DeepSeek Harness subagents: DSH plugin delegation providers

The DeepSeek Harness subagent seam lets one agent delegate work to child agents — it is an optional capability, not part of the agent loop, and a single context can host multiple providers registered by name in ctx.subagents, so you can pick a different backend per task (source).

What a DeepSeek Harness subagent is: delegating work outward

A subagent is outsourcing a relatively self-contained piece of work to another agent; unlike a seam such as bash, which allows only one executor, subagents can host multiple implementations in one context (source). Key points:

1. It is an optional capability — not part of the agent loop, so its type definitions are not in core. Expected: things work fine without it; only using it requires the matching provider.
2. Registered by name, many implementations coexist — the registry follows the pattern of an LLM adapter registry. Expected: you can mount several backends at once and choose by name.
3. Two kinds of model-facing consumer — dsh-tool-subagent delegates by provider; dsh-tool-subagent-control optionally provides global send_message, interrupt_agent, and list_agents control tools. Expected: the former starts delegation, the latter controls it.
4. Parent-child relationships are discoverable — the same ctx.subagents service discovers direct children through a parent catalog and descendants recursively through parent catalogs. Expected: you can trace a delegation tree, not just one level.

Which DeepSeek Harness subagent providers you can choose

The official docs give six sibling packages as service providers, covering in-process spawning, fork, and backends for ACP / Codex / Claude Code / DeepSeek Harness SDK (source). Meet them one by one:

1. dsh-subagent-spawn-in-process — in-process spawning. Expected: lightweight and fast to start, good for short delegations.
2. dsh-subagent-fork-in-process — in-process fork that inherits the current directory context. Expected: use it when you want to carry over the current working directory and state.
3. dsh-subagent-acp — for ACP backends. Expected: connect an agent that supports ACP.
4. dsh-subagent-codex — for Codex. Expected: reuse Codex as a subagent.
5. dsh-subagent-claude-code — for Claude Code. Expected: reuse Claude Code as a subagent.
6. dsh-subagent-dsh-sdk — for the DeepSeek Harness SDK. Expected: integrate with other DeepSeek Harness setups; once registered it is callable by name like any other backend.

Mounting a backend usually means getting its plugin into your config tree. To find community subagent-related plugins, browse DSH Plugin Hub.

One-shot vs continuable subagents, and capability mismatch errors

A provider declares its startup capabilities through a capability descriptor, and the service validates before delegating to start() — a request that depends on a capability the provider lacks is explicitly rejected rather than silently ignored (source). Key points:

1. Tell the two paths apart — one-shot delegation goes through SubagentProvider.start(), where the provider composes the child; continuable subagents are composed by the continuable execution manager through prepareContinuable. Expected: use the continuable path when you need to keep talking.
2. Five capability switches — SubagentCapabilities contains agentOptions, outputSchema, depthLimit, toolFilter, and persona, matching the startup request options one to one. Expected: check what a provider supports before choosing it.
3. Unsupported means an error — a missing required capability throws SubagentError('UNSUPPORTED_CAPABILITY'). Expected: following fail-loud rather than silent degradation surfaces problems early.
4. List existing subagents — the subagentCatalog projection extracts SubagentCatalogEntry[] from the session event stream, each with child id, creation time, mode, and labels; historical entries whose mode cannot be determined are marked mode: 'unknown'. Expected: you can list the subagents a session dispatched.

Notes and common questions

1. Align capabilities first: different providers support different capabilities, so checking SubagentCapabilities before sending a request saves one error.
2. Do not treat one-shot as continuable: if you need follow-up questions, use the continuable path, or the child conversation cannot be resumed.
3. Historical entries may have an unknown mode: mode: 'unknown' preserves identity display but does not guarantee continuable execution; do not retry based on it.
4. Subagents also create sessions and jobs: long delegations involve background jobs and session records too; see Handle long-running DeepSeek Harness tasks and Where DeepSeek Harness sessions are stored and how to query them.
5. Control tools are optional: send_message, interrupt_agent, and list_agents come from the optional dsh-tool-subagent-control; without it those tools are absent.

Sources: Subagent (official docs), Tool schema catalog (official docs)