port(upstream): #2608 - Route Claude models through GitHub Copilot native /v1/messages by bloodf · Pull Request #291 · bloodf/durindoor

bloodf · 2026-07-16T03:27:03Z

Summary

What changed: Claude models on GitHub Copilot are routed to Copilot's Anthropic-native /v1/messages shim; message_stop usage fallback folds cache_read/cache_creation into prompt_tokens. Hardened the route per Codex review: strips params upstream rejects, maps max_completion_tokens/stop/tool_choice:"none" to Anthropic shapes, bounds the fetch with the provider connect timeout, retries transient 502/503/504 + network failures, drops the Claude Code persona for non-Claude-Code Copilot clients, and strips unsigned thinking history.
Why it changed: Port of upstream decolua/9router #2608. /v1/messages is the only Copilot endpoint that surfaces prompt-cache token counts (cached_tokens); /chat/completions never gets cache_control. Non-Claude models (gpt/gemini/grok) keep existing /chat/completions (or /responses) routes untouched.

Type

feat: — new feature (port)

Breaking changes

No

Testing

What was tested: claude→/v1/messages routing + forced upstream SSE + cache-usage surfacing; failure (never [DONE]) on truncated/malformed stream; connect-timeout aborts with no in-place retry; 503 transient retries exhausted (initial+3) and 503→200 recovery; persona injected for claude-cli but absent for a generic Copilot client; unsigned thinking stripped; tool_choice:"none"→{type:"none"} contract (promoted from known-bug it.fails).
Commands run: npm run test:ci (full fail-closed gate); targeted vitest unit/github-claude-messages.test.js + unit/openai-to-claude.test.js.
Evidence: targeted suite 31/31 PASS (25 prior + 6 new); full npm run test:ci Raw failures: 0 under pinned Node 20.20.2; CI 3/3 GREEN on head e3892be0 — ✨ Lint & Build, 🧪 Vitest + No-regression gate, commitlint. git diff --check origin/dev...HEAD clean.

Checklist

Tests added or updated for behavioral changes (tests/unit/github-claude-messages.test.js +6 tests; tests/translator/bugs-toClaude-context.test.js known-bug promoted)
No new dependencies without explicit justification (none added)
Conventional commit used (port(upstream): / fix(port): / test(port):)
CHANGELOG updated if this change is user-facing (Unreleased → Features + Fixes bullets)
No secrets, keys, credentials, or tokens committed

Upstream: decolua/9router#2608

Agent disclosure

Model: durindoor/kimi/kimi-k2.7
Harness: Oh My Pi (omp) coding harness — Claude-Code-style agent
Harness version: unknown (not surfaced to agent)
Plugins installed: none visible to this agent beyond the shared skill library

This contribution is agent-generated as part of the orchestrated upstream-port cycle and was produced without interactive human review of the diff in this session.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7d1dc8bf40

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 25d7517de4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…tion_tokens/stop/tool_choice on Claude /v1/messages Codex review on PR #291 (1 P1 + 3 P2): the new native /v1/messages path translated the OpenAI body without the stripUnsupportedParams pass the /chat/completions path applies, and the generic OpenAI->Claude translator dropped max_completion_tokens, stop, and tool_choice "none". - github executor: stripUnsupportedParams("github", model, {...body}) before translateRequest so temperature/thinking/reasoning_effort no longer 400. - openai-to-claude: honor max_completion_tokens when max_tokens absent (prefer explicit max_tokens); map stop string|string[] to stop_sequences (guard stop:null); map tool_choice "none" to Anthropic {type:"none"}. Tests: route assertions in github-claude-messages.test.js; shared translator cases in openai-to-claude.test.js (30 passed across claude suites).

