Qwen Code Weekly: Agent Team Parallel Collaboration, Durable /loop Survives Restart, /cd Switch Directory

Qwen Code released four official versions v0.18.0~v0.18.3 this week, with over 100 PRs merged, covering four directions:

Multi-Agent parallel collaboration. Previously /fork ran background Agents, but each Agent operated independently with no communication, and you had to assemble the results yourself. Agent Team mode lets the model create named teams, spawn multiple teammates to work in parallel, exchange messages, share task lists, and finally the leader consolidates a single report. Teammates are persistent — after completing one task they can wait for the next, not one-shot subagents that disappear after finishing.

Scheduled tasks surviving restart. Previously /loop scheduled tasks perished with the session — silently vanishing after restart. Durable /loop persists scheduled tasks to files, automatically recovering on next launch, with missed periodic tasks running one catch-up cycle.

In-session directory switching, no restart needed. /cd <path> switches the current session’s working directory without exiting the CLI, session files auto-migrate to the new workspace, system/directory context auto-refreshes. Previously you had to restart to switch project directories.

✨ New Features

Agent Team: Multiple Agents Collaborating in Parallel, Messaging Each Other, Sharing Tasks

Previously Qwen Code’s /fork could run background Agents, but each Agent was isolated — you arranged three Agents to review code in parallel, they had no communication channel between them, and you had to manually combine the three results.

This week added Agent Team mode (experimental, opt-in). Once enabled, the model can create named teams and spawn multiple “teammates” to work in parallel. Teammates and the leader can exchange messages and share task lists, with the leader ultimately consolidating a unified report. Teammates are persistent, not one-shot subagents — they can finish one task and wait for the next, rather than disappearing after completion.

Disabled by default (enable via experimental.agentTeam setting or QWEN_CODE_ENABLE_AGENT_TEAM=1 environment variable). Once enabled, the model gains team collaboration tools (create/delete teams, message teammates, create/update/list shared tasks), and the Agent tool adds a name parameter to spawn long-lived named teammates rather than disposable subagents.

What you can do with it:

• Have three reviewers review code in parallel (one for error handling, one for race conditions, one for edge cases), and give you a consolidated report
• Coordinate a team to build a small bookmark application
• Teammates automatically report to the leader upon completion, and the leader gives you a unified summary

See PR #4844 

Durable /loop: Scheduled Tasks Surviving Restart

Previously all /loop tasks perished with the session — scheduling “check my PR every 30 minutes” would silently disappear after restarting Qwen Code. This week /loop tasks can survive restarts.

When a request implies persistence (“check my PR every hour”), the task is saved per project to ~/.qwen/tmp/<project-hash>/scheduled_tasks.json (not in the workspace, won’t appear in git status), and automatically recovered on next launch. Default remains session-only, temporary reminders don’t write files.

One-time reminders that expired during downtime: on restart the model confirms before executing, never running silently. Missed periodic tasks run one catch-up cycle after startup then resume normal rhythm. When multiple sessions exist for the same project, only one triggers persistent tasks; if the holder crashes, another takes over within seconds. Periodic tasks auto-expire after 7 days (deleted after last trigger).

What you can do with it:

• “Check my PR every hour” → continues running after restart, no need to reschedule
• Missed periodic tasks automatically catch up once on restart
• Temporary reminders /loop 5m remind me still perish with the session, don’t write files

See PR #5004 

/cd: Switch Directory Without Restarting CLI

Previously, if you wanted the Agent to work from another project directory, you had to exit and restart. Switching between multiple repos or parent/child workspaces was inconvenient, with no way to move active tool context within a session.

This week added /cd <path> slash command. It validates the target directory, handles trust confirmation, updates workspace roots and cwd service, migrates the current session’s transcript/runtime/worktree files to the new workspace session directory, and refreshes system/directory context. Path completion still works.

