replay

Autodialectics Replay

Use this skill to replay a previous pipeline run with an optional policy override.

Prerequisites

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

MCP Tool

• replay_run(run_id, policy_id?) — re-execute a stored run manifest. If policy_id is provided, the replay uses that policy instead of the original.

CLI Fallback

autodialectics replay <run_id>
autodialectics replay <run_id> --policy <policy_id>

When to Replay

• A/B policy comparison — replay the same run with two different policies and compare decisions, scores, and slop composites.
• Debugging a rejection — replay a rejected run to see if the issue was the policy, the evidence, or the task itself.
• Regression testing — after evolving a policy, replay historical runs to check for regressions.

Guidance

• Always inspect_run the original before replaying — understand what happened first.
• When comparing original vs replay, report both side by side: decision, overall score, slop composite, and any divergent artifact content.
• Replays create new run records — they don't overwrite the original.

Arguments

If the user passes a run ID after /autodialectics:replay, replay that run and summarize the comparison with the original.