[Dileep Yavanmandha (dileepyavan)]

Fixes #344706

Problem

An HTTP Responses stream can end without emitting a terminal completion, for example after a reasoning item or with an unterminated completion frame. The final classifier correctly returns unknown with Response contained no choices., but previously obtained correlation only from retained completions. With none available, it discarded the actual HTTP request ID, making service-side investigation harder.

This fixes diagnostic correlation loss, not the underlying cause of an incomplete upstream stream.

Changes

• Retain HTTP request metadata on the internal successful-HTTP response independently of its completion iterable.
• Use the retained X-Request-ID, Copilot service request ID, and raw X-GitHub-Copilot-Request-Te in the no-choices fallback.
• Preserve completion-derived metadata when present and preserve all existing response classifications. Empty header IDs remain absent rather than being replaced with invented IDs.
• Add six regression/control cases that exercise the real Responses SSE parser and final fetcher classifier: empty body, reasoning followed by EOF, unterminated completion frame, valid completion, explicit stream error, and incomplete stream with no correlation headers.

No truncated frame is accepted as a completion, and no incomplete response is converted into success.

Validation

The three incomplete-stream correlation cases were verified to fail when HTTP metadata retention was temporarily removed; the completed-response and explicit-error controls still passed. The fix was restored before final validation.

From extensions\copilot:

$env:NODE_OPTIONS='--preserve-symlinks --preserve-symlinks-main'
pnpm run test:unit src\extension\prompt\node\test\chatMLFetcherResponseApiTelemetry.spec.ts src\platform\endpoint\node\test\responsesApi.spec.ts src\extension\prompt\node\test\chatMLFetcherRetry.spec.ts src\extension\prompt\node\test\chatMLFetcherTelemetry.spec.ts

Passed: 118 tests across four files. The module-resolution options are needed for this worktree's existing dependency junctions; without them, Vitest fails during initialization, including in the existing parser suite.

From the repository root:

pnpm run eslint extensions\copilot\src\platform\openai\node\fetch.ts extensions\copilot\src\extension\prompt\node\chatMLFetcher.ts extensions\copilot\src\extension\prompt\node\test\chatMLFetcherResponseApiTelemetry.spec.ts
git diff --check

Passed: targeted ESLint and whitespace checks.

From extensions\copilot:

pnpm run compile

Failed: the installed @typescript/native package does not export @typescript/native/ast or @typescript/native/async, producing 16 resolution errors in the unchanged TypeScript-context code. The existing shared dependency installation was not modified.

No standalone type-check, live provider request, integration suite, or smoke suite was run.

[Copilot]

🟢 Approval recommended

No active review issues were identified for the correlation-metadata retention change.

No open issues 🎉

What changed in this PR

Preserves HTTP correlation metadata when Copilot Responses streams end without a terminal completion, addressing #344706 without treating incomplete streams as successful.

Changes:

• Retains HTTP request metadata independently of streamed completions.
• Uses retained metadata in the no-choices fallback while preserving completion-derived values and existing classifications.
• Adds six regression and control cases covering incomplete streams, valid completion, explicit errors, and absent correlation headers.

File	Description
extensions/​copilot/​src/​platform/​openai/​node/​fetch.ts	Adds optional HTTP correlation metadata to ChatResults.
extensions/​copilot/​src/​extension/​prompt/​node/​test/​chatMLFetcherResponseApiTelemetry.spec.ts	Tests correlation retention through the real Responses parser and final classifier.
extensions/​copilot/​src/​extension/​prompt/​node/​chatMLFetcher.ts	Captures HTTP metadata and uses it when no completion supplies correlation.

_{Review effort: Balanced}