rhei-cli 0.3.0

Command-line driver for the Rhei agent runtime.
Documentation
name: supervised-delivery
version: 1.0.0
description: A delivery pipeline one supervisor drives — implement, code review alongside product review, fix, test coverage, and documentation — briefing every step and cancelling the rounds the results made unnecessary.

inputs:
  - name: spec_path
    description: The spec, issue, or design note that says what to deliver. The supervisor reads it on its first visit and turns it into the acceptance criteria every later step is judged against. Repository-relative; rendered verbatim, so the workspace stays portable.
    type: string
    positional: 1

  - name: title
    description: Human-readable title for the workspace and for the root delivery task.
    type: string
    default: Supervised delivery

  - name: supervisor_target
    description: Agent that supervises the whole subtree. It reads the spec, writes the preparation note, briefs every child, decides which rounds to run, and writes the delivery result. Give it your strongest reasoning target - every routing decision in this workspace is its call.
    type: string
    format: execution-target
    default: claude-code[yolo]:anthropic:claude-opus-4-7

  - name: implementer_target
    description: Agent that writes the implementation the reviews then judge.
    type: string
    format: execution-target
    default: claude-code[yolo]:anthropic:claude-opus-4-7

  - name: reviewer_target
    description: Agent that runs the code review rounds. Prefer a different agent from `implementer_target` so the review is independent of the hand that wrote the code.
    type: string
    format: execution-target
    default: codex[xhigh]:openai:gpt-5.5

  - name: pm_target
    description: Agent that runs the product review rounds alongside the code review. It judges user experience, predictability, and documentation against the spec and the project's goals rather than the code.
    type: string
    format: execution-target
    default: claude-code[yolo]:anthropic:claude-opus-4-7

  - name: fixer_target
    description: Agent that applies the fixes for both review rounds and coverage rounds. One fixer answers both reviews so a single hand reconciles them.
    type: string
    format: execution-target
    default: claude-code[yolo]:anthropic:claude-opus-4-7

  - name: coverage_target
    description: Agent that audits test coverage of what was implemented and names the tests that are missing.
    type: string
    format: execution-target
    default: codex[xhigh]:openai:gpt-5.5

  - name: docs_target
    description: Agent that updates the documentation the change makes stale.
    type: string
    format: execution-target
    default: claude-code[yolo]:anthropic:claude-opus-4-7

  - name: review_rounds
    description: How many code-review + product-review + fix rounds to unroll. The supervisor may cancel the rounds it does not need, so this is a ceiling, not a schedule. 2 is the usual ceiling; 1 for a small change; 3+ for a risky one.
    type: number
    default: 2
    validate: "[1-9][0-9]*"

  - name: coverage_rounds
    description: How many coverage-audit + fix rounds to unroll after the review phase. A ceiling like `review_rounds`.
    type: number
    default: 1
    validate: "[1-9][0-9]*"

  - name: docs_rounds
    description: How many documentation rounds to unroll after the coverage phase. A ceiling like `review_rounds`.
    type: number
    default: 1
    validate: "[1-9][0-9]*"

  - name: ci_commands
    description: The commands that must be green before the fix, coverage, and documentation steps report success. Rendered into their instructions verbatim. Empty means the agents discover the project's checks themselves.
    type: array
    items:
      type: string
    default: []

  - name: review_focus
    description: Extra subsections every code review must answer (e.g. concurrency, error handling, backwards compatibility). Empty means a general correctness review.
    type: array
    items:
      type: string
    default: []

  - name: supervisor_session
    description: Give the supervisor a snapshot session so each visit continues the previous transcript. Only set this true when `supervisor_target` names a session-capable agent - of the built-in profiles that is `pi` today. With the built-in `claude-code` profile a snapshot block is a hard validation error, so this stays false by default and the supervisor runs each visit cold, carried by its checkpoints and its briefs.
    type: boolean
    default: false