Release v0.21.3

Highlights

• Enhanced /review command with test plan validation, measured failure attribution, and new verification lenses to improve code change analysis. (#8215, #8218, #8225, #8261, #8315)
• Added support for project-level named fork profiles to define tool allowlists and prompt hints via markdown configuration files. (#8148)
• Enabled workflow agents to surface approval requests for shell commands and edits directly to the parent interface with background run support. (#8240, #8303)
• Fixed Anthropic API integration issues by sanitizing tool IDs, preventing duplicate result blocks, and ensuring correct message ordering. (#8164, #8165, #8163, #8040)
• Improved Web Shell usability with artifact downloads, fixed text pasting visibility, and resolved session recap context switching errors. (#8234, #8230, #8262)
• Added native terminal image rendering via a new display_image tool and restored right-click menus with URL hyperlink handling. (#8217, #8198)

Breaking Changes

No known breaking changes.

Complete Change List

Features

• Lifecycle hooks now include optional source_type and source_id fields to help platforms distinguish sessions created through different entry points. (#8155) by @xurik
• The review workflow now verifies cache identity consistency to prevent mismatches between cache writing and reading agents in GitHub Actions. (#8205) by @wenshao
• Daemon channel adapter state is now isolated by workspace to safely support persistent browser authentication across multiple workspaces. (#8178) by @qqqys
• Added the verify-capture.mjs script to generate evidence images for PR verification with a single command instead of a complex pipeline. (#8114) by @wenshao
• Added generationConfig.cacheRetention settings to support Anthropic's extended 1-hour prompt-cache tier alongside the default 5-minute retention. (#8048) by @netbrah
• The CI startup bundle check now fails if the CLI entry point is incorrectly hoisted into a shared chunk, preventing silent execution failures. (#8203) by @wenshao
• Adds a transient 👀 emoji to GitLab notes while the agent is working and removes it upon completion. (#8119) by @OrbitZore
• Normalizes tool-call telemetry data to ensure consistent status reporting and error classification across all consumers. (#8176) by @doudouOUC
• Added a SessionDelete hook event triggered after successfully deleting a historical session via /delete or ACP. (#8059) by @xurik
• Enabled POST /session to accept and validate a caller-supplied sessionId instead of silently ignoring it. (#7836) by @qwen-code-dev-bot
• Allow safe slash commands like /status, /settings, and /help to run immediately while model responses are streaming. (#8130) by @DragonnZhang
• Adds an optional host tool invocation guard that can deny execution before a tool runs, preventing calls when configured to do so. (#8032) by @chiga0
• Adds OpenTelemetry GenAI streaming attributes to LLM spans, recording time-to-first-chunk for streaming requests while removing the private stream attribute. (#8150) by @doudouOUC
• Running the /triage command on a pull request now triggers the sandboxed verify lane in parallel to start building and running tests immediately. (#8249) by @wenshao
• The daemon default maximum session limit per workspace has increased from 20 to 32 to better support heavy multi-session workflows. (#8235) by @wenshao
• The Web Shell transcript and review views now include a Download action for artifacts, allowing users to save HTML, Markdown, and workspace files locally. (#8234) by @ytahdn
• Forked Dream workers now respect an opt-in gate that prevents writing to or editing files in top-level pinned directories during scheduled cleanup and workspace operations. (#7714) by @destire-mio
• Added support for project-level named fork profiles that define tool allowlists and prompt hints via markdown files in the .qwen/fork-profiles directory. (#8148) by @destire-mio
• Adds the skills.disabledLevels setting to hide bundled, project, user, or extension skills without affecting host-provided skills. (#8057) by @zhangxy-zju
• Extends the /autofix skill to review and repair local working trees with explicit user confirmation while retaining existing GitHub Actions support. (#8121) by @yiliang114
• Enabled workflow agents to surface approval requests for shell commands and edits directly to the parent interface. (#8240) by @qqqys
• Enhanced the verify skill to measure suggested fixes and calibrate replay tests against live code mutations. (#8242) by @wenshao
• Auto-generated Skills are now tracked for usage, marked stale after 30 days of inactivity, and archived after 90 days via a new project-scoped curator. (#7846) by @DragonnZhang
• The /review command now validates Test Plan assertions in PR descriptions against the actual code changes and workspace state. (#8215) by @wenshao
• Updated verification rules to explicitly check acceptance criteria and distinguish between empty results and unmeasured states. (#8295) by @wenshao
• Added the foundational runtime supervisor for Agent View, enabling local session management and terminal stream bridging. (#7799) by @ZijianZhang989
• Review output footers now include the Qwen Code CLI version to help trace findings to specific releases. (#8294) by @wenshao
• Plain-text messages sent during an active turn now enter the running turn by default with visible queue management. (#8229) by @ytahdn
• Added a Session Workflow view that visualizes plan execution, agent calls, and transcripts as a dependency graph. (#7580) by @yiliang114
• Added the --compaction flag to the /model command to configure a dedicated model for chat compression. (#7818) by @Aleks-0
• Connected the interactive TUI to the Goal v3 runtime with new lifecycle commands, status cards, and improved input queuing. (#8005) by @qqqys
• Workflows now support an opt-in background run mode that returns a run ID and reports results via a dedicated channel. (#8303) by @qqqys
• Added measured test-failure attribution to /review to accurately distinguish new failures from pre-existing ones. (#8218) by @wenshao
• Added a display_image tool to the TUI that renders PNGs using native terminal features or chafa fallbacks with an 8 MiB limit. (#8217) by @DragonnZhang
• Enhanced the /review command with adjudication rendering, workflow step extraction, and three new verification lenses. (#8225) by @wenshao
• The /review command now borrows recall mechanics, a fix loop, and size-derived budgeting from Claude Code to improve finding surfacing and diff cost estimation. (#8315) by @wenshao
• Review round ledgers are now embedded in the posted review body to persist across environments and enable the default effort level to recover previous round data. (#8255) by @wenshao
• The review process now detects empty or collapsed diffs to stop early and recommends closing superseded branches while applying seven new verification lenses. (#8261) by @wenshao

