تشغيل أي مهارة في Manus بنقرة واحدة

evolve-insights

النجوم٧

التفرعات١

آخر تحديث٢٢ أبريل ٢٠٢٦ في ١٣:١٠

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.

التثبيت

التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.

تشغيل في Manus

المصدر

engkimo

engkimo/open-morphic

فتح مستودع GitHub عرض مستودعات المنشئ

تنزيل

تشغيل في Manus

المهن ذات الصلةSOC

استنادا إلى تصنيف SOC المهني

مطوّرو البرمجياتمهن الحاسوب والرياضيات·SOC 15-1252

SKILL.md

readonly

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.

المزيد من هذا المستودع

نفس المستودع

constitution-check

engkimo/open-morphic

Verify a feature spec, plan, or PR against the non-negotiable principles in `.specify/memory/constitution.md`.

2026-04-227

agents-md-sync

engkimo/open-morphic

Keep CLAUDE.md (thin router) and AGENTS.md (telegraph-style) in sync. Detects drift in command lists, paths, and routing tables between the two root instruction files.

2026-04-227

cost-report

engkimo/open-morphic

Generate a monthly or on-demand cost report across LLM providers, local LLMs, and agent engines. Flags overspend and suggests downgrades.

2026-04-227

engine-e2e

engkimo/open-morphic

Run live end-to-end verification across Morphic-Agent's 6 agent CLI engines. Produces a baseline comparison table with duration and cost per engine.

2026-04-227

fractal-analyze

engkimo/open-morphic

Inspect Morphic-Agent's fractal task graph for a given task_id. Reports bypass classifier decisions, Gate 2 skips, parallelism, and regression flags.

2026-04-227

laee-dry-run

engkimo/open-morphic

Preview a LAEE action plan — classify risk, check undo availability, inspect credential touchpoints — without executing anything.

2026-04-227

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.