@argszero/cordis-plugin-subagent-effort

Per-delegation reasoning effort for deepseek-harness (dsh) subagents.

⚠️ Superseded upstream (0.1.2-alpha.1, 2026-08-28)

deepseek-harness 0.1.2-alpha.1 fixed this gap natively: AgentOptions now carries reasoningEffort (consumed by the agent-loop request builder), and dsh-tool-subagent exposes a per-call reasoning_effort parameter plus background/continuable delegation. This plugin is not needed on ≥ 0.1.2-alpha.1 and is kept for users pinned to 0.1.1-rc.2, where the silent-drop gap still exists. See discussion #4666 for the maintainer-confirmed root cause and the upstream fix.

Closes one gap end to end: @deepseek-ai/dsh-tool-subagent accepts config-level agentOptions and forwards them into the child's AgentOptions, but the agent loop's request builder only reads provider / model / maxTokens from AgentOptions — a reasoningEffort declared there was silently ignored (the model never saw it, no error). This plugin fixes the consumption side and adds a per-call effort tool.

What it does

1. Effort plumbing — extends the merge-extensible AgentOptions interface with reasoningEffort and injects it into every request of that agent via the agent/request waterfall. Config-declared effort (e.g. tool-subagent's agentOptions.reasoningEffort) now actually reaches the adapter.
2. Per-call tool — registers subagent_effort (configurable), a foreground subagent tool accepting reasoning_effort (off | low | high | max, the DeepSeek adapter vocabulary) and an optional model override per call, forwarded as child agentOptions.

Install

# add to your dsh profile (the dsh.bundle declaration mounts it as a profile layer)
pnpm add @argszero/cordis-plugin-subagent-effort

Set the provider in your profile's cordis.patch.yml (last write wins per row):

- set:
    - id: subagent-effort
      config:
        provider: spawn          # your ctx.subagents provider name
        # toolName: subagent_effort
        # defaultReasoningEffort: high

Config

key	type	default	meaning
provider	string	— (required)	the ctx.subagents provider name to delegate on (e.g. spawn, fork, acp)
toolName	string	subagent_effort	model-facing tool name
defaultReasoningEffort	'off' | 'low' | 'high' | 'max'	provider default	effort applied when a call omits reasoning_effort

Usage

Config-level effort (fixes the silent drop)

With @deepseek-ai/dsh-tool-subagent configured with agentOptions, the effort now reaches the child's requests:

- set:
    - id: subagent
      config:
        provider: spawn
        agentOptions:
          reasoningEffort: high

Without this plugin that value was accepted and silently ignored.

Per-call effort (new tool)

The subagent_effort tool lets the model vary the thinking budget per delegation:

{
  "description": "verify the proof",
  "prompt": "Check the induction step carefully…",
  "reasoning_effort": "max"
}

Optional model overrides the child's model for that delegation.

Design notes

• The tool mirrors @deepseek-ai/dsh-tool-subagent's provider lifecycle: it registers when the configured provider appears and unregisters when it leaves.
• The agent/request listener is conservative: only agents whose options explicitly carry reasoningEffort are touched; every other request passes through unchanged.
• Foreground-only by design — background/continuable delegation is available from the built-in subagent tool; this plugin is about effort control.

Peer range

This plugin targets one dsh line, and declares every harness package it reads as a peer rather than a dependency:

>=0.1.1-rc.2 <0.1.2

declared identically on @deepseek-ai/dsh-agent, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-session, @deepseek-ai/dsh-subagent and @deepseek-ai/dsh-tools. It admits exactly 0.1.1-rc.2.

The narrow scope is deliberate: upstream closed this gap natively in 0.1.2-alpha.1 (AgentOptions.reasoningEffort is consumed from there), so the plugin is only meaningful on the line where the silent drop still exists.

The declaration was the bug. v0.1.0 shipped these five as dependencies at ^0.1.1-rc.2, and a caret on a prerelease admits only its own tuple — so no newer tree could satisfy it, and npm resolves an unsatisfiable dependency by nesting its own copy. Installing the published 0.1.0 into a project running dsh 0.1.6-alpha.2 produced:

node_modules/@argszero/cordis-plugin-subagent-effort/node_modules/@deepseek-ai/dsh-tools   (=0.1.1-rc.2)
node_modules/@deepseek-ai/dsh-tools                                                        (=0.1.6-alpha.2)

Two copies of the harness runtime in one process, with no warning at all. As a peer the same narrow scope produces the correct outcome instead: npm refuses the install with ERESOLVE, which tells a user on a newer line that this plugin is not for them.

test/peer-range.spec.mjs asserts the admitted set exactly, asserts that no harness package appears in dependencies, and asserts that this README quotes the manifest range verbatim.

Development

npm install
npm run build        # tsc → lib/
npm test             # tsc + node --test (peer-range / declaration guard)

License

MIT