#18468: fix(agents): prevent infinite retry loops in sub-agent completion announcements (#18150)

## Summary Fixes #18150 - prevents 70+ duplicate sub-agent completion announcements from replaying into parent sessions by removing the duplicate resumption bypass in `retryDeferredCompletedAnnounces()`. ## Root Cause When a sub-agent completion announcement is deferred (e.g., waiting for child session to fully settle or for active descendants to complete), the completion flow is designed to retry later. However, `retryDeferredCompletedAnnounces()` was deleting the runId from `resumedRuns` before calling `resumeSubagentRun()`, which bypassed the duplicate prevention check. This created an infinite loop: 1. Sub-agent A completes, `cleanupHandled = true` 2. Announce starts (async) but is deferred (child not settled) 3. `finalizeSubagentCleanup` resets `cleanupHandled = false` to allow retry 4. Sub-agent B completes, triggers `retryDeferredCompletedAnnounces` 5. For run A: `resumedRuns.delete(runId)` + `resumeSubagentRun(runId)` 6. `resumeSubagentRun` bypasses check (runId not in Set) 7. Starts new announce, defers again, resets `cleanupHandled = false` 8. Next run completion triggers retry cascade again... With cron jobs executing periodically, each execution would trigger a retry cascade, causing 70+ duplicate announcements into the parent session, burning tokens on duplicate NO_REPLY processing. ## Changes Removed `resumedRuns.delete(runId)` from `retryDeferredCompletedAnnounces()` at line 290 of subagent-registry.ts. The `resumedRuns` Set should only be cleared by `finalizeSubagentCleanup()` when cleanup genuinely needs to be retried (!didAnnounce), not by the retry checker itself. ## Behavior **Before**: Deferred announces would retry infinitely, causing 70+ duplicate announcements **After**: Deferred announces still retry when conditions change, but duplicate resumption is prevented ## Testing - All 8 subagent-registry tests pass (nested and steer-restart scenarios) - Fix preserves legitimate retry behavior when announces are genuinely deferred - Prevents the duplicate resumption that caused infinite loops 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- greptile_comment --> <h3>Greptile Summary</h3> This PR fixes a critical infinite retry loop bug in sub-agent completion announcement handling. The root cause was in `retryDeferredCompletedAnnounces()` at line 290, which was deleting entries from the `resumedRuns` Set before calling `resumeSubagentRun()`. This bypassed the duplicate prevention check in `resumeSubagentRun()` (line 82), allowing the same run to be resumed repeatedly when announces were deferred. The fix removes the problematic `resumedRuns.delete(runId)` line, ensuring that: - The duplicate prevention check in `resumeSubagentRun()` remains effective - Only `finalizeSubagentCleanup()` clears `resumedRuns` entries when cleanup genuinely needs retry (`!didAnnounce`) - Deferred announces can still retry legitimately when conditions change, but without infinite loops Changes: - Removed `resumedRuns.delete(runId)` from `retryDeferredCompletedAnnounces()` in `src/agents/subagent-registry.ts:290` - Added detailed comment explaining the fix and referencing issue #18150 - Updated CHANGELOG.md with appropriate fix description <h3>Confidence Score: 5/5</h3> - This PR is safe to merge with no risk - it's a minimal, well-understood bug fix that removes problematic code causing infinite loops - The fix is a surgical one-line deletion that addresses a clear logic error. The removed line was bypassing duplicate prevention logic, causing 70+ duplicate announcements. The fix preserves all legitimate retry behavior while preventing the infinite loop. Existing tests cover both the nested agent scenarios and the steer-restart retry scenarios, and the PR description explicitly confirms all 8 subagent-registry tests pass. - No files require special attention _{Last reviewed commit: 8a7aef9} <!-- greptile_other_comments_section --> <!-- /greptile_comment -->