Qwen Code Files

Highlights

• Web Shell composer now supports dragging, dropping, and pasting text files as named attachments alongside images. (#9180)
• Users can now fork conversations from any specific Assistant response using durable checkpoints to ensure branch accuracy. (#8817)
• Split-view panes now inherit the host's @ mention configuration, ensuring custom categories and exclusions apply consistently. (#9052)
• The /review skill now uses dedicated platform subcommands like meta and fetch-diff instead of executing raw gh commands. (#9096)
• Stopped takeover PRs now receive an autofix/needs-human label and appear in a new Takeover pool table on the dashboard. (#8960)
• Review sessions now record session IDs and diff hashes to enable crediting agent work across interrupted runs. (#9091)

Breaking Changes

No known breaking changes.

Review Workflow & /review Skill

Enhanced the /review skill with dedicated subcommands, dynamic finding limits, and improved concurrency handling to prevent loops and ensure accurate reporting.

• Split-view panes now inherit the host's @ mention configuration, ensuring custom categories appear and builtinAtProviders exclusions apply consistently. (#9052)
• The /review skill now uses dedicated platform subcommands like meta and fetch-diff instead of executing raw gh commands via prompt prose. (#9096)
• The /review command now limits posted suggestions to Critical findings after round 5 to prevent review loops, while deferring other findings to the final report. (#9118)
• Review workflow concurrency now isolates no-op human review requests to prevent them from blocking active PR review jobs. (#9210)
• Duplicate suggestions in /review are now listed in a dedicated paragraph with links to existing comments instead of being counted as anchor failures. (#9215)
• The /review presubmit gate now correctly handles carried-id re-posts to prevent dropping valid findings that match existing comment IDs. (#9212)
• Fixed the /review pipeline to accept bracketed source tags and added a --to-anchors option to normalize inputs for final gates. (#9222)
• The PR review workflow now posts a fallback comment with retry guidance if the main review job dies abnormally. (#9255)
• Review sessions now record session IDs and diff hashes to enable crediting agent work across interrupted runs. (#9091)
• Review runners now automatically wipe and retry failed checkouts to self-heal from persistent workspace corruption. (#9220)
• PR review worktree leases now act as locks to prevent concurrent sessions from destroying active review states during setup or cleanup. (#9211)

Autofix & Takeover Management

Improved Autofix visibility with new dashboard pools and labels, while refining footprint gates and growth tracking for better takeover handling.

• autofix: deny-by-default footprint gate and positional window censuses (#9156)
• Stopped takeover PRs now receive an autofix/needs-human label and appear in a new Takeover pool table on the dashboard for better visibility. (#8960)
• The autofix-growth-now marker now uses the prepare-time measurement instant to ensure growth divergence reports accurately reflect the correct base version. (#9192)
• Prevented the triage agent from processing tracking issues created by the autofix bot by adding a user login guard to the workflow trigger. (#9271)

Reverse-Audit & Findings Tracking

Optimized reverse-audit rounds based on diff size and ensured verified findings outside PR footprints are tracked in dedicated issues.

• The reverse-audit round cap now scales dynamically based on diff topology, allowing more rounds for small diffs and fewer for huge ones. (#9183)
• The reverse-audit round cap for huge diffs is now reduced to three only when the run has a clock; otherwise, it remains at five. (#9203)
• Verified findings outside the PR footprint are now deferred to a follow-up queue and tracked in a dedicated issue instead of being lost. (#9189)
• Reverse-audit now reports retirement failures with specific diagnostics and accepts additional punctuation separators in dry-receipt parsing. (#9213)
• Presubmit overlap lists are now written to a separate file to prevent overwriting the canonical findings artifact. (#9268)

Web Shell & User Interface

Added support for dragging and pasting text files in Web Shell and improved release note presentation with bilingual digests.

• The Web Shell composer now supports dragging, dropping, and pasting text files as named attachments alongside images. (#9180)
• Release notes are now presented as a user-friendly digest grouped by capability themes with bilingual English and Chinese summaries and attached screenshots. (#9216)
• Skill bodies are now redacted from Web Shell event surfaces to reduce payload size while remaining available for native ACP clients. (#9235)

Conversation & Session Control

Enabled forking conversations from specific responses and improved session resilience to preserve active work during shutdowns.

