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

Why

Per Human direction: CLAUDE.md should contain ONLY 100% useful content (loaded into every bro session — every line costs tokens on every prompt). The consultant routing path is an optional use case — most prompts don't need it. So the row I committed in 4a4176c ("Question — expertise or architectural trade-off → Spawn a consultant via tmb_agent-creator") belongs in a slash command or skill, not the always-loaded persona.

Existing commands: scan, onboard, monitor, roundtable. No consultant command exists.

Scope

Part 1 — Add /tmb:agent-create slash command

New file: commands/consultant.md. Register via command_register MCP call (or add to schema seed). Pattern follows existing commands.

The slash command is the EXPLICIT user-typed path. The implicit autoload (description-match on tmb_agent-creator) handles the case where Human types a natural-language consultant ask without remembering the slash command.

Argument shape

/tmb:agent-create <ask>

Example: /tmb:agent-create should we break the auth service out of the monolith

Behavior

  1. Load tmb_agent-creator skill (the command's body explicitly invokes it).
  2. Branch through the existing tmb_agent-creator algorithm: agent_list → resolve target agent (named or un-named) → spawn via Agent.
  3. Surface consultant's discussion_append(kind='analysis') back to Human.
  4. Per the existing skill doctrine: Human decides based on the consultant's analysis — never the consultant.

Args schema

{"argument_hint": "<ask, free-text>"}

Part 2 — Remove the CLAUDE.md routing row

Once Part 1 lands and is verified working:

  1. Delete the row I added in commit 4a4176c from plugin/CLAUDE.md: 
    | **Question — expertise or architectural trade-off** ... | Spawn a consultant via `tmb_agent-creator` ... |
  2. The implicit autoload path stays (description-match on tmb_agent-creator covers naturalistic prompts).
  3. The explicit user-typed path is now /tmb:agent-create.

Part 3 — Doctrine update

Update docs/architecture/FLOWS.md C section (Consultant invocation) to mention the slash command:

Human types /tmb:agent-create <ask> OR a naturalistic question that matches tmb_agent-creator's description. Either path loads the skill, which runs the agent_list → resolve → spawn ceremony.

Acceptance criteria

  • /tmb:agent-create <ask> runs the consultant ceremony end-to-end (agent_list, register if needed, spawn via Agent, analysis written via discussion_append)
  • commands/consultant.md exists and is registered (visible via command_list MCP call)
  • CLAUDE.md routing-row REMOVED (per Part 2)
  • Naturalistic prompts still work via description-match (regression check: bash tests/dogfood/run-l5.sh 10-consultant passes)
  • New L5 row OR existing 10-consultant adapted to test the slash-command path explicitly
  • docs/architecture/FLOWS.md C section updated
  • L0-L4 green

Out of scope

  • Refactoring tmb_agent-creator skill itself
  • Adding more slash commands beyond /tmb:agent-create
  • Removing other CLAUDE.md content (separate audit for "100% useful content" sweep)

Constraints

  • DO NOT touch CLAUDE.md until Part 1 is verified (per feedback_claude_md_owned_by_human — Human-authored; coordinate before edit)
  • The new command + the existing tmb_concerns-protocol + tmb_agent-creator must not duplicate logic; the slash command body just loads the skill, doesn't re-implement the ceremony

Filed during

v0.7.0 release-prep, 2026-05-20. Companion to the coverage-extension issue. Block: wait for feat/sandbox-and-tri-source-schema to merge first.

Edited by Zax Shen