Release v0.27.4

Highlights

• Managed Agent sessions can access granted local directories and open an interactive Managed Web Terminal without persisting credentials or paths. (#16844, #16866)
• Send outputs to Feishu and WeCom webhooks, with long managed-channel replies delivered as ordered segments instead of being truncated. (#17012, #17013, #17223, #17224, #17002)
• Auto Mode now requires explicit authorization for risky actions, reviews classifier timeouts, avoids HTTP(S) URL filesystem lookups, and visibly marks Full Access. (#16858, #17127, #17146, #17100)
• Web Shell Goals page can merge goals from trusted workspaces, show workspace badges, and target a selected workspace when creating goals. (#17230)
• Manage memory with restored CLI /memory controls, persistent workspace settings, and typed SDK APIs to preview, submit, inspect, cancel, and retry Codex memory imports. (#17162, #17213)
• Terminal tabs, web previews, local commands, and slash commands keep working over plain HTTP when native UUID support is unavailable. (#17196)

Breaking Changes

No known breaking changes.

Complete Change List

Features

• Managed Agent sessions can use the browser user's local filesystem through temporary directory grants, without persisting credentials or paths. (#16844) by @wenshao
• Long managed-channel replies are split into ordered delivery segments instead of being truncated, with a visible truncation notice if the segment limit is exceeded. (#17002) by @wenshao
• With experimental.unifiedAgentLoop enabled, each AgentCore round runs model and tool batch in one runAgentLoop call, optionally on the event-bus kernel. (#17059) by @qqqys
• feat(cli): run each OpenTUI round in one runAgentLoop call (#17104) by @qqqys
• feat(channels): add Feishu custom bot webhook delivery (#17012) by @jkxiao
• Auto Mode now lets the session model review a fast-classifier timeout once when identifiers differ, allowing or blocking the action under existing policy. (#17146) by @callmeYe
• Exposes a typed workspace-bound SDK API for previewing, submitting, inspecting, canceling, and retrying Codex memory imports. (#17162) by @callmeYe
• Runs each Ink TUI round in one runAgentLoop call when experimental.unifiedAgentLoop is enabled, aligning TUI loop execution. (#17160) by @qqqys
• Adds per-turn Hook execution records, read-only Hook pages, SDK access, summaries, and signed snapshots for session history. (#17015) by @callmeYe
• Makes cancelled runs publish loop.cancelled and finish within a 10-second bound, preserving completed batch results when possible. (#17145) by @qqqys
• Adds a send-only Feishu Webhook connection method with secret URL, optional signing secret, channel control, and scheduled group delivery. (#17013) by @jkxiao
• Adds nightly E2E loop rollout legs for kernel and OpenTUI, including forced parity arms, while push runs keep their current legs. (#17050) by @wenshao
• Enables the serve daemon's PR-state watch to observe and deliver PR updates promptly by provisioning its dependencies and adding resume delivery. (#17032) by @wenshao
• Improves web-shell session search with a dedicated centered dialog, clearer results, keyboard selection, loading and retry states, and responsive layouts. (#16779) by @deggs7
• Adds an interactive Managed Web Terminal to Managed Agent sessions, using a short-lived ticket for authenticated WebSocket access and keeping the runtime active while attached. (#16866) by @wenshao
• feat(managed-agent): fixed-interval /loop schedules for Managed Sessions (#17088) by @wenshao
• With experimental.unifiedAgentLoop enabled, each ACP round runs in one runAgentLoop call; experimental.eventBusAgentLoop uses the event-bus kernel. (#17181) by @qqqys
• When experimental.unifiedAgentLoop is enabled, speculation model rounds now use the shared agent loop; otherwise behavior is unchanged. (#17029) by @wenshao
• Experimental loop switches now resolve per surface and can be forced on or off by environment overrides such as 1, 0, or headless,acp=0. (#17055) by @wenshao
• Adds a send-only WeCom group webhook Channel that can deliver Notify outputs, prompt results, and supported scheduled-task results to a configured WeCom group. (#17223) by @jkxiao
• Adds a send-only WeCom Webhook connection method with group URL configuration, channel control, and scheduled delivery selection. (#17224) by @jkxiao
• Makes PR watch operator-owned by adding the operator-only serve.sessionPrWatch setting and retiring GitHub channel state-transition delivery. (#17214) by @wenshao
• Restored native /memory controls and User/Project file or managed-folder actions in the CLI, with workspace settings persisting across restarts. (#17213) by @chiga0
• Adds multi-workspace support to the Web Shell Goals page, merging goals from trusted workspaces, adding workspace badges, and letting new goals target a selected workspace. (#17230) by @wenshao
• Adds experimental.unifiedAgentLoop hosted surface so Hosted Harness model rounds can use the shared agent loop while the legacy path remains default. (#17264) by @wenshao

