Leaderless multi-agent swarm with Playwright oracle. Spawns N workers in isolated worktrees with pre-partitioned DOD items. Workers implement and verify via screenshots. No orchestrator, no shared state.
From ancpluanpx claudepluginhub ancplua/ancplua-claude-plugins --plugin ancpluaThis skill is limited to using the following tools:
Guides Next.js Cache Components and Partial Prerendering (PPR) with cacheComponents enabled. Implements 'use cache', cacheLife(), cacheTag(), revalidateTag(), static/dynamic optimization, and cache debugging.
Migrates code, prompts, and API calls from Claude Sonnet 4.0/4.5 or Opus 4.1 to Opus 4.5, updating model strings on Anthropic, AWS, GCP, Azure platforms.
Details PluginEval's skill quality evaluation: 3 layers (static, LLM judge), 10 dimensions, rubrics, formulas, anti-patterns, badges. Use to interpret scores, improve triggering, calibrate thresholds.
Leaderless swarm launcher. Workers implement and verify via Playwright MCP screenshots.
Building: @$1
Workers run in isolated git worktrees. There is no shared filesystem — no lock files, no coordination directories, no status files. Coordination is structural:
This diverges from Carlini's git-push-racing locks because his agents ran for days across 2000 sessions. Ours run for minutes as subagents. Pre-partitioning is simpler and correct.
Extract DOD items from the user's prompt. Each item must be an observable outcome verifiable by screenshot (not a code quality requirement).
Example prompt: "build the dashboard with sidebar navigation and 3 data charts"
DOD items:
Calculate agent count: agents = min(max(2, ceil(dod_items / 2)), 8)
User can override by saying "use 6 agents" in the prompt.
| DOD Items | Agents |
|---|---|
| 1-3 | 2 |
| 4-5 | 2-3 |
| 6-7 | 3-4 |
| 8+ | 4-8 |
Distribute items across workers using round-robin:
Each worker also receives the full DOD for context. If a worker finishes their primary items early, they may attempt overflow items not in their assignment. Duplicate work across worktrees is harmless — the lead takes the passing version.
Spawn ALL workers in a single message using the Agent tool.
Every worker gets isolation: "worktree".
Each worker spawn prompt must include:
For each worker, use:
Agent tool:
isolation: "worktree"
prompt: |
You are Worker {K} of {M} in a carlini-jr swarm.
## Your Primary Items
{assigned items for this worker}
## Full DOD (for context and overflow)
{all DOD items}
## Project Context
{what app, framework, how to run it}
## Instructions
{inline the worker agent instructions from worker.md}
Implement your primary items first. Verify each via Playwright
MCP screenshot. If you finish early, pick overflow items from
the full DOD that are NOT in your primary assignment.
All workers launch in ONE message — maximum parallelism.
The lead does NOT monitor workers. The Agent tool blocks until all workers return.
When all workers return, report:
If workers produced changes in their worktrees, the lead merges the passing worktree branches. If two workers changed the same file, take the version backed by a passing screenshot.
subagent_type for worker — inline the worker instructions in the prompt./ancplua:carlini-jr build the dashboard with sidebar navigation and 3 data charts
That's it. Everything else is defaults.