Turn the lightpanda-backed fetch service on, bump worker concurrency to
3, and add baidu.com plus weixin.sogou.com to the allowed domains so the
knowledge URL fallback can resolve those sources without a config edit.
Postgres rejects FOR UPDATE on the nullable side of a LEFT JOIN, which
made LeaseNextQueuedDesktopTask fail under contention. Replace the join
with a NOT EXISTS subquery, scope the row lock to the dt alias, and add
a regression test plus warn-level logging on lease query failures so
similar issues are easier to diagnose.
Map dongchedi adapter failures (raw "server exception" or
dongchedi_platform_exception) to a Chinese prompt steering the user to
the dongchedi backend, so renderer/admin views and desktop-task error
payloads no longer surface the opaque platform message.
Hoists hasActiveGenerationStatus into the shared display helper so all
views agree on what counts as in-flight generation, shows the backend
generation_error_message inline on the imitation list and detail drawer
when status is failed, and drops the workspace/templates/imitation poll
interval from 10s to 3s (with a 5s fallback when streaming is enabled)
plus explicit query-cache invalidation so failed/completed states show
up promptly.
The state-consistency worker now detects articles still flagged as
generating whose latest generation_task is already failed or completed
(with a version) and aligns the article status to the terminal task,
invalidating article caches on the way out. Adds supporting partial
indexes and switches UpdateGenerationTaskStatus to COALESCE started_at
and completed_at so partial status updates don't wipe the timestamps.
Adds a new browser-fetch microservice that wraps Lightpanda for rendering
JS-heavy pages, and wires it into the knowledge URL parser as a fallback
when the upstream link reader returns 403/429/5xx for an allow-listed
domain (e.g. zhuanlan.zhihu.com). Includes config/env plumbing, hot-reload
diff support, and full deploy assets (Dockerfile target, docker-compose,
k3s manifests, image-build/package scripts).
Playwright serializes page.evaluate callbacks via Function.prototype.toString and runs them in the target page context. Obfuscating the main bundle injected string-array helper references that the browser page could not resolve, surfacing as `ReferenceError: wO is not defined` for bilibili and dongchedi publishes. Mirrors the existing Vite `minify: false` guard for main/preload.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Detect the passport.qianwen.com bridge page after a successful SSO login
and navigate the webContents back to the original returnUrl so account
binding can proceed instead of stalling on the bridge.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
setName pins the runtime app name (menu bar, task switcher) and
setAboutPanelOptions overrides the macOS About panel which is sourced
separately from Info.plist when running unpackaged. Packaged builds are
unaffected: electron-builder productName "省心推" still drives the .app
bundle and NSIS installer metadata. The macOS Dock label in dev still
shows "Electron" because that comes from the unpacked Electron.app
bundle's CFBundleName, which is outside Electron's API surface.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Baijiahao throws baijiahao_challenge_required when its risk control flags
the network environment, but PublishManagementView previously showed the
raw "code: ..., message: ..." string. Route the error through the existing
normalizer pipeline so users see a clear "需要人机验证 + 去后台手动完成
一次发稿" prompt instead.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Playwright serializes the function passed to page.evaluate via .toString()
and runs it in the browser context. With main-process minify on, esbuild
renames inlined helpers (e.g. __name) and closure references to two-letter
identifiers like FK/p/ed; their definitions stay at module scope on the
Node side, so the browser eval throws ReferenceError. Mac dev builds were
unaffected because dev mode doesn't minify; only packaged Windows builds
hit this. Main process bundle size is irrelevant (loaded once on startup)
so turning minify off is the cheapest correct fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Toutiao's wap_login redirect chain raises ERR_ABORTED (-3) without the
literal symbol in the message, so the previous regex fell through to the
fallback error page. Now match the numeric form too, and add a 7s blank-
page watchdog that auto-reloads stuck "Loading..." popups once per fresh
navigation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
K2.6 introduced new UI surfaces that polluted the answer extraction and
the busy-signal probe, breaking monitor runs even though Kimi rendered a
correct answer:
- collectCandidates now drops elements that contain a chat composer child
(<textarea>, non-hidden <input>, [contenteditable]). This filters
wrapper containers like .chat-detail-content that the wide
[class*="content"] fallback would otherwise pick, pulling editor
placeholder text ("可快捷使用技能", "Kimi K2.6 已就位") into the answer.
