agent-mux

分类: 工具与效率 | 上传者: buildoak | 下载: 0 | 版本: v1.0（最新）
Unified subagent dispatch layer for Codex, Claude Code, and OpenCode engines. Spawn workers, run parallel execution pipelines, and get one strict JSON output contract.
更新日志: Source: GitHub https://github.com/buildoak/agent-mux
目录结构

当前层级: 根目录
📁 .github/
- 📁 workflows/
  - 📄 ci.yml 342 B
📁 references/
- 📄 engine-comparison.md 3.4 KB
- 📄 get-shit-done-agent.md 9.4 KB
- 📄 installation-guide.md 5.2 KB
- 📄 output-contract.md 6.4 KB
- 📄 prompting-guide.md 3.6 KB
📁 src/
- 📁 engines/
  - 📄 claude.ts 6.3 KB
  - 📄 codex.ts 6.5 KB
  - 📄 opencode.ts 13.7 KB
- 📁 mcp-servers/
  - 📄 agent-browser.mjs 13.8 KB
- 📄 agent.ts 745 B
- 📄 core.ts 31.3 KB
- 📄 mcp-clusters.ts 5.2 KB
- 📄 types.ts 2.1 KB
📁 tests/
- 📁 engines/
  - 📄 claude.test.ts 16.9 KB
  - 📄 codex.test.ts 16.1 KB
  - 📄 opencode.test.ts 19.6 KB
- 📄 core.test.ts 39.7 KB
- 📄 mcp-clusters.test.ts 8.7 KB
- 📄 output.test.ts 9.5 KB
📄 .gitignore 109 B
📄 bun.lock 25.9 KB
📄 CHANGELOG.md 1.9 KB
📄 LICENSE 1.0 KB
📄 mcp-clusters.example.yaml 2.0 KB
📄 package.json 961 B
📄 README.md 20.6 KB
📄 setup.sh 2.9 KB
📄 SKILL.md 14.4 KB
📄 tsconfig.json 320 B
📄 UPDATE-GUIDE.md 2.1 KB
📄 UPDATES.md 1.8 KB
SKILL.md

---
name: agent-mux
description: |
  Unified subagent dispatch layer for Codex, Claude Code, and OpenCode engines.
  Spawn workers, run parallel execution pipelines, and get one strict JSON output contract.
  Use this skill when you need to: dispatch a subagent, spawn an agent worker,
  run multi-model pipelines, invoke codex/claude/opencode engines, or coordinate
  parallel execution across AI coding engines. Covers unified output parsing,
  timeout/heartbeat behavior, skill injection, and MCP cluster configuration.
  Keywords: subagent, dispatch, worker, codex, claude, opencode, parallel execution,
  multi-model, spawn agent, engine, unified output.
---

# agent-mux

One CLI for Codex, Claude, and OpenCode with one strict JSON contract.

```bash
agent-mux --engine <codex|claude|opencode> [options] "prompt"
```

> If `agent-mux` is not linked globally, run `bun run src/agent.ts ...` from this repo.

---

## Setup

```bash
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 repo into `.claude/skills/agent-mux/`
- **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](references/installation-guide.md).

---

## Quick Reference

```bash
# Codex: implementation, debugging, concrete code changes
agent-mux --engine codex --cwd /repo --reasoning high --effort high "Implement retries in src/http/client.ts"

# Codex Spark: fast grunt work and broad scan/edit tasks
agent-mux --engine codex --model gpt-5.3-codex-spark --cwd /repo --reasoning high "Add doc comments across src/"

# Claude: architecture, reasoning, synthesis
agent-mux --engine claude --cwd /repo --effort high --permission-mode bypassPermissions "Design rollout plan for auth refactor"

# OpenCode: third opinion, model diversity, cost-flexible checks
agent-mux --engine opencode --cwd /repo --model kimi "Review this patch and challenge assumptions"

# Skill injection (repeatable)
agent-mux --engine codex --cwd /repo --skill react --skill test-writer "Implement + test dark mode"

# MCP clusters
agent-mux --engine claude --cwd /repo --mcp-cluster knowledge "Find canonical docs for token rotation"

# --browser sugar for browser cluster
agent-mux --engine codex --cwd /repo --browser "Open app, inspect controls, summarize findings"

