agent-toolbox

A portable, spec-driven workflow and skill set for AI coding agents — one source of truth across Claude Code, Codex CLI, Antigravity CLI, and GitHub Copilot CLI¹.

Three plugins, each documented in detail by its own README:

• swe — the core: the spec-driven workflow skills, the capability agents, and the swe-loop conductor that ships an approved spec end to end
• lab — research: autonomous experiment loops and data-visualization guidance
• llmos — tooling, hooks, and the llmos-vault CLI for the shared llmOS Obsidian vault

Shared provider-neutral instructions live in AGENTS.md; scripts/install.sh is the manual path for non-plugin providers.

Installation

Claude Code

/plugin marketplace add kpeez/agent-toolbox
/plugin install swe@agent-toolbox
/plugin install lab@agent-toolbox      # optional: research machines
/plugin install llmos@agent-toolbox    # optional: machines with the llmOS vault

Codex CLI

codex plugin marketplace add kpeez/agent-toolbox
codex plugin add swe@agent-toolbox
codex plugin add lab@agent-toolbox

Codex plugins deliver skills only; the Codex .toml subagents come from the manual script below.

Manual install (Codex agents, Antigravity, Copilot)

./scripts/install.sh

Target	Installed to
Codex agents	~/.codex/agents/*.toml
Antigravity CLI	~/.gemini/AGENTS.md + skills symlinked from the repo
Copilot CLI	~/.copilot/copilot-instructions.md
Claude statusline	~/.claude/cc_statusline.py

Re-run after updating agent-toolbox. Skill scripts need no install — skills run them in place with uv run.

Skills

Each skill's SKILL.md is the canonical contract; this table is just the map. The per-plugin READMEs explain how the skills fit together. Skills follow the agentskills.io specification.

swe

Skill	Purpose
start-loop	Run/resume the swe-loop end to end — triage, design, approval gate, then the conductor ships the spec
sharpen	Interview the user to stress-test a plan; cross-check the code, record ADRs
write-spec	Create a feature spec — a pure-markdown design draft proven by committed tests
to-issues	Publish a spec as vertical-slice tracker issues with blocked-by relations
implement	Orchestrate implementing a spec — prove behavior with /tdd, fan work out to agents
tdd	Functional-test discipline — scratch scripts in tests/temp/, survivors become committed tests
ship-pr	Commit stable verified work as atomic commits, push, keep a draft PR current; finalize flips it ready
diagnose	Disciplined debugging — build a feedback loop, reproduce, hypothesize, instrument, fix
codebase-design	Shared deep-module vocabulary — depth, seams, adapters, the deletion test
improve-codebase-architecture	Find deepening refactors — turn shallow modules into deep ones
merge-conflicts	Resolve conflicts by tracing each side's intent; verify with the project's checks
research	Investigate a question against primary sources; capture cited findings as markdown
qmd	Search local markdown knowledge bases with the qmd CLI
setup-repo	Interview-driven repo setup — thin AGENTS.md, CLAUDE.md symlink, docs/agents/ topology

lab

Skill	Purpose
autoresearch	Autonomous experiment loops with defined metrics and stop conditions
data-viz	Research-backed guidance for designing and critiquing charts and figures

llmos

Skill	Purpose
maintain-llmos	Maintain the shared llmOS Obsidian vault through its conventions
setup-llmos	Diagnose and configure machine access to the llmOS vault
vault-cli	Route deterministic vault operations through the llmos-vault CLI

Workflow

The spine is sharpen → spec → issues → implement → review → PR. /start-loop <idea> runs it as one resumable command: spec approval is the last user prompt, after which the swe-loop conductor slices, implements, reviews, and ships with no further prompting. The swe plugin README documents the spine, the agents, and the conductor in detail, with diagrams.

Versioning

Each plugin's version lives in exactly two files, kept identical: its .claude-plugin/plugin.json and .codex-plugin/plugin.json. The marketplace catalogs carry only names and paths — regenerate them with scripts/gen-marketplaces.py, never hand-edit. Bump both manifests at once:

scripts/bump-plugin-version.sh swe 1.9.8

A bump is inert until it lands on master — both providers install from GitHub, not the working copy.

Footnotes

1. Inspired by Matt Pocock's skills repo ↩