- isAuxiliaryPanelElement skips elements with role="dialog" /
"alertdialog", aria-modal="true", or the <dialog> tag, so the K2.6
launch-announcement popover ("Kimi K2.6, 已就位 / 一键部署 OpenClaw")
no longer counts as answer content.
- Drop the class-based loading-indicator probe entirely. The previous
global [class*="loading"]/"stream"/"typing"/"spinner" sweep matched
K2.6's always-on text-stream wrapper and held busy=true forever, which
caused kimi_query_timeout even after the assistant turn had fully
rendered. The textual busySignalsPattern (思考中/搜索中/生成中…) is
the actual semantic signal and survives any UI rename.
- Surface a capacityNotice from the assistant turn when Kimi shows its
short capacity-exhausted message ("算力不够 / 请稍后再试"). The top
level falls back to it when answerText is empty so the SaaS pipeline
records the user-visible message instead of an unknown row, with
capacity_limited / capacity_notice flagged in raw_response_json.
All filters lean on HTML form / ARIA semantics and visible text rather
than Kimi-specific BEM classes, so they survive future redesigns.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- findInteractiveByText now also accepts elements carrying business
attributes [dt-button-id]/[data-button-id]. The new yuanbao toolbar
renders the deep-think toggle as <div dt-button-id="deep_think"> rather
than a <button>, so the previous text-only locator missed it and probes
silently ran in the fast model — without inline citations.
- Lower <div data-idx-list="1,2,10"> placeholders to bare "[1][2][10]"
(previously "[citation:1] [citation:2] [citation:10]") and accept both
forms in YUANBAO_CITATION_MARKER_PATTERN. Final answer text runs through
normalizeCitationMarkers so any residual SSE [citation:N] is rewritten
to [N], matching DeepSeek's chip rendering on the SaaS side.
- searchGuid frames in the new protocol carry every reference inside
docs[] (each with a 1-based index that mirrors the answer markers) and
citations is now always null. Promote docs into the citations bucket so
citations[index-1] aligns with the inline marker, eliminating the
spurious yuanbao_incomplete_citations failures.
- Update the SaaS detail view: createYuanbaoCitationPattern accepts both
"[N]" and "[citation:N]" so yuanbao answers render the same clickable
superscript chips as DeepSeek.
Tests cover the new searchGuid.docs alignment, fast-cache answers (no
markers, no sources -> still succeed), legacy [citation:N] -> [N]
rewriting, and mixed-format marker extraction.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Refactor readDeepseekAccountIdentity so the JSON payload parser is a pure
ts function (extractDeepseekIdentityFromPayload). The webContents IIFE now
just polls for userToken and returns the raw API JSON; payload picking lives
in ts and is unit-tested.
Tests lock the contract for both login types we observed in the wild:
- WeChat-bound account: id_profile.name = '阿白', picture present.
- Phone-only login: id_profile = null, mobile_number = '177******08',
email = '' (empty string, not null) — the case that originally regressed.
Plus regressions for the priority order (profile.name > mobile_number > email),
the all-empty fallback, and non-record payloads.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phone-number logins return id_profile=null, mobile_number="177******08", and
email="" (empty string, not null). The previous fallback used `??`, which only
short-circuits on null/undefined, so an empty email string clobbered the
mobile_number fallback and the displayName ended up null — leaving the
DeepSeek card stuck on the "${label} 会话" placeholder.
