Whip Simulate

Run multi-agent simulations from a user-provided scenario. Concretize the scenario into test cases, spawn agents, and analyze output patterns for consistency.

Input

Extract from $ARGUMENTS:

• Scenario: what to simulate, compare, or verify
• --runs N: number of simulation runs (default: 5)
• --agent: use inline mode (see Execution Mode below)

Execution Mode (mutually exclusive)

$ARGUMENTS determines which dispatch mode this skill uses. The two modes are mutually exclusive:

Strict rules:

1. No --agent in arguments → tracked mode. No exceptions, no inference.
2. --agent in arguments → inline mode. /whip-start, IRC, and lifecycle steps are all skipped.
3. --backend specification (e.g., user says "use codex") → implies tracked mode. Backend selection is a whip concept and is incompatible with --agent.
4. Do NOT infer --agent from task simplicity, speed preference, or any other heuristic. The flag must be explicitly present in the user's input.

If running inside an active whip workspace, use whip workspace view <workspace-name> to get the worktree path for reading code artifacts referenced in the scenario. In tracked mode, simulation tasks go in the global workspace (ephemeral — do not pollute the active workspace).

Workflow

1. Concretize

Read any files, git refs, or codebase artifacts referenced in the scenario, then transform it into concrete test cases:

The output contract is critical — all agents must produce the same structure so results are mechanically comparable:

### Result
- pattern: [short label for the approach taken]
- output:

[code block, JSON, or other structured output]

- decisions: [key judgment calls made]

For A/B comparisons, choose a strategy:

Present the test plan including:

• Test cases with output contracts
• Execution mode (tracked or inline)
• A/B strategy if applicable
• Total agent count

Wait for user approval before executing.

2. Execute

Tracked mode (default)

Hand off dispatch to /whip-start. Prepare one task spec per simulation run and let /whip-start handle IRC, creation, assignment, and monitoring.

Each simulation run becomes one task:

• Title: sim-{test-case}-{run}
• Workspace: global
• Difficulty: easy
• Description: self-contained prompt (Role + Context + Task + Output Contract)

After all tasks complete, collect outputs and proceed to analysis.

Inline mode (--agent)

Spawn one Agent tool call per run, named sim-{test-case}-{run}.

Each prompt must be self-contained — embed all context inline, not file paths:

1. Role: "You are a simulation agent. Execute the task and produce structured output."
2. Context: All file contents and reference material inline
3. Task: The test case action
4. Output contract: The exact format to produce

Batching:

• ≤ 10 runs: spawn all at once with run_in_background: true
• > 10 runs: groups of 10, next batch after previous completes

3. Analyze

Classify outputs into patterns:

1. Collect all agent outputs
2. Group by structural similarity — ignore cosmetic differences (whitespace, comment style, translation wording)
3. Label each group (A, B, C...)
4. Identify root cause of each divergent pattern
5. Flag agents with malformed output as "unclassifiable"

4. Report

## Simulation Report

### Consistency: X/N (Y%)

### Output Patterns
| Pattern | Count | Runs | Description |
|---------|-------|------|-------------|
| A       | 8     | #1-6,#8,#10 | [dominant behavior] |
| B       | 2     | #7,#9 | [variant behavior] |

### Divergence Analysis
For each non-dominant pattern:
- Runs: [list]
- Root cause: [why]
- Severity: cosmetic | functional | breaking
- Diff from dominant: [key differences]

### Summary
- Total: N runs across M test cases
- Dominant pattern: A (X%)
- Key findings: ...
- Recommendation: [if applicable]

Save the full report with raw agent outputs to /tmp/simulate-{slug}-{timestamp}.md and tell the user the path.

Rules

• Never execute before user approves the test plan
• Embed all context inline in prompts — no shared state assumptions
• For A/B comparisons, both versions receive identical inputs
• Use real file contents from the codebase — never fabricate code
• In tracked mode, use global workspace and delegate dispatch to /whip-start
• In tracked mode, clean up simulation tasks after collecting results: whip task clean
• In inline mode, each run is single-shot — no follow-up messages or shared state