Skip to content

chore: release 1.0.3b6 - #2996

Merged
jubaoliang merged 30 commits into
mainfrom
release/1.0.3b6
Nov 15, 2026
Merged

jubaoliang merged 30 commits into
mainfrom
release/1.0.3b6

Conversation

@jubaoliang

Copy link
Copy Markdown
Collaborator

[1.0.3b6] - 2026-11-15

新增

  • 支持 Helm 集群部署。多实例共用一个 PostgreSQL
  • 管理员可以限制创建专家时能选的存储后端
  • 新增托管专家,对话可以交给后端 Agent,如 OpenCode
  • 知识库支持知识图谱和重排序
  • 新增本地模型供应商 Magpie
  • 定时任务可以跟最新对话,也可以固定到某一场历史对话
  • 对话里列出本轮改过的文件、所在目录和增删行数

修复

  • 编辑托管专家时,不显示存储后端也能保存

安全

  • 不能再把宿主机目录挂进 Docker 专家
  • 腾讯云登录按用户隔离

Release checklist

  • CI green
  • Merge this PR into main
  • After merge, GitHub Action auto-pushes v1.0.3b6 tag on main tip

Made with Cursor

ideshun and others added 30 commits October 23, 2026 10:56
….0.3b5

chore: sync main into develop after 1.0.3b5
feat(knowledge): rerank retrieved chunks with a provider's rerank model
…uth lapses and resume the turn (#2929)

A lapsed custom MCP token now fails the turn with a re-authorize action, then the original message is sent again after the provider confirms.

Co-authored-by: Cursor <cursoragent@cursor.com>
…2946)

File actions move to the card footer so the name keeps the full card width, including in a narrow window.

Co-authored-by: Cursor <cursoragent@cursor.com>
…se (#2947)

The move dialog lists writable knowledge bases the same way as the page, and files can move into a folder of another base.

Co-authored-by: Cursor <cursoragent@cursor.com>
…kends (#2953)

Move the image and Compose files under deploy/, and add a Kubernetes layout plus a two-process lab that share one database. Admins can also restrict which backend types a user may choose when creating an expert.
* feat(dashboard): add a welcome guide for new users

New users land on an empty chat page with no overview of what Octop
offers. Add a short step-by-step guide that opens on first login and
introduces conversations, experts, model configuration, channels,
knowledge bases and skill packages, with a button on each step that
opens the matching page.

- Store `onboarding_completed` per user in `preferences_json`, exposed
  through GET/PATCH /api/preferences (no schema change).
- Finishing, skipping, closing or jumping to a page marks it done, so
  it does not open again by itself; "Getting started" in the account
  menu reopens it.
- Steps follow the user's permissions; members without the models page
  are pointed at the chat model picker instead.
- zh/en copy under `onboarding.*`.

* feat(dashboard): walk new users through each page with a spotlight tour

Instead of describing pages in a single dialog, the guide now opens each
page in turn and spotlights its sidebar entry with antd Tour:

- Intro card with the mascot and the list of stops, staggered in.
- Each step navigates to its page, pulses the sidebar entry under the
  spotlight, and shows a short description, a tip and a progress bar.
  Arrow keys move between steps; Esc exits.
- A wrap-up card with quick links to chat or create an expert.
- On mobile the nav lives in a closed drawer, so the card is centered.
- Animations are disabled under prefers-reduced-motion.

Sidebar items expose `data-nav-key` so the tour can find its targets.

---------

Co-authored-by: Alison <20137720+alison-xx@users.noreply.github.com>
… to 4096MB (#2948)

Updated the Dockerfile and GitHub Actions workflow to raise the memory limit from 2048MB to 4096MB during the dashboard build process. This change addresses out-of-memory issues encountered on GitHub runners, ensuring successful builds without interruptions.
…#2961)

Updated the electron-builder configuration to clarify the use of full-bleed opaque icns and replaced the icon asset in resources to improve visual quality. This change ensures better compatibility with macOS Finder display requirements.
…clean

feat(telegram): 基于 Gateway 集成配置与原生问答
The opening copy described the product, and the title sat too far below the mascot.

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
… chat (#2970)

Open expert, team, and cloud creation from the chat sidebar, and order
experts by the viewer's own latest thread. The composer chip shows only
the model name; conversation history still keeps the provider.
Replace the hand-written Kubernetes manifests with one chart, and push it to GHCR on version tags.

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Add a config.json editor on app settings, show last login, deliver team tool media to IM, and keep an in-flight chat turn after refresh. Admins can manage experts they do not own.

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Strict providers reject a system message that lands mid-thread, and a member's first dispatch used to drop the briefing. Expert create and edit drawers also keep the basic fields up front and fold the rest into More configuration.
…in retrieval (#2927)

* feat(knowledge): rerank retrieved chunks with a provider's rerank model

Knowledge retrieval returns the k nearest vectors as they are. Nearest
is not the same as most relevant, and Octop already recognises rerank
models on a provider (`is_rerank_model`) without ever calling one
(#2410, #1099).

With a rerank model selected in the knowledge settings, retrieval now
fetches a wider candidate set (4×k, between 20 and 50), sends the query
and the candidate texts to the provider's `/rerank` endpoint (the Jina /
Cohere request shape, also served by SiliconFlow, vLLM, Xinference and
TEI), and keeps the reranker's top k. If the request fails, retrieval
logs it and falls back to the vector order, so a flaky reranker cannot
break a chat turn.

