Skip to content

chore: release 1.0.3b3 - #2815

Merged
jubaoliang merged 33 commits into
mainfrom
release/1.0.3b3
Nov 4, 2026
Merged

jubaoliang merged 33 commits into
mainfrom
release/1.0.3b3

Conversation

@jubaoliang

Copy link
Copy Markdown
Collaborator

[1.0.3b3] - 2026-11-04

新增

  • PostgreSQL 可存放技能包和知识库。多实例共用定时任务、频道、头像、插件和已发布专家
  • Docker 与飞牛增加精简版。存储后端增加火山引擎 TOS
  • 可预建 LDAP 账号,聊天列表可隐藏任意专家

修复

  • 专家工作区落到允许的存储根下。共享连接器只有创建者能改
  • 聊天历史不再错序,未答完的提问不会被别的会话接走
  • 超长图片先缩小再发送。失效的默认模型会自动换掉
  • 升级记忆库:长会话分批提炼,重启后不重跑
  • 安装清理不再删掉运行环境。PostgreSQL 上的知识库检索和关怀去重恢复正常

变更

  • 升级 octop-harness 至 1.0.7
  • 内置专家并入专家市场,安装包不再自带市场目录

Release checklist

  • CI green
  • Merge this PR into main
  • After merge, GitHub Action auto-pushes v1.0.3b3 tag on main tip

Made with Cursor

weixin_58011598 and others added 30 commits October 15, 2026 11:23
….0.3b2

chore: sync main into develop after 1.0.3b2
_psycopg Connection objects put executemany on the cursor, not the
connection, so the sqlite-shaped batch insert in care_push.insert fell
through _PgConnectionProxy.__getattr__ and raised AttributeError on a
PostgreSQL control plane, leaving care_push_records empty and the
30-day dedup dead. Route executemany through a cursor and convert
? placeholders like execute already does.

Co-authored-by: wxhking <186391338+wxhking@users.noreply.github.com>
fix(docker): publish a lite image beside the full one
…2394)

* fix(agents): rebuild AUTO agents on any provider change to refresh baked fallbacks

HarnessAgentConfig.default_model is resolved once at build time (explicit
agent default, else active model, else first usable) and baked into the
harness config, because the memory extraction/promotion aux LLM needs a
concrete ref. The selective reload impact set only rebuilt AUTO agents
when the changed provider was still the active one, so disabling a
provider that owned the baked ref left those agents calling the removed
provider from background paths forever (ValueError: Unknown provider on
every session end) (#2388).

Any provider change can strand the baked ref (first-usable ordering is
landscape-derived, not active-derived), so AUTO agents are now always in
the impact set; agents with an explicit model ref to another provider
are still skipped.

Verified locally (Python 3.13 venv): new red test flipped green; full
tests/unit/agents 832 passed / 2 failures reproduced on unmodified
develop (onnx tqdm progress, skillhub page url - pre-existing); ruff
check+format clean on the three touched files.

Fixes #2388

Co-authored-by: Jay Feng <jayfeng77@outlook.com>

* fix(agents): narrow provider reloads to AUTO agents with stale baked refs

Review follow-up on #2394: rebuilding every AUTO agent on any provider
change is correct but broad — unrelated provider edits would rebuild all
running AUTO agents (bounded-parallel) while the admin request awaits.

The built ref is reachable: HarnessAgentManager.list_agents() exposes
each AgentEntry's config, so the impact set now rebuilds an AUTO agent
only when its baked ref points at the changed provider or no longer
matches what a build would resolve now. Active-model changes keep
rebuilding all AUTO agents; explicit refs to other providers are still
untouched. tests/support harness mock wires list_agents, and the impact
tests now seed baked refs — including the cheap-unrelated-edit case the
narrow rule must skip.

Verified locally: 65 passed (both suites), tests/unit/agents 814 passed
/ 7 skipped / 2 pre-existing failures reproduced on develop, ruff
check+format clean.

Co-authored-by: Jay Feng <jayfeng77@outlook.com>

* fix(agents): rebuild AUTO agents that baked no model once one appears

Review follow-up on #2394 (@balabommablock-cpu): an AUTO agent built while
no model was usable bakes `default_model=None`, and `_stale_auto_default_ids`
skipped it entirely, so when a provider later became usable the agent kept
running without a model until the process restarted.

Treat "nothing baked, but a fallback exists now" as stale; with no fallback
available there is still nothing to rebuild towards, so it stays untouched.

Verified locally: test_provider_reload_impact.py 7 passed,
test_agent_registry.py 60 passed, ruff check + format clean.

