[luuuc]

Summary

brAIn shipped every part of the memory loop and nothing ever called it — a session
started cold unless someone remembered to type brain recall first. This adds the
editor integration that closes it: a session-start hook that loads project memory
automatically, plus the rules covering what a hook cannot know.

Pitch

.doc/pitches/02-04-claude-code-hooks.md — local only, since .doc/ stays out of
the public repo.

Changes

New command

• brain hooks prints Claude Code's settings.json block and the CLAUDE.md rules;
  --tool cursor prints .cursor/rules/brain.mdc
• brain hooks session-start injects the top five memories, ranked by layer authority

Display

• Memory.Title() is now shared by recall, the MCP tools, and the hook. Effectiveness
  memories no longer render as ## Outcomes

This repository

• .brain/ seeded with sixteen memories and committed
• README documents the setup path and what a fresh clone actually gets

Architecture Highlights

• The hook is a subcommand, not a shell script. Shaped as a script; a script adds
  an install step, a chmod, a bash dependency, and a test that has to reconstruct a
  hostile PATH to exercise ten lines.
• session-start always exits 0. A SessionStart hook exiting 2 blocks the session,
  and brain recall exits 2 on an empty store — an ordinary state, not an error.
• No automatic trust recording. The placeholder called for a post-commit hook. A
  local commit is not a ship: ten of them would walk a domain from ask to notify on
  typing alone. Recording stays with merges, deploys, and Beacon.
• brAIn prints, the human installs. It never edits a settings file someone else owns.
• The matcher covers startup|clear|compact. Compaction matters most — it is where
  a correction gets summarised away. resume and fork inherit a context that already
  has the injection.

Test Plan

• Session start exits 0 and prints nothing with no .brain/, an empty one, or a stale --dir
• With memories present, output is ordered correction → decision → fact and capped at five
• Effectiveness memories render as persona and domain, never as ## Outcomes
• The settings snippet parses as JSON, carries a timeout, and the command it names executes
• Every brain <command> in the printed setup and in the README resolves against the real
  command tree — mutation-checked by injecting brain verify, which fails the test
• Cursor frontmatter parses as YAML with alwaysApply: true; the Cursor setup never claims
  a hook loads memory
• Unknown --tool exits 3
• make ci passes (build + test + lint, 0 issues); coverage internal/cmd 79.5% → 82.6%,
  internal/memory 0% → 62.5%

Follow-ups

• Two definition docs are stale: 08-ai-workflow.md shows .cursorrules, which is absent from
  Cursor's current docs, and 06-mcp-and-cli.md documents brain recall --for council, a flag
  that has never existed.
• Seeding surfaced that corrections outrank by layer rather than relevance, so a narrow
  correction permanently occupies a session-start slot ahead of any decision. Worth its own pitch.