midnite-swarm

Orchestrates parallel background subagents across multiple development phases with recurring status reports.

Updated Sep 1, 2026
One-click install
npx skills add https://github.com/bilo-io/midnite-studio --skill midnite-swarm-bilo-io
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: midnite-swarm
Source: https://github.com/bilo-io/midnite-studio/tree/main/.agents/skills/midnite-swarm
Command: npx skills add https://github.com/bilo-io/midnite-studio --skill midnite-swarm-bilo-io

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Running several development phases or tasks one at a time is slow, and manually tracking many parallel coding agents leads to stalled PRs, unwatched CI, and lost work when a subagent dies mid-task. ## Core Features & Use Cases - Parallel fan-out: Launches one background subagent per phase or ad hoc task, each invoking the midnite-create workflow with a configurable theme/size cap. - Recurring sitreps: Posts a status table (Agent | % | ETA | Doing | Notes) at a chosen interval, cross-checked against real PR and CI state plus on-disk SCRATCHPAD.md files. - Stall recovery: Detects subagents that report done without finishing, resumes or replaces them, and finishes remaining stages (CI watch, squash-merge, worktree teardown) directly when needed. - Use Case: You have four open phases in the .midnite/tasks/ tracker. Invoke the swarm with a size cap of L and a 10-minute sitrep interval, and every phase lands as its own merged PR while you watch one status table. ## Quick Start Run the midnite-swarm skill to fan out phases 12, 14, and 17 across parallel subagents with a 10-minute sitrep interval.

Frequently Asked Questions about midnite-swarm

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run multiple development phases in parallel with AI agents?▼

Invoke the swarm skill with the phases or tasks to run, a per-subagent size cap, and a sitrep interval. It spawns one background subagent per phase, each running the midnite-create workflow in its own git worktree until its PR merges.

How to monitor parallel coding subagents until their PRs merge?▼

The skill posts a recurring status table with columns Agent, %, ETA, Doing, and Notes at your chosen interval. Percentages come from observable state like worktree presence, open PRs, and CI results, never from subagent self-reports alone.

What happens when a background subagent stalls or dies mid-task?▼

Each subagent maintains a SCRATCHPAD.md in its worktree recording Done and Next lines. On each sitrep tick the orchestrator cross-checks PR and CI state, resumes stalled agents, or finishes the remaining stages itself from the scratchpad.

Can I limit how much work each parallel subagent takes on?▼

Yes. Stage 2 asks for a batch-wide cap: a theme or two, a size ceiling such as no theme over L, or all open unblocked themes. The cap is passed into each subagent's prompt, and the subagent picks specific themes within it.

Does the swarm skill work with any model or provider?▼

It asks which model or provider your CLI's parallel-agent mechanism exposes and passes that choice to each spawned subagent. It does not invent options your CLI does not support, and defaults to the current session's model when no separate choice exists.