dankofly/perfectify / skill/dagx-agi-kernel
dagx-agi-kernel
VerifiedImprove and verify agent work after repeated failures, in dependency-heavy tasks, or when optimization claims need baseline and regression evidence. Use for DAGx/Perfectify requests, failed retries, risky multi-step work, or requests to verify an improvement. Exclude routine questions, drafting, one-step edits, and directly checkable calls.
references/self-learning.md
View on GitHub# Self-Learning Protocol (V0.8+)
The playbook (`playbook/playbook.md`) is the kernel's procedural memory: itemized rules in ACE format (arXiv:2510.04618) with helpful/harmful counters.
## Loop
1. **Reflect** (after each F1+ task): read your own trace. What was decisive? What almost failed? Distill at most 3 lessons. Learn from BOTH successes and failures - failures yield preventative guardrails, often more valuable than success recipes (ReasoningBank evidence).
2. **Propose**: each lesson as a delta JSON entry:
- `{"op": "ADD", "section": "gates|verification|failure-recovery|...", "content": "<rule with Trigger and Test>"}`
- `{"op": "UPDATE", "id": "<bullet-id>", "helpful": 1}` or `"harmful": 1` when an existing bullet helped or misled.
3. **Gate before merge** (regression-awareness):
- no bullet without explicit trigger + test condition;
- no bullet contradicting an existing one (update that one instead);
- max 3 new bullets per task.
4. **Merge deterministically**: `python3 scripts/merge_deltas.py playbook/playbook.md deltas.json`. Never rewrite the playbook as a whole - monolithic rewrites cause context collapse.
5. **Retrieve lazily**: when a later task matches a section, load only that section's bullets and mark used bullets helpful/harmful after the task.
6. **Govern**: `python3 scripts/govern_playbook.py --apply` every ~15 tasks or above ~60 bullets. It retires harmful bullets (harmful >= helpful after >=5 trials), evicts lowest contributors beyond the cap, dedupes, and logs decisions to `playbook/decision-log.jsonl`.
## Hard constraints
- Never edit the playbook directly by hand; always via delta merge so counters stay truthful.
- Never add benchmark-specific answers or task-identifying shortcuts; rules must generalize behind their trigger condition.
- Held-out honesty: a rule promoted to the playbook has NOT been validated on unseen tasks until a fresh eval run confirms it; treat playbook content as UNVERIFIED until then.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 470k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |