dsh-advisor

Turn advisor for DSH — evaluates each turn (default: on error) via LLM and renders a full-width light-blue advice card in the chat.

A host+client bundle plugin for DeepSeek Harness. The host half listens to session/event for turn/end, calls ctx.llm.stream() with a configurable prompt, and appends advisor/eval (log-only). The browser half registers a ConversationNodeDefinition for advisor/eval and a keyed conversation.chat.node renderer (advisor) — a full-width card in harmonious blue (#e3eefc/#3b6ea5, no traffic-light colors) with verdict icon differentiation.

• Log-only evaluation. The verdict is appended as advisor/eval — a log-only node with no surfaceOp, so the agent never sees it.
• Errors by default. evaluateOn: ['error'] (default) means only failed turns cost an LLM call; ['all'] opts into every turn.
• Harmonious, not alarming. One blue family (#e3eefc / #3b6ea5) with icon-only verdict differentiation — no red/yellow/green traffic lights.
• One-line install. Built lib/ is committed, so dsh plugin add works without a build step.

Install

dsh plugin --profile web add github:haoliangwu/dsh-advisor

Built lib/ is committed, so the git install is one line — no prepare script, no allowBuilds permission. Restart dsh --profile web after install (bundle layer stacks compose at boot).

Configure

In your profile's cordis.patch.yml (~/.dsh/profiles/web/cordis.patch.yml):

- id: dsh-advisor
  config:
    # When to evaluate: ['error'] (default) or ['all']
    evaluateOn: ['error']
    # Optional explicit route; fallback is session's current model (request/header)
    provider: deepseek-official
    model: deepseek-chat
    # Optional prompt template with {turn} and {transcript} placeholders
    promptTemplate: |
      Evaluate turn {turn}...
      Transcript:
      {transcript}

evaluateOn accepts any turn/end reason kind (completed, error, aborted, blocked, max-tokens, interrupted) or all. An empty list disables evaluation.

promptTemplate defaults to a hard-coded JSON template that requests {"verdict":"ok"|"needs-attention"|"off-track","issues":[],"advice":"..."}. JSON parse failures degrade to needs-attention with advice set to raw output.

Disable entirely:

- id: dsh-advisor
  disabled: true

How it works

• Host: ctx.on('session/event') filters turn/end by evaluateOn; builds a short transcript from session.deriveMessages() (last 10) and the requestHeader/requestContext route (or explicit provider/model); calls ctx.llm.stream() and parses verdict/issues/advice; appends advisor/eval log-only (no surfaceOp, ignorable).
• Client: ConversationNodeDefinition (kind: advisor, target: chat) matches advisor/eval per turn and builds a chatNode('advisor', …); AdvisorCard renders a full-width light-blue card with blue-family icons (○ / ◐ / ◉) and locale-aware verdict labels (advisor namespace, zh/en).
• Card styling: background #e3eefc, border #c3dbf5, text #3b6ea5 family; icon variants differentiate verdicts within the same hue (light fill → solid fill), never red/yellow/green.

Build from source

pnpm install
pnpm build         # emits lib/index.js, lib/invariant.js, lib/client.js + sourcemaps
pnpm typecheck     # tsc --noEmit

lib/ is committed so git installs work without a build step. After changing source, run pnpm build and commit the updated lib/.

License

MIT