fix(api): surface adaptive thinking display and thinking_tokens for Anthropic models - #1327

No actionable comments were generated in the recent review. 🎉

sequenceDiagram
  participant AnthropicAPI
  participant AnthropicProvider
  participant AnthropicVertexProvider
  participant ApiStreamChunk
  AnthropicAPI->>AnthropicProvider: message_start and message_delta usage
  AnthropicAPI->>AnthropicVertexProvider: streaming usage events
  AnthropicProvider->>ApiStreamChunk: reasoningTokens from thinking_tokens
  AnthropicVertexProvider->>ApiStreamChunk: reasoningTokens from thinking_tokens

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

_{Comment @coderabbitai help to get the list of available commands.}

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

CI trace & coverage — PR #1327 (final)

CI trace (all checks on final commit b2e6883):

Failures encountered (chronological):

1. platform-unit-test (ubuntu-latest + windows-latest), run 32490112749 (commit a02b0da) — 3 assertion failures in src/api/providers/__tests__/requesty.spec.ts (uses adaptive thinking for Claude Fable 5 / Sonnet 5 / Opus 5): received thinking: { display: "summarized", type: "adaptive" }, expected the old exact shape { type: "adaptive" }. Root cause: our diff — the shared getAnthropicProviderReasoning transform changed and the Requesty handler consumes it, but those 3 exact-match assertions were not updated. Windows job: cancelled after the ubuntu job failed in the same run.
2. codecov/patch (commit a02b0da) — patch-coverage gate failed: the newly added thinking_tokens handling lines in anthropic.ts/anthropic-vertex.ts (and the new display branch in reasoning.ts) were only partially covered by the existing test suite — specifically the numeric message_start thinking_tokens path had no coverage at all.

Actions taken:

• Commit 0b5b245ee — test(requesty): update adaptive thinking assertions to include display summarized: 3 stale assertions in requesty.spec.ts updated to the new shape (test-only, mirrors the assertions the PR already updated in the anthropic/vertex suites).
• Commit b2e6883d4 — test(anthropic): add thinking_tokens message_start coverage per CodeRabbit — addresses the CodeRabbit nitpick (review 4994124213) "Add missing thinking_tokens stream coverage":
  • anthropic.spec.ts: new case where message_start usage carries numeric output_tokens_details.thinking_tokens; asserts the message_start usage chunk surfaces reasoningTokens: 5, and that a message_delta without output_tokens_details omits it (existing numeric message_delta + unset message_start case retained).
  • anthropic-vertex.spec.ts: two focused cases — numeric message_start/message_delta thinking_tokens (asserts reasoningTokens emitted as 5 / 150) and unset output_tokens_details (asserts reasoningTokens omitted).
  • Both commits verified locally before push: vitest 189/189 across the 4 affected suites, eslint --prune-suppressions --max-warnings=0 clean (no suppression-count changes), tsc --noEmit clean; repo pre-commit hooks (prettier + full lint + tsc) green.

Coverage result (100% patch coverage):
Verified with vitest run <4 suites> --coverage (v8 + lcov reporter) on the committed tree, then matched every added line of the PR diff (vs base 871bb98) against lcov.info:

• src/api/transform/reasoning.ts: 14 patch lines — 13 are comments/type declarations (non-executable, not instrumented), the single executable patch line (the return { type: "adaptive", display: "summarized" } branch) has 21 hits.
• src/api/providers/anthropic.ts: 14 patch lines — executable lines L262 (28 hits) and L285 (2 hits) both covered; the conditional-spread lines are part of the enclosing yield statements (L264/L273 = 28 hits, L287/L294 = 2 hits), and both arms of each ternary are exercised (numeric thinking_tokens by the new tests, unset by ~26 pre-existing tests).
• src/api/providers/anthropic-vertex.ts: 9 patch lines — L143 4 hits; enclosing yield statements L129/L140 = 19 hits and L145/L152 = 4 hits, both ternary arms exercised by the two new tests.
• Test files (reasoning.spec.ts, anthropic.spec.ts, anthropic-vertex.spec.ts, requesty.spec.ts): every added line is executed by the 189-test run (vitest reports the suites as passed; test files are not separately instrumented in the lcov reporter output).

Result: every executable/patch line of all 6 changed files is covered; no zero-hit executable patch line remains. (The codecov/patch CI gate now passes on the final commit.)

Follow-up for reviewers:

• No re-runs were needed at any point — every failure was a genuine miss caused by this PR's diff and was fixed by the two additive test-only commits above.
• CodeRabbit round 1 is addressed (nitpick fixed in b2e6883d4); I polled the review endpoint every 2 min for ~50 min after the push and no new review round appeared yet — if CodeRabbit posts a fresh round, any remaining actionable comments will be addressed in a follow-up.
• Out of scope / not touched: src/eslint-suppressions.json is stale relative to main (a local --prune-suppressions run prunes ~1,700 unrelated entries); deliberately left alone in this PR.

CI trace by agent — easonLiangWorldedtech