Commit Graph

798 Commits

Author SHA1 Message Date
jxxghp
9323f00510 docs(agent): clarify config safeguards 2026-11-11 13:51:21 +08:00
jxxghp
6bc19a59cc fix(agent): protect plugin config and replay receipts (#7360) 2026-11-11 13:39:35 +08:00
Penguin Secret Agency
8be08ee773 feat(llm): 新增 UberLLM provider (#7302)
* feat(llm): 新增 UberLLM provider

* fix(llm): 调整 UberLLM 排序值

与 cheapaimodel 的排序修复协同:cheapaimodel 使用 60、GitHub Copilot 顺延为 61 后,UberLLM 顺延为 62,保持目录排序值唯一。

---------

Co-authored-by: sang nguyen <sangnguyen@sangs-MacBook-Pro-5.local>
Co-authored-by: jxxghp <jxxghp@gmail.com>
2026-11-11 10:45:46 +08:00
Nicho Deng
192bd7d5b8 feat(llm): 新增 cheapaimodel provider (#7303)
* feat(llm): 新增 cheapaimodel provider

* fix(llm): 修正 cheapaimodel 排序冲突

cheapaimodel 与 Obscurify.ai 同为 59,合并后会留下重复排序值。将 cheapaimodel 调整为 60,GitHub Copilot 顺延为 61,保持新预设位于 Obscurify.ai 之后。

---------

Co-authored-by: jxxghp <jxxghp@gmail.com>
2026-11-11 10:44:50 +08:00
jxxghp
d0fb0a21a8 fix(agent): make session_search sort Gemini compatible 2026-11-10 17:24:53 +08:00
jxxghp
252d05767f feat(browser): add queued CloakBrowser watchdog recovery 2026-11-09 17:24:32 +08:00
pyss56
b44af19e8d fix(agent): 空字符串联合参数归一为 null (#7339) 2026-11-09 15:06:15 +08:00
jxxghp
263ae3fbc7 fix(llm): retain finite provider request timeouts 2026-11-07 05:52:11 +08:00
freeman
68cfc377ee feat(history): register externally completed transfers (#7293) 2026-11-04 06:24:38 +08:00
SmallSpider
7f227857d8 fix: 修复 WebAgent 会话与选择状态 (#7266)
* fix(agent): preserve WebAgent session state

* test: separate subscription pressure coordination and owner waits

---------

Co-authored-by: jxxghp <jxxghp@gmail.com>
2026-11-01 17:49:20 +08:00
ranzhigg
914abf6273 fix: tighten API authorization boundaries 2026-10-30 16:14:49 +08:00
jxxghp
9aab7c0f4c fix(agent): pin scheduled tasks to real execution accounts (#7241) 2026-10-29 22:39:08 +08:00
jxxghp
8fe8ea477e fix(agent): preserve cache keys for compatible Responses endpoints (#7236)
Forward the stable anonymized session key through the generic OpenAI-compatible Responses path so gateways can retain cache routing across tool calls.

Refs #7233
2026-10-29 17:17:48 +08:00
Tony Davis
421dd2c055 feat(llm): 新增 Obscurify.ai provider (#7229) 2026-10-29 17:06:11 +08:00
Cat
87561a90d1 fix(agent): route agent command receipts to the signed channel user (#7228)
- Agent 代渠道用户调用内部 API 时,把会话渠道与渠道用户写入宿主签发令牌的签名
  声明(agent_channel、agent_channel_user),不再新增可伪造的请求头。
- /commands/run 对 Telegram 等通知渠道只使用与请求渠道一致的签名渠道用户执行命令
  和投递回执;缺少该声明时明确拒绝,不再把 MoviePilot 用户表主键当作渠道 chat id。
  Web、WebAgent 与无渠道调用仍按 MoviePilot 用户 ID 处理。
- 只读公共资料借用宿主身份,不携带渠道发起人声明。
- 同步依赖基线与架构总览中的内部导入边数量。

Refs #7225

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: jxxghp <jxxghp@gmail.com>
2026-10-28 19:05:52 +08:00
andyxu8023
e1b276bf9b fix(agent): encode API session header as ASCII (#7210) 2026-10-27 10:36:28 +08:00
jxxghp
42edfd1da8 fix(api): bound collection views and query calendar dates (#7200) 2026-10-27 00:40:09 +08:00
JontyLee
0ba73f5b87 fix(database-operation): 拦截未显式授权的凭据配置表读取
systemconfig 按 SystemConfigKey 聚合实例级 API Key、Cookie 与口令,托管
config.system.get 默认脱敏,但数据库技能脚本与 MCP database_operation 接受
任意 SQL,一条 SELECT 就能整包取回明文。

现在由脚本本身按表拦截:引用 systemconfig 的语句(含引号标识符、模式限定名、
CTE 与子查询)在读写两种模式下都默认拒绝,只有显式 --show-secrets /
show_secrets=true 才放行,失败信息直接给出可执行的下一步,因此 MCP 工具、
技能调用与命令行入口共享同一道边界。宿主策略把开启该开关的查询判定为
SENSITIVE_READ:要求系统管理员确认,结果标记为 SECRET,回执摘要不再保留内容。

