+// Dummy sentinel used when no real thoughtSignature is available but thinking
+// mode is enabled. Gemini accepts any base64 string and won't reject this.
+const DUMMY_THOUGHT_SIGNATURE = 'c2tpcF90aG91Z2h0X3NpZ25hdHVyZV92YWxpZGF0b3I=';
+
+function isDummyThoughtSafeModel(modelName: string): boolean {
+  const normalized = asTrimmedString(modelName).toLowerCase();
+  return normalized.startsWith('gemini-') || normalized.startsWith('models/gemini-');
+}
+
+function parseInlineDataUrl(url: string): { mimeType: string; data: string } | null {
   if (!url.startsWith('data:')) return null;
   const [, rest] = url.split('data:', 2);
   const [meta, data] = rest.split(',', 2);
@@ -36,7 +68,7 @@ function normalizeFunctionResponseResult(value: unknown): unknown {
   }
 }
 
-function convertContentToGeminiParts(content: unknown): Array<Record<string, unknown>> {
+function convertOpenAiContentToGeminiParts(content: unknown): Array<Record<string, unknown>> {
   if (typeof content === 'string') {
     const trimmed = content.trim();
     return trimmed ? [{ text: trimmed }] : [];
@@ -63,7 +95,7 @@ function convertContentToGeminiParts(content: unknown): Array<Record<string, unk
     }
     if (type === 'image_url') {
       const imageUrl = asTrimmedString(item.image_url && isRecord(item.image_url) ? item.image_url.url : item.url);
-      const parsed = imageUrl ? parseDataUrl(imageUrl) : null;
+      const parsed = imageUrl ? parseInlineDataUrl(imageUrl) : null;
       if (parsed) {
         parts.push({
           inlineData: {
@@ -89,7 +121,7 @@ function convertContentToGeminiParts(content: unknown): Array<Record<string, unk
   return parts;
 }
 
-function buildGeminiTools(tools: unknown): Array<Record<string, unknown>> | undefined {
+function buildGeminiToolsFromOpenAi(tools: unknown): Array<Record<string, unknown>> | undefined {
   if (!Array.isArray(tools)) return undefined;
   const declarations = tools
     .filter((item) => isRecord(item))
@@ -109,7 +141,7 @@ function buildGeminiTools(tools: unknown): Array<Record<string, unknown>> | unde
   return [{ functionDeclarations: declarations }];
 }
 
-function buildGeminiToolConfig(toolChoice: unknown): Record<string, unknown> | undefined {
+function buildGeminiToolConfigFromOpenAi(toolChoice: unknown): Record<string, unknown> | undefined {
   if (typeof toolChoice === 'string') {
     const normalized = toolChoice.trim().toLowerCase();
     if (normalized === 'none') {
@@ -147,11 +179,10 @@ export function buildGeminiGenerateContentRequestFromOpenAi(input: {
     }
   }
 
-  // Detect if thinking mode is enabled (needed for dummy signature fallback)
   const hasThinkingEnabled = !!resolveGeminiThinkingConfigFromRequest(input.modelName, input.body);
+  const allowsDummyThoughtSignature = isDummyThoughtSafeModel(input.modelName);
+  let shouldDisableThinkingConfig = false;
 
-  // Collect thoughtSignatures from assistant tool_calls for injection.
-  // OpenAI format stores signatures in provider_specific_fields or encoded in tool call IDs.
   const thoughtSignatureById = new Map<string, string>();
   for (const message of messages) {
     if (!isRecord(message) || asTrimmedString(message.role) !== 'assistant') continue;
@@ -160,7 +191,6 @@ export function buildGeminiGenerateContentRequestFromOpenAi(input: {
       if (!isRecord(toolCall)) continue;
       const id = asTrimmedString(toolCall.id);
       if (!id) continue;
-      // Check provider_specific_fields first
       const providerFields = isRecord(toolCall.provider_specific_fields) ? toolCall.provider_specific_fields : null;
       if (providerFields && typeof providerFields.thought_signature === 'string') {
         thoughtSignatureById.set(id, providerFields.thought_signature);
@@ -177,7 +207,7 @@ export function buildGeminiGenerateContentRequestFromOpenAi(input: {
     if (!isRecord(message)) continue;
     const role = asTrimmedString(message.role).toLowerCase();
     if (role === 'system' || role === 'developer') {
-      systemParts.push(...convertContentToGeminiParts(message.content));
+      systemParts.push(...convertOpenAiContentToGeminiParts(message.content));
       continue;
     }
     if (role === 'tool') {
@@ -201,7 +231,7 @@ export function buildGeminiGenerateContentRequestFromOpenAi(input: {
       continue;
     }
 
