work-loop

# Work Loop

Queue-driven development workflow that processes requests using isolated sub-agents.

## Core Principle

**The Orchestrator never performs task work.**

Every request spawns a fresh sub-agent. Sub-agents die after completion. This ensures:
- Context isolation between tasks
- Clean state for each request
- Crash resilience
- Predictable execution

## Directory Structure

```
do-work/
├── requests/     # Pending work (input queue)
├── in-progress/  # Currently executing (max 1)
├── archive/      # Completed successfully
└── errors/       # Failed with error summary
```

If directories don't exist, create them on first run.

## Workflow

### 1. Initialize

Check for and create required directories:

```bash
mkdir -p do-work/{requests,in-progress,archive,errors}
```

### 2. Check Queue

```bash
ls -1 do-work/requests/*.md 2>/dev/null | head -1
```

If empty and `--watch` not set, exit cleanly.
If empty and `--watch` set, poll every 5 seconds.

### 3. Process Request

For each request file:

#### 3a. Claim Request

Move from `requests/` to `in-progress/`:

```bash
mv do-work/requests/FILENAME.md do-work/in-progress/
```

Only ONE file should be in `in-progress/` at a time.

#### 3b. Parse Request

Read the request file and extract:
- **Intent** - What outcome is desired
- **Context** - Relevant constraints
- **Tasks** - Steps to complete
- **Done When** - Completion criteria

#### 3c. Spawn Sub-Agent

Create a fresh execution context with:
- Request content only
- Current repo state
- No previous task context

The sub-agent operates in three phases:

**Planning Phase:**
- Analyze the request
- Identify files to modify
- Determine approach
- Validate against "Done When" criteria

**Execution Phase:**
- Implement the tasks
- Make necessary changes
- Run relevant tests

**Validation Phase:**
- Verify "Done When" criteria met
- Run tests if applicable
- Check for regressions

#### 3d. Handle Outcome

**Success:**
```bash
mv do-work/in-progress/FILENAME.md do-wor