Summary: During the week of June 18, 2026, Qwen Code shipped v0.18.0 through v0.18.3 with more than 100 merged PRs, adding experimental opt-in Agent Team collaboration, durable /loop scheduling, and /cd directory switching. Durable /loop persists only requests that plainly imply persistence and expires periodic tasks seven days after their final trigger, while /cd validates paths and requires trust confirmation. Agent Team remains off by default and its model-driven teammates can vary in collaboration quality.
In the week of June 18, 2026, Qwen Code shipped four releases in quick succession — v0.18.0 through v0.18.3 — folding in more than 100 merged PRs. Rather than scattered tweaks, the batch goes after day-to-day friction: getting multiple agents to collaborate as a team in parallel, keeping scheduled tasks alive across restarts, and letting you change working directories without killing the CLI. Below is a walkthrough of what changed, the mechanics behind each piece, where it fits in practice, and what it costs.
Four Releases, One Week: The Lay of the Land
The week’s output spans four versions (v0.18.0–v0.18.3) and lands in five headline areas: Agent Team parallel collaboration, durable /loop scheduling, /cd directory switching, a collapsible thinking display, and an MCP approval gate. Alongside them sit smaller touches — speedier context compression through /compress-fast, permission bubbling for background agents, and workflow improvements aimed at parallel subagent execution.
Agent Team: Teammates That Stick Around
Under the Hood
/fork already let you run agents in the background, yet each one toiled alone — no channel existed between them, so merging their outputs was your job. Agent Team mode (experimental, opt-in) closes that gap. Switch it on with the experimental.agentTeam setting or the QWEN_CODE_ENABLE_AGENT_TEAM=1 environment variable, and the model can form named teams and spin up several “teammates” that work side by side. Teammates and the leader message one another, keep a shared task list, and the leader folds everything into one unified report at the end.
The pivotal design choice is persistence: teammates stick around instead of behaving like one-shot subagents that vanish the moment a task is done. They idle after finishing, waiting for whatever comes next. New tools hand the model the ability to create and delete teams, message teammates, and create, update, or list shared tasks, while the Agent tool picks up a name parameter for spawning long-lived, named teammates.
Where It Fits
- Send three reviewers over the same code at once — one on error handling, one hunting race conditions, one probing edge cases — and collect a single merged report.
- Or orchestrate a small team to put together a bookmark application; when teammates finish, they report back to the leader on their own.
What to Watch Out For
The mode is experimental and stays off until you turn it on. Because teammates are model-driven, collaboration quality hinges on how well the underlying model wields the team tools. Beyond the initial prompt you get no manual lever over teammate behavior. As an opt-in feature, it will not suit every workflow.
Durable /loop: Schedules That Outlive a Restart
The Mechanics
Every /loop schedule used to be chained to its session and quietly evaporated on restart. Durable /loop writes tasks to disk so they survive the process. When a request implies persistence — say, “check my PR every hour” — the task lands in ~/.qwen/tmp/<project-hash>/scheduled_tasks.json, one file per project. That file sits outside the workspace, so git status never surfaces it, and tasks come back automatically on the next launch.
The default remains session-only: throwaway reminders such as /loop 5m remind me never touch disk. A durable task is created only when a request plainly calls for persistence.
The edge cases are covered:
- A one-time reminder that came due while you were offline: after restart, the model checks with you before running it.
- Periodic tasks missed during downtime run one catch-up cycle once you are back, then settle into their normal cadence.
- With several sessions open on the same project, exactly one of them fires the persistent tasks; should that holder crash, another picks up the duty within seconds.
- Periodic tasks expire automatically 7 days after their final trigger.
When It Helps
- An hourly “check my PR” keeps running straight through a restart — no rescheduling needed.
- Periodic tasks that got missed catch up once, automatically, on startup.
- Short-lived reminders such as “remind me in 5 minutes” still die with the session.
Boundaries of the Feature
Only tasks that imply persistence get saved; session-only reminders stay as forgettable as ever. The 7-day auto-expiry may feel tight for some use cases, though it stops tasks from piling up forever. Multi-session coordination runs itself, yet rare edge cases could still expose race conditions.
