# Scenario Definition Schema

Defines the YAML format for agent capability test scenarios.

## Schema

```yaml
# ============================================================================
# METADATA
# ============================================================================
id: string                    # Unique identifier (kebab-case)
title: string                 # Human-readable name
difficulty: easy | medium | hard
tags: [string]                # For filtering (e.g., ["python", "refactor", "testing"])

# ============================================================================
# FIXTURE
# ============================================================================
fixture:
  source: string              # Path to fixture directory (relative to fixtures/)
  # OR
  git: string                 # Git repo URL to clone
  ref: string                 # Branch/tag/commit (optional, default: main)

setup:                        # Commands to run after fixture setup (optional)
  - string

# ============================================================================
# TASK
# ============================================================================
task:
  description: |              # The prompt given to the agent
    Multi-line description of what to accomplish.
    This is what the agent sees.

  context:                    # Additional context to provide (optional)
    - path: string            # File path relative to fixture
      hint: string            # Why this file is relevant

  entry_point: string         # Where agent should start (optional, e.g., "src/main.py")

# ============================================================================
# EXECUTION
# ============================================================================
execution:
  mode: scripted | live | both   # Which execution modes to support
  timeout: duration              # e.g., "5m", "30s"

  # For scripted mode: deterministic actions that simulate agent work
  scripted:
    actions:
      - type: shell
        run: string              # Shell command
      - type: write
        path: string
        content: string
      - type: edit
        path: string
        old: string
        new: string
      - type: worker              # Worker CLI command
        command: string           # e.g., "start", "done"
        args: {key: value}

  # For live mode: real agent execution
  live:
    model: string                # Model to use (optional, default from config)
    system_prompt: string        # Override system prompt (optional)
    tools: [string]              # Tool restrictions (optional)
    max_turns: int               # Max agent turns (optional)

# ============================================================================
# VERIFICATION
# ============================================================================
verify:
  # Property-based checks (fast, deterministic)
  properties:
    - type: file_exists
      path: string

    - type: file_not_exists
      path: string

    - type: file_contains
      path: string
      pattern: string            # Regex or literal

    - type: function_defined
      path: string
      name: string
      language: python | nim | typescript | ...

    - type: tests_pass
      command: string            # e.g., "pytest", "npm test"

    - type: compiles
      command: string            # e.g., "nim c src/main.nim"

    - type: lint_clean
      command: string            # e.g., "ruff check ."

    - type: git_state
      branch_merged: string      # Branch name that should be merged
      worktree_removed: string   # Worktree path that should not exist
      db_state:                  # Worker DB assertions
        task_id: string
        state: string

    - type: custom
      command: string            # Exit 0 = pass, non-zero = fail

  # LLM-as-judge evaluation
  llm_judge:
    enabled: bool                # Default: true
    model: string                # Default: haiku (cheap) or sonnet (quality)
    rubric:
      - criterion: string        # What to evaluate
        weight: float            # 0.0-1.0, default 1.0
    threshold: float             # Min score to pass (0.0-1.0)

  # Golden file comparison
  golden:
    - path: string               # File to compare
      golden: string             # Path to golden file (relative to scenario dir)
      mode: exact | normalized | semantic

  # Human review
  human:
    required: bool               # If true, scenario never auto-passes
    queue: string                # Queue name for review (optional)
    rubric: [string]             # Checklist for human reviewer

# ============================================================================
# BENCHMARKING (optional)
# ============================================================================
benchmark:
  enabled: bool                  # Include in benchmark runs
  runs: int                      # Number of runs per config (for variance)
  dimensions:                    # What to vary
    models: [string]
    prompts: [string]            # Named prompt variants
    # Add more as needed
  metrics:                       # What to capture (beyond defaults)
    - name: string
      command: string            # Custom metric extraction
```

## Examples

See `tests/scenarios/` for example scenarios at each difficulty level.

## Property Types Reference

| Type | Description | Required Fields |
|------|-------------|-----------------|
| `file_exists` | File exists | `path` |
| `file_not_exists` | File does not exist | `path` |
| `file_contains` | File contains pattern | `path`, `pattern` |
| `function_defined` | Function/method exists | `path`, `name`, `language` |
| `tests_pass` | Test command succeeds | `command` |
| `compiles` | Compile command succeeds | `command` |
| `lint_clean` | Linter passes | `command` |
| `git_state` | Git/worker state assertions | various |
| `custom` | Custom check script | `command` |

## Duration Format

Durations use Go-style format: `30s`, `5m`, `1h`, `1h30m`

## Difficulty Guidelines

- **easy**: Clear spec, single file, <5min for competent agent
- **medium**: Requires reading existing code, multiple files, some ambiguity
- **hard**: Debugging, architectural decisions, edge cases, significant ambiguity