Repository navigation
·
8 commits
to release/v0.28.3
since this release
Highlights
See the complete change list below.
Breaking Changes
No known breaking changes.
Complete Change List
Features
- feat(web-shell): explain background tasks from the composer chips (#17398) by @wenshao
- Adds the backend-neutral unified chat surface contract and capability guards for Web Shell, with no user-visible behavior change yet. (#17520) by @wenshao
- feat(channels): add GitHub as the third managed channel adapter (#17511) by @wenshao
- Lets Generated UI cards preserve view state across page reloads, reopened conversations, and forks. (#17522) by @qqqys
- feat(web-shell): send to another session with /send (#17539) by @qqqys
- Removes daemon dependencies from shared approval and composer components so the managed chat kernel can use them without loading daemon code. (#17661) by @wenshao
- Adds a send-only Generic Webhook Channel that posts notification, prompt, and scheduled-task results as JSON to a configured HTTPS endpoint, with optional Bearer authentication. (#17610) by @jkxiao
- Managed agents now notify owners when attachments are retired and classify suspension age-out waits with control-plane wait codes. (#17681) by @doudouOUC
- Adds Generic Webhook destinations in the Web Shell Channel editor, with endpoint URL, optional Bearer token, and local/private receiver options. (#17405) by @jkxiao
- Session switching now shows the artifact count immediately from session metadata, before full loading completes. (#17518) by @wenshao
- Enables Mods in ACP sessions for qwen --acp, Zed, and the VS Code companion when experimental.mods is enabled. (#17693) by @qqqys
- Applies the --max-sse-connections admission budget to all daemon stream surfaces, with shared occupancy accounting and logged refusal when the limit is exceeded. (#17662) by @wenshao
- Adds an experimental WeCom managed-channel adapter that routes private-chat messages into bound Managed Agent Sessions and returns replies over the authenticated WebSocket. (#17441) by @wenshao
- Installs and pins default Workspace Skills from an optional catalog when a Workspace Session is created without caller-provided Skills or an explicit pin list. (#17703) by @doudouOUC
- Adds a status-tinted Goal marker to Web Shell sidebar session rows so sessions with active Goals are recognizable and can be opened from the marker. (#17515) by @wenshao
- Adds a Traditional Chinese UI locale to Web Shell and VS Code companion chrome, with zh-TW, zh-HK, zh-MO, and zh-Hant resolving to 繁體中文 [zh-TW]. (#17421) by @callmeYe
- Marks shortened inbound messages from WeCom, DingTalk, and GitHub managed adapters with a truncation notice while keeping them within character and byte limits. (#17726) by @wenshao
- Hosted Harness now suspends idle managed-agent sessions after active work drains and cold-opens them on the next request. (#17690) by @doudouOUC
- Adds an opt-in setting that skips models exhausted by capacity errors for the rest of the turn and re-probes them on the next turn. (#17377) by @wenshao
- Binds DingTalk persisted ledger to registration identity so pending messages are not replayed under a different registration of the same robot app. (#17742) by @wenshao
- Adds an inject channel dispatchMode that delivers incoming messages to the running turn so the agent can read them at the next tool-batch boundary. (#15634) by @Eric-GoodBoy-Tech
- Adds a workspace artifact library to the Managed Agent Web UI for browsing, previewing, and downloading artifacts across sessions. (#17728) by @wenshao
- Adds opt-in native screen-reader presentation for the CLI with keyboard navigation, stable completed-reply announcements, and semantic history while Ink remains default. (#17500) by @chiga0
- Moves the Artifacts (N) header entry to the right panel so clicking it toggles a dedicated session-scoped Artifacts list tab. (#17689) by @wenshao
- Adds the daemon chat intent lifecycle core so submit, cancel, approval, and question states are settled from authoritative evidence rather than transport responses alone. (#17709) by @wenshao
- Wakes the Managed Session bound to a watched pull request with private prompts when CI fails, conflicts appear, or the head changes, without posting publicly. (#17665) by @wenshao