• Users can now fork conversations from any specific Assistant response using durable checkpoints to ensure branch accuracy. (#8817)
• Established a hidden runtime boundary for Conversations to isolate ownership and lifecycle while preserving existing owner-routed compatibility paths. (#9181)
• Sessions now preserve active work when close is refused by draining queued tasks within an 8-second budget before tearing down the session. (#9134)

CLI Tools & Extensions

Introduced External Context Provider profiles and added flags to qwen review commands for incremental validation and scoping.

• Introduced the External Context Provider Extension Profile v1 to enable provider-owned retrieval integrations via Qwen Extension and MCP boundaries. (#9068)
• Added --since flag to qwen review fetch-pr for validating incremental review anchors and scoping diffs based on local cache or ledger markers. (#9100)
• The runAllChunks command now outputs a diagnostic message on stderr when --all-chunks fans out a plan identified as Step 3A. (#9249)
• Fixed a bug where the findings command could silently overwrite input files when using --to-anchors and added validation to prevent flag wiring mismatches. (#9270)

Infrastructure & Reliability

Strengthened CI pipelines, resource limits, and automation workflows to prevent build failures and ensure consistent agent settings.

• ci: skip non-bot review_requested siblings before jobs spend compute (#9204)
• The hourly spam-minimization sweep now includes inline pull request review comments to block unwanted content effectively. (#9229)
• Increased the browser daemon SDK bundle size budget to 191 KiB to accommodate new attachment metadata and fix build failures. (#9238)
• ACP HTTP pre-attach buffers are now strictly bounded by byte count and frame limits to prevent resource exhaustion. (#9007)
• Release jobs now force-push release branches to prevent non-fast-forward errors during retries after previous failed publication attempts. (#9082)
• Fixed automation workflows to correctly pass agent settings like turn caps and tool allowlists that were previously silently dropped. (#9252)

Other Fixes & Improvements

Addressed various edge cases in tracing, mocking, and goal summarization to ensure system stability.

• goal: summarise the last Goal when a turn holds no permit (#9164)
• Fixed main-agent tracing edge cases regarding budget-triggered aborts, non-streaming calls, and deferred TUI tool batch ownership. (#9121)
• Fixed mock ACP child fixtures in integration tests to correctly handle the tool-guard handshake required by qwen serve. (#9161)

Other Changes

• Added an internal limitKind field to GoalRecord to improve how stopped Goals are typed, with no change to user-facing resume behavior. (#9165)
• Added test-only pins to verify ordering invariants and write-target paths without changing production behavior. (#9225)
• Tests now verify that background artifact refresh failures silently preserve the last successful state without showing errors or clearing the panel. (#9227)
• Fixed skill-parity tests on macOS by resolving the fixture root with realpath to ensure path consistency when tmpdir is a symlink. (#9269)

---

中文摘要

亮点

• Web Shell 编辑器现在支持拖放和粘贴文本文件作为命名附件，与图片并列显示。 (#9180)
• 用户现在可以使用持久化检查点从任意 Assistant 回复分叉对话，以确保分支准确性。 (#8817)
• 分屏视图现在继承主机的 @ 提及配置，确保自定义类别显示且 builtinAtProviders 排除项一致生效。 (#9052)
• /review 技能现在使用 meta 和 fetch-diff 等专用平台子命令，不再通过提示文本执行原始 gh 命令。 (#9096)
• 停止的接管 PR 现在获得 autofix/needs-human 标签，并在仪表板的 Takeover pool 表中显示。 (#8960)
• 审查会话现在记录会话 ID 和差异哈希，以便在中断的运行中确认代理工作。 (#9091)

