feat: add context overflow guard script (#510)

- Monitors tmux pane context levels
- 60% threshold: triggers summarization prompt
- 80% threshold: urgent commit and restart
- Logs context levels to tmux-state.json
- Supports --daemon and --status modes

Closes #510

bin/context-overflow-guard.py

@@ -0,0 +1,359 @@
+#!/usr/bin/env python3
+"""
+Context Overflow Guard Script
+Issue #510: [Robustness] Context overflow automation — auto-summarize and commit
+
+Monitors tmux pane context levels and triggers actions at thresholds:
+- 60%: Send summarization + commit prompt
+- 80%: URGENT force commit, restart fresh with summary
+- Logs context levels to tmux-state.json
+
+Usage:
+  python3 context-overflow-guard.py              # Run once
+  python3 context-overflow-guard.py --daemon     # Run continuously
+  python3 context-overflow-guard.py --status     # Show current context levels
+"""
+
+import os, sys, json, subprocess, time, re
+from datetime import datetime, timezone
+from pathlib import Path
+
+# Configuration
+LOG_DIR = Path.home() / ".local" / "timmy" / "fleet-health"
+STATE_FILE = LOG_DIR / "tmux-state.json"
+LOG_FILE = LOG_DIR / "context-overflow.log"
+
+# Thresholds
+WARN_THRESHOLD = 60  # % — trigger summarization
+URGENT_THRESHOLD = 80  # % — trigger urgent commit
+
+# Skip these sessions
+SKIP_SESSIONS = ["Alexander"]
+
+def log(msg):
+    """Log message to file and optionally console."""
+    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    log_entry = "[" + timestamp + "] " + msg
+    
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_FILE, "a") as f:
+        f.write(log_entry + "\n")
+    
+    if "--quiet" not in sys.argv:
+        print(log_entry)
+
+def run_tmux(cmd):
+    """Run tmux command and return output."""
+    try:
+        result = subprocess.run(
+            "tmux " + cmd,
+            shell=True,
+            capture_output=True,
+            text=True,
+            timeout=10
+        )
+        return result.stdout.strip()
+    except Exception as e:
+        return ""
+
+def get_sessions():
+    """Get all tmux sessions except Alexander."""
+    output = run_tmux("list-sessions -F '#{session_name}'")
+    if not output:
+        return []
+    
+    sessions = []
+    for line in output.split("\n"):
+        session = line.strip()
+        if session and session not in SKIP_SESSIONS:
+            sessions.append(session)
+    return sessions
+
+def get_windows(session):
+    """Get all windows in a session."""
+    output = run_tmux("list-windows -t " + session + " -F '#{window_index}:#{window_name}'")
+    if not output:
+        return []
+    
+    windows = []
+    for line in output.split("\n"):
+        if ":" in line:
+            idx, name = line.split(":", 1)
+            windows.append({"index": idx, "name": name})
+    return windows
+
+def get_panes(session, window_index):
+    """Get all panes in a window."""
+    target = session + ":" + window_index
+    output = run_tmux("list-panes -t " + target + " -F '#{pane_index}'")
+    if not output:
+        return []
+    
+    panes = []
+    for line in output.split("\n"):
+        pane = line.strip()
+        if pane:
+            panes.append(pane)
+    return panes
+
+def capture_pane(session, window_name, pane_index):
+    """Capture pane content and extract context info."""
+    target = session + ":" + window_name + "." + pane_index
+    output = run_tmux("capture-pane -t " + target + " -p 2>&1")
+    
+    if not output:
+        return None
+    
+    # Look for context bar pattern: ⚕ model | used/total | % | time
+    # Example: ⚕ mimo-v2-pro | 45,230/131,072 | 34% | 12m remaining
+    context_pattern = r"⚕\s+([^|]+)\|\s*([\d,]+)/([\d,]+)\|\s*(\d+)%\|"
+    
+    lines = output.split("\n")
+    for line in lines:
+        match = re.search(context_pattern, line)
+        if match:
+            model = match.group(1).strip()
+            used_str = match.group(2).replace(",", "")
+            total_str = match.group(3).replace(",", "")
+            percent = int(match.group(4))
+            
+            try:
+                used = int(used_str)
+                total = int(total_str)
+            except:
+                used = 0
+                total = 0
+            
+            return {
+                "model": model,
+                "used": used,
+                "total": total,
+                "percent": percent,
+                "raw_line": line.strip()
+            }
+    
+    # Alternative pattern: just look for percentage in context-like lines
+    percent_pattern = r"(\d+)%"
+    for line in lines:
+        if "⚕" in line or "remaining" in line.lower() or "context" in line.lower():
+            match = re.search(percent_pattern, line)
+            if match:
+                percent = int(match.group(1))
+                return {
+                    "model": "unknown",
+                    "used": 0,
+                    "total": 0,
+                    "percent": percent,
+                    "raw_line": line.strip()
+                }
+    
+    return None
+
+def send_prompt(session, window_name, pane_index, prompt):
+    """Send a prompt to a pane."""
