Skip to content

runtime-qwen.md

Latest commit

 

History

History
150 lines (108 loc) · 5.61 KB

runtime-qwen.md

File metadata and controls

150 lines (108 loc) · 5.61 KB

Qwen Runtime

The Qwen Code extension lives in qwen/ (the output directory declared in src/platforms/qwen/runtime-config.js). The manifest (qwen-extension.json) and context file (QWEN.md) remain at the repo root so Qwen Code can discover the extension; generated artifacts (qwen/agents/, qwen/hooks.json) live in the subdirectory. It mirrors the Gemini CLI extension structure with Qwen-specific manifest, context, and tool mappings.

Configuration

Manifest: qwen-extension.json Version: generated from package.json Context File: QWEN.md

MCP Server

{
  "command": "node",
  "args": ["${extensionPath}/mcp/maestro-server.js"],
  "cwd": "${extensionPath}",
  "env": {
    "MAESTRO_RUNTIME": "qwen",
    "MAESTRO_WORKSPACE_PATH": "${workspacePath}"
  }
}

Qwen reuses the repo-root public server at mcp/maestro-server.js. The Qwen manifest launches that shared adapter with MAESTRO_RUNTIME=qwen; without that env var the adapter defaults to Gemini. The adapter requires canonical src/mcp/maestro-server.js directly and runs the Qwen runtime against shared source in src/. Qwen declares primary: filesystem and fallback: none.

Agent Naming

Qwen uses snake_case for agent names (matching Gemini's convention): code_reviewer, api_designer, accessibility_specialist.

Agent files are generated at qwen/agents/*.md with snake_case filenames (Qwen's own subdirectory, separate from Gemini's repo-root agents/).

Delegation

Direct function call syntax:

coder(query: "Implement the user service...")
architect(query: "Design the auth system...")

Commands

The Qwen runtime does not emit its own TOML command files. src/generator/entry-point-expander.js sets qwen: null in both ENTRY_POINT_CONFIG and CORE_COMMAND_CONFIG, so expandEntryPoints('qwen') and expandCoreCommands('qwen') return empty arrays. This is intentional: Qwen Code is Gemini-CLI-compatible and consumes the 12 TOML commands generated for Gemini at repo-root commands/maestro/ when both extensions coexist. See docs/runtime-gemini.md for the full command list.

Hooks

4 hook events in qwen/hooks.json:

Event Script Purpose
SessionStart hooks/hook-runner.js qwen session-start Initialize hook state, prune stale sessions
SubagentStart hooks/hook-runner.js qwen before-agent Detect agent, inject session context
SubagentStop hooks/hook-runner.js qwen after-agent Validate Task Report + Downstream Context
SessionEnd hooks/hook-runner.js qwen session-end Clean up hook state

SubagentStop Validation

Qwen uses the same post-delegation validation as Gemini:

  • Checks for ## Task Report (or # Task Report) and ## Downstream Context headings
  • First failure: blocks and requests retry
  • Second failure (stopHookActive=true): allows through with warning

Hook Adapter

Qwen reuses the shared Gemini-style adapter at hooks/adapters/qwen-adapter.js (or the shared Gemini adapter when the CLI JSON shape matches). Output format: { continue: boolean, systemMessage?: string }.

Policies

policies/maestro.toml — TOML-based shell guardrails evaluated by Qwen's policy engine (same rules as Gemini):

Deny (priority 950):

  • rm -rf, rm -fr, sudo rm -rf, sudo rm -fr
  • git reset --hard, git checkout --
  • git clean -fd, git clean -df, git clean -xfd, git clean -xdf
  • Heredocs (<<)

Ask User (priority 850):

  • tee commands
  • Output redirection (>, >>)

Tool Mapping

Qwen tools use canonical names with Qwen-specific overrides declared in src/platforms/qwen/runtime-config.js:

Canonical Qwen
read_file read_file
read_many_files read_many_files
list_directory list_directory
glob glob
grep_search grep_search
google_web_search web_search
web_fetch web_fetch
write_file write_file
replace edit
run_shell_command run_shell_command
ask_user ask_user_question
write_todos todo_write
activate_skill skill
enter_plan_mode enter_plan_mode
exit_plan_mode exit_plan_mode
codebase_investigator codebase_investigator

Feature Flags

The canonical feature set (same 4 flags across all runtimes, values per runtime):

exampleBlocks:             false
claudeStateContract:       false
scriptBasedStateContract:  true
codexStateContract:        false

See src/platforms/qwen/runtime-config.js for the authoritative values.

Agent Frontmatter

Qwen agent stubs share the Gemini shape:

---
name: coder
kind: local
description: "..."
tools: [read_file, write_file, ...]
temperature: 0.2
max_turns: 25
timeout_mins: 10
---

Fields: kind (always "local"), temperature, max_turns, timeout_mins.

Generated Files

qwen/
├── agents/                39 agent stubs (snake_case, Qwen tool names)
└── hooks.json             hook registration (SubagentStart, SubagentStop, …)

The Qwen extension reuses Gemini's repo-root commands/maestro/, hooks/, mcp/, and policies/ artifacts when both runtimes coexist; no duplicates are written under qwen/.

Notes

  • Qwen writes its own agent stubs and hook config to qwen/agents/ and qwen/hooks.json (separate from Gemini's repo-root outputs). For the /maestro:* command surface, Qwen reuses Gemini's repo-root commands/maestro/ TOML files — the Qwen generator does not duplicate these.
  • node scripts/generate.js derives qwen-extension.json from package.json with the other runtime manifests so release metadata stays aligned.