Switch to a truthy pickString helper that requires a non-empty trimmed string
before accepting a candidate, and prefer mobile_number over email so the
masked phone number shows up exactly the way DeepSeek's own sidebar renders it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The generic AI page-state heuristic only catches displayName via DOM patterns
like [class*="user-name"], so wenxin/yuanbao/deepseek (CSS-module hashed
classes or no exposed store) fell back to the placeholder "${label} 会话".
Add per-platform readers under adapters/{wenxin,yuanbao,deepseek}/account-identity.ts:
- wenxin: GET /eb/user/info with BDUSS cookie; fall back to __NEXT_REDUX_STORE__
and sidebar DOM if the API fails or hasn't responded yet.
- yuanbao: poll the stable BEM .yb-nav__user .nick-info-name + avatar img,
since yuanbao exposes no user-info API and the sidebar renders late.
- deepseek: poll localStorage for userToken, then call /api/v0/users/current
with Authorization: Bearer <token>; reads biz_data.id_profile.{name,picture}.
detectWenxin/Yuanbao/DeepSeekPlatform now run the credential-backed detector
first, then override the placeholder displayName/avatar via the per-platform
reader. Existing rows in DB still hold the placeholder until the account is
re-bound — probe does not write back display_name yet.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The AI 平台 badge counted distinct platforms with at least one
active-auth account, while 媒体账号 counted raw account rows
regardless of auth state. The two numbers shared a slot but meant
different things, which made them not comparable at a glance.
Switch the AI 平台 count to mirror 媒体账号: account-level count
of accounts whose platform falls in the monitoring catalog, with
no authState filter.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Business calls (heartbeat, account sync, lease/pull/resume tasks,
preempt monitoring lease) used to short-circuit any 401 into
handleAuthExpired, which tore down the runtime and switched the
window to login. That stole the screen on transient or per-request
auth glitches and conflicted with the renderer-side proactive token
renewal.
Now business 401s fall through to the existing warn/danger activity
log paths and runtime keeps running. The login redirect is owned
solely by the renderer's renewAuthenticatedSession → on refresh /
rotate failure → forceLogoutAfterRenewFailure path. Server-side
revocation via error code 40991 is unaffected.
Removed the now-orphan plumbing: handleAuthExpired,
isApiClientError helper, emitRuntimeAuthExpired /
onRuntimeAuthExpired events, and forceLoginWindowForAuthExpired
listener registration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Drop the broad path filters that triggered the desktop client build on
unrelated workspace changes, and add a swagger description for the
existing article regenerate endpoint.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Poll the knowledge item list every 5s while any item is still pending
or processing so the UI reflects ingestion progress without manual
refresh, and exclude items that have no active chunks from group
counters so groups with no usable content do not appear populated.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Intercept route changes and tab close when the template wizard has
unsubmitted content so users can save a draft, discard, or stay instead
of silently losing their progress.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Expose tenant-authenticated publish task list and retry endpoints so the
admin web can show desktop publish state without an Electron bridge, and
register a `shengxintui://` deep-link so the page can hand off to the
desktop client workbench.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mark 401/refresh-failure errors as handled so per-view catch blocks fall
through silently while a single global "登录已过期" warning is shown.
Centralize ops-web error rendering via showOpsError.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Remove the 1400px max-width cap on Home/Accounts/AiPlatforms/PublishManagement
so content stretches with the window, and clean up the unused online-clients
badge from the settings sidebar.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Emit a 'publish-task-lease' runtime invalidation event whenever the
scheduler activates a publish task, and have PublishManagementView
listen for it to trigger an immediate refresh. The view also switches
to a 5s active-task poll only while a task is in_progress, replacing
the unconditional 20s timer.
Move the Qwen page-state probe, session detection, and silent probe
helpers out of account-binder into apps/desktop-client/src/main/adapters/qwen,
mirroring the Kimi adapter layout. Generic AI auth no longer special-cases
qwen since the platform now owns its own credential rules.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
root
Xu Liang