🏛️ feat: Lego templates — plugin ships zero global agents

[ZaxShen]

Summary

Implements #60 — the Lego templates restructure.

Plugin ships ZERO global agents. Bro is the only persona (in CLAUDE.md). Every other agent (swe, pr-reviewer, architect, cto, ceo, pm) is a Lego template that bro copies into <project>/.claude/agents/ on demand, with explicit Human approval. Project customization happens by extending the agent's skills: array via tmb_skill-creator — never by editing the agent body.

Lego rule

Layer	Mutability
Template body	Immutable — bro copies verbatim, never edits
skills: array	Additive — extended by tmb_skill-creator
Spawn prompt	Ephemeral — fresh per call

What's in this PR

New templates (6 agents, ≤25 lines body each)

templates/agents/ — swe, pr-reviewer, architect, cto, ceo, pm. Each is identity + role + boundary + skills: []. The 30-line Lego cap is enforced by tests/lint/agent-line-budget.sh.

New skill templates (7 default skills)

templates/skills/ — swe-checklist, review-protocol, review-findings, code-quality, docs-conventions, git-conventions, naming-conventions. Bootstrap copies these into the project alongside swe + pr-reviewer.

tmb_ prefix on all 14 plugin protocol skills

Plus 2 NEW protocol skills:

• tmb_bootstrap — first-run template copy (with Human approval)
• tmb_skill-creator — generates project skills, extends agent skills: arrays (NEVER edits agent body)

tmb_agent-creator rewritten with template-copy mode (PRIMARY) + from-scratch fallback.

CLAUDE.md updates

• Bootstrap-check added to bro's first-action chain
• Code-touching chain references tmb_* skill names
• Routing table now describes the template-copy-on-demand flow
• New 'Templates shipped with the plugin' section enumerates all 6

MCP role enforcement (still in place from #59)

• task_create_batch, issue_create, issue_close → bro only
• task_update_status → bro + swe
• validation_record → pr-reviewer only
• discussion_append → bro + swe + pr-reviewer + any consultant

Layer 1 + 2 status: GREEN

• Onboarding contract lint: PASS
• Lego cap (≤30 lines on templates/agents/): PASS — all ≤25 lines
• MCP server unit tests: 0 failures
• MCP integration suite: 0 failures
• Hook tests: PASS

Test plan

• Layer 1 lint
• Layer 2 MCP unit + integration + hook tests
• Layer 3 dogfood — see tests/manual/scenarios.md. Specifically:
  • 1.1 (onboarding) — should NOW also trigger tmb_bootstrap after onboarding completes (project gets swe + pr-reviewer + default skills copied)
  • 2.1.bro (canonical simple-task smoke test) — should work end-to-end with the project-local swe + pr-reviewer
  • C.1 (consultant invocation) — should now trigger tmb_agent-creator template-copy mode for architect/cto/ceo/pm

Followup

• Open question: should agents have a shared skill bag (inheritance) or pure per-agent listing? #61 — open question on agent-shared skill bags
• Open question: should plugin templates be split into a separate marketplace channel? #62 — open question on splitting templates into a separate marketplace channel

🤖 Generated with Claude Code

[ZaxShen]

Layer 3 dogfood verification — end-to-end PASS

Walked the canonical scenario in fresh Mode B (`/tmp/tmb-smoke`) against this branch. First time the bro→swe→pr-reviewer chain has completed end-to-end with a clean lifecycle.

Trigger

Wiped `/tmp/tmb-smoke`, fresh `git init`, fresh Mode B session, fresh DB. Single ask: `@bro write a cli todo by Python`.

