QwenLM/qwen-code v0.22.2

Release v0.22.2

on GitHub

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. (#9394) by @qqqys
• feat(review): add a --topology minimal single-pass A/B arm (#9919) by @wenshao
• feat(cli): Manage agent view session lifecycle (#9986) by @ZijianZhang989
• feat(goal): make get_goal's default view a summary instead of the whole catalog (#9973) by @qqqys
• feat(goal): stop a Goal whose checkpoints stall three times in a row (#9975) by @qqqys
• feat(web-shell): make compact view the only mode (#9993) by @ytahdn
• feat(daemon): Support current-session scheduled tasks (#9838) by @doudouOUC
• feat(core): trust a generated-scripts root for workflow scriptPath loads (#9987) by @qqqys
• feat(providers): load model recommendations before editing (#9980) by @qqqys
• feat(goal): stop autonomous continuation at a token budget the user re-arms (#9891) by @qqqys
• feat(acp): enable managed auto-memory lifecycle (#9992) by @zgxkbtl
• feat(cli): Add standalone sessions for projectless tasks (#9978) by @doudouOUC
• feat(web-shell): add session token usage panel (#9988) by @ytahdn
• feat(core): accept cross-session messages behind an inbound gate (#9576) by @qqqys
• feat(goal): let a Goal stop early when its objective is infeasible (#10125) by @qqqys
• feat(goal): grant one hand-off turn before a spent budget stops the Goal (#10132) by @qqqys
• feat(core): add a bundled goal-draft skill that writes verifier-judgeable Goals (#10002) by @qqqys

