[diegosouzapw]

Summary

• The official OpenAI Chat Completions API rejects tool-call IDs longer than 64 characters (Invalid 'messages[N].tool_calls[0].id': string too long. Expected a string with maximum length 64). Clients such as Cursor replay longer IDs minted elsewhere, so multi-turn tool conversations routed to the openai provider fail with a 400.
• For the first-party openai provider only, every distinct overlong ID is now rewritten once per request to <sanitized 20-char prefix>_<sha256 base64url> (exactly 64 chars, [A-Za-z0-9_-]), keeping assistant.tool_calls[].id and tool.tool_call_id linked. IDs of 64 chars or fewer, and every other provider (openrouter, openai-compatible-*, azure-openai, nvidia), are untouched.

Attribution

Thanks to @mwpn for the original implementation.

Changes

• open-sse/executors/default/openaiToolCallIds.ts (new leaf): normalizeOpenAIToolCallIds() — request-scoped map, so repeated IDs and tool results that precede their assistant call map identically; full-ID SHA-256 digest, so IDs that share a long prefix never collide; deterministic across turns. Copy-on-write: the caller's body/messages (shared with logging, replay and combo fallback) are never mutated.
• open-sse/executors/default.ts: one-line wiring in DefaultExecutor.transformRequest() next to the existing NVIDIA tool-call-ID normalization (net line count unchanged, file stays at its frozen ceiling of 1213).
• The max_tokens → max_completion_tokens half of the original change already exists here and was not re-ported.
• tests/unit/openai-tool-call-id-normalization.test.ts: 10 cases (86-char ID linked on both sides, 64-char ID untouched, repeated IDs, result-before-call ordering, shared-prefix collision, unsafe-char prefix sanitizing, determinism, no mutation of the caller body, openrouter / openai-compatible untouched).

Test plan

• New test fails before the fix: node --import tsx/esm --test tests/unit/openai-tool-call-id-normalization.test.ts → ℹ tests 10 ℹ pass 4 ℹ fail 6
• Same test after the fix → ℹ tests 10 ℹ pass 10 ℹ fail 0
• NVIDIA regression: tests/unit/nvidia-tool-compatibility-2840.test.ts → ℹ tests 7 ℹ pass 7 ℹ fail 0
• All 121 unit test files that reference DefaultExecutor / executors/default (node --import tsx/esm --test --test-concurrency=4 …) → ℹ tests 1012 ℹ pass 1012 ℹ fail 0
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> → exit 0; npx prettier --check <changed files> → all formatted
• npm run typecheck:core → exit 0
• GITHUB_BASE_REF=release/v3.8.52 npm run check:test-masking → OK — 0 modificado(s) … sem enfraquecimento; check-pr-test-policy.mjs (PR context) → Result: PASS
• npm run check:file-size → default.ts within its ceiling; the only violation (open-sse/services/combo/executeTargetAttempt.ts 1288 > 1284) is pre-existing on the base and untouched here

⚠️ base-red inherited: #15306 — unit/vitest/integration/package-artifact gates on release/v3.8.52

[github-actions]

CI Coverage Report

• Coverage job: skipped
• PR test policy: success

Coverage artifact was not available for this run.