Best possible solution:

Land this focused commitments-runtime fix after exact-head checks and mergeability are accepted, then let it close the linked bug; keep broader background retry or circuit-breaker policy work separate.

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

Yes. The linked bug provides concrete config and log steps, and current main gives a static reproduction: enable commitments with agents.defaults.model.primary = openai-codex/gpt-5.5; the extractor omits provider/model and the embedded runner defaults that omission to direct OpenAI.

Is this the best way to solve the issue?

Yes. Reusing resolveDefaultModelForAgent inside defaultExtractBatch and passing the resolved provider/model is the narrowest maintainable fix for the default-model regression; the cooldown remains commitment-local and does not change shared background-run policy.

What I checked:

• Current main omits extractor model routing: defaultExtractBatch calls runEmbeddedPiAgent with session, config, prompt, and safety flags but no provider or model, so the configured agent/default model is not carried into hidden extraction on current main. (src/commitments/runtime.ts:179, bd20f8e07e92)
• Embedded runner defaults explain the reported fallback: When embedded runs receive no provider/model, runEmbeddedPiAgent initializes provider and model from DEFAULT_PROVIDER and DEFAULT_MODEL, matching the direct OpenAI fallback reported by the linked bug. (src/agents/pi-embedded-runner/run.ts:404, bd20f8e07e92)
• Existing resolver is the right local seam: resolveDefaultModelForAgent already resolves the effective agent/default configured model before falling back to global defaults, so using it avoids new model-selection policy. (src/agents/model-selection.ts:200, bd20f8e07e92)
• PR applies the narrow runtime fix: The PR diff imports resolveDefaultModelForAgent, resolves modelRef in defaultExtractBatch, and passes provider: modelRef.provider plus model: modelRef.model to runEmbeddedPiAgent. (src/commitments/runtime.ts:211, 8ffa3a8291e3)
• Regression tests cover model routing and cooldown: The PR adds a runtime test asserting openai-codex/gpt-5.5 is split into provider: openai-codex and model: gpt-5.5, plus a test that terminal auth/model failures suppress only the affected agent until cooldown expiry. (src/commitments/runtime.test.ts:153, 8ffa3a8291e3)
• Docs support provider-qualified model refs: The config docs specify model.primary as provider/model and explicitly cite openai-codex/gpt-5.5 for Codex OAuth, while the OpenAI provider docs describe openai-codex/* as the Codex OAuth subscription route. Public docs: docs/gateway/config-agents.md. (docs/gateway/config-agents.md:371, bd20f8e07e92)

Likely related people:

• vignesh07: Recent main history shows Vignesh authored the immediately preceding commitments runtime/test safety work, and the changelog credits this area to @vignesh07. (role: recent commitments maintainer; confidence: high; commits: b277ae3f4c40; files: src/commitments/runtime.ts, src/commitments/runtime.test.ts, src/commitments/commitments-full-chain.integration.test.ts)
• steipete: Current blame and recent log history tie Peter to the embedded runner defaulting behavior, resolveDefaultModelForAgent, and adjacent external credential discovery work. (role: model-routing and embedded-runner maintainer; confidence: medium; commits: b40c679630d6, 90419df6633f; files: src/agents/pi-embedded-runner/run.ts, src/agents/model-selection.ts, src/commitments/runtime.ts)
• pashpashpash: Recent adjacent history changed embedded runner state handling for Codex app-server isolation, which is near the provider/model runtime boundary touched by this PR. (role: adjacent embedded runtime contributor; confidence: low; commits: 027ea5f08bd9; files: src/agents/pi-embedded-runner/run.ts)

Remaining risk / open question:

• No local tests were run because this was a read-only review; exact-head CI/check-run data and the PR’s stated validation should remain the merge gate.

Codex review notes: model gpt-5.5, reasoning high; reviewed against bd20f8e07e92.