daemon-backend-qwen-code

# Qwen Code Daemon Backend — Flag Discovery Entrypoint The installed CLI's own help is the authority for Qwen Code (npm package `@qwen-code/qwen-code`, binary `qwen`) flags; this page is only the entrypoint. Conversion rules, key safety, and persistence live in the parent [`reference/cli-backends/SKILL.md`](../../../SKILL.md). The alias `qwen` canonicalizes to the `qwen-code` backend id. ## Discover flags from the installed CLI 1. Run, in bash: `qwen --version` and `qwen --help`. The daemon backend wraps the top-level `qwen` binary directly — it spawns `qwen --yolo <backend_argv...> -p <prompt>`, no subcommand — so `qwen --help` is the whole relevant flag surface. These are local read-only commands; no session is started. 2. Translate what you found into `backend_options` with the parent's generic conversion rules. Nothing Qwen-specific is added to that contract here. Verify the installed `qwen --help` before passing `backend_options`; do not confuse provider setup with shell execution. ## Example: model selection via the generic route Your options land between the harness-owned `--yolo` and the final `-p <prompt>` (argv placement is pinned by `tests/test_daemon_backend_options.py::test_qwen_code_cmd_appends_backend_argv_before_prompt`): ```jsonc { "backend": "qwen", "tasks": [{ "task": "Implement and validate the change.", "tools": [], "backend_options": { "model": "qwen3-coder-plus" } }] } // argv: qwen --yolo --model qwen3-coder-plus -p <prompt> ``` The model vocabulary belongs to the installed CLI and the configured provider — LingTai does not validate, enumerate, or simulate model names. ## Subscription & auth Authenticates via the installed CLI's provider config (Qwen/DashScope keys); LingTai does not set or rotate credentials. Official docs: https://github.com/QwenLM/qwen-code ## Harness boundary Qwen Code reserves `--prompt`/`-p`, `--yolo`/`-y`, and `--approval-mode`: they drive LingTai's non-interactive headless harness, and passing any of them in `backend_options` refuses the whole batch before spawn. Beyond that, do not re-point harness-owned surfaces: the daemon writes a per-run `<run>/qwen-daemon-settings.json` (carrying `mcpServers.daemon_common` plus parent stdio MCP registrations) and injects it via the `QWEN_CODE_SYSTEM_SETTINGS_PATH` environment variable — overriding settings paths silently breaks completion enforcement. Plan flags at emanate time: `daemon(action='ask', input={'id': ..., 'message': ...})` is intentionally unsupported for this backend (no stable headless resume contract), so there is no later chance to adjust a running session. Output parsing is verbatim text — Qwen Code headless mode has no machine-readable event stream here, so stdout/stderr are recorded as-is and the result is the final stdout text; the `daemon_common` `finish(status="done")` call is still required for success.