-    const textParts = convertContentToGeminiParts(message.content);
+    const textParts = convertOpenAiContentToGeminiParts(message.content);
     const fcParts: Array<Record<string, unknown>> = [];
     const toolCalls = Array.isArray(message.tool_calls) ? message.tool_calls : [];
     for (const toolCall of toolCalls) {
@@ -222,29 +252,21 @@ export function buildGeminiGenerateContentRequestFromOpenAi(input: {
       const fcPart: Record<string, unknown> = {
         functionCall: { name, args },
       };
-      // Inject thoughtSignature if available from provider_specific_fields.
-      // If no signature is found, use a dummy sentinel so Gemini doesn't reject
-      // the request outright. Gemini 3+ requires thoughtSignature on all
-      // functionCall parts when thinking is enabled.
-      // Ref: https://ai.google.dev/gemini-api/docs/thought-signatures
       const id = asTrimmedString(toolCall.id);
       const signature = thoughtSignatureById.get(id);
       if (signature) {
         fcPart.thoughtSignature = signature;
-      } else if (hasThinkingEnabled) {
+      } else if (hasThinkingEnabled && allowsDummyThoughtSignature) {
         fcPart.thoughtSignature = DUMMY_THOUGHT_SIGNATURE;
+      } else if (hasThinkingEnabled) {
+        shouldDisableThinkingConfig = true;
       }
       fcParts.push(fcPart);
     }
 
     const geminiRole = role === 'assistant' ? 'model' : 'user';
-
-    // Gemini requires that parts WITH thoughtSignature and parts WITHOUT
-    // are not mixed in the same message. Split if needed.
-    // Ref: https://ai.google.dev/gemini-api/docs/thought-signatures
-    const hasSigned = fcParts.some((p) => 'thoughtSignature' in p);
+    const hasSigned = fcParts.some((part) => 'thoughtSignature' in part);
     if (hasSigned && textParts.length > 0 && fcParts.length > 0) {
-      // Emit text parts first, then functionCall parts in a separate message
       request.contents = [
         ...(Array.isArray(request.contents) ? request.contents : []),
         { role: geminiRole, parts: textParts },
@@ -287,18 +309,18 @@ export function buildGeminiGenerateContentRequestFromOpenAi(input: {
     generationConfig.stopSequences = input.body.stop.filter((item): item is string => typeof item === 'string' && item.trim().length > 0);
   }
   const thinkingConfig = resolveGeminiThinkingConfigFromRequest(input.modelName, input.body);
-  if (thinkingConfig) {
+  if (thinkingConfig && !shouldDisableThinkingConfig) {
     generationConfig.thinkingConfig = thinkingConfig;
   }
   if (Object.keys(generationConfig).length > 0) {
     request.generationConfig = generationConfig;
   }
 
-  const geminiTools = buildGeminiTools(input.body.tools);
+  const geminiTools = buildGeminiToolsFromOpenAi(input.body.tools);
   if (geminiTools) {
     request.tools = geminiTools;
   }
-  const toolConfig = buildGeminiToolConfig(input.body.tool_choice);
+  const toolConfig = buildGeminiToolConfigFromOpenAi(input.body.tool_choice);
   if (toolConfig) {
     request.toolConfig = toolConfig;
   }
@@ -306,91 +328,225 @@ export function buildGeminiGenerateContentRequestFromOpenAi(input: {
   return request;
 }
 
-export function wrapGeminiCliRequest(input: {
-  modelName: string;
-  projectId: string;
-  request: Record<string, unknown>;
-}) {
-  const { model, ...requestPayload } = input.request;
+function canonicalPartToGeminiPart(
+  part: CanonicalContentPart,
+  toolNameById?: ReadonlyMap<string, string>,
+): Record<string, unknown> | null {
+  if (part.type === 'text') {
     return {
-    project: input.projectId,
-    model: input.modelName,
-    request: requestPayload,
+      text: part.text,
+      ...(part.thought === true ? { thought: true } : {}),
     };
   }
 
-export function unwrapGeminiCliPayload<T>(payload: T): unknown {
-  if (!isRecord(payload)) return payload;
-  if (payload.response !== undefined) {
-    return payload.response;
+  if (part.type === 'image') {
+    const source = typeof part.dataUrl === 'string' && part.dataUrl.trim()
+      ? part.dataUrl
+      : (typeof part.url === 'string' ? part.url : '');
+    if (!source) return null;
+
+    const dataUrl = parseDataUrl(source);
+    if (dataUrl) {
+      return {
+        inlineData: {
+          mimeType: dataUrl.mimeType,
+          data: dataUrl.data,
+        },
+      };
     }
+
+    return {
+      fileData: {
+        fileUri: source,
+        ...(part.mimeType ? { mimeType: part.mimeType } : {}),
+      },
+    };
+  }
+
+  if (part.type === 'file') {
+    if (part.fileData) {
+      return {
+        inlineData: {
+          mimeType: part.mimeType || 'application/octet-stream',
+          data: part.fileData,
+        },
+      };
+    }
+
+    const fileUri = part.fileUrl || part.fileId;
+    if (!fileUri) return null;
+    return {
+      fileData: {
+        fileUri,
+        ...(part.mimeType ? { mimeType: part.mimeType } : {}),
+      },
+    };
+  }
+
+  if (part.type === 'tool_call') {
+    return {
+      functionCall: {
+        id: part.id,
+        name: part.name,
+        args: parseJsonString(part.argumentsJson),
+      },
+    };
+  }
+
+  if (part.type === 'tool_result') {
+    const response = part.resultJson ?? parseJsonString(part.resultText ?? '');
+    return {
+      functionResponse: {
+        name: toolNameById?.get(part.toolCallId) || 'unknown',
+        response: {
+          result: response,
+        },
+      },
+    };
+  }
+
+  return null;
+}
+
+export function buildCanonicalRequestToGeminiGenerateContentBody(
+  request: CanonicalRequestEnvelope,
+): Record<string, unknown> {
+  const contents: Array<Record<string, unknown>> = [];
+  const systemParts: Array<Record<string, unknown>> = [];
+  const toolNameById = new Map<string, string>();
+
+  for (const message of request.messages) {
+    if (message.role === 'system' || message.role === 'developer') {
+      systemParts.push(
+        ...message.parts
+          .map((part) => canonicalPartToGeminiPart(part, toolNameById))
+          .filter((part): part is Record<string, unknown> => !!part),
+      );
+      continue;
+    }
+
+    for (const part of message.parts) {
+      if (part.type === 'tool_call') {
+        toolNameById.set(part.id, part.name);
+      }
+    }
+
+    const parts = message.parts
+      .map((part) => canonicalPartToGeminiPart(part, toolNameById))
+      .filter((part): part is Record<string, unknown> => !!part);
+
+    if (parts.length <= 0) continue;
+
+    if (message.role === 'tool') {
+      contents.push({
+        role: 'user',
+        parts,
+      });
+      continue;
+    }
+
+    contents.push({
+      role: message.role === 'assistant' ? 'model' : 'user',
+      parts,
+    });
+  }
+
+  const payload: Record<string, unknown> = {
+    contents,
+  };
+
+  if (systemParts.length > 0) {
+    payload.systemInstruction = {
+      role: 'user',
+      parts: systemParts,
+    };
+  }
+
+  const generationConfig: Record<string, unknown> = {};
+  if (request.reasoning?.budgetTokens !== undefined) {
+    generationConfig.thinkingConfig = {
+      thinkingBudget: request.reasoning.budgetTokens,
+    };
+  } else if (request.reasoning?.effort) {
+    generationConfig.thinkingConfig = reasoningEffortToGeminiThinkingConfig(
+      request.requestedModel,
+      request.reasoning.effort,
+    );
+  }
+  if (Object.keys(generationConfig).length > 0) {
+    payload.generationConfig = generationConfig;
+  }
+
+  if (Array.isArray(request.tools) && request.tools.length > 0) {
+    const functionTools = request.tools.filter(isCanonicalFunctionTool);
+    if (functionTools.length > 0) {
+      payload.tools = [{
+        functionDeclarations: functionTools.map((tool) => ({
+          name: tool.name,
+          ...(tool.description ? { description: tool.description } : {}),
+          ...(tool.inputSchema ? { parameters: tool.inputSchema } : {}),
+        })),
+      }];
+    }
+  }
+
+  if (request.toolChoice) {
+    if (request.toolChoice === 'none') {
+      payload.toolConfig = { functionCallingConfig: { mode: 'NONE' } };
+    } else if (request.toolChoice === 'auto') {
+      payload.toolConfig = { functionCallingConfig: { mode: 'AUTO' } };
+    } else if (request.toolChoice === 'required') {
+      payload.toolConfig = { functionCallingConfig: { mode: 'ANY' } };
+    } else if (isCanonicalNamedToolChoice(request.toolChoice)) {
+      payload.toolConfig = {
+        functionCallingConfig: {
+          mode: 'ANY',
+          allowedFunctionNames: [request.toolChoice.name],
+        },
+      };
+    }
+  }
+
   return payload;
 }
 
-function rewriteGeminiCliSseEventBlock(block: string): string {
-  const lines = block.split(/\r?\n/g);
-  return lines.map((line) => {
-    if (!line.startsWith('data:')) return line;
-    const data = line.slice(5).trim();
-    if (!data || data === '[DONE]') return line;
-    try {
-      const parsed = JSON.parse(data);
-      return `data: ${JSON.stringify(unwrapGeminiCliPayload(parsed))}`;
-    } catch {
-      return line;
-    }
-  }).join('\n');
-}
-
-export function createGeminiCliStreamReader(reader: {
-  read(): Promise<{ done: boolean; value?: Uint8Array }>;
-  cancel(reason?: unknown): Promise<unknown>;
-  releaseLock(): void;
-}) {
-  const decoder = new TextDecoder();
-  const encoder = new TextEncoder();
-  const outputQueue: Uint8Array[] = [];
-  let buffer = '';
-  let done = false;
-
-  async function fillQueue() {
-    while (outputQueue.length <= 0 && !done) {
-      const result = await reader.read();
-      if (result.done) {
-        done = true;
-        const tail = decoder.decode();
-        if (tail) buffer += tail;
-        if (buffer.trim()) {
-          outputQueue.push(encoder.encode(`${rewriteGeminiCliSseEventBlock(buffer)}\n\n`));
-          buffer = '';
-        }
-        break;
-      }
-      if (!result.value) continue;
-      buffer += decoder.decode(result.value, { stream: true });
-      let separatorIndex = buffer.indexOf('\n\n');
-      while (separatorIndex >= 0) {
-        const block = buffer.slice(0, separatorIndex);
-        buffer = buffer.slice(separatorIndex + 2);
-        outputQueue.push(encoder.encode(`${rewriteGeminiCliSseEventBlock(block)}\n\n`));
-        separatorIndex = buffer.indexOf('\n\n');
-      }
-    }
-  }
-
+export function parseGeminiGenerateContentRequestToCanonical(
+  body: unknown,
+  ctx?: ProtocolParseContext,
+): { value?: CanonicalRequestEnvelope; error?: { statusCode: number; payload: unknown } } {
+  const rawBody = isRecord(body) ? body : {};
+  const requestedModel = asTrimmedString(rawBody.model ?? ctx?.metadata?.requestedModel);
+  if (!requestedModel) {
     return {
-    async read() {
-      await fillQueue();
-      if (outputQueue.length > 0) {
-        return { done: false, value: outputQueue.shift() };
-      }
-      return { done: true, value: undefined };
+      error: {
+        statusCode: 400,
+        payload: {
+          error: {
+            message: 'model is required',
+            type: 'invalid_request_error',
           },
-    cancel(reason?: unknown) {
-      return reader.cancel(reason);
         },
-    releaseLock() {
-      reader.releaseLock();
       },
     };
   }
+
+  const stream = rawBody.stream === true || ctx?.metadata?.stream === true;
+  const normalizedBody = geminiGenerateContentInbound.normalizeRequest(rawBody, requestedModel);
+  const openAiBody = buildOpenAiBodyFromGeminiRequest({
+    body: normalizedBody,
+    modelName: requestedModel,
+    stream,
+  });
+
+  return {
+    value: canonicalRequestFromOpenAiBody({
+      body: openAiBody,
+      surface: 'gemini-generate-content',
+      cliProfile: ctx?.cliProfile,
+      operation: ctx?.operation,
+      metadata: ctx?.metadata,
+      passthrough: ctx?.passthrough,
+      continuation: ctx?.continuation,
+    }),
+  };
+}

