Best possible solution:

Use one complete refusal-to-visible-text fix across both OpenAI-completions assemblers, with green type checks, focused regression coverage, and redacted real runtime proof while leaving fallback/error policy to separate work.

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

Yes. Source inspection gives a high-confidence path: feed either OpenAI-completions parser a refusal-only delta or message and current main reads content, reasoning, and tool calls but not refusal.

Is this the best way to solve the issue?

No. The narrow parser-parity direction is right, but this branch is incomplete until it typechecks and covers both live completions assemblers with proof.

Full review comments:

• [P1] Fix the aggregated refusal fixture type — src/agents/openai-transport-stream.test.ts:2787-2793
  The current check-test-types job fails here because processOpenAICompletionsStream accepts AsyncIterable<ChatCompletionChunk>, but this fixture yields object: "chat.completion" with a message choice and no required delta. Please make the aggregated shape an explicit tolerated cast/helper or otherwise keep the fixture assignable so type checks pass.
  Confidence: 0.94
• [P2] Mirror refusal handling in the shared completions provider — src/agents/openai-transport-stream.ts:3169-3170
  This fixes only the agent transport parser, but packages/ai/src/providers/openai-completions.ts is a live built-in openai-completions assembler and still handles content/tool calls without reading choice.delta.refusal. Users routed through that shared provider path can still get an empty assistant turn, so the sibling parser and package-level regression coverage need the same repair.
  Confidence: 0.9

Overall correctness: patch is incorrect
Overall confidence: 0.92

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning high; reviewed against 6630138598b3.

Label changes:

• add P2: This PR addresses a normal-priority provider-output bug with limited blast radius, but the current patch still needs correctness and proof follow-up.
• add rating: 🧂 unranked krab: Overall readiness is 🧂 unranked krab; proof is 🧂 unranked krab and patch quality is 🦪 silver shellfish.
• add status: 📣 needs proof: The PR needs real behavior proof before ClawSweeper can clear the contributor ask. Needs real behavior proof before merge: The PR body lists only a Vitest command, so it still needs redacted real behavior proof such as terminal output, logs, copied live output, or a recording from an OpenAI-compatible runtime turn. After adding proof, update the PR body; ClawSweeper should re-review automatically. If it does not, the PR author or someone with repository write access can comment @clawsweeper re-review.

Label justifications:

• P2: This PR addresses a normal-priority provider-output bug with limited blast radius, but the current patch still needs correctness and proof follow-up.
• rating: 🧂 unranked krab: Overall readiness is 🧂 unranked krab; proof is 🧂 unranked krab and patch quality is 🦪 silver shellfish.
• status: 📣 needs proof: The PR needs real behavior proof before ClawSweeper can clear the contributor ask. Needs real behavior proof before merge: The PR body lists only a Vitest command, so it still needs redacted real behavior proof such as terminal output, logs, copied live output, or a recording from an OpenAI-compatible runtime turn. After adding proof, update the PR body; ClawSweeper should re-review automatically. If it does not, the PR author or someone with repository write access can comment @clawsweeper re-review.

PR surface:

Source +3, Tests +80. Total +83 across 2 files.

What I checked:

• Repository policy applied: Root AGENTS.md and scoped src/agents/AGENTS.md were read; their provider sibling-surface and real-proof requirements affected this review. (AGENTS.md:1, 6630138598b3)
• Current main agent parser gap: Current main resolves choice.delta or choice.message, then handles content, reasoning, and tool calls without reading refusal. (src/agents/openai-transport-stream.ts:3135, 6630138598b3)
• Current main shared provider gap: The shared packages/ai OpenAI-completions stream handles reasoning fields, content, and tool calls, but does not read choice.delta.refusal. (packages/ai/src/providers/openai-completions.ts:427, 6630138598b3)
• Shared provider is reachable: The built-in registry lazily registers openai-completions to packages/ai/src/providers/openai-completions.ts, so the sibling parser is live runtime code. (packages/ai/src/providers/register-builtins.ts:103, 6630138598b3)
• PR typecheck failure: The check-test-types job fails because the aggregated test fixture yields object "chat.completion" with a message-only choice; ChatCompletionChunk choices require delta. (src/agents/openai-transport-stream.test.ts:2787, 97fd038f9b79)
• PR diff is one-sided: The proposed production change adds refusal handling only in src/agents/openai-transport-stream.ts and leaves packages/ai/src/providers/openai-completions.ts unchanged. (src/agents/openai-transport-stream.ts:3169, 97fd038f9b79)

Likely related people:

• vincentkoc: Current blame on both affected parser loops resolves to a recent merged carrier commit, and earlier history shows provider transport seam work in the same agent transport surface. (role: recent area contributor; confidence: high; commits: eb5d8f85c458, 4798e125f40c; files: src/agents/openai-transport-stream.ts, packages/ai/src/providers/openai-completions.ts)
• steipete: Recent merged history touches the shared packages/ai OpenAI-completions provider and built-in registration boundary. (role: adjacent package-runtime contributor; confidence: medium; commits: 4bf70be01a21; files: packages/ai/src/providers/openai-completions.ts, packages/ai/src/providers/register-builtins.ts)
• balajisiva: Authored a nearby merged OpenAI-compatible completions silent-drop repair for vLLM reasoning and empty tool-call response parsing. (role: adjacent bug-fix author; confidence: medium; commits: efab9763dc87; files: src/agents/openai-transport-stream.ts)

• 🦀 challenger crab: rare, exceptional readiness with strong proof, clean implementation, and convincing validation.
• 🦞 diamond lobster: very strong readiness with only minor maintainer review expected.
• 🐚 platinum hermit: good normal PR, likely mergeable with ordinary maintainer review.
• 🦐 gold shrimp: useful signal, but proof or patch confidence is still limited.
• 🦪 silver shellfish: thin signal; proof, validation, or implementation needs work.
• 🧂 unranked krab: not merge-ready because proof is missing/unusable or there are serious correctness or safety concerns.
• 🌊 off-meta tidepool: rating does not apply to this item.

Shiny media proof means a screenshot, video, or linked artifact directly shows the changed behavior. Runtime, network, CSP, and security claims still need visible diagnostics.

• ClawSweeper keeps one durable marker-backed review comment per issue or PR.
• Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
• A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
• PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
• Maintainers can also comment @clawsweeper review to request a fresh review only.
• Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
• Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
• Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.