# Full access mode
agent-mux --engine codex --cwd /repo --full "Install deps and implement requested fix"

# System prompt from file
agent-mux --engine claude --cwd /repo --system-prompt-file prompts/reviewer.md "Review this patch"

# Coordinator persona
agent-mux --engine codex --cwd /repo --coordinator reviewer "Audit this change for regressions"

# Combine coordinator + extra skill + inline system prompt
agent-mux --engine claude --cwd /repo --coordinator reviewer --skill perf-auditor --system-prompt "Prioritize latency risks." "Evaluate this service update"
```

---

## Engine Selection Protocol

Use this decision tree:

1. Code execution, file edits, implementation -> **Codex** with `--reasoning high`.
2. Fast grunt work, filesystem scanning, parallel worker throughput -> **Codex Spark** with `--model gpt-5.3-codex-spark`.
3. Architecture, deep reasoning, multi-file analysis, synthesis, writing -> **Claude**.
4. Model diversity, third-opinion checks, cost-flexible runs -> **OpenCode**.

_Note: Pratchett-OS coordinator uses Codex + Claude only._

> For engine-specific prompting tips, model variants, and comparison tables, see [references/prompting-guide.md](references/prompting-guide.md).

---

## CLI Flags

Source of truth: `src/core.ts` (`parseCliArgs`) + `src/types.ts`.

### Common (all engines)

| Flag | Short | Type | Values | Default | Notes |
| --- | --- | --- | --- | --- | --- |
| `--engine` | `-E` | string | `codex`, `claude`, `opencode` | required | Engine selector |
| `--cwd` | `-C` | string | path | current directory | Working directory |
| `--model` | `-m` | string | model id | engine default | Model override |
| `--effort` | `-e` | string | `low`, `medium`, `high`, `xhigh` | `medium` | Effort level |
| `--timeout` | `-t` | string | positive integer (ms) | effort-mapped | Hard timeout override |
| `--system-prompt` | `-s` | string | text | unset | Appended system context |
| `--system-prompt-file` | -- | string | path to file | unset | Loads file as system prompt, joined with other prompts |
| `--coordinator` | -- | string | coordinator name | unset | Loads spec from `<cwd>/.claude/agents/<name>.md` |
| `--skill` | -- | string[] | repeatable names | `[]` | Loads `<cwd>/.claude/skills/<name>/SKILL.md` |
| `--mcp-cluster` | -- | string[] | repeatable names | `[]` | Enables MCP cluster(s) |
| `--browser` | `-b` | boolean | true/false | `false` | Adds `browser` cluster |
| `--full` | `-f` | boolean | true/false | `false` | Full access mode |
| `--version` | `-V` | boolean | true/false | `false` | Print version |
| `--help` | `-h` | boolean | true/false | `false` | Print help |

### Codex-specific

| Flag | Short | Type | Values | Default | Notes |
| --- | --- | --- | --- | --- | --- |
| `--sandbox` | -- | string | `danger-full-access`, `workspace-write`, `read-only` | `danger-full-access` | `--full` also forces `danger-full-access` |
| `--reasoning` | `-r` | string | `minimal`, `low`, `medium`, `high`, `xhigh` | `medium` | Model reasoning effort |
| `--network` | `-n` | boolean | true/false | `true` | Enabled by default; `--full` also forces `true` |
| `--codex-path` | -- | string | path to Codex binary | unset | Overrides the Codex CLI binary; `AGENT_MUX_CODEX_PATH` is the env var equivalent |
| `--add-dir` | `-d` | string[] | repeatable paths | `[]` | Additional writable dirs |

### Claude-specific

| Flag | Short | Type | Values | Default | Notes |
| --- | --- | --- | --- | --- | --- |
| `--permission-mode` | `-p` | string | `default`, `acceptEdits`, `bypassPermissions`, `plan` | `bypassPermissions` | `--full` also resolves to `bypassPermissions` |
| `--max-turns` | -- | string | positive integer | effort-derived if unset | Parsed to number when valid |
| `--max-budget` | -- | string | positive number (USD) | unset | Parsed to `maxBudgetUsd` |
| `--allowed-tools` | -- | string | comma-separated tool list | unset | Split into string array |

### OpenCode-specific

| Flag | Short | Type | Values | Default | Notes |
| --- | --- | --- | --- | --- | --- |
| `--variant` | -- | string | preset/model string | unset | Used if `--model` absent |
| `--agent` | -- | string | agent name | unset | OpenCode agent selection |

### Canonical enum values (from `src/types.ts`)

- Engine names: `codex`, `claude`, `opencode`
- Effort levels: `low`, `medium`, `high`, `xhigh`

> For timeout/effort mapping, sandbox modes, and permission details, see [references/engine-comparison.md](references/engine-comparison.md).

---

## Output Contract

All engines emit one JSON payload to `stdout`. Parse JSON, never text.

Success shape: `{ success: true, engine, response, timed_out, completed, duration_ms, activity, metadata }`
Error shape: `{ success: false, engine, error, code, duration_ms, activity }`

Error codes: `INVALID_ARGS`, `MISSING_API_KEY`, `SDK_ERROR`.

Heartbeat: every 15s on `stderr` (`[heartbeat] 45s -- processing`). `stdout` is reserved for final JSON.

> For full JSON schema, field descriptions, and examples, see [references/output-contract.md](references/output-contract.md).

---

## Skills

Use `--skill <name>` (repeatable). Resolves from `<cwd>/.claude/skills/<name>/SKILL.md`.

- Skill content prepended as `<skill>` XML blocks
- If `<skillDir>/scripts` exists, prepended to `PATH`
- For Codex, skill directories auto-appended to sandbox `addDirs`
- Path traversal names are rejected

---

## Coordinator Mode

Use `--coordinator <name>` to load a coordinator spec from `<cwd>/.claude/agents/<name>.md`.

Coordinator file format: Markdown with optional YAML frontmatter.

- Frontmatter `skills` (array): auto-injected as repeated `--skill`
- Frontmatter `model` (string): default model unless CLI `--model` is set
- Frontmatter `allowedTools` (string or array): Claude-only, merged with CLI `--allowed-tools`
- Markdown body after frontmatter: coordinator system prompt content

Example coordinator file (`<cwd>/.claude/agents/reviewer.md`):

