@skill tool lets the agent grow its own capabilities: when it learns a reusable procedure, a project convention or a workflow you repeat, it writes a skill that auto-activates on its triggers in this and every future session. It’s hermes-agent’s living skill layer, native to ChatCLI’s skill format.
@memory vs @skill: @memory stores facts (“the user prefers tabs”); @skill stores a reusable procedure/knowledge with triggers (“how to deploy this project”). Short fact → memory; “how to do X” → skill. And to author skills from documentation, attach the corpus as a knowledge base: the agent researches with @knowledge and writes the skill with @skill.Subcommands
name is a validated kebab-case slug (no path traversal). create won’t clobber; update requires the skill to exist.
Evolution loop
skill-authoring skill guides the agent on when and how to author (capture a repeated workflow, good descriptions/triggers, evolve instead of duplicate).
Self-evolution
Skills used to appear only when the model remembered to call@skill — so you had to ask. Now they author and evolve themselves, riding the same memory-extraction pass (no extra LLM call): each turn the prompt carries only a compact skill index card; when the pass detects a new reusable procedure it authors the skill, and when an insight improves an existing one it reuses that exact name and pulls only that skill’s body on demand for an additive merge (preserving what already works).
Safety and reversibility:
- Every engine write is tracked by a content-hash manifest (like the built-in seed).
- A skill you authored by hand (or edited) is backed up once before the first modification —
@skill restore <name>undoes it and restores your original. - A backup failure downgrades to a suggestion; it never overwrites without a safety net.
CHATCLI_SELFEVOLVE_MODE:
Observability under
/config selfevolve (mode + how many skills the engine authored vs. the total); per-skill activation analytics in @skill stats.
Usage analytics
@skill stats shows the activation ranking (most used first, with last used) and flags authored skills that never activated — candidates to evolve or remove. Data lives in ~/.chatcli/skill-usage.json, written asynchronously by the manager on each activation (FindAutoActivatedSkills), never slowing a turn.
Shareable packs
SKILL.md (lossless round-trip). On import, existing skills are skipped unless overwrite:true. Useful for moving skills across machines and teams.
Safety
- Slug-validated name → no writes outside the skills directory.
- Never store secrets/tokens in a skill (that’s an environment/credential concern).
- Builtin skills seeding never clobbers agent-authored skills.