random-thought（随意的想法）

Random Thought

A thinking engine that reads your workspace and writes about what it finds — not summaries, not status updates, but genuine reflection on what's unresolved, alive, or worth questioning.

Two Stages

Writer

Pick a random file from the corpus, read it, write a reflective observation, and deliver it.

Curator

Read all recent Writer outputs, synthesize patterns, classify observations using configurable action tags, and produce a digest.

Quick Start

Manual invocation

Run the writer for a single reflection:

bash scripts/corpus-pick.sh        # see what file gets selected
# Then the agent reads the file and writes a reflection

Cron-driven (recommended)

Wire two cron jobs — the writer runs frequently, the curator runs periodically:

# Writer: every hour
0 * * * *   openclaw skill run random-thought --stage writer

# Curator: daily at a quiet hour
0 5 * * *   openclaw skill run random-thought --stage curator

Each cron invocation runs in an isolated session — no context bleed between runs.

Configuration

Create random-thought.config.json in your workspace root (all fields optional):

{
  "corpus": {
    "watchDirs": ["."],
    "excludePatterns": [
      "node_modules", ".git", ".next", "dist", "build",
      "venv", "__pycache__", "*.png", "*.jpg", "*.gif",
      "*.mp3", "*.ogg", "*.pdf", "*.zip", "*.env",
      "*.pem", "*.key", "package-lock.json", "*.lock"
    ],
    "minFileSize": "100c",
    "maxFileSize": "500k"
  },
  "freshness": {
    "enabled": true,
    "days": 7,
    "historyFile": ".random-thought-history"
  },
  "actionTags": [
    { "name": "you-decide", "description": "Needs human judgment" },
    { "name": "agent-execute", "description": "Agent can act autonomously" },
    { "name": "spark", "description": "Interesting but no action needed" }
  ],
  "output": {
    "dir": "random-thought-output",
    "digestFormat": "markdown"
  }
}

Configuration Reference

• corpus.watchDirs — Directories to scan (relative to workspace). Default: ["."]
• corpus.excludePatterns — Glob patterns to exclude. Sensible defaults included.
• freshness.enabled — Prevent revisiting files within N days. Default: true
• freshness.days — Cooldown period per file. Default: 7
• freshness.historyFile — Where to track visited files. Default: .random-thought-history
• actionTags — Labels the Curator uses to classify observations. Override to match your workflow.
• output.dir — Where digests are written. Default: random-thought-output/

Writer Stage

What it does

1. Run scripts/corpus-pick.sh to select a random file (respecting freshness gate)
2. Read the selected file (up to 200 lines if large)
3. Write a reflective observation — not a summary, not a review. Follow the thread wherever it leads.
4. Deliver the output (post to configured channel, write to file, or return to caller)

Writer output format

• First line: 📂 [absolute path of the file selected in step 1]
• Blank line
• Flowing prose. No bullets. No headers. No preamble.
• Follow what's unresolved, surprising, half-finished, or alive.
• Make connections to other things in the workspace if they surface naturally.
• Last line: 🖊️ *Writer*

Avoid these crutches

The Writer should push past generic phrasing:

• "quietly elegant" / "speaks to something deeper" / "testament to" / "tapestry"
• "there's something almost [adjective] about"
• "beautiful in its simplicity"

Concrete observations beat ornamental ones.

Curator Stage

What it does

1. Read all Writer outputs from the configured period (default: last 24 hours)
2. Classify each observation using configured action tags
3. Identify recurring themes or convergent patterns across observations
4. Write a structured digest to output.dir/YYYY-MM-DD.md

Digest format

# Random Thought Digest — YYYY-MM-DD

## Action Items
### you-decide
- [observation summary] — [why it needs human judgment]

### agent-execute
- [observation summary] — [what the agent should do]

## Patterns
[Recurring themes across today's observations — what's converging?]

## Sparks
- [Interesting observations with no action needed]

Curator guidelines

• Classify every observation into exactly one action tag
• Surface theme convergence — if multiple observations point at the same gap, name it explicitly
• Keep the digest scannable — each item should be one line with context
• Do NOT re-surface items from previous digests unless they've evolved

Scripts

scripts/corpus-pick.sh

Selects a random file from the corpus, respecting the freshness gate.

Usage: bash scripts/corpus-pick.sh [workspace_root] [config_path]

• Reads configuration from random-thought.config.json if present
• Falls back to sensible defaults if no config exists
• Records selected file to the history file for freshness tracking
• Outputs the absolute path of the selected file

scripts/freshness-gate.sh

Manages the file visit history for the freshness gate.

Usage:

• bash scripts/freshness-gate.sh check <file> [config_path] — exits 0 if file is fresh (OK to visit), 1 if recently visited
• bash scripts/freshness-gate.sh record <file> [config_path] — records a file visit
• bash scripts/freshness-gate.sh prune [config_path] — removes entries older than the configured window

References

• architecture.md — Design rationale: why two stages, why cron-driven isolation matters, the hybrid skill+cron model. Read when customizing the pipeline or understanding trade-offs.