---------

Co-authored-by: Jay Feng <jayfeng77@outlook.com>
Co-authored-by: 薄生 <67680641+Bosheng0422@users.noreply.github.com>
…ity (#2745)

The terminal-prompt mark read as a command cursor. Use a monitor in steel blue, and give authenticator, passkey, and recovery the same alignment and row spacing as change password.

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
The lite publish workflow replaced the ARM image missing warning, so the arm64 publish guard now checks the missing-CPU message.
…#2757)

* feat: keep skill packages and knowledge on the control-plane database

PostgreSQL stores package files, knowledge originals, and vectors beside the
control-plane tables. Setup can keep an existing PostgreSQL account, search
reads that index, and the library can show slices and rename files.
* fix(db): stop fsyncing every SQLite write on Windows

WAL was left on SQLite's default synchronous=FULL, so each autocommit
called FlushFileBuffers. That made the Windows suite several times slower
than Linux. NORMAL is the recommended WAL setting; the history archive
still opts into FULL.

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

* docs: note the Windows SQLite sync change in the changelog

The unreleased fix list should say control-plane writes no longer fsync
every commit, while the history archive stays durable across power loss.

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

* chore: drop the changelog note from the SQLite sync fix

* fix(db): adjust SQLite synchronous settings for Windows

Updated the SQLite connection settings to ensure that the synchronous mode is set to NORMAL only on Windows, while other platforms retain the default FULL setting. This change improves performance on Windows without compromising data integrity during power loss. Updated related test to reflect this behavior.

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
… at pack time (#2751)

* fix: name the x86 FnOS native package with amd64

The x86 package was still Octop-fnos-native while the ARM package already carries arm64. Put amd64 in the filename, workflow artifact, and download links.

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

* feat: enhance workflow for temporary package builds

Updated the GitHub Actions workflow to support temporary package builds. Introduced a `temporary` input option that allows users to create packages from the current branch without pushing Docker images, uploading to COS, or attaching to a GitHub Release. Adjusted job dependencies and conditions to accommodate this new feature, ensuring that temporary builds are clearly marked and managed separately from standard builds. Updated README to reflect changes in the build process and clarify the temporary package functionality.

* chore: update dependencies and version placeholders in manifests

Updated the Pillow dependency version in pyproject.toml to 13.0.0. Bumped the fastembed package version to 0.9.1 in uv.lock and adjusted the corresponding source URLs. Changed version placeholders in the Docker and native manifests to 0.0.0, ensuring consistency across files. Updated related documentation to reflect these changes and clarified version mapping rules.

* refactor: streamline GitHub Actions workflow and update README

Refactored the GitHub Actions workflow for building Octop FPK to enhance clarity and efficiency. Removed unnecessary conditions related to temporary package builds and updated job dependencies. Revised the README to reflect changes in the build process, emphasizing the reuse of Docker images and the handling of package uploads. Improved documentation for better understanding of the workflow mechanics.

* test: match amd64 native packages against the amd64 host token

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(dashboard): serve Monaco from the page origin

@monaco-editor/loader otherwise downloads min/vs from jsDelivr, which hangs on a LAN with no outbound network.

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

* feat(dashboard): integrate i18n for loading messages in CodeEditor and FileEditModal

Added translation support for loading messages in both CodeEditor and FileEditModal components using the react-i18next library. This enhances user experience by providing localized loading indicators.

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(market): load plugins and experts from the remote market

Serve catalogs from the configured market CDN with versioned installs,
private-market setup, and network errors instead of a bundled fallback.

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

* refactor(market): update environment variable names for portal and market origins

Changed `OCTOP_PORTAL_URL` to `OCTOP_PORTAL_ORIGIN` and `OCTOP_MARKET_REMOTE_URL` to `OCTOP_MARKET_REMOTE_ORIGIN` across documentation, code, and tests to improve clarity and consistency. This update ensures that the environment variables accurately reflect their purpose in configuring the plugin and expert catalog origins.

* feat(skill-market): introduce SKILL_MARKET_ORIGIN for configurable skill market source

Added support for the `SKILL_MARKET_ORIGIN` environment variable to configure the skill market source, allowing users to set a custom endpoint for skill searches and downloads. Updated documentation to reflect this change and included a new guide for setting up a private skill market. Enhanced error handling for skill market requests to provide appropriate HTTP status codes based on failure types.

* refactor(skill-market): remove CLI fallback for skill market requests

Eliminated the fallback to the local `skillhub` command line for skill market requests when HTTP failures occur. Updated related documentation and tests to reflect this change. Enhanced error handling to ensure appropriate HTTP status codes are returned for skill market errors.

