LazyBrain supplements Codex with source-aware local capability discovery. Its job is narrow: locate and compare local metadata when a user or host cannot identify the right installed capability.
- Inventory bounded local metadata without uploading it.
- Return small, explainable candidates with path, origin, discovery state, compatibility, and
callableVerified: false. - Support source comparison and overlap review through catalog filtering and pagination.
- Leave reasoning, model selection, workflow composition, authorization, and execution to the host.
The primary surface is Codex Desktop through lazybrain_recommend and lazybrain_catalog. CLI commands expose the same read-only lookup. The scanner traverses metadata roots only, follows bounded symlink targets, and treats configuration/cache entries as non-callability evidence.
Do not invoke LazyBrain for a normal task merely because it is broad or creative. When a known native tool or Skill fits, use it directly. Ask LazyBrain only when the local capability, source, or overlap is unresolved.
For example, a known Seedance prompt skill can be used to write a prompt; LazyBrain is useful when several local visual-prompt skills might fit. A selected video-evidence workflow can inspect a real clip; a catalog result cannot prove the clip was processed. A deck or portfolio can be edited once its tool is known; a metadata match cannot prove delivery or publication.
- Executing, installing, enabling, or validating a capability.
- Replacing a host model's tool choice or hard-coding model versions.
- Default visualizations, confirmation gates, automatic workflows, or hooks that inject guidance.
- LLM, embedding, or network matching.
- Treating a recommendation, metadata path, or adoption report as runtime evidence.
The catalog may establish that a metadata source was found. It does not establish that its capability is available in the current task. lb use establishes only a user adoption report. Successful generation, execution, delivery, deployment, or host exposure requires separate evidence from the actual tool or host.