@yiln-dsh/dsh-plugin-web-daemon

A DSH plugin that manages the dsh web worker as a real systemd unit and adds a compact server-health monitor above the New Session button. The plugin generates the unit file, maps the GUI buttons onto systemctl start/stop/restart/reset-failed, and lets the Settings page edit the few fields that matter. It also keeps a durable registry of top-level sessions attached to the daemon so they can be resumed after a process restart.

Crash recovery (Restart=always — it keeps restarting unless explicitly stopped with systemctl stop), boot autostart (systemctl enable), and logs (journald) are handled by systemd itself. Session recovery is handled by the plugin's registry at $DSH_HOME/plugins/dsh-plugin-web-daemon/active-sessions.json: attached top-level agents are recorded atomically with their live status. Only records marked running: true are resumed after the next daemon start; idle sessions are removed from the registry. A final turn terminated by a daemon shutdown is explicitly queued once for continuation after the Agent is restored. The registry stores only session identity, workspace, preset, and model options; it contains no credentials. Only the managed systemd worker (DSH_WEB_DAEMON_WORKER=1) owns this registry; a foreground owner sharing the same DSH_HOME never resumes the same identity. A small active-sessions.lock lease beside the registry prevents two managed workers from resuming the same session. A normal shutdown keeps the records for the next start, while an explicitly disposed session is removed.

What it does

• Adds a Web daemon card under Settings > Plugins > Plugin configuration with live unit status (active state, PID, restart count) and Start/Stop/Restart/Reset buttons plus the editable fields.
• Generates /etc/systemd/system/<unit> (system scope, needs root) or ~/.config/systemd/user/<unit> (user scope) on save/start.
• Adds a compact Server status panel above New Session with live CPU percentage, memory percentage, network download/upload rates, and mounted physical block-backed filesystem usage (device, mount point, used/total bytes, and percentage). The panel grows with the number of disk rows and collapses to a status dot with the sidebar rail.
• Exposes daemon state through /_dsh/web-daemon/* JSON routes and server metrics through /_dsh/web-daemon/metrics.
• Disk metrics list local block-backed filesystems visible to the worker and mounted in its namespace; virtual, network, overlay, loop, and zram filesystems are omitted. Unmounted disks cannot report filesystem occupancy. Byte counters are returned as decimal strings for large-volume precision.
• Records top-level sessions and their live status in $DSH_HOME/plugins/dsh-plugin-web-daemon/active-sessions.json. Only sessions recorded with running: true are resumed; idle sessions are not started just because they have a transcript. Persisted sessions that were deleted are pruned; subagent sessions are intentionally excluded.
• A resumed session restores its transcript and Agent. If the previous process stopped while the session was marked running, the plugin queues one internal recovery notice to continue it. Tool calls marked as unknown are explicitly left for the model to verify before retrying.
• Registers the web-daemon settings namespace in the Host settings service (the card is keyed by that namespace on settings.plugin.item).
• The worker runs dsh web --profile <profile> --no-open --port <port> bound to loopback; LAN exposure is the job of @yiln-dsh/dsh-plugin-auth-webserver.
• The unit gets DSH_WEB_DAEMON_WORKER=1; a daemonized GUI detects this and keeps only Restart available (e.g. to pick up plugin updates) — it asks systemd to restart its own unit, so the fresh process comes up even though the requesting one dies mid-request. Start/Stop and configuration stay with the unit owner's GUI.
• On a Linux worker without DISPLAY/WAYLAND_DISPLAY (and not WSL), the plugin owns the /api/host.openPath and /api/host.openTextFile guard. It returns a structured message directing users to the file explorer preview or download instead of leaking a native xdg-open failure into the GUI.

Install

The published package is @yiln-dsh/[email protected].

npm package

dsh plugin --profile web add @yiln-dsh/dsh-plugin-web-daemon@latest

The package is a dual-face plugin: its host row talks to systemd and its dsh.client export registers the Settings plugin-configuration card. Client module changes need a page refresh; the running dsh web process does need to be restarted once so the host row is composed into the profile.

Configure

Open Settings > Plugins > Plugin configuration > Web daemon in the GUI. There are five fields:

• enabled: maps to systemctl enable/disable plus start on boot.
• systemdScope: system or user.
• systemdUnit: unit file name (default dsh-web.service).
• profile: DSH profile the worker runs.
• port: worker listen port on loopback.

Everything else is intentionally fixed: Restart=always with RestartSec=2, start-rate limiting left to systemd's defaults, logs in the journal (journalctl -u <unit> -f). Saving rewrites the unit, reloads the daemon, and restarts the worker if it was running.

Session recovery

The registry lives at $DSH_HOME/plugins/dsh-plugin-web-daemon/active-sessions.json and stores one record per top-level session together with its live status:

{
  "version": 1,
  "updatedAt": "2026-08-27T03:22:27.680Z",
  "sessions": [
    { "sessionId": "session-…", "running": true, "cwd": "/path", "agentPreset": "yiln", "agentOptions": { "provider": "…", "model": "…", "maxTokens": 128000 } }
  ]
}

Recorded state is the single source of truth:

• agent/status = running adds or refreshes the record; idle or an explicit disposal removes it.
• On SIGTERM / SIGINT the plugin snapshots every Agent that is still running before the process exits, so systemctl restart keeps the exact set that was live.
• On startup the systemd worker (DSH_WEB_DAEMON_WORKER=1 — the only process that owns the registry) loads the file and resumes each running: true record: the Agent is rebuilt with the recorded agentPreset and the model options the session was actually using before the restart — taken from the session log's latest request/header, not the creation-time snapshot stored in the registry — and one internal recovery notice is queued so the model continues the interrupted task instead of sitting idle.
• Legacy records without a running field are migrated once: they are resumed only when the persisted log ends in an interrupted/disposed turn.
• A foreground owner sharing DSH_HOME never resumes; active-sessions.lock (pid + boot identity + token) prevents two workers from racing.
• While recovery is in progress, Host API calls that would create or wake an Agent (session.prompt, session.models, …) wait behind the same barrier, so an auto-reconnecting browser cannot claim a session before resume.

Each boot writes recovery-diagnostics.json next to the registry, recording the lock result, every session's decision (resumed / skipped-already-live / skipped-not-running / failed + error), and the API calls the gate held:

cat "$DSH_HOME/plugins/dsh-plugin-web-daemon/recovery-diagnostics.json"
journalctl -u dsh-web.service -f   # look for "resumed session …"

Layout

| File | Content | | --- | --- | | index.js | Host half: systemd unit generation, session registry and resume lifecycle, CPU/memory/network/filesystem metrics sampling, settings namespace, JSON API, and the sidebar client-bundle patch. | | lib/client.js | Browser half: server status panel above New Session plus the Settings plugin-configuration card. | | cordis.patch.yml | Adds the host row and default configuration to the composed profile. |