mycelium-hq/ai-brain-starter/evolve

/evolve — promote a cluster of instincts into a structure

`{SKILL_DIR}` = this skill's own folder (locally: the directory this SKILL.md lives in; a served brain substitutes the real absolute path before you read this). Shared starter files live at the repo root two levels up: {SKILL_DIR}/../... If a path does not resolve, name the missing file and stop — never guess another location.

When many related, high-confidence instincts pile up in one domain, that is a signal to promote them into ONE reusable structure instead of leaving them as loose memories. /evolve finds those clusters deterministically and drafts a proposal you refine.

ECC source pattern: "/evolve clusters related instincts into higher-level structures: Commands / Skills / Agents." Reimplemented clean per license-hygiene.

Step 1 — run the clusterer (deterministic, zero LLM cost)

python3 "{SKILL_DIR}/../../scripts/instinct.py" evolve

It groups every instinct by inferred domain, computes each cluster's median effective confidence, and for clusters that clear the propose bar (>= 2 instincts AND median confidence >= 0.80) writes a scaffold to <vault>/⚙️ Meta/Instinct Proposals/proposed-skill-<domain>.md. Clusters below the bar print as watch (not yet ripe).

Step 2 — judge each PROPOSE cluster

For each proposal the script wrote, decide the structure:

• Command — the cluster is ONE repeatable procedure with a clear trigger.
• Skill — the cluster is a coherent body of domain guidance (most common).
• Agent — the cluster describes an autonomous multi-step workflow.

A cluster that is really just 2-3 facets of an existing rule should be CONSOLIDATED into that rule, not promoted. Reject those.

Step 3 — draft (only on confirm)

For an accepted cluster, draft the real skill/command body from the member instincts (keep each instinct's Action + Evidence). Place it in the right repo per the install rules, wire its discoverability + automation + verification in the SAME session (the three-layer wiring rule), then retire or link the source memories. Delete the proposal file once adopted or rejected — proposals are scratch, not a backlog.

Headless/auto mode

In a cron/--print session: run Step 1, report the PROPOSE clusters and the proposal paths, and STOP. Promotion to a real skill is a judgment call that needs a human — never auto-create skills.