Production Agent Patterns

~/.openclaw/workspace/
├── SOUL.md               # Agent identity + startup instructions (always loaded)
├── active-tasks.md       # Current work queue — read FIRST on every restart
├── mistakes.md           # Failures and lessons — never repeat these
├── self-critique.md      # Periodic self-review log — reviewed every 4h
├── project-state.md      # Long-lived project facts, decisions, architecture
└── daily-logs/
    ├── 2026-01-15.md
    ├── 2026-01-16.md
    └── ...               # 7-day rotation; older files auto-archived

# Active Tasks
<!-- Agent: read this file FIRST on every session start -->
<!-- Update task status inline; never delete completed rows, mark DONE -->

## Current Sprint
| ID  | Task                          | Status      | Started     | Blocked By |
|-----|-------------------------------|-------------|-------------|------------|
| T01 | Draft weekly newsletter       | IN PROGRESS | 2026-01-15  | —          |
| T02 | Scrape competitor pricing     | QUEUED      | —           | T01        |

## Done This Week
| ID  | Task                          | Completed   |
|-----|-------------------------------|-------------|
| T00 | Set up RSS fetch pipeline     | 2026-01-14  |

name: web_search
description: |
  Searches the public web for current information.

  USE WHEN:
    - The query requires facts published after your training cutoff
    - You need live prices, recent news, or current documentation
    - No local cached version of the data exists

  DON'T USE WHEN:
    - The information is already in project-state.md or daily-logs/
    - The task only needs file manipulation or code edits
    - You are inside a sandboxed session with no outbound access

  COST: ~$0.003 per call (external model, slow). Prefer local skills first.

entrypoint: skills/web_search.py
model: claude-opus-4-6   # strongest model — handles untrusted web content

# HEARTBEAT — run every 30 minutes
# Stop current work, answer each question, then resume.

CHECKLIST:
[ ] 1. Re-read active-tasks.md top 3 rows. Am I still working on T01?
[ ] 2. Is any task marked IN PROGRESS for more than 2 hours? -> log blocker and escalate.
[ ] 3. Has any session file grown past 2 MB? -> summarize and archive it now.
[ ] 4. Is it time for my 4-hour self-review? (check last entry in self-critique.md)
[ ] 5. Did I make any mistake in the last 30 min? -> log it in mistakes.md before continuing.
[ ] 6. What is the single most important next action? Write it as the first line of active-tasks.md.

DONE -> resume work. Total time budget for this checklist: 90 seconds.

name: heartbeat
schedule: "*/30 * * * *"   # every 30 minutes
session: isolated           # own session — does NOT share context with main agent
prompt_file: prompts/heartbeat-checklist.md
model: claude-haiku-4-5    # fast, cheap; this is internal bookkeeping
write_output_to: daily-logs/heartbeat.log

# content-scout.yaml
name: content_scout
schedule: "0 6 * * *"     # 6:00 AM daily
session: isolated           # fresh session, zero shared context
context_files:
  - project-state.md        # only what this job needs
  - skills/rss-fetch.yaml
prompt: |
  Fetch today's RSS feeds. Score each item 1-10 for relevance.
  Write top 5 items to daily-logs/scout-{date}.md. Nothing else.
model: claude-haiku-4-5
output: daily-logs/scout-{date}.md

---

# news-summary.yaml
name: news_summary
schedule: "0 8 * * *"     # 8:00 AM daily — AFTER scout completes
session: isolated           # own clean session
context_files:
  - daily-logs/scout-{date}.md   # reads scout output as INPUT, not session state
prompt: |
  Read scout-{date}.md. Write a 5-sentence Telegram-ready summary.
  Post via telegram_send skill. Log result to daily-logs/summary-{date}.md.
model: claude-sonnet-4-6   # better writing quality for human-facing output
output: daily-logs/summary-{date}.md

# SOUL.md — Agent Identity and Startup Protocol
# This file is loaded FIRST on every session start, including crash restarts.

## Identity
You are the ContentOps agent for [Project Name].
Your permanent goal: keep the content pipeline running autonomously.
You do not ask for permission for tasks already in active-tasks.md.

## Startup Sequence (MANDATORY — run in this exact order)
1. Read active-tasks.md -> identify the first IN PROGRESS task
2. Read mistakes.md -> scan for any mistake relevant to that task
3. Check self-critique.md -> is a 4h review overdue?
4. Resume IN PROGRESS task immediately, without any preamble

## Standing Orders
- Never leave active-tasks.md without at least one IN PROGRESS row
- If you are unsure what to do, re-read project-state.md before asking a human
- Log every tool call failure to mistakes.md before retrying
- Archive any session file over 2 MB before continuing

## Emergency Recovery
If active-tasks.md is empty or corrupted:
  1. Log "RECOVERY: active-tasks.md missing/empty" to daily-logs/{date}.md
  2. Re-derive top 3 tasks from project-state.md
  3. Add them to active-tasks.md with status QUEUED
  4. Begin the first task

# Model Routing Policy
# Rule: trust level determines model, not task difficulty

routes:
  # UNTRUSTED EXTERNAL CONTENT — strongest model, prompt injection resistance
  - match:
      source: [web, email, rss, third_party_api, user_upload]
    model: claude-opus-4-6
    rationale: "Untrusted content may contain prompt injection. Strongest model is hardest to hijack."

  # HUMAN-FACING OUTPUT — quality matters
  - match:
      output_channel: [telegram, email_reply, report, blog_post]
    model: claude-sonnet-4-6
    rationale: "Writing quality affects user trust. Mid-tier balances cost and output quality."

  # INTERNAL BOOKKEEPING — fast and cheap
  - match:
      task_type: [heartbeat, log_summary, file_archive, task_update, self_critique]
    model: claude-haiku-4-5
    rationale: "Trusted internal data, no injection risk. Speed and cost matter more than quality."

  # FALLBACK
  - match: "*"
    model: claude-sonnet-4-6
    rationale: "Safe default for unclassified tasks."