Summary: Qwen Code’s June 11, 2026 v0.18.0-preview update adds /fork background tasks, a /skills panel, and user-level cross-project memory to reduce long-task blocking, skill sprawl, and repo-switching context loss. The changes matter to teams managing agents, though file-boundary conflicts can erase fork time savings and prompt-driven memory routing plus stale preferences require human review. The experimental Workflow tool stays opt-in and off by default.
Stay with a coding agent long enough and three irritations show up fast: long tasks hold the main conversation hostage, the skill list outgrows what anyone can recall, and every repo switch wipes your preferences back to zero. Qwen Code’s weekly update of June 11 — v0.18.0-preview, 50+ PRs merged — takes on all three at once. /fork background tasks, the /skills panel, and user-level cross-project memory carry the main line of the release, with an interactive /stats panel, sleep prevention, and an experimental workflow tool at the edges.
/fork: Background Long Tasks, Context Fully Carried Over
Run /fork <instruction> and a background agent starts running at once, taking the entire working state along: system prompt, conversation history, tools, model settings, and the prompt cache. What it holds is your live context, not a distilled brief of the task. The main conversation keeps running unblocked; when the fork completes, the background task panel sends a notification and the results land in a JSONL transcript so you can review it afterwards.
The distinction from /branch is worth stating precisely: branch clones the conversation into a new session to explore a different direction; fork posts a background worker inside the current session, suited to independently verifiable subtasks such as “add tests for this module” or “investigate this error first.” Several forks may run in parallel, bounded by concurrency limits, and empty instructions or missing config produce explicit error messages rather than silent failure.
The cost of all this shows up at file boundaries: forks editing the same batch of files can burn every minute saved on conflict resolution. Parallelism lowers the price of waiting, not the rigor of acceptance — what each fork covers and what counts as done still need to be decided up front.
/skills: A Searchable, Lockable Panel Replaces Memorized Names
A skill once demanded two things of you: recall its exact name to use it, and hand-edit settings.json to disable it. /skills now opens a picker dialog — every skill listed with name, description, and source level, searchable by keyword, toggled via the Space key, and Enter placing the command into the input box, ready to run.
Scope is where the governance detail sits. The disabled state lands in the workspace-level skills.disabled setting and takes hold in the current session at once: <available_skills> and slash completions sync in real time, and no restart is needed. A skill disabled at the user level shows [locked: User] in the workspace panel, beyond the project level’s power to override. For teams, “which capabilities are available” stops being a verbal agreement and becomes a governable settings layer — upper levels lock, lower levels adjust, and the permission direction is unambiguous.
Cross-Project Memory: The Prompt Decides the Routing, Not Hardcode
The new user-level memory directory ~/.qwen/memories/ works beside the project-level .qwen/memories/ — no more “agent forgets everything on repo switch.” Routing is the key piece of the design: the existing four memory categories drive judgment, and the rules live in the memory extraction agent’s prompt rather than hardcode — user-type entries (personal preferences) go to user level, project-type entries (deadlines, events) stay at project level, and it adds zero new configuration. The two middle categories get defaults as well: feedback-type entries go to user level unless they encode explicitly project-specific conventions; reference-type entries stay at project level unless they name company-wide resources.
Both scope indexes render into the system prompt: user level first as background, project level after with more specificity, forming a natural general-to-specific order. Even an empty user-level MEMORY.md still gets its placeholder, so the agent knows the directory exists and accepts writes. The switch is the existing memory.enableManagedAutoMemory; no extra setup here either, and the practical gain is cheaper migration: categories, toggle, and rendering mechanism all come from the existing design, so you upgrade once and no new config file awaits.
The trade-offs deserve an honest statement: prompt-driven routing is not deterministic and will misclassify now and then; user-level memory collects stale preferences and calls for periodic human review. Still, the line sits in the right spot — personal style crosses repos with you, and repo-specific information never reaches the next workspace.
More in This Update
/stats turns into an interactive three-tab panel; usage data goes to ~/.qwen/usage_record.jsonl at session end and on /clear, giving token counts, costs, and cache hit rates a reviewable history at last. System sleep prevention defaults to on (general.preventSystemSleep: true) via caffeinate on macOS, systemd-inhibit on Linux, and SetThreadExecutionState on Windows — long tasks no longer fear a sleeping machine. A skill’s declared allowedTools are auto-authorized now, so /review and its kind run without a chain of permission prompts.
The experimental Workflow tool (P1) is opt-in, off unless enabled:
QWEN_CODE_ENABLE_WORKFLOWS=1With that environment variable set, the model writes JavaScript inside a node:vm sandbox to orchestrate several subagents. The idea is promising, yet the project itself tags it experimental — production should keep watching for now.
Advice for Builders
Three things to tackle today: give independently verifiable subtasks to /fork and save the main conversation for work that needs your judgment; disable the skills you never touch in /skills to shorten the agent’s available-command list; teach your preferences to the agent in one repo and let user-level memory carry them into the next project. Fold the memory content into routine maintenance too: check ~/.qwen/memories/ now and then, drop stale preferences, and re-file the entries that were misrouted. Scattered as they look, the three features share one thesis: the productivity bottleneck of agents is moving from model capability to management friction, and time (fork), capability (skills), and context (memory) each now have a governance surface of their own.