Qwen Code Weekly: Parallel Agent Panel, Auto-Memory On by Default, Worktree Phase D

The coding agent space heated up another notch this week. Cognition announced over $1B in funding at a$26B valuation, and Simon Willison declared that Anthropic and OpenAI have found PMF through coding agents —at the cost of significantly higher enterprise spending. Meanwhile, NVIDIA open-sourced Polar, a reinforcement learning framework , using Qwen Code as its test subject—boosting SWE-Bench scores from 3.8% to 26.4% based on Qwen3.5-4B.

Qwen Code shipped v0.16.2 this week with 30+ PRs merged, covering three directions users feel most directly:

Parallel agents are finally visible. Previously when multiple agents ran concurrently, all you could do was watch logs scroll. Now there’s a compact panel—one line per sub-agent, keyboard navigation, real-time progress. This is the first truly usable interaction interface for “agents evolving from single-threaded assistants to multi-threaded teams.”

Memory requires zero configuration. auto-dream and auto-skill are now on by default. The agent learns your preferences and project conventions as you use it, persisting across sessions. New /memory command to view and toggle anytime. Compared to Claude Code still requiring manual CLAUDE.md maintenance, this is a significant UX gap.

Isolated environment in one parameter. Worktree advanced from last week’s Phase C (session persistence) to Phase D—qwen-code --worktree launches directly into an independent work tree, --worktree=#<N> fetches remote PR code. Optional worktree.symlinkDirectories config symlinks heavy directories like node_modules to the main repo, avoiding redundant installs. Combined with /goal + Auto Approval, the agent runs a full task autonomously in isolation while keeping main branch clean.

Additionally, long conversation overflow saw real progress this week: auto-compression was redesigned from a single threshold to a warn / auto / hard three-tier ladder, reducing wasted headroom for large-window models from 30% to a fixed 33K. Headless mode gained --max-wall-time and --max-tool-calls runtime budgets—finally a hard limit for running agents in CI.

✨ New Features

Parallel Agent Panel: All Agents Visible at a Glance

If you’ve been following coding agent trends, “parallel” is the most mentioned keyword. Cognition’s $26B valuation sells the story of multiple agents working simultaneously. But when multiple agents are running, how do you know who’s doing what and where they are?

Qwen Code’s approach: when 2 or more agents execute in parallel, an inline compact panel (InlineParallelAgentsDisplay) renders automatically. Each sub-agent gets one line showing status, name, current operation, elapsed time, and token consumption. Keyboard ↑↓ switches focus, Enter expands details for full output, Esc returns to input box, and typing automatically cancels focus back to input—fully mouse-free.

What you can do with it:

• When /review and similar commands launch multiple sub-agents, see all progress at a glance (status/name/operation/time/tokens)
• Keyboard navigation: ↑↓ to select, Enter to expand details, Esc to return to input—typing auto-cancels focus
• Panel updates in real-time during live phase, then freezes as scrollable history on completion

See PR #4477 

Auto-Memory On by Default + /memory Toggle

Claude Code is pushing hard on auto-memory—users have to manually create CLAUDE.md and write rules for the agent to remember. The problem is many people don’t even know this file exists, let alone how to maintain it.

Qwen Code turned auto-dream and auto-skill on by default this week. This means: things the agent learns during conversation (your preferences, project conventions, common patterns) are automatically written to memory and loaded in the next session. No configuration, no file creation, works out of the box.

If you want manual control, the /memory dialog now includes an Auto-skill toggle row alongside existing Auto-memory and Auto-dream—three toggles navigated with arrow keys, toggled with Enter, settings persisted to workspace’s .qwen/settings.json.

What you can do with it:

• Zero configuration needed—all three managed-memory subsystems (summary memory, staged consolidation, skill extraction) are on by default
• Preferences and project conventions persist across sessions without repeating yourself
• /memory dialog provides three toggle rows to enable/disable any subsystem anytime
• Existing explicit settings are unaffected—new defaults don’t override your previous choices

See PR #4547 

Worktree Phase D: Launch Directly Into Isolation

Last week Worktree reached Phase C—session persistence + status bar display. This week jumps straight to Phase D: add --worktree at launch and Qwen Code automatically creates and enters an independent work tree. With PR refs, you can pull any PR’s code into an isolated environment for review or continued development without touching main branch.

