@buberlo/dsh-jev

@buberlo/dsh-jev

The Jev decision layer plugin and bundle for DeepSeek Harness (DSH).

DSH plans, calls tools, and runs them. Jev makes the small, fast decisions in between — and this plugin maps those decisions onto supported DSH behavior. It can only narrow or gate, never widen a permission.

pre-step  ──▶ Jev: which tools are relevant?   ──▶ tools.restrict
tool call ──▶ Jev: is this call safe to run?   ──▶ ask / hold / deny
result    ──▶ deterministic loop guard (no model)
request   ──▶ Jev: which configured model?     ──▶ verified route

Install

dsh plugin --profile  add @buberlo/dsh-jev
dsh --profile  --dump-config   # shows the "# == @buberlo/dsh-jev" layer

The bundle inserts one row; defaults are provider: mock + mode: shadow: deterministic, offline, and behavior-neutral.

Configure

- id: jev
  name: '@buberlo/dsh-jev'
  config:
    provider: mock            # mock | live
    mode: shadow              # off | shadow | enforce
    selection:                # dynamic tool preselection
      enabled: true
      alwaysAllow: [read_file]
      categories:
        files: 'Reading or writing workspace files'
        tests: 'Running or inspecting tests'
      toolCategories:
        read_file: [files]
        run_tests: [tests]
    assessment:               # per-call semantic check
      enabled: true
      onFailure: ask          # ask | hold  (never an allow)
    skills:
      enabled: true
    modelRouting:
      enabled: false
      routes:
        fast: { provider: deepseek-official, model: deepseek-v4-flash }
        reasoning: { provider: deepseek-official, model: deepseek-v4-pro }

Live API — two explicit settings, never implicit:

    provider: live
    mode: shadow              # live + shadow still transmits state
    apiKey: !!js process.env.TYPESAFE_API_KEY

Web client

In a web/desktop profile the bundle ships a configuration page on the Plugins page: provider, live mode selector, a write-only API key field, and the five feature toggles. Writes go to the host settings document and reconfigure the running service immediately.

What it does

Extension point	Decision
agent/pre-step	tool selection across independent categories; scoped tools.restrict, recomputed every turn
tools/pre-execute	call assessment: task fit, missing information, explicit restriction conflict
tools/result + tools/guard	bounded loop detection; identical repeated calls are denied
agent/request	model routing to a configured route, only when verified available
ctx.skills	at most one skill hint per turn; bodies are never auto-loaded

Guarantees

• No implicit live access. provider: live without an explicit key fails at plugin load.
• No model-derived permissions. Failures, timeouts, stale snapshots, and validation errors produce ask/hold.
• No premature allow. The listener always delegates first and composes monotonically; later policies are never skipped.
• No widening. A Jev selection intersects with existing restrictions.
• No stale decisions, no cache. Decisions are bound to a turn/catalog snapshot; changed arguments are assessed again.
• Bounded data. Only task text, tool metadata, and bounded/redacted arguments leave the process.

Documentation

• Project:
• docs/policy.md · docs/architecture.md · docs/upstream-compatibility.md
• docs/getting-started.md · docs/skills.md · docs/publishing.md

License

MIT