审查工作流与 /review 技能

增强了 /review 技能，引入专用子命令、动态发现限制及改进的并发处理，以防止循环并确保报告准确。

• 分屏视图现在继承主机的 @ 提及配置，确保自定义类别显示且 builtinAtProviders 排除项一致生效。 (#9052)
• /review 技能现在使用 meta 和 fetch-diff 等专用平台子命令，不再通过提示文本执行原始 gh 命令。 (#9096)
• /review 命令在第 5 轮后仅发布 Critical 建议以避免循环，其他发现将延迟至最终报告。 (#9118)
• 审查工作流并发现在隔离无效的人类审查请求，防止其阻塞活跃的 PR 审查作业。 (#9210)
• /review 中的重复建议现在列在带有现有评论链接的专用段落中，不再计为锚点失败。 (#9215)
• /review 预提交网关现在正确处理携带 ID 的重新发布，防止丢弃匹配现有评论 ID 的有效发现。 (#9212)
• 修复了 /review 管道以接受带括号来源标签，并添加了 --to-anchors 选项以规范化最终网关的输入。 (#9222)
• 如果主审查作业异常终止，PR 审查工作流现在会发布包含重试指南的备用评论。 (#9255)
• 审查会话现在记录会话 ID 和差异哈希，以便在中断的运行中确认代理工作。 (#9091)
• 审查运行器现在会自动清除并重试失败的检出，以从持久性工作区损坏中自我修复。 (#9220)
• PR 审查工作树租约现在充当锁，防止并发会话破坏正在进行的审查状态。 (#9211)

Autofix 与接管管理

通过新仪表板池和标签提升 Autofix 可见性，同时优化足迹门控和增长追踪以更好地处理接管任务。

• autofix: deny-by-default footprint gate and positional window censuses (#9156)
• 停止的接管 PR 现在获得 autofix/needs-human 标签，并在仪表板的 Takeover pool 表中显示。 (#8960)
• autofix-growth-now 标记现在使用准备时的测量时刻，确保增长差异报告准确反映正确的基准版本。 (#9192)
• 通过在工作流触发器中添加用户登录守卫，防止审查代理处理 autofix bot 创建的跟踪问题。 (#9271)

逆向审计与发现追踪

根据差异大小优化逆向审计轮次，并确保 PR 范围外的已验证发现在专用 issue 中追踪。

• reverse-audit 轮次上限现在根据 diff 拓扑动态调整，小 diff 允许更多轮次，大 diff 则减少。 (#9183)
• 只有当运行有时钟时，巨大差异的反向审计轮次上限才会降至三，否则保持为五。 (#9203)
• PR 范围外已验证的发现现在推迟到后续队列，并在专用 issue 中跟踪，避免丢失。 (#9189)
• 逆向审计现在报告具体的退休失败诊断信息，并在干接收解析中接受更多标点分隔符。 (#9213)
• 预提交重叠列表现在写入单独的文件，以防止覆盖规范 findings 工件。 (#9268)

Web Shell 与用户界面

在 Web Shell 中支持拖放和粘贴文本文件，并通过双语摘要改进发布说明展示。

• Web Shell 编辑器现在支持拖放和粘贴文本文件作为命名附件，与图片并列显示。 (#9180)
• 发布说明现在按功能主题分组，提供双语摘要和截图，更易于用户阅读。 (#9216)
• 已从 Web Shell 事件表面中剔除技能正文以减少负载大小，同时保留对原生 ACP 客户端的可用性。 (#9235)

对话与会话控制

支持从特定回复分叉对话，并增强会话弹性以在关闭时保留活跃工作。

• 用户现在可以使用持久化检查点从任意 Assistant 回复分叉对话，以确保分支准确性。 (#8817)
• 为 Conversations 建立隐藏运行时边界以隔离所有权和生命周期，同时保留现有的所有者路由兼容路径。 (#9181)
• 当关闭被拒绝时，会话现在会在 8 秒内排空排队任务再销毁会话，从而保留正在进行的活跃工作。 (#9134)

CLI 工具与扩展

引入外部上下文提供者配置，并为 qwen review 命令添加标志以支持增量验证和范围限定。

• 引入 External Context Provider Extension Profile v1，支持通过 Qwen Extension 和 MCP 边界提供独立的检索集成。 (#9068)
• 为 qwen review fetch-pr 添加 --since 标志，用于验证增量审查锚点并基于本地缓存或账本标记限定差异范围。 (#9100)
• 当 --all-chunks 分发被识别为 Step 3A 的计划时，runAllChunks 命令现在会在 stderr 输出诊断信息。 (#9249)
• 修复了 findings 命令在使用 --to-anchors 时可能静默覆盖输入文件的问题，并增加了标志验证。 (#9270)

