auto-resume-fixes.md

Auto-Resume Script Fixes

Problems Found

1. Processes stuck for 40+ hours - Claude processes had no output for 140,000+ seconds
2. Timeout not enforced - SUBPROCESS_TIMEOUT constant was defined but never used
3. No auto-resume on stuck processes - Script didn't detect and restart stuck sessions
4. No context compaction handling - Didn't detect when Claude hit session limits
5. No auto-continue mode - Never switched to -c flag to resume sessions

Fixes Applied

1. Actual Timeout Enforcement

def heartbeat_logger():
    # Check for timeout - kill if stuck too long
    if elapsed > SUBPROCESS_TIMEOUT:
        log(f"[TIMEOUT] No output for {int(elapsed)}s (timeout: {SUBPROCESS_TIMEOUT}s)")
        log(f"[TIMEOUT] Killing stuck subprocess PID {proc[0].pid}...")
        proc[0].kill()

Now the heartbeat thread actually kills stuck processes after 1 hour of no output.

2. Context Compaction Detection

def check_context_compaction(output):
    """Check if output indicates context window compaction or session limit."""
    patterns = [
        r"context.*window.*full",
        r"compacting.*context",
        r"session.*limit.*reached",
        r"message.*history.*truncated",
        r"conversation.*too.*long",
    ]

Detects when Claude hits session limits and automatically switches to continue mode.

3. Stuck Process Detection

def check_stuck_or_timeout(ret_code, output):
    """Check if Claude got stuck or timed out."""
    # Exit code -9 means killed by SIGKILL (our timeout killer)
    if ret_code == -9 or ret_code == 137:
        return True

Detects when process was killed by timeout and automatically restarts with -c.

4. Auto-Continue Mode

# After first successful run, always use continue flag
if not current_continue:
    log("[SESSION] Switching to continue mode (-c) for subsequent runs...")
    current_continue = True

After the first run completes, all subsequent runs use -c to continue the session.

5. Error Recovery

if ret_code != 0:
    log(f"[ERROR] Will retry with continue flag (-c) in case session needs resuming...")
    current_continue = True
    run_number -= 1  # Retry with continue flag
    time.sleep(10)
    continue

Any error now triggers a retry with continue mode after 10 seconds.

Why --fg Mode?

The script runs in foreground mode by default when called with --fg, but when you run it WITHOUT --fg, it:

1. Spawns a background process with --fg
2. Detaches it from your terminal
3. Redirects stdout/stderr to the log file
4. Returns immediately so you can use your terminal

The --fg in your ps output is normal and expected - it means the background wrapper is working correctly.

How to Use (Correct Way)

Background Mode (Recommended)

./claude-auto-resume "Implement all features in plan/ directory"

This will:

• Start in background
• Return immediately
• Keep running forever
• Log to .claude-auto-resume/session.log

Check Status

./claude-auto-resume --status

Follow Logs

tail -f .claude-auto-resume/session.log

Stop Gracefully

./claude-auto-resume --stop

Force Stop (If Stuck)

./claude-auto-resume --force-stop

Testing the Fixes

Run this to test:

# Clean up old sessions
./claude-auto-resume --force-stop

# Start fresh session
./claude-auto-resume "Implement all features in plan/ directory"

# Follow logs in another terminal
tail -f .claude-auto-resume/session.log

What Changed

1. Timeout enforcement: Kills stuck processes after 1 hour
2. Auto-continue: Switches to -c flag after first run
3. Context detection: Detects session limits and resumes
4. Error recovery: Retries with continue flag on any error
5. Stuck detection: Detects killed processes and restarts

Expected Behavior

• First run: Uses original prompt
• Subsequent runs: Uses -c to continue
• On timeout: Kills process, restarts with -c
• On error: Waits 10s, restarts with -c
• On context limit: Immediately restarts with -c
• On rate limit: Waits until reset, then continues

The script will never get stuck for 2 days again!