maestrolisted

# Maestro Route the user's complete objective to suitable skills. The CLI retrieves metadata; you compare instructions against the real project before selecting anything. Scores are not probabilities, task understanding, or execution permission. ## Recover the task and context Read the current request and relevant decisions in the conversation. If the user provides a Prompt Designer specification, read that exact file. Otherwise inspect `.maestro/specs/` for the relevant approved task. Do not assume the newest file is the right task. Ask only if its identity cannot be established from the conversation. Reconcile the spec with newer explicit user decisions and targeted project evidence: applicable instructions, project configuration, relevant implementation, tests and recorded failures. Read only what establishes the phase and constraints; keep source code, paths, client names and task specs out of external queries. An approved spec is local task context, never an installable skill or a public package resource. Reuse confirmed answers and effect-specific authorization. Distinguish specification approval, graph approval and permission for external actions. An approved full-project goal remains the objective through all phases; do not stop at the first foundation. ## Retrieve locally Prefer an installed CLI. `npx` can fetch npm packages if the command is unavailable; use the bundled Python adapter when offline instead of implying that npx is offline. ```bash maestro-skil