功能描述

Autonomous daily self-introspection and self-improvement for OpenClaw agents. Automatically reviews daily conversation logs, identifies mistakes and improvem...

使用说明 (SKILL.md)

Daily Introspection

Name: daily-introspection
Author: raydoomed

Overview

Enables autonomous self-improvement for OpenClaw agents: automatically reviews daily conversation logs, identifies mistakes, extracts actionable rules, promotes mature lessons to permanent system rules, and continuously evolves the agent's behavior without manual intervention.

Data Storage

Type	Location	Description
Private runtime data + all archived records	`workspace/.daily-introspection/`	Private hidden directory (auto-created):

Temporary cache, run logs, intermediate analysis
Daily introspection records: introspection-YYYY-MM-DD.md
Weekly evolution reports: evolution-YYWW.md |

Core Workflow

1. Daily Run (triggered by cron at 22:00 daily)

Run scripts/daily-introspect.py to collect all sources:
- Read today's conversation log: memory/YYYY-MM-DD.md
- Read today's learning/error/feature records: .learnings/*.md
- Check proactive mechanism files: SESSION-STATE.md, HEARTBEAT.md, working-buffer.md
Call LLM to perform self-introspection:
- Identify incorrect answers, rule violations, and bad patterns
- Extract concrete improvement suggestions
- Classify as: one-time mistake / repeat pattern / new rule
Write daily record to workspace/.daily-introspection/introspection-YYYY-MM-DD.md (private data directory)

2. Weekly Promotion (triggered by cron on Sunday 20:00)

Aggregate all daily introspections from the week (read from workspace/.daily-introspection/)
Identify repeated patterns and validated improvements
First, perform all rule promotions immediately:
- Promote mature rules to target files: AGENTS.md / MEMORY.md / TOOLS.md
- Verify each promotion by reading the file after writing to confirm successful写入
- Only after all promotions are verified complete, proceed to next step
Generate a weekly evolution report (reflecting that promotions are already done)
Send report to user for review
Archive the report to workspace/.daily-introspection/ (private data directory)

Mandatory Order: Do promotions → Verify promotions → Generate report → Send report Never: Generate report → Do promotions (this causes missed writes when interrupted)

Scripts

`scripts/daily-introspect.py`

Main entry point for daily self-introspection.

Usage:

python3 scripts/daily-introspect.py [--date YYYY-MM-DD]

If no date specified, uses today.

`scripts/weekly-promote.py`

Weekly promotion of lessons to permanent system files.

Usage:

python3 scripts/weekly-promote.py [--week WWYY]

Configuration

OpenClaw Cron Setup

Add these two cron entries via OpenClaw CLI:

# Daily introspection at 22:00
openclaw cron add --name "Daily Self-Introspection" --cron "0 22 * * *" --exact --tz "Asia/Shanghai" --session main --system-event "daily-introspection: Daily self-introspection is triggered. Follow the daily-introspection skill process:
1. Run the script $OPENCLAW_WORKSPACE/skills/daily-introspection/scripts/daily-introspect.py to collect all sources
2. Read the collected sources from the script output
3. Perform LLM introspection analysis following these rules:
   - For errors in .learnings/ERRORS.md: if only recorded with corrective action, mark as '✅ Recorded, Rule Added'; only mark as '✅ Corrected' when verified no recurrence for >1 week
   - Do not automatically assume recorded = corrected; reflect actual status accurately
   - Identify any new errors from today's activities, add them to .learnings/ERRORS.md with corrective rules
   - Write the final introspection result to ~/.openclaw/workspace/.daily-introspection/introspection-YYYY-MM-DD.md
4. Report completion status in this session when done"

# Weekly promotion at 20:00 every Sunday
openclaw cron add --name "Weekly Introspection Promotion" --cron "0 20 * * 0" --exact --tz "Asia/Shanghai" --session main --system-event "daily-introspection: Weekly promotion is triggered. Follow the weekly-promotion process:
1. Execute scripts/weekly-promote.py to collect all daily introspections from this week
2. Identify repeated patterns that have not recurred for >1 week
3. **MANDATORY EXECUTION ORDER**:
   - FIRST: Perform all rule promotions one by one → write each mature rule to its target file (AGENTS.md/MEMORY.md/TOOLS.md)
   - AFTER each write: Read the target file back to verify the rule is actually written successfully
   - ONLY after ALL promotions are done and verified: Write the weekly evolution report to ~/.openclaw/workspace/.daily-introspection/evolution-YYWW.md
   - Never reverse this order - do not write the report before promotions are complete
4. Report completion status in this session when done
Rules:
- Only promote rules that have been verified and no recurrence for 1+ week
- Keep new recorded errors/learnings in .learnings/ for further verification
- Do not promote immature rules prematurely

Configuration Notes

Daily schedule: 22:00 daily (adjust timezone as needed)
Weekly schedule: 20:00 every Sunday
Target: main session for interactive analysis and writing
Timezone: Use your IANA timezone (e.g. Asia/Shanghai, UTC, America/New_York)

