Skill Evolver

Skill evolution system. Analyzes agent execution traces to generate Evolution Units (EUs — exploit/explore subtypes under the adaptive type), deploys them to SKILL.md via evidence-based scoring, making agents learn from their own experience. Trigger: 'run skill-evolver' / 'run skill evolution' / 'evolver status' / 'skill patch'

Skills should adapt to you — and prove it.

Watches how you actually use each skill, identifies where it falls short or succeeds, attaches those lessons to the skill, and tracks whether each lesson holds up over time.

CLI: node scripts/evolver-cli.js <command>

All paths below are relative to this skill's directory.

---

When to Use

First-time setup: User explicitly requests skill self-improvement, e.g. "run skill-evolver", "run skill evolution", "evolver status", or "skill patch".

Ongoing (lightweight): Once initialized, evolution runs automatically. Direct interaction only for:

• Status queries, unit review, pin/evict/feedback/clear/block
• Pause/resume, config changes, manual runs

Most of the time, skill-evolver works silently. Don't proactively mention it unless asked.

Boundaries

• Reads: session JSONL files, existing SKILL.md and Evolution Unit (EU) files
• Writes: EU .md files in eu/, inline blocks in SKILL.md, state/config JSON, evidence/trace history files
• Executes: Node.js pipeline scripts (generator.js, validator.js, reviewer.js, monitor.js) via OS cron or direct invocation
• Network: trace and unit/skill text are sent only to the LLM configured for generation/validation/review — either your own provider (evolver-config.json) or OpenClaw's subagent channel, which reuses the agent's existing model with no separate API key (auto-detected by default). No telemetry, no hidden endpoints. Set privacy.allowRemoteLLM: false to disable all outbound calls. Optional http(s) webhook for notifications.
• Scheduling: optional OS crontab entry for periodic runs; remove it to stop.
• Does NOT: modify skill source code, monitor calls in real-time, make agent decisions, or run user-supplied shell commands.
• Safety: worst case is a bad inline block — evict with one command.

---

Initialization Flow

When no evolver-state.json exists:

Step 1: Collect User Preferences (ask the user before apply)

Three configuration decisions (all have defaults):

Mode (default: continuous)

• continuous: monitor every N hours, auto-trigger when enough new skill invocations accumulate
• scheduled: run at fixed schedule (user specifies frequency)
• manual: only on explicit request

Notification level (default: brief)

• off: no push notifications, results logged to file only
• brief: periodic natural-language summary of what was learned (requires LLM + delivery)
• verbose: brief reports + immediate push of raw evolution results after each run

If brief or verbose: frequency (default: 1d) — how often reports are sent ("1d", "2d", "7d")

Push channel (default only for openclaw: current session)

• Which channel to deliver notifications through (e.g. wecom, discord)
• Platforms without channels: push via webhook URL or shell command configured in evolver-config.json

Step 2: Run Setup

Pass user choices to setup.js:

bash

node scripts/setup.js --mode <mode> --notify <off|brief|verbose> [--notify-frequency 1d] [--notify-channel wecom] [--interval 4h]

All flags optional — omitted flags use defaults. setup.js detects the environment, writes config, installs scheduling, and outputs JSON to stdout.

Review the output before confirming to user. Key things to check:

• checks.sessions.reason — explains what was found and where. If sample is empty but user has conversation history, the detected path may be wrong.
• checks.llm.detail — how LLM was resolved. "subagent" tier works but is lower quality; worth mentioning.
• checks.notifications.hookAvailable — whether push was auto-configured. If false, notifications go to log file only.
• blockers[] — must-fix issues. Help user resolve, then re-run.
• warnings[] — non-blocking issues worth mentioning.

Step 3: Optional Bootstrap

If setup output shows session files were found, a one-time bootstrap can seed initial learnings:

bash

bash scripts/bootstrap.sh --window-days 7

Not required — the system works from zero. If stdout contains <brief_data>, deliver a brief report to the user using brief-prompt.md as the prompt template.

Diagnostics

bash

node scripts/preflight.js          # detailed per-component health check
node scripts/preflight.js --fix    # auto-create missing dirs
node scripts/schedule.js status    # scheduling state + full detection audit

---

Runtime Commands

Manual Evolution Run

bash

bash scripts/bootstrap.sh

Runs one evolution cycle (incremental from last evolution) and outputs <brief_data> to stdout. If output is non-empty, deliver a brief report using brief-prompt.md as the prompt template.

Options:

• --max-si 20 — more traces per run
• --window-days 14 — look further back
• --filter-skill <name> — only process SIs for one skill (targeted run)
• --hint "<text>" — prepend a guidance block to each SI trace; the generator's LLM uses it as steering when extracting Evolution Units. The agent is expected to translate the user's verbal intent into a concise hint before passing it in.

Targeted manual run (e.g. user just finished a wecom-doc workflow and wants the lesson codified immediately, with a steering hint):

bash

bash scripts/bootstrap.sh --filter-skill wecom-doc \
  --hint "User wants the MCP path to become the default; cookie path is fallback only."

For raw pipeline output without brief formatting:

bash

bash scripts/run-daily.sh --state evolver-state.json

Then read eu/evolution-report.json for structured results.

