feat(sse): route GitHub Copilot Claude models through native /v1/messages

[diegosouzapw]

Summary

• GitHub Copilot's /chat/completions and /responses endpoints never surface prompt-cache token counts (cached_tokens) for Claude models, and round-tripping Claude tool_use/tool_result/thinking content blocks through the OpenAI shape is lossy.
• Claude models on the github (Copilot) provider now route through Copilot's Anthropic-native /v1/messages shim instead, reusing OmniRoute's existing per-model targetFormat mechanism (the same one already used by opencode/zen's Qwen entries and opencode/go) so chatCore translates the request to Claude shape before the executor ever sees it.
• The github executor's request-transform quirks that only make sense for /chat/completions (content-part flattening, trailing-assistant-prefill drop, the response_format-as-system-prompt workaround) are now skipped on the native path — the flattening in particular would otherwise destroy native tool_use/tool_result blocks.

Attribution

Thanks to @yidecode for the original implementation.

Changes

• open-sse/config/providers/registry/github/index.ts: add messagesUrl transport field + targetFormat: "claude" on every claude-* model entry.
• open-sse/config/providers/shared.ts / open-sse/config/providerRegistry.ts / open-sse/executors/base.ts: thread the new messagesUrl registry field through to ProviderConfig (mirrors how responsesBaseUrl is threaded).
• open-sse/executors/github.ts: buildUrl() routes claude-targeted models to messagesUrl; buildHeaders() adds the required anthropic-version header for those models; transformRequest() gates its /chat/completions-only transforms off the native path.
• tests/unit/provider-models-config.test.ts: updated the pre-existing registry-lineup assertion to reflect the new targetFormat.

Test plan

• new test tests/unit/github-copilot-claude-native-messages.test.ts — 10/10, fails before the change (asserted the old /chat/completions URL, missing header, flattened content, dropped prefill) and passes after
• node --import tsx/esm --test on the full set of github/registry-adjacent test files (108/108 pass, including the updated provider-models-config.test.ts and unmodified github-claude-reasoning-effort-granular.test.ts)
• npm run typecheck:core — clean
• npx eslint (with suppressions) on all changed files — clean

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[diegosouzapw]

Aprovado. Revert-proof isola bem o fix: sem os 5 arquivos de produção, 5/10 testes caem exatamente nos pontos que o PR descreve (tool_use virando text, prefill perdido). Fiação do parâmetro 'model' em buildHeaders confirmada — o call site genérico em base.ts já suportava o 4º argumento, então não há regressão pros outros executors. Pronto para merge.

[diegosouzapw]

Babysit summary — CI green ✅

Reds fixed (2 root causes, both real, both caused by this PR's intentional routing change):

1. Unit Tests fast-path (1/4 – 4/4) → 2bd77f37c
   Four pre-existing tests hard-coded registered claude-* Copilot ids (claude-sonnet-4.5/4.6) to exercise the /chat/completions legacy path. This PR tags those ids targetFormat: "claude", so they now route to the native /v1/messages shim, which intentionally skips the /chat/completions-only workarounds those tests assert (content-part flattening, prefill drop, response_format-as-system-prompt). Fixed by re-pointing three of them at an unregistered claude-sonnet-4 id (the pattern the sibling test at executor-github.test.ts:123 already used), and splitting copilot-gemini-claude-route-no-responses.test.ts into a Claude case (now expects /v1/messages) and a Gemini case (still /chat/completions). No assertion was weakened or removed — the /chat/completions invariants are still asserted, just on a model that still takes that path. The native path's coverage lives in github-copilot-claude-native-messages.test.ts.

2. Fast Quality Gates (complexity + cognitive ratchets) → 2bd77f37c
   Real regression from this PR: 2058 > 2056 and 891 > 890. transformRequest hit complexity 18 / cognitive 16 and buildHeaders hit 17 after the isClaudeNative branches were added. Extracted, not rebaselined — applyChatCompletionsOnlyQuirks() and resolveInitiatorHeader(). Both ratchets now measure exactly at baseline (2056 / 890). Behavior identical; verified by the 57 tests above.

Also: merged origin/release/v3.8.49 (4524bbe3e) — the branch was 45 commits behind, clean merge, no conflicts.

Policy check — no collision. #7223 gates its routing on targetFormat === "claude" for the GitHub Copilot provider; the open GPT-5.x nest (#7012/#7101/#7242 → /v1/responses, #7095/#7176 reasoning summaries) targets OpenAI/Codex model families and never reaches the claude branch. #7012/#7242 do touch providers/shared.ts, but in the GPT_5_6_* capability constants (~L285-300); #7223 only adds the messagesUrl?: string interface field (~L107/L178). Disjoint regions, disjoint semantics.

Gate: all green — Fast Quality Gates, 4/4 unit shards, Vitest, ESLint, Docs, Merge integrity, semgrep, dast-smoke.
Tests: 4 updated (re-pointed to the legacy-path model, none weakened), 57 passing across the 6 affected files.

Ready for human review & merge — not merging.