Uv Skill Evolution Manager

Skill evolution manager (pkbllm)

This skill helps you continuously improve skills based on real sessions, without losing fixes when skills are updated.

It does this by:

1. extracting structured learnings into evolution.json, and
2. stitching those learnings into a bounded section in SKILL.md (idempotent re-apply).

This is adapted from KKKKhazix/Khazix-Skills (MIT) and made pkbllm-compatible.

Canonical vs installed (pkbllm rules)

• Canonical source of truth (PKB_PATH) lives under: bootstrap/, common/, human/, knowledge/, productivity/.
• skills/ is generated. Never edit it by hand.
• Local skill installs (e.g. a project .agents/skills or ~/.codex/skills) are not canonical and are often gitignored.

If you evolve canonical skills, regenerate mirrors:

python bootstrap/scripts/update_skills_mirror.py all

Non-negotiables (requirements)

• Evidence-first: only add learnings supported by the session (errors, user feedback, successful commands).
• No arbitrary prompt injection: store structured constraints, pitfalls, and verification steps (see schema).
• Idempotent: repeated stitching must not duplicate content.
• No “install for all agents”: update existing local installs if found; do not install globally unless explicitly requested.
• Never edit skills/ by hand: only update canonical + regenerate.

Data model (evolution.json)

Each evolved skill keeps a sibling file:

<skill_dir>/
├── SKILL.md
└── evolution.json

Schema (see assets/evolution.schema.json):

• preferences: user preferences / defaults
• fixes: concrete fixes (platform quirks, known issues)
• pitfalls: common mistakes to avoid
• verification: commands/checks to run before claiming success
• examples: short “worked examples” (command + expected snippet)

Workflow (the evolution loop)

1) Review the session

For each invoked skill:

• What worked?
• What failed? (exact error)
• What did the user request as a preference?
• What verification would have caught the issue earlier?

2) Extract structured learnings

Produce a JSON object matching the schema. Keep each item short, specific, and testable when possible.

3) Apply + stitch (canonical and/or local)

Apply to a single skill name:

python scripts/apply_evolution.py --skill-name uv-hands-on-learning --scope both --json '{"pitfalls":["Always keep raw logs under results/ (gitignored) and copy only small excerpts into evidence/."], "verification":["Confirm no tracked files under any hands_on/**/results/." ]}'

Scopes:

• pkb: update canonical skill under PKB_PATH (or inferred repo root)
• local: update locally installed copies (default: project scope + ~/.codex/skills)
• both: do both

4) Align after updates

When skills are updated/refactored, re-stitch to re-apply the learned section:

python scripts/align_all.py --scope pkb
python scripts/align_all.py --scope local

Review policy (must follow)

Before saying “evolution applied”:

• The target SKILL.md contains a single Learned section bounded by markers.
• evolution.json is valid JSON and contains only schema fields.
• Canonical changes are mirrored by running python bootstrap/scripts/update_skills_mirror.py all.
• Local installs were updated only if found; otherwise, report “no local installs found”.

Checklist: references/review.md

Common pitfalls

• Editing skills/ directly instead of canonical.
• Adding vague learnings (“be careful”, “run tests”) rather than concrete items.
• Adding session-specific secrets or environment-specific paths into canonical skills.
• Forgetting to re-stitch after a skill update (learned section disappears).

Expanded list: references/pitfalls.md

Scripts

• scripts/apply_evolution.py: apply a JSON delta to pkb, local, or both scopes.
• scripts/merge_evolution.py: merge JSON delta into evolution.json (dedupe, stable).
• scripts/smart_stitch.py: update/insert the bounded Learned section in SKILL.md.
• scripts/align_all.py: re-stitch all skills that have an evolution.json.

References / assets

• Schema: assets/evolution.schema.json
• Example: assets/evolution.example.json
• Learned section format: assets/learned_section_template.md

Converted and distributed by TomeVault — claim your Tome and manage your conversions.