- Adds an opt-in setting, qwen.managed-agent.runtime-broker.mcp-parking.enabled, to keep idle MCP connections across suspension so reopened sessions reuse them. (#17749) by @doudouOUC
- Shows a workflow's declared meta.phases as Not started lanes before they begin, then replaces each lane when the script reaches that phase. (#17782) by @qqqys
- Adds managed-agent support for external Shell workers to publish control-plane events. (#17478) by @yiliang114
- Adds an off-by-default QQ bot option exposing tools for agents to send local files, images, videos, and voice messages. (#15450) by @Eric-GoodBoy-Tech
- Adds a Microsoft Teams channel extension with personal chats, group mentions, thread isolation, and createTeamsIngress host webhook setup. (#15620) by @XIQIXIQIXIQI
- feat(telemetry): record OpenRouter routing backend and generation id (#16495) by @2SinceJuly
- WebShell transcripts expose an optional overscan prop so hosts can tune or disable the offscreen row buffer for virtualized message lists. (#17807) by @callmeYe
- DingTalk managed-agent replies now render as a single editable native card that updates as more reply segments arrive, instead of plain text messages. (#17748) by @wenshao
- feat(core): name the output budget when truncation rejects a write (#16632) by @xxwtiancai
- feat(web-shell): let hosts disable the composer glow (#17814) by @callmeYe
- feat(core): run ACP sessions on the shared agent loop by default (#17744) by @qqqys
- feat(cli): run background workflows in headless mode (#17809) by @qqqys
- Shows Mod-rewritten assistant reply rows in the Web Shell while preserving original text for history, copy, search, and branching. (#17718) by @qqqys
- Adds a --timeout option to /model --vision for setting visionBridgeTimeoutMs, accepting seconds or ms-style values and applying immediately. (#17830) by @qqqys
- Adds the daemon chat controller so incoming daemon events are bound to the correct attempt before settling, preventing late events from affecting finished chat attempts. (#17794) by @wenshao
- Binds the managed GitHub adapter ledger to its registration identity so pending comments are not replayed under a different registration. (#17775) by @wenshao
- feat(managed-agent): clarify Web UI parity and wire local terminal (#17795) by @wenshao
- feat(managed-agent): stamp clean suspensions for wake dispatch (#17908) by @doudouOUC
- feat(core): run OpenTUI on the shared agent loop by default (#17745) by @qqqys
- feat(acp): end a tool-calling turn when queued input is sent now (#17825) by @qqqys
- feat(serve): add daemon restart recovery inbox (#17677) by @wenshao
- feat(web-shell): worktree carry metadata + snapshot-then-delete cleanup (#17237, P6 partial) (#17533) by @wenshao
- feat(acp): deliver pending async hook output on the default legacy round (#17890) by @wenshao
- feat(web-shell): bind the daemon chat controller to the session provider (#17895) by @wenshao
- feat(managed-agent): degrade process components per package in content-only extension runtimes (#17944) by @doudouOUC
- feat(channels): bind the email ledger to its registration identity (#17925) by @wenshao
Bug Fixes
- fix(verify-pr): harden the watch-state shape guard, container nonce, and VM prose (#17440) by @wenshao
- fix(external-context): deflake cumulative-deadline deletion timing (#17342) by @qwen-code-dev-bot
- fix(serve): reserve listener headroom from SSE and reap stalled writers by default (#17502) by @wenshao
- fix(triage): fail closed when QWEN_TRIAGE_DRAFT_DIR is unset in CI (#17549) by @yiliang114
- fix(managed-agent): recover vanished Harness admissions safely (#17521) by @doudouOUC
- Fixes Web Shell so PR watch deep links can be served before authentication, improving direct link access. (#17119) by @Afloat16
- Updates the Read unchanged-file message to tell users to re-read with a different offset or limit when earlier content is unavailable or out-of-band changes are suspected. (#17465) by @callmeYe
- On touch devices and narrow viewports, the composer model picker now opens as a bottom sheet with model search and reasoning controls, while desktop behavior stays unchanged. (#17401) by @callmeYe
