**Description**: Unlike `ai-idle-checker.ts`, the plan checker passes the prompt directly as a shell argument rather than via a temp file. This can cause `E2BIG` (argument list too long) errors when the terminal buffer is large.
**Code**:
```typescript
// ai-plan-checker.ts - PROBLEMATIC
constescapedPrompt=prompt.replace(/'/g,"'\\''");
constclaudeCmd=`claude -p ${modelArg} --output-format text '${escapedPrompt}'`;
// ai-idle-checker.ts - CORRECT
writeFileSync(this.checkPromptFile,prompt);
constclaudeCmd=`cat "${this.checkPromptFile}" | claude -p ${modelArg} --output-format text`;
```
**Impact**: With `maxContextChars: 8000`, the prompt can easily exceed shell argument limits (~128KB on most systems), causing the AI plan check to fail with a cryptic error.
**Suggested Fix**: Modify `ai-plan-checker.ts` to use the same temp file approach as `ai-idle-checker.ts`:
1. Add `checkPromptFile` instance variable
2. Write prompt to temp file
3. Use `cat | claude -p` pattern
4. Clean up temp file in `cleanupCheck()`
---
### Issue 2: Missing Prompt File Cleanup on Timeout (MEDIUM)
**Description**: When the AI check times out, the `cleanupCheck()` function is called via `finally`. However, the prompt file deletion in `cleanupCheck()` may fail silently if the file is still being read by the `cat` command in the spawned screen.
**Impact**: Temp files may accumulate in `/tmp` if checks frequently timeout.
**Suggested Fix**: Add a small delay before file deletion or use a unique timestamp-based naming scheme that guarantees no collisions (already partially implemented but the shell command `rm -f` in the fullCmd also handles this).
---
### Issue 3: Race Condition in cancel() with Promise Resolution (MEDIUM)
**Description**: The `cancel()` method resolves the pending promise and then calls `cleanupCheck()`. However, the interval poll timer might fire between these two operations and attempt to resolve an already-resolved promise.
**Code**:
```typescript
cancel():void{
if(this._status!=='checking')return;
this.checkCancelled=true;
// Race: poll timer might fire here
if(this.checkResolve){
this.checkResolve({verdict:'ERROR',...});
this.checkResolve=null;
}
this.cleanupCheck();// This clears the poll timer
this._status='ready';
}
```
**Impact**: Could theoretically cause a double-resolve, though the guard `if (this.checkCancelled)` in the poll handler mitigates this.
**Suggested Fix**: Set `checkCancelled = true` and call `cleanupCheck()` (which clears the poll timer) before resolving the promise.
**Description**: Screen names are generated using only `sessionId.slice(0, 8)`, without a unique suffix. While there's a `screen -X quit` to kill leftover screens, two rapid checks could conflict.
**Description**: The `outputSilent` calculation uses `config.completionConfirmMs`, but when in `ai_checking` state, the output silence threshold should conceptually be the AI check timeout, not the completion confirm time.
**Impact**: UI display may show incorrect "silence" status during AI check.
---
## 4. Coverage Gaps
### Functions Not Fully Tested
1.**`sendClear()`** - Never reaches execution in tests because the cycle is cut short
2.**`sendInit()`** - Same as above
3.**`completeCycle()`** - Tests don't allow cycles to complete fully
4.**`checkClearComplete()`** - Requires mocking prompt detection after /clear
3. Kickstart prompt flow when `/init` doesn't trigger work
4. AI checker returning `WORKING` verdict (only test with timeout)
5. AI plan checker `PLAN_MODE` verdict (real check, not just pre-filter)
### Edge Cases Not Tested
1. Multiple rapid AI checks before cooldown
2. AI checker disabled mid-check
3. Session terminal buffer being `undefined` on start
4. Buffer trimming behavior at MAX_RESPAWN_BUFFER_SIZE
5.`signalElicitation()` called multiple times
---
## 5. Timing and Race Condition Analysis
### Potential Timing Issues
1.**Pre-filter timer vs AI check**: If pre-filter fires while an AI check is already running, it correctly skips (line 1613).
2.**Completion confirm during AI check**: Output during AI check cancels it (line 1173-1188), which is correct.
3.**Auto-accept during respawn cycle**: Correctly guards against non-watching state (line 1741).
4.**Step confirm timer**: Uses same `completionConfirmMs` as idle confirm, which may be too short for complex operations.
### Timer Cleanup
All timers appear to be properly cleaned up in `clearTimers()`:
-`idleTimer`
-`stepTimer`
-`clearFallbackTimer`
-`completionConfirmTimer`
-`stepConfirmTimer`
-`autoAcceptTimer`
-`preFilterTimer`
-`noOutputTimer`
-`detectionUpdateTimer` (interval)
---
## 6. Recommendations (Prioritized)
### Critical (Should Fix)
1.**Fix E2BIG vulnerability in ai-plan-checker.ts**
- Use temp file for prompt like ai-idle-checker.ts does
- Add `checkPromptFile` instance variable
- Update `cleanupCheck()` to delete prompt file
### High Priority
2.**Add integration tests for full respawn cycle**
- Mock session to simulate prompt detection after each step
- Test: update -> clear -> init -> back to watching
- Test: kickstart prompt when init doesn't trigger work
3.**Fix cancel() race condition in AI checkers**
- Clear poll timer before resolving promise
- Or use a mutex/flag check in poll handler
### Medium Priority
4.**Add tests for AI checker WORKING verdict**
- Mock the screen process to return WORKING
- Verify cooldown is started
- Verify controller returns to watching
5.**Add tests for plan checker PLAN_MODE verdict**
- Mock the screen process to return PLAN_MODE
- Verify Enter is sent
- Verify state transitions
6.**Add timestamp to screen session names**
- Prevents potential collisions during rapid operations
### Low Priority
7.**Improve coverage of edge cases**
- Test `sendInit: false` and `sendClear: false` combinations
- Test buffer trimming behavior
- Test pause/resume with activity detection
8.**Consider reducing step confirm timeout**
- Currently uses `completionConfirmMs` (10s default)
- May want separate config for step confirmation
---
## 7. MockSession Limitations
The current `MockSession` class in tests is simplified and doesn't accurately simulate:
1.**Delayed responses**: Real sessions have processing time between input and output
2.**State persistence**: Real sessions maintain state between commands
3.**Screen integration**: Real sessions use GNU screen for persistence
4.**Token tracking**: Real sessions parse and track token usage
Consider adding a more sophisticated mock that can:
- Queue delayed responses
- Simulate multi-step workflows
- Return realistic completion messages with timing
---
## 8. Conclusion
The respawn controller tests are comprehensive for basic functionality but lack coverage for:
- Full respawn cycle completion
- AI checker verdicts (IDLE/WORKING/PLAN_MODE)
- Kickstart functionality
- Edge cases and error handling
The main bug found is in `ai-plan-checker.ts` which can fail with E2BIG errors on large terminal buffers. This should be fixed by using the temp file approach already implemented in `ai-idle-checker.ts`.
All 91 tests pass, and there are no type errors, indicating a stable codebase but with room for deeper integration testing.
