Skip to content

[pull] main from lobehub:main - #495

Open
pull[bot] wants to merge 9302 commits into
code:mainfrom
lobehub:main
Open

pull[bot] wants to merge 9302 commits into
code:mainfrom
lobehub:main

Conversation

@pull

@pull pull Bot commented Jan 27, 2026 •

Copy link
Copy Markdown

See Commits and Changes for more details.


Created by pull[bot] (v2.0.0-alpha.4)

Can you help keep this open source service alive? 💖 Please sponsor : )

@pull pull Bot locked and limited conversation to collaborators Jan 27, 2026
@pull pull Bot added ⤵️ pull merge-conflict Resolve conflicts manually labels Jan 27, 2026
tjx666 and others added 27 commits November 7, 2026 00:36
…22382)

* 💄 style(settings): match settings skeletons to their settled layouts

Every personal settings tab now registers a skeleton that mirrors its page:
same container width, list vs block shape, row counts, and static titles and
labels rendered as text. All tabs share one navigation header with the tab
title; page actions portal into its right slot.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(settings): move page actions into the content column and align more skeletons

Page actions now render in a row at the top of the content column instead of
the navigation header, so they stay within the page body. Proxy labels use the
electron namespace, the labels skeleton draws its agent/task tabs, the trash
filter and proxy controls match the settled widths, and the connector builtin
list shows five rows.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(integrations): render the directory when a connection lookup fails

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(settings): use the global section class after the CSS Modules removal

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…#22411)

* ✨ feat(chat): send the operation id with client-runtime chat requests

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(chat): attribute context-compression calls to their run

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…22434)

* 🐛 fix(page-collab): keep personal pages out of collaboration rooms

A personal page has a single editor, so a room only adds its costs and its
bootstrap race. New clients edit personal pages on the stored row, and server
writers no longer open a room for one; a room an older client still holds
open keeps receiving server writes so they are not projected over.

* ✅ test(page-agent): cover initPage on personal and workspace pages with rooms on

The collaboration cases used a page row without a workspace while its permission metadata had one; give the row its workspace, and add the personal page case, which now writes the stored row.

* 🐛 fix(page-editor): refresh the page header title after an agent edits a stored page
…alias (#22366)

Co-authored-by: lobehub-coder <340605552+lobehub-coder@users.noreply.github.com>
#22362)

Context compression has always run on the conversation model. Drop the
historyCompress system-agent slot and chatConfig.compressionModelId, which
had no runtime consumer.

Refs T-11702

Co-authored-by: lobehub-coder <340605552+lobehub-coder@users.noreply.github.com>
…for custom MCP connectors (#22456)

* ✨ feat(connector): support OAuth Client ID Metadata Documents

Custom MCP connectors now identify LobeHub with a Client ID Metadata
Document (SEP-991) when the authorization server advertises
client_id_metadata_document_supported, before falling back to RFC 7591
dynamic registration. Private deployments whose APP_URL an external
server cannot fetch keep using DCR.

* 🐛 fix(connector): keep server-resolved OAuth clients out of the edit form

The custom connector edit form pre-filled oidcConfig.clientId for every
OAuth connector, and saving it marked any filled Client ID as
pre-registration. A Client ID Metadata Document URL got pinned as a
hand-registered client, and a DCR-issued id lost its client secret,
which the edit read never returns. Only show the Client ID the user
registered themselves; automatic clients are re-resolved on save.
…ile loading (#22462)

* 🐛 fix(desktop): stop content buttons punching holes in the titlebar drag region

Chromium reports every app-region element's *unclipped* border box in layout order and
the later rect wins (`gfx::DraggableRegionsToSkRegion`). The page-wide
`button { -webkit-app-region: no-drag }` in global.css therefore turned every button in
the page into a hole in the titlebar, including buttons in the content that scroll up
under it. Such a button cancelled the drag region over an arbitrary band of the tab
strip: dragging there did nothing while the rest of the row still moved the window.

Scope the opt-out to buttons inside a drag region, the only place it is needed. StyleX
cannot express a descendant selector, so `draggable` also composes the `lobe-drag-region`
hook class and global.css keys the rule on it.

Verified on Electron 44.5.1 with ELECTRON_DEBUG_DRAGGABLE_REGIONS plus synthetic window
drags: a content button whose unclipped box sat at y 0..18 left the titlebar dead at
y=8 while y=30 still dragged; with the scoped rule the same drag at y=8 moves the window
and the titlebar's own buttons stay non-draggable.

* 🐛 fix(desktop): keep the titlebar draggable on the static loading screen

The windows are frameless, so the titlebar band is only draggable through CSS. Until React
mounts, the static loading screen is the whole window and declared no app-region at all:
measured 0 app-region elements for the first ~2.5 s in the dev app, so the titlebar could
not move the window while the loading screen was up.

