agent-context-kit

Vendor-neutral, persistent project context for AI coding agents. Shared state lives under docs/context/, stays reviewable in Git, and survives chat resets, compaction, and switching agents (Claude Code, Codex, Antigravity).

Layout

skills/engineering/agent-context/   # the installable skill (SKILL.md + references/ + assets/)
docs/context/                       # per-project context this skill manages
AGENTS.md / CLAUDE.md / GEMINI.md   # short pointers so any agent auto-discovers docs/context/

docs/context/:

docs/context/
├── index.md                # generated routing table
├── summary.md              # generated rolling summary — read this first
├── project.md              # durable project-wide context
├── backlog.md              # things to do next, captured with /ac-note
├── config.yaml             # tracking and summary policy
├── topics/                 # reusable, evidenced knowledge
├── work/                   # active task state and checkpoints
├── archive/                # completed task records
└── generated/repo-tree.md  # reproducible repository structure

.context.local/             # gitignored machine-local/private context

The skill is instruction-driven — no program to install or run. The agent performs every action with its own file tools plus date/git/find (optionally tree), following SKILL.md and references/procedures.md. Never stores secrets, raw transcripts, or chain-of-thought.

Install

With npx skills (recommended — works for every agent)

# Claude Code
npx skills add ThanakritAof/agent-context-kit -a claude-code

# Other agents (codex, antigravity, gemini-cli, cursor, etc.)
npx skills add ThanakritAof/agent-context-kit -a 

Alternative — symlink

ln -s "$(pwd)/skills/engineering/agent-context" .claude/skills/agent-context   # Claude Code
ln -s "$(pwd)/skills/engineering/agent-context" .agents/skills/agent-context   # Codex

Either way, tell the agent to run the skill’s Initialize action once per target repo — it scaffolds docs/context/, the AGENTS.md/CLAUDE.md/GEMINI.md pointers, the /ac- slash commands, and the .gitignore entry. Native install is optional: the skill also works by just telling any agent to “use the agent-context skill” — it’s designed for explicit invocation, not description-matching.

If you installed the skill globally (as npx skills add does by default), also drop /ac-init at the same personal scope so it works in any repo without a prior Initialize run there:

mkdir -p ~/.claude/commands
curl -fsSL https://raw.githubusercontent.com/ThanakritAof/agent-context-kit/main/skills/engineering/agent-context/assets/templates/commands/ac-init.md \
  -o ~/.claude/commands/ac-init.md

Continuity hook (optional)

Everything else here is loaded once at session start and relies on the agent choosing to re-read it later — nothing re-asserts it mid-session, so it can fade on a long conversation, especially around context compaction. Claude Code, Codex, and Antigravity each support a real SessionStart hook that re-injects docs/context/summary.md on session start, resume, /clear, and right after compaction — so state survives even if nobody remembered to refresh it first. Ask the agent to “install the continuity hook” and it merges the JSON below into whichever vendor(s) you name; each is independent, so installing (or skipping) one never affects the others.

The command is the same for all three:

cat docs/context/summary.md 2>/dev/null || true

Claude Code and Codex share this exact shape (merge into the file’s existing hooks object — don’t overwrite other hooks/settings):

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "startup|resume|clear|compact",
        "hooks": [
          { "type": "command", "command": "cat docs/context/summary.md 2>/dev/null || true" }
        ]
      }
    ]
  }
}

Antigravity’s hooks.json is assumed to use the same shape. See Continuity hook for the exact merge/removal steps.

Checkpoint nudge hook (optional)

The continuity hook above only re-reads summary.md back in — it doesn’t make the agent write checkpoints. This second, separate hook nudges that side: a PostToolUse hook counts implementation edits since the last checkpoint and, past a threshold (default 8), injects a one-line reminder into the agent’s context. It never writes a checkpoint itself — the summary/decision text still needs the agent’s judgment. Ask the agent to “install the checkpoint nudge hook”.

Claude Code and Codex only. Antigravity’s PostToolUse hook has no channel back to the model (its stdout is always {}), so this one isn’t available there — Antigravity users get the continuity hook only.

See Checkpoint nudge hook for the exact merge/removal steps and why the two scripts differ (Claude Code reports a plain edited-file path; Codex reports a patch blob that has to be parsed for *** Update File:-style markers).

Reference

• agent-context — Initialize and maintain vendor-neutral repository context: project overview, durable topics, active work, checkpoints, rolling summaries, a backlog of things to do next, and resumable handoffs across Claude Code, Codex, and Antigravity. Trigger by asking to initialize, save, checkpoint, summarize, restore, or resume project context.

Lifecycle actions

The summary refresh interval is configured by every_checkpoints: in docs/context/config.yaml (default 3).

Slash commands

Optional shortcuts — the same actions work in plain language. Installed per vendor at .claude/commands/, .agent/workflows/, .codex/prompts/. /ac-init is the exception: it’s only installed by Initialize, so a never-initialized repo won’t have it yet. Install it at the personal scope too (~/.claude/commands/ac-init.md) so it works before that first run — see Install.