What you can do with it:

• /cd ../another-project switch to adjacent project, session files auto-migrate
• /cd /path/to/parent-repo navigate up/down within session, no restart needed
• Untrusted paths trigger confirmation, consistent with the trust flow on first startup

See PR #4890 

Thinking Collapsible: Reasoning Process No Longer Dominates Screen

Previously the model’s thinking blocks always displayed the full reasoning text, occupying massive screen space. This week changed to a three-state display:

• Streaming phase: 4-line fixed-height scrolling window showing live reasoning content + timer (”∴ Thinking… 8s”)
• Collapsed after completion (default): one line showing total duration (”∴ Thought for 15s”)
• Expand to view: a dedicated shortcut key for thinking blocks will be assigned later; currently Ctrl+O triggers compactMode toggle, sharing the same key with thinking expand, independent expand/collapse shortcuts will be decoupled in a subsequent PR

thinkingDisplayMode setting was removed, replaced by a single collapsible mode. Thinking blocks are UI history only, not sent to the API, session resume behavior unchanged.

What you can do with it:

• Complex reasoning collapses to one line, not cluttering the view
• Ctrl+O to expand and view the model’s complete reasoning process anytime
• Streaming phase has a timer, showing how long the model has been thinking

See PR #4598 

MCP Approval Gate: Project-Level MCP Config No Longer Auto-Connects

Previously MCP servers declared in project .mcp.json or workspace .qwen/settings.json would auto-connect — opening a repo containing a malicious .mcp.json could start the server before you reviewed it.

This week added an approval gate. Project .mcp.json and workspace scope MCP servers must be approved before connecting. Approval decisions are persisted in <QWEN_HOME>/mcpApprovals.json, indexed by project root + server name, and bound to a config hash. Changing a server’s command/args/URL/environment/headers changes the hash, reverting it to pending — won’t silently inherit old approvals.

MCP server priority model: user/default settings < project .mcp.json < workspace/system settings < session (ACP/IDE) < --mcp-config. Project config can override user config but cannot override enterprise system config.

What you can do with it:

• Opening a new repo, project-level MCP servers show as “Pending approval”, no auto-connect
• qwen mcp approve <name> or qwen mcp approve --all for batch approval
• Modifying server config auto-requires re-approval (hash binding)
• qwen mcp list to view all server approval statuses

See PR #4713 

Background Agent Permission Bubbling: Background Tasks No Longer Auto-Deny Confirmations

Previously background Agents requiring permission confirmation for tool calls (like git push, rm) would be auto-denied with “background agents cannot prompt for confirmation”, then the parent session would re-run in foreground — essentially the background run was wasted, queuing back to foreground.

This week added approvalMode: bubble. When a subagent config sets approvalMode: bubble, background Agents encountering tool calls needing confirmation will bubble the request to the parent session’s Background tasks UI. You can open details from the footer pill to approve. After approval the Agent continues execution; after denial it exits normally. bubble is a subagent-only approvalMode value, not a session-level option. In headless/ACP/SDK contexts it still auto-denies (no UI to answer).

What you can do with it:

• Custom subagent declares approvalMode: bubble, background tasks no longer auto-deny
• Footer pill shows ”⚠ needs approval”, press ↓ to open details for approval (single authorization only, no “always allow” option)
• After approval the Agent continues and returns results, no need to re-run in foreground

See PR #4955 

/compress-fast: No LLM Call, Instant Context Compression

Previously /compress relied on a LLM side query to generate a summary (~2-5 seconds, consuming ~30K tokens). Too slow for local model deployments or users who just want to quickly reclaim space.

This week added /compress-fast, purely rule-driven: forces microcompaction to clean old tool results and media content (keeping the most recent N), then strips all thinking parts from model replies. Zero API calls, zero token overhead, instant response. Writes chat_compression checkpoint to JSONL, --resume behavior identical to /compress.

What you can do with it:

• /compress-fast immediately reclaim thousands of tokens, no delay no cost
• Ideal for long sessions after many tool calls when you want to quickly reclaim space
• /compress still available — use it when you need semantic summary quality
• Two consecutive /compress-fast calls, the second prompts “No compression needed”

See PR #4893 

More New Features

🔧 Key Fixes

🎉 Contributors

Thanks to this week’s contributors:

For questions or suggestions, feel free to share feedback at GitHub Issues !