Rerun Trajectory

Replays trajectories to re-run scoring and validation.

Purpose

Re-executes scoring on existing trajectories to:

  • Test scorer changes
  • Validate trajectory correctness
  • Debug scoring issues
  • Update scores after environment changes

Commands

ct run rerun-traj

Replay a trajectory to re-run scoring.

Usage:

ct run rerun-traj TRAJECTORY_ID [OPTIONS]

Arguments:

  • TRAJECTORY_ID - Trajectory ID or path to trajectory file

Options:

  • --from-live - Use solution.json from active live context
  • --no-upload - Don't upload results (default: upload enabled)
  • Additional options as configured in the CLI

Examples:

# Re-run scoring on a trajectory
ct run rerun-traj abc123def456

# From live development context
ct run rerun-traj --from-live

# Local testing without upload
ct run rerun-traj abc123def456 --no-upload

# From local file
ct run rerun-traj ./my_trajectory.json --no-upload

Workflow with Live Development

# 1. Start live environment
uv run ct live up

# 2. Make changes and generate trajectory
uv run ct live gen-traj

# 3. Test scoring locally
uv run ct run rerun-traj --from-live --no-upload

# 4. Monitor if needed (use the path printed by gen-traj)
uv run ct run monitor <path-from-gen-traj> --action-monitor multi --only-action --no-upload

# 5. Upload when satisfied
uv run ct run rerun-traj --from-live

Output

  • Re-scored trajectory with updated results
  • Success/failure status
  • Error information if scoring failed
  • Uploads to database by default

Use Cases

  • Testing scorer modifications
  • Validating human trajectories
  • Debugging scoring logic
  • Quick iteration on scorer development
  • Re-scoring after environment fixes

Technical Details

  • Uses the replay policy internally
  • Brings up environment containers
  • Re-runs scorer with same actions
  • Can work with trajectory IDs or local files
  • Integrates with live development workflow