[vincentkoc]

Adds two qualified historical repair workloads for Gemini SSE negotiation and durable managed-job observation loss. The instructions define required wire behavior and new positional API arguments before inference; acceptance tests remain outside the repair loop.

Verification: Gemini passes 12 original tests on both revisions, with 2 baseline failures out of 15 acceptance tests; the reference passes 15/15. Managed observation loss passes 10 original tests on both revisions, with 10 baseline failures out of 21; the reference passes 21/21. Every baseline failure is an assertion, with no skips or cancellations. Workload/local-runner checks: 31 passed, three optional Docker checks skipped. Independent P0–P2 review found no actionable issues.

Deployment: none. This changes benchmark manifests, documentation and the qualification job timeout. These are coordinator-selected development tasks; qualification does not establish model quality or a fusion gain. No routing defaults change.

Hosted qualification passed every corpus at the original head, then the job exceeded its 25-minute limit during the final isolated candidate checks. Raise the job allowance to 35 minutes without removing tests or changing individual test deadlines. The timeout follow-up passed independent P0–P2 review. Fresh full CI is required before landing.

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.

ClawSweeper review complete

ClawSweeper finished reviewing this revision. The review result is being finalized.

View the workflow run.

[clawsweeper]

Codex review: needs maintainer review before merge.

What this changes

Adds two pinned historical repair workloads for Gemini SSE negotiation and managed-job observation loss, with qualification documentation.

Example: A coordinator selects the gemini-sse-negotiation benchmark.

• Before: The corpus has no dedicated task grading Gemini alt=sse negotiation.
• After: The task exposes the original query tests and independently grades the repair against the pinned streaming contract.

Review scores

Measure	Result	What it means
Overall readiness	🦐 gold shrimp (3/6)	Focused, coherent corpus extension with no identified defect; complete hosted qualification was not yet available.
Proof confidence	🦐 gold shrimp (3/6)	This benchmark-only change requires no shipped-entry-point proof; reported qualification results match the inspected suite structure, while hosted qualification remains pending.
Patch quality	🐚 platinum hermit (4/6)	No actionable review findings were identified.

Product

Kind: Test only · Worth it: Yes · Fix scope: Complete
User problem: Fusion comparisons need additional repair tasks covering streaming negotiation and durable lifecycle behavior.
Reason: Two distinct historical tasks extend an established owner-directed benchmark without new runtime machinery or unsupported model-quality claims.

Merge readiness

✅ Ready for maintainer review

Keep open: this maintainer-authored PR adds distinct benchmark coverage absent from current main; no actionable defect was found.

Priority: P3
Reviewed head: 77a1db776be0c3e6f909bb0ec447f44d76b873fc

Before merge

None.

Findings

None.

Agent review details

How this fits together

Fusion's offline benchmark harness consumes pinned task manifests, presents editable source and visible tests to repair agents, and grades submitted repairs with separate acceptance tests.

flowchart TD
 Manifest[Workload manifest] --> Qualification[Baseline and reference qualification]
 Qualification --> Agent[Repair agent]
 Agent --> Visible[Visible tests]
 Agent --> Repair[Submitted source repair]
 Repair --> Acceptance[Isolated acceptance tests]
 Acceptance --> Receipt[Benchmark receipt]

Loading

Technical review

Best possible solution:

Land the focused corpus extension once its existing qualification check succeeds.

Do we have a high-confidence way to reproduce the issue?

This is a benchmark addition rather than a current bug report; pinned source contracts and acceptance tests were inspected without executing target code.

Is this the best way to solve the issue?

The existing manifest and qualification machinery provide the smallest complete extension, with explicit API requirements and separate acceptance tests.

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning medium; reviewed against 1f12686c1448.

Testing

Proof path: in-process harness.

Security

None.

Evidence

What I checked:

• Focused additive change: The pinned introduction adds only the workload manifest and documentation; it changes no production source, dependencies, configuration defaults, or stored data. (benchmarks/fusion/stream-lifecycle-workloads.json:1, 77a1db776be0)
• Existing qualification owner: The existing harness checks each oracle's parent, preserves original visible tests, and installs acceptance tests only in a separate verifier; the qualification workflow discovers every workloads.json manifest. (scripts/fusion-workloads.mjs:262, 77a1db776be0)
• Gemini historical contract: The parent pin and editable files match the merged repair at fix(worker): enforce Gemini SSE streaming contract #377; its acceptance tests exercise native and manifest requests, invalid alt values, final usage, and unchanged unary behavior. (worker/test/google-stream-contract.test.mjs:7, e166a8db3119)
• Lifecycle historical contract: The pins and five editable modules match fix(worker): freeze managed observation loss once #487, including the sixth observeManaged argument, fourth collector callback argument, immutable uncertainty receipt, stopped polling, and first bodyless 204 acknowledgement. (benchmarks/fusion/stream-lifecycle-workloads.json:24, d3c91747625e)
• Distinct current-main coverage: Current main contains the earlier discovery, streaming-accounting, lifecycle, and quota workloads, but lacks this manifest; searching task identities and oracle pins found no competing corpus entry. (benchmarks/fusion, 1f12686c1448)
• Validation state: The exact-head Fusion test step passed; historical task qualification remains in progress, with no failed check reported. Some historical blobs were unavailable locally; relevant repair patches and original visible tests were read through GitHub REST. (.github/workflows/fusion-workloads.yml:49, 77a1db776be0)

Likely related people:

• Vincent Koc: Raw commit a7999d7 adds docs/fusion-small-models.md:493 relative to its recorded parents. This identifies author metadata, not feature responsibility or a PR merger. (role: source-line author; confidence: high; commits: a7999d7de81b; files: docs/fusion-small-models.md)

Preflight source

Source unknown. Source comparison at the observed PR snapshot only. Completion is self-reported; body, policy, tests, execution, and report bytes are not verified.

Labels

Label changes:

• add P3: Offline benchmark coverage and documentation; no production behavior changes.
• add rating: 🦐 gold shrimp: Overall readiness is 🦐 gold shrimp; proof is 🦐 gold shrimp and patch quality is 🐚 platinum hermit.
• add status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR.

Label justifications:

• P3: Offline benchmark coverage and documentation; no production behavior changes.
• rating: 🦐 gold shrimp: Overall readiness is 🦐 gold shrimp; proof is 🦐 gold shrimp and patch quality is 🐚 platinum hermit.
• status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR.

Rating scale

6/6 🦀 challenger crab · 5/6 🦞 diamond lobster · 4/6 🐚 platinum hermit · 3/6 🦐 gold shrimp · 2/6 🦪 silver shellfish · 1/6 🧂 unranked krab. Overall follows the weaker of proof and patch quality; ✨ marks media proof (a screenshot, video, or linked artifact) that directly shows the changed behavior.

Workflow

ClawSweeper edits this one comment on every review. Comment @clawsweeper re-review for a fresh review only. Repair and merge execution lanes are retired.

Reviewed November 13, 2026, 12:57 PM ET / 17:57 UTC.