@qwen-code/qwen-code 0.26.3

Release v0.26.3

on Node.js NPM

Highlights

See the complete change list below.

Breaking Changes

No known breaking changes.

Core runtime, memory, and context

Improves tool call recovery, memory extraction, and context-window detection for qwen models.

• memory: opt-in extraction cadence after a no-op run (#13571)
• core: recover outer XML calls with quoted call content (#13579)
• core: shrink tool results to the headroom left before auto-compaction (#13599)
• Context-window detection now treats invite-only qwen-latest-series- aliases as 1M-context models. (#15758)
• Context-window detection now treats qwen4 model variants as having a 1,000,000-token context window when no upstream catalog entry defines them. (#15762)

Shell, OpenTUI, and Generated UI

Smoother Web Shell composer, OpenTUI rendering, and more reliable Generated UI cards.

• web-shell: retain Generated UI instances across same-page navigation (#15611)
• web-shell: add configurable simple composer (#15663)
• OpenTUI terminals now draw the Mods band above the prompt, matching Ink behavior and supporting ui.render hooks with layout limits. (#15699)
• generated-ui: offer Generated UI only to clients that render MCP Apps (#15720)
• Live Generated UI cards in Web Shell now attach immediately and show their interactive view when the tool result arrives. (#15787)
• generated-ui: check a view's initial calculation before display (#15790)

Managed Agent hosting and admin

Updates for Java embedding, daemon channels, scan summaries, maintenance CLI, and runtime artifacts.

• Java hosts can now opt into embedding Managed Agent in the same JVM and Spring context, using a dedicated connection pool or supplied DataSource. (#15718)
• Documents that managed-agent definitions sharing a Session intentionally queue in that Session and adds slow-probe guidance. (#15724)
• Managed-agent scan summaries now show the exception class for message-less row failures instead of failure=null. (#15725)
• Daemon-managed channels now honor approvalMode even without webhooks, matching standalone startup. (#15754)
• The reference distribution adds bin/qwen-managed-agent-admin to run managed-agent offline maintenance commands by short name, with help and unknown-command errors. (#15772)
• qwen serve --managed now receives verified runtime artifacts from GitHub and OSS without waiting for Maven Central publication. (#15782)

Workspaces and Skills

Faster Workspace Skill inventory and administrator reuse of idle resident CSI Workspaces.

• Workspace Skill inventory now returns a retained snapshot immediately and refreshes it in the background, while cold reads, deletion, and enablement still wait for fresh data. (#14890)
• Adds administrator-driven reclamation and reuse of idle resident CSI Workspaces with cleanup verification and receipts. (#15729)

Sessions, ACP, and conversations

Adds persistent session Workflow mode, ACP identity propagation, and archived conversation controls.

• add persistent session Workflow mode (DW-04A) (#15422)
• acp: carry rewind identity through daemon clients (#15441)
• Archived conversations are now managed under Settings → Archived conversations, keeping existing archive groups, filters, and lifecycle actions. (#15557)
• ACP now projects subagent identity through live frames, persisted replay, SDK normalization, and transcript reduction for thoughts, messages, and tool events. (#15657)

CLI and PR automation

CLI sandbox workflow switches, safer autofix-pr starts, and clearer /verify-pr failure attribution.

• cli: keep autofix-pr starts off the user's own branch work (#15736)
• cli: forward workflow switches into the CLI sandbox (#15740) (#15743)
• /verify-pr now attributes failures to the PR, harness, or host, requires a failures ledger, and rejects findings or blocked verdicts with no PR-attributed failures. (#15764)

Other Changes

• Release publishing checks out with a run-scoped token while keeping the bot PAT only for write steps, reducing checkout auth failures. (#15728)
• Pull request CI now checks out the exact head SHA from the triggering event, keeping reruns pinned to the original commit. (#15734)
• Internal tests now treat git worktree prune dry-run output as an unordered set, avoiding failures caused by filesystem-dependent ordering. (#15767)
• Acceptance tests now wait for the focused shell tab byte to reach the child before asserting input bytes, reducing flaky OpenTUI failures. (#15763)
• Managed snapshot pool test setup now allows three seconds for the first database connection, reducing CI setup timeouts. (#15749)
• Installed npm and private Bun acceptance tests now run nightly and by manual dispatch instead of on every PR update, while source TUI PR behavior stays unchanged. (#15719)
• Managed Agent failover validation now includes G3 terminal recovery gates for reply loss and write failure. (#15756)
• CLI review cleanup tests now pin process liveness to avoid flaky orphan-reap assumptions. (#15751)
• Desktop release preparation now resets self-hosted runner Git metadata before checkout to avoid stale metadata failures. (#15761)
• Hosted gates now use one shared zombie-aware liveness check for managed-agent processes, making running and stopped detection consistent. (#15777)

