mcpField
D
dispatch-new-session

dispatch-new-session

github.com/wkentaroAgent workflows0176 installsUpdated 1d ago
Repository

About dispatch-new-session

This skill starts one independent agent session in a background Herdr tab, confirming the target harness and compatible model and reporting its startup state. It is for dispatching a defined task to Claude, Codex, or Pi while leaving the session running.

Key features

  1. 01Starts a session in a background Herdr tab
  2. 02Selects a target harness and compatible model
  3. 03Uses the current workspace and directory
  4. 04Checks for working or blocked state
  5. 05Reports session and agent details

Use cases

  1. 01Dispatch a defined task to an independent agent session
  2. 02Choose a model for a different harness
  3. 03Start background work and check for blockers

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/wkentaro/agent-skills --skill dispatch-new-session

After installing, type/dispatch-new-sessionin Claude Code to confirm it loaded.

Bundled filesagents/openai.yaml

Details

From the project README

Dispatch New Session

Launch one independent session in a background Herdr tab.

Choose

Use the invocation argument or current request as the task brief. If neither contains a concrete task, ask for it first.

Ask for the target harness with native structured input, then ask for a model compatible with that harness. Recommend the path matching the harness running this skill:

CallerRecommended target
ClaudeClaude with Opus
CodexCodex with Sol
PiPi with its configured default model

For a cross-harness dispatch, recommend that target's same default. Offer these useful alternatives:

  • Claude: Opus, Fable, or another Claude model.
  • Codex: Sol or another Codex model.
  • Pi: its configured default or a model present in pi --list-models.

Put the recommendation first and label it Recommended. When structured input is unavailable, ask one plain-text question at a time. Never silently replace an unavailable or incompatible choice.

Dispatch

Read herdr --skill, require HERDR_ENV=1, and use the installed CLI syntax. Create a new tab in the current workspace and directory with --no-focus. Derive a short tab label and unique agent name from the task.

Start the selected harness in the tab's root pane:

  • Claude: --allow-dangerously-skip-permissions --model <model>
  • Codex: --yolo -m <model>; use gpt-5.6-sol for Sol.
  • Pi: omit --model for the Pi default; otherwise pass the exact provider/model identifier reported by pi --list-models.

Prompt the new agent with the task brief without expanding its scope. Wait only long enough to confirm the prompt enters working or blocked; leave the session running. If startup fails before an agent starts, close only the tab created by this invocation. If blocked, inspect and report the blocker without answering it.

Report the tab, agent name, harness, model, and observed state.

Comments

Markdown supported