基础设施与可靠性

强化 CI 流水线、资源限制及自动化工作流，以防止构建失败并确保代理设置一致。

• ci: skip non-bot review_requested siblings before jobs spend compute (#9204)
• 每小时垃圾邮件最小化扫描现在包含内联 PR 审查评论，以有效阻止不需要的内容。 (#9229)
• 将 browser daemon SDK 包大小预算提升至 191 KiB 以容纳新的附件元数据并修复构建失败。 (#9238)
• ACP HTTP 预附加缓冲区现在严格按字节数和帧数限制，以防止资源耗尽。 (#9007)
• 发布任务现在强制推送 release 分支，防止因之前失败尝试导致的非快进错误在重试时阻塞发布。 (#9082)
• 修复了自动化工作流，正确传递之前被静默丢弃的代理设置（如轮次上限和工具允许列表）。 (#9252)

其他修复与改进

解决了追踪、模拟和目标摘要中的各种边缘情况，以确保系统稳定性。

• goal: summarise the last Goal when a turn holds no permit (#9164)
• 修复了 main-agent 追踪中关于预算触发中止、非流式调用和延迟 TUI 工具批处理所有权的边缘情况。 (#9121)
• 修复了集成测试中的 mock ACP child 以正确处理 qwen serve 所需的 tool-guard 握手。 (#9161)

其他变更

• 在 GoalRecord 中添加了内部 limitKind 字段以改进已停止 Goal 的类型定义，不影响用户可见的恢复行为。 (#9165)
• 添加了仅用于测试的固定项以验证顺序不变量和写入目标路径，不改变生产行为。 (#9225)
• 测试验证后台构件刷新失败时静默保留上次成功状态，不显示错误也不清空面板。 (#9227)
• 通过在 macOS 上对 fixture 根目录使用 realpath 修复技能一致性测试，确保 tmpdir 为符号链接时的路径一致性。 (#9269)

Complete Change List (43 pull requests)