- Limits Grep fallback searches to tracked and non-ignored untracked files so results match git-grep and avoid excluded build output. (#17530) by @callmeYe
- Recovers resubmitted managed-agent turns that are already accepted and, when the parked execution is ready, drives its continuation instead of leaving it stalled. (#17680) by @doudouOUC
- Managed Web Shell now shows request, session, and recovery errors as normal conversation messages, with retry via message link or Ctrl+Y and fewer separate recovery panels. (#17557) by @yiliang114
- Keeps managed DingTalk replies sent in multiple chunks from being truncated and documents why no aggregate send limit is applied. (#17529) by @wenshao
- Cold worktree restoration now delays automatic work until ownership validation and working-directory restore finish, preventing blocked restores and preserving pending questions. (#17654) by @doudouOUC
- Web Shell now removes the extra gap under /send records and keeps each reply visible beside its Cross-session message card when turns are collapsed. (#17686) by @qqqys
- /verify-pr --watch now registers its watch round before prepare, and --start no longer requires --artifact-dir, improving liveness without breaking existing callers. (#17550) by @wenshao
- Long-running processes now release idle JSONL append locks after successful or failed writes settle, reducing lock accumulation. (#17577) by @Afloat16
- Web Shell Memory panel now has a Refresh control that reloads the file list and retries the selected file without closing the panel, restoring editing when reads succeed. (#17108) by @Afloat16
- SDK header merging now treats HTTP header names case-insensitively, so caller headers override transport defaults instead of creating duplicate values. (#17579) by @Afloat16
- CLI exit cleanup now preserves pending cleanup work when handlers unregister, preventing skipped or repeated cleanup during shutdown. (#17572) by @Afloat16
- Stops delivering buffered SSE events after a stream is cancelled, preventing stale session updates. (#17569) by @Afloat16
- Cancels heartbeats for sessions removed or replaced during heartbeat registration, preventing leaked timers. (#17117) by @Afloat16
- Custom commands now preserve argument text literally during prompt expansion, so $$ and $& are not reinterpreted. (#17367) by @Afloat16
- Keeps Python SDK initialization running when one waiting control call is cancelled, allowing other calls and prompt input to continue. (#17499) by @Afloat16
- Shows an amber busy marker in Web Shell sidebar session rows when background work is running and aligns row actions with that state. (#17676) by @wenshao
- Delivers completed background async hook output to the model on the next turn instead of silently dropping it. (#17477) by @wenshao
- Labels shell tool results as Output (stdout and stderr combined) and Execution Error, while keeping compatibility with earlier labels. (#17685) by @qqqys
- CLI and Web Shell copy extraction now keeps fence-like lines inside copied code and does not extract math from code blocks, reducing truncated copies. (#17578) by @Afloat16
- ACP Java SDK now returns selected file lines joined by normal newlines instead of path separators, preventing corrupted file content shown to the agent. (#17688) by @Afloat16
- The SDK now clears permission callback deadlines after failures, preventing stray timeout rejections, crashes, or leftover timers after a denial. (#17581) by @Afloat16
- Web Shell change overview now limits its commit fallback to the current session's own commits and hides it when no reliable session start exists. (#17514) by @wenshao
- Raises the document export renderer JavaScript size budget so newly added Web Shell transcript copy does not fail the build. (#17746) by @wenshao
- SDK now rejects oversized peer frames before dispatch, even when completed in one chunk, and stops overflowing batches while preserving valid separate frames. (#17576) by @Afloat16
- Managed DingTalk streams now centralize reconnection so a single drop starts at most one reconnect chain, and concurrent connect calls join the in-progress attempt. (#17691) by @wenshao
