Loading ToolBench
DeepSeek Harness plugin: dispatch work to DSH agents from Claude Code / Codex, as native subagents with live progress
dsh-crew provides 4 tools with detailed descriptions and structured schemas. Tool naming follows verb_noun convention (dsh_run_worker, dsh_worker_config, dsh_worker_status, dsh_cancel_worker). Descriptions are comprehensive (200-400 chars), exceeding baseline averages. However, critical gaps exist: no error handling guidance (tools lack recovery hints or error classification), no output schemas documented, and no tool annotations (readOnlyHint/destructiveHint). Parameters are well-typed with enums and constraints, but descriptions lack actionable format guidance. The server is domain-specific (DSH worker dispatch) and well-scoped, but lacks production-grade error resilience patterns.
Cancel a running worker job by ID. Sends SIGTERM to the worker process (or hub session) and waits for it to settle. Returns the final job view.
Run one DSH (DeepSeek Harness) coding agent task and block until it finishes. Tier guidance: flash for mechanical, well-scoped work (single-file edits, lookups, formatting); pro for multi-file changes, debugging, and design judgment. The task string must be self-contained: the worker has zero conversation context, so include absolute paths, exact acceptance criteria, and constraints. Set worker="agy" or "grok" only when the user explicitly asks for that CLI, never as a default. Guardrails refuse (never queue) worker→worker recursion beyond the origin-chain depth cap and a second writer on the same cwd; allow_concurrent_cwd: true only for read-only tasks.
Get and set the session-level worker dispatch configuration: default tier, reasoning effort, execution mode (hub vs standalone), timeout, tier policy, escalation on failure, agent presets per tier, LLM provider/model bindings, and available worker profiles. Returns the current config and the list of named CLI worker profiles (agy, grok, etc.) with their backend, model, effort, and permission metadata.
List all running and recently finished worker jobs in this session, with their status, progress (turn/step/tool calls), token usage, and result/error. Includes workspace grouping, cwd locks held by running jobs, and orphan detection (jobs that vanished from the status source without reaching a terminal state).
No output schemas documented. Tools describe inputs well but omit return type definitions. LLMs cannot plan downstream calls or extract required fields (e.g., job_id from dsh_run_worker result).
No error handling guidance. Tools lack recovery hints (e.g., 'If timeout occurs, retry with longer timeout_seconds'). Errors are not classified as retryable, user-fixable, or fatal.
No tool annotations. dsh_run_worker and dsh_cancel_worker are destructive (WRITE risk) but lack destructiveHint annotation. dsh_worker_status is read-only but lacks readOnlyHint. Agents cannot infer safety properties.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | <=2025-11-25 | v2 |
Parameter descriptions lack actionable format guidance. E.g., 'task' param says 'self-contained' but does not specify max length, required sections, or syntax rules. 'timeout_seconds' says 'positive, max 7200' but omits minimum.
dsh_worker_config accepts a partial config object but does not document the schema of that object or which fields are mutable. LLMs cannot determine valid config keys or types.