门禁范围收敛为单表:userconfig 只承载界面偏好,plugindata 是文档化的诊断路径,
其余表的凭据残留风险改由 SKILL.md 的窄投影要求约束,避免以不可绕过的方式破坏
既有可运维性。MCP 文档、技能说明与相关契约测试同步更新。
2026-10-26 20:20:16 +08:00
Nirupam Bhowmick
363b8d9e5d feat(llm): 新增 Runcrate provider
新增 Runcrate OpenAI-compatible 预设(https://api.runcrate.ai/v1),沿用共享
openai_compatible 通道,模型目录通过 /v1/models 发现;附带模型发现的边界回归测试。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-25 08:02:17 +00:00
Aqr-K
7584d4b9fc fix: isolate agent sessions and retain runtime resource ownership (#7149)
* fix: arbitrate subscription identity and restore workflow state

* fix: isolate agent sessions and retain runtime resource ownership

* test: preserve legacy empty-owner subscription transactions

* test: handle reaped terminal descendants during shutdown polling

* fix: align shared atomic configuration method with lifecycle prerequisite

* test: cover subagent cancellation and completion lifecycle boundaries

---------

Co-authored-by: jxxghp <jxxghp@gmail.com>
2026-10-23 08:20:59 +08:00
jxxghp
ec5ff575f0 fix(agent): restore API bodies before direct request dispatch 2026-10-23 07:55:16 +08:00
jxxghp
17a866f47b fix(transfer): restore AI takeover for failed transfers 2026-10-22 09:43:03 +08:00
jxxghp
cc2e02793e fix(telegram): wait for final streaming edit interval
Wait once for the remaining message edit interval during Agent finalization, without bypassing Telegram 429 cooldown or replaying confirmed text.

Refs #7124
2026-10-22 08:25:25 +08:00
jxxghp
6d4ce83837 fix(types): clear host mypy debt and prevent typing regressions
严格检查全部宿主源码,清零类型错误预算并移除按模块类型豁免。
按原因、所属方法、错误码及语法指纹锁定必要兼容例外和站点类型桩,
CI 阻断新增忽略、检查范围退化、本地桩替代真实接口及新的类型错误。

修复空值、异步和供应方类型边界,保持插件与 API 兼容合同,
并补充存储、订阅、通知、Agent 和工作流边界的回归覆盖。
2026-10-21 22:03:18 +08:00
jxxghp
53c5344f4e fix(agent): avoid replaying delivered streaming progress 2026-10-21 00:20:13 +08:00
alexcastgonz
14b9ea6d3e fix(setup): 更新 Google 默认模型并兼容 Gemini 3.8 思考级别 (#7109)
* fix(setup): 更新 Google 默认模型为 gemini-3.8-flash

安装向导中 Google 的预设模型 gemini-2.5-flash 对新建的 Gemini API 密钥
已返回 404,在模型选择处直接回车会写入无法使用的 LLM_MODEL。改为
Google 文档中推荐给新项目的稳定模型 gemini-3.8-flash。

新增用例校验预设值不再是 Gemini 2.5 系列。

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

* fix(llm): honor Gemini 3.8 minimum thinking level

---------

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-authored-by: jxxghp <jxxghp@gmail.com>
2026-10-20 18:41:57 +08:00
paysen
1f7e91b645 fix(llm): serialize ChatGPT OAuth token refresh (#7092) 2026-10-20 11:25:52 +08:00
aiapienthusiast
75ee313f4c feat(llm): add Cheaper Inference provider (#7088) 2026-10-19 23:02:38 +08:00
jxxghp
c85f9b262f feat(agent): isolate interactive execution from background services (#7083) 2026-10-19 18:12:05 +08:00
jxxghp
6e5c9d0afb fix(storage): 修复 reflink 同存储回退语义与测试稳定性 (#7081)
- transhandler: 恢复同存储 copy/move 的 elif 回退语义,不再在 item 操作未命中后重复调用远端操作
- local: reflink 空源路径直接失败,避免新增 mypy 存量债务
- tests: 管道用例改为模拟克隆,系统用例改用真实能力探测,并新增不支持时严格失败的用例
- mcp: 整理历史模式说明补充 reflink
2026-10-19 17:07:02 +08:00
秋澪Akimio
114c2cebe5 feat(storage): 本地存储增加引用链接(reflink)整理方式 (#7077) 2026-10-19 16:48:41 +08:00
Aqr-K
bf4ef4214e perf(mediaserver): add caller-controlled query projections (#7068)
* perf(mediaserver): add caller-controlled query projections

* test(mediaserver): cover query projection option contracts

---------

Co-authored-by: jxxghp <jxxghp@gmail.com>
2026-10-19 08:33:45 +08:00
DDSRem
e1b703aacc refactor(agent): centralize graph resource ownership (#7062) 2026-10-18 21:49:57 +08:00
DDSRem
e0642eafa7 chore(types): reduce v3 host mypy debt (#7057)
* chore(types): clear AniList endpoint mypy debt

