dsh-capability-optimizer

External-expert consultation for DeepSeek Harness: the agent headlessly invokes the Claude Code CLI with role personas — advisor, reviewer, designer, or your own — and weighs the subscription model's replies as reference answers.

English | 中文

Why this exists

A single harness has one perspective. When the DSH agent hits a decision point, finishes risky work, or starts significant new code, a second opinion from a different model — Claude, on your existing Claude Code subscription — is cheap insurance and a genuine quality lift. This plugin makes that a first-class tool call instead of a copy-paste detour: the agent consults, Claude answers in role, and the reply is framed as advice to weigh, not obey (the same contract Oh My Pi's advisors use).

Phase 1 speaks only to the Claude Code CLI. The settings schema (v2, one workspace per harness), the UI tab catalog, and the runner seam are already multi-backend: codex, zcode, kimi-code, pi, opencode, and omp each land later as a runner behind the same three tools.

Features

Install

# from npm (recommended)
dsh plugin --profile web add dsh-capability-optimizer

# or straight from the GitHub repo
dsh plugin --profile web add github:hezhongtang/dsh-capability-optimizer

Restart dsh web (or your profile of choice). Works in any profile — web, tui, headless — because the tools are host-side agent tools.

Requirements: the claude CLI (npm i -g @anthropic-ai/claude-code) on PATH, logged in.

Usage

Ask your agent:

"consult the reviewer on this diff before we call it done"

The agent picks the role, packs the material into context, and calls consult_expert. Claude's reply returns as a reference answer with run metadata (session, turns, duration, cost) — advice the agent weighs, not an order it obeys.

* Read-only for your workspace; each call spends your Claude subscription quota — the tool descriptions themselves tell the model to batch material instead of machine-gunning calls.

Auto consult

The composer toolbar (the permissions control's row) carries an Expert Consult toggle: check roles for this session and the host proactively works them into the loop — no prompting, no re-explaining.

• Policy section: every model request carries a short policy block naming the checked roles and when each applies (advisor at decision points, reviewer before declaring work done, designer after the first file write of a turn, on the next step). Auto-consult mode is off | remind | required (default remind).
• Lifecycle nudges: the first file write of a turn arms the designer anchor (the nudge rides the next step of that turn); a turn that changed files and is about to finish without a reviewer pass is steered one more step to consult first.
• Budget: capPerRole (default 3) counts real consult_* calls per role per session — nudges and the model's own discretionary calls share it. At the cap the promise drops out of the policy text and the anchors go quiet.
• Soft by design: a nudge guarantees the instruction is delivered, never the tool call — dsh has no forced-call API. A model that declines must state the reason in one line.
• The popover shows live usage counts (used/cap) per role; the last selection is remembered per browser. Settings → Expert Consult → Auto consult edits the default checked set and budget (row-config key autoConsult) — the same layer tui/headless profiles consume.

Settings UI

Settings → Expert Consult is organized as one workspace per harness CLI — a tab bar over the catalog (claude-code live; codex, zcode, kimi-code, pi, opencode, omp reserved with a planned-status page and no settings stored until their runners land). The Claude Code workspace manages everything at runtime:

• General — CLI path, default model (full catalog: follow-CLI-default, latest aliases, and versioned ids like claude-opus-5 — extracted from the CLI itself), thinking effort (--effort: low/medium/high/xhigh/max), fallback model, per-call timeout, max turns, panel size, extra CLI args.
• Roles workspace — add / edit / delete roles, each with name, label, description, system prompt, a dedicated model, a dedicated fallback, and a dedicated thinking effort. A role's switch disables it omp-style: it stays in the roster but leaves the tools' enum until re-enabled.
• Auto consult — the default checked set plus the per-role per-session call budget; the composer toggle overrides it per session.
• Connectivity test — one real consultation end-to-end (CLI + auth + proxy), with turns, duration, cost, and a fallback-used marker.
• Save & apply persists to ~/.dsh/dsh-capability-optimizer/settings.json (atomic writes, 0600) and hot-applies: the agent tools re-register immediately. Reset removes the file and restores defaults.

