Code in Concert — version-pinned launcher for the spec-driven Maestro TUI. A premium terminal UI and scriptable pipeline for spec-driven agent orchestration with built-in tools for file operations, searching, shell execution, and workflow management.
Maestro defines 6 tools with schemas and descriptions, but quality is inconsistent. Tool names lack action verbs (read, write, grep are acceptable; ask and stipulate are vague). Descriptions vary widely: read/grep/write are clear (50-80 chars), but stipulate is a 400+ char wall of text mixing multiple concerns. Parameters are typed but many lack descriptions (e.g., 'workdir' in bash has no guidance on defaults or format). The stipulate tool conflates 4 distinct actions (engine, contract, delegate, contribution) into one tool, violating single-responsibility. Error handling is absent, no recovery guidance, no categorization of retryable vs fatal errors. Output schemas are not documented. The write tool's read-before-edit contract is clever but underdocumented for LLM understanding.
Request user input with a question and multiple choice options. Returns the selected option index.
Execute a shell command and return its output.
Search file contents recursively for a pattern.
Read a file and return its contents.
Maestro's Stipulate contract authority. Inspect status/plan, bootstrap, explore, validate, start an approved change, record fresh check evidence, document it, or delegate an approved DAG task. action='engine' takes argv and optional stdin (for plan --file -); action='contract' writes proposal.md or spec.md in an existing change; action='delegate' takes change and task. Approval and archive are human commands only. Returned worker contributions are not accepted; action='contribution' records accepted/rejected with a review reason.
Write content to a file. Enforces read-before-edit contract: existing files must be read first, and file must not have changed on disk since read.
stipulate tool violates single-responsibility principle: combines engine execution, contract authoring, task delegation, and contribution recording into one tool with 4 enum actions. Should be split into separate tools (execute_engine, write_contract, delegate_task, record_contribution) so LLM can reason about each independently.
bash tool lacks critical safety documentation: no mention of timeout, no guidance on shell injection risks, no output size limits stated in description (code shows 64KB cap but LLM cannot see it). Description does not warn that this is irreversible.
write tool's read-before-edit contract is underdocumented for LLM understanding. Description mentions the contract exists but does not explain what happens if LLM violates it (rejection? error message?). No guidance on how to recover from stale-file errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract structured data. E.g., grep returns what? Array of matches? Objects with line numbers? read returns raw string? Undocumented outputs force LLMs to guess.
ask tool name is vague (does not start with action verb). Should be 'request_user_input' or 'prompt_user_choice'. Current name does not signal to LLM that this is a user-interaction tool requiring human response.