* feat(expert-market): integrate user-published templates into the expert market

Merged the built-in expert templates into the expert market, removing the separate listing for built-in experts. Updated related UI components and localization files to reflect this change, ensuring users can now access both user-published and market templates seamlessly. Adjusted the Experts page and documentation to clarify the new structure and functionality.

* refactor(expert-market): streamline expert template management and UI updates

Removed built-in expert templates from the installation package, consolidating all templates within the expert market. Updated the user interface to reflect these changes, ensuring that new users pull templates directly from the market. Enhanced the plugin market functionality by removing fallback displays for unavailable plugins and adjusting the refresh button. Documentation has been revised to clarify the new structure and installation processes for expert templates and plugins.

* feat(expert-market): update localization and enhance UI components

Modified localization files for English and Chinese to simplify the "Create from Market" label. Enhanced the UI in the Experts page by adding a loading skeleton for dark mode and adjusting the layout of the Published Expert Card. Improved the search functionality and loading states for better user experience.

* feat(expert-market): add skill count to published experts and enhance UI

Introduced a new optional field `skill_count` to the `PublishedExpert` interface to track the number of skills associated with each expert. Updated localization files for English and Chinese to include new labels and improved existing ones. Enhanced the UI components in the Experts page, including adjustments to layout and styling for better user experience. Removed unnecessary management logic from the Published Expert Card component.

* refactor(catalog): simplify skill count calculation in expert directory

Condensed the skill count calculation logic in the `skill_count_in_expert_dir` function for improved readability and efficiency. This change streamlines the code while maintaining functionality.

* feat(expert-market): add framer-motion and motion libraries, update localization and UI components

Integrated the framer-motion and motion libraries into the project for enhanced animations. Updated localization files for English and Chinese to reflect changes in terminology, including "Published" to "Shared" and "Expert Market" to "Discover Experts." Enhanced the UI components in the Experts page, introducing new styles for better user interaction and experience.

* feat(pill-scroller): implement PillScroller component for enhanced navigation

Added a new PillScroller component to facilitate horizontal scrolling of options in the expert market. This component includes customizable scroll arrows and integrates with existing UI styles for improved user interaction. Updated the PluginMarketPanel and ExpertMarketTab to utilize the new PillScroller, enhancing the overall navigation experience. Additionally, removed redundant styles and components related to previous scrolling implementations.

* fix(expert-market): update button labels and styles for market source tabs

Modified the ExpertMarketTab component to switch the source selection logic between "Published" and "Remote," updating the corresponding button labels for clarity. Adjusted styles in the index.module.less file to enhance the appearance of the market source row, including a font size increase for better readability. Updated the PublishedExpertCard component to reflect the new label for the installation button.

* feat(market): update download URL requirements and enhance documentation

Modified the requirements for the `download_url` in the plugin and expert markets to allow any path on the configured `OCTOP_MARKET_REMOTE_ORIGIN`. Updated related documentation to clarify the new rules for URL structure and improved the handling of archive URLs in the codebase. Added tests to ensure the new behavior is correctly implemented.

* feat(skill-market): introduce SKILLSET_ORIGIN for skillset management and update documentation

Added the SKILLSET_ORIGIN variable to manage skillset details and imports, allowing for a customizable source for skillset metadata. Updated related documentation to clarify the usage of SKILLSET_ORIGIN alongside SKILL_MARKET_ORIGIN. Enhanced error handling in the skillhub_market module to validate the skillset origin. Adjusted tests to ensure correct functionality for fetching skillsets from the new origin.

* refactor(supplements): rename and update supplement handling to install skills

Renamed the supplement handling functions to clarify their purpose, changing `append_supplements` to `install_supplements`. Updated the logic to install packaged skills directly into the workspace instead of appending to `AGENTS.md`. Adjusted related tests and documentation to reflect these changes, ensuring a clearer understanding of the installation process for workspace skills.

* feat(changelog): update changelog with new features for plugin and skill markets

Added details about the ability to install plugins and experts from remote markets, including customizable source configurations for both plugin and skill markets. Updated documentation references for private market setups and clarified changes in installation processes for new users. Enhanced the overall structure of the changelog to reflect these updates.

* refactor(plugins): improve code readability in plugins API installation URL construction

Refactored the URL construction logic in the plugins API to enhance readability by formatting the template string. This change maintains the existing functionality while making the code cleaner. Additionally, simplified the props destructuring in the FadeInImg component for better clarity.

* fix(tests): ensure newline consistency in supplement test files

Updated the test for installing supplements to explicitly set newline characters when writing to test files. This change ensures consistent handling of line endings across different environments, improving test reliability.

