Problem
flagged_retriable in packages/platform-apple/src/runner/runner-error-classification.ts derives the transport's "same-session resend" verdict (retryable) from the caller-facing details.retriable flag. These are different facts:
details.retriable tells a caller's poll (for example wait, wait-polling.ts) to try again.retryable tells runner-client.ts to resend the same request.
#2852 had to add an order-dependent app_not_running row just to undo the conflation for APP_NOT_RUNNING.
Today the conflation also resends:
Proposal
Make flagged_retriable key on the fact it means: details.runnerErrorCode === RUNNER_BUSY_RUNNER_CODE (renamed runner_busy). Then delete the app_not_running row. Move the pin in runner-error-classification.test.ts (the generic-flag resend case) and decide explicitly whether startup-phase and provider errors should be resent.
Found in the thermo-nuclear review of #2852; kept out of that PR because it changes retry behavior beyond it.
Problem
flagged_retriableinpackages/platform-apple/src/runner/runner-error-classification.tsderives the transport's "same-session resend" verdict (retryable) from the caller-facingdetails.retriableflag. These are different facts:details.retriabletells a caller's poll (for examplewait,wait-polling.ts) to try again.retryabletellsrunner-client.tsto resend the same request.#2852 had to add an order-dependent
app_not_runningrow just to undo the conflation forAPP_NOT_RUNNING.Today the conflation also resends:
retriable: true(runner-cache-metadata.ts:115,:320:runner_phase_budget_exhausted,apple_toolchain_probe_unavailable) for read-only commands, up to 3 times, each with a fresh startup budget, which works against the fix(ios): budget cold toolchain probes for the first-exec signature stall #2423 cap;retriable: trueerrors.Proposal
Make
flagged_retriablekey on the fact it means:details.runnerErrorCode === RUNNER_BUSY_RUNNER_CODE(renamedrunner_busy). Then delete theapp_not_runningrow. Move the pin inrunner-error-classification.test.ts(the generic-flag resend case) and decide explicitly whether startup-phase and provider errors should be resent.Found in the thermo-nuclear review of #2852; kept out of that PR because it changes retry behavior beyond it.