Add a TITLE_BAR_HEIGHT drag strip to the loading screen of both renderer entries (main and
popup). It is removed together with `#loading-screen`, and the app shell skeleton's own
drag region takes over once React mounts.

Verified in the Electron dev app: dragging y=12 of the titlebar band 350 ms into a reload
moves the window, and the loading phase now reports a single 38px drag region instead of
none.
* ✨ feat(task): let CLI/API task creates set an origin that is notified on completion

task.create accepts context.origin {agentId, topicId, messageId?, operationId?, toolCallId?}
so TaskResultBridgeService reports the result back for tasks created outside the
createTask tool. The server rejects an origin the caller does not own: the topic must
be the caller's own conversation, the agent usable by the caller and the topic's agent
(or a member of its group), and the message inside that topic. task.update, which
already replaced the whole context column, now applies the same check to a changed origin.

lh task create gains --origin-agent/--origin-topic/--origin-message/--origin-operation/--origin-tool-call.

* 🐛 fix(task): keep context.origin optional on task updates

An update whose context carries no origin, or one the result bridge cannot
deliver to (missing agent or topic), is stored as sent again, as before the
origin check. Only a changed deliverable origin is checked against the
caller's conversations. A create without an origin no longer passes a
context key at all.

---------

Co-authored-by: lobehub-coder <340605552+lobehub-coder@users.noreply.github.com>
…ersationWake service (#22447)

* ♻️ refactor(bot-wake): move the wake pipeline into a platform-neutral conversationWake service

The wake policies, aggregator, config, conversation helpers, scheduler machine and the classifier
and duty-judge model calls lived under the bot code and the Agent Signal scenario, so nothing but a
chat-SDK bot could reuse them. Channel needs the same pipeline to decide when an Agent member
speaks.

- New `apps/server/src/services/conversationWake/` holds everything that does not know which chat
  platform a message came from.
- Discord / Lark specifics (mention parsing, channel and log keys, snowflake time) move to
  `bot/wake/platformFacts.ts`.
- `agentSignal/scenarios/botWake` keeps only the Agent Signal scenario wiring.
- Persisted identifiers stay as they were: policy ids `bot-wake:*`, pipeline `bot-wake@3`, signal
  sources and settings keys.

Path-only move: no behaviour change and no assertion changed; tests move with their modules.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* ♻️ refactor(bot-wake): decide unaddressed wakes in a platform-neutral engine

The "should this participant speak up on its own?" flow (named in the text, a follow-up the
classifier confirms, a message within its duties) lived inside the bot recorder, next to Redis
reads, parked-message replay and decision rows. Channel needs the same decision for its Agent
members without any of that.

- `conversationWake/initiation.ts`: `decideWakeInitiation` takes the policies, the signal and a
  state snapshot and returns the wake (if any) plus the verdicts for the record. Also
  `canInitiateWake`, `initiationFor` and the `WakeInitiation` types, renamed from `BotWake*`.
- The bot recorder keeps its IO (state reads, replay memory, evaluation cache, persistence) and
  calls the engine. The order of policies and every outcome are unchanged; the existing bot-wake
  tests pass as they were.
- Contract tests for the engine itself.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* 🐛 fix(linear): archive the task of an issue archived in Linear

A sync read archived issues but only mapped their state, so their tasks
landed in the working lists. Archiving now follows Linear one way, once per
Linear archive: a task restored in LobeHub stays restored, and a live task
waits for a later run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(linear): recheck the applied archive under the task's row lock

An overlapping sync working from an older read could archive again a task
restored since another run applied the same Linear archive. Archiving and
recording it now happen together in one transaction, after a fresh read.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(linear): archive closed tasks whose issue is archived later

A rule run without archived issues never reread a closed task, so an issue
archived after closing, as Linear does on its own, left its task out. The
closing pass now asks Linear for only the archived ones among those links.

A task that has to wait is reported with a reason, not as already up to date.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: rdmclin2 <4705237+rdmclin2@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…s stop growing quadratically (#22522)

Every call_llm step stored its full `contextEngine.output` (the whole message array sent to the
model) and only skipped it when identical to the previous step. In queue mode each step also reads
the whole partial, appends, and rewrites it. A ~740-step sub-agent run produced a 316MB (raw) partial,
and each step's read-modify-write added ~1.5GB of RSS to the worker.