Bug Fixes

• Fixed an issue preventing the Web Shell from creating scratch workspaces on loopback daemons when bearer authentication is not configured. (#8204) by @wenshao
• Updated setPermissionMode E2E tests to use the shared TEST_TIMEOUT constant, resolving intermittent timeouts on CI runners. (#8135) by @qwen-code-dev-bot
• GitHub channel final replies are now queued and retried on restart if GitHub definitely rejected the initial write due to rate limits. (#8087) by @yiliang114
• fix(core): Tolerate transcript timestamp drift (#7886) by @doudouOUC
• Added Ctrl+Tab and Ctrl+Shift+Tab as alternative shortcuts for switching @ completion tabs to avoid conflicts with terminal word-movement bindings. (#8074) by @qwen-code-dev-bot
• Added an uncaught exception handler to prevent silent crashes in VP mode by ensuring errors are logged and displayed on the terminal for debugging. (#8088) by @chiga0
• Fixed compactString to strictly respect character limits even when the truncation marker itself exceeds the specified budget. (#7872) by @chinesepowered
• Localized built-in agent type names in the Web Shell UI for Chinese users to prevent mixed-language display issues. (#8209) by @wenshao
• Fixed the release workflow to anchor notes correctly and cap the release body length to prevent GitHub API validation errors. (#8199) by @wenshao
• Fixes 400 errors for assistant-turn prefill and restores visible thinking output on Claude 4.6+ and 5.x models. (#8040) by @netbrah
• Ensures tool results appear before other content in user messages to prevent Anthropic API rejections. (#8165) by @netbrah
• Sanitizes tool IDs to meet Anthropic character requirements and prevents empty ID errors. (#8164) by @netbrah
• Fixes model selector dropdown truncation by adapting width to long model names and adding hover tooltips. (#8220) by @wenshao
• Stabilizes a flaky integration test for async SDK MCP tool handlers by asserting on deterministic tool results. (#8223) by @qwen-code-dev-bot
• Fixed a test isolation issue to ensure temporary directories for worktree symlink tests are unique per run and fully cleaned up. (#8228) by @wenshao
• Fixed memory usage display to correctly show units like MB or GB when rounded values reach the next boundary. (#7871) by @chinesepowered
• Prevented subagents from asking users interactive questions to avoid indefinite waiting during background tasks. (#8219) by @DragonnZhang
• Fixed command splitting to correctly treat the bare & operator as a command boundary for permission checks. (#7864) by @chinesepowered
• Added Option+V as a macOS shortcut for pasting clipboard images alongside the existing Ctrl+V binding. (#8120) by @qwen-code-dev-bot
• Fixes a bug where mouse wheel events could scroll the background viewport briefly while a dialog was opening. (#8089) by @kagura-agent
• Fixes a CI failure where conflicting color environment variables caused extra warning output that broke PNG capture tests. (#8236) by @ytahdn
• Adds workspace-qualified memory operations and a new CLI option to isolate managed project memory by selected workspace. (#8056) by @qqqys
• Fixes the Web Shell to ensure pasted plain text remains fully visible and editable in the composer regardless of length. (#8230) by @ytahdn
• Qwen Autofix now resolves review threads only when tied to a verified commit, preventing resolution if the PR head changes or results are ambiguous. (#8231) by @wenshao
• Autofix now allows up to ten rounds of suggestion-level changes before restricting further edits to critical findings, failed checks, or base conflicts. (#8247) by @wenshao
• The system now recovers XML-style tool calls embedded in plain text model responses, preventing agent loops from breaking during long sessions. (#8037) by @qwen-code-dev-bot
• Upgraded external-context dependencies to MCP SDK 1.30.0 and patched Hono versions to resolve security advisories while maintaining Node 18 compatibility for mobile packages. (#8206) by @doudouOUC
• Fixed an issue where duplicate tool result blocks sharing the same ID caused Anthropic API requests to fail with a 400 error by ensuring only the first result is retained. (#8163) by @netbrah
• Background task polling in web-shell now silently handles transient network failures to prevent repetitive error notices while still reporting persistent hard failures. (#7923) by @han-dreamer
• Channel ACP bridges now automatically recover after long host sleep periods or event-loop stalls without requiring a full messaging adapter reconnection. (#8211) by @yiliang114
• Fixed an issue where raw SGR mouse escape sequences appeared in the input box at startup by correctly filtering them during early input capture. (#8268) by @qwen-code-dev-bot
• Stabilizes the thinking block height to prevent layout shifts and replaces the full-screen transcript overlay with an inline Ctrl+O toggle. (#8077) by @chiga0
• Completes image @ reference routing across TUI, ACP, and non-interactive CLI entry points with consistent security validation. (#7206) by @yiliang114
• Fixes flaky ACP cron integration tests by using QWEN_CODE_TEST_CRON_FAST to trigger jobs after a 5-second delay instead of waiting for real time. (#8243) by @qwen-code-dev-bot
• Prevents stale thinking blocks from appearing by removing them when their associated tool calls are orphaned or contain empty text. (#8166) by @netbrah
• Compacts the advanced table toolbar in narrow message areas by hiding labels and statistics while keeping essential actions visible. (#8264) by @ytahdn
• Fixes the mobile composer display after resuming by removing broken WebGL overlays and limiting animated placeholders to new sessions. (#8263) by @ytahdn
• Fixed test suite failures on Windows by handling path separators, line endings, and temporary directories correctly. (#8050) by @yiliang114
• Rendered verification reports as formatted Markdown instead of escaped raw text to improve readability in comments. (#8147) by @wenshao
• Prevented automatic session recaps from appearing in the wrong Web Shell session when switching contexts rapidly. (#8262) by @wenshao
• Corrected re-run summary messages to accurately describe existing bot reviews that deferred without voting. (#8273) by @wenshao
• Fixed self-hosted runner failures caused by file permission errors left behind by previous containerized jobs. (#8115) by @qwen-code-dev-bot
• Fixed an issue in the Web Shell where duplicate permission buttons with identical labels appeared in the approval dialog. (#8250) by @qwen-code-dev-bot
• Fixed the notices generator to correctly include license text for packages installed at multiple distinct versions within the dependency tree. (#8272) by @yiliang114
• QQ Bot group messages now display the full sender openid in prompts when mention support is enabled to improve model context. (#8233) by @Eric-GoodBoy-Tech
• Fixed AutoFix timeouts by explicitly setting the primary agent budget to align with the enclosing step's time limits. (#8257) by @wenshao
• Restored right-click menus and URL clicks by adding the ui.mouseTracking setting and implementing application-level hyperlink handling. (#8198) by @qwen-code-dev-bot
• Replaced hard truncation for long tool outputs with a collapsible view that preserves full content and improves scrolling. (#8251) by @destire-mio

