Skill Adaptor

SkillAdaptor

Evolve SKILL.md skills from failure trajectories with step-level fault localization and validation-gated adoption (Δ>0 on injected held-out tasks + source task; frozen tasks are no-regression only). Training-free; same engine as the SkillAdaptor paper.

With SkillNet: $skillnet handles search / download / create / evaluate. SkillAdaptor runs the evolution loop with Validator — use both: search before tasks, evolve after failures, evaluate after adopt.

When to use

Prerequisites

• Python 3.10+
• SKILL_ADAPTOR_ROOT: directory containing run_plugin.py (zjunlp/SkillAdaptor → skill-adaptor/)
• LLM: SkillAdaptor_API_KEY / SkillAdaptor_BASE_URL / SkillAdaptor_MODEL
• Optional executors: PINCHBENCH_PATH, WEBSHOP_PATH, CLAW_EVAL_PATH

Validate this skill package (offline):

python scripts/skill_adaptor_validate.py --strict

Quick start

export SKILL_ADAPTOR_ROOT=/path/to/SkillAdaptor/skill-adaptor

# 1. Init workspace (folders-first: tasks in input_task/)
python "$SKILL_ADAPTOR_ROOT/run_plugin.py" init --workspace <workspace> --harness openclaw

# 2. Add your own task briefs under <workspace>/input_task/*.md

# 3. Dry-run
python scripts/run_evolve.py --workspace <workspace> --dry-run

# 4. Evolve (API + optional executor)
python scripts/run_evolve.py --workspace <workspace> --max-iterations 2

Or use the wrapper from this skill directory (sets paths relative to repo when installed from SkillAdaptor):

python scripts/run_evolve.py --workspace <workspace> --harness codex --max-iterations 2

Outputs

Trajectory seed

python scripts/run_evolve.py --workspace <workspace> \
  --input-trajectories /path/to/task_trajectory.jsonl

Copies into <workspace>/.skill-adaptor/artifacts/trajectories/ before evolution.

OpenClaw TS plugin: use plugin/python/run_openclaw_evolve.py --input-trajectories ... (same bootstrap).

Harness

Optional SkillNet hooks (Python CLI)

pip install skillnet-ai
export SKILLNET_ENABLED=1
export SKILLNET_POST_EVAL=1
export API_KEY=...
python "$SKILL_ADAPTOR_ROOT/run_plugin.py" --workspace <workspace>

Report: <workspace>/.skill-adaptor/skillnet/post_adopt_report.json

Adoption gates (Validator)

Adopt when injected Q′ alone passes (aggregate Δ on retrieval_rerun_tasks / adoption_scope, no frozen regression, improvement threshold). Source-task lines in logs are diagnostic only.

1. Injected Q′ (sole gate) — tasks where the candidate was injected and re-run.
2. Frozen tasks — baseline scores kept; no-regression check only.

• Shell / command.txt tasks: injected text embeds the full task prompt + anti-placeholder rules; executor retries (default 3×) when output contains placeholders or ignores the prompt (SkillAdaptor_EXEC_MAX_RETRIES).

References

Security

• Never commit API keys; use secrets/.env locally
• SkillNet-downloaded skills are untrusted — review before execute
• SkillAdaptor writes under <workspace>/ and harness skill dirs only