AI Agent Hub
Back to plugins
⚙️

dsh-session-supervisor

Workflow Updated 2026.08.14

Run the following command in DeepSeek Harness:

dsh plugin install acosmi/dsh-session-supervisor

Paste the following prompt into your AI chat to install this plugin:

Please help me install the dsh-session-supervisor plugin (source at https://github.com/acosmi/dsh-session-supervisor) using dsh plugin install acosmi/dsh-session-supervisor, then explain its guard policies and the guardian toolset.

About this plugin

When a DeepSeek Harness root Agent session runs unattended, it can silently slip into inactivity, blow past a hard deadline, or rack up a streak of error, blocked, max-tokens, or interrupted turn-end events—without anyone noticing. dsh-session-supervisor was built for exactly this gap: it attaches a read-only lifecycle supervision contract to an active session and turns the vague feeling of "something is off" into structured, auditable, notifiable facts.

Driven by periodic evaluation, the plugin watches three policies simultaneously—lifecycle_silence (no qualifying durable lifecycle event for N consecutive seconds), deadline_unclosed (an RFC 3339 absolute deadline has passed while its Guard remains unclosed), and abnormal_turn_streak (K consecutive turn/end events ending in error, blocked, max-tokens, or interrupted). The first confirmed breach within a grace window becomes a persistent incident with bounded evidence, and the plugin queues at most one follow-up notification to the owner session when configured to do so. Completed turns and fresh activity serve as recovery evidence; acknowledge and resolve are explicit, auditable actions. The plugin strictly observes and notifies: it never executes shell commands, sends network requests, retries tools, or cancels your Agent. Time elapsing is not an authorization.

It is designed for teams and individuals who let Harness Agent sessions run for extended periods and want a quiet sentinel in the background. Declare the supervision contract, and the plugin watches lifecycle facts for you—surfacing a stable incident id and a structured notification to the owner session when something goes wrong, making deduplication, auditing, and downstream handling straightforward. All plugin state lives in its own directory, is uninstallable at any time, writes nothing to the session log, and does not affect session recovery.

Use Cases

  • Root Agent sessions running unattended need continuous monitoring for silence, missed deadlines, and abnormal turn streaks
  • Multiple consecutive turns ending in error or blocked should automatically form an incident and notify the owner
  • Teams need auditable, traceable lifecycle facts for background sessions without altering session behavior

Best For

  • Developers or teams running Harness Agents autonomously for extended periods
  • Ops or architecture staff who need auditable, traceable session lifecycle facts
  • Individuals who want anomaly notifications without the plugin executing any action