|
| 1 | +# 040 — wp4: PR #2433, combo zero-output failover |
| 2 | + |
| 3 | +## Item |
| 4 | + |
| 5 | +`Ingwannu` PR #2433 `ingw/fix-combo-zero-output-failover-2431` -> `dev`, head |
| 6 | +`3ec2b1a6`. Addresses issue #2431, "failover combos stop on zero-output SSE |
| 7 | +terminal failures and model-EOL HTTP 410 responses". |
| 8 | + |
| 9 | +Adds a bounded SSE preflight so failover combos can retry terminal failures |
| 10 | +before any output is committed, plus model-lifecycle HTTP 410 classification and |
| 11 | +synchronized documentation. 12 files. |
| 12 | + |
| 13 | +## Reviewer verdict: FAIL |
| 14 | + |
| 15 | +`gpt-5.6-sol` high, read-only, exact head `3ec2b1a6`: |
| 16 | + |
| 17 | +- VERDICT FAIL, disposition NEEDS_CHANGES, risk medium. |
| 18 | +- `bun test tests/combos.test.ts tests/combo-stream-preflight.test.ts tests/server-combo-failover-e2e.test.ts tests/core-lab-boundary.test.ts` -> 126 passed, 0 failed, 694 assertions. |
| 19 | +- `gh pr checks 2433` -> 23 passed, 1 skipped, 0 failed. |
| 20 | + |
| 21 | +Green tests and green CI, and still not landable. That gap is the point of the |
| 22 | +finding: the defect is in accounting the existing tests do not assert. |
| 23 | + |
| 24 | +## The blocker |
| 25 | + |
| 26 | +`src/server/responses/core.ts:1974` — the new preflight manually records a failed |
| 27 | +terminal. Native forward and pool passthrough streams already record that same |
| 28 | +physical terminal through their eager and tee inspectors at `:3785` and `:3868`. |
| 29 | +The once-guard added at `:1930` wraps only the exported callback, so it never |
| 30 | +observes the inspector's direct invocation. |
| 31 | + |
| 32 | +One physical 502 therefore increments native account health twice. A production |
| 33 | +recorder diagnostic confirmed `consecutiveFailures: 2` for a single terminal. |
| 34 | +The practical effect is that soft-avoid and credential rotation fire at half the |
| 35 | +configured threshold, on an account that is healthy. |
| 36 | + |
| 37 | +Everything else checked clean: stream commit boundary, attempt receipts, usage |
| 38 | +handling, marker preservation, export surface, ESM/Bun constraints, no core-to-Lab |
| 39 | +edge, no import cycle, no sensitive logging, translated docs aligned, template |
| 40 | +complete. |
| 41 | + |
| 42 | +## Disposition |
| 43 | + |
| 44 | +Hold. Not closed, not merged. The required change is a single shared |
| 45 | +once-guarded recorder owning both the preflight path and the inspector path, with |
| 46 | +a regression asserting exactly one health transition per streamed attempt. |
| 47 | + |
| 48 | +That fix is being built on `codex/fix-2433-exactly-once-terminal` rather than |
| 49 | +asked of the contributor, because the PR is otherwise complete and the defect is |
| 50 | +in a seam the original author had no reason to suspect. wp4 closes on the review |
| 51 | +being posted; the landing of the corrected work is its own later work-phase. |
| 52 | + |
| 53 | + |
| 54 | +## Execution record |
| 55 | + |
| 56 | +Review posted on #2433 on 2026-08-23 naming the blocker, the reproduction, and the |
| 57 | +required shape of the fix. PR left open, not closed. Fix in flight on |
| 58 | +`codex/fix-2433-exactly-once-terminal`. |
0 commit comments