Original release notes
Highlights
See the complete change list below.
Breaking Changes
- refactor(node-repl)!: deliver the persistent Node REPL as a standalone MCP server (#9499) by @LaZzyMan
Complete Change List
Features
Added temporal-reachability and incident-replay lenses to review briefs to analyze whether values exist at the correct moments to steer decisions. (#9708) by @wenshao Review findings now include specific test acceptance criteria for their fixes and explicitly rule on non-convergence to reduce review-fix loops. (#9596) by @wenshao Daemon sessions now restore to their last selected model across restarts and detaches, while new sessions continue to use the current settings default. (#9687) by @doudouOUC Validates inline anchors for /review --comment on Aone targets before posting, relocating Critical findings to the summary body and discarding Suggestions if anchors are invalid. (#9634) by @wenshao Adds a tokensUsed field to GoalRecord to report the total tokens consumed by a Goal alongside its turn count in the lastGoal summary. (#9583) by @qqqys Enables dynamic workflows via the tools.workflowsEnabled user setting, replacing the previous undocumented environment variable method. (#9098) by @qqqys Enables PR context fetching for Aone Code targets in the /review command, restoring full review capabilities including native approval and issue fidelity. (#9621) by @wenshao Web Shell now correctly refreshes composer skills after toggles, preventing disabled skills from reappearing and surfacing failed refreshes instead of stale suggestions. (#9131) by @samuelhsin The review skill now supports comment-status and presubmit flows for Aone Code targets, generating the same JSON report schema as GitHub for merge-gate and CI state checks. (#9627) by @wenshao feat(models): support dual-role image generation models (#9650) by @callmeYe Review compose now surfaces a land-with-residual-risk advisory when critical findings persist across rounds, signaling that manual risk acceptance is required to exit the loop. (#9526) by @wenshao The Computer Use SDK for Node.js is now updated to v0.20.0, providing versioned accessibility observations and stable element tokens across macOS, Windows, and Linux platforms. (#9587) by @LaZzyMan Added support for MCP 2026 core protocol and introduced an MCP Apps host for rendering inline applications in daemon-backed WebShell sessions. (#8992) by @samuelhsin Adds a do-not-refute list and constructible rejection bar to the /review skill verifier to prevent invalid rejections of speculative findings. (#9799) by @wenshao feat(core): declare create_sub_session only under qwen serve (#9425) by @DragonnZhang Added support for creating scheduled tasks that reuse an existing live session by providing an optional sessionId. (#9361) by @yiliang114 Compact mode now folds adjacent thinking, tool activity, and parallel agents into a single expandable summary row in the web shell. (#9657) by @ytahdn Kimi (Moonshot AI) is added as a built-in third-party provider with support for international and China endpoints and editable model catalogs. (#9814) by @TianYuan1024 The Workflow tool now displays a detailed approval dialog showing the script's name, phases, and arguments before execution. (#9837) by @qqqys WebShell now exposes an optional callback to report the active session's complete subagent task snapshot for embedding hosts. (#9637) by @callmeYe Added an optional async prepareSubmit callback for Web Shell hosts to resolve context dynamically before submission. (#9802) by @dreamWB Adds four built-in output styles (Concise, Proactive, Explanatory, Learning) to control how the agent reports work throughout a session. (#9565) by @qqqys Introduces qwen serve --open-with-auth to open a loopback Web Shell with bearer authentication using a provided or auto-generated token. (#9738) by @doudouOUC Adopts the shared WebShell transcript renderer as the default and only conversation timeline in the VS Code companion, replacing the legacy renderer. (#9719) by @yiliang114 Replaced built-in Computer Use tools with a bundled skill that automatically configures the external Node REPL MCP server and CUA SDK on first use. (#9856) by @LaZzyMan Promoted language-pitfall and wrapper/proxy checks from Agent 1a to dedicated high-effort Step 3A roles for more thorough code review. (#9805) by @wenshao The /review command now adds an advisory note when the overall approach, rather than the current patch, is identified as the primary open question. (#9340) by @qqqys Added a new find-simplifications sweep skill that identifies dead code and orphaned resources, generating evidence-backed proposals for maintainer review. (#9384) by @qqqys Required explicit user opt-in before launching workflows by defining five specific triggers to prevent unintended multi-agent orchestration. (#9806) by @qqqys Review executions now run the repository's own commands inside a container boundary defined by operator policy instead of relying on local installation. (#9723) by @wenshao feat(core): add a per-project outside-repo artifact landing (#9776) by @wenshao feat(review): report findings to clients as a typed contract (#9794) by @wenshao Engages the critical severity floor immediately after two consecutive rounds of sustained finding rates instead of waiting for the fixed round-6 schedule. (#9938) by @wenshao Removed the Electron desktop package and related sync scripts since the code has been forked to the OpenWork repository. (#9085) by @yiliang114 Adds an opt-in liveness check for daemon child processes using ACP channels, with retries and timeouts to detect failures without false positives from host suspension. (#9976) by @doudouOUC Introduces a built-in DingTalk Workspace channel supporting direct messages, mentions, document notifications, and todo tracking using authenticated DWS CLI profiles. (<a href="https://github.com/QwenLM/qwen-code/pull/9394" data-hovercard-type="pull_request" dat