src/server/transformers/gemini/generate-content/responseBridge.test.ts

@@ -0,0 +1,98 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  extractGeminiGenerateContentResponseMetadata,
+  extractGeminiGenerateContentTransformerMetadata,
+  geminiGenerateContentResponseBridge,
+  serializeGeminiGenerateContentAggregateResponse,
+} from './responseBridge.js';
+import { geminiGenerateContentOutbound } from './outbound.js';
+
+describe('gemini generate-content response bridge', () => {
+  it('serializes aggregate responses with usage metadata', () => {
+    const payload = serializeGeminiGenerateContentAggregateResponse({
+      responseId: 'resp-1',
+      modelVersion: 'gemini-2.5-pro',
+      finishReason: 'STOP',
+      parts: [{ text: 'hello' }],
+      candidates: [],
+      groundingMetadata: [],
+      citations: [],
+      thoughtSignatures: [],
+      usage: {
+        promptTokenCount: 10,
+        candidatesTokenCount: 4,
+        totalTokenCount: 14,
+      },
+    });
+
+    expect(payload).toEqual({
+      responseId: 'resp-1',
+      modelVersion: 'gemini-2.5-pro',
+      candidates: [{
+        index: 0,
+        finishReason: 'STOP',
+        content: {
+          role: 'model',
+          parts: [{ text: 'hello' }],
+        },
+      }],
+      usageMetadata: {
+        promptTokenCount: 10,
+        candidatesTokenCount: 4,
+        totalTokenCount: 14,
+      },
+    });
+  });
+
+  it('extracts transformer and response metadata from aggregate payloads', () => {
+    const payload = {
+      responseId: 'resp-1',
+      modelVersion: 'gemini-2.5-pro',
+      finishReason: 'STOP',
+      parts: [{ text: 'hello', thoughtSignature: 'sig-1', thought: true }],
+      candidates: [],
+      groundingMetadata: [{ webSearchQueries: ['cat'] }],
+      citations: [{ citations: [{ uri: 'https://a.example' }] }],
+      thoughtSignatures: ['sig-1'],
+      usage: {
+        promptTokenCount: 10,
+        candidatesTokenCount: 4,
+        totalTokenCount: 14,
+      },
+    };
+
+    expect(extractGeminiGenerateContentTransformerMetadata(payload, {
+      systemInstruction: { parts: [{ text: 'system prompt' }] },
+    })).toMatchObject({
+      citations: [{ citations: [{ uri: 'https://a.example' }] }],
+      groundingMetadata: [{ webSearchQueries: ['cat'] }],
+      thoughtSignature: 'sig-1',
+      thoughtSignatures: ['sig-1'],
+      usageMetadata: {
+        promptTokenCount: 10,
+        candidatesTokenCount: 4,
+        totalTokenCount: 14,
+      },
+      passthrough: {
+        systemInstruction: { parts: [{ text: 'system prompt' }] },
+      },
+    });
+
+    expect(extractGeminiGenerateContentResponseMetadata(payload)).toMatchObject({
+      citations: [{ citations: [{ uri: 'https://a.example' }] }],
+      groundingMetadata: [{ webSearchQueries: ['cat'] }],
+      thoughtSignature: 'sig-1',
+      thoughtSignatures: ['sig-1'],
+      usageMetadata: {
+        promptTokenCount: 10,
+        candidatesTokenCount: 4,
+        totalTokenCount: 14,
+      },
+    });
+  });
+
+  it('keeps the outbound facade pointed at the response bridge object', () => {
+    expect(geminiGenerateContentOutbound).toBe(geminiGenerateContentResponseBridge);
+  });
+});

