Users were asked to sign in again every ~30 minutes, even while active. Session lifetime - Browser: UDS_SESSION kept its login-time Max-Age and was never re-issued, so the cookie vanished 30 min after login. Re-issue it on use (throttled by slidingInterval). - Desktop: remote.mux WS traffic never passed the HTTP auth middleware, so the idle TTL never slid. assertPrincipalLive now reports activity and slides the session and the account's bridges (throttled). - New setting sessionIdleTimeoutHours (Settings page + config). Default 0 = no automatic sign-out; only a manual sign-out ends a session, so conversations and night-time cron are never cut off. 0.5-168 enables an idle timeout with the 7-day cap. The TTL is read live. - Stores support long TTLs: the absolute cap follows the TTL, older 7-day sessions are lifted on activity, and expiry timers re-arm instead of overflowing setTimeout (~24.8 days). UAC token check - Every 2 hours (and 1 min after start) each QR session's token is checked against icenterapi. A rejected token is only flagged, never a sign-out. Network errors, 5xx and bad JSON keep the previous verdict (the old 6h check revoked all of an account's sessions on any failure, including outages). - GET /uds-auth/session-status (non-sliding) returns alive + tokenValid. The client probes it every minute and on focus: a gone session signs the UI out with a notice; a rejected token opens a refresh QR once. - Scanning in refresh mode writes the new token into every live session of the account and the skill credential cache in place: no new session, no reload. Other accounts are rejected (token_refresh_emp_mismatch). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
6 KiB
uds-auth
0.3.10: 升级与协议变更说明 · SBOM · 第三方许可证
UDS authentication plugin for DeepSeek Harness (trusted single-instance access control — not strong multi-tenant isolation).
See package.json for the current 0.3.x version (breaking auth protocol vs 0.2.x — re-login required).
Support boundary
| Item | Notes |
|---|---|
| Runtime | Node ^22.19.0 || >=24 (aligned with DSH Host) |
| Mode | Single Host, trusted enterprise users, controlled skills |
| Out of scope | Untrusted multi-tenant execution isolation; multi-instance without shared session store |
Install
dsh plugin --profile web add -w "<abs>/oclaw/uds-auth"
# or clean tarball
dsh plugin add -w "<abs>/uds-auth-0.3.9.tgz"
Empty roles table requires UDS_AUTH_INITIAL_ADMIN or initialAdminEmpNo. Restart Harness after install.
ctx.get('udsAuth').runAsHost(cb) — Host background ALS.
ctx.get('udsAuth').mintTaskCapability({ empNo, dshSessionId, scopes, ttlSeconds }) — inject UDS_TASK_CAPABILITY for cron/skill.
Config
uacBaseUrl: https://uac.zte.com.cn # HTTPS required at startup/merge
userSearchUrl: https://icenterapi.zte.com.cn/...
loginSystemCode: '100000455558'
initialAdminEmpNo: ''
allowOpenRegistration: true
allowRawAgentToken: false
outboundAllowedHosts: icenterapi.zte.com.cn,icentermsg.dt.zte.com.cn
trustedHosts: '' # empty = loopback Host only; proxies must set allowlist
retainSkillCredentialsOnLogout: true
skillCredentialTtlSeconds: 604800
sessionIdleTimeoutHours: 0 # 0 = never auto sign out; 0.5–168h idle timeout; editable in Settings
Durable state defaults to ~/.uds-auth/ (roles / owners / credentials / bridge / sessions / config.runtime.json; override with UDS_AUTH_DATA_DIR). Security toggles take effect on the next request after Settings save. Empty outbound allowlist denies all.
Sessions
- Lifetime: by default (
sessionIdleTimeoutHours: 0) sessions never sign out on their own — only a manual sign-out ends them, so running conversations and night-time scheduled tasks are never cut off. Setting 0.5–168 hours in Settings enables an idle timeout that slides on activity (HTTP requests and Desktop WS calls) with a 7-day absolute cap. The browser cookie and the Desktop bridge token slide with the session. - UAC token check: every 2 hours (and 1 minute after start) each QR session's UAC token is re-checked against the icenterapi department profile. A rejected token (401/403, business error code, employee not visible/mismatched) never signs out: the client shows a QR prompt, and the scan writes the new token into every live session of that account in place — no new session, no reload. Network errors, 5xx and malformed responses keep the previous verdict.
- Client probe:
GET /uds-auth/session-status(never slides the TTL) runs every minute and on window focus. It returnsaliveandtokenValid; a gone session signs the UI out, a rejected token opens the refresh QR once. - Survives restarts: the default in-memory store is persisted to
~/.uds-auth/sessions.jsonand restored on Host start (emergency admins have no SSO re-login, so they depend on this).- Only the SHA-256 hash of each session token is stored; user data (may include the UAC token) is AES-256-GCM encrypted with the plugin key
.uds-auth-secret-key. - Expired, undecryptable or tampered rows are dropped on load; disabled-account and emergency-password rotation checks still run on every request.
- Delete
sessions.jsonand restart to sign everyone out. If the key file is unavailable, sessions are simply not persisted.
- Only the SHA-256 hash of each session token is stored; user data (may include the UAC token) is AES-256-GCM encrypted with the plugin key
- Not used when a Redis session store is configured.
Admin / emergency
super_admin/admin/fallback_adminshare the same admin-class permissions (identity labels, not stronger isolation).- The emergency account is enabled by default: username
administrator, passwordAdmin@123. Change it in Settings after first sign-in. Admins can also disable it. Older roles files without an explicit emergency-account state receive this default; an explicitly disabled account stays disabled. Historical SHA-256 hashes ofAdmin@123migrate to scrypt. - Local unlock:
node scripts/seal-local-admin.mjs(interactive /--stdin— no argv passphrase) →UDS_AUTH_LOCAL_ADMIN_BOX. - Removing a user disables the account by default (open registration cannot auto-revive).
DSH compatibility
Verified on DeepSeek Harness Desktop 0.2.0-rc.2. The plugin relies on DSH internals that are not a public API; after a DSH upgrade run npm test and check login, workspaces and Settings → Models on a real install:
- Gateway
prepareInvocation/resolveDescriptor/openRemoteEvents/receiveRemoteEventResult(if missing, the ACL refuses to install instead of failing open); - the client
connection/resetevent (emitted after login so boot-time caches such as the settings mirror re-read); - endpoint names
workspace/follow,workspace/initializeDefault, and the Desktop welcome readssettings/describe,llm/listConfigurableProviders,credentials/describe,account/getState.
Desktop opens its connection anonymously at boot and reconnects after login. The plugin therefore:
- parks anonymous stream calls (e.g. account state) without data instead of failing them (a failure is terminal for DSH and shows "sign-in failed / add API key");
- answers anonymous
workspace/followwith an empty list and anonymousworkspace/initializeDefaultwith "nothing created"; - holds
/api/*requests for up to 15 s while the gateway ACL is still installing, instead of returning 403.
Skill / cron
Prefer Host-minted UDS_TASK_CAPABILITY + DSH_SESSION_ID with /uds-auth/outbound. Raw /agent-credentials stays off unless explicitly enabled. Python helper: skill-helpers/python/uds_skill_auth.py.
Chinese docs: README.zh.md, docs/skill-auth-standard.zh.md.
License
MIT