Performance

• File search crawling is now faster by caching directory ignore lookups to avoid redundant checks for sibling files. (#8253) by @dexhunter

Documentation

• Adds a runnable Python example demonstrating how to implement an external judgment service for PreToolUse HTTP hooks. (#8202) by @babyblueviper1
• Reorganized Web Shell design documents into a dedicated directory to improve navigation and naming consistency. (#8304) by @water-in-stone
• Documentation now covers the /learn command, Skill discovery sources, and clarifies that personal and project Skill edits hot-reload without restarting unless in bare mode. (#8298) by @DragonnZhang

Internal Changes

• Simplified plugin manager button labels in the Web Shell to Add, Upload, and Create for a more concise interface in English and Chinese. (#8174) by @ytahdn
• Refactors internal workflow execution ownership to improve session management and cancellation handling. (#8140) by @qqqys
• Migrates the acp-cron integration test to use a fake OpenAI server for deterministic responses and isolates agent home directories. (#8082) by @qwen-code-dev-bot
• Subagent execution E2E tests now inherit the suite's default timeout instead of a restrictive 60-second limit, preventing failures during multi-turn model workflows. (#8246) by @qwen-code-dev-bot
• Two flaky SDK E2E cases relying on specific live model tool choices have been skipped to stabilize the main branch test suite. (#8259) by @qwen-code-dev-bot
• No user-facing changes in this release. (#7967) by @doudouOUC
• No user-facing changes; this release refactors internal CLI dependencies to improve code organization. (#8141) by @yiliang114
• Fixed a flaky E2E test for abort signals in tool permission callbacks to ensure deterministic execution. (#8300) by @qwen-code-dev-bot
• Fixed a test harness argument order to correctly pass the locale parameter, resolving failing Chinese localization checks. (#8314) by @yiliang114
• Fixed a silent startup failure for the ACP process when the extension is loaded from a symlinked directory. (#8309) by @Wchoi189
• SDK orchestration end-to-end tests are now deterministic by using scripted local responses, reducing test cases from 22 to 8 while maintaining core coverage. (#8312) by @yiliang114

New Contributors

• @babyblueviper1 made their first contribution in #8202
• @Wchoi189 made their first contribution in #8309

Full Changelog: v0.21.2...v0.21.3