SkillClaw
Keep as an important session-to-skill learning system and reference implementation. Do not globally install its current default proxy or allow it to overwrite canonical skills directly.
Routing Summary
Use SkillClaw to study or prototype how agent sessions can produce candidate skills: intercept model/tool traffic, preserve trajectories, summarize repeated strategies, deduplicate candidates, replay them, and publish reviewed versions. Use Controlled Skill Evolution - GEPA and SkillOpt and Skill Sync Workflow as the governing owners. SkillClaw may collect evidence and propose candidates; it does not own the evaluator or promotion decision.
Active deployment remains held until a scoped environment proves loopback-only binding, authentication, prompt/tool-result redaction, storage ACLs, immutable raw evidence, versioned proposals, held-out evaluation, human review, and rollback.
Current Source Snapshot
| Surface | Reviewed state |
|---|---|
| Repository | AMAP-ML/SkillClaw at e9c745466f142721977735e92a291619ac536ee0 |
| Package | 0.4.0, Python >=3.10, MIT |
| Reach | 2,400 stars and 239 forks in the 2026-08-10 API snapshot |
| Runtime | OpenAI/Anthropic-compatible proxy, Codex/Claude/Hermes integration, session store, evolve server, optional dashboard/replay/shared registry |
| Default sharing | Disabled, but optional local/S3/OSS shared storage can upload complete recorded sessions |
| Verification | 129 passed, 1 skipped, 1 dashboard test failed; test extras were incomplete |
| Preserved evidence | Frozen source, source digest, test receipt, and manifest under .brain/artifacts/x/2076581641267671487/ai-builder-repos/ |
Source: frozen repository, README/config/source, and capture manifest, reviewed 2026-08-10
What Is Valuable
SkillClaw separates several jobs that should remain distinct:
- Capture model requests, responses, tool calls, results, and session context.
- Extract repeated techniques and candidate procedural knowledge.
- Aggregate and deduplicate candidates across sessions or agents.
- Validate a candidate through replay or a task-specific evaluator.
- Review and publish a versioned skill to a registry.
- Distribute and activate an approved revision in target harnesses.
That shape is directly relevant to Kevin's brain: runtime memory becomes durable only after evidence-backed writeback, and recurring agent mistakes should improve executable procedures rather than create more disconnected notes.
Current Adoption Blockers
Network exposure
The proxy defaults to host 0.0.0.0 with an empty API key. On a laptop or shared network, that can expose captured agent traffic beyond the intended process. A safe local profile must bind to 127.0.0.1 or a Unix socket and require an unpredictable credential for every non-loopback path.
Captured data
Recorded sessions include prompts, model outputs, tool arguments, and full tool results. Those may contain source code, paths, secrets, customer data, private messages, or credentials. Shared storage cannot be enabled until redaction, classification, retention, deletion, encryption, and per-tenant access are proven.
Self-modification
Setup and activation can mutate Codex, Claude, and Hermes config/skill directories. Candidate quality cannot be inferred from recurrence or a model summary. Promotion must occur through a proposed diff, independent evaluation, held-out tasks, owner review, versioned registry entry, and reversible runtime sync.
Test signal
Most upstream tests passed, but one dashboard API contract test failed because the response lacked the expected status field. This is not a reason to discard the project; it is evidence that current operational surfaces still need pinned verification before adoption.
Kevin Integration Contract
| Stage | Kevin owner | Required proof |
|---|---|---|
| Session capture | runtime-specific adapter | scoped source, consent, redaction, immutable content hash |
| Candidate generation | SkillClaw/other proposer | evidence locators and bounded proposed diff |
| Optimization | Controlled Skill Evolution - GEPA and SkillOpt / SkillOpt route | task dataset, baseline, held-out evaluator, cost |
| Promotion | Skill Sync Workflow | independent pass, human review when authority changes, version and rollback |
| Projection | generated skill registry/runtime links | registry, resolver, trigger, config-sync, and doctor checks |
No candidate writes directly to a canonical SKILL.md. Rejected candidates and failure traces remain useful negative evidence.
Timeline
- 2026-08-10 | Captured and source-audited SkillClaw 0.4.0, ran its tests, kept the session-derived skill architecture, and held global installation on proxy, data, evaluation, and promotion boundaries. Source: X/@aashatwt; frozen source and capture manifest