LegoFlow

BlocksTerminal CuratorUsage

Downstream Handoff

Complete validation and semantic review first. Use the corrected publication's manifest and registry together; retain immutable versions while downstream jobs reference them.

Tracer

Set these fields under Tracer's runtime_info.input:

task_source:
  provider: local
  local_block: terminal-curator
  terminal_tasks_dir: ../terminal-curator/artifacts/tasks

Keep the other required Tracer settings, including its existing SWE dataset_name; the producer selector chooses the terminal pool. Run /tracer:check then the authorized /tracer:run (Codex: $tracer:check, $tracer:run). Tracer stages only manifest-listed IDs and locks that batch.

For root orchestration, use /root:run / $root:run, which includes the producer's semantic audit. The legacy shell helper refuses an implicit terminal-to-Tracer chain; explicit terminal-only, audit/revalidation, then tracer-only operations remain available.

Inspect each trial's actual reward and solving trajectory. Oracle validation and a successful launcher exit do not establish that a solving agent succeeded. See Tracer Getting Started.

Trainer

Use Tracer's normal trajectory conversion and Trainer input workflow; no terminal-specific training runtime is needed. Count solved trials and retained LF records separately: existing role, tool-pair and reasoning checks may exclude a reward-1 trace. Inspect conversion statistics and retained records before data preparation, keeping those quality thresholds in effect.

Evaluator

Select the native local registry in a full Evaluator config or EVAL_CONFIG profile. These fields belong under runtime_info.input:

task_source:
  provider: harbor_registry
  dataset_name: terminal-curator
  version: "1.0"
  registry_path: ../terminal-curator/artifacts/tasks/registry.json
  no_hack: false
harbor_job:
  n_tasks: 2
  n_concurrent: 2
  max_retries: 0
job_analysis:
  enabled: false

Retain the remaining endpoint/model/agent settings; supply credentials through the environment. n_tasks: 2 is a small starting run; null selects the whole batch. SWE-specific no_hack remapping and gold-data analysis stay disabled.

Run /evaluator:check then /evaluator:run (Codex: $evaluator:check, $evaluator:run) and inspect real per-task results. Evaluator locks the physical registry version before launch. Its absolute task paths must remain readable on the Docker host; using another node requires making that batch available there. See Evaluator Getting Started.

On this page