Bug Fixes

• AUTO permission checks now judge explicit user authorization, especially for destructive, discarding, shared-history, or bulk actions, rather than mere relevance. (#16858) by @yanyukang2003-ux
• When Full Access (yolo) is active, the web shell composer mode trigger now shows its icon and label in the destructive color, keeping the risky mode visibly marked. (#17100) by @callmeYe
• Synchronous managed-session rename settlement no longer waits on the stale tenant guard, allowing concurrent first renames to settle independently within or across tenants. (#17074) by @doudouOUC
• fix(web-shell): preserve observer activity during recovery (#17098) by @callmeYe
• fix(web-shell): show timeline on narrower screens (#16986) by @callmeYe
• fix(deps): 修复 MCP、DOMPurify 和压缩请求解析漏洞 (#17105) by @chiga0
• fix(managed-agent): stop the legacy CSI guard from gap-locking other Sessions (#17069) by @doudouOUC
• Updated qs, fast-uri, ip-address, Hono, and Node adapter security dependencies across workspace and Mobile MCP trees. (#17134) by @chiga0
• AUTO mode no longer tries filesystem resolution for HTTP(S) URLs, preventing Windows network lookup delays and session termination before requests start. (#17127) by @callmeYe
• Native Anthropic conversations now omit unsigned thinking blocks from completed history while keeping visible answers, signed thinking, and completed tool exchanges. (#17031) by @callmeYe
• Updated fast-uri and ip-address in the independent Live Host and CUA Electron installations to secure versions, closing remaining URI and IPv6 advisories. (#17147) by @chiga0
• Goal status messages now show a compact one-line objective preview with Expand details, keeping long objectives from filling the chat. (#17133) by @callmeYe
• An unconfirmed send from another conversation no longer locks the composer; it remains editable, but submission is blocked until retry or discard. (#17140) by @yiliang114
• Edit now recovers matches between non-ASCII text and Unicode escape spellings when exact matching fails, preserving the file's original representation. (#17096) by @callmeYe
• Raises the document export renderer size budget to 2,136,000 bytes so current builds pass export validation. (#17153) by @chiga0
• Upgrades glob and minimatch together to secure versions, improving matching reliability and preventing search regressions. (#17149) by @chiga0
• Hover timestamps no longer cover message controls in the web shell, keeping buttons such as Goal details readable. (#17168) by @callmeYe
• Removes obsolete ExcelJS archive dependency paths and updates unzipper to 0.12.5 for cleaner dependency resolution. (#17163) by @chiga0
• Upgrades minimatch and glob dependencies in independent Electron build trees to address security advisories. (#17165) by @chiga0
• fix(external-context): deflake malformed proxy env test (#17176) (#17179) by @qwen-code-dev-bot
• Prevents table toolbar overlap in narrow web-shell views by compacting controls, shortening summaries, and preserving row counts. (#17020) by @Lilian0122
• Clears non-sensitive OpenTUI extension-setting drafts when Ctrl+C is pressed, fixing text accumulation in ordinary setting fields. (#15951) by @chiga0
• Prepares file watchers before final interactive exit so /quit can complete cleanly instead of leaving the CLI process alive. (#17148) by @chiga0
• Fixes Unicode Edit recovery so it no longer damages syntax by matching inside backslash, hexadecimal, or Unicode escape sequences. (#17152) by @callmeYe
• fix(ci): route classify_pr by association and admit dispatch to the fleet (#17178) by @yiliang114
• Terminal tabs, web previews, local and slash commands, and related shell flows now keep working over plain HTTP by generating IDs when native UUID support is unavailable. (#17196) by @wenshao
• Managed Sessions /loop schedules now recover more efficiently and only expose the cron tools declared for the session profile. (#17197) by @wenshao
• Fixes the agent tab context usage meter to use the agent's own model window instead of the main session window. (#13614) by @Shizoqua
• Fixed the CLI feedback prompt to preserve typed input and made newly registered workflow commands appear immediately in the model-facing catalog. (#17233) by @qqqys
• Fixed managed-agent configuration reads to accept PUBLISHED as a settled state for the relevant Agent resources. (#17207) by @yiliang114
• Fixed the memory-manager integration suite so it can choose an output root without requiring a special environment variable. (#17265) by @wenshao
• Returns correlated error responses for Java SDK control callback failures, including request ID, subtype, and error value. (#17113) by @Afloat16
• Fixes release validation failures in interactive memory-manager tests by ensuring fake editor and opener scripts run correctly. (#17294) by @qwen-code-dev-bot
• Fixes managed agent child sessions to use inherited extension configuration at committed state and report failures before Runtime acquisition. (#17281) by @yiliang114