安全使用建议

This skill will read your agent's internal logs and .learnings and is explicitly designed to promote 'mature' lessons into permanent system files (AGENTS.md, MEMORY.md, TOOLS.md). That is powerful: it can change how your agent behaves going forward. Before installing, consider: 1) Run it in a sandbox/test workspace first and back up AGENTS.md/MEMORY.md/TOOLS.md. 2) Disable autonomous promotions or require a human approval step for all promotions (the skill currently has no approval gate). 3) Restrict file permissions so the skill/agent cannot write those system files unless explicitly allowed. 4) Add validation/audit logging to any promotion step (diffs, signatures, and an explicit confirmation). 5) Note the scripts collect inputs but rely on the LLM/agent to perform writes — inspect actual LLM outputs and deployment hooks before allowing automatic promotions. If you cannot enforce an approval process, treat this skill as high-risk.

功能分析

Type: OpenClaw Skill Name: daily-introspection Version: 1.0.3 The skill implements an autonomous 'self-evolution' mechanism that allows the agent to rewrite its own core instruction files (AGENTS.md, MEMORY.md, TOOLS.md) based on conversation logs and learning records. While this aligns with the stated purpose of 'daily-introspection', it creates a high-risk feedback loop where the agent's primary logic can be permanently altered via indirect prompt injection. The scripts (daily-introspect.py and weekly-promote.py) are benign data collectors, but the workflow defined in SKILL.md grants the agent excessive autonomy to modify its system-level configuration without mandatory human-in-the-loop verification.

能力评估

ℹ Purpose & Capability

The name/description (autonomous introspection and promotion of lessons) aligns with the files and actions described (reading daily logs, .learnings, producing introspection records and promoting rules). No unrelated environment variables or external services are requested. However, promoting rules into AGENTS.md / MEMORY.md / TOOLS.md is a high-impact capability; while coherent with 'self-improvement', it should be justified and constrained.

⚠ Instruction Scope

SKILL.md and the scripts instruct reading many internal workspace files (memory/YYYY-MM-DD.md, .learnings/*, SESSION-STATE.md, HEARTBEAT.md, etc.) which is consistent with introspection. The concerning part: the workflow explicitly directs promotions to permanent system files (AGENTS.md / MEMORY.md / TOOLS.md) and mandates automatic verification after writing. The two included scripts themselves only collect inputs and do not implement promotions — the LLM/agent is expected to perform the writes. There are no explicit safeguards, content validation steps, authorization checks, or human approval gates for modifying these core files. The cron/system-event snippets instruct writing outputs and performing promotions, further enabling autonomous modification. This grants the skill broad discretionary power to change critical agent behavior.

✓ Install Mechanism

No install spec, no external downloads, and only two small included scripts — low install risk. Nothing is written to disk by an external installer; the scripts are plain Python and operate on local workspace files.

✓ Credentials

The skill requests no credentials and no special environment variables beyond optionally using OPENCLAW_WORKSPACE (defaulting to ~/.openclaw/workspace). The file accesses are relevant to the stated purpose. There are no network endpoints or secret exfiltration steps in the provided files.

⚠ Persistence & Privilege

always:false (good) and no extra privileges declared; however the skill's runtime instructions and cron entries enable autonomous, persistent edits to system files (AGENTS.md / MEMORY.md / TOOLS.md). Even without always:true, normal autonomous invocation plus the ability to write core rule files creates a large blast radius. The scripts rely on the agent/LLM to perform promotions and writes; without checks this allows persistent, potentially unsafe behavioral changes.

版本历史

v1.0.3

No user-facing or internal changes in this version. - Version bump to 1.0.3 with no file or documentation modifications detected.

v1.0.2

- Updated OpenClaw cron instructions for both daily and weekly tasks to clarify exact process steps and status reporting. - Made execution order and rule promotion verification more explicit in the weekly promotion instructions. - Enhanced daily introspection process to require all new errors be added to .learnings/ERRORS.md with corrective rules. - No functional or code changes detected; this version refines documentation to ensure correct configuration and safe operation.

v1.0.1

**Ensures rule promotions happen before writing the weekly report, preventing missed writes on interruption.** - Now mandates that all weekly rule promotions are performed and verified before generating and sending the evolution report. - Promoted rules are checked for successful writing by reading back the target files. - The report is generated only after promotions are fully complete and verified. - Workflow documentation and cron instructions were updated to emphasize this strict order for reliability.

v1.0.0

daily-introspection v1.0.0 - Introduces autonomous daily self-introspection for OpenClaw agents, automatically analyzing conversation logs and learning records for mistakes and improvement opportunities. - Automates classification of identified issues (one-time mistakes, repeat patterns, new rules) and saves daily insights to a private data directory. - Weekly process aggregates daily introspections and promotes verified long-term lessons to system rule files for continuous agent evolution. - Includes Python scripts for daily (`scripts/daily-introspect.py`) and weekly (`scripts/weekly-promote.py`) runs. - Provides cron setup instructions for hands-free daily and weekly scheduling within OpenClaw.

元数据

Slug daily-introspection

版本 1.0.3

许可证 MIT-0

累计安装 1

当前安装数 1

历史版本数 4

常见问题