* chore(types): clear Douban endpoint mypy debt

* chore(types): clear recommendation endpoint mypy debt

* chore(types): clear subscription notification mypy debt

* chore(types): clear SDK string compatibility mypy debt

* chore(types): ratchet v3 host mypy across remaining files

* chore(types): reduce v3 host mypy debt and retire ledger

* chore(types): reconcile host snapshot with upstream HTTP pool

* test(types): cover host paths exposed by mypy cleanup
2026-10-18 09:37:49 +08:00
DDSRem
2de976f6b3 Merge pull request #7050 from DDSRem-Dev/codex-comment-migration
docs: 统一全项目中文注释和结构化契约说明
2026-10-17 19:26:29 +08:00
jxxghp
14680dd4ad fix: narrow tool schema model type 2026-10-17 12:36:48 +08:00
jxxghp
0658c7f384 fix(agent): clean removed bundled subagents 2026-10-17 09:09:33 +08:00
jxxghp
f0b8bf75b8 refactor(agent): type provider manager initialization 2026-10-17 02:27:01 +08:00
jxxghp
8df7bd40ad refactor(agent): type sanitizer container delimiter stack 2026-10-17 02:26:56 +08:00
jxxghp
226ecd7fd1 refactor(agent): type streaming callback task lifecycle 2026-10-17 02:10:50 +08:00
wenking
5ec9516a89 fix(mcp): 支持显式 opt-in 的 tools/list schema 紧凑投影
默认保留 MCP tools/list 的完整 schema;显式携带 X-MCP-Schema-Mode: compact 时,为超限工具提供保留顶层对象和 operation/action 枚举的紧凑投影。REST 合同、tools/call 校验与权限判定沿用完整合同。

合并前修复嵌套定义引用与 patternProperties 投影边界,使用 app.agent.tools.projection 作为纯函数 owner,补充回归和相关 Skill 说明,并同步当前 v3 的架构快照与类型低水位。

验证:113 项相关回归、167 项架构策略测试、45 项架构快照测试通过;Pylint 10/10,类型、复杂度及并发等相关棘轮通过。最终 PR head 6d25e1039d 的全部 10 项必需 GitHub 检查通过;待合并树与 CI 测试树一致。

投影是 best-effort,不承诺压入 5000 字节以内;未进行生产实例或镜像发布验证。

Co-authored-by: clone <clone_fan@foxmail.com>
2026-10-16 19:52:57 +08:00
jxxghp
225da1a7f0 fix(agent): prefer model catalog for image input support 2026-10-16 12:22:31 +08:00
jxxghp
9d5ad08e08 refactor(types): govern agent display contract with strict mypy 2026-10-16 10:03:27 +08:00
jxxghp
4955b7b9ea fix: harden backend isolation and execution safety 2026-10-16 09:11:28 +08:00
jxxghp
edc719596d feat(update): prepare latest updates for a single restart 2026-10-13 12:50:01 +08:00
DennyHo0917
f4cd7d02ad feat(llm): add API Route provider (#6972)
Signed-off-by: dennyho0917 <dennyho0917@gmail.com>
Co-authored-by: jxxghp <jxxghp@gmail.com>
2026-10-13 07:25:47 +08:00
jxxghp
c8e935fc6c fix(agent): point capability entrypoints at canonical manager module (#6969) 2026-10-12 17:47:56 +08:00
jxxghp
bd1e6eb899 fix(agent): separate public API queries from bound user operations 2026-10-12 07:30:35 +08:00
jxxghp
ec93101289 Merge pull request #6932 from binyangzhu000-sudo/feat/atlas-cloud-provider 2026-10-09 18:52:23 +08:00
binyangzhu000-sudo
6d479317c1 feat(llm): 新增 Atlas Cloud provider
Atlas Cloud (https://api.atlascloud.ai/v1) 是 OpenAI 兼容的聚合网关,一个
Key 覆盖 DeepSeek、智谱、月之暗面、通义千问与 MiniMax 等开源权重模型,
形态与 OpenRouter 一致,sort_order 取 50 与 60 之间的 55。

用 openai_provider() 而非 catalog_openai_provider():后者的
model_list_strategy 是 models_dev_only,而 Atlas Cloud 尚未被 models.dev
收录;前者走 openai_compatible,直接读网关自己的 /v1/models。
models_dev_provider_id 保留默认值 atlascloud,这样等 models.dev 收录后
_models_dev_reserved_provider_ids 会认到它,不会再自动生成一个重复条目。

实测:用 discovery 内部同款 AsyncOpenAI 构造调用,models.list() 返回 117
个模型,deepseek-ai/deepseek-v4-flash 的对话正常返回。

Signed-off-by: binyangzhu000-sudo <binyangzhu000@gmail.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Signed-off-by: binyangzhu000-sudo <binyangzhu000@gmail.com>
2026-10-09 16:03:47 +08:00