[nik13]

Summary

Follow up on the independent recall of #486. Keep Parallel Responses on the maintained shared OpenAI instrumentor; add no second patch or duplicate spans.

• Remove MCP authorization and headers from telemetry-only request and response tool definitions. Preserve the actual SDK wire payload, returned objects, and function schemas.
• Keep echoed instructions, metadata, input and other request-derived fields out of Responses output, including the streaming JSON fallback. Preserve answer content, status and usage; honor hidden error details.
• Redact known client/request credentials from status descriptions, attributes, exception messages and stack traces independently of FI_HIDE_ERROR_DETAILS. Re-raise the original SDK exception unchanged.
• Mark failed/incomplete Responses and failure/error terminal events as ERROR, not successful generations. EOF or close before a terminal event is also ERROR.
• Finish explicit sync/async stream closes and the high-level responses.stream() context exactly once. That context closes the underlying HTTP response rather than the raw SDK stream, so it needs a per-call response-close proxy as well.
• End request/stream spans on cancellation or other local aborts before re-raising.
• Mark interrupted Parallel Task workflows as ERROR, without attaching raw exception text; update both the example and its README recipe.

Reproduction and verification

On current dev 5ae2b9b4a3678af52abc01d90e4fe2fe1d4c2099, the original independent recall suite reproduced all seven findings: 20 failed, 2 passed at each SDK bound, before implementation changes.

The checked-in regressions cover sync/async, non-stream/stream, raw Responses objects, SDK high-level stream context exit, repeated/empty/terminal closes, request/stream cancellation, failure/incomplete/error terminal events, unchanged wire payloads and SDK exceptions, escaped credentials, input privacy flags, and hidden error details. Task interruption tests use the actual registered processor, not only a bare OTel provider.

All Python tests and builds ran inside the prescribed macOS sandbox denying external networking (loopback allowed). SDK requests use synthetic credentials and MockTransport; existing export/example tests use loopback OTLP receivers. No live vendor or Cloud calls.

Environment	Result
Python 3.10, openai 1.69.0, parallel-web 1.0.1, core 1.2.2, source	496 passed, 2 skipped
Python 3.13, openai 3.42.1, parallel-web 1.3.7, core 1.2.2, source	496 passed, 2 skipped
Installed wheels, floor SDKs	496 passed, 2 skipped
Installed sdists, latest SDKs	496 passed, 2 skipped

• Built both wheel and sdist distributions for both packages.
• Ruff passes for new source and regression files; git diff --check passes.
• python3 .github/release/release_plan.py --check: 129 packages, versions and coverage consistent.
• Update the existing Parallel CI matrix to openai 3.42.1, the latest resolved public release.

The key-echoing 401 is a synthetic regression, not a claim about an observed live Parallel response. No registry publication or version bump; release-please owns the patch release.

Related open work

Checked existing OpenAI/Parallel PRs before creating this follow-up. #437 handles Perplexity response extensions, not these findings. #411 adds explicit raw stream close() methods for audio; it does not cover the high-level Responses HTTP-close path, early-vs-terminal status, request/stream cancellation, or the privacy findings. This PR retains compatible close delegation and adds the Responses-specific behavior. Reconcile the shared stream methods when landing those PRs; do not drop the Responses lifecycle tests.

E2E coverage

Not applicable: no application e2e harness is changed. Package contracts, real SDKs, registered processor behavior, built distributions and loopback OTLP export are exercised.

Recall round 2 fixes

• Collect MCP authorization and arbitrary header values from effective extra_body.tools, honoring shallow replacement/removal without mutating the request or exporting extra-body metadata.
• Observe sync/async raw read(), text() and json() consumption. Successful non-SSE bodies no longer report a missing SSE terminal event. Failed/cancelled/incomplete JSON statuses remain ERROR, read exceptions retain their identity and finalize exactly once, and unread/SSE responses still require completion. Reads stay lazy; no implicit SDK parse() or post-parser invocation.
• New regression cases on the reviewed head: 31 failed, 42 passed at each SDK bound. All 73 now pass.
• Current standalone full OpenAI suite: 727 passed, 8 SDK-feature skips at OpenAI 1.69.0; 735 passed at latest 3.42.1. Python 3.12, OTel 1.29.0/1.45.2; sandboxed synthetic transports.
• Merged current origin/dev normally. fix(openai): mark interrupted agents and truncated chat streams as errors #536 now contains this exact branch and the resolved shared-request-wrapper conflict, with both fix sets tested together. No force push or release.

Generated with Claude Code