compound-apply-spec

Why this exists

The plugin is deterministic. The agent is not. So we separate:

• Agent: decides what to learn (writes the spec).
• Tool (compound_apply): validates and applies changes safely.

CompoundSpec v2

Top-level keys (all optional except schema_version):

• schema_version: must be 2
• auto: { reason, sessionID }

instincts

• create[]: { id, title, trigger, action, confidence }
• update[]: { id, confidence_delta, evidence_note }

skills

• create[]: { name, description, body }
• update[]: { name, description?, body }

Notes:

• name must match ^[a-z][a-z0-9]*(?:-[a-z0-9]+)*$.
• body is markdown without frontmatter.
• If updating a skill: skills.update[].body must be the entire final managed body (no snippets/diffs).

docs

• sync: boolean (refresh derived indexes/blocks)
• blocks.upsert[]: { file, id, content }

Only upsert AI-managed blocks; do not rewrite human-owned text.

When writing markdown content inside docs.blocks.upsert[].content:

• Use repo-root-relative paths when referencing files/dirs (no absolute paths).

changelog

• { note }: a single sentence describing the memory delta.

Apply

1. Draft the CompoundSpec v2 as a single JSON object.
2. Output JSON only (no code fences, no commentary).
3. Run compound_apply() to validate and write the memory/doc updates.

Manual notes

This section is preserved when the skill is updated. Put human notes, caveats, and exceptions here.