src/server/transformers/gemini/generate-content/responseBridge.ts

@@ -0,0 +1,287 @@
+import {
+  applyGeminiGenerateContentAggregate,
+  createGeminiGenerateContentAggregateState,
+  type GeminiGenerateContentAggregateState,
+} from './aggregator.js';
+import {
+  toTransformerMetadataRecord,
+  type TransformerMetadata,
+} from '../../shared/normalized.js';
+import {
+  resolveGeminiGenerateContentUrl,
+  resolveGeminiModelsUrl,
+  resolveGeminiNativeBaseUrl,
+} from './urlResolver.js';
+
+type GeminiRecord = Record<string, unknown>;
+
+function isRecord(value: unknown): value is GeminiRecord {
+  return !!value && typeof value === 'object' && !Array.isArray(value);
+}
+
+function cloneJsonValue<T>(value: T): T {
+  if (Array.isArray(value)) {
+    return value.map((item) => cloneJsonValue(item)) as T;
+  }
+  if (isRecord(value)) {
+    return Object.fromEntries(
+      Object.entries(value).map(([key, item]) => [key, cloneJsonValue(item)]),
+    ) as T;
+  }
+  return value;
+}
+
+function isAggregateState(value: unknown): value is GeminiGenerateContentAggregateState {
+  return isRecord(value) && Array.isArray(value.parts) && Array.isArray(value.groundingMetadata) && Array.isArray(value.citations);
+}
+
+function ensureAggregateState(payload: unknown): GeminiGenerateContentAggregateState {
+  if (isAggregateState(payload)) return payload;
+
+  const state = createGeminiGenerateContentAggregateState();
+  const chunks = Array.isArray(payload) ? payload : [payload];
+  for (const chunk of chunks) {
+    applyGeminiGenerateContentAggregate(state, chunk);
+  }
+  return state;
+}
+
+function buildUsageMetadata(state: GeminiGenerateContentAggregateState): GeminiRecord | undefined {
+  const usage = state.usage;
+  const next: GeminiRecord = {};
+  if (typeof usage.promptTokenCount === 'number') next.promptTokenCount = usage.promptTokenCount;
+  if (typeof usage.candidatesTokenCount === 'number') next.candidatesTokenCount = usage.candidatesTokenCount;
+  if (typeof usage.totalTokenCount === 'number') next.totalTokenCount = usage.totalTokenCount;
+  if (typeof usage.cachedContentTokenCount === 'number') next.cachedContentTokenCount = usage.cachedContentTokenCount;
+  if (typeof usage.thoughtsTokenCount === 'number') next.thoughtsTokenCount = usage.thoughtsTokenCount;
+  return Object.keys(next).length > 0 ? next : undefined;
+}
+
+function buildCandidates(state: GeminiGenerateContentAggregateState): GeminiRecord[] {
+  if (state.candidates.length > 0) {
+    return state.candidates
+      .slice()
+      .sort((left, right) => left.index - right.index)
+      .map((candidate) => {
+        const next: GeminiRecord = {
+          index: candidate.index,
+          finishReason: candidate.finishReason || 'STOP',
+          content: {
+            role: 'model',
+            parts: candidate.parts,
+          },
+        };
+        if (candidate.groundingMetadata) next.groundingMetadata = candidate.groundingMetadata;
+        if (candidate.citationMetadata) next.citationMetadata = candidate.citationMetadata;
+        return next;
+      });
+  }
+
+  const fallback: GeminiRecord = {
+    index: 0,
+    finishReason: state.finishReason || 'STOP',
+    content: {
+      role: 'model',
+      parts: state.parts,
+    },
+  };
+  if (state.groundingMetadata.length > 0) {
+    fallback.groundingMetadata = state.groundingMetadata[0];
+  }
+  if (state.citations.length > 0) {
+    fallback.citationMetadata = state.citations[0];
+  }
+  return [fallback];
+}
+
+function extractOrderedCandidateMetadata(
+  state: GeminiGenerateContentAggregateState,
+  key: 'groundingMetadata' | 'citationMetadata',
+): GeminiRecord[] {
+  if (state.candidates.length > 0) {
+    return state.candidates
+      .slice()
+      .sort((left, right) => left.index - right.index)
+      .map((candidate) => candidate[key])
+      .filter((item): item is GeminiRecord => isRecord(item))
+      .map((item) => cloneJsonValue(item));
+  }
+
+  const fallback = key === 'groundingMetadata' ? state.groundingMetadata : state.citations;
+  return fallback.map((item) => cloneJsonValue(item));
+}
+
+function extractOrderedThoughtSignatures(
+  state: GeminiGenerateContentAggregateState,
+): {
+  all: string[];
+  preferred: string | undefined;
+} {
+  const ordered = new Set<string>();
+  const preferredThoughts = new Set<string>();
+
+  const candidates = state.candidates.length > 0
+    ? state.candidates.slice().sort((left, right) => left.index - right.index)
+    : [{ index: 0, finishReason: state.finishReason, parts: state.parts }];
+
+  for (const candidate of candidates) {
+    for (const part of candidate.parts) {
+      if (!isRecord(part) || typeof part.thoughtSignature !== 'string' || !part.thoughtSignature.trim()) {
+        continue;
+      }
+      ordered.add(part.thoughtSignature);
+      if (part.thought === true) {
+        preferredThoughts.add(part.thoughtSignature);
+      }
+    }
+  }
+
+  const orderedList = [...ordered];
+  const preferred = [...preferredThoughts][0]
+    ?? (orderedList.length === 1 ? orderedList[0] : undefined);
+
+  return {
+    all: orderedList.length > 0 ? orderedList : [...state.thoughtSignatures],
+    preferred,
+  };
+}
+
+function ensurePassthrough(metadata: TransformerMetadata): Record<string, unknown> {