中文摘要

核心运行时、记忆与上下文

改进工具调用恢复、记忆提取节奏，以及 qwen 模型的上下文窗口检测。

• memory: opt-in extraction cadence after a no-op run (#13571)
• core: recover outer XML calls with quoted call content (#13579)
• core: shrink tool results to the headroom left before auto-compaction (#13599)
• 上下文窗口检测将 qwen-latest-series- 邀请制别名识别为 1M 上下文。 (#15758)
• 上下文窗口检测现在将未收录的 qwen4 模型变体统一按 1,000,000 token 上下文处理。 (#15762)

Shell、OpenTUI 与 Generated UI

改进 Web Shell composer、OpenTUI 渲染与 Generated UI 卡片的可靠性。

• web-shell: retain Generated UI instances across same-page navigation (#15611)
• web-shell: add configurable simple composer (#15663)
• OpenTUI 终端现在会在提示符上方绘制 Mods 横幅，与 Ink 行为一致，并支持 ui.render 钩子。 (#15699)
• generated-ui: offer Generated UI only to clients that render MCP Apps (#15720)
• Web Shell 中的实时 Generated UI 卡片现在会立即附加并显示交互视图。 (#15787)
• generated-ui: check a view's initial calculation before display (#15790)

Managed Agent 托管与管理

更新 Java 嵌入、daemon 通道、扫描摘要、维护 CLI 与运行时工件获取。

• Java 主机可选择在同一 JVM 和 Spring 上下文中嵌入 Managed Agent，使用专用连接池或提供的 DataSource。 (#15718)
• 文档说明共享 Session 的 managed-agent definitions 会在该 Session 排队，并给出慢速探测建议。 (#15724)
• managed-agent 扫描摘要会显示无消息行失败的异常类，不再显示 failure=null。 (#15725)
• daemon 管理的 channel 即使没有 webhook 也会遵循 approvalMode。 (#15754)
• 参考发行版新增 bin/qwen-managed-agent-admin，可按短名称运行 managed-agent 离线维护命令，并提供 help 与未知命令错误。 (#15772)
• qwen serve --managed 现从 GitHub 和 OSS 获取已验证运行时，不再等待 Maven Central 发布。 (#15782)

Workspace 与 Skill

加快 Workspace Skill 清单刷新，并支持管理员回收复用空闲 resident CSI Workspace。

• Workspace Skill 清单现在会立即返回保留的快照并在后台刷新；冷读取、删除和启用仍等待实时观测。 (#14890)
• 支持管理员回收并复用空闲 resident CSI Workspace，包含清理校验和凭据。 (#15729)

会话、ACP 与对话管理

新增持久会话 Workflow 模式、ACP 身份传递，并支持归档对话管理。

• add persistent session Workflow mode (DW-04A) (#15422)
• acp: carry rewind identity through daemon clients (#15441)
• 归档对话现可在 Settings → Archived conversations 中管理，保留原有分组、筛选和操作。 (#15557)
• ACP 现在会在实时帧、持久化回放、SDK 规范化和会话精简中保留子代理身份，用于思考、消息和工具事件。 (#15657)

CLI 与 PR 自动化

CLI 沙箱会接收 workflow 开关，autofix-pr 启动更安全，/verify-pr 失败归因更清晰。

• cli: keep autofix-pr starts off the user's own branch work (#15736)
• cli: forward workflow switches into the CLI sandbox (#15740) (#15743)
• /verify-pr 现在将失败归因到 PR、harness 或 host，要求 failures ledger，并拒绝没有 PR 归因失败的 findings 或 blocked 结论。 (#15764)

其他变更

• 发布检出改用 run-scoped token，bot PAT 仅用于写入步骤，降低认证失败。 (#15728)
• PR CI 现在检出触发事件中的 head SHA，重跑仍固定于原始提交。 (#15734)
• 内部测试现在将 git worktree prune dry-run 输出作为无序集合校验，避免因文件系统顺序不同导致失败。 (#15767)
• 验收测试现在会先等待聚焦 shell 标签的字节送达子进程，再断言输入流，减少 OpenTUI 偶发失败。 (#15763)
• 快照池测试初始化现在给首个数据库连接 3 秒租约，减少 CI 超时失败。 (#15749)
• 已安装 npm 和私有 Bun 验收测试改为每晚或手动运行，不再随每个 PR 更新运行；source TUI 行为不变。 (#15719)
• Managed Agent 故障切换验证新增 G3 终端恢复测试门，覆盖回复丢失和写入失败。 (#15756)
• CLI review cleanup 测试现在固定进程存活检查，避免孤儿进程清理用例不稳定。 (#15751)
• 桌面发布准备现在会在检出前重置自托管 runner 的 Git 元数据，避免陈旧元数据导致失败。 (#15761)
• Hosted gates 现在对 managed-agent 进程使用统一的 zombie-aware 存活检查，使运行与停止判断更一致。 (#15777)

Full Changelog: v0.26.2...v0.26.3