drive-qwen

name	drive-qwen
description	Use when delegating a coding or research task to the Qwen Code CLI (`qwen`) from Claude Code. Covers one-shot positional prompts, approval modes, sessions (and the --chat-recording gotcha), output formats, and tmux interactive driving. Load delegate-to-agents first for the shared mechanism.

Drive Qwen Code CLI

Delegate to Qwen Code, a coding agent CLI (a Gemini-CLI-style fork). Read delegate-to-agents first for the shared Claude Code mechanism.

Flags verified against qwen 0.15.6 (qwen --help).

When to use

Coding, refactors, review, research — when you want the Qwen models or a cheap, fast second engine.

Preflight

command -v qwen && qwen --version
qwen auth        # configure auth: OpenRouter, Coding Plan, API Key, or Qwen-OAuth

Auth backends: OpenRouter, Qwen Coding Plan, API key, or Qwen-OAuth — set up via qwen auth. No valid backend → headless runs fail; fall back.

One-shot (PREFERRED): positional prompt

Unlike Gemini, Qwen's positional prompt defaults to one-shot ("use -i/--prompt-interactive for interactive"). So a bare prompt is already headless.

# one-shot, capture stdout
Bash(command="cd /repo && qwen 'Summarize the responsibilities of src/scheduler.go'", timeout=300000)

# write task, auto-approve edits
Bash(command="cd /repo && qwen 'Add input validation to the signup handler and tests' --approval-mode auto-edit", timeout=600000)

# stdin appended to the prompt
Bash(command="git -C /repo diff | qwen 'Review this diff; flag bugs and missing tests'", timeout=300000)

# long task: detach + log
Bash(command="cd /repo && qwen 'Refactor the parser' --approval-mode auto-edit > /tmp/agent-qwen.log 2>&1", run_in_background=true)

-p/--prompt still works but is deprecated in favor of the positional form.

Key flags

Flag	Effect
<prompt> (positional)	One-shot run (default)
-i, --prompt-interactive <text>	Run prompt, then stay interactive
-m, --model <name>	Model override
--approval-mode <mode>	plan (read-only), default (prompt), auto-edit (auto-approve edits), yolo (approve all)
-y, --yolo	Approve all actions
-o, --output-format <fmt>	text, json, stream-json
-s, --sandbox	Run tools in a sandbox
--system-prompt / --append-system-prompt	Replace / extend the system prompt
--include-directories, --add-dir <dirs>	Add workspace directories
--bare	Minimal mode: skip startup auto-discovery
--max-session-turns <n>	Cap agent turns

Note: --approval-mode auto-edit (hyphen). Gemini uses auto_edit (underscore). They differ — don't copy across drivers.

Sessions / resume — the --chat-recording gotcha

Resume/continue only work if chat history was recorded. --continue/--resume do nothing unless --chat-recording was enabled on the original run.

qwen --chat-recording 'Start the migration'   # record so it can be resumed
qwen -c 'Continue the migration'               # -c is a boolean (most-recent); the string is the positional prompt
qwen -r <session-id> 'Continue the migration'  # -r takes the ID; the prompt stays positional
qwen --session-id <session-id> 'Continue …'

Interactive (tmux)

Bash(command="tmux new-session -d -s qwen -x 200 -y 50")
Bash(command="tmux send-keys -t qwen 'cd /repo && qwen -i \"Add a health-check endpoint\"' Enter")
Bash(command="sleep 6 && tmux capture-pane -t qwen -p -S -60")
Bash(command="tmux send-keys -t qwen 'now add a test for it' Enter")
Bash(command="sleep 25 && tmux capture-pane -t qwen -p -S -80")
Bash(command="tmux kill-session -t qwen")

Rules

1. Positional prompt = one-shot — no extra flag needed for headless.
2. --approval-mode auto-edit (hyphen) for building; plan for read-only review.
3. Enable --chat-recording if you intend to resume later.
4. --output-format json to parse results.
5. One worktree per parallel run; clean up tmux/logs.
6. Review the diff and re-run tests yourself.