redo-reflection

name	redo-reflection
description	Use when mining Hermes/OpenClaw/agent conversation history for automation opportunities, recurring workflows, cron candidates, skill candidates, wiki/archive patterns, proactive questions, or low-value repeated work to rethink. Proposal-only by default.
version	0.1.0
author	Simon Meng + Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["automation","reflection","redo","hermes","openclaw","skills","cron"],"related_skills":["hermes-agent","structured-task-runner"]}}

Redo Reflection

Overview

Redo Reflection is a thin Redo layer for Agent systems. Instead of watching the whole desktop, it starts from the highest-signal surface: the user's conversations with an Agent and the Agent's resulting actions.

It detects opportunities to turn repeated requests into:

• cron jobs
• skills
• scripts / workflows
• wiki/archive templates
• proactive questions
• stop/rethink prompts

Default rule: proposal-only. Do not execute external actions unless Simon explicitly confirms.

When to Use

Use when Simon asks:

• “看看最近哪些事情可以自动化”
• “把 Redo 反思层跑一下”
• “找找我最近重复让 Hermes 做什么”
• “哪些流程该变成 skill / cron / wiki 模板”
• “帮我从最近对话里发现主动提醒点”

Do not use for:

• Full desktop screen observation
• Browser monitoring
• Publishing / sending / config-changing without explicit confirmation

Commands

Repository path:

__REPO_PATH__

Run daily Hermes reflection:

cd __REPO_PATH__
python3 -m redo_reflection.cli run --source hermes --profile default --days 1 --output-dir outputs

Analyze a generic JSONL event file:

python3 -m redo_reflection.cli analyze --source jsonl --input sample.jsonl

Output Contract

Each report should include at most three opportunity cards:

• discovered pattern
• evidence snippets, redacted and truncated
• suggested next action
• risk level
• user burden
• estimated leverage

Safety Rules

1. Never auto-create cron jobs from a Redo report.
2. Never auto-send messages from a Redo report.
3. Never commit raw session exports or private transcripts.
4. If a suggestion requires side effects, ask Simon to confirm that exact action.
5. Cool down repeated rejected suggestions.

Verification Checklist

• CLI ran successfully
• Report path exists
• Evidence is redacted/truncated
• Suggestions are actionable, not just summaries
• No external actions were taken automatically