* feat(experts): add AI Coding Coach and AI Safety Guardian with comprehensive skills and prompts

Introduced two new expert agents: AI Coding Coach and AI Safety Guardian. The AI Coding Coach provides techniques for effective AI programming, including context management, task decomposition, and debugging strategies. The AI Safety Guardian focuses on AI tool safety and compliance, offering checks for dangerous commands, data export risks, and compliance audits. Both agents include detailed manifest files, prompt files, and skills documentation to enhance user interaction and support.

* feat(marketplace): introduce new expert agents and update documentation

Added the AI Coding Coach and AI Safety Guardian expert agents, each equipped with comprehensive skills and prompts for effective AI programming and safety compliance. Updated the marketplace documentation to include new README files, manifest files, and detailed skill descriptions. Adjusted .dockerignore and .gitignore to reflect changes in directory structure, ensuring proper handling of the new marketplace catalog.

* feat(marketplace): remove outdated expert avatars and add new academic gaokao expert

Deleted obsolete SVG files for academic, lifestyle, and mysticism avatars. Introduced a new expert agent for academic gaokao advising, complete with a manifest file, skill documentation, and various supporting resources. Updated the AvatarPicker presets to reflect the removal of the old avatars and included new skills for enhanced user guidance in academic decision-making.

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
The overflow menu can hide owned experts as well as shared ones. A hidden expert stays in the hidden group while you chat with it, and the disclosure label stays centered.
Multiple processes sharing one control-plane database were both firing
cron jobs and connecting the same IM channels. Claim a same-row lease
before either, and keep CLI sessions from taking over background work.

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Experts and plugins now live in the OctopMarketplace repository. Octop does not load this tree at startup and does not pack it into the wheel.
)

* fix(gateway): keep ask_user_question answers on the paused thread

Dashboard sessions share one session_key per agent+user, so a new chat
message was auto-routed into an open ask pause on another thread.
Require an inbound thread_id match before treating text as the answer.

Also rename backendFlag.ts to backendFlagModel.ts so it no longer
collides with BackendFlag.tsx on case-insensitive filesystems.

* docs: add PR #2779 local verification screenshot
)

Saving LDAP settings crashed when group search was off because unmounted fields were undefined. Admins can now pre-create a passwordless directory account, and literal private addresses may use ldap:// without StartTLS.

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Conditional classes were built with template strings and array joins, so
empty tokens and nested ternaries were easy to get wrong. cx() drops
falsy values in one call per element.
WeChat long screenshots can stay under the 2 MB cap while exceeding
provider side limits, then get replayed from history. Fit those images
to 8000 px on the outgoing request so later text turns keep working.

Co-authored-by: Cursor <cursoragent@cursor.com>
fix(gateway): scale long vision images before the provider call
…OS (#2793)

init --force and clean --all were deleting ~/.octop/venv with the data, which bricks the official installer CLI. The pinned harness still cannot create, delete, or move OpenSandbox paths, so bootstrap and skill removal fail. Storage also needs a Volcengine TOS backend.

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
…2792)

* feat: share avatars, plugins, and published experts across instances

Keep those control-plane file groups in the database so a second Octop process follows the shared copy instead of its own local directory.

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

* test: write plugin fixtures as bytes so Windows keeps LF

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

---------

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Replaced outdated URLs pointing to Tencent Cloud with new URLs from the official Octop download site across various documentation files and code snippets, ensuring users have access to the latest installation scripts.
* fix(chat): reconcile cached turns with canonical history

* fix(chat): narrow history ordering fix to frontend
Readers can find 腾讯Octop in WeChat from the README, alongside the existing WeCom customer group.

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix: let only the connector owner edit a shared instance

Sharing still lets other users select the account. Admins no longer receive manage rights on someone else's connector.

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

* fix: stop the bubblewrap schedule test from stalling on Windows

Restore sys.platform before waiting, and poll from the event loop so the test does not pin the only free worker thread.

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

---------

Co-authored-by: jubaoliang <jubaoliang@tencent.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
fix(desktop): use native tray context menu
Bosheng0422 and others added 3 commits November 4, 2026 20:42
Experts created without a local root, or with the host filesystem root, ignored workspace_root_dir and landed outside it. Snap local_shell and filesystem roots, including composite routes, onto that directory at create time.
Co-authored-by: Cursor <cursoragent@cursor.com>
@jubaoliang
jubaoliang merged commit 4811eb4 into main Nov 4, 2026
14 of 15 checks passed
@jubaoliang
jubaoliang deleted the release/1.0.3b3 branch November 4, 2026 16:24
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.

10 participants