/cd: Change Directories Without Leaving the CLI
What Happens Behind the Command
/cd <path> is a new slash command that moves the current session’s working directory without exiting the CLI. Under the hood it validates the target directory, runs the trust confirmation, refreshes workspace roots and the cwd service, carries the session’s files — transcript, runtime, and worktree — into the new workspace’s session directory, and updates the system/directory context. Path completion behaves as before.
An untrusted path brings up the same trust confirmation you met at first startup. You can navigate both up and down within a session; the command never detects directory changes on its own.
Everyday Scenarios
/cd ../another-projecthops to a neighboring project, with session files migrating on their own./cd /path/to/parent-repomoves you up and down a directory tree inside one session — zero restarts.
Constraints
A directory switch happens only when you type /cd; nothing is detected automatically. The command lives in the CLI and is not offered in headless or daemon modes. After moving between projects, confirm the active workspace before running anything.
More Worth Knowing About
- Thinking Collapsible: Reasoning blocks used to hog the screen; not anymore. While output streams, live reasoning scrolls inside a fixed-height, 4-line window beside a timer. Once finished, the block shrinks to a single line carrying the total duration (for instance, “Thought for 15s”); press
Ctrl+Oto unfurl the full reasoning. ThethinkingDisplayModesetting is gone, replaced by this single collapsible mode. - MCP Approval Gate: MCP servers declared at the project level — in
.mcp.jsonor workspace settings — no longer connect on their own; approval must come first. Each decision is recorded in<QWEN_HOME>/mcpApprovals.json, keyed by project root and server name, and tied to a config hash. Edit a server’s command, args, URL, environment, or headers and the hash shifts, sending it back to pending. That keeps hostile MCP configs from connecting without anyone looking. - Background Agent Permission Bubbling: When a background agent hits a tool call that needs your sign-off —
git pushandrm, for example — it can now bubble the request up to the parent session’s background tasks UI. SetapprovalMode: bubblein the subagent config: after approval the agent carries on; after denial it exits normally. Previously the agent auto-denied, which forced re-runs in the foreground. One caveat:bubbleapplies to subagents only; headless, ACP, and SDK contexts still auto-deny. /compress-fast: Context compression with zero LLM involvement — pure rules. It forces microcompaction, cleaning out old tool results and media content (the most recent N survive) and stripping every thinking part from model replies. No API calls, no token overhead, instantaneous. A quick way to reclaim space in a long session.- Workflow Enhancements:
parallel()andpipeline()bring concurrent fan-out to workflows (opt in viaQWEN_CODE_ENABLE_WORKFLOWS=1). Workflow P3 contributes schema, agentType, and model selection, plus worktree isolation for subagents. - Declarative Agent Frontmatter: Per-agent configuration now lives in
.qwen/agents/*.mdfrontmatter, which supportspermissionMode,maxTurns,color,mcpServers, andhooks. - Tool Output Truncation: Each tool gets an output budget — shell 30k, grep 20k, agent 32k — so an oversized dump cannot blow up the context. Whatever overflows is saved to temp files.
- Other Fixes: OOM protection, SSH clipboard, the missing Windows printf command, cursor movement at wrap boundaries, and more — the full changelog sits in the PR references.
Trade-offs to Weigh
Every one of these features earns its place, and each carries a cost:
- Agent Team remains experimental and model-dependent; not every model may use the team tools effectively.
- Durable
/loopintroduces a new system file under~/.qwen/tmp. Tasks bind to one project and do not survive a project hash change — rename the project directory and they are gone. - /cd switches the active workspace from inside the CLI. Great for moving between project directories, yet unavailable in headless or daemon modes.
- The MCP approval gate adds one more step when you open a new project, but it is a security measure worth that step.
- Permission bubbling for background agents works only in interactive CLI mode; headless and SDK contexts keep auto-denying.
/compress-fasttrades semantic quality for speed; reach for/compresswhen the summary’s quality matters.
Bottom Line
Already working in Qwen Code day to day? Pick up Durable /loop and /cd first — nothing else here changes your routine faster. Run parallel tasks that benefit from agents talking to each other? Turn on Agent Team. Give the MCP approval gate a read so the new security posture holds no surprises. And in a long session packed with tool calls, /compress-fast buys back tokens on the spot.
Taken together, the batch shows the Qwen Code team going after workflow interruptions methodically: a scheduled task should not evaporate because of a restart, changing directories should not cost you the CLI, and a background agent should not burn an entire run over a missing permission. Practical, thought-through, and worth adopting.