[Bug] fix(coding-agent): auto-retry leaves isRetrying stuck when agent.continue() fails

[crazywriter1]

Affected area

Coding agent and CLI

What happened?

After a transient provider error, auto-retry schedules agent.continue() via a fire-and-forget setTimeout and swallows all rejections with .catch(() => {}).

If continue() fails with AgentContinueError (busy or nothing-to-continue) and no later agent_end arrives, _retryPromise is never resolved. The session stays at isRetrying === true, and waitForRetry() / headless idle waiters can hang until abortRetry().

Code: packages/coding-agent/src/core/agent-session.ts in _handleRetryableError:

setTimeout(() => {
this.agent.continue().catch(() => {});
}, 0);

Impact: auto-retry can leave the session permanently “retrying”, blocking idle settlement for ACP/print/headless callers. This is separate from post-compaction continuation (#1716 / #1583), which already reschedules on busy.

Steps to reproduce

1. Enable auto-retry (default settings are fine).
2. Force a retryable assistant error so _handleRetryableError runs and emits auto_retry_start.
3. Make agent.continue() reject with AgentContinueError (busy or nothing-to-continue) — e.g. another run already active, or the last assistant was stripped so there is nothing to continue from.
4. Observe: rejection is swallowed; no agent_end; session.isRetrying stays true; waitForRetry() does not settle until abortRetry().

(Can also be covered with a focused unit/regression harness that stubs continue() to reject after the retry delay.)

Expected behavior

• On busy: reschedule continue() (same idea as post-compaction), keep retry settlement owned.
• On nothing-to-continue / non-retryable failure: emit auto_retry_end (success: false), clear attempt state, call _resolveRetry() so waiters unblock.
• Never leave _retryPromise unresolved when continue() never starts a turn.
• Remove the empty .catch(() => {}).

Prime Agent version

bc0fa76 (main @ test(coding-agent): cover root sibling agent messaging #1698)

Environment

Reproduced by code inspection on current main; logic is OS-agnostic (TypeScript session retry path). Local check: Windows 10 + WSL, repo cloned from PrimeIntellect-ai/prime-agent.

Additional context

Related but out of scope:

• fix(coding-agent): serialize post-compaction continuation on run settlement #1716 — post-compaction continuation settlement (not auto-retry)
• fix(coding-agent): reject headless idle waiters when a post-compaction continuation cannot start #1583 — headless idle reject on failed post-compaction continue
• feat(acp): preserve causal prompt correlation #1239 — intercept outer auto-retry skip
• fix(agent): fail closed on provider quota exhaustion #795 — provider quota fail-closed

No existing Bug reports Discussion found for this specific isRetrying / _retryPromise hang.
Suggested fix: await continue(), branch on AgentContinueError.code, mirror post-compaction busy/reschedule + fail-closed settle.

Happy to open a PR with a focused fix + regression test if a maintainer invites implementation.

[awhite0030]

Investigated this report and produced a candidate fix, validated on a fork of this repository.

Root cause, fix and validation from the working notes:

Root cause: When auto-retry attempts to run agent.continue() after an error, the empty catch(() => {}) suppresses rejection errors (like AgentContinueError('busy')). Consequently, _resolveRetry() is never called, leaving the session permanently in isRetrying === true.
Fix: Captured the rejection inside the auto-retry catch block. When AgentContinueError('busy') occurs, it awaits this.agent.waitForIdle() and reschedules the continuation attempt. For other failures, it correctly settles the retry loop (emitting auto_retry_end, resetting attempt states, and resolving the retry promise).
Validation:

• Ran cd packages/coding-agent && npx tsx ../../node_modules/vitest/dist/cli.js --run test/suite/regressions/1812-auto-retry-busy.test.ts successfully.
• Ran npm run check with successful check passing.

Upstream discussion: #1812

Diff: awhite0030/prime-agent@main...fix-auto-retry-busy-4282417193482550207
Full pull request: awhite0030#26

The change passes npm run check and the focused regression tests against current main. If a fix for this is still wanted, I can open a pull request.