| name | context-monitor |
| description | Context window usage monitoring and auto-compact recovery. Use when context is filling up or after auto-compact occurs. Optional for strong models. |
Context Monitor
Thresholds
| Level | Action |
|---|
| < 70% | Normal operation |
| 70-85% | Persist current state to files, update task list |
| 85-95% | Save intermediate results, minimize new file reads |
| ~95% | Auto-compact triggers — context will be compressed |
Note: The thresholds above are behavioral guidelines for agents.
The context_monitor.py script uses warning_threshold (70%) and
auto_compact_threshold (95%) for visualization and alerting.
The 85% intermediate action is the agent's own responsibility.
Pre-Compact Checklist
- Write work state summary to file (SOTA values, current version, next action)
- Update ChangeLog.md with latest results
- List reference file paths for post-compact recovery
Post-Compact Recovery
- Read task file / ChangeLog.md
- Restore minimal working context
- Resume from interruption point
CLI Script
python3 ${CLAUDE_SKILL_DIR}/scripts/context_monitor.py --help
python3 ${CLAUDE_SKILL_DIR}/scripts/context_monitor.py claude --find-latest --status
python3 ${CLAUDE_SKILL_DIR}/scripts/context_monitor.py codex --find-latest --status --json
python3 ${CLAUDE_SKILL_DIR}/scripts/context_monitor.py gemini /path/to/telemetry.jsonl --visualize -o ./viz
python3 ${CLAUDE_SKILL_DIR}/scripts/context_monitor.py claude /path/to/session.jsonl --visualize --context-limit 200000
Supported parsers: claude, codex, gemini, opencode
Multi-Agent Considerations
- Delegate large tasks to sub-agents to distribute context load
- SE/PM should monitor agent context consumption trends
- Use
PG_visible_dir.md to limit unnecessary file reads
