3t-init

Initialize this project for 3-tier agents. The plugin supplies the protocol, skills, and the implementor agent — this skill only scaffolds the per-project writable files and turns the project's activation marker on.

Step 1 — Create directories

mkdir -p .claude/context/cold docs/adr

Step 2 — Copy the blank memory templates (skip any that exist)

The blank templates ship with the plugin. Copy each only if it does not already exist (never clobber a populated memory file):

SRC="${CLAUDE_PLUGIN_ROOT}/templates"
for f in MEMORY.md CONTEXT.md EXECUTOR_MEMORY.md IMPLEMENTOR_MEMORY.md OVERRIDE_LOG.md; do
  [ -f ".claude/context/$f" ] || cp "$SRC/$f" ".claude/context/$f"
done
[ -f ".claude/context/cold/INDEX.md" ] || cp "$SRC/cold/INDEX.md" ".claude/context/cold/INDEX.md"

.claude/context/cold/INDEX.md is the activation marker — once it exists, the plugin's SessionStart hook engages the 3-tier protocol in this project (and stays dormant everywhere else). It is committed, so a fresh clone of an already-3t project auto-activates without re-running this skill. That is the per-project opt-in switch.

Step 3 — Write settings.json (model config)

If .claude/settings.json does not exist, ask the user which model setup they want, then write it. Default recommendation: opusplan + advisor on Opus.

{
  "model": "opusplan",
  "advisorModel": "opus"
}

If .claude/settings.json already exists, do NOT overwrite it. Instead show the two keys and tell the user to merge them in manually: "model": "opusplan", "advisorModel": "opus".

The advisor is the architect tier — advisorModel is what makes hard/irreversible design decisions route to Opus. Confirm it is set before finishing.

Step 4 — Update .gitignore

Append (only if not already present) the per-developer memory exclusions. CONTEXT.md and cold/ stay committed — they are shared institutional knowledge.

grep -q "CLAUDE.local.md" .gitignore 2>/dev/null || cat >> .gitignore <<'EOF'

# Claude Code 3-tier — per-developer memory (not version-controlled)
.claude/context/MEMORY.md
.claude/context/EXECUTOR_MEMORY.md
.claude/context/IMPLEMENTOR_MEMORY.md
.claude/context/OVERRIDE_LOG.md
.claude/CLAUDE.local.md
# 3-tier transient flags (session activation + deactivation)
.claude/.3t-active
.claude/.3t-disabled
EOF

Step 5 — Confirm

Show the user the resulting tree and tell them:

Project initialized for 3-tier agents.

  .claude/
  ├── settings.json              ← model config (commit)
  └── context/
      ├── CONTEXT.md             ← domain glossary + ADRs (commit)
      ├── cold/INDEX.md          ← cold-storage index (commit)
      ├── MEMORY.md              ← activation marker + project state (gitignored)
      ├── EXECUTOR_MEMORY.md       ← delegation lessons (gitignored)
      ├── IMPLEMENTOR_MEMORY.md        ← execution lessons (gitignored)
      └── OVERRIDE_LOG.md        ← override tracking (gitignored)
  docs/adr/                      ← ADRs authored by the implementor

The protocol, slash commands, and the implementor agent come from the
claude-3t plugin — nothing else to install.

Restart Claude here (so the SessionStart hook picks up the new marker),
then run /claude-3t:3t-start to begin.
The canonical flow: /grill-with-docs → advisor → /to-prd → /to-issues → execution