- Store the output as a content-addressed list node `{ base, refs }` that extends the longest prefix
  shared with the last recorded list; message bodies go into per-step `pool`s by hash, once each.
  Top-level input keys are pooled too, so a small key change no longer re-copies agentDocuments.
- Add `resolveContextEngineSteps` (order-independent; reads the old changed-fields format and mixed
  partials); the viewer, ctx-lint/ctx-map/replay and `inspect -j` use it.
- Replaying the real 734-step trace: every step resolves identically, CE 236MB -> 9.6MB,
  partial 316MB -> 19.8MB, per-step peak ~41MB.
…22524)

* 🐛 fix(agent-runtime): refuse to commit a step after its lock is lost

A step whose LLM call outlives the step lock TTL can be claimed by a queue
redelivery while it still runs. Both attempts then saved the step result,
the stale one with older state, and each scheduled its own next step, so two
chains continued the same operation and every later LLM call was paid twice.

Fence the commit on lock ownership:

- re-check the lock (owner-scoped refresh) once the LLM call returns, before
  anything writes on the step's behalf, and again right before
  `saveStepResult`; a refresh that fails or cannot reach Redis counts as lost
- on loss, drop the in-memory trace partial so the `finally` flush cannot roll
  back the steps the new owner records, and throw `StepOwnershipLostError`
- `failStep` rethrows it without writing an error state, emitting completion
  signals or finalizing the trace: the run belongs to the new owner
- the run-step handler already answers a thrown step with 500, so QStash
  redelivers it and the existing claim decides who runs it (conflict →
  reschedule/ACK, stale step → ACK, parked inline envelope → resume)

Once the step result is saved, a redelivery of the same step is skipped as
already done, so the committing attempt stays the only one that hands the
next step on; nothing after the commit needs fencing.

Not atomic with the save, and LLM/tool calls already in flight are not
interrupted: losing the lock only stops the commit.

* 🐛 fix(agent-runtime): fence step commits only where the queue redelivers

The local queue drops a failed callback and the sync path has no queue, so a
refused commit there is never retried: a Redis blip during the ownership
refresh left the step uncommitted and the run without a next step. Neither
mode can have a second delivery owning the step either, so fence commits only
on a redelivering queue and keep the previous behaviour elsewhere.
… Redis event stream (#22528)

Every `step_start` carries the whole UI message list (megabytes on a long topic),
and the Redis stream stored a copy per step: one long operation's stream reached
323MB, and an SSE read from `'0'` (no COUNT) pulled all of it into one process,
~700MB of heap per read. `getStreamHistory(…, 50)` read dozens of snapshots too.

Clients take this snapshot from the gateway push, which publishes the event
before the stream strip. Nothing reads it back from the stream: the SSE route
serves the CLI, whose `step_start` handler only prints the step number, and
`getStreamHistory` is debug output. So drop `uiMessages` from `step_start` at the
stream chokepoint (`stripFinalStateInEventData`); the gateway push is unchanged,
and the single `agent_runtime_end` snapshot per operation stays.
…its viewport (#22532)

A Canvas page that pins itself with `html, body { height: 100%; overflow: hidden }`
and no inner scroll area clipped everything past the first viewport. The preview
frame is exactly one viewport tall, so the side panel had no scroll host at all:
wheel, trackpad and scrollbar did nothing.

The host already owns the page frame (it forces the body's gutter and reading
width with a layered `!important`), so it now owns the scroll axis too: the
document's vertical axis stays visible while the horizontal clipping a page wrote
for itself is left alone. An app-shell page's own `overflow: auto` region keeps
scrolling and a clipped report grows instead.
* ✨ feat(task): open cross-workspace task links in the side panel

A task link pointing at another workspace used to navigate the whole page
into that workspace. When the host has a portal, it now opens the task
beside the conversation, read through the owning workspace's scope and
shown read-only, with an action to open it in its workspace.

* ♻️ refactor(task): share the foreign task not-found error

---------

Co-authored-by: lobehub-coder <340605552+lobehub-coder@users.noreply.github.com>
* 🔨 chore: add local DeepSeek i18n command

* 💄 style: refresh locale translations with DeepSeek Flash

* 🔨 chore: use DeepSeek in the existing i18n command
…22253)

* 🐛 fix(spend): keep Agent Signal verbatim and bill task-result wakeups apart

Every locale had machine-translated the feature name Agent Signal; both labels
now read it verbatim.

The task-result creator wakeup was stamped agent_signal although it is neither
Agent Signal work nor gated by Labs → Self-evolving. It gets its own
task_callback trigger, so spend rows label it separately. Every branch that read
agent_signal for the wakeup keeps its behavior for task_callback: composer
supersede, background-run handling, hidden sub-agent threads, and the inline
skill lane.