Bug Fixes

• Clicking a session in the Session Overview panel now correctly opens it within its original workspace directory instead of defaulting to the primary workspace. (#9730) by @wenshao
• Web Shell now automatically retries messages via the ordinary prompt path if a mid-turn insertion is rejected while the stream is idle, preventing unnecessary error toasts. (#9732) by @ytahdn
• The review pipeline no longer posts fallback comments when a run is cancelled because a newer superseding run started, while still commenting on job-level timeouts. (#9716) by @wenshao
• Pinned sessions in the Web Shell sidebar now maintain a stable order based on pinning time and provide instant visual feedback when pinned or unpinned. (#9560) by @yiliang114
• Updates the workflow size baseline for qwen-autofix.yml to 397656 bytes to match the actual shipped file size and unblock CI checks. (#9747) by @wenshao
• fix(ci): gate the fork signal on fields the review payload delivers (#9469) by @qqqys
• Rejects session titles that exactly match the prompt's example strings, triggering the existing fallback behavior for auto-title and /rename --auto. (#9709) by @yiliang114
• Replaces VM evaluation of workflow meta literals with a static parser to prevent hangs or unbounded allocation caused by malicious model-generated code. (#9325) by @qqqys
• Enables workspace artifact downloads using relative daemon base URLs to fix Invalid URL errors in Web Shell integrations configured with same-origin paths. (#9734) by @ytahdn
• Increases the repair attempt agent budget to 45 minutes and adjusts related step caps to prevent opaque check failures during deterministic rejections. (#9691) by @qqqys
• Extends the workflow stall watchdog window from 60s to 180s to prevent false timeouts during long transport retries or slow model responses. (#9397) by @qqqys
• Gates skill announcements on whether the model explicitly declared the Skill tool, preventing wasted turns for subagents with restricted tool lists. (#9718) by @wenshao
• Counts fix-induced re-reports as new work in review statistics, distinguishing them from simple re-posts of existing findings. (#9744) by @wenshao
• Enables previewing for Markdown, HTML, and raster images in Web Shell by correctly classifying them before generic document handling. (#9760) by @ytahdn
• fix(daemon): keep restored ask_user_question valid after load (#9763) by @doudouOUC
• Public GitHub extensions now work on systems with Git older than 2.37 by resolving refs to immutable commits and downloading via the secure archive path instead of native Git. (#9690) by @yiliang114
• The review workflow now repairs permissions and retries before failing to clean up leftover worktrees, ensuring jobs do not fail due to removal errors. (#9748) by @wenshao
• Session title echo comparisons now use symmetric normalization for both candidate and example titles to prevent false mismatches. (#9809) by @yiliang114
• Fixed a React dev-mode memory leak in Web Shell by capping performance.measure accumulation and clearing the timeline every 16,384 entries. (#9770) by @wenshao
• The DingTalk channel now correctly parses forwarded chat records and includes their titles and summaries in the agent's context. (#9339) by @qqqys
• Web Shell now renders bold text correctly when emphasis markers are adjacent to CJK punctuation by adding the remark-cjk-friendly plugin. (#9457) by @cactuser-Lu
• Updates the settings schema to accept prompt hooks with prompt and model fields, aligning VS Code validation with runtime behavior. (#8779) by @zjunothing
• Adds a read-only reasoning effort preview for the qwen3.8-max model in WebShell welcome state before session creation. (#9599) by @callmeYe
• Enables Vertex AI authentication using Application Default Credentials when a Google Cloud project is configured without an API key. (#9017) by @axiom-of-choice
• Updates the workflow size baseline for cd-cua-driver.yml to prevent CI failures caused by its increased shipped file size. (#9822) by @wenshao
• Clears the tool display list immediately before invoking the completion callback to prevent duplicate rendering in the TUI. (#9602) by @qwen-code-dev-bot
• Normalizes Windows drive-letter casing in MCP approval keys to ensure consistent server recognition between CLI and VS Code. (#9779) by @zhou2024NAU
• File reading now validates image content via magic bytes before trusting extensions, preventing text files with image extensions from being misclassified. (#9113) by @rbalachandar
• Session exports on Windows now correctly preserve drive-letter paths in file links, allowing them to open locally instead of failing in the browser. (#8953) by @yefuyou
• The CLI input box border now adapts to terminal resizing by truncating long titles, preventing layout overflow and row jitter. (#8991) by @tlysanhuo