Query Status

bash

node scripts/evolver-cli.js status

Returns: { mode, paused, lastEvolution, totalEvolutions, pendingSI, units: { total, byTrack, topSkills } }. Summarize for user: current mode, how many units are active, when evolution last ran, how many pending SIs.

List / Explain Units

bash

node scripts/evolver-cli.js units [skill-name]
node scripts/evolver-cli.js explain <unit-keyword>

(patches is a deprecated alias for units and works identically for one release cycle.)

Pin / Unpin / Evict

bash

node scripts/evolver-cli.js pin <unit-keyword>
node scripts/evolver-cli.js unpin <unit-keyword>
node scripts/evolver-cli.js evict <unit-keyword>

After evicting, re-run lifecycle to update SKILL.md:

bash

node scripts/lifecycle.js --eu-dir eu --evidence /dev/null --traces /dev/null

User Feedback

bash

node scripts/evolver-cli.js feedback <unit-keyword> good|bad

Clear / Block (per-skill user control)

For when you want a skill left alone. clear detaches a skill's learned units from its SKILL.md without deleting them (the units stay archived on disk and can be restored later). block additionally excludes the skill from all future evolution.

bash

# Detach one skill's units from its SKILL.md (archived, not deleted).
# NOTE: re-attached on the next evolution run unless you also block it.
node scripts/evolver-cli.js clear <skill>

# Detach ALL skills' units at once.
node scripts/evolver-cli.js clear --all

# Detach AND block in one step (stays off).
node scripts/evolver-cli.js clear <skill> --block

# Block a skill: detach now + skip it in all future evolution runs.
node scripts/evolver-cli.js block <skill>

# Re-allow a blocked skill (its archived units re-attach on the next run).
node scripts/evolver-cli.js unblock <skill>

# Show blocked skills.
node scripts/evolver-cli.js blocklist

Blocked skills are recorded in evolver-config.json under blockedSkills (hand-editable). Blocking never deletes lessons — it only stops them from being attached and stops new ones from being generated. Trace extraction skips blocked skills upstream, so no LLM cost is spent on them.

Pause / Resume / Config

bash

node scripts/evolver-cli.js pause
node scripts/evolver-cli.js resume
node scripts/evolver-cli.js config <key> <value>
node scripts/evolver-cli.js history --limit 5

---

Unit Matching

Users don't use filenames. Match their intent:

• If units were just displayed → match by position ("the second one")
• Skill + keyword → match against title/condition
• Ambiguous → show candidates and ask

Unit Types & Scoring

Evolution Units currently all live under type adaptive, with two subtypes:

• Exploit (✅): Proven strategy — "when X, do Y"
• Explore (⚠️): Dead end warning — "when X, don't do Y"
• Pinned (📌): User-protected, always active

New units start in Nursery (score capped) until positive evidence promotes them. Top 3 per skill are inlined to SKILL.md. Negative evidence → evict. Explore units retire when a matching exploit covers the same scenario.

---

Continuous Mode Behavior

monitor.js runs via OS crontab (default every 4h):

1. First run: initializes state, exits. First evolution is user-initiated (Step 3).
2. Counts new SIs since last evolution. If ≥ threshold (default 5) or maxInterval elapsed: triggers run-daily.sh.
3. Processes up to 20 SIs per batch. Remaining carry over to next cron cycle.
4. After evolution with new units, writes pending_brief to evolver-state.json.

---

Brief Report Delivery

Evolution results are written to evolver-state.json field pending_brief. To deliver:

• If hook is configured: already auto-triggered. No action needed.
• If platform supports recurring agent tasks: create a recurring task at notifications.frequency interval. Task: check pending_brief in evolver-state.json → if present, format with brief-prompt.md → deliver to user → clear field.
• Otherwise: check pending_brief on any evolver-related interaction and deliver if present.

---

File Layout

text

scripts/
  setup.js               One-shot initialization wizard
  preflight.js           Detailed health check (diagnostics)
  run-daily.sh           Pipeline runner (extract → generate → validate → patch)
  monitor.js             Continuous mode trigger + notification hooks
  schedule.js            OS crontab installer
  evolver-cli.js         Query & control CLI
  trace-extractor.js     Session → trace extraction
  generator.js           LLM: trace → Evolution Units
  validator.js           LLM: evidence attribution
  lifecycle.js           Scoring + lifecycle + SKILL.md inline writeback
                         (formerly patcher.js, renamed in v2.4.4)
  lib/
    llm-client.js        Shared LLM client (multi-platform)
    platform-detect.js   Platform detection + adapter resolution
  adapters/
    base.js              PlatformAdapter base class
    openclaw.js          OpenClaw adapter
    claude-code.js       Claude Code adapter
  tests/                 Adapter fixture tests
evolver-config.json      Config (optional, overrides auto-detection)
                         Key fields: `llm` (endpoint/apiKey/model), `skillsPaths` (array of
                         absolute paths to skills root dirs — reviewer uses this to locate
                         SKILL.md descriptions; auto-inferred from install path if omitted)
evolver-state.json       State
eu/<skill>/              Evolution Unit storage (files are eu-YYYYMMDD-<slug>.md;
                         directory was renamed from patches/ in v2.4.0)
CHANGELOG.md             Version history