feat: Phase 3 — bootstrap commands, skills, and state issue

.gitignore

@@ -1,5 +1,10 @@
 .DS_Store
 .claude/settings.local.json
+.claude/settings.json
 .claude/worktrees/
 node_modules/
 *.tsbuildinfo
+
+# middle's per-repo bootstrap dir is local-only (mm init step 7). Skills under
+# .claude/skills/ stay committed; the operational .middle/ does not.
+.middle/

docs/dogfooding.md

@@ -0,0 +1,84 @@
+# Dogfooding crossover — operator runbook
+
+From Phase 3 onward, middle dispatches its own work. The crossover is the moment
+the operator runs `mm init` **on the middle repo itself** and hands the wheel to
+middle. The spec calls this out as a deliberate chicken-and-egg step
+(build spec → "Dogfooding rules" #5): bootstrap is manual up to here, and the
+crossover is performed once, by hand, by the operator — not by an agent that is
+itself mid-dispatch.
+
+## Why an in-flight agent must not self-run the crossover
+
+When middle dispatches an agent for an Epic, the agent runs inside a worktree
+whose `.middle/` directory holds that dispatch's operational state (the brief,
+the hook script, session bookkeeping). The crossover commands conflict with that
+in two load-bearing ways:
+
+1. **`mm dispatch . <issue>` nests agents.** It needs the dispatcher process
+   (`mm start`, port 8822), creates a *second* worktree, and spawns a *second*
+   `claude`/tmux session. Run from inside a live dispatch it collides with the
+   running dispatcher and consumes a slot against itself. It also needs a
+   **manually created** Epic to target.
+2. **`mm init .` creates live GitHub state.** It opens the real
+   `agent-queue:state` issue and label on `thejustinwalsh/middle`. Done from an
+   unmerged PR branch, that issue is orphaned if the PR is reworked. The state
+   issue is meant to be "real from Phase 3" (dogfooding rule #2) — created once,
+   deliberately, at the canonical checkout.
+
+So the capability is built and verified (PR #81, sub-issues #22–#25); the live
+crossover is the operator's to run. The committable slice of the self-bootstrap —
+gitignoring `.middle/` (mm init step 7) — is already in the repo.
+
+## Runbook
+
+Run from a **clean** checkout of `thejustinwalsh/middle` (not a middle worktree):
+
+```bash
+# 0. Preconditions — all green.
+mm doctor
+
+# 1. Inspect the plan first (no mutation).
+mm init . --dry-run
+
+# 2. Perform the crossover. Creates the agent-queue:state label + state issue,
+#    stages skills to .claude/skills/ and .codex/skills/, writes .middle/hooks/
+#    hook.sh, per-CLI hook config, and .middle/config.toml (with the issue number).
+mm init .
+
+# 3. Commit the shared skills (.claude/skills/ and .codex/skills/ are committed;
+#    .middle/ is gitignored).
+git add .claude/skills .codex/skills .gitignore
+git commit -m "chore: install middle skills (dogfooding crossover)"
+
+# 4. Verify the state issue exists and conforms.
+gh issue list --repo thejustinwalsh/middle --label agent-queue:state
+#    Its body parses against schemas/state-issue.v1.md (validated by mm init).
+
+# 5. Start the dispatcher and dispatch a manually created Epic against middle.
+mm start
+mm dispatch . <epic-number>     # the Epic (see "Creating a dispatchable Epic" below)
+
+# To reverse the crossover entirely:
+mm uninit .                     # closes the state issue, removes staged files
+```
+
+### Creating a dispatchable Epic
+
+`mm dispatch . <epic-number>` expects an **Epic**: a GitHub issue that has
+**sub-issues**, where each sub-issue is one phase of the work and carries its own
+acceptance criteria. The agent works the open sub-issues down, one per phase, on a
+single branch/PR.
+
+Don't hand-author these ad hoc — use the **`creating-github-issues`** skill (in
+`.claude/skills/creating-github-issues/`): give it a plan and it files the parent
+Epic plus sub-issues with acceptance criteria, the right labels, and the proper
+parent/child hierarchy that the recommender and implementer expect. Minimum the
+dispatch needs:
+
+- a **parent** issue with one or more **sub-issues** attached (the phases),
+- **acceptance criteria** on each sub-issue (an Epic whose children lack them is
+  classified `needs-human` and won't be dispatched),
+- optionally a `phase:N` / `dogfood` label for grouping (not required to dispatch).
+
+`mm init` is idempotent: a re-run with a matching `bootstrap.version` refreshes
+skills/hooks but keeps the config and the existing state issue.

package.json

@@ -10,7 +10,9 @@
   ],
   "scripts": {
     "test": "bun test",
-    "typecheck": "tsc --noEmit"
+    "typecheck": "tsc --noEmit",
+    "sync-skills": "bun scripts/sync-skills.ts",
+    "prepare": "git config core.hooksPath scripts/hooks 2>/dev/null && echo 'middle: enabled repo git hooks (core.hooksPath=scripts/hooks)' || true"
   },
   "devDependencies": {
     "@types/bun": "latest",

packages/cli/package.json

@@ -11,6 +11,8 @@
     "@middle/adapter-claude": "workspace:*",
     "@middle/core": "workspace:*",
     "@middle/dispatcher": "workspace:*",
-    "commander": "^14.0.3"
+    "@middle/state-issue": "workspace:*",
+    "commander": "^14.0.3",
+    "smol-toml": "^1.6.1"
   }
 }