Performance

• Improves CLI performance in ACP by skipping interactive extension watching, reducing startup contention. (#17177) by @callmeYe
• Speeds up the first native ACP session by overlapping initialization work and deferring unneeded command and dependency loading. (#17209) by @callmeYe
• Improves Java transcript performance by sharing immutable event snapshots instead of duplicating parsed page data. (#17115) by @Afloat16
• Speeds up the Web Shell Daemon Status Usage dashboard by caching per-transcript usage summaries and rescanning only changed transcripts. (#17239) by @wenshao

Documentation

• Adds bilingual Managed Agent design documentation inventorying lifecycle claim mirrors and tenant guards without changing runtime behavior. (#17073) by @doudouOUC
• Documents streaming tool dispatch decisions in English and Chinese, with opt-in implementation deferred to a follow-up. (#17171) by @qqqys
• README Capabilities now presents a grouped overview of Qwen Code features, with links to user docs and labels for experimental and opt-in features. (#17226) by @wenshao
• Added design documentation evaluating the Hosted Harness agent-loop migration and its separation into journaled tool batch and model round paths. (#17248) by @wenshao
• Updated README availability claims to distinguish default features from experimental or opt-in features and expanded listed chat channels. (#17263) by @wenshao
• Updates README capability labels to avoid repeated pitch text and links /resume to the commands reference. (#17277) by @wenshao
• Updates PR watch design documents to record shipped status, settled review questions, and the removal of the enablement switch. (#17250) by @wenshao

Internal Changes

• Web shell test now opens Session Overview from the footer entry instead of the More menu, matching single-column layouts. (#17067) by @qwen-code-dev-bot
• Managed Agent schema initialization now uses ManagedSchema.initialize directly after removing the redundant JdbcRuntimeBrokerSchema alias. (#17075) by @doudouOUC
• refactor(managed-agent): drop the catalog probe and repeated root lock from binding control (#17072) by @doudouOUC
• test(cli): compare the kernel off and on in every headless parity scenario (#17130) by @qqqys
• Improves ACP turn handling by replacing host-ownership flags with explicit turn inputs and caller-owned interaction spans. (#17070) by @yiliang114
• Adds an embedding regression test and removes obsolete release repair steps to preserve Managed Agent scheduler isolation. (#17139) by @doudouOUC
• Stabilizes the TUI background notification parity test by normalizing launch size fields and PID-width variations. (#17114) by @qwen-code-dev-bot
• Stabilizes headless and TUI agent-loop parity tests to reduce flaky CI failures without weakening assertions. (#17172) by @yiliang114
• Moves the mid-width toolbar compaction test into the intended container-width band; no product behavior changes. (#17187) by @wenshao
• Stabilizes the managed runtime terminal replay test by attaching the output listener before WebSocket reconnection; no product behavior changes. (#17194) by @qwen-code-dev-bot
• Adds a CI lane that verifies shared-loop and event-bus agent execution across Managed headless, AgentCore, and ACP entry points. (#17180) by @yiliang114
• CI stability improves for large transcript export tests on shared ARM64 runners without weakening regression detection on GitHub-hosted CI or local runs. (#17205) by @qwen-code-dev-bot
• Adds tests verifying Managed Sessions automation handling of null creator keys and tenant isolation, with no production behavior change. (#17235) by @wenshao
• test(external-context): clear lowercase proxy vars in the fail-fast pack test (#17240) by @wenshao
• Internal test change: deflakes the queued-input parity case by treating process-derived shell result size telemetry as volatile. (#17220) by @qwen-code-dev-bot
• Internal test change: strengthens CI coverage for tree-ref encoding guards and corrects misleading test comments. (#17191) by @wenshao
• Internal test change: makes the external-context fail-fast pack test self-contained on proxied runners and adjusts its subprocess timeout. (#17246) by @qwen-code-dev-bot
• Improved web-shell retention tests so wheel events target the transcript and virtualized card restore checks run reliably. (#17242) by @wenshao
• Improved ACP integration tests by counting loop rounds with a dedicated marker for each runAgentLoop call. (#17227) by @qqqys
• Removes an unused per-call client ID override from Web Shell QR-login wiring without changing login behavior. (#16999) by @wenshao
• Moves trusted pull request CI jobs to persistent runners while keeping untrusted PRs on disposable runners. (#17279) by @yiliang114

Full Changelog: v0.27.3...v0.27.4