v0.7.0 — step 11 (roundtable) restructure: from-scratch agent via /tmb:agent-create with base template + roundtable of mixed templated + scratch consultants

[ZaxShen]

Why

Per Human direction, step 11 row needs to test the FROM-SCRATCH agent-create path (Branch C — base template.md) AND verify a roundtable works with a mix of templated + scratch consultants. Sibling to #50/!2937/GH#198 (which restructures step 10 — /tmb:agent-create cto template-copy).

Current step 11 (per tests/dogfood/rows/11-roundtable/README.md) tests /roundtable with architect+cto+pm already registered as project-local. The restructure folds in the create-from-scratch path.

Scope

L5 mode

Pre-state via setup-l5.sh:

• Copy templates/agents/cto.md → .claude/agents/cto.md (i.e. seed cto as already-registered; simulates "step 10's template-copy already happened")
• Register cto in DB via SQL insert (mirror what tmb_agent-creator would have written)

Turn 1: Human types /tmb:agent-create <NEW-NAME> where <NEW-NAME> is NOT in templates/agents/. Suggested: legal-reviewer (already cited in tmb_agent-creator description as un-templated example; clean domain for roundtable). Bro runs Branch C: scaffolds from base templates/agents/template.md → AUQ for role specifics → writes .claude/agents/legal-reviewer.md → agent_register → tmb_agent_created audit row.

Turn 2: Human types /roundtable <topic that needs both cto + legal-reviewer>. Bro orchestrates: roundtable_create(participants=[cto, legal-reviewer], topic=...) → spawns each via Agent → each writes discussion_append(kind='analysis') + roundtable_vote.

L6 mode (chain)

Pre-state inherits from chain step 10:

• cto is ALREADY registered as project-local (step 10 left this state)
• NO setup-l5.sh runs (chain mode ignores it)

Turn 1: same as L5 Turn 1 — /tmb:agent-create legal-reviewer → from-scratch ceremony.
Turn 2: same as L5 Turn 2 — /roundtable with cto + legal-reviewer.

Only difference between L5 + L6: who seeded cto. L5 = setup-l5.sh, L6 = step 10 chain inheritance.

Scorer updates

• tools-required:
  • Turn 1: Skill (for /tmb:agent-create), agent_register, Agent (when later spawned), agent_list (Branch C reads registry to check name collision)
  • Turn 2: mcp__plugin_tmb_trajectory-server__roundtable_create, mcp__plugin_tmb_trajectory-server__roundtable_vote (×N participants), Agent (×N spawns)
• outcome.sql additions:
  • assert legal-reviewer registered in agents table with scope='project-local' (proves from-scratch path completed)
  • assert tmb_agent_created audit row with content_json LIKE '%legal-reviewer%' AND content_json LIKE '%from-scratch%' (or 'mode':'from-scratch')
  • assert roundtables table has ≥1 row with participants including both cto + legal-reviewer
  • assert discussion_append(kind='analysis') rows from BOTH cto + legal-reviewer
  • assert roundtable_votes table has rows from BOTH agents
• outcome-coherence.json: agent count ≥3 (swe + pr-reviewer + cto + legal-reviewer), roundtable_votes ≥2
• script.json: max_turns=2

Files to update

• tests/dogfood/rows/11-roundtable/prompt.txt — turn-1 + turn-2 prompts
• tests/dogfood/rows/11-roundtable/setup-l5.sh — seed cto (L5 only)
• tests/dogfood/rows/11-roundtable/script.json — max_turns=2, user_after_bro for turn 2
• tests/dogfood/rows/11-roundtable/outcome.sql — new assertions
• tests/dogfood/rows/11-roundtable/outcome-coherence.json — coherence checks
• tests/dogfood/rows/11-roundtable/README.md — rewrite for two-phase from-scratch + roundtable scenario
• tests/EVALUATION.md — update row 11 entry

Dependencies

• BLOCKS on v0.7.0 — add /tmb:agent-create command + refactor templates to Lego model (base template + slim role bodies) + drop redundant CLAUDE.md routing row #198 / !2937 / issue MCP role enforcement gap: task/validation/issue/ledger/audit/skill tools lack requireRoles wrappers #50 landing first (need /tmb:agent-create command + base template.md + tmb_agent-creator from-scratch ceremony updates)
• Can land as ONE MR with v0.7.0 — add /tmb:agent-create command + refactor templates to Lego model (base template + slim role bodies) + drop redundant CLAUDE.md routing row #198 work OR as a follow-up MR; SWE judgment

Acceptance criteria

• L5 step 11 passes: cto seeded → /tmb:agent-create legal-reviewer (from-scratch) → /roundtable orchestrates with both → all scorer assertions green
• L6 chain step 11 passes: same scenario with cto inherited from step 10 chain state
• L0-L4 stay green
• v0.7.0 stable gated on this (per feedback_rc_policy — no rc until L6 14/14)

Out of scope

• Adding more roundtable participants beyond cto + legal-reviewer (architect + pm dropped from current 3-way to keep test focused)
• Multi-turn deliberation (roundtable_finalize_decisions) — single round of analyses + votes is sufficient for this test
• The roundtable-slash-detect.sh hook coverage stays as it is

Filed during

v0.7.0 release-prep, 2026-05-20. Companion to #50 (/tmb:agent-create + Lego refactor) and #45-derived test-infra work. Blocks v0.7.0 stable.