Qwen Code Weekly: /learn Turns Knowledge into Skills, Major Code Review Quality Upgrades, Web Shell UI Overhaul

Qwen Code shipped v0.19.9, v0.19.10, and v0.19.11 this week — 3 stable releases with 160+ PRs merged.

/learn command is live. Users can now proactively create Skills — provide a knowledge source (directory, URL, conversation, or text), and Qwen Code automatically organizes it into a SKILL.md for reuse later.

Code Review leveled up from “works” to “works well.” Suggestions are now posted directly as PR inline comments. Reviews gained procedural correctness checks, pre-release validation, a Step 4 verifier, and a Step 5 reverse-audit.

Web Shell got a major overhaul. Built on the shadcn component library, it now features an Artifact review panel, an extension management page, and Mermaid zoom/pan/drag. Multi-workspace support is fully wired end-to-end, with split panes color-coded by workspace.

Voice Bridge brings audio to ACP. Even a text-only primary model can now handle voice prompts — audio is first transcribed to text before reaching the model. Models that natively support audio get the raw audio passed through directly.

Scanned PDFs no longer stall the Agent. When text extraction exceeds the budget, pages are automatically rendered as images for the vision model to process.

CLI workflows rounded out. /update for one-click upgrades, Ctrl+O to freeze the transcript, Ctrl+S to stash your prompt, Explore inherits the primary model, Alt+T for full thinking output.

New Features

/learn: Turn Knowledge into Reusable Skills

/learn is a new slash command that lets you create reusable Skills from knowledge sources. A knowledge source can be a local directory, a URL, conversation history, or text you write directly. The primary model collects the knowledge and automatically compiles it into a SKILL.md, saved under .qwen/skills/learned-skill-<name>/ with a source: learned tag — completely separate from hand-written or auto-generated Skills.

What you can do with this:

• /learn Deploy process: ... — turn your deployment workflow into a Skill you can invoke next time
• Point Qwen Code at a docs directory to learn your project’s coding conventions, then follow them automatically
• Extract knowledge from a conversation and distill it into a reusable instruction template

See PR #6440 

Code Review Quality Gets a Major Upgrade

Code Review received quality improvements across multiple dimensions this week — from inline comments to review quality verification.

/review now posts Suggestion-type findings directly as PR inline comments — no more manually copying suggestions from review output to the PR. Reviews also gained procedural correctness finders, adjustable effort levels, and guardrail validation that verifies each finding actually holds before posting.

Large diff reviews are no longer assigned wholesale — every line of diff has a clearly assigned reviewer, ensuring full review coverage. The review chunk agent now receives fully constructed prompts instead of running on empty context. PR triage comments gained confidence scores, sequence diagrams, file overviews, and a review footer for richer reviewer context. Automated review default timeout is now 180 minutes with up to 240 minutes override — large PRs no longer get cut short.

What you can do with this:

• /review automatically posts inline suggestion comments on the PR, eliminating manual copy-paste
• Adjustable effort levels to match different review scopes
• PR triage confidence scores help you identify which PRs need closer attention

v0.19.11 further adds Step 4 verifier and Step 5 reverse-audit — the review workflow evolved from “find issues, post comments” into a complete loop: “find issues, verify each finding holds, reverse-audit to confirm nothing was missed.” The findings list is folded into the verify/reverse-audit prompt to reduce prompt size. Agent 7’s build/test scope is now narrowed to the workspaces actually touched by the diff, so large repos no longer trigger full builds.

See PRs #6593 , #6711 , #6612 , #6789 , #6771 , #6840 , #6942 , #6965 , #6892 , #6955 

Web Shell UI Overhaul

Web Shell received 15+ PRs this week, covering everything from the underlying UI component library to the sidebar, settings page, and model selector.

shadcn was introduced as the foundational UI component library, unifying design tokens and component standards. The multi-workspace sidebar got a redesigned, cleaner layout. The settings page moved to modern components — user-level settings can now be edited directly in the panel, and model management is integrated into the settings panel. The model selector was redesigned with clearer model metadata and faster switching. See PRs #6760 , #6804 , #6817 , #6758 

Multi-workspace support is now fully wired end-to-end. Daemon ACP endpoints support workspace scoping, channel workers are grouped and bound by workspace, split view supports cross-workspace sessions and workspace tabs. Dynamically added workspace registrations persist across daemon restarts — no reconfiguration needed. Workspaces can be removed at runtime without restarting the server. See PRs #6621 , #6635 , #6746 , #6716 , #6745 