• Background agent notifications now include counts of remaining active agents for the same owner to improve coordination status visibility. (#9018) by @tao943
• Reasoning effort tiers are now automatically capped to the maximum level supported by each specific model endpoint to prevent request errors. (#9501) by @harjothkhara
• Microphone permission checks now occur only when recording starts rather than during voice warmup, preventing unnecessary early permission prompts. (#8912) by @Nas01010101
• Fixed loop detection to halt turns when the model repeats long verbatim blocks in content or reasoning streams. (#9668) by @yiliang114
• Secured the review process by ensuring the verified git identity remains consistent across all residue probe commands. (#9742) by @wenshao
• Fixed an issue where nested sub-agent approval requests were not surfaced to the UI, causing background agents to hang. (#9793) by @wenshao
• Ensured agent launch failures now correctly populate the error field to be reported as failed tool calls. (#9519) by @yiliang114
• Prevented ghost shutdown pending states by gating task assignment during teammate shutdown request processing. (#9550) by @AaronZ345
• Cached fork snapshots are now isolated by session ID to prevent cross-session data leakage during background reads. (#9471) by @tomsen02
• Evidence catalog previews are now capped at 240 UTF-8 bytes instead of characters to prevent Goal failures in non-ASCII languages. (#9835) by @qqqys
• Fixed text duplication in WSL terminals by skipping the redraw optimizer when running under ConPTY. (#7897) by @C0d3N1nja97342
• Fixed token count tracking to invalidate data when the active model route changes, ensuring accurate session limits and compression banners. (#9506) by @yiliang114
• Bounded conditional-close refusal holds to a 1,024-hold per-Session limit to prevent oversized responses from breaking cache retention. (#9820) by @doudouOUC
• Enabled workflow reapproval after hook bounces by making approval handling incarnation-aware and preventing stale events from affecting newer instances. (#9547) by @AaronZ345
• Fixed debug log routing in daemon mode to use async-local session IDs, preventing logs from different sessions being written to the same file. (#9538) by @tomsen02
• Added support for per-provider streamIdleTimeoutMs settings to handle varying latency requirements across different model providers. (#9795) by @yu-xin-c
• Fixes memory management by allowing forget operations to scan the same uncapped document universe as recall, preventing stale entries from persisting. (#9530) by @harjothkhara
• fix(core): make team shutdown a leader-only tool (#9401) by @yiliang114
• CI now moves undeletable workspace residue aside via rename instead of failing when standard deletion methods encounter permission issues. (#9868) by @wenshao
• Live task bridge operations now consistently use canonical lowercase UUIDs to prevent session lookup failures caused by case mismatches. (#9819) by @doudouOUC
• fix(core): map OpenAI-compatible finish_reason case-insensitively (#9884) by @yiliang114
• After retry exhaustion, the system now accepts quiet model completions following a tool result instead of aborting the run with an API error. (#9196) by @yiliang114
• Added stream-json to the allowed values for the output.format setting in the schema, ensuring VS Code settings match the CLI runtime capabilities. (#8966) by @harjothkhara
• The system now validates completed Anthropic tool arguments as strict JSON objects, rejecting malformed inputs to prevent execution of invalid calls. (#9013) by @tlysanhuo
• Fixed CI to wipe only A/B checkout directories instead of the entire workspace, preventing 900 MB git history re-downloads and runner hangs. (#9228) by @qwen-code-dev-bot
• Disabled the default permission-response timeout so requests remain pending until user decision, while still supporting explicit timeouts via --permission-response-timeout-ms. (#9933) by @doudouOUC
• Removed unsupported uniqueItems constraints from function schemas for OpenAI-compatible requests to prevent DashScope endpoint rejections. (#9869) by @qqqys
• Unified Goal continuation prompts across all interfaces with a guarded JSON block to prevent prompt injection and ensure objective consistency. (#9834) by @qqqys
• Banned unexecuted mutation claims in coverage briefs, requiring unrun mutations to be labeled as hypotheses with witness notes. (#9923) by @wenshao
• Downgraded workflow-size ratchet failures to warnings when the PR did not modify the grown file, preventing false positives from baseline changes. (#9931) by @wenshao
• Reduces rendering jank in Web Shell by keeping streamed thinking text out of the top-level render path and unmounting collapsed tool details. (#9914) by @ytahdn
• Neutralizes both modern ::name:: and legacy ##[name] workflow commands in autofix stdout echoes to prevent injection attacks in job logs. (#9871) by @wenshao
