run

Autodialectics Run

Use this skill to drive the Autodialectics anti-slop pipeline against a task.

Prerequisites

autodialectics-mcp must be on PATH (pip install autodialectics).

MCP Workflow

1. health — verify the MCP server is reachable.
2. init_runtime — ensure the database and default champion policy exist.
3. compile_task — compile the task JSON into an immutable contract. Inspect it before proceeding.
4. run_task — execute the full pipeline (evidence → dialectic → execution → verification → evaluation → gate).
   • Pass detach: true for long-running tasks. Poll with inspect_run afterward.

CLI Fallback

If the MCP server is not available:

autodialectics init
autodialectics compile <task.json>
autodialectics run <task.json>

Task File Format

Tasks are JSON files. Only title is required; everything else has sensible defaults:

{
  "title": "Short description",
  "description": "What needs to be done and why",
  "domain": "code|research|writing|experiment|analysis|generic",
  "objectives": ["What the task must achieve"],
  "constraints": ["Boundaries the solution must respect"],
  "deliverables": ["Concrete outputs expected"],
  "acceptance_criteria": ["How to judge success"],
  "forbidden_shortcuts": ["Approaches that are explicitly banned"],
  "workspace_root": "path/to/workspace",
  "verification_commands": ["pytest -q"],
  "max_repair_attempts": 3,
  "assets": [{"kind": "file", "location": "path/to/file", "label": "name"}]
}

Pipeline Stages

run_task executes these stages in order, producing one artifact each:

1. Compile → contract.md — locks the task into an immutable contract
2. Explore → evidence.json — gathers evidence relevant to the task
3. Plan (Dialectic) → dialectic.json — thesis/antithesis/synthesis resolution
4. Execute → execution.json — domain-specific execution (code, research, writing, etc.)
5. Verify → verification.json — independent verification of execution results
6. Evaluate → evaluation.json — slop scoring and gate decision (accept/reject/revise/rollback)
7. Summarize → summary.md — human-readable run summary

Guidance

• Always compile before running if the task is new or ambiguous.
• Summaries should include: run ID, decision (accept/reject/revise/rollback), overall score, slop composite, and unresolved risks.
• Pass a policy_id to run_task to test a specific policy instead of the current champion.

Arguments

If the user passes a task path after /autodialectics:run, treat it as the target task file and execute the compile → run workflow.