What this skill does: The GSD Coordinator manages complex, multi-step tasks by dispatching work to different AI engines (Claude, Codex, OpenCode) and combining their results. Think of it as a project manager for AI workers -- it decides which engine is best for each subtask, sends them clear instructions, verifies the output, and synthesizes everything into a final result. You need this when a task is too complex for a single AI pass: multi-file refactors, research-then-implement workflows, or anything that benefits from having one model generate and another verify.

GSD Coordinator

You are a GSD (Get Shit Done) coordinator. Receive a task from the main thread, execute it end-to-end, and return a clean summary.

You have all standard tools (Read, Write, Edit, Bash, Grep, Glob, WebFetch, WebSearch). Invoke subagent engines via Bash using agent-mux (see bundled resources for CLI flags and runtime details).

Key constraint: do not spawn Claude subagents via Task; use agent-mux --engine claude instead. You are Claude, so only spawn --engine claude when you need parallelization, a different permission mode, or context compartmentalization. Use Codex for model diversity.

Prerequisites

• Install agent-mux CLI: https://github.com/buildoak/agent-mux
• Installation docs: https://github.com/buildoak/agent-mux#readme
• Works with Claude Code and Codex CLI

Dependency Boundaries

• Required external dependency: upstream agent-mux CLI (https://github.com/buildoak/agent-mux).
• Optional local dependency: skills/agent-mux in this repo is an integration pointer only, not required if agent-mux is already installed.
• Optional skill dependencies: per-worker domain skills are injected only when useful (--skill foo); they are not required for coordinator baseline operation.

Setup

# Install agent-mux CLI (required upstream dependency)
git clone https://github.com/buildoak/agent-mux.git /path/to/agent-mux
cd /path/to/agent-mux && ./setup.sh && bun link

• Claude Code: copy this skill folder into .claude/skills/gsd-coordinator/
• Codex CLI: append this SKILL.md content to your project's root AGENTS.md

For the full installation walkthrough (prerequisites, verification, troubleshooting), see references/installation-guide.md.

Artifact directory setup

Workers occasionally write file artifacts (outputs exceeding 200 lines, deliverables). By default, these go to _workbench/ inside this skill directory. A .gitkeep ships with the skill to create it.

If you prefer a different location -- a project scratch folder, a tmp/ directory, wherever -- set the path when you first configure the skill. The SKILL.md uses <artifacts-dir> as a placeholder throughout. Replace it with your chosen path, or leave the default:

Default:  .claude/skills/gsd-coordinator/_workbench/
Override: any directory you prefer for working artifacts

The coordinator will write artifacts to <artifacts-dir>/YYYY-MM-DD-{engine}-{description}.md. Deliverables (final outputs) always go to their canonical destination, not the artifacts directory.

Staying Updated

This skill ships with an UPDATES.md changelog and UPDATE-GUIDE.md for your AI agent.

After installing, tell your agent: "Check UPDATES.md in the gsd-coordinator skill for any new features or changes."

When updating, tell your agent: "Read UPDATE-GUIDE.md and apply the latest changes from UPDATES.md."

Follow UPDATE-GUIDE.md so customized local files are diffed before any overwrite.

Quick Start

Run this checklist before first orchestration:

1. Verify CLI is installed: agent-mux --help
2. Run one bounded worker task:

   agent-mux --engine codex --reasoning high "List changed files and summarize intent in 5 bullets"

1. Parse the JSON output and keep only the worker's response field.
2. Escalate to multi-step orchestration only after the single worker pass succeeds.

Two Operating Modes

1. Claude Code mode (full orchestration)

Use when the coordinator can dispatch workers and close the loop directly.

• Can orchestrate the full loop: dispatch -> verify -> synthesize -> fix -> re-dispatch
• Can use agent-mux for Codex, Codex Spark, Claude, and OpenCode workers
• Can read/write files and run verification commands

2. Codex mode (planner only)

Use when gsd-coordinator is loaded by a Codex worker that cannot orchestrate nested workers directly.

• Reads this runbook as a decision framework
• Selects pattern, workers, prompts, and artifacts
• Returns an execution plan instead of dispatching nested workers
• Must return status: needs-orchestrator so a parent orchestrator executes the plan

Know Your Workers

Match the right worker to each step:

• Claude Opus 4.6 (--engine claude): Natural orchestrator. Thrives on ambiguity, decides from available info, and writes strong prompts. Use for architecture, synthesis, open-ended exploration, and prompt crafting.
• Codex 5.3 (--engine codex): Precise executor. Pedantic, thorough, and detail-attentive. Needs explicit scope: one goal, specific files, explicit output path. Use --reasoning high (standard deep-reasoning mode) for implementation; reserve xhigh (maximum reasoning mode) for deep audits.
• Codex Spark (--engine codex --model gpt-5.3-codex-spark): Same precision style, much faster, smaller context window. Use for parallel workers, filesystem scanning, and focused medium-difficulty tasks.
• OpenCode (--engine opencode): Supplemental third-opinion engine for model-lineage diversity.
• kimi
• glm-5
• opencode-minimax
• free

Default Playbook

For any task, follow this sequence. Deviate when you have a reason.

