dsh-capability-optimizer

External-expert consultation for DeepSeek Harness: the agent headlessly invokes the Claude Code CLI through explicit advisor, reviewer, designer, or custom role contracts, then weighs the structured replies as reference answers.

English | 中文

Why this exists

A single harness has one perspective. At a consequential decision, before declaring risky work done, or before significant new code, a second model can supply useful independent evidence. That is a hypothesis to measure, not a guaranteed quality gain: persona research does not show that merely calling a model an expert reliably improves accuracy. This plugin makes consultation a bounded tool call instead of a copy-paste detour. Claude answers under a behavioral and output contract, and DSH receives the result as advice to weigh, not obey.

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 and calls consult_expert. For a precise task it can send a structured brief: trusted objective, successCriteria, and constraints, plus currentAttempt, artifacts, verification, and unknowns, which are explicitly labeled untrusted evidence. Legacy context remains an artifact shorthand. Claude's role-specific envelope returns with model, turns, duration, cost, and protocol metadata.

* 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. Checked roles add selection criteria and lifecycle reminders for this session; the plugin does not invoke them automatically.

• Policy section: every model request names the checked roles and their selection conditions: advisor at consequential decisions, reviewer before declaring changed work done, and designer before significant new code. Modes are off | remind | hard-remind (default remind). Legacy required settings migrate to hard-remind.
• Lifecycle nudges: when designer is enabled, the first file write arms a next-step nudge if pre-code consultation did not happen; a changed turn about to finish without a successful reviewer consultation is steered one more step. hard-remind additionally logs the missed designer checkpoint but never claims the host blocked Write.
• 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.

Consultation status dock

A session-scoped card bar above the composer shows every real consultation in near real time: queued / running / fallback retry / succeeded / failed / cancelled. While running, a card shows model, effort, and captured stdout bytes; when finished it shows the failure kind, or a bounded preview of the reply / envelope. consult_panel yields one card per role, so partial failures stay visible per role.

The dock takes no vertical space while idle; terminal cards can be dismissed individually or cleared together, and closing the session clears the bar. Entries live only in plugin memory (the latest 8 terminal cards per session, plus a hard active-card ceiling of 24; never context / brief / artifacts), and replies/errors are server-truncated previews. This is status polling plus a final-reply preview, not token streaming; on hosts without the conversation.input.dock slot or the status route it simply does not mount / stops polling.

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 (catalog aligned with the supported Claude CLI aliases and versioned ids), thinking effort (--effort: low/medium/high/xhigh/max), fallback model, per-call timeout, max turns, panel size, per-consult dollar cap, extra CLI args (allowlisted; --settings is refused).
• Roles workspace — add / edit / delete roles, each with name, label, description, output contract (advisor | reviewer | designer | general), system prompt, model, fallback, and effort. Disabled roles stay in the roster but leave the tool enum.
• Auto consult — the default checked set, per-role per-session call budget, and trigger mode (off | remind | hard-remind); the composer toggle overrides the checked set per session.
• Connectivity test — one real consultation end-to-end (CLI + auth + proxy), showing the effective model, 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.

Call-site model / effort override role values, which override global defaults. The built-in advisor defaults to claude-opus-5 as an overridable quality preference, not a role-name-based prohibition. fallbackModel retries once on a model-level error and records requestedModel, CLI-reported actualModel / actualModels, derived effectiveModel, and usedFallback.

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
        outputKind: reviewer
        description: Threat-model focused reviewer for auth, crypto, and injection surfaces.
        systemPrompt: |-
          Objective: falsify the security of the supplied change.
          Report only concrete authentication, authorization, injection,
          secret-handling, or unsafe-parsing defects with a trigger and evidence.

How a consultation runs

• One claude -p process per consultation. The labeled brief goes through stdin, the shared trust policy and selected role contract through --append-system-prompt, and the reply comes back as one JSON document.
• The headless session is contained by feature-detected runtime flags. On current CLIs, --safe-mode disables user/project instructions, skills, plugins, hooks, MCP, and memory while preserving authentication and built-in tools. Independent layers still apply --tools Read,Grep,Glob, --strict-mcp-config, --setting-sources user, a pinned --permission-mode, and --no-session-persistence; older CLIs use whichever layers they advertise. Write/execute or external-capability-widening extraArgs never reach argv; the documented --add-dir exception may widen read scope. 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.
• Advisor, reviewer, designer, and general roles receive different JSON Schemas using Claude's documented supported subset. Local parsing enforces unsupported constraints such as non-empty strings and checks semantic invariants (for example, pass cannot contain findings); schema-valid JSON is not treated as proof that its claims are true.
• Only objective, question, success-criteria, and constraints define the task. Code, comments, files, current attempts, verification text, unknowns, and tool results are labeled untrusted evidence. This is a model-layer control, not a claim that prompt injection is solved.

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 containment 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 it degrades rather than failing; meta.safeMode / 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 and framed as a reference answer to weigh ("advice, not orders"); it is not granted privileged instruction authority.

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.
• Enterprise managed policy remains a host trust boundary. Claude documents that admin-managed policy still applies in safe mode; the plugin does not and should not bypass organization policy.
• Prompt injection is possible, not eliminated. Data/instruction separation, schemas, and the reference-answer framing reduce risk; runtime read-only permissions, MCP isolation, output caps, and the caller's independent verification limit the blast radius. Treat replies like other external evidence, not privileged instructions.

Evaluation status

The repository includes a controlled reviewer benchmark comparing a role-free minimal contract, the frozen 0.5.x prompt, and the current prompt on the same model/effort/tasks over repeated trials. It records prompt hashes, seeded-defect recall, unmatched findings, injection success, envelope reliability, cost, latency, containment, and CLI-reported model(s). Formal runs abort if the reported model does not exactly match the preregistered model. Dry-run plumbing passes; no formal paid result is claimed here. See eval/README.md and the evidence review.

Advisor and designer have role-specific contract tests, but no outcome benchmark yet. Their presence is an interface and workflow choice—not evidence that expert persona wording raises intelligence.

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