self-improve — ChaosEngine learning & adapting

Lean CE-native skill informed by Task Observer methodology (Eoghan Henn / rebelytics, CC BY 4.0 — see LICENSE and UPSTREAM.md). Not a blind clone.

When

• Primary: root-owned Learning Session after confirmed delivery.
• Secondary: explicit operator request mid-session.
• Not: every casual turn — keep always-on cost low.

Dual track

1. Harness — skills, hooks, MemPalace, Graphify, Headroom, installer/doctor.
2. Product — enhancements for the product under development (queued issues).

Details: references/observation-taxonomy.md. Activation: references/activation.md. Adopt/reject research: references/research-adopt-reject.md.

How (wraps learning.py)

1. Classify each finding (harness vs product; category allow-list).
2. Write minimal fields only: category, title, lesson, proposedChange, benefit, estimatedTokens.
3. Queue through learning.py so privacy gates + GitHub filing invariants hold.
4. Never auto-install skill patches; stage proposals for human/CI review.
5. "Nothing durable" is a valid outcome.

Example:

cat > /tmp/learning-candidate.json <<'EOF'
{
  "category": "tooling",
  "title": "Doctor headroom fix-next missing",
  "lesson": "Operators lacked a single install command after pin landed",
  "proposedChange": "Surface uv tool install pin in doctor fix-next",
  "benefit": "Faster Headroom provisioning on adopter hosts",
  "estimatedTokens": 120
}
EOF
python3 .chaos-engine/learning.py queue \
  --state .chaos-engine-state/learning \
  --upstream Owner/ExampleRepo \
  --candidate /tmp/learning-candidate.json

Local smoke

# From a temp project with ChaosEngine installed, or the source tree:
python3 -c "from pathlib import Path; assert Path('chaos-engine/skills/self-improve/SKILL.md').is_file()"
# Queue one harness + one product candidate (privacy-safe fixtures) via learning.py
# then confirm queue.json grew by two items without secrets/paths.