- Cancelling a composer image batch now stops queued image reads that have not started, preventing unnecessary reads and base64 allocation after cancellation. (#17705) by @cactuser-Lu
- Restores managed-agent class-size and facade-population gate checks without changing runtime behavior, fixing deterministic gate failures left on main. (#17753) by @doudouOUC
- SDK now waits for actual child termination, classifies intentional shutdown, preserves failures, and tears down transport before rejecting pending Query and MCP requests. (#17580) by @Afloat16
- Fixes QQ bot direct chats so voice attachments from qqbot.ugcimg.cn are downloaded and delivered as audio instead of empty turns. (#15339) by @mvanhorn
- Shows the managed composer attachment entry for unbound active Sessions when their Harness supports input attachments. (#17722) by @wenshao
- Stabilizes the composer toolbar near width thresholds so approval and model labels no longer shift or disappear when the chat-width toggle appears. (#17729) by @callmeYe
- Fixes qwen-managed-agent-server tool-result publication failures when embedded with json-schema-validator 2.x. (#17752) by @yiliang114
- The delete dialog now blocks deleting any session with running background work and shows the reason, while the current session remains deletable once no work is running. (#17713) by @wenshao
- Time to first token for streamed requests now starts when the request is dispatched, making TTFT and related timing metrics more accurate. (#17790) by @qqqys
- Sessions that contain only side records now show their artifact count on first open, and remaining artifact-summary display issues are fixed. (#17711) by @wenshao
- fix(web-shell): Group approval scopes in a dropdown (#17801) by @callmeYe
- fix(ci): raise the CodeQL JavaScript cap to 180m after the nightly overran 90m (#17808) by @yiliang114
- fix(serve): pin remote-access admission and non-REST shutdown drain (#17712) (#17737) by @wenshao
- Completes Russian UI coverage in Desktop and CLI and keeps long account-menu labels on one line. (#17719) by @callmeYe
- Fixes permission rules so Windows drive-letter and UNC absolute paths match correctly, including Always Allow entries. (#17810) by @yiliang114
- Allows LocalDirectory writes of valid text beginning with a UTF-8 BOM while still rejecting invalid unpaired surrogates. (#17820) by @yiliang114
- fix(web-shell): evict aborted mention cache lookups (#17826) by @yiliang114
- Ensures JSONL logs remain readable after an interrupted append by adding a leading newline before the next record when the previous record is unterminated. (#17771) by @wenshao
- Adds missing Russian translations for experimental.mods, visionBridgeTimeoutMs, and /model --vision --timeout CLI messages. (#17858) by @wenshao
- Compression now reduces or refuses summarization requests that cannot fit the model window and recovers once from context-length rejections. (#15936) by @yanyukang2003-ux
- Web Shell markdown images that point to local paths or fail to load now render as selectable text chips instead of broken placeholders. (#17866) by @doudouOUC
- Caps the joined async hook context block sent to the model at 10,000 characters to prevent oversized requests. (#17736) by @wenshao
- Preserves MCP tool-reported failure diagnostics in failed tool responses so CLI and ACP replay the exact tool error content. (#17803) by @callmeYe
- Fixes six deferred managed /loop issues, improving behavior and related test coverage. (#17594) by @wenshao
- Allows editing a draft in Web Shell while a saved session loads, while sending remains disabled until the session is ready. (#17842) by @callmeYe
- fix(managed-agent): settle recovery from durable native terminals (#17886) by @doudouOUC
- fix(web-shell): decouple first prompt from automatic title writes (#17843) by @callmeYe
- fix(core): bound JSONL tail diagnostic recovery and cache churn (#17901) by @wenshao
- fix(managed-agent-server): keep journal blob reclamation moving past sealed heads (#17907) by @doudouOUC
- fix(memory): rebuild extraction indexes in the runtime (#17785) by @callmeYe
- fix(daemon): include memory metrics in detailed status JSON (#17906) by @callmeYe
- fix(agents): reject incompatible Host replacements (#17702) by @water-in-stone
- fix(i18n): unify ru timeout spelling and vision-bridge term (#17867) by @wenshao
- fix(web-shell): keep the model sheet's rows and search field touch-sized (#17892) by @wenshao
- fix(managed-agent): settle a parked turn whose runtime is terminally lost (#17932) by @doudouOUC
- fix(ci): route autofix reviews through live bot (#17881) by @yiliang114
- fix(core): resolve dotted DeepSeek catalog limits (#17876) by @yiliang114
- fix(channels): address Aone channel deferred review findings (#17517) (#17519) by @wenshao
- fix(channels): own DingTalk endpoint discovery so a hung connect stops muting the process (#17773) by @wenshao
- fix(cli): attribute daemon log lines and bind late channel restores to the boot conflict decline (#17904) by @wenshao
- fix(web-templates): re-track renderer byte budget (#17940) by @wenshao
- fix(managed-agent): create the ephemeral Runtime state directory before spawning (#17949) by @yiliang114
- fix(models): support Gemini proxy streaming and model refresh (#17947) by @wenshao
- fix(managed-agent): load transport wakes through the passive verb (#17955) by @doudouOUC
- fix(core): tie async hook output consumption to the accepted turn (#17930) by @wenshao
Performance
- Speeds up follow-up suggestions by limiting context to a compact 4,000-character projection when cache sharing is unavailable or a different model is used. (#17532) by @callmeYe
- Improves session catalog performance by loading sidecar data only for returned rows instead of all sessions. (#17738) by @callmeYe
- Indexes bundled startup chunks to speed up packaged serve and ACP startup, cutting median workspace-ready time by 41.16% in one measured sandbox. (#17720) by @callmeYe
- Improves Web Shell theme switching performance by reducing repeated style recalculations. (#17786) by @callmeYe
- Reduces Web Shell input lag during history replay, with a recorded case improving key-to-editor delay by up to 1,706 ms. (#17788) by @callmeYe
- Improves web-shell typing performance by preventing composer input from triggering unnecessary transcript relayouts. (#17821) by @callmeYe
- Web Shell loads Simplified Chinese, Traditional Chinese, and Russian language catalogs on demand to reduce startup JavaScript while keeping English fallback. (#17833) by @callmeYe
- Improves managed-agent-server responsiveness by limiting database scans and round trips that grew with unrelated history. (#17856) by @doudouOUC
- perf(web-shell): load xterm on demand (#17840) by @callmeYe
- fix(acp-bridge): merge background-turn live text chunks (#17862) by @wenshao
- perf(web-shell): render large workspace PDFs on demand (#17899) by @callmeYe
- perf(web-shell): gzip serve responses for large transcript loads (#6181) (#11220) by @now-ing
Documentation
- docs(design): propose idle Session suspension toward a stateless Hosted Harness (#17527) by @doudouOUC
- docs(design): drop the 20-run parity requirement from the loop rollout steps (#17535) by @qqqys
- Updates managed-agent documentation and regression tests to match current read-retry and probe-removal behavior. (#17082) by @wenshao
- Clarifies that /btw is the recommended side-question syntax for Ink and OpenTUI and explains how ?btw behaves in each. (#17759) by @Barzabel
- Adds tests and design documentation for Web Shell sidebar goal marker double-click protection and clearing transitions. (#17733) by @wenshao
- Updates English and Chinese design documentation to mark the PR-watch bound-session feature as merged and record its follow-up tracking details. (#17793) by @wenshao
- Updates docs to say new sessions start in Auto Mode and how to start Ask Permissions Mode via setting or --approval-mode default. (#17799) by @qqqys
- docs(security): add GitHub private vulnerability reporting to SECURITY.md (#17806) by @yiliang114
- feat(serve): let create_sub_session reach any session, skip its workspace prompt in YOLO, and drop the thread glyph (#17818) by @qqqys
- Adds bilingual documentation for Phase 3 control-plane wake scheduling and Runtime-owned background observation in the Hosted Harness roadmap. (#17805) by @doudouOUC