Another practical config: worktree.symlinkDirectories settings support symlinking heavy directories like node_modules and .venv in the worktree back to the main repo, avoiding redundant dependency installs. This is opt-in—configure path list in .qwen/settings.json, and all subsequent worktrees (--worktree, enter_worktree tool, Agent isolation) apply automatically. Path traversal, absolute paths, .git and .qwen directories are safely rejected.

What you can do with it:

• qwen --worktree enters isolation with one parameter—no manual git worktree add
• --worktree=#<N> or GitHub PR URL: fetches remote PR via git fetch origin pull/<N>/head into a worktree for review or continued development
• Configure worktree.symlinkDirectories to share large directories like node_modules via symlinks without reinstalling
• Forms a complete loop with last week’s session persistence + /resume

See PR #4381 

Headless Protection: Runtime Budgets for Non-Interactive Mode

Anthropic published a Zero Trust AI Agent Security Framework  this week, noting that a major risk of autonomous agents is “running away with no one to stop them.” Baoyu (@dotey ) is saying the same thing: the key to using coding agents well lies at both ends—planning at the start and review at the end.

Qwen Code added two runtime budget parameters: --max-wall-time (maximum runtime, supporting formats like 90, 30s, 5m, 1h) and --max-tool-calls (maximum tool call count). When budget is exhausted, it terminates with exit code 55, distinct from user cancellation (130) and turn limits (53)—CI scripts can pinpoint the termination reason.

Additionally, when using --yolo (full auto-approval) without a sandbox in headless mode, a security warning is printed to stderr at startup, suppressible with QWEN_CODE_SUPPRESS_YOLO_WARNING=1.

What you can do with it:

• qwen -p "refactor module" --max-wall-time 5m --max-tool-calls 50—hard limits for agents in CI
• Auto-stops on budget exhaustion, exit code 55, CI scripts can distinguish “budget exhausted” from “completed normally”
• All budgets default to -1 (unlimited), existing usage unaffected
• --yolo without sandbox startup warning helps discover unsafe configurations

See PR #4502 

Three-Tier Auto-Compression: No More Long Conversation Overflow

Simon Willison says coding agents are significantly increasing enterprise costs. A core reason: long conversation context keeps growing, and when it approaches the model’s context window it either crashes or loses critical information.

Qwen Code redesigned auto-compression from a single threshold (triggered at 70% window) to a three-tier ladder, using both occupancy ratio and absolute byte count:

• warn tier (≈60% window or 20K from auto tier): alerts user that context is approaching limits—a UX-level early warning
• auto tier (≈70% window or 13K from window edge): proactively triggers compression, low cost, nearly lossless conversation quality
• hard tier (3K from window edge): forced compression—the last safety net before API rejects an oversized prompt

Small-window models (32K/64K) use the ratio branch, large-window models use the absolute byte branch—wasted headroom drops from 30% of window to approximately 33K fixed overhead.

What you can do with it:

• Long conversations no longer crash from context overflow—hard tier force-compresses before API rejection
• No manual configuration needed, system auto-calculates three-tier thresholds based on model window size
• Old chatCompression.contextPercentageThreshold config is deprecated, outputs deprecation warning at startup

See PR #4345 

Command Substitution No Longer Rejected Outright: Confirmation + ⚠️ Warning

Anthropic’s Zero Trust Security Framework  notes that agent security boundaries need to balance “usable” and “secure.” Qwen Code previously went too strict: if agent-generated shell commands contained $(), backticks, <() or other command substitution syntax, execution was flatly refused—even YOLO mode couldn’t bypass it.

Now it shows a confirmation dialog with a ⚠️ “Contains command substitution” warning. You see the warning and choose to allow or deny—the decision is yours, no more blanket rejections. YOLO mode executes normally with an audit record in DEBUG logs for post-hoc review.

What you can do with it:

• Legitimate scenarios like $(git rev-parse HEAD) to get commit hash are no longer blocked
• Confirmation dialog shows ⚠️ warning so you immediately know the command contains substitution syntax
• YOLO mode executes normally, DEBUG logs retain audit records
• Monitor tool fixes the same overly-strict rejection issue

See PR #4386 

More New Features

🔧 Notable Fixes

🙏 Contributors

Thanks to all contributors this week:

For questions or suggestions, feel free to open an issue on GitHub Issues !