swarm

Splits work across multiple Pstack agents to explore options in parallel or double-check results independently.

Installation
Run `npx skills add "https://github.com/christofferbergj/agent-skills" --skill "swarm"` to install this skill, then follow its SKILL.md instructions for my next request.

Paste this into Claude Code, Cursor, or any agent that can run commands.

SKILL.mdShow the author's original SKILL.md
---
name: swarm
description: "Use for Pstack delegation or independent review; organize coverage slices, races, or parallel exploration."
---

# Swarm and host execution

For every Pstack delegate or cross-vendor review, read [the execution contract](references/execution.md). It owns native host APIs, model settings, isolation, review restrictions, and disclosed fallback. A caller can use that contract for one delegate without starting a coverage swarm.

To hand a check that needs Computer Use to Codex, follow [the Computer Use route](references/computer-use.md).

## 1. Frame

Open a task list with one entry per phase before dispatching anything. Name the done predicate, report or artifact, and required coverage. Choose slices, an identical-brief race, or a mix. Declare a race rule before dispatch: first passing result, rank all, or best-of. Workers take the `swarm workers` role from [the execution contract](references/execution.md#model-roles); for a model race, name each arm's model before dispatch instead. Choose the smallest useful worker count within the user's budget and host capacity; total jobs can exceed concurrent slots.

Give each writer its own worktree or output directory. A verification or measurement brief names the exact revisions, sample definition, method, and acceptance predicate. Complete when every required slice or arm has an independent brief and output location.

## 2. Dispatch

Use the host's native delegation tools and actual schema. Start independent jobs before waiting, respecting the available concurrency window. In T3 Code, a job that needs another machine's tools, or would load this one for long, runs there through [the T3 route's On another machine](references/t3.md#on-another-machine). Each brief supplies the goal, boundary, repository guidance, resolved rubric or reference paths, evidence to return, and allowed writes. A worker that can prove a defect reports `ISSUES` and lists every issue it can prove, not only the first. Prefer cross-vendor critics through the execution contract when available.

Complete when each job has a handle or an explicit dispatch failure. A dropped worker is a gap, never an implicit success.

## 3. Aggregate

Collect terminal results. Accept `PASS`, `ISSUES`, or `BLOCKED` only with the brief's required revisions, method, and evidence. Retry a missing or malformed result once when recoverable; after another miss, record the gap. Inspect each finding rather than treating agreement as proof. Coverage requires every mandatory slice; a race uses the declared selection rule.

Complete when every dispatched job is accounted for and every required slice has evidence or a named gap.

## 4. Report

Return one compact table of jobs, verdicts, revisions, and evidence, followed by actionable issues, dropouts, and the race result, without pasting raw worker output. Disclose sequential execution, same-vendor review, or unavailable delegation. Complete when the report distinguishes actual independent passes from local fallback work.

Ships with 10 supporting files:

  • agents/openai.yaml
  • LICENSE
  • references/computer-use.md
  • references/execution.md
  • references/t3.md
  • scripts/review.py
  • scripts/target.py
  • scripts/tests/test_review.py
  • scripts/tests/test_target.py
  • UPSTREAM.md

Mirrored from the author's public source. Install counts from the open skills registry.

The systems behind these skills get built for partners every week.

Partner with us