agent-kit
Concepts

Session context

What every session starts with, how work resumes, and how the context budget is kept.

The SessionStart hook injects, in about 1.3k tokens:

  • the profile — voice for your technical level, reply language, git flow, method, stack;
  • the lanes and one line about the living specs (capabilities and requirement counts) and the QA setup;
  • Resume — where the last session stopped (branch, uncommitted files) and each active work item with its next task;
  • the knowledge index (one line per entry) and the script index.

Claude Code caps hook context at 10,000 characters, so the hook renders under 9,500: when needed it collapses global knowledge groups, then the scripts list, then project groups, and clips summaries to 160 characters.

Resume

A Stop hook writes .agent-kit/checkpoint.json after every turn. Keep status and the next open task of work files accurate — a new session picks them up automatically.

Budget

agent-kit context measures what loads every session: the CLAUDE.md / AGENTS.md chain, skill listings, the session context, and skills Claude Code can't see. --sections ranks every heading by tokens. agent-kit doctor fails above contextBudget (default 8000 tokens). The adopt guide shows how to shrink a big AGENTS.md.

Debugging hooks

Start Claude Code with AGENT_KIT_HOOK_LOG=/tmp/agent-kit-hooks.log: every agent-kit hook appends one line when it runs.

Guardrails

Hooks and checks that enforce the rules instead of asking for them.

How the kit improves

From projects to the kit and back — retros, proposals, the weekly scout, releases.