agent-kernel

name

agent-kernel

description

Local-first governance kernel for AI coding agents — gives Claude Code, Codex, Cursor, Antigravity, and Gemini CLI a shared memory, an approval inbox for new rules, generated AGENTS.md / CLAUDE.md / cursor-rule.mdc, Claude + git hooks, and a deterministic policy guard. Adds a JSON-first memory layer (rules + preferences + workflows + project-notes + skills), an episodic memory archive, and MCP tools for search/read/capture/sync of past sessions. Use when the user asks to "remember this rule", "save this as a memory", "what did we do last time", "search past episodes", "propose a new rule", "make agent follow our standards automatically", or any request to give an agent persistent local memory across sessions. Bundled inside delegate-team v2.5.0 — installed via `./install.sh --kernel`. Triggers on: "agent kernel", "ak", "remember this", "save this rule", "memory tool", "episodic memory", "approval inbox", "guard policy", "rule inbox".

agent-kernel — local-first governance + memory layer

The memory backbone of delegate-team. A vendored, version-pinned copy of @mamdouh/agent-kernel@0.0.5 ships inside this repo at agent-kernel/dist/cli.mjs. Installed by default with ./install.sh --all in delegate-team v2.5.0+.

What this skill IS

A single Node.js CLI (agent-kernel / ak) that gives every coding agent you use:

1. A shared local memory at ~/.agent-kernel/source/memories/*.json
2. An episodic memory archive at ~/.agent-kernel/episodes/ (with search/show/stats)
3. An approval inbox so agents can propose rules but only the kernel publishes them
4. Generated instruction files for every agent:
   • AGENTS.md (Claude Code / Codex / Cursor / OpenCode)
   • CLAUDE.md
   • .cursor/rules/00-agent-kernel.mdc
   • .agents/agents.md
   • GEMINI.md
5. Hooks — Claude PreToolUse + PostToolUse, git pre-commit, optional CI guard
6. MCP tools — agent_kernel_search_episodes, agent_kernel_read_episode, agent_kernel_capture_episode, agent_kernel_sync_episodes
7. A deterministic policy guard — blocks dangerous rm -rf, curl|sh, force-push to main, secret leaks, and any rule you add

Why bundle it inside delegate-team

delegate-team orchestrates many agents. Without a shared memory, each session starts cold — the same standards get repeated every conversation. agent-kernel solves that:

Without agent-kernel	With agent-kernel
Standards repeated in every prompt	Standards live in ~/.agent-kernel/source/memories/*.json and auto-attach
Lost context after session end	Episodes saved locally; searchable later
Agent writes whatever rule it wants	Proposal inbox; you approve before publish
Manual git commit may leak secrets	Pre-commit hook + agent-kernel guard --staged blocks
Different agents see different rules	One JSON-first source compiles to all platforms

Quick start (inside delegate-team)

# One command — link + init + link the project + install hooks
./install.sh --kernel

# Or auto-installed via:
./install.sh --all

# Verify
agent-kernel doctor
agent-kernel status

After installation, in any project:

agent-kernel init --sync --enforce   # First time in a project
agent-kernel link . --hooks         # Link AGENTS.md + git hook

# Save a rule
agent-kernel remember "Never add local SQLite fallback to production Supabase apps." \
    --type policy --level critical --tags supabase,database --publish

# Search episodes
agent-kernel episode search "SQLite fallback Supabase"
agent-kernel episode show <episode-id>

Files in this vendored copy

agent-kernel/
  dist/cli.mjs             # The CLI — single ~85KB ESM file
  docs/                    # 8 architecture + protocol docs
  examples/                # CI guard workflow, sample memory rules, sample episode
  develpment/              # Backlog + epics + milestones + sprint plan + backlog.json
  LICENSE                  # MIT
  README.md                # Upstream README (verbatim)
  scripts/build.mjs        # Build script (not used here — cli.mjs is shipped pre-built)
  SKILL.md                 # THIS FILE — Skills.sh manifest
  install.sh               # Local installer
  wrapper.sh               # Bash shim — prefers global agent-kernel, falls back to vendored
  MEMORY.md                # How delegate-team uses this for memory
  VERSION                  # 0.0.5

Compatibility

Agent	Memory source	Hook install	Compile target
Claude Code	✅ via AGENTS.md + CLAUDE.md	✅ ~/.claude/hooks/	PreToolUse + PostToolUse
Codex	✅ via AGENTS.md	n/a (read-only)	AGENTS.md
Cursor	✅ via .cursor/rules/00-agent-kernel.mdc	n/a	.mdc rule
OpenCode	✅ via AGENTS.md	n/a	AGENTS.md
Antigravity	✅ via .agents/agents.md	n/a	.agents/
Gemini CLI	✅ via GEMINI.md	n/a	GEMINI.md
60+ others	✅ see Skills.sh index	depends on agent	via AGENTS.md

Memory layout: fully backward compatible with v0.0.1 (auto-migrates flat files via agent-kernel migrate json --publish).

Installation paths

Path	When
Global npm install -g @mamdouh/agent-kernel (when published)	Preferred when user wants the latest
Vendored at agent-kernel/dist/cli.mjs	Default with delegate-team — guaranteed version match
npx -y agent-kernel	One-off use, no install

The wrapper.sh shim picks the right one:

1. If agent-kernel on $PATH → use it (user's preferred version)
2. Else if node agent-kernel/dist/cli.mjs works → use vendored copy
3. Else → install on first use via install.sh

See also

• MEMORY.md — how delegate-team uses agent-kernel for the /apeiron memory layer
• docs/MEMORY_PROTOCOL.md — rule + episode data model
• docs/MCP_SERVER.md — MCP tool definitions for non-CLI integration
• docs/EPISODIC_MEMORY.md — episode sync from Claude/Codex JSONL transcripts
• docs/STRICT_MODE.md — when to flip --enforce
• docs/JSON_FIRST_STORAGE.md — storage schema

License

MIT (© Mamdouh Aboammar) — upstream agent-kernel v0.0.5.

Repository

https://github.com/imMamdouhaboammar/delegate-team/tree/master/agent-kernel