---

name: qwen-code description: Use Qwen Code CLI as a subagent. Lets the main agent prompt Qwen Code CLI from the terminal in headless mode with the qwen command, and send the goal, the decisions and the scope with the task. Use when the user asks to delegate work to Qwen Code CLI, when a task needs a second independent agent, or when a plan needs a fresh perspective. allowed-tools:

• Bash
• Read
• Write
• Grep
• Glob

---

Qwen Code CLI (subagent / task delegation)

What is it

This skill teaches your agent to use an agent from another vendor as a subagent.

Almost every coding agent has a headless mode. Headless mode runs the agent from the terminal with one command and returns the answer to stdout. This skill gives the command pattern for one such agent, and the protocol to transfer enough context with the task.

You stay the main agent. You keep the plan, the decisions and the conversation with the user. The subagent does one bounded task and reports back.

When to delegate

• The user asks for it. The user names another agent, or asks for a second opinion.
• Large or multi-step work. The task covers several files, phases or checkpoints.
• Code review with a fresh mind. Another agent reviews the work with a different lens and no memory of the decisions that produced it.
• Planning that needs different ideas. A second agent proposes options that you did not consider.
• An internal council. Several subagents answer the same question, and you compare the answers.
• Deep research. The other CLI has web search, web extraction or browser automation.
• Background investigation. A long task runs in the background while you continue in the foreground.

When not to delegate

• Small or single-step tasks. One or two edits, or a short explanation, are faster in this thread.
• Tight feedback loops. The user wants quick back-and-forth in one conversation.
• Secrets or policy-sensitive work. Do not pipe credentials into a subagent. Redact first.
• Context you already hold. Repeating the whole plan in a prompt adds no value.
• Low return on investment. If writing the handoff costs more than doing the edit, do the edit.
• Work that needs high precision and full judgement. A subagent without your context will break a task that depends on a detail only this session knows.

Delegation and context transfer protocol

An isolated subagent saves tokens, but it splits the story. The subagent does not see this conversation. It has no memory of the goal you agreed with the user, the decisions you already made, or the paths you agreed not to touch. A poor handoff causes misread tasks, wrong assumptions about the stack or the style, and wasted edits.

Treat every call as stateless delegation: the subagent keeps nothing between calls, so each command must carry its own state. A good handoff happens when you keep these practices in mind:

• Shared state handoff. The prompt carries the original goal, the decisions already made, and the explicit scope, to bridge the subagent's missing session history.
• Contextual preservation. The high-volume work stays in the isolated process, so the bulk never enters this conversation and no "split story" forms from missing history.
• Result reconciliation. What comes back is folded into this thread afterwards, so there is a single source of truth.

Write the prompt as a transfer of shared state, not as a title. Include all six fields:

Include	Why
Goal	The same objective as the user, not only the immediate micro-task.
Decisions	Framework, patterns, naming, "use X not Y" — anything that would otherwise be guessed wrong.
Scope	The absolute paths and modules to touch, and the areas to leave alone. A relative path may not resolve where you expect.
Constraints	Performance, accessibility, compatibility, review gates, "no new dependencies".
Verification	The exact command the subagent must run and pass before it returns.
Output	For example: "report only, no edits", "apply edits with a minimal diff", "list the files changed".

Prefer sequential delegations with explicit carry-over. Parallel runs diverge unless every run gets the same briefing.

Run the subagent in the mode that does the work without asking for approval. Headless mode has nobody to answer a permission prompt. Any mode that stops to ask will stall, or return an empty answer with a success exit code. Each CLI names this mode differently — take the flag from the vendor card below.

Never call a subagent in plan mode. Many CLIs have a plan or read-only mode. That mode makes the other agent write a plan for itself, which is not what you asked for: you want its work or its answer, and you keep the planning. Plan mode also waits for someone to approve that plan, so the run comes back with nothing. If you want no file changes, keep the auto-approving mode and write "report only, no edits" in the Output field. Do not use plan mode to make a run read-only.

After the subagent returns:

• Report to the user. Give a short summary. Do not paste long logs unless the user asks.
• Reconcile the context. Record the decisions it made, the files it changed, and the open risks, so this session stays the single source of truth.

CLI failure modes

Every vendor has its own terminal rules. A headless run fails quietly more often than it fails loudly.

