dsh-healthcheck

Runtime health check for DeepSeek Harness (dsh). Read-only. Zero config.

A tiny plugin that helps you spot the two things that silently kill long dsh sessions: runaway session size (long-task OOM) and crash-torn session logs.

What it does

One tool: health_check

• health_check (no arguments) — scans your whole session library:
  • totals: sessions / size / events
  • the largest sessions
  • warnings for oversized sessions (≥10 MB watch, ≥30 MB danger) and event-count blowups (≥3k watch, ≥8k danger) — the usual precursors of long-task OOM
• health_check session_id="<id>" — deep-inspects one session log:
  • tail balance (an unclosed turn means a crash tail; dsh auto-repairs it on resume — this just tells you it happened)
  • tool-call / result pairing
  • event-type distribution

Everything is read-only — it never modifies a single byte of your data.

Install

# requires pnpm on PATH (dsh plugin forwards to pnpm)
dsh plugin --profile web add github:neufagents/dsh-healthcheck

Then restart dsh and ask your agent:

run a health check on my dsh sessions

Compatibility

• Tested with dsh 0.1.5-rc.2 (developer preview). The rc line moves fast; issues welcome.
• Use v0.1.1+. v0.1.0 shipped raw TypeScript sources, which Node refuses to load from node_modules (ERR_UNSUPPORTED_NODE_MODULES_TYPE_STRIPPING); v0.1.1 ships compiled JS in dist/.

Development

npm i
npm run build     # tsc -> dist/ (compiled JS is what gets shipped)
npm test          # 11 tests (pure functions + contract)
npx @deepseek-ai/dsh --profile web --patch ./dev.patch.yml --dump-config

License

MIT · made by NeufAgents