…p/tool_choice on /v1/messages Codex review on PR #291 (1 P1 + 3 P2): the new native /v1/messages path translated the OpenAI body without the stripUnsupportedParams pass the /chat/completions path applies, and the generic OpenAI->Claude translator dropped max_completion_tokens, stop, and tool_choice "none". - github executor: stripUnsupportedParams("github", model, {...body}) before translateRequest so temperature/thinking/reasoning_effort no longer 400. - openai-to-claude: honor max_completion_tokens when max_tokens absent (prefer explicit max_tokens); map stop string|string[] to stop_sequences (guard stop:null); map tool_choice "none" to Anthropic {type:"none"}. Tests: route assertions in github-claude-messages.test.js; shared translator cases in openai-to-claude.test.js (30 passed across claude suites).

bloodf · 2026-07-16T06:53:47Z

All 8 Codex findings addressed on e3892be0:

Original 4 (already in code, this push confirms):

stripUnsupportedParams("github", body) — github.js:221 → drops temperature/thinking/reasoning_effort upstream 400s
max_completion_tokens → max_tokens fallback — openai-to-claude.js:23-33
stop (string|array) → stop_sequences, stop:null guarded — openai-to-claude.js:35
tool_choice:"none" → {type:"none"} — openai-to-claude.js:332 (promoted it.fails known-bug in bugs-toClaude-context.test.js:38 to a passing contract)

New 4 (added this push):
5. Connect-timeout bound — per-attempt AbortController + FETCH_CONNECT_TIMEOUT_MS around proxyAwareFetch (github.js:235)
6. Persona guard — detectClientTool(requestContext.clientHeaders) !== "claude" drops the CLAUDE_SYSTEM_PROMPT block for non-Claude-Code Copilot clients (github.js:269)
7. Unsigned-thinking strip — normalizeClaudePassthrough(..., "claude") validating path (github.js:300)
8. Transient retry loop — 502/503/504 + network-as-502 with per-status counters and waitForRetryDelay (github.js:315)

Defended by 6 new tests in tests/unit/github-claude-messages.test.js (timeout aborts with 1 dispatch; 503 exhausted after initial+3 retries; 503→200 transient recovery; persona absent for non-Claude-Code client; persona kept for claude-cli; unsigned thinking stripped). Targeted suite 31/31 green; full npm run test:ci gate Raw failures 0; CI 3/3 green.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e3892be073

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…p/tool_choice on /v1/messages Codex review on PR #291 (1 P1 + 3 P2): the new native /v1/messages path translated the OpenAI body without the stripUnsupportedParams pass the /chat/completions path applies, and the generic OpenAI->Claude translator dropped max_completion_tokens, stop, and tool_choice "none". - github executor: stripUnsupportedParams("github", model, {...body}) before translateRequest so temperature/thinking/reasoning_effort no longer 400. - openai-to-claude: honor max_completion_tokens when max_tokens absent (prefer explicit max_tokens); map stop string|string[] to stop_sequences (guard stop:null); map tool_choice "none" to Anthropic {type:"none"}. Tests: route assertions in github-claude-messages.test.js; shared translator cases in openai-to-claude.test.js (30 passed across claude suites).

…ry parity (Codex #291)

…ssages

…tive /v1/messages

…p/tool_choice on /v1/messages Codex review on PR #291 (1 P1 + 3 P2): the new native /v1/messages path translated the OpenAI body without the stripUnsupportedParams pass the /chat/completions path applies, and the generic OpenAI->Claude translator dropped max_completion_tokens, stop, and tool_choice "none". - github executor: stripUnsupportedParams("github", model, {...body}) before translateRequest so temperature/thinking/reasoning_effort no longer 400. - openai-to-claude: honor max_completion_tokens when max_tokens absent (prefer explicit max_tokens); map stop string|string[] to stop_sequences (guard stop:null); map tool_choice "none" to Anthropic {type:"none"}. Tests: route assertions in github-claude-messages.test.js; shared translator cases in openai-to-claude.test.js (30 passed across claude suites).