* 🔨 chore(i18n): keep Agent Signal untranslated in generated locales

* 🔥 chore(i18n): drop the unused agentSignal.receipts.agentSignalLabel key

No UI has read it since the receipt marker was removed in #15179.
YuTengjing and others added 30 commits November 12, 2026 02:07
* 📝 docs(agents): track work in LobeHub tasks instead of Linear

Remove the linear skill and point PR bookkeeping and deep-review
hand-offs at LobeHub tasks managed with lh task.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

* 📝 docs(pr): pause linked subtasks only after their bookkeeping

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 📝 docs(deep-review): draft legacy hand-offs as LobeHub tasks

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
…eamInterrupted (#22827)

When the upstream resets the connection while the SDK iterator is reading the body,
undici throws `TypeError: terminated` (cause `read ECONNRESET`). The first-chunk error
handler defaulted it to `ProviderBizError`, a fallback bucket, so the router rewrapped
it as a harness `StreamChunkError` (E7001, countAsFailure). Classify these disconnects
as `ProviderStreamInterrupted` (E6005, provider), the same as a stream that ends
without a terminal event.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…22746)

* ♻️ refactor(brief): move the day news digest onto @lobechat/replica

- `useFetchNewsByDay` is a `briefNews` replica keyed by local day (IndexedDB,
  identity-scoped, maxEntries 7, 10s dedupe); Home Inbox reads it through
  `briefNewsSelectors` and keeps the previous day on screen while paging.
- A restored empty day stays "loading" until the server confirms it.
- Read / resolve / delete link the digest with the unresolved feed, so writes
  land on both (persisted rows included); mark-all-read marks the digest.
- `decideGoal` revalidated a `brief:list` SWR key that no longer exists; it
  now revalidates the `briefList` replica.
- Drop the unused `briefKeys` SWR key.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* ⚡️ perf(brief): pre-hydrate the inbox feed and today's digest in the Home route loader

Home's index routes (personal and workspace) seed the persisted briefList
and today's briefNews rows before the route commits, bounded by
PRE_PAINT_HYDRATE_TIMEOUT, so a reload paints the last inbox instead of a
brief skeleton until the mount-time hydrate lands.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
… client clock (#22842)

* 🐛 fix(agent-runtime): order gateway step rows by server time, not the client clock

A gateway step inserts its assistant row from the `stream_start` seed with the
client's `Date.now()`, while the user message it answers carries the server's
time. The flat list sorts by `createdAt`, so on a client whose clock runs about
0.5s or more behind the server, every reply of the run sorts above the user
message and the message sits at the bottom until the topic is reloaded.

- `stream_start` seeds carry the stored `createdAt` (epoch ms) of the assistant
  row and of a joined steer parent.
- The client keeps that time when inserting the seed, and never files a seeded
  row at or before its parent (covers older servers that ship no time).
- The `createMessage` dispatch accepts an explicit `createdAt`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* ✅ test(agent-runtime): expect the seeded createdAt on stream_start createMessage

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…22848)

A web client inside the `agent_llm_relay` rollout sends `llmExecutor` with every
execAgent call, and the deferred-init gate counted it as an orchestrated run, so
no send from those users was ever deferred. The init never reads the executor:
`createOperation` stores it on the pre-init state either way.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
An unaddressed Channel message now has exactly one primary responder. Every
other member woken for it becomes a `followup` deferred behind it
(`channel_jobs.wake.afterMemberId`). It is queued as `awaiting_primary` and
claimed only after the primary's job settles, with the primary's published
reply in its prompt (`channel_reply`), so it can stay silent or add only what
is distinct.

- Decision model: the most likely member at or above 0.7 leads; other
  confident and uncertain members follow it. When nobody reaches 0.7, only
  the likeliest uncertain, non-quiet member is woken.
- Own rules: the member named first leads among several named ones;
  arbitrated claims are led by the most likely member, and otherwise by
  follow-up window, then duties.
- Explicit @mentions, paired groups and the all-candidate fallback are
  unchanged.
- The routing audit records primaryMemberId/deferredMemberIds, and decision
  rows record whether a member was the primary or deferred.
- Reply activity does not count jobs waiting on the primary.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…22836)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…panel (#22850)

* ✨ feat(channel): open a sidebar thread as the page instead of a side panel

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(channel): keep thread attachments and hide the reply topic when a thread opens as the page

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…top on cancel (#22819)

Both attempt loops now share createRouteRun for preparing channels, reporting
attempts, the fallback decision and route success; only the commit point
differs. Hook payloads keep their existing shape per path. The shared fallback
decision stops on an aborted signal, so cancelled non-chat requests are no
longer replayed on the remaining channels.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…e work location picker (#22853)

* 💄 style(project): show device presence as a dot beside the name in the work location picker

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(project): keep the presence label beside the dot in the work location picker

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(project): drop the sandbox and local description lines in the work location picker

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…mail on contact card (#22845)

* 💄 style(channel): show the date on message timestamps like Discord

Day dividers show the full date; message headers show the time for today,
"昨天 HH:mm" for yesterday and the full date for older messages, so a message's
day stays readable after its divider scrolls away. Rewrites LOBE-UX-020.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* ✨ feat(contacts): sort contacts and show an agent's email on its card

