budget-executor

Files (1)

• SKILL.md

SKILL.md (first lines)

## Subagent binding

"executor subagent" → `Agent(subagent_type: budget-executor, run_in_background: true, prompt: <prompt>)`.

## Model resolution

The subagent model is controlled by the `model` frontmatter of the resolved `budget-executor.md` agent file: `.claude/agents/budget-executor.md` takes precedence over `~/.claude/agents/budget-executor.md`. Do not pass a per-spawn model parameter.

## Dispatch mode

Dispatch in the background only from a main agent, routed SAI coordinator, or routed SAI worker whose lifetime outlives the child. Capture and await the continuation on the dispatcher's own turn.

## Spawn prompt

Give the executor the exact command(s) to run, in order, and the output you need back. A goal without a command also works: the executor then picks the narrowest command that meets it. Either way it runs and reports; diagnosis and fixes stay with you.

A report without a final result for a requested command (`pending`, `running`, or no entry) is an obstacle under `sai/policies/obstacle-escalation.md`: read the result from the evidence the report names, such as its output log, or dispatch again, and do not wait for a later message from that executor.

## Warming and dispose

Fetch @sai/policies/cache-warming.md

The executor never arms: its report declares `warming: false`, so it is never sent a warming line or a `warm-up`, and there is nothing to dispose.

## Execution contract

The agent fetches `@sai/policies/executor-agent.md`, which owns exact-command execution, narrow low-output command choice, parallel independent commands, no self-correction, structured failure reporting, and the raw-output boundary. There is no tool-call cap.

Fetch @sai/policies/executor-agent.md

From mmadariaga/shared-ai, MIT.

More from mmadariaga/shared-ai