… + thinking guards

…to passing contract

…ry parity (Codex #291)

…ssages

…ndled rejection

bloodf force-pushed the port/upstream-2608 branch from 7d1dc8b to 5ccf065 Compare July 16, 2026 03:35

chatgpt-codex-connector Bot reviewed Jul 16, 2026

View reviewed changes

Comment thread open-sse/executors/github.js Outdated

Comment thread open-sse/executors/github.js Outdated

Comment thread open-sse/executors/github.js Outdated

Comment thread open-sse/executors/github.js Outdated

bloodf force-pushed the port/upstream-2608 branch from 5ccf065 to 25d7517 Compare July 16, 2026 04:33

chatgpt-codex-connector Bot reviewed Jul 16, 2026

View reviewed changes

Comment thread open-sse/executors/github.js Outdated

Comment thread open-sse/executors/github.js Outdated

Comment thread open-sse/executors/github.js Outdated

Comment thread open-sse/executors/github.js Outdated

bloodf force-pushed the port/upstream-2608 branch from 25d7517 to 3624109 Compare July 16, 2026 05:17

bloodf force-pushed the port/upstream-2608 branch from 3624109 to e6a4ff6 Compare July 16, 2026 05:18

bloodf force-pushed the port/upstream-2608 branch from e6a4ff6 to 1c59908 Compare July 16, 2026 05:23

chatgpt-codex-connector Bot reviewed Jul 16, 2026

View reviewed changes

Comment thread open-sse/executors/github.js Outdated

Comment thread open-sse/executors/github.js

Comment thread open-sse/executors/github.js

Comment thread open-sse/executors/github.js

bloodf pushed a commit that referenced this pull request Jul 16, 2026

fix(executors): github /v1/messages tool-choice, empty turns, and ret…

030716d

…ry parity (Codex #291)

bloodf pushed a commit that referenced this pull request Jul 16, 2026

test(port): #2608 - add Codex #291 regression tests for github /v1/me…

6e167a9

…ssages

bloodf force-pushed the port/upstream-2608 branch from e3892be to 362d896 Compare July 16, 2026 19:20

CortexOS and others added 7 commits July 16, 2026 22:02

port(upstream): #2608 - Route Claude models through GitHub Copilot na…

b46d23e

…tive /v1/messages

fix(port): #2608 - bound /v1/messages fetch; retry transient; persona…

c088651

… + thinking guards

test(port): #2608 - promote tool_choice=none from known-bug it.fails …

9a41528

…to passing contract

fix(executors): github /v1/messages tool-choice, empty turns, and ret…

0e9b65a

…ry parity (Codex #291)

test(port): #2608 - add Codex #291 regression tests for github /v1/me…

da961b7

…ssages

test(port): #2608 - fix github-claude-messages fixture shape and unha…

938336e

…ndled rejection

bloodf force-pushed the port/upstream-2608 branch from 362d896 to 938336e Compare July 16, 2026 22:02

bloodf merged commit cb1939c into dev Jul 16, 2026
3 checks passed

bloodf deleted the port/upstream-2608 branch July 17, 2026 02:46

bloodf mentioned this pull request Jul 20, 2026

merge: DurinDoor 2.2.0 release #360

Merged

7 of 8 tasks

bloodf mentioned this pull request Aug 27, 2026

port(upstream): #3506 - harden Claude trailing prefill and tool types #623

Merged

7 of 13 tasks

bloodf mentioned this pull request Oct 8, 2026

Source inventory: OmniRoute PRs, issues and port candidates #1073

Open

4 tasks

Navigation Menu

Repository navigation

port(upstream): #2608 - Route Claude models through GitHub Copilot native /v1/messages - #291

port(upstream): #2608 - Route Claude models through GitHub Copilot native /v1/messages#291
bloodf merged 7 commits into
devfrom
port/upstream-2608

bloodf commented Jul 16, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

bloodf commented Jul 16, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

bloodf commented Jul 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Type

Breaking changes

Testing

Checklist

Agent disclosure

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

bloodf commented Jul 16, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

bloodf commented Jul 16, 2026 •

edited

Loading