- Adds a verification guide for Web Shell changes, covering daemon capability fallback, partial settings writes, channel session startup, and model turn evidence. (#17762) by @Barzabel
- Updates the H5e card-surfaces design status to record the merged DingTalk interactive-card private-chat vertical, its verification ledger, and deferred follow-ups. (#17815) by @wenshao
- Adds documentation for the Phase 3 seal-time wake stamp addendum that replaces the earlier wake-scheduling mechanism. (#17869) by @doudouOUC
- docs(web-shell): link the deferred background-work reporting gap to its issue (#17850) by @wenshao
- Adds English and Simplified Chinese verification documentation for legacy JSONL append isolation behavior and follow-up gaps. (#17853) by @wenshao
- docs(managed-agent): mark the deferred loop pins design merged with its verification ledger (#17905) by @wenshao
- docs(core): mark JSONL tail diagnostics design as merged (#17913) by @wenshao
- docs(managed-agent): design Runtime-owned background observation (Phase 3b) (#17912) by @doudouOUC
- docs(managed-agent): record the workspace artifact library's merged status (#17779) by @wenshao
- docs(worktree): record merged delivery and verification (#17928) by @wenshao
- docs(managed-agent): record web UI parity delivery (#17915) by @wenshao
- docs(design): record background-turn live replay landing (#17916) by @wenshao
- docs(design): record recovery merge and fix workflow test isolation (#17941) by @wenshao
Internal Changes
- test(managed-agent): observe a real scheduled /loop fire in the full-stack E2E (#17506) by @wenshao
- Raises the timeout for a DingTalk webhook-cache test to reduce flaky CI failures without changing product behavior. (#17615) by @qwen-code-dev-bot
- Refactors the Web Shell composer status stack into a shared component with no user-visible behavior change. (#17684) by @wenshao
- Adds a 120-second per-test budget for Web Shell browser smoke tests that boot the app twice, covering setup and test body. (#17487) by @qwen-code-dev-bot
- Adds tests for disabling a Skill by workspace-qualified name when the workspace catalog is empty, including settings write failure handling. (#17766) by @Barzabel
- Stabilizes the CLI SSE pairing-listener clamp test by shutting down the daemon before deleting its temporary log directory. (#17764) by @qwen-code-dev-bot
- Improves end-to-end test isolation by cloning gitRemotes fixtures and correcting the fixture-ownership comment, with no user-facing behavior change. (#17800) by @wenshao
- Adds a CI meta-test ensuring scheduled workflows keep non-empty cron triggers. (#17829) by @yiliang114
- Internal test changes improve coverage for listener stream limits and socket pruning without changing user-visible behavior. (#17841) by @yiliang114
- Adds a test verifying that the Web Shell mobile composer retains editor focus when switching between simple and advanced modes. (#17828) by @Barzabel
- Strengthens e2e fixture-ownership tests to pin cloned fixture values and content, helping catch regressions that could leak shared test data. (#17824) by @wenshao
- Managed-agent file sessions now use the unified lifecycle v2 close path and related preflight checks. (#17871) by @doudouOUC
- Removes pre-production rolling-upgrade version machinery and unifies managed-agent version-shaped formats. (#17864) by @doudouOUC
- test(managed-agent): re-home publication lifecycle IT v1 witness to external-files (#17909) by @doudouOUC
- test(integration): deflake the export document budget on all CI runners (#17884) (#17887) by @qwen-code-dev-bot
- test(managed-agent): scope the workspace artifact index assertion to its own database (#17893) by @wenshao
- refactor(managed-agent): key tenant guard on readable VARCHAR scope (identity de-hash increment 1) (#17927) by @doudouOUC
- test(managed-agent): run the wake stamp suites on MySQL and pin each suspension gate clause (#17954) by @doudouOUC
- refactor(managed-agent): drop redundant storage keys (#17958) by @doudouOUC
- refactor(managed-agent): drop redundant in-process locks and the archived legacy Session guard stub (#17959) by @doudouOUC
- test(managed-agent): use canonical schema initializer (#17921) by @wenshao
New Contributors
Full Changelog: v0.28.0...v0.28.3