Capability-aware reasoning and extended thinking translation

[lyzgeorge]

Summary

• Adds a capability-gated reasoning/thinking translation layer so both the OpenAI (/v1/chat/completions) and Anthropic (/v1/messages) surfaces route reasoning_effort, thinking_budget, and Anthropic thinking config to whichever upstream knob each Copilot model advertises under capabilities.supports.
• Propagates reasoning_text / reasoning_opaque through non-stream responses and emits native content_block_start / thinking_delta / signature_delta / content_block_stop events on the streaming path, so Claude Code and similar clients see real thinking UIs.
• Adds a new src/routes/reasoning-context.ts helper shared by both handlers, plus handler-level tests for /v1/chat/completions and /v1/messages that verify adaptive-thinking, reasoning-effort-only, disabled-thinking, and unsupported-model paths.

Why

Copilot advertises reasoning capabilities under capabilities.supports.adaptive_thinking and capabilities.supports.reasoning_effort, but neither surface read those fields. As a result, thinking: enabled on /v1/messages was silently dropped, and /v1/chat/completions would happily forward thinking_budget to models that don't accept it. The new helper reads the advertised capabilities and translates requests accordingly — e.g. Anthropic thinking: enabled becomes reasoning_effort: high on gpt-5 family models and reasoning_effort: high + thinking_budget on Claude Sonnet 4.6.

What changed

• src/routes/reasoning-context.ts (new): buildAnthropicReasoningContext and buildOpenAIReasoningContext, gated on supports.reasoning_effort and supports.adaptive_thinking.
• src/routes/messages/handler.ts: resolves the selected model, builds the reasoning context, passes it into translateToOpenAI, and debug-logs when an unsupported model causes thinking to be stripped.
• src/routes/messages/non-stream-translation.ts: translateToOpenAI now takes a ReasoningContext and emits reasoning_effort / thinking_budget; translateToAnthropic preserves reasoning_text as a thinking block and reasoning_opaque on the response, maintaining per-choice ordering.
• src/routes/messages/stream-translation.ts: tracks currentBlockType explicitly and emits Anthropic-native thinking/text/tool-use transitions, including signature_delta scoped to each thinking block. Forwards cached prompt tokens as cache_read_input_tokens on message_delta.
• src/routes/chat-completions/handler.ts: builds an OpenAI reasoning context, drops unsupported thinking_budget with a debug log, and normalizes the payload before sending to Copilot.
• src/routes/messages/count-tokens-handler.ts: reuses the same reasoning-context helper so token counts match what we actually send upstream.
• src/services/copilot/create-chat-completions.ts: types reasoning_effort, thinking_budget, stream_options, and reasoning fields on Delta / ResponseMessage.
• src/services/copilot/get-models.ts: extends ModelSupports with adaptive_thinking and reasoning_effort.
• Tests: new tests/chat-completions-handler.test.ts and tests/messages-handler.test.ts (integration-style using fetch mocking so no module-level mocks leak across suites), plus reasoning-context helper coverage in tests/anthropic-request.test.ts and streaming/non-stream reasoning cases in tests/anthropic-response.test.ts.
• README: new hook in the intro, a bullet in Features, and a dedicated Reasoning & Extended Thinking section with curl examples for both surfaces.

Test plan

• bun test tests/ — 48 tests, all passing
• bun run typecheck
• bun run lint
• Live /v1/chat/completions against gpt-5-mini and claude-sonnet-4.6 with reasoning_effort in minimal / low / medium / high
• Live /v1/messages against gpt-5-mini and claude-sonnet-4.6 with thinking: enabled and thinking: disabled
• Verified reasoning_text / reasoning_opaque flow through non-stream responses
• Verified thinking-block stream events for Claude Sonnet 4.6

Notes

• The fix for capability reads (capabilities.supports.*) is split out as its own commit (3a65946) so the history shows the root cause clearly.
• .worktrees/ was added to .gitignore in a separate chore commit so local worktree directories don't pollute git status.