mward4/deepseek-harness
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 32 Wiki Activity
Files
b3b0cd95cc9300cc4d4804eaaa7a866e250bfcc9
deepseek-harness/packages/subagent/subagent-claude-code
T
History

README.md

description, kind
description kind
The one-shot Claude Code subagent provider for users and maintainers choosing a product backend, installing a Profile bundle, or configuring an unattended Claude Code delegation. package-bundle

@deepseek-ai/dsh-subagent-claude-code

English | 中文

Summary

Install this Profile Bundle when a delegated task should run as a fresh, unattended Claude Code session in the parent workspace. Each run accepts one self-contained text task and returns the final answer or a safe failure diagnostic; reasoning, tool traffic, stderr, usage, and workspace diffs stay out of the parent Session. Native Claude settings and authentication remain authoritative, while Profile configuration selects the model, environment, and permissionMode. The platform-pinned runtime starts on demand and never falls back to the host claude executable. Choose it when isolation and genuine Claude Code behavior matter more than continuation or prompts.

Table of Contents

Use this package

Mount this provider when a delegation should run as a real Claude Code session in the parent's workspace. The common path is explicit: install the Bundle into a Profile, optionally configure the provider row, and expose it to the model through a delegation tool row.

Installing the Bundle

Install the package into the target Profile, then restart that Profile. The installation brings the pinned Agent SDK and one compatible platform CLI payload into the Profile; the declared patch layer registers only the dormant provider and starts no Claude process.

dsh plugin --profile <name> add @deepseek-ai/dsh-subagent-claude-code
dsh plugin --profile <name> remove @deepseek-ai/dsh-subagent-claude-code
dsh --profile <name>

Removing the package withdraws the provider and its private runtime closure on the next Profile start. Installation controls Host availability, not model permission: the model can only reach the provider through a delegation tool row you compose.

Configuration

Field Default Meaning
providerName claude-code Non-empty registry name on ctx.subagents; each mounted instance needs a unique value
model native Claude settings Optional non-empty model name fixed for every run from this provider instance; omission sends no SDK override
env {} Explicit SDK/CLI environment layered over the credential-scrubbed parent environment
permissionMode dontAsk Native non-interactive permission policy fixed for every run from this provider instance
disposeGraceMs 3000 Grace between the shared managed-range owner's termination tiers
permissionMode value Native behavior
dontAsk Deny operations that are not already authorized instead of prompting
acceptEdits Accept file edits; any remaining permission prompt is denied by the unattended callback
auto Let Claude Code's native classifier allow or deny permission requests
plan Run in native planning mode, deny execution approval, and return the completed plan as the final answer
bypassPermissions Explicitly set the SDK's dangerous confirmation and bypass permission checks

The generated configuration catalog is the exhaustive source for every accepted field and its JSDoc. A configured model passes unchanged to every query from that provider instance; omission leaves native model selection in force. Credential-shaped ambient variables are removed before the explicit env overlay, so an API key intended for the child must be supplied there. The provider omits the SDK settingSources option, so Claude Code reads the host's normal user, project, and local settings relative to the parent Session cwd. It does not copy or filter those files, create or modify login state, inspect PATH, or fall back to a host claude executable.

Exposing the tool

Each delegation tool row names one provider and needs its own toolName, so the model sees static tools rather than a dynamic provider selector. Full Agent Presets carry a matching default tool row with disabled: true; copy a preset and remove that field to expose subagent_claude_code only to agents composed from the copy.

- id: tool-subagent-claude
  name: '@deepseek-ai/dsh-tool-subagent'
  config:
    provider: claude-code
    toolName: subagent_claude_code
    maxDepth: provider-managed

The tool returns a child id after accepting the task and sends its result to the parent Agent on completion. Each external activation executes once and accepts no follow-up input or resume.

What you get

The completion notice contains the final Claude Code answer, or the stop reason and optional safe diagnostic. The parent Session also retains the external task identity and complete terminal result independently of the notice; no local child Session is created. Product reasoning, tool activity, raw stderr, and workspace diffs do not enter the parent Session.

Failure and recovery

An install that omits optional dependencies, uses an unsupported platform, or loses the selected payload leaves the provider dormant and fails the first delegation at the SDK startup boundary with a safe query-start / unknown failure fact; there is no host-CLI fallback. The original product error stays on the internal cause chain and in the provider's Host log. A cancelled run settles as aborted.

Understand the implementation

Implementation internals — click to expand

This section explains how the provider drives a real Claude Code CLI and where the observable behavior comes from; the full contract lives in Use this package.

Design concept

  • One fresh query per run. Every run has an independent SDK query, cancellation controller, CLI process, and non-persisted product session; there is no continuation, resume, or pooling.
  • Native settings are authoritative. The provider deliberately omits the SDK settingSources option, so Claude Code reads the host's normal user, project, and local settings; an optional model and the required permissionMode are the only query-level overrides.
  • Unattended by design. AskUserQuestion is disabled and permission prompts are denied except in bypass mode, so the query never waits for a user interface.

