[diegosouzapw]

Summary

• When a client disconnects mid-stream (499 client_disconnected) or the upstream fails mid-answer, the request is now recorded with the tokens the provider had already generated, instead of zero.
• Provider-reported usage wins; a side not reported yet (e.g. Claude's output count, sent only in the final message_delta) is estimated from the streamed output and flagged estimated. If nothing was reported and nothing was streamed, the zero-token record stays as it was.
• Status 499 / success:false, the fix(backend): Completed SSE streams recorded as 499 client-disconnected with 0 tokens (finalization race) #9653 disconnect grace window and the completion-wins guard stay as they are. Only the token columns of the failure records change.

Attribution

Thanks to @brimob-sowax, @ntdat812 and @ntdatt812 for the original implementations (three PRs merged into one port here).

Changes

• open-sse/utils/streamPartialUsage.ts (new): WeakMap side table registerStreamPartialUsage / readStreamPartialUsage (same pattern as registerStreamTiming), plus the pure resolvePartialStreamUsage({ reported, contentLength, body }).
• open-sse/utils/stream.ts: createSSEStream registers a lazy reader that closes over the transform's live state.usage (translate) / usage (passthrough) and totalContentLength. It does no per-chunk work.
• open-sse/utils/streamFailureFinalization.ts: createStreamFailureFinalizers takes an optional readPartialUsage (errors are swallowed and fall back to zero). When partial usage exists, it goes to persistFailureUsage(status, code, aggregate) and onto the completion payload as partialUsage. When there is none, the calls are byte-identical to before.
• open-sse/handlers/chatCore/streamingTail.ts: wires readPartialUsage: () => readStreamPartialUsage(transformStream). The call log takes streamUsage || partialUsage for its token columns, and estimates are marked through the existing isEstimatedUsage path.

Design notes:

• usage stays null on the failure completion payload on purpose. A non-null usage there would make recordStreamingUsageStats write a second usage row next to persistFailureUsage, and would also run cost and input-calibration (calibration against an estimate) for a failed stream.
• Partial tokens are not added to API-key USD spend or token budgets. Those aggregations count only successful rows (success = 1 in apiKeyUsageLimits.ts, and recordTokenUsage runs only on 200). Counting aborted streams against budgets is a separate policy change.
• usage_history has no "estimated" column, so an estimated count shows up there as a plain number. The call log keeps the estimated provenance.

Test plan

• tests/unit/stream-aborted-partial-usage.test.ts (15 tests): finalizer carries reported usage on 499 and on a mid-stream 502; no usage keeps persistFailureUsage(status, code) with exactly 2 args; throwing reader still finalizes; completion-wins guard reads/writes nothing; resolver cases (verbatim, estimate, null, Claude input-only + estimated output, input/output_tokens shape); real passthrough and translate (OpenAI → Claude) streams pumped without flush; end-to-end disconnect → 499 row with non-zero tokens.
  • Against the base code (git show HEAD: of stream.ts + streamFailureFinalization.ts): ℹ tests 15 ℹ pass 9 ℹ fail 6, with the 6 failing tests being the ones that check carried partial usage.
  • With the fix: ℹ tests 15 ℹ pass 15 ℹ fail 0 (also exits cleanly without --test-force-exit).
• Sibling sweep (75 test files that reference streamFailureFinalization / createStreamFailureFinalizers / streamingTail / persistFailureUsage / the SSE stream factories): ℹ tests 563 ℹ pass 556 ℹ fail 2. Neither failure comes from this change: integration-wiring "keeps usage quota wired through A2A and MCP tools" (an MCP server source regex) is inherited, and a fixture file was picked up by the glob.
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files>: clean. npx prettier --check <changed files>: clean.
• npm run typecheck:core: exit 0.
• GITHUB_BASE_REF=release/v3.8.52 node scripts/check/check-pr-test-policy.mjs: PASS. npm run check:test-masking: OK (0 modified / 0 deleted).

⚠️ 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.