Skip to content

Repository files navigation

agent-toolkit-bundle

The single source-of-truth for engineering rules, skills, hooks, agents, commands, and governance — symlinked into every AI coding tool you use.

One repository holds the standards your agents must follow (CLAUDE.md + templates), the domain knowledge they load on demand (skills/), the deterministic guardrails that enforce the rules (hooks/ + plugins/), the role-specific sub-agents (agents/), the workflow commands (commands/), the shared policy data (policy/), and the MCP integrations for memory and code intelligence. install.sh symlinks it into Claude Code and OpenCode, and the same corpus is consumed by Codex and Gemini. Edit a file here once — it changes everywhere.


Why this exists (the differentiators)

Most agent-asset collections are a pile of advisory skills. This one is a governed, enforced, multi-tool platform:

  • Governance & compliance layer. A PII guard that blocks PANs/IBANs/national IDs before they reach a tool call; an egress allowlist; a tamper-evident, hash-chained audit log (each entry commits to the previous entry's hash); DORA and PCI-DSS 4.0 control mappings (skills/compliance, docs/data-classification.md); and OpenTelemetry agent tracing (docs/otel-agent-tracing.md). Nothing else in the ecosystem ships this.
  • Deterministic enforcement, not suggestions. Rules are enforced by blocking hooks (Claude Code) with TypeScript plugin twins (OpenCode) that share one policy source (policy/guard-patterns.json). Skills advise; hooks stop you — non-parameterised SQL, print() in library code, force-pushes to main, secret literals, and unformatted files are caught at the tool boundary.
  • No-AI-attribution + mode circuit-breaker. Commits and PRs are scrubbed of AI attribution deterministically. A company/private "mode" guard hard-separates work and personal paths at every tool call.
  • Spec-driven workflow. OpenSpec (/opsx:*) gives every non-trivial change a paper trail: explore → propose → apply → archive.
  • Persistent memory + code intelligence over MCP. MemPalace (cross-session memory: wings/rooms/halls/drawers) and CodeGraph (symbol search, call graphs, blast-radius) are wired as MCP servers for all tools.
  • Multi-tool source-of-truth. One repo drives Claude Code, OpenCode, Codex, and Gemini agents — no divergent copies.
  • The deepest reliability vertical in the ecosystem. First-class SRE, observability, multi-cloud (AWS/Azure/GCP), and Datadog skills — SLO/error-budget, incident response, capacity forecasting, runbooks, alerting, dashboards.

What's inside (verified against the tree)

Asset Count Location
Skills 166 skill directories (159 top-level loadable SKILL.md + nested sub-skills; 194 SKILL.md files total) skills/
Agents — Claude Code 19 agents/claude/
Agents — OpenCode 21 (the 19 + opsx, refactor) agents/opencode/
Agents — Gemini 15 (agents-only, experimental) agents/gemini/
Hooks (shell) 18 hooks/
Plugins (TypeScript) 13 plugins/
Commands — Claude Code 13 commands/claude/
Commands — OpenCode 14 (the 13 + model-report) commands/opencode/
Shared policy data guard patterns, model-tier map policy/
MCP sub-packages MemPalace, CodeGraph, OpenSpec mempalace/, codegraph/, openspec/

The exhaustive, categorised skill catalogue lives in docs/skills.md; auto-activation keyword mappings live in skill-rules.json.


Multi-tool support

Claude Code OpenCode Codex Gemini
Instructions file CLAUDE.md AGENTS.md AGENTS.md GEMINI.md
Agents 19 (agents/claude/) 21 (agents/opencode/) reused as role prompts 15 (agents/gemini/)
Enforcement 18 blocking shell hooks 13 TypeScript plugins advisory (no native hook runtime) advisory (no native hook runtime)
Commands 13 native slash commands 14 native slash commands markdown playbooks reused as playbooks
Skills shared from skills/ shared from skills/ shared from skills/ shared from skills/
Memory / code intel MemPalace + CodeGraph (MCP) MemPalace + CodeGraph (MCP) MemPalace + CodeGraph (MCP) via MCP where supported
Install install.sh symlinks install.sh symlinks starter templates (below) agent files + GEMINI.md template

Codex consumes the shared corpus directly: use templates/AGENTS.md.example as your project AGENTS.md, and create a config.toml from templates/codex.config.toml.example (which wires MemPalace and CodeGraph as [[mcp_servers]]). Codex has no native hook/plugin runtime here, so the hooks and plugins act as documented policy rather than deterministic enforcement — see docs/codex.md.

Gemini ships 15 agent definitions in agents/gemini/ (frontmatter uses Gemini tool names: read_file, write_file, replace, glob, grep_search, run_shell_command). It is agents-only and experimental — there is no Gemini hook/plugin runtime and the installer does not auto-symlink Gemini. Use templates/GEMINI.md.example for project instructions and read shared skills/ on demand.


Install

git clone https://github.com/mwigge/agent-toolkit-bundle.git
cd agent-toolkit-bundle
./install.sh --help     # inspect options first
./install.sh            # symlink into Claude Code + OpenCode

install.sh creates symlinks from ~/.claude/ and ~/.config/opencode/ into this checkout (agents, commands, hooks, plugins, tools, and skills), so edits here are live everywhere immediately. It is idempotent and backs up any existing non-symlink files. Optional flags:

  • --templates copies CLAUDE.md.example, AGENTS.md.example, GEMINI.md.example, and a Codex .codex/config.toml into the current directory.
  • --components a,b,c installs a subset (agents,skills,hooks,plugins,tools,scripts,commands,mempalace,codegraph,openspec).

settings.json for Claude Code is copied, not symlinked (it carries per-machine hook paths). See docs/installation.md for the full cross-platform guide, prerequisites, and the mode circuit-breaker shell function.

Hooks fail open. Every hook that depends on an external tool (jq, python3, formatters, codegraph, mempalace) guards on it and exits cleanly if it is absent — a fresh box never errors, it just enforces less.


Slash commands

Native in Claude Code (commands/claude/) and OpenCode (commands/opencode/):

Command Purpose
/commit Analyse the diff, draft a conventional commit, validate, commit
/pr Validate branch, push, fill the MR/PR template, open the request
/story INVEST check, draft a user story with ACs, hand off to @jira-story
/review Four-lens adversarial review of the current branch
/test Generate and run tests for a target, closing coverage gaps
/debug Systematic root-cause debugging of an error (reproduce -> isolate -> fix -> verify)
/refactor Safe, behaviour-preserving refactor of a target, one verified step at a time
/docs Generate or update docs (README, reference, docstrings, ADR) for a target
/index Update the docs index and memory.md after a work session
/opsx:propose OpenSpec — create a change with proposal, design, specs, tasks
/opsx:explore OpenSpec — thinking mode, no implementation
/opsx:apply OpenSpec — implement the next unchecked task
/opsx:archive OpenSpec — promote specs, archive the change
/model-report OpenCode only — tiered model usage / cost summary

MemPalace ships an opt-in /mempalace-mine command in its sub-package (mempalace/commands/) — installed only when you enable the mempalace component.


Agents

19 Claude Code agents, 21 in OpenCode (adds opsx and refactor), 15 in Gemini. Read-only-by-design agents are scoped to least privilege: @architect gets Read, Grep, Glob; @debugger, @reviewer, and @security get Read, Grep, Glob, Bash (no write); coder/tester agents keep their write tools. In Claude Code agents are leaf nodes (human-triggered handoffs); in OpenCode the orchestrator can spawn mode: subagent agents autonomously. Full inventory and collaboration flow: docs/agents.md. Per-agent model routing for OpenCode: docs/model-tier.md and docs/local-models.md.


Enforcement: hooks and plugins

18 shell hooks (Claude Code) and 13 TypeScript plugins (OpenCode) enforce the same rules from one policy source (policy/guard-patterns.json). Highlights: mode-guard, no-ai-attribution, security-guard, pii-guard (PreToolUse, blocking); format-on-save, inline-quality, codegraph-sync, mempalace-ingest (PostToolUse); quality-gate (Stop, blocking); observe (tamper-evident audit chain); permission-autoapprove (GREEN/YELLOW/RED tiers). Lifecycle, exit codes, and the plugin API: docs/hooks.md and docs/plugins.md.


OpenSpec workflow

Every non-trivial change gets a paper trail:

/opsx:explore   →  think before committing to a solution
/opsx:propose   →  proposal.md, design.md, specs/, tasks.md
/opsx:apply     →  implement the next unchecked task
/opsx:archive   →  promote specs, archive the change

OpenSpec is a prerequisite: npm install -g @fission-ai/openspec@latest, then openspec init --tools <claude|opencode|codex> in a project. See openspec.dev and openspec/docs/.


MemPalace & CodeGraph (MCP)

  • MemPalace — persistent cross-session memory (facts, events, discoveries, preferences, advice) backed by ChromaDB, entirely local. Setup: docs/install-mempalace.md.
  • CodeGraph — structural code intelligence: codegraph_search, codegraph_callers, codegraph_impact (blast radius). Prefer it over Grep/Glob in large repos. See codegraph/docs/.

Both are wired as MCP servers for Claude Code, OpenCode, and Codex.


Further reading

About

Bundled agents, skills, hooks, commands, and OpenCode plugins for Claude Code and OpenCode — sibling to agent-circuit-breaker

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages