| name | evolve-insights |
| description | Extract session insights after a task completes and feed them into Morphic's self-evolution engine. Captures what worked, what didn't, and proposes prompt/routing tweaks. |
| when_to_use | After completing a non-trivial task. Also on `/loop` boundary to run evolution continuously. |
| argument-hint | [session_id | last] |
| allowed-tools | ["Read","Write","Edit","Grep","Glob","Bash(morphic session export *)"] |
| model | sonnet |
Evolve Insights
You harvest learnings from a completed session and route them to the evolution engine.
Input
$ARGUMENTS = session id (default: last).
Procedure
- Export the session timeline:
morphic session export <id> --format json > /tmp/session.json
- Delegate analysis to
harness-optimizer subagent.
- The subagent classifies each step into:
- WIN: went faster/cheaper than baseline. Capture the pattern.
- LOSS: regression, retry, or extra-cost path.
- SURPRISE: unexpected behavior (good or bad).
- For each WIN:
- Distill to a portable rule (model/prompt/route hint).
- Store in
.morphic/evolution/wins.jsonl as an append-only record.
- For each LOSS:
- Identify the proximate cause (prompt, model, tool, classifier).
- Draft a counter-measure (prompt patch, routing weight shift, new tool).
- Store in
.morphic/evolution/losses.jsonl.
- Output a session postmortem.
- If ≥3 losses of the same kind, propose a harness change and stop — user decides whether to apply.
Output
# Evolve Insights — session <id>
## Wins (N)
- W-01 bypass classifier skipped 3 subtasks correctly, saved $0.04
- W-02 Gemini CLI beat Claude on structured extraction, 2.3s vs 4.1s
## Losses (N)
- L-01 fractal planner over-split a 2+2 task into 5 nodes (TD-167 regression)
- L-02 OpenHands retry after temperature error — use Gemini engine instead
## Surprises (N)
- S-01 ollama qwen3:8b matched Sonnet quality on JSON extraction
## Proposed harness changes
1. Tighten bypass threshold for arithmetic tasks (+confidence weight 0.2)
2. Blacklist OpenHands+Claude until OH upgrade
## Recommendation
Apply #1 immediately (low risk, reversible). Defer #2 until next release.
Guardrails
- Never auto-apply harness changes. Always require user approval.
- Preserve raw session data — evolution entries are derivative, not replacements.
- If ≥3 losses share a root cause, escalate to user before adding more detections.