• Compiles workflow scripts before registering runs to ensure invalid scripts fail cleanly without creating phantom rows or telemetry events. (#9873) by @qqqys
• Repairs persisted session lifecycle operations to handle empty, malformed, or orphaned transcript files while returning conflicts as per-session errors. (#9626) by @doudouOUC
• The CLI now gracefully handles unreadable prompt files by logging a warning and using the default rewrite prompt instead of crashing. (#9753) by @yiliang114
• Goal ownership now hydrates independently from session metadata, allowing Slash commands to load faster without waiting for slower metadata requests. (#9977) by @ytahdn
• The /goal resume command now correctly resumes evidence-limited Goals by resetting the evidence cursor while preserving objective and revision data. (#9840) by @qqqys
• Restricts Git credential configuration to scoped operations initiated by Qwen Code, ensuring credentials remain isolated from URLs, logs, and environment variables. (#9870) by @callmeYe
• Ensures branched sessions inherit the source session's display name with numeric suffixes instead of random UUIDs, improving session recognition and consistency. (#9764) by @water-in-stone
• Fixes telemetry aggregation errors during failed session swaps by using explicit transactions to prevent inflated usage records from replayed history. (#9844) by @yiliang114
• Corrects QQ Bot cron message handling by tracking active prompts in a dedicated set instead of relying on unreliable stream state indicators. (#9971) by @yiliang114
• fix(cli): mark estimated compression token counts in banners (#9568) by @yiliang114
• Ensures OpenRouter requests explicitly disable reasoning features when configured off, by emitting the native reasoning parameter in the API payload. (#9758) by @yiliang114
• fix(core): validate skill commands against the live provider to survive late attach (#9824) by @yiliang114
• fix(cli): keep in-flight tool calls pending during live session replay (#9705) by @yiliang114
• fix(core): reject unverifiable validated read inodes (#9857) by @AaronZ345
• fix(review): the sandbox default named an image that does not exist (#9972) by @wenshao
• fix(web-shell): localize vision bridge notices (#10003) by @ytahdn
• fix(web-shell): allow reasoning selection before session creation (#10008) by @callmeYe
• fix(core): subordinate the coordinate skill's teammate count to the configured cap (#9403) by @yiliang114
• fix: repair the Windows and macOS test lane failures (#9728) by @wenshao
• fix(ci): give the macOS and Windows lanes a trigger again (#9370) by @wenshao
• fix(core): make permissions.allow restrict the tool schemas sent to the model (#9829) by @yiliang114
• fix(webui): preserve hydrated Goal state during event startup (#10012) by @ytahdn
• fix(cli): graft the review anchor forward across fail-closed rounds (#9932) by @wenshao
• fix(live): restore Live Host after desktop removal (#9994) by @yiliang114
• fix(goal): cite this turn's delivered output instead of refusing over it (#9880) by @qqqys
• fix(serve): let channel workers reach TLS-enabled daemons (#9392) by @qqqys
• fix(cli): drop duplicate fake properties breaking the main build (#10022) by @yiliang114
• fix(core): register report_findings in the permission alias table (#10033) by @yiliang114
• fix(cli): preserve bridge timeouts for mid-turn media (#9995) by @ytahdn
• fix(mcp): recover restarted HTTP MCP servers in-session and in CLI (#9962) by @yiliang114
• fix(core): sync loaded-skill state with history eviction (#9500) by @ZijianZhang989
• fix(core): preserve images for multimodal DeepSeek (#9854) by @shenyankm
• fix(ci): yield the event loop between script tests to avoid vitest RPC timeouts (#10037) (#10050) by @qwen-code-dev-bot
• fix(vscode): sync companion token limits (#9850) by @AaronZ345
• fix(ui): suppress duplicate identical TodoList panels in a single turn (#9692) by @PratikWayase
• fix(ci): retry sandbox image builds and file an issue when a release build fails (#9916) by @yiliang114
• fix(web-shell): constrain responsive sidebar drawer (#10020) by @dreamWB
• fix(core): dedupe hierarchical memory files by canonical identity (#9600) by @yiliang114
• fix(cua-driver): support sandboxed Windows SDK clients (#10120) by @LaZzyMan
• fix(node-repl): make package verification idempotent (#10133) by @LaZzyMan
• fix(review): give cancelled runs an accurate fallback body instead of the failure comment (#10114) by @wenshao
• fix(cua-driver): synchronize release version marker (#10135) by @LaZzyMan
• fix(core): keep allowlist-uncovered tools registered as deferred instead of removing them (#10075) (#10082) by @yiliang114
• fix(daemon): Persist empty sessions before scheduled-task binding (#10144) by @doudouOUC
• fix(channels): Make same-chat delivery session-aware (#10145) by @doudouOUC