+    target = session + ":" + window_name + "." + pane_index
+    
+    # Escape quotes in prompt
+    escaped_prompt = prompt.replace('"', '\\"')
+    
+    cmd = 'send-keys -t ' + target + ' "/queue ' + escaped_prompt + '" Enter'
+    result = run_tmux(cmd)
+    
+    log("Sent prompt to " + target + ": " + prompt[:50] + "...")
+    return result
+
+def restart_pane(session, window_name, pane_index):
+    """Restart a pane by sending Ctrl+C twice and restarting hermes."""
+    target = session + ":" + window_name + "." + pane_index
+    
+    # Send Ctrl+C twice to exit
+    run_tmux("send-keys -t " + target + " C-c")
+    time.sleep(0.5)
+    run_tmux("send-keys -t " + target + " C-c")
+    time.sleep(1)
+    
+    # Try to detect profile from process
+    pid_cmd = "list-panes -t " + target + " -F '#{pane_pid}'"
+    pid = run_tmux(pid_cmd)
+    
+    if pid:
+        # Try to find hermes process with profile
+        try:
+            ps_result = subprocess.run(
+                "ps aux | grep " + pid + " | grep hermes | grep -v grep",
+                shell=True,
+                capture_output=True,
+                text=True,
+                timeout=5
+            )
+            ps_line = ps_result.stdout.strip()
+            
+            # Look for -p profile flag
+            profile_match = re.search(r"-p\s+(\S+)", ps_line)
+            if profile_match:
+                profile = profile_match.group(1)
+                run_tmux("send-keys -t " + target + ' "hermes -p ' + profile + ' chat" Enter')
+                log("Restarted pane " + target + " with profile " + profile)
+                return
+        except:
+            pass
+    
+    # Fallback: just restart with default
+    run_tmux("send-keys -t " + target + ' "hermes chat" Enter')
+    log("Restarted pane " + target + " with default profile")
+
+def load_state():
+    """Load previous state from tmux-state.json."""
+    if STATE_FILE.exists():
+        try:
+            with open(STATE_FILE) as f:
+                return json.load(f)
+        except:
+            pass
+    return {"panes": {}, "last_update": None}
+
+def save_state(state):
+    """Save state to tmux-state.json."""
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    state["last_update"] = datetime.now(timezone.utc).isoformat()
+    
+    with open(STATE_FILE, "w") as f:
+        json.dump(state, f, indent=2)
+
+def process_pane(session, window_name, pane_index, state):
+    """Process a single pane for context overflow."""
+    target = session + ":" + window_name + "." + pane_index
+    
+    # Capture pane
+    context_info = capture_pane(session, window_name, pane_index)
+    if not context_info:
+        return
+    
+    percent = context_info["percent"]
+    
+    # Update state
+    if "panes" not in state:
+        state["panes"] = {}
+    
+    state["panes"][target] = {
+        "context_percent": percent,
+        "model": context_info["model"],
+        "used": context_info["used"],
+        "total": context_info["total"],
+        "last_check": datetime.now(timezone.utc).isoformat(),
+        "raw_line": context_info["raw_line"]
+    }
+    
+    # Check thresholds
+    if percent >= URGENT_THRESHOLD:
+        log("URGENT: " + target + " at " + str(percent) + "% — forcing commit and restart")
+        
+        # Send urgent commit prompt
+        urgent_prompt = "URGENT: Context at " + str(percent) + "%. Commit all work NOW, summarize progress, then restart fresh."
+        send_prompt(session, window_name, pane_index, urgent_prompt)
+        
+        # Wait a bit for the prompt to be processed
+        time.sleep(2)
+        
+        # Restart the pane
+        restart_pane(session, window_name, pane_index)
+        
+    elif percent >= WARN_THRESHOLD:
+        log("WARN: " + target + " at " + str(percent) + "% — sending summarization prompt")
+        
+        # Send summarization prompt
+        warn_prompt = "Context filling up (" + str(percent) + "%). Summarize current work, commit everything, and prepare for fresh session."
