Best possible solution:

Land one targeted subagent-aware cron fallback fix after maintainer review, and keep the linked issue open until the chosen PR merges.

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

Yes, at source level. Current main can select subagents.model for isolated cron agentTurn runs while resolving fallbacks and embedded fallbackConfigured from the normal model policy; the linked issue includes a matching timeout log.

Is this the best way to solve the issue?

Yes. The PR takes the narrow maintainable path by carrying the selected subagent fallback policy into both the outer fallback runner and embedded failover flag while preserving payload, hook, session, and agent-primary behavior.

Acceptance criteria:

• node scripts/run-vitest.mjs src/cron/isolated-agent/run.message-tool-policy.test.ts src/cron/isolated-agent/run-fallback-policy.test.ts src/cron/isolated-agent/run.payload-fallbacks.test.ts src/agents/pi-embedded-runner/run/fallbacks.test.ts src/agents/agent-scope.test.ts src/cron/isolated-agent.model-formatting.test.ts src/cron/isolated-agent/run.cron-model-override-forwarding.test.ts src/agents/pi-embedded-runner/run/failover-observation.test.ts src/agents/pi-embedded-runner/run/failover-policy.test.ts
• pnpm exec oxfmt --check --threads=1 CHANGELOG.md src/agents/agent-scope.ts src/agents/pi-embedded-runner/run.ts src/agents/pi-embedded-runner/run/fallbacks.ts src/agents/pi-embedded-runner/run/fallbacks.test.ts src/cron/isolated-agent.model-formatting.test.ts src/cron/isolated-agent/model-selection.ts src/cron/isolated-agent/run-execution.runtime.ts src/cron/isolated-agent/run-executor.ts src/cron/isolated-agent/run-fallback-policy.test.ts src/cron/isolated-agent/run-fallback-policy.ts src/cron/isolated-agent/run.payload-fallbacks.test.ts src/cron/isolated-agent/run.test-harness.ts src/cron/isolated-agent/run.ts src/cron/isolated-agent/run.message-tool-policy.test.ts
• node scripts/run-oxlint.mjs CHANGELOG.md src/agents/agent-scope.ts src/agents/pi-embedded-runner/run.ts src/agents/pi-embedded-runner/run/fallbacks.ts src/agents/pi-embedded-runner/run/fallbacks.test.ts src/cron/isolated-agent.model-formatting.test.ts src/cron/isolated-agent/model-selection.ts src/cron/isolated-agent/run-execution.runtime.ts src/cron/isolated-agent/run-executor.ts src/cron/isolated-agent/run-fallback-policy.test.ts src/cron/isolated-agent/run-fallback-policy.ts src/cron/isolated-agent/run.payload-fallbacks.test.ts src/cron/isolated-agent/run.test-harness.ts src/cron/isolated-agent/run.ts src/cron/isolated-agent/run.message-tool-policy.test.ts
• git diff --check

What I checked:

• Protected-label gate: The hydrated GitHub context shows this PR is open and labeled maintainer, which requires explicit maintainer handling rather than cleanup close; it also has proof: override. (d6f2bb8173ea)
• Current main selects subagent models without recording source: Current main chooses agentConfigOverride.subagents.model, then agentConfigOverride.model, then agents.defaults.subagents.model, but the success result only returns provider/model, so the later fallback policy cannot tell when the selected model came from subagent config. (src/cron/isolated-agent/model-selection.ts:85, cce12697a1ee)
• Current main fallback resolver ignores subagent fallback lists: Current main resolves cron fallbacks from payload fallbacks or resolveEffectiveModelFallbacks, which reads the normal agent/default model policy rather than subagents.model.fallbacks. (src/cron/isolated-agent/run-fallback-policy.ts:14, cce12697a1ee)
• Current embedded runner failover flag uses normal model fallback config: runEmbeddedPiAgent currently computes fallbackConfigured with hasConfiguredModelFallbacks, so embedded timeout failover decisions do not see subagent-specific fallback arrays unless an override is forwarded. (src/agents/pi-embedded-runner/run.ts:482, cce12697a1ee)
• Existing config contract supports subagent fallbacks: AgentModelConfig already supports { primary, fallbacks }, and agents.defaults.subagents.model plus per-agent subagents.model use that type, so this is a contract gap rather than a new config option. (src/config/types.agents-shared.ts:8, cce12697a1ee)
• PR diff addresses the implicated path: The PR adds model-source tracking, resolves subagent fallback overrides, forwards them to runWithModelFallback, and passes the same override into the embedded runner's modelFallbacksOverride path. (src/cron/isolated-agent/run-fallback-policy.ts:1, d6f2bb8173ea)

Likely related people:

• vincentkoc: Current-line blame and git show route the extracted isolated cron model-selection and fallback-policy files to commit 4e10969, which introduced the current modular cron runtime surface. (role: recent cron runtime contributor; confidence: high; commits: 4e10969aded0; files: src/cron/isolated-agent/model-selection.ts, src/cron/isolated-agent/run-fallback-policy.ts, src/agents/agent-scope.ts)
• steipete: Recent local history shows multiple cron/fallback refactors and tests by Peter Steinberger, including owner-tool cron coverage and fallback assertion cleanup; this relevance is from merged history, not only authorship of this PR. (role: recent cron and fallback area contributor; confidence: medium; commits: 03a6e3b460a7, 8fd043abac71, b37dc4224072; files: src/cron/isolated-agent, src/agents/model-fallback.ts)
• yinghaosang: Commit f902697 added the per-job cron payload.fallbacks behavior that the PR preserves and composes with the new subagent fallback path. (role: fallback-policy feature contributor; confidence: medium; commits: f902697bd5c4; files: src/cron/isolated-agent/run-fallback-policy.ts, src/cron/isolated-agent/run.payload-fallbacks.test.ts)
• Aviral: Commit d0ca02e is titled as the fix for respecting subagents.model in isolated cron sessions, the selected-model behavior whose fallback policy this PR extends. (role: subagent cron model behavior contributor; confidence: medium; commits: d0ca02e963b6; files: src/cron/isolated-agent/model-selection.ts)

Remaining risk / open question:

• No focused tests or live provider-timeout run were executed during this read-only review.
• The PR body reports mocked regression tests only; the real-behavior proof gate is waived by proof: override, not satisfied by attached live cron/provider output.

Codex review notes: model gpt-5.5, reasoning high; reviewed against cce12697a1ee.