/learn

/learn

You are the epistemic learning agent. When the user runs /learn, execute these four phases in order.

Phase 1: Save the Session

Write a session summary to knowledge/session_log/. Create a file named with the current timestamp (YYYYMMDDTHHmmSS.jsonl).

Write one JSON line per event. Include:

{"event":"session_start","timestamp":"...","project":"..."}
{"event":"agent_invoked","agent":"spring-engineer","trigger":"tuning spring physics on modal"}
{"event":"claim_consulted","claim_id":"claim-007","relevance_score":0.85}
{"event":"suggestion","suggestion_id":"sug-001","file":"components/AnimatedDrawer.tsx","claim_refs":["claim-007"]}
{"event":"suggestion_outcome","suggestion_id":"sug-001","outcome":"accepted"}
{"event":"candidate_claim","description":"Spring stiffness 300 + damping 25 feels natural for slide-in drawers across projects"}
{"event":"session_end","timestamp":"...","duration_minutes":25,"files_changed":["components/AnimatedDrawer.tsx"]}

Reconstruct this from your memory of the session. Include:

• Which agents you loaded
• Which knowledge claims you consulted (if any)
• What suggestions you made and whether they were accepted
• What files were changed
• Any patterns you observed that are not in the knowledge base (log these as candidate_claim events)

Phase 2: Learn

Run the fast learning script:

python -m scripts.epistemic.learn --plugin animation-pro

Read the JSON output from stdout. This is the review queue.

Phase 3: Review

Present the review queue items to the user. Handle each type:

Confidence Changes

Report them. No action needed from the user.

Format:

Knowledge updates:
  claim-007 strengthened: 0.71 -> 0.79 (suggestion accepted)
  claim-012 weakened: 0.65 -> 0.52 (suggestion rejected)

New Tensions

Show the tension and ask the user to resolve it.

Format:

Tension detected:
  "claim-003 says use spring physics for all modals, but this session
   used CSS transitions for the same component type"
  Related claims: claim-003, claim-018

  Options:
  1. Update claim-003
  2. Keep both (context-dependent)
  3. Dismiss (one-off exception)

Wait for the user's response. If they choose to update a claim, edit the claim in knowledge/claims.jsonl.

Candidate Claims

Show proposed claims and ask if they should be added.

Format:

Observed pattern:
  "Spring stiffness 300 + damping 25 feels natural for slide-in drawers across projects"
  Add to knowledge? [yes / no / edit first]

If yes, append to knowledge/claims.jsonl as a new claim with status "active" and default confidence (0.67).

Attention Needed

Show claims that need human judgment.

Format:

Needs attention:
  claim-012 (confidence 0.38): "Always use layoutId for shared element transitions"
  This claim has been rejected more often than accepted.
  Options: [keep / retire / edit]

If retire, set the claim's status to "retired" in claims.jsonl.

Phase 4: Report

Print a one-line summary at the end:

[animation-pro]: N active claims, avg confidence X.XX, M items reviewed

If there were no review items, just print:

[animation-pro]: N active claims, avg confidence X.XX, session saved