Performance

• Terminal virtual scrolling now renders at up to 60 FPS, eliminating choppiness and improving scroll smoothness in large PTY windows compared to the previous 30 FPS cap. (#9681) by @DragonnZhang

Documentation

• Clarifies documentation by explicitly naming axes when referring to "rank 3" to resolve ambiguity between trimming and keeping operations. (#9759) by @wenshao
• docs(agent): clarify parameter preconditions (#9580) by @tao943
• Corrected the VS Code extension README to state the actual minimum required version of 1.96.0 instead of 1.85.0. (#9852) by @AaronZ345
• docs(core): fix goal judge timeout unit (#9861) by @AaronZ345
• Updated the ReadFile tool documentation to reference the truncateToolOutputLines setting for the default paging limit instead of a hardcoded line count. (#9863) by @AaronZ345
• Fixed TypeScript SDK timeout examples to use the correct query({ prompt, options }) factory signature instead of the outdated positional argument format. (#9867) by @AaronZ345
• docs: consolidate assorted documentation corrections (#9855) by @AaronZ345

Internal Changes

• Migrated autofix prose to a design record to reduce workflow file size and implemented a ratchet mechanism to limit future growth visibility in reviews. (#9677) by @wenshao
• Adds tests confirming subagent registries exclude MCP instructions, ensuring buildMcpServerInstructionsReminder remains safe without production code changes. (#9720) by @wenshao
• Enforced architectural boundaries by moving domain-coupled modules out of the CLI utils leaf layer and adding an ESLint rule to prevent upward imports. (#9737) by @yiliang114
• Refactors ACP skill management by extracting source acquisition and mutation logic into cohesive internal modules. (#8865) by @callmeYe
• Image generation now supports MiniMax endpoints by using the correct request schema and handling both URL and base64 responses. (#8322) by @octo-patch
• Refactored Goal continuation prompts to use a single core renderer for consistent output across all interfaces. (#9581) by @qqqys
• The review skill manifest is split into a core body and verdict-gated reference files to optimize loading based on posting status. (#9804) by @wenshao
• chore(ci): refresh the core issue-owner pool (#9808) by @yiliang114
• Refactors core utils to be a strict leaf layer with no runtime dependencies on other core modules, ensuring cleaner architectural boundaries. (#9778) by @yiliang114
• Centralizes approval mode contracts across core, CLI, and SDKs to ensure consistent validation and prevent silent mismatches in future updates. (#9796) by @yiliang114
• Establishes explicit ownership for derived Config instances via a new deriveConfig factory boundary and guards for lifecycle mutations. (#8100) by @yiliang114
• ACP cron tests now terminate the entire CLI process tree during cleanup to prevent directory deletion errors on POSIX and Windows. (#9815) by @dream-creator
• Removed unused token-counting and summarized-thinking capabilities from the core content generator interface to simplify provider implementations. (#9676) by @DragonnZhang
• Updated the Vertex-mode lazy generator test to use embedContent instead of the removed countTokens method to fix the TypeScript build. (#9888) by @AaronZ345
• Skipped the unreadable-ledger test when running as root since permission checks are bypassed, ensuring test stability in root environments. (#9913) by @wenshao
• Updated internal comments to correctly explain that replay safety relies on discarding failed attempt turns rather than excluding thought parts from history. (#8861) by @ComplexSimply
• Refactored the Core client to remove a runtime import cycle by importing the approval-mode enum directly from its definition. (#9959) by @DragonnZhang
• Workspace Git and status routes now share a single trusted-runtime resolver to ensure consistent security validation across all endpoints. (#9958) by @DragonnZhang
• chore(github): add @qqqys as core package codeowner (#9982) by @yiliang114
• test(cli): teach session-swap telemetry fakes the /branch title prerequisites (#9998) by @yiliang114
• refactor(acp-bridge): narrow workspace event capabilities (#9957) by @DragonnZhang
• refactor(web-shell): extract shared contexts (#9954) by @DragonnZhang
• refactor(web-shell): reuse canonical todo parser (#9956) by @DragonnZhang
• refactor(cli): remove unused EnumSelector component (#9997) by @qqqys
• ci: take the macOS and Windows lanes off pull requests (#10059) by @wenshao
• refactor(core): remove unused LruCache utility (#9926) by @AaronZ345
• refactor(cli): remove unused useInputHistoryStore hook (#10041) by @qqqys
• chore(ci): migrate CUA npm packages to trusted publishing (#9836) by @yiliang114
• chore(cua-driver): bump SDK to 0.20.1 (#10127) by @LaZzyMan

New Contributors

• @cactuser-Lu made their first contribution in #9457
• @axiom-of-choice made their first contribution in #9017
• @zhou2024NAU made their first contribution in #9779
• @octo-patch made their first contribution in #8322
• @yefuyou made their first contribution in #8953
• @tao943 made their first contribution in #9018
• @AaronZ345 made their first contribution in #9550
• @yu-xin-c made their first contribution in #9795
• @dream-creator made their first contribution in #9815
• @zgxkbtl made their first contribution in #9992

Full Changelog: v0.22.0...v0.22.2