dsh-subagent-concurrency-limit

dsh-subagent-concurrency-limit

Limit how many DeepSeek Harness subagent runs and local child-agent turns run at the same time in one process. The root Agent is not counted.

Why it is needed

In DeepSeek Harness dsh-v0.1.1-rc.2 (commit b150a551b8d465e31e418e1b2eaf5e79bbb7d28e), the workflow engine's maxConcurrentAgents and maxTotalAgents apply to agent() calls inside each workflow run. The standard subagent tool has no process-wide limit for root-level direct runs outside a workflow. maxParallelToolCalls limits the dispatch pool for parallel-safe tool calls in one Agent step, not the number of background or continuable runs.

This plugin fills that gap by limiting concurrent root delegation admissions and local descendant Agent turns in one process.

Scope

The plugin limits the delegation tools listed in subagentToolNames and local Agent turns in the same process where delegationDepthOf(agent) > 0. They share one process-local pool. The default limit is 2.

• A root Agent's delegation call waits in FIFO order before creating a child.
• Add nonstandard delegation tool names to subagentToolNames.
• Nested children that call the same delegation tool also consume the pool. If a permit is available, they take it immediately; when the pool is full, they fail without waiting so a child cannot block the ancestor that is waiting for it.
• Workflow agent() calls and resumed local child turns use the same pool when their Agent turn starts.
• A permit is released when the child turn becomes idle, the run ends, or the Agent is disposed.
• Pool waits are bounded: a root delegation call or a resumed local child turn that has been queued for queueTimeoutMs (default 300000, five minutes) fails with a clear error instead of waiting forever.
• A locally published child whose reservation was lost (for example a background job that published outside the delegating call's async context) never queues: it takes a permit immediately when one is free and fails loudly when the pool is full, so it cannot block the ancestor that is waiting for it.
• For remote-process providers such as Codex, Claude Code, and ACP, only the configured root delegation admission and the observable subagent/start and subagent/end boundaries are visible. Nested execution and restarts inside the remote process cannot be counted.
• maxParallelToolCalls is independent because it limits tool calls, not live runs.

Install

These examples use the web profile. Replace web with your profile name in each command and path.

Install the plugin:

dsh plugin --profile web add dsh-subagent-concurrency-limit

Add the plugin configuration to $DSH_HOME/profiles/web/cordis.patch.yml:

- id: subagent-concurrency-limit
  name: dsh-subagent-concurrency-limit
  config:
    maxConcurrentSubagents: 2
    queueTimeoutMs: 300000
    subagentToolNames:
      - subagent

maxConcurrentSubagents must be a positive safe integer. queueTimeoutMs must be a positive safe integer; the default is 300000 (five minutes), which bounds every pool wait. subagentToolNames must contain at least one nonempty delegation tool name. The bundle patch applies the defaults above; edit the same keys to override them.

Start the profile:

dsh --profile web

Update

To update the installed plugin to the latest npm release, run:

dsh plugin --profile web add dsh-subagent-concurrency-limit@latest

Restart the profile after updating.

Development

pnpm install
pnpm run check

To inspect the package before publishing:

npm pack --dry-run

prepare builds lib/ when packing or publishing. npm publish runs the checks through prepublishOnly before publishing to npm.

License

The project is licensed under MIT.

Harness compatibility

The declared compatibility range is DeepSeek Harness >=0.2.0-rc.2.