+  if (!metadata.passthrough) metadata.passthrough = {};
+  return metadata.passthrough;
+}
+
+function extractRequestSemantics(requestPayload: unknown): TransformerMetadata {
+  if (!isRecord(requestPayload)) return {};
+
+  const metadata: TransformerMetadata = {};
+  const passthrough = ensurePassthrough(metadata);
+  if (requestPayload.systemInstruction !== undefined) {
+    passthrough.systemInstruction = cloneJsonValue(requestPayload.systemInstruction);
+  }
+  if (requestPayload.cachedContent !== undefined) {
+    passthrough.cachedContent = cloneJsonValue(requestPayload.cachedContent);
+  }
+  if (requestPayload.safetySettings !== undefined) {
+    metadata.geminiSafetySettings = cloneJsonValue(requestPayload.safetySettings);
+  }
+  if (requestPayload.toolConfig !== undefined) {
+    passthrough.toolConfig = cloneJsonValue(requestPayload.toolConfig);
+  }
+
+  const generationConfig = isRecord(requestPayload.generationConfig) ? requestPayload.generationConfig : null;
+  if (generationConfig) {
+    const preservedKeys = [
+      'stopSequences',
+      'responseModalities',
+      'responseMimeType',
+      'responseSchema',
+      'candidateCount',
+      'maxOutputTokens',
+      'temperature',
+      'topP',
+      'topK',
+      'presencePenalty',
+      'frequencyPenalty',
+      'seed',
+      'responseLogprobs',
+      'logprobs',
+      'thinkingConfig',
+      'imageConfig',
+    ];
+    for (const key of preservedKeys) {
+      if (generationConfig[key] !== undefined) {
+        if (key === 'imageConfig') {
+          metadata.geminiImageConfig = cloneJsonValue(generationConfig[key]);
+        } else {
+          passthrough[key] = cloneJsonValue(generationConfig[key]);
+        }
+      }
+    }
+  }
+
+  if (Array.isArray(requestPayload.tools)) {
+    const requestTools = requestPayload.tools
+      .filter((item) => isRecord(item))
+      .map((item) => {
+        const next: GeminiRecord = {};
+        if (item.googleSearch !== undefined) next.googleSearch = cloneJsonValue(item.googleSearch);
+        if (item.urlContext !== undefined) next.urlContext = cloneJsonValue(item.urlContext);
+        if (item.codeExecution !== undefined) next.codeExecution = cloneJsonValue(item.codeExecution);
+        if (item.functionDeclarations !== undefined) next.functionDeclarations = cloneJsonValue(item.functionDeclarations);
+        return next;
+      })
+      .filter((item) => Object.keys(item).length > 0);
+
+    if (requestTools.length > 0) {
+      passthrough.tools = requestTools;
+    }
+  }
+
+  if (metadata.passthrough && Object.keys(metadata.passthrough).length <= 0) {
+    delete metadata.passthrough;
+  }
+
+  return metadata;
+}
+
+export function serializeGeminiGenerateContentAggregateResponse(
+  payload: GeminiGenerateContentAggregateState | unknown,
+): GeminiRecord {
+  const state = ensureAggregateState(payload);
+  const response: GeminiRecord = {
+    responseId: state.responseId || '',
+    modelVersion: state.modelVersion || '',
+    candidates: buildCandidates(state),
+  };
+
+  const usageMetadata = buildUsageMetadata(state);
+  if (usageMetadata) {
+    response.usageMetadata = usageMetadata;
+  }
+
+  return response;
+}
+
+export function extractGeminiGenerateContentTransformerMetadata(payload: unknown, requestPayload?: unknown): TransformerMetadata {
+  const state = ensureAggregateState(payload);
+  const metadata = extractRequestSemantics(requestPayload);
+
+  const citations = extractOrderedCandidateMetadata(state, 'citationMetadata');
+  const groundingMetadata = extractOrderedCandidateMetadata(state, 'groundingMetadata');
+  const thoughtSignatures = extractOrderedThoughtSignatures(state);
+
+  if (citations.length > 0) metadata.citations = citations;
+  if (groundingMetadata.length > 0) metadata.groundingMetadata = groundingMetadata;
+  if (thoughtSignatures.all.length > 0) {
+    if (thoughtSignatures.preferred) {
+      metadata.thoughtSignature = thoughtSignatures.preferred;
+    }
+    metadata.thoughtSignatures = thoughtSignatures.all;
+  }
+  const usageMetadata = buildUsageMetadata(state);
+  if (usageMetadata) metadata.usageMetadata = usageMetadata;
+
+  return metadata;
+}
+
+export function extractGeminiGenerateContentResponseMetadata(payload: unknown, requestPayload?: unknown): GeminiRecord {
+  return toTransformerMetadataRecord(extractGeminiGenerateContentTransformerMetadata(payload, requestPayload)) ?? {};
+}
+
+export const geminiGenerateContentResponseBridge = {
+  resolveBaseUrl: resolveGeminiNativeBaseUrl,
+  resolveModelsUrl: resolveGeminiModelsUrl,
+  resolveActionUrl: resolveGeminiGenerateContentUrl,
+  extractTransformerMetadata: extractGeminiGenerateContentTransformerMetadata,
+  extractResponseMetadata: extractGeminiGenerateContentResponseMetadata,
+  serializeAggregateResponse: serializeGeminiGenerateContentAggregateResponse,
+};
+
+export const geminiGenerateContentOutbound = geminiGenerateContentResponseBridge;
+
+export {
+  extractGeminiGenerateContentResponseMetadata as extractResponseMetadata,
+  extractGeminiGenerateContentTransformerMetadata as extractTransformerMetadata,
+  serializeGeminiGenerateContentAggregateResponse as serializeGeminiAggregateResponse,
+};

