Add Qwen Code as a supported agent

[QuanCheng-QC]

Why

Request, translated from the original:

Integrate Qwen Code (Alibaba Tongyi's coding CLI) into openagents as a new agent. Node side only (packages/agent-connector), no Python SDK. End-to-end with a real key: send a message, reuse the same session across turns, call an MCP tool, Stop.

Qwen Code (@qwen-code/qwen-code, binary qwen) is the first agent in the plan to add Chinese coding agents.

Before and after

On develop, Qwen Code is not in the catalog, the launcher cannot install it, and a workspace cannot run it: createAdapter('qwen-code') throws "Unknown agent type".

	Before	After
Catalog and launcher	Not listed	Listed and installable. Includes an icon, en/zh descriptions, and a model picker that lists from Model Studio
Workspace	Cannot be added	Replies to messages, keeps one CLI session per channel, uses workspace tools over MCP, and stops on Stop
Detection of a CLI the user installed	Not detected	Detected for npm global, nvm, and both official standalone installers (~/.local/bin, %LOCALAPPDATA%\qwen-code\bin, and QWEN_INSTALL_ROOT / QWEN_INSTALL_BIN_DIR)

How it works

One headless run per message:

qwen --output-format stream-json --approval-mode yolo \
  --exclude-tools cron_create cron_list cron_delete loop_wakeup \
  --auth-type openai --model <model> --mcp-config <file> \
  [--resume <session>] --append-system-prompt <workspace briefing>

• Input and secrets. The prompt is piped over stdin and the key goes in the environment, so neither appears in argv.
• Output. The frames have the same shape as Claude Code's. The reply is read from the result frame; assistant frames only drive the progress ticker.
• Sessions. Each channel's session is resumed with --resume. Qwen files sessions by working directory, so a session is only resumed from the directory it was created in. If the CLI cannot find the saved session, the turn is retried once from scratch. A stopped session stays resumable.
• Workspace tools. They come from the same stdio MCP server the Claude adapter uses. The config file is written 0600 and deleted when the run ends.
• Stop. On POSIX, Stop sends SIGINT to the process group, and the CLI kills its own shell tools. On Windows the npm entry script relaunches the CLI as a child process, so Stop uses taskkill /T.
• Binary. The npm package's cli-entry.js runs under the launcher's Node (the CLI needs Node 22 or newer), and a standalone install runs on the Node it bundles. Either way, the system prompt never passes through cmd.exe and its command-line length limit.
• Decision log and glossary. Both are pinned into the system prompt on every turn, including resumed turns.

CLI behaviour this relies on

Verified against Qwen Code 0.25.1. The minimum supported version is 0.16.0, the first release that accepts --mcp-config.

• Auto-memory delays every answer. With the default memory.enableManagedAutoMemory, a memory-extractor subagent runs after each turn and the result frame waits for it. On a resumed session the answer was ready after 7s but the frame arrived after 194s, and the extractor used 100k+ extra tokens. A system-defaults settings layer (QWEN_CODE_SYSTEM_DEFAULTS_PATH) turns auto-memory and self-update off. The user's own settings.json still takes precedence, and a machine-wide system-defaults file is merged in, not replaced. The CLI writes a $version stamp back into the file; the adapter ignores it, so the file is written only once.
• Credentials. The CLI only infers its OpenAI provider when both a key and a model are set, so --auth-type openai is always passed with a key. Key, base URL and model are taken as one set, either all from the agent's LLM_* fields or all from OPENAI_*, never mixed.
  • A blank base URL means Model Studio only for the agent's own key.
  • A key from the environment with no base URL goes explicitly to api.openai.com, so it is never sent to another provider.
• Saved providers override CLI arguments. A modelProviders entry in ~/.qwen/settings.json with the same model id redirects the request to its own endpoint, and can even swap in its own key. When the agent supplies a key, a per-run system-settings file clears modelProviders. Machine-wide system settings are carried over, and the temporary file is deleted after the run.
• API errors. Errors are streamed as [API Error: …] assistant text, followed by an is_error result frame. Release 0.16.0 instead wraps them in a successful result and exits 0. Both forms are reported as errors and never posted as an answer.
• Text encoding. stdout is decoded as a UTF-8 stream, so multi-byte characters split across chunks are not garbled.
• Scheduling tools are excluded. They schedule inside a one-shot process that exits when the turn ends, so a reminder set that way would never fire.

Configuration

LLM_API_KEY, LLM_BASE_URL and LLM_MODEL map to OPENAI_*. The base URL is required and prefilled with Model Studio China. The model is required for two reasons:

• Left blank, the CLI runs qwen3.7-max, its most expensive tier, which a Coding Plan account may not offer.
• The launcher's "Test connection" probes whatever the form holds, so a blank field tested a different model or endpoint than the agent would use.

The workspace's own agent form can still create the agent with only a key. The adapter then falls back to Model Studio and the CLI's default model.

Compatibility

• This only adds a new agent. Other adapters are unchanged, and paths.js only gains detection directories.
• No package or launcher version is changed. The launcher offers Qwen Code once a core release containing this adapter is published and the launcher's core dependency is updated.
• The branch merges cleanly with current develop (v1.1). The Qwen tests and the full agent-connector suite were re-run on the merged tree.
• Follow-up, not in this PR: v1.1 gave the Claude adapter's MCP env OPENAGENTS_REQUESTER_EMAIL for escalation tools. The Qwen adapter does not pass it yet, and neither does CodeBuddy's.

Tests

• qwen-code-stream.test.js and qwen-code.test.js: 77 tests, using a mock CLI that replays real transcripts (success, MCP call, missing key, 401, a 0.16.0-style error, a UTF-8 character split across chunks, stale resume, Stop, settings handling).
• qwen-code-cli.test.js: an opt-in real-CLI regression test. Set QWEN_CODE_TEST_CLI to a qwen binary; it uses a local HTTP endpoint and no paid key. It passes on 0.25.1 and 0.16.0 and fails on the code before the saved-provider fix.
• stop-control.test.js, agent-detection-matrix.test.js and the registry sync check now cover qwen-code.
• End to end with a real Model Studio key against a local workspace backend:
  • A message gets a reply.
  • Three turns run on one CLI session (--resume).
  • workspace_put_todos called over MCP writes todos to the workspace.
  • Stop during a long shell command ends the CLI and its tool process within about 1–2s.
  • The conversation continues on the same session after Stop.
  • The settings file is not rewritten after the first turn.
  • On the final commit, one check missed only because the model answered "PONG" when asked for "PONG-1", and the adapter relayed that verbatim.
• Tested manually on Windows.
• Known failures that also occur on develop:
  • wsl.test.js "an agent that only exists inside the distro" fails on the test machine because it reads the real ~/.openagents/runtimes.
  • The first test in launcher index.test.tsx is flaky.
  • Launcher tsc passes.

Docs: docs/agents/qwen-code.md.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
openagents-workspace	Ready	Preview	Oct 15, 2026 9:56am UTC