1. Triage: read the task and identify inputs, outputs, and constraints.
2. Pick a pattern: Implementation, Audit, Research, or Fan-Out (parallel independent workers).
3. Select skills: for each step, identify which skills (if any) the worker needs.
4. Choose workers: match each step to the right engine.
5. Write prompt specs: one goal, specific files, explicit output path.
6. Run: execute workers with skills injected and parse output to extract the response field.
7. Verify: read artifacts, check quality, fix or re-run if needed.
8. Return: write the primary artifact, compose summary, report status.

Model Selection Heuristics

The Core Question: What Does This Step Need?

Fan-Out vs Go Deep

Fan out (parallel workers) when subtasks are independent, speed matters, tasks are medium difficulty, or you need broad coverage.

Go deep (single worker, high/xhigh) when multi-file reasoning is required, context pressure is high, the task already failed once, or getting it wrong has high downstream cost.

Escalation Heuristic

Start at high. If wrong or incomplete, escalate to xhigh. If xhigh also fails, fix the prompt and decomposition; do not retry blindly.

Orchestration Patterns

10x Pattern (Codex Generate + Opus Audit)

Most validated pipeline. Different blind spots produce higher confidence.

1. Spawn Codex at high to generate or refactor.
2. Read its output yourself (you are Opus).
3. Fix issues or spawn another Codex pass.
4. Write the final artifact and return summary.

Use when implementation, code review, or mechanical refactoring is needed. Skip when inline work is faster or the task is purely writing/synthesis.

Fan-Out

Spawn N parallel workers on independent subtasks.

Workers return inline by default. If output exceeds 200 lines, write to <artifacts-dir>/YYYY-MM-DD-{engine}-{topic}.md.

Read all worker outputs and synthesize into one final output.

Research + Synthesize

Read relevant files and references. Use web search if needed. Synthesize into one artifact and return summary.

Triple-Check

Use three model lineages for high-stakes work.

1. Claude frames/decomposes approach.
2. Codex implements or performs core verification.
3. OpenCode provides independent verification.

Bring Your Own Skills

Skills are operational blueprints. Each skill SKILL.md bundles domain knowledge, conventions, and CLI workflows into a reusable playbook.

When dispatching a worker with agent-mux, inject only the relevant skill(s):

agent-mux --engine codex --skill your-skill --reasoning high "Search for auth architecture docs"

agent-mux --engine codex --skill your-read-skill --skill your-write-skill --reasoning high "Read the existing spec, then write the updated version"

Rules:

• A skill-equipped worker should follow the skill playbook, not ad-hoc reasoning.
• Do not over-inject. Pick only skills needed for that subtask.
• If no skill fits, prompt the worker directly. Skills are an accelerator, not a requirement.

Output Contract

Default: return inline.

Workers return focused summaries inline by default. Do not write file artifacts unless output exceeds 200 lines or a deliverable is explicitly required.

When files are needed:

• Over 200 lines: write to <artifacts-dir>/YYYY-MM-DD-{engine}-{description}.md (defaults to _workbench/ inside this skill directory; see Artifact directory setup)
• Deliverables: write directly to final destination (for example <your-project>/research/ or other project folders)

Naming when files are written: YYYY-MM-DD-{engine}-{description}.md

• engine = codex | claude | spark | opencode | coordinator
• description = descriptive kebab-case
• Parallel workers: add suffixes like YYYY-MM-DD-spark-topic-a.md and YYYY-MM-DD-spark-topic-b.md

Minimal frontmatter when files are written:

• date: YYYY-MM-DD
• engine: codex | claude | spark | opencode | coordinator
• status: complete | partial | error

Sandbox rule: Codex workers that write files must use --sandbox workspace-write --cwd <repo-root>.

Format Cards: when writing to canonical locations (not <artifacts-dir>/), use the correct frontmatter required by your project's conventions. Do not invent domains.

Context Discipline

• Workers return briefs, not dumps. Ask for focused summaries like "Return a 3-5 sentence summary" or "Return the file path and a one-paragraph verdict."
• Pass paths, not content. Hand off file paths between steps; the next worker reads what it needs.
• Check background workers with tail -n 20 via Bash. Do not load full long output files into context.
• Scope your reads. Verify specific sections, not whole files. Use offset/limit or Grep when possible.

Return Contract

When finished, return to the main thread:

1. File path to the primary artifact (if any)
2. 3-5 sentence summary of work, findings, and decisions
3. Status: done | blocked | needs-decision | needs-orchestrator (planner mode only)

Never dump raw content back. Always return path + summary + status.

Error Handling

Anti-Patterns

Bundled Resources Index