Configuration

config.yaml follows LegoFlow's meta_info / runtime_info block schema. It configures one run; progress and status are written to artifacts. Paths within the block config resolve from the block directory, including when --config selects an overlay stored elsewhere.

Inputs

All fields below live under runtime_info.input.

The checked-in configuration is the exact field reference, including runtime paths and defaults. Do not change the upstream scripts to configure a run. Use --config <file> with block commands for a local overlay; it is a full block configuration with the same schema, not a partial YAML patch.

Source modes

stackoverflow collects questions and accepted answers through the Stack Exchange API, then generates candidates. Set tags or question IDs and keep max_questions / max_pages small for a first run. An invalid API key or exhausted quota is a source failure; the block does not silently switch to another source or crawl indefinitely.

json reuses saved Terminal-Lego input. The file has a top-level questions array. Each item contains a question and accepted-answer body:

{
  "questions": [{
    "question_id": 123456,
    "title": "A concrete shell task",
    "body": "<p>The source problem.</p>",
    "tags": ["bash"],
    "score": 10,
    "link": "https://stackoverflow.com/questions/123456",
    "accepted_answer_id": 123457,
    "accepted_answer": {
      "answer_id": 123457,
      "body": "<p>The accepted source answer.</p>",
      "score": 12
    }
  }]
}

This is a format example, not a real source to collect. Use actual question provenance in runs. The pinned repository's examples/sample_input.json illustrates the same shape. JSON generation still needs an LLM, then the same Harbor acceptance gate as live collection.

tasks imports local Terminal-Lego-style task candidates. Set source.path to their pool. Each candidate needs instruction.md, environment/Dockerfile, solution/solve.sh, task.toml, and tests/test_outputs.py containing pytest tests. Candidate directories and their contents must not be symlinks. The adapter installs a canonical tests/test.sh that runs the pytest checks and writes a JUnit report and Harbor reward; an existing entrypoint is preserved as tests/test.sh.upstream. Arbitrary verifier frameworks are not supported by this import mode. Imported tasks are normalized and validated; an incoming manifest or previous upstream validator result is not trusted as this block's fresh Oracle/NOP evidence. No generation LLM key is required.

Credentials

Environment values override their corresponding configured fields:

Use your shell or secret manager to supply values. Do not paste credentials into commands that will be archived, commit a populated key field, or put a secret in a source JSON file. Reports and dashboards must contain no keys. Archived configuration is redacted, including when an overlay is used.

Resources and dependencies

meta_info.resources.ip: local runs on the current host. A remote host means SSH and tmux on that host, with directory naming the synchronized checkout's block directory. The normal default is local; choose a Docker-capable node when the orchestration node lacks Docker.

The legacy root launcher maps an in-block profile to the same relative location under the remote resources.directory; a profile outside the block keeps its explicit absolute path. Prepare that profile at the mapped path with the authorized code/input synchronization, and supply credentials in the execution host's environment.

repos/terminal-lego is pinned under meta_info.repos. Setup may initialize the submodule; generation and checks never modify it. Harbor validation uses Tracer's pinned repository and artifacts/env/harbor-uv environment rather than a new floating runtime. Source tasks must be available on that same host.

Outputs and Tracer selection

runtime_info.output.harbor_tasks_dir.path identifies the published pool, normally artifacts/tasks. The report output is artifacts/reports/latest.json. The verified manifest inside the pool is the authoritative downstream allowlist.

In blocks/tracer/config.yaml:

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

The task_source snippet belongs under runtime_info.input. Retain Tracer's existing SWE dataset_name value; the source resolver chooses the terminal path when local_block selects it. curator remains the default local producer, so installing this plugin does not require terminal credentials for an existing SWE run.

The same publication also exports runtime_info.output.eval_registry.path, normally artifacts/tasks/registry.json, for Evaluator's existing harbor_registry provider. Select terminal-curator@1.0, point task_source.registry_path at that file, and set task_source.no_hack: false. See the full Evaluator example.