### Features - autofix: deny-by-default footprint gate and positional window censuses ([#9156](https://github.com/QwenLM/qwen-code/pull/9156)) by @wenshao - review: absorb prose gh commands into platform-backed subcommands ([#9096](https://github.com/QwenLM/qwen-code/pull/9096)) by @wenshao - support fork from any conversation ([#8817](https://github.com/QwenLM/qwen-code/pull/8817)) by @water-in-stone - web-shell: support text file attachments in the composer ([#9180](https://github.com/QwenLM/qwen-code/pull/9180)) by @doudouOUC - review: adopt a round-aware convergence posture for posted findings ([#9118](https://github.com/QwenLM/qwen-code/pull/9118)) by @wenshao - autofix: escalate stopped takeover PRs and age out unanswered pauses ([#8960](https://github.com/QwenLM/qwen-code/pull/8960)) by @wenshao - review: scale the reverse-audit round cap to the diff topology ([#9183](https://github.com/QwenLM/qwen-code/pull/9183)) by @wenshao - review: apply the huge round reduction only when the run has a clock ([#9203](https://github.com/QwenLM/qwen-code/pull/9203)) by @wenshao - autofix: defer verified out-of-footprint findings to a surviving follow-up queue ([#9189](https://github.com/QwenLM/qwen-code/pull/9189)) by @wenshao - review: run-session ledger and cross-session agent evidence ([#9091](https://github.com/QwenLM/qwen-code/pull/9091)) by @wenshao - external-context: Add provider extension profile ([#9068](https://github.com/QwenLM/qwen-code/pull/9068)) by @doudouOUC - review: validate and scope the incremental anchor inside fetch-pr ([#9100](https://github.com/QwenLM/qwen-code/pull/9100)) by @wenshao - daemon: Isolate the Conversations runtime boundary ([#9181](https://github.com/QwenLM/qwen-code/pull/9181)) by @doudouOUC - release: user-facing bilingual digest for release notes ([#9216](https://github.com/QwenLM/qwen-code/pull/9216)) by @wenshao ### Bug Fixes - web-shell: share at mention providers with split-view panes ([#9052](https://github.com/QwenLM/qwen-code/pull/9052)) by @samuelhsin - ci: skip non-bot review_requested siblings before jobs spend compute ([#9204](https://github.com/QwenLM/qwen-code/pull/9204)) by @yiliang114 - ci: minimize spam inline review comments ([#9229](https://github.com/QwenLM/qwen-code/pull/9229)) by @yiliang114 - ci: keep no-op review requests out of the PR review concurrency group ([#9210](https://github.com/QwenLM/qwen-code/pull/9210)) by @wenshao - integration-tests: ack daemon tool-guard handshake in the mock ACP child (#9159) ([#9161](https://github.com/QwenLM/qwen-code/pull/9161)) by @qwen-code-dev-bot - sdk: raise daemon browser bundle budget to 191KB ([#9238](https://github.com/QwenLM/qwen-code/pull/9238)) by @wenshao - review: note when --all-chunks fans out a plan whose numbers say 3A ([#9249](https://github.com/QwenLM/qwen-code/pull/9249)) by @yiliang114 - telemetry: Address main agent tracing edge cases ([#9121](https://github.com/QwenLM/qwen-code/pull/9121)) by @doudouOUC - review: give duplicate-dropped Suggestions their own compose state and body sentence ([#9215](https://github.com/QwenLM/qwen-code/pull/9215)) by @wenshao - autofix: re-anchor growth divergence on measurement time and external head moves ([#9192](https://github.com/QwenLM/qwen-code/pull/9192)) by @wenshao - ci: stop dropping agent settings in resolve and follow-up workflows ([#9252](https://github.com/QwenLM/qwen-code/pull/9252)) by @wenshao - goal: summarise the last Goal when a turn holds no permit ([#9164](https://github.com/QwenLM/qwen-code/pull/9164)) by @qqqys - serve: redact skill bodies from the Web Shell event surface ([#9235](https://github.com/QwenLM/qwen-code/pull/9235)) by @wenshao - review: exempt carried-id re-posts from the presubmit overlap drop ([#9212](https://github.com/QwenLM/qwen-code/pull/9212)) by @yiliang114 - review: normalize last-gate inputs and anchor mid-line fragments ([#9222](https://github.com/QwenLM/qwen-code/pull/9222)) by @wenshao - review: fix silent reverse-audit retirement failures and keep non-converged evidence ([#9213](https://github.com/QwenLM/qwen-code/pull/9213)) by @wenshao - ci: keep a fallback comment when the PR review runner dies ([#9255](https://github.com/QwenLM/qwen-code/pull/9255)) by @wenshao - serve: Bound ACP HTTP pre-attach buffers by bytes ([#9007](https://github.com/QwenLM/qwen-code/pull/9007)) by @doudouOUC - ci: self-heal failed checkouts on the reused review runners ([#9220](https://github.com/QwenLM/qwen-code/pull/9220)) by @wenshao - review: keep the presubmit overlap list out of the canonical findings artifact ([#9268](https://github.com/QwenLM/qwen-code/pull/9268)) by @wenshao - ci: force-push release branch so retries replace failed attempts (#9076) ([#9082](https://github.com/QwenLM/qwen-code/pull/9082)) by @qwen-code-dev-bot - ci: stop triaging the autofix bot's own deferred-finding tracking issues (#9264) ([#9271](https://github.com/QwenLM/qwen-code/pull/9271)) by @yiliang114 - daemon: Preserve sessions when active-work close is refused ([#9134](https://github.com/QwenLM/qwen-code/pull/9134)) by @doudouOUC - review: close out the four leftover findings from the [#9222] review ([#9270](https://github.com/QwenLM/qwen-code/pull/9270)) by @wenshao - review: lock the PR review worktree lease against concurrent sessions ([#9211](https://github.com/QwenLM/qwen-code/pull/9211)) by @wenshao ### Internal Changes - refactor(goal): type the limit that stopped a Goal ([#9165](https://github.com/QwenLM/qwen-code/pull/9165)) by @qqqys - test(review): close confirmed pin gaps from [#9194] (batch 1) ([#9225](https://github.com/QwenLM/qwen-code/pull/9225)) by @yiliang114 - test(web-shell): pin silent failure of background artifact refreshes (#7427) ([#9227](https://github.com/QwenLM/qwen-code/pull/9227)) by @yiliang114 - test(review): realpath the skill-parity fixture root ([#9269](https://github.com/QwenLM/qwen-code/pull/9269)) by @wenshao

Full Changelog: https://github.com/QwenLM/qwen-code/compare/v0.21.12...v0.21.13