AGENT SKILL

Autoskill (workflow-to-skill miner)

暂无简介

License 许可
MIT
所属仓库
rhCat/skillChip
最近更新
2026-07-31
Stars
★ 0
Forks
0

01 前言 Front Matter

name
Autoskill (workflow-to-skill miner)
perks
[ "doctor", "fetch", "redact", "cluster", "match", "synthesize", "promote" ]
skill
autoskill

02 资源文件 56

03 SKILL.md 正文

autoskill — Autoskill (workflow-to-skill miner)

Mine the user's own local screenpipe workflow history into proposed scientific skills. The pipeline fetches a time window from the local screenpipe daemon, redacts PII, clusters repeated workflows, matches each cluster against the sibling skill library, asks an LLM to classify reuse/compose/novel, and stages drafts the user can review and promote. Each operation is a governed perk with explicit inputs and outputs.

What to look out for

Each tool emits one line of structured JSON (the audit + debug log) and writes its artifacts under record_store. LOGS TO CHECK: that line + the named output file + the executor run-ledger. Detection runs locally; only redacted cluster summaries reach the LLM. The fetch, match, and synthesize perks reach out (screenpipe loopback / heavy embedding model / LLM backend); their porters degrade gracefully and still write a well-formed output when the dependency is absent offline.

Perks

perk tool nature
doctor doctor read-only preflight (config / skills_dir / screenpipe / llm)
fetch fetch_window read-only — paginates screenpipe /search (localhost) into a timeline JSON
redact redact read-only / pure — scrubs emails, keys, bearer tokens, JWTs, phones, SSNs
cluster cluster read-only / pure — segment sessions on idle gaps + cluster by app-signature
match match_skills read-only — top-k cosine match of clusters vs sibling SKILL.md descriptions
synthesize synthesize read-only — LLM judge: reuse / compose / novel + drafts a SKILL.md body
promote promote local file move — stages an approved proposal into skills/<name>/

redact and cluster are pure stdlib transforms (fully hermetic). fetch talks to the local screenpipe daemon, match loads sentence-transformers, and synthesize calls the configured LLM backend — each declared destructive: false (read-only / local) and each porter writes a valid output even when its dependency is unavailable. promote only moves a local directory; it refuses to overwrite an existing skill.

How to use it

Pick a perk, copy ledger.json → task-ledger.json, fill its vars + record_store, then validate → compose → compile → oversight → executor.

Localized from K-Dense-AI/scientific-agent-skills autoskill — MIT (see LICENSE.txt).

04 同仓库更多 Skills