Agent Factory

A skill-first factory for creating minimal reusable agent customizations and modernizing legacy packs without losing domain knowledge. It supplies missing context, capabilities, constraints and contracts; the model/harness supplies intelligence and runtime orchestration.

No customization is a valid result. This pack has one portable skill, selective references, and optional artifact examples. It has no custom factory agent, fixed team, thought-stage state machine, or obligatory review/delegation loop.

The legacy copilot-factory and roo-agent-factory packs remain available, but init.cmd now installs only this factory. Invoke the new agent-factory skill, not @copilot-factory.

Installation

On Windows, run the repository's initializer to copy the canonical skill into .github\skills\agent-factory:

.\init.cmd

No Node.js installer is required. Only this skill is copied; legacy factories are neither installed nor uninstalled. Re-running init.cmd overwrites matching files in the installed skill, so review or back up local edits first. It does not remove additional destination files or provide a sandbox.

For another project or host, copy the complete skills\agent-factory directory to the appropriate project-local location:

The source here is canonical and tracked; the checkout's installed .github\skills\ copy is ignored. Do not edit separate vendor copies independently.

Copilot CLI plugin distribution:

copilot plugin marketplace add srulyt/srulys-agent-packs
copilot plugin install agent-factory@srulys-agent-packs

The root manifest uses portable Agent Plugins 1.0.0, shared by documented current Copilot/Codex support. Claude Code uses the thin .claude-plugin\plugin.json wrapper and the same skills\ tree; for local development, use claude --plugin-dir .\agent-packs\agent-factory. Check your exact client/surface before installing. Codex and Claude runtime execution has not been verified here.

Do not use the legacy npm pack installer for this skill-only plugin: it currently discovers agent-based/Roo layouts.

Usage

Ask the default agent:

Use the agent-factory skill to create the smallest customization for our
incident briefs. Here are our severity definitions and required brief fields...

Use the agent-factory skill to modernize <pack>. Preserve its business rules,
required approvals and output contracts. Write the migrated result to <destination>.

Codex supports $agent-factory; Claude Code uses the native slash skill entry (plugin installs namespace it). Restart/reload skills using the selected host's supported discovery controls. Do not install both plugin and project copies unnecessarily; duplicate names can confuse discovery/eval tooling.

What it chooses

Nothing for a task already handled well; instructions for small broadly applicable invariants; skills for specialized reusable knowledge; prompts for invocation convenience; custom agents for enduring isolated context/tool/permission boundaries; MCP for missing capabilities; hooks/code/policy for actual enforcement; plugins for distribution only.

Modernization classifies meaning inside legacy files, preserves essential rules and domain-required ordering, and writes a concise migration record with source-to-destination provenance and unresolved behavior. Generic personas/choreography are not retained merely for familiarity.

The root skill directly links its selection, modernization, evaluation and platform references. They are loaded by relevance, not as a mandatory reading roster. Platform freshness records dated first-party evidence and refresh triggers; model defaults remain inherited.

Checks and evidence

node .\scripts\run-evals.mjs agent-factory -- --tags structural

See factory evaluations for live scenarios, source-grounded preservation checks, comparisons, and honest runtime limitations. The legacy pack linter skips plugin-root skills; the new structural eval covers this pack explicitly without changing old contracts.