fix(github): route Claude models to Copilot's native format via chatCore

[yidecode]

Follow-up to #2608 (same author). That PR sent Claude models to Copilot's Anthropic-native /v1/messages shim by branching inside the executor: chatCore still resolved github's target format to "openai", and executeWithMessagesEndpoint translated openai → claude on the way out and claude → openai on the way back.

That made targetFormat a lie, and Claude-format clients paid for it.

The bug

A request from Claude Code / Zed (directly, or through a downstream Anthropic proxy) went claude → openai → claude. translator/request/claude-to-openai.js has no case for thinking blocks — its switch handles text, image, tool_use and tool_result only. The assistant turn's thinking block and its signature were therefore dropped on the way in, so on every follow-up turn the model lost its own prior reasoning.

It failed silently: Copilot accepts a history with no thinking block (verified), it only rejects a thinking block whose signature is missing. Same body, two routes:

claude → claude   assistant blocks: ['thinking', 'tool_use']   signature: preserved verbatim
claude → openai   assistant keys:   ['role', 'tool_calls']     thinking: gone

The fix

The provider now declares the routing decision where chatCore can see it:

transport.resolveTargetFormat(model) -> "claude" | null

Resolved at request time from the model name, because Copilot's live catalog ships claude-* variants long before the static models[] list catches up — it currently serves claude-sonnet-5, claude-opus-4.8, claude-opus-4.8-fast and claude-fable-5, none of them listed. A static per-entry targetFormat would miss exactly those, which is why #2608 avoided one; a runtime hook does not have that problem.

Providers without the hook are unaffected — the helper returns null and the existing precedence chain is untouched.

With the target format honest:

• A claude-source request short-circuits in translateRequest (sourceFormat === targetFormat) and reaches the shim untouched — thinking blocks and real signatures survive verbatim.
• prepareClaudeRequest still runs — it is gated on targetFormat === CLAUDE, outside the short-circuit — so feat(github): route Claude models through Copilot's native /v1/messages #2608's cache_control injection and its prompt-cache token counts are preserved. There is a test pinning this.
• OpenAI-format clients still translate once, now in chatCore rather than the executor.

executeWithMessagesEndpoint goes away entirely (-90 lines), along with its hand-rolled SSE transform and its duplicate _toolNameMap threading: the generic streaming handler already does both, keyed off the same target format. Net -72 lines.

Verification

• Live end-to-end against Copilot: the exact body 9router now produces for a claude-source multi-turn request — thinking block with a real 400-char signature, tool_use, tool_result — returns 200.
• Deployed smoke test through the running gateway: multi-turn thinking replay, single-turn thinking text, and OpenAI-format clients all pass.
• Full suite: failure set identical to the master baseline, zero regression.
• 13 new tests, including a reverse guard that explicitly asserts the claude → openai hop drops thinking blocks, so a future refactor cannot quietly route back through it.

Scope

Deliberately does not touch request/claude-to-openai.js or openai-to-claude.js — patching those would affect every provider that pivots through OpenAI, for a benefit this routing fix already delivers. Note that tests/translator/bugs-toClaude-context.test.js already has a red test (assistant reasoning_content becomes a thinking block) documenting that separate gap.

OpenAI-format clients still cannot round-trip thinking — that format has nowhere to carry a signature. Inherent, not addressed here.

本 PR 由 Claude Code 辅助完成