Trajectory log (from MCP DB monitor)

```
19:29:53 identity ZaxShen written
19:29:56 config branching_model="github-flow", pr_target="main"
19:29:56 config protected_branches=["main"] ← raw JSON array, NOT "[\"main\"]" — config_set defense fix verified
19:54:43 issue 1 created (objective: "Build a Python CLI todo app")
19:55:13 discussion 2 (note): "Beginning planning on branch_id feat/cli-todo, triage: simple"
← updated tmb_branch-id-proposal wording (was "Routed to architect" pre-cleanup)
19:55:30 ledger 1 (scope_gate_waived): "simple-triage personal CLI: defaulted to argparse + unittest..."
← simple-fast-lane waiver fired correctly
19:55:31 task 1 (feat/cli-todo, pending) created via task_create_batch
19:55:34 ledger 2 (planning_complete)
──────── 51s planning phase total ────────
19:59:10 ledger 3 (tmb_bootstrap_complete): "Copied 2 agent templates (swe, pr-reviewer) + 7 skill templates"
← bro detected missing .claude/agents/swe.md, invoked tmb_bootstrap, AskUserQuestion approved, copied verbatim
20:03:48 task 1 → completed (SWE atomic close with commit_sha)
← worktree at .claude/worktrees/agent-acb891639c9f96486, committed:
── caeb93d ✨ feat(cli-todo): minimal argparse + JSON-store todo CLI
── files: todo.py, test_todo.py, README.md
20:06:44 validation 1 (task_id=1, attempt_n=1, verdict=pass) ← pr-reviewer signed off
20:06:44 task 1 → closed (bro flipped status atomically)
20:08:11 issue 1 → closed (bro called issue_close after all tasks closed)
```

Design checkpoints — all green

Checkpoint	Status
Onboarding on fresh DB (identity + 3 config writes)	✓
`protected_branches` stored as raw `["main"]` array (not double-encoded `"[\"main\"]"` — config_set defense)	✓
Updated routing-note wording ("Beginning planning on..." not "Routed to architect on...")	✓
Simple-fast-lane: no env probe, no ceremonial Q+A, defaults table picked	✓
Scope-gate waiver with valid reason logged to ledger	✓
`tmb_bootstrap` auto-invoked on missing `.claude/agents/swe.md`	✓
Template copy verbatim (2 agents + 7 skills, no body edits)	✓
Bro→swe spawn finds project-local agent	✓
Worktree isolation + atomic commit + status='completed'	✓
Bro→pr-reviewer spawn + `validation_record(verdict='pass')`	✓
Bro flips task → `closed` after pass verdict	✓
Bro flips issue → `closed` after all tasks closed	✓
Zero `forbidden` events (MCP `requireRoles` enforcement clean)	✓
Every state transition logged to ledger (replayable)	✓

Lifecycle final state

```
issue.status: open → closed
task.status: pending → completed → closed
verdict: pass (pr-reviewer)
ledger events: scope_gate_waived → planning_complete → tmb_bootstrap_complete
```

Every state machine ended in a terminal position. No orphan `running` tasks. No missing verdicts. No abandoned ledger entries.

Comparison to baselines

Run	Total time	Outcome
Original `bro→architect→swe` chain (pre-#58)	7+ minutes	Hung at SWE spawn, never completed, Ctrl-C didn't interrupt
Pure Claude Code on the same ask (no plugin)	~32s	Code generated, no audit, no governance
This run (this PR's Lego architecture)	~12min total (one-time bootstrap + chain), `~8.5min` projected for subsequent tasks in same project	Closed cleanly with full audit + role enforcement

The runtime is meaningfully heavier than pure Claude — multi-agent + audit trail + gates aren't free — but the chain now actually completes, every transition is replayable from SQLite, and consultants can never write workflow state by construction. Subsequent tasks in the same project skip bootstrap entirely.

What this verifies for #60

• ✅ Plugin ships zero global agents (`agents/` is empty)
• ✅ `tmb_bootstrap` skill copies templates verbatim, no body edits
• ✅ All `tmb_*` skill renames work — no name collisions, all references resolve
• ✅ Bro's first-action chain correctly invokes `tmb_bootstrap` when `.claude/agents/swe.md` absent
• ✅ Lego rule enforced in practice: agent files copied verbatim from templates, no edits
• ✅ Layer 3 dogfood confirms what Layers 1 + 2 asserted statically

Open follow-ups (not blockers for this PR)

• The 3.5min interactive bootstrap pause is wall-clock from the AskUserQuestion approval round-trip; that's user-bound, not chain-bound.
• pr-reviewer ran in ~3min including bro's wrapper time; could potentially be tightened in a future iteration but doesn't block.
• No `forbidden` events surfaced because the chain stayed in role; the negative-path tests in role-matrix.test.mjs cover the rejection case at Layer 2.

Recommend merging. Lego templates work end-to-end in dogfood.