dsh-sentinel

dsh-sentinel

English | 中文

Condition-driven wakeup for DeepSeek Harness: the agent registers a watch, goes to sleep — even closes the session — and the sentinel wakes it when the condition happens. Every subscription and every fire is a user-visible session event, and the browser dock shows what is on duty.

How it works

The node half owns one server-lifetime runtime that folds a plugin-owned sidecar log ($DSH_HOME/sentinel.jsonl) into live subscriptions, probes every sensor on a shared 5s heartbeat, and delivers wakeups through the official followup channel — resuming a dormant session's agent first when needed. Subscriptions therefore survive process restarts, and conditions that become true while the server is down late-fire on the next probe.

Watching is a resident-process concern: probing and fire delivery only run while a long-running dsh process (typically dsh web) is up. Headless one-shot runs load the plugin and can create, list and cancel watches, but nothing probes after the process exits — those watches become active once a resident process starts.

One duty owner per $DSH_HOME: a lease file (sentinel.lease) makes the first process own probing and delivery; a second dsh process on the same home stays passive (tools work, writes persist to the shared sidecar) and takes over within one lease TTL of the owner dying. The owner re-reads the sidecar every heartbeat, so watches created on a passive instance are adopted automatically. Delivery is at-least-once: a fire logged but not delivered before a crash is requeued on the next boot from its delivered watermark.

The browser half is a dock card above the composer (the conversation.input.dock family) listing the session's active watches — sensor, target, live probe state, fire budget, next-probe countdown — plus recent fire history when expanded. It polls the read-only state route and renders nothing when the session has no watches.

Two surfaces make the server-global watch set visible. A Sentinel entry in the sidebar's global panel list (sidebar.panellist, id sentinel) opens the watch table in the central column (main, the same id) — its glyph carries a state dot while any watch is active. The dashboard is a standalone table of every watch across every session: session (active/dormant), sensor, target, pattern, fire budget, last probe state, next probe.

Sensors

Kind	Engine	Fires on
file	path snapshot + inotify push	snapshot change (sub-second); accelerated by fs events
command	read-only shell line, probed on an interval	output/exit-code change
http	URL probed on an interval	status/body change
process	pgrep -f pattern, probed on an interval	match-set change
port	TCP connect to [host:]port, probed on an interval	reachability change (open/closed/timeout)
webhook	pure push	any POST to the returned hook URL

With pattern, probe kinds fire on the no-match→match edge of that regex and webhooks accept only matching payloads; without it, probe kinds fire on any change after the baseline.

Configuration

All deployment-tunable knobs live in the plugin's config schema (defaults in parentheses); override them on the bundle row in your profile's cordis.patch.yml:

- id: dsh-sentinel
  name: dsh-sentinel
  config:
    heartbeatMs: 5000            # probe round interval
    probeConcurrency: 8          # in-flight probes per round
    maxSubscriptionsPerSession: 16
    maxPendingWakeups: 8         # queued wakeups per session before dropping oldest
    defaultIntervalSeconds: 30   # when a watch does not specify one (5–86400)
    defaultCooldownSeconds: 60
    dutyLeaseTtlMs: 30000        # passive-instance takeover window after the owner dies
    notifyWebhookUrl: ''         # optional: POST every fire here as JSON

Invalid values fail plugin load with a schema error rather than misbehaving at runtime.

notifyWebhookUrl fans every fire out of the harness as a JSON POST ({plugin, event, sessionId, id, kind, target, note, fireNumber, maxFires, summary, after}) — point it at a Lark/WeCom/Slack bot or any receiver. Delivery is at-most-once: a failed POST warns in the log and never blocks the in-harness wakeup.

Tools

• sentinel_watch — register a watch: kind, target, optional pattern, interval (1–3600s, default 30), note (delivered verbatim with every wakeup), maxFires (default 1: one-shot), cooldown (default 60s), optional ttl.
• sentinel_list — active watches with live probe state.
• sentinel_cancel — cancel one watch by id.

Routes

• GET /plugins/dsh-sentinel/state?sessionId=… — read-only state for the dock and the sidebar panel (omit sessionId for every session).
• GET /plugins/dsh-sentinel/dashboard — the server-global watch table.
• POST /plugins/dsh-sentinel/hook?id=watch-N&s=<sessionId> — webhook entry; put a curl into a CI job, git hook, or another machine's script to wake the agent. Watch ids are per session, so the s qualifier is what keeps two sessions' watch-1 hooks from colliding (the tool hands out the full URL). URLs without s still work and resolve to the first matching webhook watch.
• POST /plugins/dsh-sentinel/cancel?sessionId=…&id=watch-N — manual cancel. The dashboard table and every UI row carry a ✕ that calls this, so a watch can always be stopped by hand — including orphaned ones whose session (and agent) is long gone; the host has no session-deleted event, so this is the kill switch of last resort.
• All four routes enforce a browser-trust fence: browser-marked cross-site requests (a malicious page can form-POST to localhost) and DNS-rebinding attempts (Host/Origin naming a DNS host) get 403. Headerless clients such as curl and CI jobs are unaffected. The state route also reports duty (lease heartbeat age) and droppedWakeups per session (queued wakeups dropped by the maxPendingWakeups cap).

First-probe semantics: a pattern-less watch absorbs its first observation as the baseline (no fire), while a pattern watch whose target already matches fires on the first probe — the condition already holds.

