LegoFlow

BlocksTerminal CuratorUsage

Configuration Guide

Edit blocks/terminal-curator/config.yaml, or select a complete private profile with --config <file> / TERMINAL_CURATOR_CONFIG. Profiles use the same meta_info / runtime_info schema; they are not partial patches. Relative paths resolve from the block directory.

Source modes

source.typeInputGeneration LLM required?
stackoverflowQuestions and accepted answers selected by tags or IDs.Yes
jsonSaved Terminal-Lego questions at source.path.Yes
tasksLocal candidate directories at source.path.No

For live collection, start with a small batch under runtime_info.input:

source:
  type: stackoverflow
  tags: [bash]
  question_ids: []
  max_questions: 1
  min_score: 10
  max_pages: 1
generation:
  workers: 1
  request_timeout_sec: 180
  max_retries: 2
validation:
  n_concurrent: 1
  timeout_multiplier: 1.0
  job_timeout_sec: 1800

Keep the remaining fields from the full config, including validation.harbor_bin. Explicit question IDs still obey max_questions. Tags and IDs are arrays; limits/concurrency are positive integers, timeouts are positive finite numbers, and max_retries counts extra attempts (2 allows at most three API attempts per call).

Saved questions

JSON input contains a questions array with question and accepted-answer bodies. This abbreviated format example uses placeholder IDs:

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

Use real source attribution in runs. The pinned repository's examples/sample_input.json provides a complete example.

Local candidates

Each candidate needs instruction.md, task.toml, environment/Dockerfile, solution/solve.sh, and pytest tests in tests/test_outputs.py. Symlinked candidates are rejected. The adapter creates the canonical tests/test.sh, preserving an existing entrypoint as test.sh.upstream. Imports receive fresh validation even if they already have an upstream manifest; arbitrary verifier frameworks are not supported.

Credentials

Set llm_api.api_base_url and llm_api.model for generation. Keep tracked key fields empty and supply secrets on the execution host:

Environment variablePurpose
OPENAI_API_KEYGeneration API key.
OPENAI_API_BASE_URL / OPENAI_API_BASEEndpoint override.
MODEL_NAMEModel override.
SO_API_KEYOptional Stack Exchange key.
SO_API_BASE_URLStack Exchange base override.

Environment values override configured values. Keep secrets out of source JSON, commands, reports and commits. source.type: tasks needs no LLM key.

Host and output paths

meta_info.resources.ip: local selects the current machine. For a remote Docker host, set its block directory, synchronize code/input/profile files, and run there through SSH and tmux. An in-block profile keeps its relative location under the remote block directory; an external absolute profile must also exist on that host.

Setup reuses pinned Terminal-Lego and Tracer's pinned Harbor environment. Keep both vendored repositories unchanged.

Outputs must stay under this block's artifacts/, without overlapping inputs, reports, lifecycle files or immutable versions. eval_registry.path must be the task pool's registry.json. See Output Format for the defaults and Downstream Handoff for consumer settings.

On this page