The interface is localized. Skill descriptions and documents are machine-translated by AI into your language; code, commands, paths and technical terms stay in the original, and the author's document governs.
Explore skills
Development · SKILL.md

conductor

conductor

Use when the user explicitly asks for Conductor, wants track/spec/plan/review artifacts on disk, or needs long-lived project context management. Prefer start-dev for small or one-off tasks. If Conductor is not configured in the project, start with setup.

Independent GitHub snapshot · Automated classification · Original documentation governs

Skill documentation

GitHub

name: conductor description: Use when the user explicitly asks for Conductor, wants track/spec/plan/review artifacts on disk, or needs long-lived project context management. Prefer start-dev for small or one-off tasks. If Conductor is not configured in the project, start with setup. version: 0.1.0 source: fork checksum: adf8bf4bc4f26b82d39d35ce6699ecd65504ead8143a243de1e0c59cc3b03620 updated_at: 2026-05-29T11:50:00+08:00 intent: execution layer: orchestration

Conductor

Conductor is a Context-Driven Development (CDD) workflow for managing durable project context on disk. The working loop is:

Refresh -> Spec -> Plan -> Implement -> Review -> Reconcile

This skill is the Codex-native adaptation of the upstream Conductor repository. Keep the workflow compatible with upstream concepts, but translate execution into Codex conventions instead of copying upstream-only behavior literally.

When To Use

Use Conductor when the user wants:

  • persistent artifacts under conductor/
  • a track with spec.md and plan.md
  • a governed multi-step implementation flow
  • a lightweight preflight summary before important work
  • review/cleanup/archive operations tied to the track lifecycle

Use start-dev instead when the task is small, one-off, or does not need new artifacts on disk.

Native Subagent Protocol (Codex)

Codex supports native subagents. Delegate with spawn_agent, coordinate with send_input, collect via wait_agent, and clean up with close_agent.

Execution preference:

  1. Use native subagents first for independent workstreams (parallel when possible).
  2. Merge results in main thread and run final verification.
  3. Fallback only when delegation is blocked: use the [ANALYST]/[ARCHITECT]/[EXECUTOR]/[REVIEWER] structure in a single response.

Minimal orchestration pattern:

spawn_agent -> send_input (optional) -> wait_agent -> close_agent

Upstream Alignment (Conductor)

When adapting or extending this skill, inspect the upstream repository as a protocol bundle, not just the README. The highest-value upstream files are:

  • upstream project instruction files
  • upstream extension metadata
  • policies/conductor.toml
  • commands/conductor/*.toml
  • templates/workflow.md

Preserve these upstream ideas when they improve Codex behavior:

  • conductor/ is the default plan directory
  • commands are protocol files; read the TOML prompt field before executing
  • review is a first-class command, not an afterthought
  • file resolution should work from semantic names, not only hard-coded paths
  • implementation and verification should prefer isolated review contexts when possible

The compatibility boundary for this skill is defined in:

  • docs/CONDUCTOR_V2_COMPATIBILITY_CONTRACT.md

Universal File Resolution Protocol (Codex Adaptation)

Before using Conductor files, resolve them in this order:

  1. If the project contains upstream metadata such as project instruction files, extension metadata, or policies/conductor.toml, read those first and honor their plan-directory declarations.
  2. Otherwise, default to the fixed conductor/ directory.
  3. Resolve semantic aliases as follows:
    • Tracks Registry -> conductor/tracks.md
    • Product Definition -> conductor/product.md
    • Product Guidelines -> conductor/product-guidelines.md
    • Tech Stack -> conductor/tech-stack.md
    • Workflow -> conductor/workflow.md
    • Track Folder -> conductor/tracks/<track_id>/
    • Archive -> conductor/archive/
    • Temp Workspace -> conductor/tmp/
    • Code Styleguides -> conductor/code_styleguides/

If the user asks for a custom artifact path, state that Conductor defaults to conductor/ and either:

  • proceed with the default directory, or
  • pause for a deliberate protocol refactor

Core Concepts

  • Track: A unit of work (feature or bug fix) with its own spec and plan
  • Spec: Detailed requirements document (spec.md)
  • Plan: Phased task list with checkboxes (plan.md)
  • Workflow: Rules for task lifecycle, TDD, commits, and quality gates

Directory Structure

When initialized, Conductor creates this structure in the project:

conductor/
├── product.md              # Product vision and goals
├── product-guidelines.md   # UX/brand guidelines
├── tech-stack.md           # Technology choices
├── workflow.md             # Development workflow rules
├── tracks.md               # Master list of all tracks
├── code_styleguides/       # Language-specific style guides
├── tracks/                 # Active tracks
│   └── <track_id>/
│       ├── metadata.json
│       ├── spec.md
│       └── plan.md
└── archive/                # Completed tracks

Available Commands

CommandPurpose
SetupInitialize Conductor in a project (new or existing)
RefreshRefresh the facts layer from code, config, and git state
New TrackCreate a new feature/bug track with spec and plan
ImplementExecute tasks from a track's plan following TDD workflow
ReviewReview a track or current changes against spec, plan, workflow, and code guidelines
StatusShow progress overview of all tracks
RevertGit-aware rollback of tracks, phases, or tasks

Protocol References

The detailed protocols are in TOML format. Read the prompt field from each file:

ActionProtocol File
Setup projectcommands/conductor/setup.toml
Refresh factscommands/conductor/refresh.toml
Create new trackcommands/conductor/newTrack.toml
Implement taskscommands/conductor/implement.toml
Review implementationcommands/conductor/review.toml
Check statuscommands/conductor/status.toml
Revert changescommands/conductor/revert.toml

How to read: Each .toml file has a prompt field containing the full protocol instructions.

Task Status Markers

  • [ ] - Pending
  • [~] - In Progress
  • [x] - Completed

Key Workflow Principles

  1. The Plan is Source of Truth: All work tracked in plan.md
  2. Test-Driven Development: Write tests before implementing
  3. High Code Coverage: Target >80% coverage
  4. Commit After Each Task: With git notes for traceability
  5. Phase Checkpoints: Manual verification at phase completion

When to Use Each Protocol

  • "set up conductor" or "initialize project" -> Read commands/conductor/setup.toml
  • "refresh", "sync current context", "update conductor facts" -> Read commands/conductor/refresh.toml
  • "new feature", "new track", "plan a feature" -> Read commands/conductor/newTrack.toml
  • "implement", "start working", "next task" -> Read commands/conductor/implement.toml
  • "review", "audit", "check this track" -> Read commands/conductor/review.toml
  • "status", "progress", "where are we" -> Read commands/conductor/status.toml
  • "revert", "undo", "rollback" -> Read commands/conductor/revert.toml
  • If the user does not mention Conductor and the task is small -> Use start-dev instead

Assets

  • Code Styleguides: templates/code_styleguides/ (general, go, python, javascript, typescript, html-css)
  • Workflow Template: templates/workflow.md

Critical Rules

  1. Validate every tool call - If any fails, halt and report to user
  2. Check setup first - Resolve and verify the required Conductor files before any operation
  3. Codex-first interaction - Prefer reasonable assumptions; ask one concise blocking question only when necessary
  4. Plain-language UX - Do not suggest slash commands. Tell the user what to ask for directly in natural language
  5. Codex-native delegation - Replace upstream-only flows such as ask_user or isolated CLI shells with Codex chat, native subagents, and local commands
Saved skills