Repository navigation
Description
状态:频道 Agent 凭证、独立回复和固定 health-check 已合并并有 UAT/native 证据;WebSocket、完整活动连接撤权及剩余平台验收未完成。 Human 于 2026-10-18 授权合并既有 Draft;不授权 Prod 部署。
Parent
What to build
群管理员用手机创建外部 Agent 凭证;另一电脑上的独立参考 Agent 接入群。手机 @它请求已预授权 health-check,结果回到同群;手机可撤销它的接入。
所有业务走生产目标 owner、持久存储及公开接口;同时交付所需界面、CLI 与回归证据,不交付隔离的横向层。
Acceptance criteria
- 手机与 CLI 均能创建群内 Agent、签发/撤销凭证、查看成员。凭证限群、限 actor、限能力;秘密安全交付,普通 stdout/日志不回显。
- 参考程序不依赖 DSH/BotHarness,通过公开 HTTP 和主动连接 WebSocket 收发与补历史。无需 Tunnel 或公网接收端。
- 新 Agent 默认不能读加入前历史;不能冒充 Human/其他 Agent、跨群读写或管理实例/组织。
- 仅执行预授权的固定 health-check,无任意 shell/路径参数;校验真实请求人、目标、授权和有效期,独立持久记录请求及实际执行计数。
- 手机看到来自正确 Agent/环境的结果。普通聊天和结果消息不会触发工作,重复已完成请求只返回既有结果。
- 凭证撤销阻止现有连接及新请求继续读写;主人资格失效停止新的工作授权。重连不重新授予资格。
- 固定能力的预授权配置有无模型命令路径;默认入群/群管理员身份不等于获得本机执行权。
- 自动契约/权限回归及真实手机→第二电脑→手机结果通过 Human QA。只证明参考 Agent,不宣称已适配所有第三方产品。
后续群管理方向:Agent 将可用自己的 actor/限范围凭证担任群主或管理员,通过同一 TOB API/CLI 执行已授权管理;不借用 Human 会话或供应商控制面密钥。该能力及入群申请审批不属于本票,后续验收见 spec §8。本票继续验证普通 Agent 越权被拒绝。
Blocked by
#3 — T2 完成并通过 Human QA。
TOB 公开仓库及规划迁移已获授权;本任务仍未启动,不授权实现、部署、配置 WorkOS、发信或收费。
合并检查点 — 2026-10-18
TOB #17–24 已依次合并,main 44dbb505ae6ba36a82372d7e11f3c356e3651931;DeepSeekBot #2103/#2121/#2131/#2138/#2146/#2150/#2152/#2154 的原始提交由 #2157 整合,main f9ed710b5a718526cc97a30583af80fa14e28424。上述最新 Human 授权取代历史“暂不启动/不合并”描述;未运行的验收仍不算通过。此次没有部署或发布,Prod 未变。
频道 Agent 凭证、独立回复和固定 health-check 已合并并有 UAT/native 证据;WebSocket、完整活动连接撤权及剩余平台验收未完成。 本总任务保持开放。
Activity
DoodleBears commentedon Oct 18, 2026
Agent work claim
codex/local/01a148ae-ec7f-74b2-8157-961411ce54c9(not a DSH Session)codex/4-channel-agentsHuman's latest goal authorizes autonomous implementation and UAT testing, superseding the old ticket's not-started wording. No Prod or merge; no email delivery or implicit channel bridging. Owner/member and local execution permission remain separate.
docs: record first channel Agent UAT verification (#4)
DoodleBears commentedon Oct 18, 2026
Agent work continuation — fixed health-check
Agent-Task:
codex/local/01a148ae-ec7f-74b2-8157-961411ce54c9; branchcodex/4-agent-health-check, stacked on #22.The identity/echo sub-slice passed actual UAT and native Batch17; evidence in #22 and BotHarness/DeepSeekBot#2152. T3 remains open.
Next narrow behavior: phone/CLI submits a structured, expiring request to one Agent; a separately configured reference process validates the actual Human requester, target and fixed capability, durably claims execution, performs a bounded environment health-check, and posts its typed result to the same channel. No arbitrary shell/path/input. Local owner policy is separate from cloud membership. Ordinary text and Agent results cannot start work. Completed-result replay does not rerun; ambiguous started records require review. Current requester/issuer qualification and revoke are checked before starting.
Public API/CLI and mobile entry ship together with focused workerd/client/ledger checks and real UAT evidence. WebSocket, organization-wide Agents, DSH Messaging binding, two-Host execution and broader failure coverage remain subsequent increments; no claim that echo alone completes them. UAT only, no Prod, no merge. Latest Human autonomous goal authorizes this continuation.
docs: record fixed work UAT and native evidence (#4)
DoodleBears commentedon Oct 18, 2026
T3 fixed health-check slice handoff: Draft mobile #2154, head e63bcdc8a5e2c5d5f5fec15fa4e835a150332ef5; service #23, runtime23db30d, UAT schema5/4/4.
Actual phone member request → independent reference process → same-channel result passed. Fresh process replay executes0; removing local allowed-users produces refusal0. Owner CLI confirms exactly4messages and identity/request-result correlations. Six Dynamic Type/theme controls reachable, max action used; cold restart/refresh retains same session/results. Own mobile logout/remote/latest-refresh/local deletion and otherCLI isolation pass. RuntimeCI39088289282 passed; latest evidence-only CI tracked on PR. Six embedded PR images independently verified rendered.
Local HTML review aid is saved under the ignored task folder and bound to this exact head; it is not a hosted reviewer artifact. Public PR outline plus evidence README contain the same boundaries. First failures and limits are recorded in the evidence README. Android/physical/Human/HostedAuthKit/WS/DSH/two-Host unqualified. Batch18 released, own testAgent revoked and mobile fixture removed; no Prod/merge.
Next autonomous scope is the DSH adapter/two isolated Hosts through real UAT, with an explicit local work policy and durable authority. Keep T3/overall goal open. This issue is ready for Human review, not closed by source or native evidence alone.