- `infra/knowledge/rerank.py`: settings, readiness, the request and the
  response parsing (`results`, `output.results`, bare list).
- `GET /knowledge-bases/rerank-options` lists enabled providers that
  offer a rerank model; the capability payload gains `rerank`; `PUT
  /knowledge-bases/feature` accepts `rerank_enabled` /
  `rerank_provider_id` / `rerank_model` and refuses a model that is not
  a reranker (KNOWLEDGE_RERANK_NOT_READY) before saving anything.
- Dashboard: a "Rerank retrieved passages" section in the knowledge
  settings. Changing it does not reindex.

Off by default; nothing changes for existing installs.

* feat(knowledge): build an entity graph per knowledge base and use it in retrieval

A knowledge base can now build a knowledge graph. Each indexed passage is
sent once to a chat model that returns entities and relations as JSON;
entities are merged by name and stored next to the vector index (a
graph.sqlite sidecar, or public.knowledge_graph_* tables on PostgreSQL).

Retrieval finds the entities named in the question by name, appends their
relations to the context and gives up to a quarter of the passages to
linked passages the vector search missed. The dashboard shows the graph as
a node-link view with entity search, relations and source passages.

Building is opt-in per knowledge base and needs the extraction model chosen
in the knowledge settings. Off by default.

Refs #1099

* fix(knowledge): lay out the graph, rerank and OCR settings like the rest of the drawer

The knowledge-graph section showed its switch without the card the
'enable knowledge bases' switch uses, and two unlabeled dropdowns whose
values can read the same (provider and model). It now uses the card for the
switch, an 'Extraction model' label, the provider dropdown and the
selectable model list used for the embedding model. The OCR and rerank
switch rows get the same card so every switch in the drawer matches.

---------

Co-authored-by: Alison <20137720+alison-xx@users.noreply.github.com>
Co-authored-by: Jubaoliang <jubaoliang@gmail.com>
…xpert files (#2969)

Creating an expert from the market builds file_overrides with
diffComposerFiles(originalFileContents, fileContents), but the market
drawer had no file baseline while ensurePromptFiles injected an empty
AGENTS.md editor placeholder into fileContents. The untouched stub
diffed as a user edit, so the create request shipped
{name: "AGENTS.md", content: ""} and overwrote the template-seeded
AGENTS.md with an empty file — SOUL.md and friends survived, and the
harness template fallback then skipped AGENTS.md because it "existed".

An empty file without a baseline counterpart is now treated as an
editor placeholder rather than an override, and both drawer branches
baseline the placeholders alongside the fetched file contents.

Fixes #2968

Co-authored-by: Jubaoliang <jubaoliang@gmail.com>
…2986)

Jobs that do not start a fresh thread can follow the channel's newest
conversation or stay on a chosen thread. The tasks page, CLI, and cronjob
tools share that choice.

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
#2988)

Title and conversation stay one full-height column, with a corner control to open the dock. Secondary composer actions and duplicate float icons hide when that column is narrow, and the knowledge graph is marked beta.
Show each file a turn wrote or edited, with its parent directory and the net line delta, and open that file from the chip.
* feat(providers): add Magpie as a local model provider

Discover the gateway on the Octop host, import its chat catalog with capabilities, and attribute each turn with the thread id.

* fix(providers): keep Magpie refresh tied to fetched connection

* feat(dashboard): check Magpie from its drawer and brand iFlytek

Move the server probe into the Magpie configuration drawer so it runs only when asked, and show iFlytek Astron and Magpie with their own marks.
…ers as experts (#2992)

* feat: add expert compaction and runtime limits, and register ACP runners as experts

Each expert can tune automatic context compaction, a per-turn tool-call cap, and model or tool retries. A managed expert keeps the normal agent surface and forwards every turn to a connected ACP runner.

* fix(dashboard): hide context usage for managed experts

A managed expert does not run a local model, so the chat composer should not show a context-window ring.
* fix: block docker host mounts and isolate cloud logins

Drop user-supplied docker volumes before they reach the daemon, and keep
Tencent Cloud credentials on a per-user profile behind a cloud permission.
Depend on published octop-harness 1.0.10 and octop-gateway 1.0.3 instead of
the draft git pins.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: satisfy mypy on docker spec and cloud import
* fix: finish Cursor turns when several approvals overlap

Cursor can ask for two permissions in one ACP turn. The second request
replaced the only permission future, so the first approval never came
back and the chat stopped on the preamble. Queue those requests, and
keep the original tool title when a later update omits it.

Co-authored-by: Cursor <cursoragent@cursor.com>

* docs: note the Cursor approval queue fix in the changelog

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
…2995)

The edit form hides the backend picker, so save was calling startsWith on an unregistered backend_choice and failed for welcome text and other fields.
Co-authored-by: Cursor <cursoragent@cursor.com>
if not name:
return None
if os.path.isabs(name) or "/" in name or "\\" in name:
return name if os.path.isfile(name) else None
@jubaoliang
jubaoliang merged commit bf134db into main Nov 15, 2026
15 of 16 checks passed
@jubaoliang
jubaoliang deleted the release/1.0.3b6 branch November 15, 2026 15:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

9 participants