Run

Autodialectics

Use this skill when the user wants to work through the Autodialectics harness instead of asking the model to freestyle a task.

Repository facts:

• Repo root: the repository root for the active checkout
• Primary config: autodialectics.yaml
• MCP entrypoint: uv run autodialectics-mcp
• CLI entrypoint: uv run autodialectics
• Python fallback: .venv/bin/python -m autodialectics.cli.main

Preferred MCP workflow:

1. health
2. init_runtime
3. compile_task
4. run_task
5. inspect_run or read_artifact
6. benchmark, evolve_policy, promote_policy, rollback_policy, or replay_run

CLI fallback command forms:

uv run autodialectics init
uv run autodialectics compile examples/code_fix/task.json
uv run autodialectics run examples/code_fix/task.json
uv run autodialectics benchmark
uv run autodialectics inspect <run_id>
uv run autodialectics replay <run_id>
uv run autodialectics evolve
uv run autodialectics promote <policy_id>
uv run autodialectics rollback

Working rules:

• Prefer the MCP server when it is available through the plugin.
• Run CLI commands from the repo root when the MCP server is unavailable.
• Prefer uv run autodialectics ... over ad hoc module invocation.
• Use --config autodialectics.yaml if config resolution looks ambiguous.
• When a run finishes, inspect the generated artifacts/run_* directory and summarize the gate decision, score, slop composite, and unresolved risks.
• For benchmark work, report aggregate results and call out canary failures explicitly.
• For policy work, do not promote a challenger unless the benchmark output supports it.

Expected examples:

• Compile before a risky run if the task contract seems underspecified.
• Use inspect after run or benchmark instead of guessing from stdout.
• Use replay when the user wants the same task rerun under another policy.