v0.7.0 — add /tmb:agent-create command + refactor templates to Lego model (base template + slim role bodies) + drop redundant CLAUDE.md routing row

[ZaxShen]

Full content on GitLab: https://gitlab.com/trustmybot/plugin/-/issues/2937

(Tri-source sync schema landed in MR !208 but mirror logic not yet bidirectional; this issue is a stub for v0.7.x board review.)

[ZaxShen]

Step 10 restructure added to scope

Per Human direction: step 10 row becomes a two-phase scenario.

Phase 1: /tmb:agent-create cto invocation (uses the new command this issue ships). Bro runs template-copy ceremony → registers cto → tmb_agent_created audit row lands.

Phase 2: Human asks a tech-strategy question about ACTUAL CODE:

• L5 mode: setup-l5.sh pre-seeds an app.py substrate; bro spawns just-created cto consultant against that code
• L6 mode: chain state carries forward code from earlier steps (todo CLI from step 4 → grown by step 5/8 → cto evaluates it)

Scorer updates:

• tools-required: add Skill (for /tmb:agent-create), keep agent_register, Agent, agent_list
• outcome.sql: keep tmb_agent_created assertion; ADD assertion that cto's analysis landed via discussion_append(agent='cto', kind='analysis')
• script.json: max_turns=2 (instead of current 1)

Files to update:

• tests/dogfood/rows/10-consultant/prompt.txt — two-phase prompt
• tests/dogfood/rows/10-consultant/setup-l5.sh — code substrate for L5
• tests/dogfood/rows/10-consultant/script.json — max_turns=2
• tests/dogfood/rows/10-consultant/outcome.sql — assert cto analysis discussion
• tests/dogfood/rows/10-consultant/README.md — rewrite for two-phase scenario
• tests/EVALUATION.md — update row 10 entry

v0.8.0 stable gated on this: no rc until L6 14/14 (per feedback_rc_policy).

Discarded work: prior in-flight fixes on fix/consultant-spawn-prescriptive-in-headless (hook headless suffix + TEST_MODE prefix prescription) targeted the WRONG SHAPE of step 10. Branch deleted. The /tmb:agent-create command restructure is the proper structural fix.

[ZaxShen]

Phase 1 ships structurally; L5 step 10 surfaces compounding bugs

SWE delivered all 7 spec items (Lego base + 6 slim templates + /tmb:agent-create command + Branch C scaffold + step 10 row two-phase restructure). pr-reviewer PASS (8/8 structural).

L5 step 10 pre-push gate FAILS with 1/3 outcome assertions passing. Diagnosed bugs:

Bug 1 — bro doesn't follow skill algorithm step 1

Skill body says step 1 = agent_list(). But bro's first action is a Glob for templates/agents/cto.md (relative to CWD = scratch project). Returns no match → bro concludes "no template found" → goes Branch C (from-scratch) instead of Branch B (template-copy).

Fixed the skill body to explicit ${CLAUDE_PLUGIN_ROOT}/templates/agents/<name>.md path. Bro STILL Glob'd relative path because the LLM's natural pattern is "verify-then-act" not "trust-the-registry". Skill body step 1 needs to be more prescriptive.

Bug 2 — agent_register INSERT OR IGNORE doesn't UPDATE scope

When cto already exists in agents table with scope='template' (from schema seed), agent_register(name='cto', scope='project-local', ...) keeps the existing row unchanged (template scope). The scorer asserts scope='project-local' which can never be satisfied with the current MCP server.

Known issue — was actually documented in the original outcome.sql comment: "agent_register is INSERT OR IGNORE in the server impl, so the cto registry row's scope stays 'template' — the scope-update is a known server-side gap, not a bro-behaviour gap."

Server-side fix needed: agent_register should UPDATE scope when re-registering an existing agent.

Bug 3 — user_after_bro turn 2 never fired

script.json has max_turns: 2 + user_after_bro with the turn-2 prompt. But the test ended with stop_reason: 'end_turn', num_turns: 16 (internal tool uses), and the second user message was never injected. Other rows (misc-reonboard-redirect, misc-roundtable-routing-redirect) use user_after_bro successfully with max_turns: 3 + simpler continuation prompts.

May need: max_turns bumped higher OR collapse to single-turn prompt that combines both phases.

Bug 4 — CC may need /plugin-reload mid-session

Even if turn 2 fired correctly, CC may not discover the newly-created .claude/agents/cto.md without a plugin reload. Agent(subagent='cto') would fail to find the agent. Skill already emits a "Run /plugin-reload" reminder; but the test runs in single claude -p session with no reload between phases.

Options

Option	What	Cost
A	Drop cto-project-local + cto-analysis assertions; ship phase 1 with only audit assertion	Smallest; weakest test
B	Server fix (agent_register UPDATEs scope) + skill body more prescriptive + collapse to single-turn + plugin-reload mid-test	Hours of iteration
C	Accept step 10 as v0.8.1 work + cut v0.8.0 with documented partial coverage	Requires explicit waiver of rc policy

Awaiting Human direction.