A new right-side Artifact review panel — edited files appear as turn cards with line counts and expandable per-file diffs. The right panel supports both list and tree navigation with draggable width adjustment. Artifacts and scheduled tasks also have their own turn output cards, and HTML artifacts render in a sandbox iframe. See PR #6591 

v0.19.11 adds an extension management page — view and manage installed extensions directly in the Web Shell. Mermaid diagrams now support zoom, pan, and drag — flowcharts and sequence diagrams are no longer just static images. Split panes can be maximized individually, and each pane is color-coded by workspace. Composer controls moved to popovers, and the session list now shows sessions awaiting user action. See PRs #6815 , #6881 , #6951 , #6971 , #6877 , #6956 

Other improvements: Markdown table cells can be double-clicked to show full content in a popup for easy copying; the input toolbar now shows a Git branch chip; session groups support custom hex color codes; scheduled tasks from all workspaces are aggregated into a unified view; Composer supports custom placeholder text; workspace paths are now locked. See PRs #6530 , #6725 , #6752 , #6759 , #6765 , #6853 

Voice Bridge: Audio Input Meets ACP

Previously, sending voice prompts through ACP (Agent Communication Protocol) would fail if the primary model was text-only — audio simply couldn’t be processed. Now there’s a voice bridge: audio is first transcribed to text by the configured voice model, then passed to the primary model. If the primary model natively supports audio input, the raw audio is passed through directly without transcription. When transcription is unavailable or fails, the primary model receives a safe “audio unavailable” message instead of raw audio.

What you can do with this:

• Send voice commands through an ACP client — even text-only models can understand your speech
• Primary models with native audio support (e.g. multimodal models) keep the raw audio stream intact — no information loss
• Graceful fallback when transcription fails — audio issues won’t break your conversation

See PR #6576 

Automatic PDF-to-Image: Scanned Documents No Longer Stall the Agent

Previously, scanned or text-dense PDFs that exceeded the 12,000-token extraction budget would trigger a FILE_TOO_LARGE death loop — the Agent kept retrying with smaller page ranges, each time still exceeding the limit, inflating the context window until the session crashed. Now there’s a page-to-image rendering fallback: when text extraction exceeds the budget or fails, pages are automatically rendered as JPEG images via pdftoppm and handed to the vision model. Token cost per page is fixed regardless of text density. Text extraction remains the default path (grep-friendly, precise extraction); rendering only kicks in when the budget is exceeded or extraction fails. For text-only primary models paired with a vision model, scanned attachments via @ are rendered as a small set of pages and transcribed by the vision bridge.

What you can do with this:

• Scanned PDFs no longer trigger FILE_TOO_LARGE — the vision model reads the images directly
• Dense PDFs that exceed the token budget automatically fall back to images without blowing up the context
• Plain-text PDFs still go through text extraction — no precision lost

See PR #6585 

CLI Workflow Upgrades: One-Click Upgrades, Never Lose a Conversation

Previously, upgrading Qwen Code meant remembering your install method — npm install for npm installs, re-downloading for standalone. This week adds the qwen update CLI command and /update slash command, which auto-detect the install type and run the appropriate upgrade. Standalone installs pull the latest version directly from the npm registry with SHA256 checksum + Ed25519 signature verification (optional) + smoke test (the new binary must pass --version) + atomic replacement with automatic rollback on failure. npm installs display the corresponding npm install -g command. i18n support across 9 languages. See PR #5780 

Ctrl+O adds a frozen transcript view — press it to open an alternate screen showing the full conversation record with all tool outputs expanded, scrollable up and down. Exit with Esc/q/Ctrl+C/Ctrl+O to return to the main interface. Tool output rendering was also reworked: read/search/list tool results collapse into single-line summaries, while mutation tools stay expanded. The original compactMode setting is silently ignored. See PR #5666 

Ctrl+S adds prompt stashing — press Ctrl+S while the input box has content, and the prompt is saved to a project-level stash file. Next time you launch from the same project, it’s automatically restored. The stash is auto-deleted after submission. Large pasted placeholders are expanded before saving, so what you get back is the original content. Different projects don’t interfere with each other. See PR #6709 

Shell command timeout is now configurable. A new tools.shell.defaultTimeoutMs setting lets you adjust the default foreground command timeout at the project or session level — no need to specify it per command. Default remains 120 seconds; set to 0 to disable timeout. See PR #6628 

Explore subagent now inherits the primary model by default. The built-in Explore subagent no longer auto-selects fastModel — it inherits the primary session model. If you’re running Claude Opus as your primary, Explore also runs Claude Opus — no compromise on reasoning power. To keep using the fast model, set agents.builtin.exploreModel to fast or a specific model ID. See PR #6807 

Alt+T expands full thinking output. During streaming, pressing Alt+T expands the thinking block to show the complete reasoning content streaming in real time — no more hardcoded 4-line tail preview. Following the model’s chain of thought finally keeps up. Very long reasoning is capped by viewport height to prevent infinite expansion. See PR #6678 

More New Features

Important Fixes

Contributors

If you have questions or feedback, feel free to open an issue on GitHub Issues !