```md
---
skills:
  - code-review
  - test-writer
model: claude-sonnet-4-20250514
allowedTools:
  - Bash
  - Read
  - Write
---
You are a senior code reviewer. Focus on correctness, edge cases, and test coverage.
```

Example invocations:

```bash
# Load coordinator persona
agent-mux --engine claude --cwd /repo --coordinator reviewer "Review recent changes"

# Coordinator + file prompt + inline prompt (all three compose in order)
agent-mux --engine codex --cwd /repo --coordinator reviewer --system-prompt-file prompts/repo.md --system-prompt "Prioritize auth paths" "Audit this module"
```

Prompt composition order:

1. Coordinator body (`--coordinator`)
2. File content (`--system-prompt-file`)
3. Inline text (`--system-prompt`)

Flag interactions:

- `--model` on CLI overrides frontmatter `model`
- Frontmatter `skills` merge with explicit `--skill` flags (coordinator skills first)
- For Claude, frontmatter `allowedTools` merge with `--allowed-tools`

Validation and errors:

- Coordinator path traversal is rejected
- Missing coordinator file returns `INVALID_ARGS`
- Malformed frontmatter returns `INVALID_ARGS`
- `--system-prompt-file` resolves relative to `--cwd`
- Missing prompt file or directory path returns `INVALID_ARGS`

### GSD Coordinator Reference

agent-mux ships with a reference GSD (Get Shit Done) coordinator spec at [references/get-shit-done-agent.md](references/get-shit-done-agent.md).

This is a multi-step task coordinator that orchestrates Codex and Claude workers for complex pipelines. It includes:
- Model selection heuristics (when to use Codex vs Claude vs Spark)
- Orchestration patterns (10x Pattern, Fan-Out, Research + Synthesize)
- Output contracts and context discipline rules
- Anti-patterns to avoid

To use the GSD coordinator:

```bash
# Copy to your project
cp references/get-shit-done-agent.md <project>/.claude/agents/get-shit-done-agent.md