Adds a per-device sort (name / recently updated) to the 助理 and 群组 headers, and
a contact-info section with copy and mailto actions for agents that own an address.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(channel): center the composer placeholder on the attach button

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(messages): say typing in a replying row instead of a green running spinner

A row whose Agent is replying showed a green spinner and "执行中…" while the
conversation header said "正在输入中". The row now shows the header's typing
label and dots in the preview's quiet colour; a blocked reply keeps the preview.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(channel): set the author line on one baseline, a step quieter

Channel messages opt into a quiet ChatItem header: name and time share a baseline, the
name drops to the secondary text colour and the time to tertiary. The main chat is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(messages): show typing beside a row's name and keep its preview

Matches the conversation header: "正在输入中 •••" follows the name, and the preview line stays.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(messages): give the Lobe AI row the same preview and time as other Agents

The inbox row previewed only its direct Channel or description and hid its time, so it
rendered as a bare title. It now shares the Agent rows' preview rule and time, with its tagline
as the last fallback.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(messages): show a project Channel with its project's logo

A Channel grouped under a project uses the project avatar; a standalone one keeps "#".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* ✨ feat(contacts): pin contacts into their own section

Pinned agents and groups lead the contact list in a 置顶 section and leave their own
sections; each contact gets a pin / unpin context menu.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(channel): keep a quiet author name one step above its time

`type="secondary"` resolves to the description colour, the same step as the time, so the
Channel author name dropped two steps and matched the time. Use colorTextSecondary.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(messages): keep a row's typing dots clear of its time

With `flex: 1` and `minWidth: 0` the typing label only got the width left after the name, so
beside a longer name its dots ran into the time. It now keeps its own width and the name
ellipsizes instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
… and valid PDF parsing (#22862)

* fix: count parsed knowledge chunks by physical file id

* 🐛 fix: skip storage URL resolution for native document placeholders

* 🐛 fix: pass a plain Uint8Array to the PDF chunk loader
…content reads (#22863)

* feat: add business knowledge access seam and scoped content reads

* 🐛 fix: read ordinary knowledge content in the explicit caller scope

* ✨ feat: extend the business knowledge seam with per-item access lists

* ⏪ revert: keep the business knowledge seam library-level

Per-item access lists need the distribution's own content authorization;
the community content reads only enforce library-level access, so the
seam should not advertise item-level policies it cannot enforce.

* 🐛 fix: pass the executing user to knowledge and attachment tool reads

* ♻️ refactor: bound concurrent library access resolution

* 🔒 fix: only accept file-reading credentials for managed files at the file proxy

* 🔒 fix: keep managed file reads personal-scope and move reader policy into the server service
)

* 🔒 fix(server): fail closed on unauthenticated workflow callbacks

* 🐛 fix(server): sign direct callback fallback when QStash publish fails

* ✅ test(server): sign widget workflow callbacks in integration tests
* ✨ feat(desktop): serve the remote terminal from the desktop app

The chat terminal's device picker only offered devices with a live `lh connect`
(cli) connection, because only the CLI daemon hosted the PTY behind the
terminal RPCs. A machine running just the desktop app never showed up, so the
picker was stuck on "This computer".

- Move `TerminalSessionManager` from the CLI into
  `@lobechat/device-control/terminal` with an injected `node-pty` loader, plus
  `createTerminalDeps()` for the dispatcher handlers.
- The desktop's gateway connection now serves `createTerminalSession` /
  `writeTerminal` / `readTerminal` / `resizeTerminal` / `closeTerminal` and
  reports `metadata.terminal = '1'` when it registers its device.
- `canHostTerminal` accepts a desktop connection whose row reports that key, so
  older desktops (which refuse the RPCs) stay hidden; the desktop's own row is
  left out of its picker since "This computer" already covers it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(terminal): keep a session's calls on the client that created it

A device can hold both `lh connect` and the desktop app, and each keeps its
own terminal sessions. The gateway re-picked a connection on every call
(preferring the CLI), so a shell opened on the desktop would have its reads
sent to the CLI once `lh connect` came up; the CLI answers an unknown id with
`exited`, which closed the pane while the desktop shell kept running.

The host now reports its `client` when it creates a session. The panel sends
it back with write/read/resize/close, and the server forwards it as
`clientKinds` so the gateway restricts those calls to that client. Gateways
that predate `clientKinds` ignore it, which leaves today's behavior unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(terminal): move an ordinary agent's terminal entry into the Overview

The header terminal button now shows only for coding agents (Claude Code,
Codex, …), which work in a shell. Other agents open the terminal from a
"Terminal" row next to Skills / Documents in the workspace Overview.
`useTerminalEntry` owns the placement and the "somewhere to run a shell" check.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…22844)

MiniMax speech models served through an OpenAI-compatible gateway (e.g.
OpenRouter `minimax/speech-2.8-hd`) were billed one character per code
point, while the upstream bills a Chinese character as 2. Share the
MiniMax counter from usageConverters/tts and use it in the
OpenAI-compatible speech path for MiniMax speech models.