Source map

File Role
src/index.ts Plugin entry: config schema, provider registration
src/run.ts The SDK query lifecycle, result acceptance, and permission handling
src/process.ts Managed-range termination escalation on disposal
cordis.patch.yml The Profile patch layer that registers the dormant provider

Run flow

A start accepts only a non-empty sequence of text blocks and derives the child cwd from the parent session. It creates a private AbortController, calls the official SDK query() with the exact concatenated task, and publishes the run only after the SDK's custom-spawn hook has supplied a live CLI handle owned by the subprocess seam. The provider iterates the complete message stream and accepts only a result message with subtype: "success", is_error: false, and a nonblank result, followed by normal iterator completion. Every other outcome maps to a fixed-category error diagnostic naming the lifecycle stage and observed process outcome — the category set lives in src/run.ts. Local cancellation wins the result race and maps to aborted without a failure diagnostic.

Further Exploration

Read these pages when the package-level contract is not enough. They move from this provider to the seam it plugs into and the sibling product provider.

Model Experience

Child request

What the model sees

The Claude Code child receives the standalone text task as one fresh SDK query. Its workspace is the parent Session cwd; the selected provider instance fixes the query's configured model, environment, and non-interactive permission mode, while an omitted model and every other product setting come from native Claude configuration. The executable version comes from the Bundle's pinned SDK platform payload.

Token effect

The child pays for an independent Claude Code context and query. Child tokens do not enter the parent's context.

KV Cache effect

Independent of the parent request cache. Reuse depends only on Claude Code's own model, instructions, tools, native settings, and fresh query.

Parent scheduling and results, indirectly

What the model sees

Through dsh-tool-subagent, the parent model first receives a child id, then the final Claude Code answer or a failure notice with its stop reason and safe diagnostic. Diagnostics contain only fixed stage, category, and observed protocol or process facts. Product reasoning, intermediate messages, tool activity, stderr, usage, product identifiers, commands, paths, and raw protocol payloads are not copied into the parent Session.

Token effect

Parent input grows by the start acknowledgement and completion notice, including the final answer or failure detail. Child tokens do not enter the parent context. This provider adds no parent tool schema by itself.

KV Cache effect

Start acknowledgements and completion notices append after the reusable parent request prefix. A notice may wake another turn without rewriting the existing prefix.

Known Limitations and Deferred Work

These limits define when this provider is a poor fit or needs special operational care. They are current package constraints, not a general Claude Code comparison or a task backlog.

  • One fresh query and process per run — there is no continuation, resume, pooling, progress stream, or product-session persistence.
  • Static instance selection — Profile rows fix provider names, optional models, and tool bindings; calls cannot choose or change either a provider or model dynamically, and every exposed tool needs a unique toolName.
  • Host settings are intentionally authoritative — when model is omitted, project and user settings choose it; native settings always retain the remaining tools and behavior, and the provider does not provide a filtered or hermetic production mode.
  • Authentication and account state remain native — the Bundle supplies the CLI but does not create an account, log in, or rewrite Claude settings; configuration and authentication failures surface with their lifecycle stage and the safe unknown fallback rather than a separate public classification.
  • The SDK platform payload is required at delegation time — installs that omit optional dependencies, unsupported platforms, and missing or damaged payloads fail at the first query; there is no host-CLI fallback.
  • No human interaction path — AskUserQuestion is disabled, permission prompts are denied, MCP elicitation is declined, and blocking dialogs fail closed instead of suspending.
  • Assistant payload is final text only — failed runs can also expose a separate safe diagnostic; reasoning, intermediate messages, tool traffic, usage, stderr, and workspace diffs remain outside the parent Session. Task identity and terminal results are retained in the parent log.
  • No optional shared capabilities — agentOptions, output schemas, child personas, tool filtering, and harness depth enforcement are rejected by the shared service for this provider.
  • No wall-clock timeout or side-effect rollback — the caller cancels long work, and files or external systems changed before cancellation are not restored.

Dev Note

Working context for maintainers — click to expand

This Dev Note is working context for maintainers: open questions and undecided directions. It is explicitly non-authoritative — shipped behavior and limits live in the sections above and in the package code.

  • Payload size disclosure — the current darwin-arm64 platform payload packs to about 92 MB and unpacks to about 325 MB; these are disclosure numbers, not installation thresholds.
  • Version-pinned protocol — the runtime dependency is pinned to Agent SDK 0.3.263; upgrading pins a new SDK version and requires re-running the keyless real-product and loader-composition evidence.

Runtime invariant: No companion is published. Lifecycle pairing belongs to the shared subagent service, and managed-range ownership belongs to the subprocess service.

Reference in new issue View Git Blame Copy Permalink