# Invoke via agent-mux
agent-mux --engine claude --cwd <project> --coordinator get-shit-done-agent "Complex multi-step task"

# Or from Claude Code via Task subagent
Task(subagent_type="gsd-coordinator")
```

Customize the frontmatter `skills` list and output paths for your project. The spec is a template designed to be adapted.

---

## MCP Clusters

Config search: `./mcp-clusters.yaml` then `~/.config/agent-mux/mcp-clusters.yaml`.

`--mcp-cluster` is repeatable. `all` merges all clusters. `--browser` is sugar for `--mcp-cluster browser`.

Bundled server: `src/mcp-servers/agent-browser.mjs` (browser automation).

See `mcp-clusters.example.yaml` for config format.

---

## Bundled Resources Index

| Path | What | When to load |
| --- | --- | --- |
| `references/output-contract.md` | Full output schema, examples, field descriptions | Parsing agent output, debugging response shape |
| `references/prompting-guide.md` | Engine-specific prompting tips, model variants, comparison | Crafting prompts for specific engines |
| `references/engine-comparison.md` | Detailed engine table, timeouts, sandbox/permission modes | Choosing engine config, debugging options |
| `references/get-shit-done-agent.md` | GSD coordinator spec (reference template) | Setting up multi-step task coordination |
| `references/installation-guide.md` | Full installation walkthrough | First-time setup, prerequisites, troubleshooting |
| `src/agent.ts` | CLI entrypoint and adapter dispatch | Trace invocation path |
| `src/core.ts` | parseCliArgs, timeout, heartbeat, output assembly | Always for behavior truth |
| `src/types.ts` | Canonical engine/effort/output types | Always for contract truth |
| `src/mcp-clusters.ts` | MCP config discovery and merge logic | MCP cluster setup/debug |
| `src/engines/codex.ts` | Codex adapter | Codex option/event behavior |
| `src/engines/claude.ts` | Claude adapter | Claude permissions/turn behavior |
| `src/engines/opencode.ts` | OpenCode adapter + model presets | OpenCode model routing |
| `src/mcp-servers/agent-browser.mjs` | Bundled browser MCP wrapper | Browser automation integration |
| `setup.sh` | Bootstrap script | First install or environment repair |
| `mcp-clusters.example.yaml` | Starter MCP config | Creating cluster config |
| `CHANGELOG.md` | Release history | Verify version-specific behavior |
| `tests/` | Test suite | Validate changes/regressions |

---

## Timeout Alignment

When calling agent-mux from a wrapper (Claude Code `Task`, Bash `timeout`, shell scripts), the wrapper's timeout **must exceed** agent-mux's internal timeout by at least 60 seconds. Otherwise the wrapper kills the process before agent-mux's graceful timeout path fires, losing activity logs and the `timed_out: true` JSON response.

| Effort | agent-mux timeout | Wrapper minimum timeout |
|--------|-------------------|------------------------|
| `low` | 2 min (120s) | 3 min (180s / 180000ms) |
| `medium` | 10 min (600s) | 11 min (660s / 660000ms) |
| `high` | 30 min (1800s) | 31 min (1860s / 1860000ms) |
| `xhigh` | 45 min (2700s) | 46 min (2760s / 2760000ms) |

**Rule:** `wrapper_timeout = agent_mux_timeout + 60_000ms`

## Completed Field

The `completed` boolean on success output is the single source of truth for whether work finished:
- `completed: true` — work ran to completion (normal exit)
- `completed: false` — work was interrupted (timeout or shutdown)

Callers MUST check `completed` (or at minimum `timed_out`) before treating output as final. Checking only `success` will treat timeouts as completions.

---

## Anti-Patterns

- Do not parse agent-mux output as text. Always parse JSON from stdout.
- Do not run parallel browser workers. One browser session at a time.
- Do not read agent output files with full Read; use `tail -n 20` via Bash.
- Do not use `--reasoning minimal` with MCP tools (Codex rejects them).
- Do not send exploration tasks to Codex; use Claude for open-ended work.
- Do not use `xhigh` effort for routine tasks; `high` is the workhorse.

---

## 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 agent-mux 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.
登录后下载/点赞/收藏 ❤ 28 | ★ 0
agent-mux

目录结构

SKILL.md

举报内容

提示
