fix(apple-runner): resend a read only on the runner's busy refusal

[thymikee]

Summary

A read-only runner command is now resent inside one request only when the runner refused it with RUNNER_BUSY. Before, any COMMAND_FAILED with details.retriable: true was resent. That flag is for a caller's poll (wait); #1126 added the check for RUNNER_BUSY only. The rule is renamed flagged_retriable → runner_busy and keys on details.runnerErrorCode. The app_not_running row and its ordering note are deleted.

Producers of retriable: true that reach the classifier:

• RUNNER_BUSY: still resent.
• APP_NOT_RUNNING: not resent, as before.
• runner_phase_budget_exhausted and apple_toolchain_probe_unavailable (thrown in ensureRunnerSession): no longer resent. Each resend opened a fresh startup budget, which defeated the fix(ios): budget cold toolchain probes for the first-exec signature stall #2423 cap. A probe timeout already gets one retry inside that budget.
• External AppleRunnerProviders: no resend on the flag alone. A provider that forwards the runner's RUNNER_BUSY code is still resent.

All keep retriable, so the caller retries on its next request. CHANGELOG updated. 4 files. Closes #2862.

With #2804: its runner_busy_refusal row matches this runner_busy row. When it rebases, it adds drainResend: true here and drops its duplicate row. The new busy test does not assert the status probe that #2804 removes.

Validation

Tested at 7a24d34091. pnpm check:affected --run passed: 387 files, 2669 tests. The first run had one timeout in the unrelated runner-client.test.ts; it passed on the rerun.

All tests go through runAppleRunnerCommand. Mutations that each turn tests red:

• Key the rule on retriable: APP_NOT_RUNNING, startup-phase and provider tests fail.
• Restore main's classifier: the three retriable-producer tests fail.
• Set retryable: false on runner_busy: the busy-resend and alert-provider tests fail.

No device run: no Swift or wire change.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.80 MB	4.80 MB	-280 B
Package (unpacked)	4.80 MB	4.80 MB	-280 B
Package (download)	1.44 MB	1.44 MB	-27 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.7 ms	26.2 ms	-0.5 ms
CLI --help	79.1 ms	77.3 ms	-1.8 ms

[thymikee]

Adversarial review: clean. Applied its one note in 105702ed69: the comment now names the toolchain probe among the retriable producers. Merge-order note: #2804 declares its own hasRunnerBusyCode/runner_busy_refusal row. Whichever of the two merges second must fold them into one runner_busy row carrying retryable: true and drainResend: true; otherwise the duplicate declaration fails to compile.

[thymikee]

Reviewed at 105702e. The code looks correct: only a RUNNER_BUSY refusal now earns a read-only resend, and the spent startup budget, unreadable toolchain and external-provider cases are no longer resent. The new tests fail on the old classifier and pass here.

Not blocking: removing the app_not_running row from #2852 also removes its explicit connectRetry: false, and shouldRetryRunnerConnectError falls back to true when no row matches. I found no route where APP_NOT_RUNNING reaches that check today. Would you keep a narrowed row with only connectRetry: false, or add a classification test for that verdict?

Smoke Tests was still queued, so there is no CI result to attribute yet. No device run is needed for a classification-table change. #2804 adds its own runner-busy row in the same file, so whichever PR merges second should fold the two into one rule.

[thymikee]

Rebased onto main after #2804 merged, and squashed to 31eecfffcf. As predicted, the two busy rules are now one: #2804's runner_busy_refusal row (retryable, connectRetry, drainResend, keyed on the RUNNER_BUSY code) stays. This PR deletes the generic flagged_retriable row and the app_not_running row, and keeps main's isRetryableRunnerError doc. The classification and recovery-wiring tests pass 47/47, and pnpm check:affected --run passed.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-09-24 09:10 UTC