vancsj/bootgear

> An agent harness framework — the gear between the LLM motor and your skills, tools, configs, and docs.
What is bootgear?
bootgear is a curated collection of general-purpose, extensible **skills**, **agents**, and **hooks** for coding-agent workflows.
Think of an LLM as a motor: raw power, no direction. A motor alone doesn't move anything — it needs a transmission that converts spin into motion, mounted on a chassis you can actually load your cargo onto. bootgear is that layer. It sits between the model and your assets, routes work through an opinionated loop, and lets you bolt on your own gear where it matters.
A plugin that serves both hosts ships from one source directory under plugins/,
with a .claude-plugin/plugin.json and a .codex-plugin/plugin.json side by
side (see memory-ledger, advisor). codex-plugins/ holds plugins with no
Claude Code counterpart at all (e.g. engine's Codex-CLI port). Both retain
platform-native manifests, hooks, wrappers, and invocation contracts.
converge is the exception to both: it ships as a Claude Code plugin from
plugins/converge and is bundled into the Codex engine rather than carrying
its own .codex-plugin, so there is no separate Codex install.
Install
Prerequisites: git, Python 3.11 or later, and uv
(engine and converge run through it). An engine run's advisor roles (a debate
reviewer, or the converge refuter and gap hunter) go to the other vendor's host through
advisor: on Claude Code they need the Codex CLI installed and signed in; on Codex, a
Claude Code session listening through advisor.
Each plugin installs on its own — you do not need the rest of bootgear. The table's
"installs … with it" notes describe Claude Code's declared dependencies; Codex CLI packages
differ (see Codex CLI).
| gear | what it does |
|---|---|
| memory-ledger | a ledger of settled questions, where confidence counts independent verifications instead of asserting itself |
| engine | an autonomous task-execution loop — start a run with /engine:start (Codex: $engine:start); installs converge, spec, test, review, advisor and memory-ledger with it |
| converge | settles the facts behind a findings slate: find, refute, and cut each claim with separate agents, then compute what is reported; installs memory-ledger and advisor with it, and is installed automatically with engine and review |
| review | PR review: lens finders and a gap hunt, with every finding settled through converge; installs converge with it |
| spec | requirements and design for engine's ticket-writing and ticket-to-pr runs; works on its own too |
| test | test writing, running, and failure triage for engine's bug-triage and ticket-to-pr runs; works on its own too |
| advisor | pairs Claude Code and Codex CLI through a shared mailbox, or makes a one-shot call to Codex |