Skip to content

Automation with Qwen Code

Running little-loops orchestration on the Qwen Code CLI via QwenRunner (scripts/little_loops/host_runner.py, ENH-3156). Flag behavior on this page was live-verified on qwen 0.21.6 by the FEAT-3155 spike (thoughts/research/qwen-code-surface.md).

Running orchestration CLIs

All orchestration tools (ll-auto, ll-parallel, ll-action, ll-harness, ll-loop, ll-sprint, plus the FSM evaluators and handoff) support qwen as the backing host CLI. Set LL_HOST_CLI=qwen to select it explicitly:

LL_HOST_CLI=qwen ll-auto
LL_HOST_CLI=qwen ll-parallel --workers 3
LL_HOST_CLI=qwen ll-action manage-issue --json
LL_HOST_CLI=qwen ll-loop run my-loop
LL_HOST_CLI=qwen ll-sprint run v2-launch

Auto-detection

You do not need to set LL_HOST_CLI explicitly if qwen is on PATH. resolve_host() in host_runner.py probes available binaries and selects QwenRunner when qwen is found and no override is set.

Detection order (first match wins):

  1. LL_HOST_CLI environment variable
  2. LL_HOOK_HOST environment variable
  3. Binary probe: claude → codex → pi → gemini → omp → kimi → qwen → error

You can also set the host permanently in .ll/ll-config.json (or .qwen/ll-config.json):

{
  "orchestration": {
    "host_cli": "qwen"
  }
}

What the runner emits

QwenRunner.build_streaming assembles argv as:

qwen --yolo --output-format stream-json [--continue] [--include-directories <dir>] [--model <alias>] -p "<prompt>"

with LL_NON_INTERACTIVE=1, DANGEROUSLY_SKIP_PERMISSIONS=1, and QWEN_CODE_SUPPRESS_YOLO_WARNING=1 (keeps stderr clean for error diagnostics) in the child environment. Qwen-specific facts:

  • --yolo (alias of --approval-mode yolo) is hidden from --help but live — the permission-skip path for headless runs.
  • -p <prompt> always terminates the argv (Claude/Kimi convention).
  • --continue resumes the most recent session in the project (requires general.chatRecording, default on).
  • --include-directories is additive, not a jail — workspace_root widens access rather than confining it (workspace_sandboxed=False).

Structured output — the headline capability

Qwen is the second host after Claude Code with structured_output=True:

  • The FSM evaluators append the inline --json-schema <json> flag when HostCapabilities.structured_output is true. The schema is Ajv-validated through a synthetic structured_output tool; the verdict arrives as a JSON string in the final envelope's result field (there is no claude-style structured_output envelope key), and the evaluators' existing json.loads(result) path handles it.
  • Session persistence for these one-shot evaluator calls is opted out via --chat-recording false — qwen rejects claude's --no-session-persistence flag (argv parse error), so _structured_output_args() keys the persistence opt-out on invocation.binary.
  • Hosts without the capability fall back to prompt-and-parse (BUG-2626 tag recovery) — nothing changes for them.

This re-enables the schema-constrained verdict path for ll-loop FSM evaluators on a non-Claude host for the first time.

Output shapes

  • Streaming (--output-format stream-json): JSONL — system/init → interleaved stream_event deltas (Qwen-internal, ignorable) → assistant messages (message.content[], message.usage) → final {"type":"result","subtype":"success","result":"…","usage":{…},"is_error":false}.
  • Blocking: build_blocking_json also streams (qwen's --output-format json buffers a JSON array that breaks single-envelope consumers — Kimi posture); consumers take the last result envelope.
  • Token reporting: usage (including total_tokens) is present on assistant messages and on the final result envelope.

Capability limitations (warn-and-drop)

Parameter Behavior
agent= Dropped with a CapabilityNotSupported warning — qwen has no --agent CLI flag (documented upstream as planned future work)
tools= Dropped with a warning — --exclude-tools is a denylist, not allowlist semantics
disable_background_tasks= Accepted, no-op (Protocol conformance; no qwen equivalent)

Current limitations

  • agent_select and tool_allowlist are ✗ (see above).
  • SessionEnd hooks do not fire under -p — session cleanup in headless runs rides the Stop legacy scripts (see Hook Events).
  • Qwen chat-file wire format is not parsed by ll-session backfill extraction yet (folder resolution works; ENH-3161 follow-up).
  • --include-directories widens rather than confines — treat workspace_root as access expansion.

See also