name: compound-autolearn-compoundspec description: Generate a CompoundSpec v2 JSON-only autolearn proposal from git diffstat + existing skills, while respecting strict file-scope constraints. license: MIT compatibility: opencode,claude metadata: created_at: "2026-01-30T16:42:50.393Z" updated_at: "2026-01-30T16:42:50.393Z" version: "1"

Purpose

Turn an autolearn prompt (recent activity + constraints) into a valid CompoundSpec v2 JSON object.

When To Use

• Prompt says "Background Autolearn" or "learning agent".
• Prompt requires JSON-only output and limits edits to memory/docs artifacts.

Procedure

• Read the constraints and allowed paths; assume anything else is forbidden.
• If a system reminder indicates read-only / Plan Mode:
  • Do not propose product code changes.
  • Keep proposals strictly within skills/instincts/docs/changelog.
  • Do not propose actions that imply file modifications outside the allowed set.
• Use the git summary (changed_files + diffstat) as evidence for what to learn.
• If the git summary indicates an in-progress merge/conflict state (e.g., diffstat lines labeled "Unmerged <path>"):
  • Treat the diff as incomplete/low-signal.
  • Avoid inferring product behavior changes.
  • Prefer a small instinct/skill update about conflict-state hygiene.
  • If there is no other stable evidence, emit empty skill/instinct changes and set docs.sync: false.
• If the diffstat is empty (or there are no meaningful file changes):
  • Do not invent learnings.
  • Emit empty instincts.create/update and skills.create/update.
  • Set docs.sync: false.
  • Use a changelog note that explains the skip (e.g., "Autolearn ran with no diff evidence; skipped memory updates.")

Evidence hygiene (low-signal diffs)

Treat generated/derived artifacts as low-signal evidence:

• If services/index.json changes without corresponding services/*.md edits, assume it was refreshed and do not infer new dependency learnings.
• If .opencode/memory/instincts.json / .opencode/memory/INSTINCTS.md show large rewrites or deletions, assume cleanup and avoid inventing new heuristics from it.
• If changes appear duplicated under both .claude/skills/ and .opencode/skills/, assume a mirror/sync artifact:
  • Do not infer new behavior from the duplication.
  • In CompoundSpec proposals, only target .opencode/skills/** (never propose changes to .claude/skills/**).
• If the diff is primarily additions/deletions under src/*.egg-info/, assume packaging metadata cleanup/regeneration noise and avoid learning anything beyond "egg-info is generated" unless the prompt explicitly states an intentional packaging change.
• If the diffstat shows changes only under examples/ with no corresponding src/ or tests/ changes, treat it as low-signal for durable learnings; emit empty skills/instincts, set docs.sync: false, and use a changelog note explaining the skip.

Ticket/process-heavy diffs

• If the diff is primarily ticket/process artifacts (e.g. .tickets/*.md, LOOM_ROADMAP.md, LOOM_CHANGELOG.md, AGENTS.md) with no corresponding product code changes:
  • Prefer instincts.update[] that strengthen workflow heuristics (triage, scoping, read-only constraints).
  • Avoid creating new skills unless a repeated procedural gap is clearly demonstrated.
  • Avoid proposing docs block updates unless a stable always-on principle changed.
  • Do not infer or describe product behavior changes.

Prefer updates over creation

Prefer:

• instincts.update[] to strengthen an existing heuristic
• skills.update[] to refine an existing skill
• skills.create[] only if there is no close match

Skill updates

• If updating a skill:
  • Re-emit the entire final managed body (no diffs/snippets).
  • Keep it checklist-like.

Instinct quality bar

• Keep instincts crisp:
  • Trigger is a concrete situation.
  • Action is a concrete behavior.
  • Confidence is 0.6-0.85 unless repeated evidence.

Keep proposals small

• Max 3 skills per run.
• Max 8 instinct updates per run.

Output hygiene

• Output exactly one JSON object.
• Do not wrap in code fences.
• Do not include commentary.
• Use repo-root-relative paths in any markdown content.

Suggested Minimal Template

• auto.reason from prompt (often session.idle).
• auto.sessionID from prompt.
• docs.sync: true if you touched skills/instincts/docs; otherwise docs.sync: false.
• changelog.note as a single sentence describing the memory delta.

Manual notes

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