Co-authored-by: rdmclin2 <4705237+rdmclin2@users.noreply.github.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
…ag (#22846)

* 💄 style(channel): show composer mentions as highlighted text, not a tag

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(channel): register the themed mention plugin last so the composer theme applies

<Editor> mounts its own ReactMentionPlugin after `plugins`, and a later
registration replaces the earlier config, so the theme passed through
withProps was dropped. Render the themed plugin as a child (registered last),
keep the @Label markdown writer, and outline a selected mention.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 💄 style(channel): tint a selected composer mention instead of outlining it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…ry is taken (#22855)

* ✨ feat(channel): move a branch into its own worktree when its directory is taken

A Channel reply that found another run in its member's device directory used to queue
as "waiting for this directory to be released". Now the coordinator asks the device to
cut `lobe-channel/<member>-<branch>` from that directory's HEAD under
`~/.lobehub/worktrees/`, records it per member and conversation branch, and claims the
job there. Later replies in the same branch keep the worktree; nothing merges back.
Directories that cannot host a worktree keep waiting and are retried after 10 minutes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(channel): keep non-git directories out of worktree creation

The device's addGitWorktree creates the target's parent folder before git runs, so a
directory outside git left an empty ~/.lobehub/worktrees/<dir>-<hash> behind each
attempt. List the directory's worktrees first: a repository always lists its own
checkout, and an existing channel checkout is reused without another add.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(channel): keep the subdirectory and environment in channel worktrees

A member configured on a repository subdirectory now works in the same subdirectory of
its worktree, and the checkout is keyed by the repository. The worktree branch name
carries a hash of the device and source directory, so moving the member to another
checkout of the same repository no longer collides with the branch the earlier
environment created (`git worktree add -b` refuses an existing branch).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(channel): place a worktree directory through the checkout's realpath

The configured directory is the device's realpath, but git can spell the checkout that
holds it differently (a symlinked repository root). When the lexical containment check
misses, ask the device for the checkout's realpath and retry before giving up.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(channel): keep waiting when the subdirectory is not in the worktree

A member configured on an untracked or ignored subdirectory got a worktree whose
matching path does not exist, so the run failed and later replies kept reusing that
assignment. Check the relocated directory on the device before returning a new or
reused checkout, and keep the waiting fallback when it is missing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…c and topic share info onto @lobechat/replica (#22849)

* ♻️ refactor(chat): move Agent Signal receipts, the document chat topic and topic share info onto @lobechat/replica

- agentSignalReceipts: new agentSignal store replica (per agent + topic,
  persisted, maxEntries 30); polling/backoff unchanged; the first read after
  a reload is a full head; a confirmed rollback is written onto the replica.
- agentDocumentChatTopic: joins the agentDocument store, keyed by
  (agentId, documentId), persists only the resolved topicId; seeded by the
  docs route loader.
- topicShareInfo: joins the chat topic slice; only an existing share record
  is persisted, writes land the server record on the replica.
- Remove agentSignalKeys.receipts, shareKeys.topicInfo and
  agentDocumentSWRKeys.documentChatTopic.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(chat): read the full receipt head when a topic is reopened

An incremental (sinceCreatedAt) read never returns a change to an older
receipt, so an Undo made in another tab kept its Undo action after
switching topics. Opening a topic (again) now reconciles the whole head;
polls stay incremental.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(chat): keep a share change in the identity that started it

A visibility change that resolved after a workspace or account switch
wrote (and persisted) the previous identity's share record into the new
scope. Capture the replica scope before the request and land the result
only when it is still active.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(chat): keep the full-head reconcile intent until it succeeds

A reopened topic whose full-head read failed lost its marker, so the
retry read only receipts newer than the head and missed an Undo made
elsewhere. Clear the marker only after a full head answered.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…22867)

* 💄 style(messages): list only agents and groups with a conversation

The IM message list gave every agent and group a row, so it read like
Contacts. A row now needs a direct Channel or a topic; pinned rows and the
open conversation stay, and a search still lists every agent so a new
conversation can start from it. The list waits for recents before deciding.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(messages): refresh recents on route change and retry failed recents

A first topic started in an idle agent never reached the conversation recents
while the list stayed mounted, so the row disappeared once the person moved on.
The list now revalidates recents whenever the routed conversation changes. A
recents failure with nothing cached showed an endless skeleton; it now shows an
inline error with a retry for that request.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…22868)

Reviewers preferred the earlier layout: the row says "typing •••" in
place of the preview instead of beside the name.

