self-improving-agent

name	self-improving-agent
description	Capture corrections, errors, feature requests, and recurring best practices in a local `.learnings/` workspace so future agent runs can avoid repeating mistakes. Use when a command fails, a user corrects the agent, a missing capability is requested, a better repeatable approach is discovered, or before major work to review prior lessons.

Self-Improving Agent

Use this skill to keep a lightweight local learning loop for Wee Orchestrator work and related repos.

Quick Start

Initialize a learning workspace in the current project root:

./scripts/init_learnings.sh

That creates:

.learnings/
├── LEARNINGS.md
├── ERRORS.md
└── FEATURE_REQUESTS.md

Do not overwrite existing files. Keep entries short, specific, and sanitized.

Wee Orchestrator Adaptation

This skill is adapted for Wee Orchestrator rather than OpenClaw.

• Use .learnings/ in the repo or agent workspace you are actively changing.
• For Wee Orchestrator dev work, prefer /opt/n8n-copilot-shim-dev/.learnings/ on the dev host.
• For local agent repos like /opt/wee-dev, /opt/wee-qa, or /opt/skills/<skill-name>, keep .learnings/ inside that repo.
• Do not store secrets, tokens, raw environment files, or full logs. Store summaries or redacted excerpts.

When to Log

Situation	Destination
Unexpected command or tool failure	.learnings/ERRORS.md
User correction or outdated assumption	.learnings/LEARNINGS.md
Missing capability request	.learnings/FEATURE_REQUESTS.md
Better recurring implementation pattern	.learnings/LEARNINGS.md
Repeated bug pattern from multiple tasks	Update existing learning and increase recurrence metadata

Before Major Work

Before a substantial task, scan local learnings:

grep -n "^## \[" .learnings/*.md 2>/dev/null

If you are working in a familiar area, also search by topic:

grep -Rni "auth\|scheduler\|queue\|frontend" .learnings/ 2>/dev/null

Promotion Targets

When a learning is stable and broadly useful, promote it out of .learnings/:

• Project conventions → CLAUDE.md
• Agent workflow guidance → AGENTS.md
• Copilot-specific reminders → .github/copilot-instructions.md
• Cross-session durable lessons → Total Recall agent_lessons
• Repository-wide implementation facts → repo memory/store-memory workflow

For Wee Orchestrator specifically, the most common promotion targets are /opt/AGENTS.md, /opt/CLAUDE.md, and the relevant repo-local AGENTS.md.

Entry Guidelines

Use IDs like LRN-YYYYMMDD-001, ERR-YYYYMMDD-001, and FEAT-YYYYMMDD-001.

Always include:

• timestamp
• priority
• status
• area
• concise summary
• concrete suggested action
• related files when known

Prefer updating an existing entry over duplicating it. Link related items with See Also.

Status Values

Use these statuses consistently:

• pending
• in_progress
• resolved
• wont_fix
• promoted
• promoted_to_skill

Hook / Reminder Support

This skill includes small helper scripts:

• scripts/activator.sh — reminder at prompt start
• scripts/error-detector.sh — reminder after shell/tool failures

For Claude-style environments with hooks, see references/wee-orchestrator-integration.md. For Copilot CLI, add a short reminder to AGENTS.md or copilot-instructions.md instead of relying on hooks.

Resources

• references/examples.md — example entries for each log type
• references/wee-orchestrator-integration.md — setup guidance for local agent repos
• references/upstream.md — upstream source references and adaptation notes
• assets/ — starter templates for .learnings/ files