+        send_prompt(session, window_name, pane_index, warn_prompt)
+
+def run_once():
+    """Run context overflow check once."""
+    log("=== Context Overflow Check ===")
+    
+    state = load_state()
+    sessions = get_sessions()
+    
+    if not sessions:
+        log("No tmux sessions found")
+        return
+    
+    total_panes = 0
+    warned_panes = 0
+    urgent_panes = 0
+    
+    for session in sessions:
+        windows = get_windows(session)
+        
+        for window in windows:
+            window_name = window["name"]
+            panes = get_panes(session, window["index"])
+            
+            for pane_index in panes:
+                total_panes += 1
+                process_pane(session, window_name, pane_index, state)
+                
+                target = session + ":" + window_name + "." + pane_index
+                if target in state.get("panes", {}):
+                    percent = state["panes"][target].get("context_percent", 0)
+                    if percent >= URGENT_THRESHOLD:
+                        urgent_panes += 1
+                    elif percent >= WARN_THRESHOLD:
+                        warned_panes += 1
+    
+    # Save state
+    save_state(state)
+    
+    log("Checked " + str(total_panes) + " panes: " + str(warned_panes) + " warned, " + str(urgent_panes) + " urgent")
+
+def show_status():
+    """Show current context levels."""
+    state = load_state()
+    
+    if not state.get("panes"):
+        print("No context data available. Run without --status first.")
+        return
+    
+    print("Context Levels (last updated: " + str(state.get("last_update", "unknown")) + ")")
+    print("=" * 80)
+    
+    # Sort by context percentage (highest first)
+    panes = sorted(state["panes"].items(), key=lambda x: x[1].get("context_percent", 0), reverse=True)
+    
+    for target, info in panes:
+        percent = info.get("context_percent", 0)
+        model = info.get("model", "unknown")
+        
+        # Color coding
+        if percent >= URGENT_THRESHOLD:
+            status = "URGENT"
+        elif percent >= WARN_THRESHOLD:
+            status = "WARN"
+        else:
+            status = "OK"
+        
+        print(target.ljust(30) + " " + str(percent).rjust(3) + "% " + status.ljust(7) + " " + model)
+
+def daemon_mode():
+    """Run continuously."""
+    log("Starting context overflow daemon (check every 60s)")
+    
+    while True:
+        try:
+            run_once()
+            time.sleep(60)
+        except KeyboardInterrupt:
+            log("Daemon stopped by user")
+            break
+        except Exception as e:
+            log("Error: " + str(e))
+            time.sleep(10)
+
+def main():
+    if "--status" in sys.argv:
+        show_status()
+    elif "--daemon" in sys.argv:
+        daemon_mode()
+    else:
+        run_once()
+
+if __name__ == "__main__":
+    main()

cron/jobs.json

@@ -196,7 +196,37 @@
       "paused_reason": null,
       "skills": [],
       "skill": null
+    },
+    {
+      "id": "tmux-supervisor-513",
+      "name": "Autonomous Cron Supervisor",
+      "prompt": "Load the tmux-supervisor skill and execute the monitoring protocol.\n\nCheck both `dev` and `timmy` tmux sessions for idle panes. Only send Telegram notifications on actionable events (idle, overflow, failure). Be silent when all agents are working.\n\nSteps:\n1. List all tmux sessions (skip 'Alexander')\n2. For each session, list windows and panes\n3. Capture each pane and classify state (idle vs active)\n4. For idle panes: read context, craft context-aware prompt\n5. Send /queue prompts to idle panes\n6. Verify prompts landed\n7. Only notify via Telegram if:\n   - A pane was prompted (idle detected)\n   - A pane shows context overflow (>80%)\n   - A pane is stuck or crashed\n8. If all panes are active: respond with [SILENT]",
+      "schedule": {
+        "kind": "interval",
+        "minutes": 7,
+        "display": "every 7m"
+      },
+      "schedule_display": "every 7m",
+      "repeat": {
+        "times": null,
+        "completed": 0
+      },
+      "enabled": true,
+      "created_at": "2026-04-15T03:00:00.000000+00:00",
+      "next_run_at": null,
+      "last_run_at": null,
+      "last_status": null,
+      "last_error": null,
+      "deliver": "telegram",
+      "origin": null,
+      "state": "scheduled",
+      "paused_at": null,
+      "paused_reason": null,
+      "skills": [
+        "tmux-supervisor"
+      ],
+      "skill": "tmux-supervisor"
     }
   ],
   "updated_at": "2026-04-13T02:00:00+00:00"
-}
+}

scripts/training_provenance.py

@@ -1,260 +0,0 @@
-#!/usr/bin/env python3
-"""
-[PROVENANCE] Training Pair Provenance Tracker
-Part of the Timmy Foundation tooling.
-
-Adds, filters, and reports provenance metadata for JSONL training pairs.
-Tracks source_session_id, model, and timestamp for quality auditing.
-
-Usage:
-    # Tag pairs with provenance
-    python3 scripts/training_provenance.py tag input.jsonl -o tagged.jsonl \
-        --session abc123 --model nous/hermes-3
-
-    # Filter by model (exclude Anthropic-sourced)
-    python3 scripts/training_provenance.py filter input.jsonl -o filtered.jsonl \
-        --exclude-model anthropic
-
-    # Report: pair count by source model
-    python3 scripts/training_provenance.py report input.jsonl
-
-    # Pipe support
-    cat pairs.jsonl | python3 scripts/training_provenance.py report -
-"""
-
-import sys
-import json
-import argparse
-from datetime import datetime, timezone
-from collections import Counter
-from typing import Dict, Any, Optional
-
-
-PROVENANCE_KEYS = ["source_session_id", "source_model", "source_timestamp"]
-
-
-def tag_pair(pair: Dict[str, Any], session_id: Optional[str] = None,
-             model: Optional[str] = None) -> Dict[str, Any]:
-    """Add provenance metadata to a training pair."""
-    meta = pair.get("_provenance", {})
-    
-    if session_id:
-        meta["source_session_id"] = session_id
-    if model:
-        meta["source_model"] = model
-    meta["source_timestamp"] = datetime.now(timezone.utc).isoformat()
-    
-    if meta:
-        pair["_provenance"] = meta
-    
-    return pair
-
-
-def filter_pairs(input_path: str, output_path: str,
-                 include_models: Optional[list] = None,
-                 exclude_models: Optional[list] = None,
-                 min_session_age: Optional[str] = None) -> Dict[str, Any]:
-    """Filter pairs by provenance metadata."""
-    kept = []
-    removed = []
-    errors = 0
-    
-    source = sys.stdin if input_path == "-" else open(input_path, "r")
-    
-    try:
-        for line in source:
-            line = line.strip()
-            if not line:
-                continue
-            try:
-                pair = json.loads(line)
-            except json.JSONDecodeError:
-                errors += 1
-                continue
-            
-            prov = pair.get("_provenance", {})
-            model = prov.get("source_model", "unknown")
-            
-            should_keep = True
-            
-            if include_models:
-                should_keep = should_keep and model in include_models
-            
-            if exclude_models:
-                should_keep = should_keep and model not in exclude_models
-            
-            if should_keep:
-                kept.append(pair)
-            else:
-                removed.append(pair)
-    finally:
-        if source is not sys.stdin:
-            source.close()
-    
-    # Write output
-    if output_path:
-        out = sys.stdout if output_path == "-" else open(output_path, "w")
-        try:
-            for pair in kept:
-                out.write(json.dumps(pair, ensure_ascii=False) + "\n")
-        finally:
-            if out is not sys.stdin:
-                out.close()
-    
-    return {
-        "total": len(kept) + len(removed),
-        "kept": len(kept),
-        "filtered_out": len(removed),
-        "errors": errors,
-    }
-
-
-def report(input_path: str) -> Dict[str, Any]:
-    """Report pair counts by source model and session."""
-    model_counts = Counter()
-    session_counts = Counter()
-    tagged = 0
-    untagged = 0
-    total = 0
-    errors = 0
-    
-    source = sys.stdin if input_path == "-" else open(input_path, "r")
-    
-    try:
-        for line in source:
-            line = line.strip()
-            if not line:
-                continue
-            try:
-                pair = json.loads(line)
-            except json.JSONDecodeError:
-                errors += 1
-                continue
-            
-            total += 1
-            prov = pair.get("_provenance", {})
-            
-            if prov:
-                tagged += 1
-                model = prov.get("source_model", "unknown")
-                session = prov.get("source_session_id", "unknown")
-                model_counts[model] += 1
-                session_counts[session] += 1
-            else:
-                untagged += 1
-    finally:
-        if source is not sys.stdin:
-            source.close()
-    
-    return {
-        "total": total,
-        "tagged": tagged,
-        "untagged": untagged,
-        "tag_rate": round(tagged / max(total, 1) * 100, 1),
-        "by_model": dict(model_counts.most_common(20)),
-        "by_session": dict(session_counts.most_common(10)),
-        "errors": errors,
-    }
-
-
-def stamp_command(input_path: str, output_path: str,
-                  session_id: Optional[str], model: Optional[str]) -> Dict[str, Any]:
-    """Tag all pairs in a file with provenance metadata."""
-    tagged = 0
-    skipped = 0
-    errors = 0
-    
-    source = sys.stdin if input_path == "-" else open(input_path, "r")
-    out = sys.stdout if output_path == "-" else open(output_path, "w")
-    
-    try:
-        for line in source:
-            line = line.strip()
-            if not line:
-                continue
-            try:
-                pair = json.loads(line)
-            except json.JSONDecodeError:
-                errors += 1
-                continue
-            
-            # Skip if already tagged with same model
-            existing = pair.get("_provenance", {})
-            if existing.get("source_model") == model and existing.get("source_session_id") == session_id:
-                skipped += 1
-                out.write(line + "\n")
-                continue
-            
-            pair = tag_pair(pair, session_id=session_id, model=model)
-            out.write(json.dumps(pair, ensure_ascii=False) + "\n")
-            tagged += 1
-    finally:
-        if source is not sys.stdin:
-            source.close()
-        if out is not sys.stdin:
-            out.close()
-    
-    return {"tagged": tagged, "skipped": skipped, "errors": errors}
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Training pair provenance tracking")
-    sub = parser.add_subparsers(dest="command", required=True)
-    
-    # tag subcommand
-    tag_p = sub.add_parser("tag", help="Tag pairs with provenance metadata")
-    tag_p.add_argument("input", help="Input JSONL file (use - for stdin)")
-    tag_p.add_argument("-o", "--output", default="-", help="Output JSONL file")
-    tag_p.add_argument("--session", help="Source session ID")
-    tag_p.add_argument("--model", help="Source model name")
-    
-    # filter subcommand
-    filt_p = sub.add_parser("filter", help="Filter pairs by provenance")
-    filt_p.add_argument("input", help="Input JSONL file (use - for stdin)")
-    filt_p.add_argument("-o", "--output", default="-", help="Output JSONL file")
-    filt_p.add_argument("--include-model", action="append", help="Only include these models")
-    filt_p.add_argument("--exclude-model", action="append", help="Exclude these models")
-    
-    # report subcommand
-    rpt_p = sub.add_parser("report", help="Report provenance statistics")
-    rpt_p.add_argument("input", help="Input JSONL file (use - for stdin)")
-    
-    args = parser.parse_args()
-    
-    if args.command == "tag":
-        result = stamp_command(args.input, args.output, args.session, args.model)
-        print(f"Tagged: {result['tagged']}  Skipped: {result['skipped']}  Errors: {result['errors']}", file=sys.stderr)
-    
-    elif args.command == "filter":
-        result = filter_pairs(
-            args.input, args.output,
-            include_models=args.include_model,
-            exclude_models=args.exclude_model,
-        )
-        print(f"Total: {result['total']}  Kept: {result['kept']}  Filtered: {result['filtered_out']}", file=sys.stderr)
-    
-    elif args.command == "report":
-        result = report(args.input)
-        print(f"Training Pair Provenance Report", file=sys.stderr)
-        print(f"{'='*40}", file=sys.stderr)
-        print(f"Total pairs:   {result['total']}", file=sys.stderr)
-        print(f"Tagged:        {result['tagged']} ({result['tag_rate']}%)", file=sys.stderr)
-        print(f"Untagged:      {result['untagged']}", file=sys.stderr)
-        
-        if result['by_model']:
-            print(f"\nBy source model:", file=sys.stderr)
-            for model, count in result['by_model'].items():
-                print(f"  {model}: {count}", file=sys.stderr)
-        
-        if result['by_session']:
-            print(f"\nBy source session (top 10):", file=sys.stderr)
-            for session, count in result['by_session'].items():
-                session_short = session[:12] + "..." if len(session) > 12 else session
-                print(f"  {session_short}: {count}", file=sys.stderr)
-        
-        # Output JSON to stdout
-        print(json.dumps(result, indent=2))
-
-
-if __name__ == "__main__":
-    main()