Reverts 1f79bd9 and d6e44b8.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Creating a project skipped its default General channel while the
`channel` runtime flag was off, so projects created before the flag was
turned on had no channel. The default channel is now opened regardless
of the flag; the direct channel of a new agent stays gated.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The agent and group conversation composers dropped the control bar under
`im_layout`, so normal agents lost device and working-directory selection.
Keep it in both layouts; the composer itself stays slim.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
#22859)

* ✨ feat(channel): sort sidebar channels by created time, activity or name

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* ✨ feat(channel): highlight unread channels and count @mentions in the sidebar

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(channel): keep the unread pill inside the row so the accordion does not clip it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* 🐛 fix(channel): clear the cached mention count when a Channel is read

`markRead` refreshed `unreadCount` optimistically but left `mentionCount` alone, so a
Channel read and then left re-showed its @mention badge until the idle list poll (up to
60s) recomputed it from the database.

Mentions are counted from the same read position as `unreadCount`, so a read that reaches
the known `sequence` consumes all of them. A partial read cannot be split client-side
without per-message mention data, so the cached count stays until the next poll.

Refs: Codex review thread on src/features/Channels/ChannelList.tsx:260

* 🚨 chore(channel): sort the systemStatus type imports

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…asoning reads (#22645)

A key the snapshot confirmed as nothing saved got no replica entry, so a later
IndexedDB hydration repainted an older row. A per-model read started before the
seed could also resolve after it and persist the older value as confirmed.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…patch (#22872)

* 🐛 fix(agent-runtime): settle operations whose first step fails to dispatch

When the first step cannot be scheduled, the caller sees an error, but the
operation row stayed running, so the stale-operation reaper later redrove it
and the request ran a second time after the user had already retried.

- Record a failed first-step dispatch as an error terminal state
  (QueueDispatchError) with a running-only conditional update before
  rethrowing the original error.
- Refuse to execute a step when the durable status cannot be read; the
  queue retries it instead of spending on models or tools in an unknown state.
- Cover dispatch failure, terminal preservation, concurrent reaper claims and
  redrive/abandon races.

* 🐛 fix(agent-runtime): keep unreadable durable status retryable after deferred init

The post-init status re-check runs inside the step's main try block, so a
transient read failure reached failStep and settled the run as an error.
Raise DurableStatusUnavailableError from the status read and let failStep
rethrow it without settling, like a lost step lock, so the queue retries.

* 🐛 fix(agent-runtime): keep intervention continuations retryable on dispatch failure

A prepared intervention continuation republishes the same deterministic
operation through its own recovery, so a failed first publish must leave it
running instead of settling it as QueueDispatchError.

* 🐛 fix(agent-runtime): retry refused durable reads for local queue deliveries

The local queue drops a callback that throws, so a step that refused to run
on an unreadable durable status was never retried. Retry it in the local
execution callback with the same budget as QStash's default redelivery;
other step failures are not retried.
…22877)

The contact card at /agents/:id now offers Message / View / More action
tiles and lists the handle, mail addresses and bio in one info list,
like an IM contact profile. Message opens the direct Channel when
Channels are on; View opens the Agent; More holds the profile entry and
Copy Agent ID.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…, and drag-and-drop reordering (#22866)

* ✨ feat(knowledge-base): support natural prefix sorting, sort controls, and drag-and-drop reordering

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* 🐛 fix(knowledge-base): move reorder transaction to KnowledgeRepo and align search/custom sort

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* ✨ feat(knowledge-base): add sidebar sort button and align multi-run/paginated ordering

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* 🐛 fix(knowledge-base): align Drive vs library folder grouping between server and client

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* 🐛 fix(knowledge-base): use compact move payload and batched SQL for sibling reorder

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* 🐛 fix(knowledge-base): disable sibling insertion in flat search results to preserve folder drops

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
# 🚀 LobeHub Release (20261112)

**Release Date:** November 12, 2026  
**Since v2.3.2:** 253 merged PRs · 252 commits · 12 contributors

> A week spent turning chat into a real workspace: direct Channels with
their own reply rules, an inbox identity for every agent, and projects
that can be created straight from a repository.

---

## ✨ Highlights

- **Direct Channels** — every agent and group now has an IM-style
conversation surface: stacked header avatars, typing in the header, card
composer with a `+` menu, day dividers, reply activity on list rows.
(#22559, #22583, #22556, #22440)
- **Per-agent reply strategy** — each Agent decides when to wake: own
rules first, follow-ups stay silent, unaddressed messages get one
primary responder, and the runtime records why a wake happened. (#22551,
#22454, #22847)
- **Agent identity & mail** — agents get an address, an inbox, and a
Profile page under an Identity group; mail-woken runs post what they
concluded back to the owner's Channel and only ever see their own
mailbox. (#22009, #22010, #22578, #22686)
- **Projects from GitHub** — create a project from a repository, pick
its work location, edit or remove working directories, and link tasks to
projects with picker, badges and grouping. (#21552, #22612, #22725,
#22383)
- **Knowledge base ordering** — natural prefix sorting, sort controls
and drag-and-drop reordering, plus a caller-scoped business knowledge
access seam. (#22866, #22863)
- **Desktop serves the remote terminal** — the desktop app can now serve
a remote terminal to the web and mobile clients, alongside local-file
preview for remote devices. (#22852, #22647)
- **New surfaces behind flags** — Lobe Bench workspace pages
(benchmarks, datasets, experiments), a meeting data model with TRPC API,
and versioned skill archives. (#22483, #21259, #22590)

---

## 🏗️ Channel & Conversations

- Sidebar channels sort with unread and @mention highlights, and a
project's General channel always opens. (#22859, #22870)
- Threads open as the page instead of a side panel; entity links render
and acceptance reports open beside the feed. (#22850, #22584)
- Composer mentions highlight as text rather than a tag, and the
composer reply bar softens while the feed fades into it. (#22846,
#22839)
- Agent profile cards appear on hover in message avatars and join event
lines. (#22515, #22836)

## 🤖 Agent Runtime & Identity

- Relayered model runtime: Router attempts run through one lifecycle and
stop on cancel, parallel tool-call arguments stay on their own call, and
a mid-stream socket drop is classified as `ProviderStreamInterrupted`.
(#22819, #22718, #22827)
- Anthropic, OpenAI and OpenAI-compatible code is organized by protocol,
with shared Router helpers extracted. (#22785, #22731, #22818)
- Gateway steps order by server time, step commits are refused after a
lost lock, and background sub-agent results are delivered again.
(#22842, #22524, #22530)
- Memory work stops for banned users, failed extractions record their
real cause, and long consolidation merges run in small batches. (#22691,
#22797, #22596)
- Claude Code sessions can be steered with queued follow-ups, OpenCode
2.x is supported, and Copilot / Cursor ACP run with full access by
default. (#22575, #22740, #22574)

## 📱 Projects, Tasks & Goals

- New projects default to a dedicated main agent, the sidebar collapses
projects into one link with favorites and sorting, and creating from
GitHub can also create that agent. (#22401, #22320, #22696)
- Code projects run in the sandbox inside the project's own repository.
(#22761)
- Scheduled tasks accept an exact minute; cross-workspace task links
open in the side panel; a redelivered completion settles a run once.
(#22406, #22473, #22637)
- Goal fixes throughout: gate-held rounds read as waiting for release,
planning-turn stalls close, and `goal:manage` is granted to runs the
user sent. (#22750, #22760, #22726)

## 🖥️ Desktop & Clients

- AppImage core OTA relaunches through a detached spawn, the titlebar
stays draggable over the tab strip, and dev instances are isolated and
labelled. (#22793, #22462, #22823)
- The task topic drawer header leaves the drag region, and the sidebar
agent header follows the tab's agent after Home. (#22831, #22825)
- `lh` gains `--visibility` / `--json` on agent create, a `mcp` alias,
meeting commands and CLI requirement editing for goals. (#22801, #22626,
#22593, #20568)

## 🔒 Security & Reliability

- **Security:** workflow callbacks now fail closed when unauthenticated;
a mail-woken run is scoped to its own mailbox; the SWR key registry is
frozen so new server data routes to local-replica. (#22864, #22686,
#22727)
- **Reliability:** page collaboration keeps one Yjs tree per page with
durable rooms; context engine snapshots are pooled so long-run traces
stop growing quadratically; file parsing no longer blocks the event
loop. (#22695, #22522, #22796)

## 🔧 Tooling & Maintenance

- StyleX migration continues with a style-equivalence verifier that
shows where a migrated descendant rule moved. (#22041, #22658)
- Test consolidation across modules, database coverage back to 100%, and
dead v8 fallbacks removed. (#22337, #22719, #21893)
- Retired the `new_tier_plans` flag, the unused agent stream endpoint
and the daily free video quota hint. (#22611, #22601, #22811)

---

## 👥 Contributors

Huge thanks to **12 contributors** who shipped **253 merged PRs** this
cycle.

@MarioJames · @R0BiN00Z · @IveTian · @arvinxx · @Innei · @tjx666 ·
@rdmclin2 · @ImSingee · @sudongyuer · @AmAzing129 · @canisminor1990 ·
@nekomeowww

Plus @lobehubbot, @lobehub-coder and renovate[bot] for maintenance.

---

**Full Changelog**: v2.3.2...release/weekly-20261112
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Labels

⤵️ pull merge-conflict Resolve conflicts manually

Projects

None yet

Development

Successfully merging this pull request may close these issues.