Per-role model and effort beat the global defaults; a call-site effort argument beats both. fallbackModel (role-level or global) retries once when Claude fails with a model-level error (unrecognized_model, model-not-found, ...), recording usedFallback in the run metadata.

Configure (composition layer)

The row's config still works as the base layer (settings file wins once saved):

Example — a security-focused custom role:

- id: dsh-capability-optimizer
  name: 'dsh-capability-optimizer'
  config:
    model: sonnet
    roles:
      - name: security
        description: Threat-model focused reviewer for auth, crypto, and injection surfaces.
        systemPrompt: |-
          Role: security reviewer.
          Threat-model the material: authentication, authorization, injection,
          secrets handling, and unsafe parsing. Rate each finding by exploitability.

How a consultation runs

• One claude -p process per consultation; the question (plus optional material) goes in via stdin, the role persona via --append-system-prompt, the reply comes back as one JSON document.
• The headless session is contained by five feature-detected flags: --tools Read,Grep,Glob (built-in tools), --strict-mcp-config (MCP tools, which --tools does not govern), --setting-sources user (drops the consulted repo's own settings), a pinned --permission-mode (outranks permissions.defaultMode from any settings source) and --no-session-persistence. Permission-widening extraArgs never reach argv. Caller cancellation (AbortSignal) and the wall-clock timeout stay distinguishable.
• Wall-clock timeout (default 5 min) with SIGTERM → SIGKILL escalation; --max-turns (default 8) caps agentic turns inside the CLI.
• Every reply carries a shared framing for Claude — this is a reference answer another agent will weigh — so even custom roles inherit the "advice, not orders" contract.

Security & data flow

What the plugin guarantees:

• The plugin never passes --dangerously-skip-permissions or any permission-bypassing flag. extraArgs is an allowlist, not a passthrough. On a CLI that advertises the flags, a consultation gets exactly Read, Grep, Glob and no MCP server — verified against a real CLI, including against a project whose own .claude/settings.json asks for bypassPermissions (evidence, reproducible with DCO_LIVE_CLI=1 npm test). Each flag is feature-detected, so a CLI too old to advertise them degrades rather than failing; meta.tools / meta.strictMcp / meta.permissionMode report what was actually enforced.
• Prompts travel as argv/stdin to the local CLI only — the plugin itself adds no third-party service, no telemetry, no credential storage. Routes enforce same-origin; the settings file is 0600 and atomically written; subprocess output is size-capped and timeouts always reap the child.
• Claude's reply is returned to the DSH agent as tool-result data, framed as a reference answer to weigh ("advice, not orders"), never as an instruction channel.

What you should know (inherent to any agent-consults-agent setup):

• Your material leaves this machine to your own Claude account. The question plus any code/diff/plan passed as context is processed by the Claude Code CLI under the account you logged in with — the same data flow as running claude -p yourself. Do not paste secrets you would not send to Claude directly.
• Prompt injection is possible, not eliminated. If consulted material (e.g. a malicious file Claude reads from the workspace) manipulates its reply, that reply reaches the DSH agent as text. The reference-answer framing and the "weigh, don't blindly obey" contract mitigate this, but treat expert replies with the same skepticism as web search results — the same residual risk class as every dual-model workflow.

Limitations

• Phase 1 is Claude Code only; the multi-backend settings schema (v2, one workspace per harness) and UI tabs for codex / zcode / kimi-code / pi / opencode / omp are already in place — each lands as a runner behind the same tools.
• Each consultation spends your Claude subscription quota; the tool descriptions tell the model to batch material instead of machine-gunning calls.
• No streaming — one JSON result per call.

Contributing

Issues and PRs welcome at hezhongtang/dsh-capability-optimizer. The codebase is intentionally small and dependency-free — plain ESM on the host, a hand-authored CJS bundle in the browser, no build step to set up. Adding a harness backend = one runner module (see lib/claude.js) + flipping available in lib/backends.js.

License

MIT © 2026 hezhongtang