Track observed skill use and explicit outcomes, then inspect whether a derived routing hint is supported. The statistical analysis path runs locally without an additional LLM call. Exposure alone is not evidence that a skill worked.
Three stages — observe, analyze, evolve. The statistical path runs locally; optional providers and connectors retain their own network behavior.
When installed and supported by the client, hooks can capture selected tool events and context. Verify event coverage, truncation, redaction, storage, and runtime cost for the specific client.
SkillPerformanceMiner can build traces and outcome heuristics from available observations. Measure consolidation cost and distinguish explicit outcomes from inferred signals.
Configured soft prompts can surface bounded routing hints. Treat derived assertions as untrusted context until their evidence base and effect are verified.
Treat derived skill hints as evidence, not authority.
Verify the installed V4 runtime with slm --version, inspect the configured observation paths, and review the evidence before relying on a derived routing hint.
The backend is IDE-agnostic. Any client can POST tool events. The shipped hook currently supports Claude Code.
| IDE | Status | Integration |
|---|---|---|
| Claude Code | Supported | Auto-registered via slm init |
| Any IDE | API available | POST to /api/v3/tool-event |
| Cursor | Planned | Adapter in development |
| Windsurf | Planned | Adapter in development |
| VS Code / JetBrains | Planned | Extension adapter |
Everything Claude Code (ECC) is an external project whose observation files can be imported by the documented SLM ingest path. Its behavior, license, and compatibility remain independently versioned.
slm ingest --source ecc can import supported ECC observation files. Preview with --dry-run, inspect counts and omissions, and do not assume every ECC artifact is covered.
ECC is optional. The importer reads external data — review consent, provenance, redaction, and retention before writing it into SLM.
These external references are design context, not proof that the current SLM package reproduces their results.
External preprint; inspect its protocol and scope before transferring any result to SLM.
arXiv:2604.01687External repository; inspect its current code, license, and release evidence directly.
github.com/HKUDS/OpenSpaceExternal preprint; its benchmark does not establish current SLM product performance.
arXiv:2602.12670External survey; use its taxonomy and security findings only within the paper's stated protocol.
arXiv:2602.12430Install SLM V4, enable only the observation paths you consent to, and verify signal quality before changing routing.
Open source under AGPL v3 · A Qualixar Research Initiative