RepositoriesAgent Dispatch

Agent skill · MIT

Agent Dispatch

The model is picked by the type of task, never by the importance of the project.

Unless something sets a model, Claude Code runs every subagent on the model you chose for the session, so the agents translating a strings file cost the same per token as the one designing the architecture. This skill makes the split explicit before anything launches, and reports what was spent after.

  • Claude Code · Cursor · Codex · Gemini CLI
  • One SKILL.md
  • English and Spanish
  • MIT
Dispatch of a small Chrome extension: design and code inline on the session's model, the tests on a mid model, four translations on small models, and the final review inline.
Dispatch: 5 agents · 1 mid (tests) · 4 small (translations) · inline: design, code, lint + tests, final review · cap: 1 fix round

The rule

One table

Type of taskExamplesModel · effort
MechanicalTranslations of fixed strings, listing, counting, extracting fields, cutting a long log down to its errorsSmall (e.g. Haiku) · low
Bounded analysisSummarising a module, mapping dependencies, tests from a clear spec, one mechanical change across many filesMid (e.g. Sonnet) · medium
JudgmentArchitecture, the code that decides how something works, choosing between options, the final reviewThe session's model

Guardrails

What it tells the agent to keep

Delegate at all?

The default is no, and the count starts at zero. Checking what a command can check is never a reason for a subagent.

The reviewer never gets cheaper

Judgment stays on the session's model, and so does the final review.

One retry, a tier up

A small or mid agent that fails gets its one retry on the next tier, never again on the same model.

Self-contained briefs

Every subagent gets the goal, the minimum context, what it must not touch, the output format and when it's done. It never sees your conversation.

Every loop declares its cap

The success condition and the maximum number of rounds, before it starts.

Declare, then measure

A dispatch line before launching and the spend your tool reports after. No invented savings.

Install

One file

Claude Code, as a personal skill

mkdir -p ~/.claude/skills/agent-dispatch
curl -fsSL https://raw.githubusercontent.com/Octonove/agent-dispatch/main/SKILL.md \
  -o ~/.claude/skills/agent-dispatch/SKILL.md

Or as a plugin

/plugin marketplace add Octonove/agent-dispatch
/plugin install agent-dispatch@agent-dispatch

In Spanish: es/SKILL.md, or the agent-dispatch-es plugin. Cursor, Codex, Gemini CLI and other clients: see the README.

You need

  • Claude Code, Cursor or any client that reads Agent Skills

What it doesn't do

Where it stops

  • It doesn't give you a savings number. It can't measure what your tool doesn't report, and it won't invent one.
  • Judgment never gets cheaper. If you want everything on the smallest model, this isn't the skill for you.
  • The install paths for Cursor and other clients come from each vendor's documentation and are untested.

It is also section 1G of The Orchestrator, the Invokard card that routes every request: if you use that card, the dispatch is already inside it.

The code, the issues and the history are on GitHub.

Star it if it helps; open an issue if it breaks.

View on GitHub