Subagent forge

Replaces the built-in delegation tools with one configurable tool: the calling agent writes the child's system prompt, optionally its first user message, picks its model from a configurable allowlist (with descriptions the agent reads), sets a reasoning effort, and decides whether that child may start subagents of its own.

# from a GitHub repository
dsh plugin --profile <profile> add "github:<owner>/dsh-subagent-forge"

# from a local checkout
dsh plugin --profile <profile> add /absolute/path/to/dsh-subagent-forge

Requires DeepSeek Harness >=0.2.1 (engines.dsh). No build step: the package ships plain ESM.

Tools

Two symmetric tools replace the built-in pair:

Tool	Provider	Child
spawn_agent	spawn	fresh session, gets your system_prompt
fork_agent	fork	seeded with this conversation's completed turns; inherits the system prompt and tool set, so it takes no system_prompt

Shared parameters:

Parameter	Meaning
description	short label, required
prompt	first user message / task; omitted uses emptyPrompt
system_prompt	only spawn_agent: durable role and rules for this child
model	provider/model, exactly one allowlist route
reasoning_effort	reasoning effort advertised for that model (effort is an alias)
can_delegate	required: whether the child may start its own subagents
cwd	initial child working directory

Recursion

There is no depth counter. A child sees the delegation tools only when its direct parent passed can_delegate: true; a child that receives the grant decides the same question for its own children. Children without the grant also lose workflow, because it can start agents indirectly.

Built-in delegation is blocked twice

1. tools.guard denies execution of subagent / subagent_fork in every scope, including steps whose tool list was resolved before this plugin loaded.
2. tools.restrict hides them from each agent's tool list, so the model does not see a tool it cannot use.

Both lists are configuration (blockedTools, blockBuiltinDelegation).

Settings

The page lives in Settings → Built-in plugins → Subagent forge (a tab in the Plugins settings section). It edits the model allowlist: one row per route with the description the agent reads when it chooses.

Edits are validated, applied to the live tool at once, and stored in $DSH_HOME/subagent-forge/config.json, which takes precedence over the bundle patch on the next start. The write path is:

page → POST /subagent-forge/config (loopback only)
     → reconfigure(value): Config(value) → file → re-register the tool

The route is registered by the second bundle row (subagent-forge-bridge).

Durability

Each child's spec is stored as JSON under $DSH_HOME/subagent-forge/children/<sessionId>.json and re-applied when the child resumes (the plugin re-registers the persona section and the restrictions on agent/created, which fires for resumes too). Losing that directory does not break a child; it only loses the custom persona on later resumes.

Known limits

• The child's composition (tool set) is inherited from the parent; only the persona and the delegation grant are overridden per child.
• workflow is not blocked for agents that hold the grant.
• Changing code of the installed package needs an application restart; the settings page itself needs one page refresh after installation.
• After installation the browser must be reloaded once to load the client half.