src/server/transformers/gemini/generate-content/stream.ts

@@ -1,253 +1,21 @@
-import {
-  applyGeminiGenerateContentAggregate,
-  createGeminiGenerateContentAggregateState,
-  type GeminiGenerateContentAggregateState,
-} from './aggregator.js';
-import { serializeGeminiAggregateResponse } from './outbound.js';
-
-type ParsedSsePayloads = {
-  events: unknown[];
-  lines: string[];
-  rest: string;
-};
-
-type GeminiGenerateContentStreamFormat = 'sse' | 'json';
-
-type ParsedGeminiStreamPayload = {
-  format: GeminiGenerateContentStreamFormat;
-  events: unknown[];
-  lines?: string[];
-  rest: string;
-};
-
-type AppliedGeminiStreamPayloads = ParsedGeminiStreamPayload & {
-  state: GeminiGenerateContentAggregateState;
-};
-
-function serializeSsePayload(payload: unknown): string {
-  return `data: ${JSON.stringify(payload)}\n\n`;
-}
-
-function parseSsePayloads(buffer: string): ParsedSsePayloads {
-  const events: unknown[] = [];
-  const lines: string[] = [];
-  let cursor = 0;
-
-  while (cursor < buffer.length) {
-    const boundaryMatch = /\r?\n\r?\n/.exec(buffer.slice(cursor));
-    if (!boundaryMatch || typeof boundaryMatch.index !== 'number') break;
-
-    const boundary = cursor + boundaryMatch.index;
-    const block = buffer.slice(cursor, boundary);
-    const rawBlock = buffer.slice(cursor, boundary + boundaryMatch[0].length);
-    cursor = boundary + boundaryMatch[0].length;
-
-    if (!block.trim()) continue;
-    lines.push(rawBlock);
-
-    const data = block
-      .replace(/\r\n/g, '\n')
-      .split('\n')
-      .filter((line) => line.startsWith('data:'))
-      .map((line) => line.slice(5).trimStart())
-      .join('\n')
-      .trim();
-
-    if (!data || data === '[DONE]') continue;
-
-    try {
-      events.push(JSON.parse(data));
-    } catch {
-      // Ignore malformed event payloads so aggregation remains tolerant.
-    }
-  }
-
-  return {
-    events,
-    lines,
-    rest: buffer.slice(cursor),
-  };
-}
-
-function isRecord(value: unknown): value is Record<string, unknown> {
-  return !!value && typeof value === 'object' && !Array.isArray(value);
-}
-
-function parseJsonArrayPayload(payload: unknown): unknown[] {
-  if (Array.isArray(payload)) return payload;
-  if (isRecord(payload)) return [payload];
-  return [];
-}
-
-function parseGeminiStreamPayload(
-  payload: unknown,
-  contentType?: string | null,
-): ParsedGeminiStreamPayload {
-  if (geminiGenerateContentStream.isSseContentType(contentType)) {
-    const parsed = parseSsePayloads(String(payload ?? ''));
-    return {
-      format: 'sse',
-      events: parsed.events,
-      lines: parsed.lines,
-      rest: parsed.rest,
-    };
-  }
-
-  if (typeof payload === 'string') {
-    const trimmed = payload.trim();
-    if (trimmed) {
-      try {
-        const parsedJson = JSON.parse(trimmed);
-        return {
-          format: 'json',
-          events: parseJsonArrayPayload(parsedJson),
-          rest: '',
-        };
-      } catch {
-        const parsed = parseSsePayloads(payload);
-        return {
-          format: 'sse',
-          events: parsed.events,
-          rest: parsed.rest,
-        };
-      }
-    }
-    return {
-      format: 'json',
-      events: [],
-      rest: '',
-    };
-  }
-
-  return {
-    format: 'json',
-    events: parseJsonArrayPayload(payload),
-    rest: '',
-  };
-}
-
-function applyParsedPayloadToAggregate(
-  state: GeminiGenerateContentAggregateState,
-  parsed: ParsedGeminiStreamPayload,
-): AppliedGeminiStreamPayloads {
-  for (const event of parsed.events) {
-    applyGeminiGenerateContentAggregate(state, event);
-  }
-
-  return {
-    ...parsed,
-    state,
-  };
-}
-
-function applyJsonPayloadToAggregate(
-  state: GeminiGenerateContentAggregateState,
-  payload: unknown,
-): GeminiGenerateContentAggregateState {
-  applyParsedPayloadToAggregate(state, parseGeminiStreamPayload(payload, 'application/json'));
-  return state;
-}
-
-function applySsePayloadsToAggregate(
-  state: GeminiGenerateContentAggregateState,
-  buffer: string,
-): AppliedGeminiStreamPayloads {
-  return applyParsedPayloadToAggregate(state, parseGeminiStreamPayload(buffer, 'text/event-stream'));
-}
-
-function consumeUpstreamSseBuffer(
-  state: GeminiGenerateContentAggregateState,
-  buffer: string,
-): AppliedGeminiStreamPayloads & { lines: string[] } {
-  const applied = applySsePayloadsToAggregate(state, buffer);
-  return {
-    ...applied,
-    lines: applied.lines ?? [],
-  };
-}
-
-function serializeAggregateJsonPayload(
-  payload: GeminiGenerateContentAggregateState | unknown,
-): unknown {
-  return serializeGeminiAggregateResponse(payload);
-}
-
-function serializeAggregateSsePayload(
-  payload: GeminiGenerateContentAggregateState | unknown,
-): string {
-  return serializeSsePayload(serializeAggregateJsonPayload(payload));
-}
-
-function serializeAggregatePayload(
-  payload: GeminiGenerateContentAggregateState | unknown,
-  format: GeminiGenerateContentStreamFormat = 'json',
-): unknown {
-  return format === 'sse'
-    ? serializeAggregateSsePayload(payload)
-    : serializeAggregateJsonPayload(payload);
-}
-
-function serializeUpstreamJsonPayload(
-  state: GeminiGenerateContentAggregateState,
-  payload: unknown,
-  streamAction = false,
-): unknown {
-  if (streamAction) {
-    const events = parseJsonArrayPayload(payload);
-    for (const event of events) {
-      applyGeminiGenerateContentAggregate(state, event);
-    }
-    return payload;
-  }
-
-  applyJsonPayloadToAggregate(state, payload);
-  return serializeAggregateJsonPayload(state);
-}
-
-export const geminiGenerateContentStream = {
-  isSseContentType(contentType: string | null | undefined): boolean {
-    return (contentType || '').toLowerCase().includes('text/event-stream');
-  },
-
-  parseJsonArrayPayload,
-  parseGeminiStreamPayload,
-  parseSsePayloads,
-  serializeSsePayload,
-  serializeAggregateJsonPayload,
-  serializeAggregatePayload,
-  serializeAggregateSsePayload,
-  serializeUpstreamJsonPayload,
-  applyParsedPayloadToAggregate,
-  applyJsonPayloadToAggregate,
-  applySsePayloadsToAggregate,
-  consumeUpstreamSseBuffer,
-
-  createAggregateState(): GeminiGenerateContentAggregateState {
-    return createGeminiGenerateContentAggregateState();
-  },
-
-  applyAggregate(state: GeminiGenerateContentAggregateState, payload: unknown): GeminiGenerateContentAggregateState {
-    return applyGeminiGenerateContentAggregate(state, payload);
-  },
-};
-
 export {
-  applyParsedPayloadToAggregate,
   applyJsonPayloadToAggregate,
+  applyParsedPayloadToAggregate,
   applySsePayloadsToAggregate,
+  consumeUpstreamSseBuffer,
+  geminiGenerateContentStream,
   parseGeminiStreamPayload,
   parseJsonArrayPayload,
   parseSsePayloads,
   serializeAggregateJsonPayload,
   serializeAggregatePayload,
   serializeAggregateSsePayload,
-  serializeUpstreamJsonPayload,
   serializeSsePayload,
-  consumeUpstreamSseBuffer,
-};
+  serializeUpstreamJsonPayload,
+} from './streamBridge.js';
 export type {
   AppliedGeminiStreamPayloads,
   GeminiGenerateContentStreamFormat,
   ParsedGeminiStreamPayload,
   ParsedSsePayloads,
-};
+} from './streamBridge.js';

