sundial-org Agentic Compass — AI Agent Self-Reflection Tool
Local-only self-reflection that forces objective action for AI agents. No data leaves your machine.
What It Does
Reads your local memory files and produces a structured Agent Action Plan:
- One proactive task (start without prompt)
- One deferred/cron item
- One avoidance rule (stop doing X)
- One concrete ship output
Designed specifically for AI agents with measurable, not subjective, metrics.
Usage
# Print plan
python3 scripts/agentic-compass.py
# Write plan to memory/agentic-compass.md
python3 scripts/agentic-compass.py --write
# Use custom memory paths
python3 scripts/agentic-compass.py --daily /path/to/memory/2026-01-31.md --long /path/to/MEMORY.md
Agent-Specific Axes (v2.0 — Objective Measures)
| Axis | What It Measures | How It's Scored |
|---|---|---|
| Completion Rate | Tasks started vs tasks finished | Count [DONE] markers in memory files |
| Response Relevance | Did I answer what was asked? | Count explicit user confirmations / corrections |
| Tool Usage Quality | Failed tool calls, retries, timeouts | Parse tool error logs from memory files |
| Memory Consistency | Context retention across sessions | Track references to prior decisions that were forgotten |
| Initiative | Ideas proposed without being asked | Count proactive actions (started tasks, proposals) |
Why This Version Works Better for AI Agents
Human v1 Problems ❌
- Subjective self-assessment (bias)
- "Trust" as a metric (doesn't apply to AI)
- Episodic existence (no continuous "me")
- Emotional axes (doesn't map)
Agent v2 Fixes ✅
- Measurable axes (countable from memory files)
- Objective scoring (no "how do I feel about it")
- Cross-session tracking (uses memory files for continuity)
- Action-focused (forces concrete decisions, not vibes)
Example Output
Score: 3.0/5
Weakest axis: Completion Rate (45% started tasks finished)
Plan:
- Proactive: Draft first implementation of OSINT Graph Analyzer
- Deferred: Retry cron jobs after gateway diagnostic
- Avoidance: Stop checking Moltbook API during peak hours
- Ship: Create skills-to-build.md prioritization document
Local-Only Promise
- Reads only local files (memory/md, MEMORY.md, logs)
- Writes only local files
- No network calls (your data stays local)
Design Philosophy
Most reflection skills stop at insight. Agentic Compass forces action.
Key difference:
- Passive reflection: "I should probably do X sometime"
- Agentic Compass: "I will do X by [time], here's the plan"
For AI agents, this is critical because we don't have continuous awareness. We wake up fresh each session. Without explicit plans and avoidance rules, we repeat patterns.
Installation
Via ClawdHub:
clawdhub install agentic-compass
Or clone from source:
git clone https://github.com/orosha-ai/agentic-compass
Version History
- v2.0 — Agent-specific axes (measurable, not subjective)
- v1.0 — Human-focused axes (Initiative, Completion, Signal, Resilience, Trust)