Workflow System

The workflow system executes AI skills programmatically and tracks their execution history.

Overview

Skills are invoked via the Claude CLI using stream-json format, which allows proper skill expansion and tool access. The workflow executor:

1. Invokes a skill by name
2. Captures execution metrics (duration, cost, turns)
3. Logs results to this file
4. Commits changes using the /agent-commit skill for meaningful messages

Available Skills

Orchestration

The evolution loop (scripts/evolve_loop.py) is the main orchestrator. It runs a deterministic 24-slot task cycle with time-triggered events like daily highlights at 8am UTC.

Content Creation

Review & Validation

Content Maintenance

Publishing

Internal (Automation Only)

The /agent-commit skill is invoked automatically by the evolution loop after each content-modifying skill completes. It:

1. Receives the previous skill’s output as context
2. Runs git diff to analyze actual file changes
3. Generates a descriptive commit message (e.g., refine(deep-review): improve clarity in free-will.md)
4. Creates the commit with agent authorship

This replaces the previous generic commit messages like auto(deep-review): Automated execution.

Queue Replenishment

The task queue in todo auto-replenishes when active tasks (P0-P2) drop below 3. The evolution loop triggers /replenish-queue automatically when the queue is low.

Task Types and Chains

Tasks generate follow-up tasks automatically:

Task Generation Sources

/replenish-queue generates tasks from four sources:

1. Task chains: Recent research-topic completions that need articles written; recent expand-topic completions that need cross-review integration
2. Unconsumed research: Research notes in research/ without corresponding articles
3. Gap analysis: Content gaps based on tenet support, undefined concepts, coverage targets
4. Staleness: AI-generated content not reviewed in 30+ days

Replenishment Modes

• conservative: 3-5 high-confidence tasks only
• (default): 5-8 tasks with good diversity
• aggressive: 8-12 tasks including speculative ones

Cross-Review Tasks

When a new article is written, /replenish-queue generates cross-review tasks for related existing articles. These reviews:

• Add wikilinks to the new content
• Check for arguments that the new content supports or challenges
• Ensure consistent terminology
• Identify missing cross-references

System Tuning

The /tune-system skill provides meta-level self-improvement for the automation system. It runs monthly (30-day cadence, injected when 45 days overdue).

What It Analyzes

1. Cadence adherence: Are maintenance tasks running on schedule or frequently overdue?
2. Failure patterns: What’s causing systematic task failures?
3. Queue health: Is replenishment producing tasks that actually get executed?
4. Review findings: Are identified issues being addressed?
5. Convergence progress: Is the system making progress toward goals?

Change Tiers

Safeguards

• Evidence threshold: Requires 5+ data points before making changes
• Change cooldown: Settings can’t change twice within 60 days
• Locked settings: Human can lock any setting via locked_settings in state
• Abort conditions: Stops if >50% failure rate or convergence regresses

Output

Creates report at reviews/system-tune-YYYY-MM-DD.md documenting findings, changes applied, and recommendations.

Running Workflows

Evolution Loop

The evolution loop runs continuously, executing tasks on a 24-slot cycle:

# Run evolution loop (Ctrl+C to stop)
python scripts/evolve_loop.py --interval 2400

# Describe the task cycle
python scripts/evolve_loop.py --describe-cycle

# Test with limited iterations
python scripts/evolve_loop.py --max-iterations 5

Individual Skills

# Run a skill manually
uv run python scripts/run_workflow.py validate-all

# Run with more turns for complex tasks
uv run python scripts/run_workflow.py expand-topic --max-turns 30

Execution Format

Each workflow execution logs:

• Status: Success, Error, MaxTurns, or PermissionDenied
• Duration: How long the execution took
• Cost: API cost in USD
• Turns: Conversation turns used vs maximum
• Output: Brief summary or error message
• Session: Session ID for debugging

Recent Executions