[nik13]

Summary

Follow up on merged #1069. The normal TrueFoundry streaming application now refuses to print an incomplete answer as a successful result.

• Require the requested choice's finish_reason="stop"; EOF, [DONE], empty/usage-only endings, token limits and content filtering do not count as a complete answer.
• Use the SDK stream context manager so success, iteration failures and local aborts close the stream.
• Preserve completed answers and trailing usage chunks. main() returns 1 with no partial answer on stdout when completion validation fails.
• Keep tracing in the maintained package. No copied instrumentation, additional spans or public QA diary.

Verification

• New application regressions before the change: 8 failed, 1 passed, including the completed-answer control.
• Entire TrueFoundry recipe suite: 64 passed with the exact existing released requirements (OpenAI 3.42.0, traceai-openai 0.1.11, fi-instrumentation-otel 1.2.2).
• Same 64 tests pass on OpenAI 1.69.0 and current latest 3.42.1 with the maintained integration fixes.
• Independent exact main() truncation/completion probes pass at both SDK bounds.
• Docs integration tests: 20 passed. Link audit: zero broken navigation/content links. Full Astro/Pagefind build passes.
• Python tests and docs build ran with outbound networking denied and synthetic credentials only. No live gateway or Cloud verification is claimed. No rendered-page changes.

Release dependency

This fixes application behavior without waiting for a package release. The separate span-status/cancellation/privacy fixes remain in future-agi/traceAI#520 and future-agi/traceAI#536. The public registry still reports traceai-openai 0.1.12 and fi-instrumentation-otel 1.2.2, without those fixes. Existing recipe pins are intentionally unchanged: update them to genuine fixed releases when available, not invented versions. This PR is not a claim that the entire released telemetry recall is complete.

E2E coverage

The real SDK application entry point, mock HTTP transport, exit status/stdout, stream closure and existing loopback OTLP recipe tests are exercised; no application UI or e2e harness is changed.

Generated with Claude Code