• Wrap the call in an external timeout. A headless CLI can stall before its own timeout starts.
• Exit code 0 is not success. If stdout is empty, treat the run as failed and read stderr. A permission the CLI could not ask for, and a prompt that never arrived, both look like success.
• Never carry a flag habit from one CLI to another. The same short flag means different things in different tools. In OpenCode, -p is --password: passing a prompt to it empties the message and the run waits on stdin forever. Confirm every flag against the CLI's own --help.
• An unrecognized-flag error means this skill is stale, not that the task is impossible. Read --help, continue with the flags that exist, and tell the user which line in this file is wrong.

Vendor card

Binary and prompt form

• Binary: qwen
• Prompt form: flag. The prompt is the value of -p/--prompt "prompt". Content can also be piped in, e.g. cat file.ts | qwen -p "summarize".

Headless and output flags

Need	Flag
Run once and exit	-p, --prompt "prompt"
Output format	--output-format text|json|stream-json
Streamed input	--input-format stream-json
Wall-clock / call limits	--max-wall-time, --max-tool-calls, --max-subagent-depth, --deadline
Reasoning depth	--effort <level>

Approvals and permissions

This CLI does document an approval-mode ladder and a CLI flag for it — do not treat it as undocumented. --approval-mode <mode> accepts plan, default ("Ask Permissions"), auto-edit, auto, or yolo; --yolo is the shorthand for --approval-mode yolo. In session, modes also cycle with Shift+Tab (Tab on Windows) or /approval-mode [mode].

Mode	Behavior
plan	Read-only analysis; no edits or commands — do not use for delegation
default (Ask Permissions)	Manual approval required for edits and commands
auto-edit	File edits (edit, write_file, notebook_edit) auto-approved; shell commands still ask
auto	A classifier auto-approves read-only/build/test/in-workspace edits, blocks destructive patterns (rm -rf /, credential theft, external code execution)
yolo	Every tool call, including shell commands and file writes, auto-approved

The docs state plainly: "When running headless commands, Ask Permissions Mode is the default behavior" — i.e. without --approval-mode/--yolo, a plain qwen -p "..." defaults to a mode that asks for approval. What exactly happens to that ask with no TTY to answer it is not spelled out by the vendor (no explicit stall-vs-deny statement was found) — treat it as a stall risk and always pass --yolo or --approval-mode yolo for unattended delegation. Note also that --yolo only skips approval prompts; it does not enable a sandbox — --sandbox, QWEN_SANDBOX, or tools.sandbox control that separately, and untrusted delegated work should still run in a disposable environment.

Models and how to list them

No model-listing command is documented for this CLI. Select with --model <identifier>; there is no fixed, stable set of accepted strings recorded here — current identifiers split roughly into a coding-tuned tier for implementation and a higher-reasoning tier for complex architecture, but exact names shift often. --fallback-model <identifier> sets a fallback if the primary is unavailable. Search the web and consult Artificial Analysis for current identifiers, pricing, and benchmarks before picking one; if the user names a model, use it as given.

Command pattern

qwen -p "GOAL: [goal] | DECISIONS: [decisions] | SCOPE: [paths] | CONSTRAINTS: [constraints] | VERIFICATION: [test_command] | OUTPUT: [format]" \
  --yolo --output-format text --model <identifier> 2>&1

Prompt examples

• Implement: qwen -p "GOAL: [goal] | SCOPE: [paths] | VERIFICATION: [test_command] | OUTPUT: [format]" --yolo --output-format text --model <identifier>
• Heavy analysis: qwen -p "GOAL: Analyze complex architecture for potential deadlocks | SCOPE: [paths] | VERIFICATION: [check_command] | OUTPUT: analysis report" --yolo --model <higher-reasoning identifier>
• Investigate: qwen -p "GOAL: Map how [feature] works | SCOPE: [paths] | OUTPUT: concise file:line map" --yolo --output-format text

Docs and reference

• Flags, approval modes, skills, JSON output: reference.md
• Vendor documentation: https://qwenlm.github.io/qwen-code-docs/en/users/features/approval-mode/ and https://qwenlm.github.io/qwen-code-docs/en/users/configuration/settings/
• Checked against the official Qwen Code documentation on 2026-09-19. Not run against an installed binary — confirm with qwen --help before trusting a flag.