src/server/transformers/gemini/generate-content/streamBridge.test.ts

@@ -0,0 +1,40 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  consumeUpstreamSseBuffer,
+  geminiGenerateContentStream,
+  parseGeminiStreamPayload,
+  serializeAggregateSsePayload,
+} from './streamBridge.js';
+
+describe('gemini generate-content stream bridge', () => {
+  it('parses SSE payloads and preserves trailing rest blocks', () => {
+    const parsed = parseGeminiStreamPayload([
+      'data: {"responseId":"resp-1","candidates":[{"content":{"parts":[{"text":"hello"}]}}]}',
+      '',
+      'data: [DONE]',
+      '',
+      'data: {"responseId":"partial"',
+    ].join('\n'), 'text/event-stream');
+
+    expect(parsed.format).toBe('sse');
+    expect(parsed.events).toHaveLength(1);
+    expect(parsed.rest).toBe('data: {"responseId":"partial"');
+  });
+
+  it('aggregates upstream SSE blocks while preserving raw lines and serializable final output', () => {
+    const state = geminiGenerateContentStream.createAggregateState();
+    const result = consumeUpstreamSseBuffer(state, [
+      'data: {"responseId":"resp-1","candidates":[{"content":{"parts":[{"text":"hello"}]},"finishReason":"STOP"}]}',
+      '',
+      'data: [DONE]',
+      '',
+      '',
+    ].join('\n'));
+
+    expect(result.lines).toHaveLength(2);
+    expect(result.events).toHaveLength(1);
+    expect(result.rest).toBe('');
+    expect(serializeAggregateSsePayload(result.state)).toContain('"responseId":"resp-1"');
+  });
+});