Compatibility

Verified against these harness versions (plugin loads, duty lease is held, web routes answer):

• 0.2.1-alpha.1 — 2026-10-07, rehearsal on a copy of the live profile: the plugin activates, holds the duty lease, serves its client half and answers the state route (68 boot rows). Reaching it needed the peer range widened — see the 0.2.0-rc.2 note below — and the host's version gate was checked directly against evaluatePluginCompatibility: the current range admits 0.2.0-rc.2 and 0.2.1-alpha.1 while still refusing 0.2.0-rc.1 and older.
• 0.2.0-rc.2 — 2026-09-30, runtime and client build dependencies aligned with this release; npm and pnpm lockfiles regenerated. Dormant-session wakeups restore the saved preset from the session projection and preserve the complete default model selection, including reasoning effort. Duty leases stay in the runtime's original home; startup and probe rounds settle before disposal releases the lease. 2026-10-07: the three harness peers were widened from a single exact version to >=0.2.0-rc.2 <0.3.0, so one build serves the whole 0.2 line instead of exactly one release — an exact pin left the plugin uninstallable on the very next 0.2.x, which is what happened going from 0.1.7 to 0.2.0.
• 0.1.7-rc.2 — 2026-09-29, clean-profile upgrade rehearsal against a copy of the live profile: 0.1.7 removed the shared catch-all plugin message-source kind (every producer now declares its own), so a wakeup carries { kind: 'sentinel' } — same context placement in the transcript, and it renders as "Sentinel" on both lines. The harness dependency range was also re-pinned to the 0.1.7 line, because under strict semver >=0.1.5-rc.2 <0.2.0 does not admit 0.1.7-rc.2 (the prerelease rule); left alone, a 0.1.7 host would have resolved this plugin's harness imports to 0.1.5 copies — the exact drift the 0.1.5 alignment removed. Verified: pnpm typecheck and all 63 tests pass, the plugin activates and holds the duty lease, the web routes answer, and the served client bundle carries sidebar.panellist (65 boot rows)
• 0.1.5-rc.2 — 2026-09-15, live web deployment after the 0.1.5 alignment: the plugin's whole runtime import closure resolves to the deployed line (its harness imports are declared dependencies, so a profile's older hoisted copies can no longer shadow them), the client half builds against the real 0.1.5 types with no shims, pnpm typecheck and all 63 tests pass, and a live file watch fired through inotify 1s after the change and the wakeup was delivered into the session as a plugin-sourced message; after the restart the deployment serves the new client half (bundle rev changed, sidebar.panellist present, 54 boot rows)
• 0.1.5-alpha.2 — 2026-09-09, temporary web-profile smoke: Node plugin load, duty lease, state/dashboard routes, and the browser plugin bundle all worked with no browser-console errors; conversation.input.dock remains a supported session-scoped list slot, and the plugin sidecar is unaffected by the Session V3 migration
• 0.1.1-rc.2 — 2026-08-26, source-build smoke: git install into a web profile, duty lease held, state and dashboard routes answer
• 0.1.0-rc.8 — 2026-08-20, scratch-profile smoke
• 0.1.0-rc.7 — 2026-08-20, live web deployment

Compatibility means the cordis loader entries, the ctx.agents followup channel, the declared slot seats, and the web routes keep working; file an issue if a harness version breaks any of them. Its harness imports (@deepseek-ai/dsh-tools, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-scope) are declared dependencies pinned to the verified line, so the plugin carries aligned copies instead of inheriting whatever a profile's hoisted store happens to hold; @deepseek-ai/cordis stays a peer because service identity must come from the running host.

Install

One line through the official bundle channel:

dsh plugin --profile web add dsh-sentinel

Or straight from git (build artifacts are committed, so the git-source install runs no build):

dsh plugin --profile web add "github:fuhefei/dsh-sentinel#v0.11.0"

Alternatively, add the node half manually through a patch-list configuration over the shipped base:

# cordis.patch.yml
- insert:
    - id: dsh-sentinel
      name: dsh-sentinel

The browser half ships in the same package (./client) and is injected by the Web UI's plugin loader.

Sidebar surface (0.1.5 and later)

The dock, the sidebar entry and the dashboard all work on a stock host: the entry registers into the official sidebar.panellist seat and its panel into the layout's main slot, with no host patch.

Through 0.1.2 the global view instead grew a branch under each watched session row, which needed the session-row holes the official tree never declared; that path is retired with patches/session-row-holes.patch kept only for those older trees. 0.1.5 dropped the hole, so a plugin built for it must use the panel seat above.

better-sidebar integration (optional)

When dsh-better-sidebar is installed in the same profile, sentinel registers its global watch table as a sidebar tab (dsh-sentinel:watches, in the + menu) through better-sidebar's documented ctx.betterSidebar.registerTab extension surface: every watch server-wide with live probe state, fire budgets and recent fire history, fed by one shared poller. No configuration needed; without better-sidebar the registration is silently skipped and the dock / panel / dashboard keep working as before.

Plays well with

Install dsh-notification alongside sentinel and the whole wakeup loop reaches your desktop: sentinel wakes the agent, the agent works the turn, and the turn's completion fires a desktop notification — no integration needed, the two plugins compose on their own.

Develop

npm install
npm run build     # tsc -b + tsdown (lib/index.js, lib/client.js)
npm test          # vitest: domain fold/normalize, sensors, dashboard escaping, e2e wakeup flow

License

BSD 3-Clause. See LICENSE.