baldart 4.33.1 → 4.33.2

package/CHANGELOG.md

@@ -5,6 +5,16 @@ All notable changes to BALDART will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [4.33.2] - 2026-06-12
+
+**`/new` runs its two purely-mechanical spawns on haiku instead of opus.** Audit of every agent `/new` spawns (not `new2`) found that the spawns inherit each agent's frontmatter default and none ever override `model:` — correct for the reasoning agents (`coder`/`ui-expert` = opus; the review/analysis agents = sonnet), but the **background worktree-setup subagent** that runs `/nw` is a `general-purpose` agent with no `model:`, so it inherited the **session model (opus)** to do pure git plumbing (create worktree, install deps, allocate port, write registry). Same shape for the **file-scoped revert agent** (was spawned as `coder`/opus to do a mechanical revert). Both now run on **haiku** — cheaper and faster on the background barrier, with no quality loss because neither does any reasoning or code authoring. `qa-sentinel` deliberately stays on sonnet (failure interpretation + test-tier selection). `/mw` is invoked inline by the orchestrator (not a subagent) so it is unaffected. No `baldart.config.yml` key involved (model selection is native) ⇒ schema-change propagation rule does not apply. **PATCH** (cost/perf, no new capability, no breaking change).
+
+### Changed
+
+- **`framework/.claude/skills/new/references/setup.md`** — the §4b background worktree-setup subagent (`general-purpose`, runs `/nw`) now spawns with `model: "haiku"`, with a rationale noting it is deterministic plumbing.
+- **`framework/.claude/skills/new/references/implement.md`** — the Phase 2.4 unauthorized-file revert is now a `general-purpose` + `model: "haiku"` agent with an explicit ROLE BOUNDARY (mechanical file-scoped revert, never code authoring), replacing the prior `coder`/opus revert spawn.
+- **`framework/.claude/agents/REGISTRY.md`** — the Model Selection Matrix "never haiku" rule is scoped to **named specialist agents**, with an explicit **plumbing carve-out** sanctioning `model: "haiku"` for mechanical `general-purpose` plumbing (the two `/new` uses above); `qa-sentinel`'s stay-on-sonnet rationale made explicit.
+
 ## [4.33.1] - 2026-06-12
 
 **Discoverability follow-up to v4.33.0** — the new `stack.schema_deploy_from_trunk_only` gate guard was enforced in the `/new` reference modules + `new2.js`, but the **"Reads from `baldart.config.yml`"** manifest line at the top of each skill body still omitted it. Surface it so the key is discoverable from the skill's config-reads declaration (parity with `stack.database` / `stack.deployment`). Manifest-only — **no behavior change**. **PATCH**.

package/VERSION

@@ -1 +1 @@
-4.33.1
+4.33.2

package/framework/.claude/agents/REGISTRY.md

@@ -198,7 +198,9 @@ Use this table when spawning agents via the `Task` tool. The `model` field in ea
 | **hybrid-ml-architect** | opus | — (always opus) | ML architecture is complex by nature |
 | **senior-researcher** | opus | — (always opus) | Research depth requires strongest model |
 
-**Rules**: Never use haiku for any agent. Opus for code writing and creative/complex work. Sonnet for analysis, review, and documentation.
+**Rules**: Never use haiku for any **named specialist agent** in the matrix above. Opus for code writing and creative/complex work. Sonnet for analysis, review, and documentation. `qa-sentinel` stays on sonnet even though it is a mechanical gate runner — its failure interpretation and SCOPED-vs-FULL tiering benefit from sonnet.
+
+**Plumbing carve-out (haiku allowed):** the "never haiku" rule above governs the specialist agents. **Mechanical plumbing spawned as `general-purpose`** — with no reasoning and an explicit ROLE BOUNDARY — MAY use `model: "haiku"` via override. Two sanctioned uses today, both in `/new`: (1) the background **worktree-setup** subagent that runs `/nw` (create worktree, install deps, allocate port, write registry — `references/setup.md` §4b); (2) the **file-scoped revert agent** that restores unauthorized files to their pre-commit state (`references/implement.md` Phase 2.4). Both are deterministic git/file ops, never code authoring.
 
 ## Notes
 

package/framework/.claude/skills/new/references/implement.md

@@ -358,4 +358,4 @@
    (`$TRUNK` is the trunk branch resolved in Phase 0; if the worktree branched from a different base, the `HEAD~1..HEAD` fallback covers the last commit.)
    Compare against this card's allowed files in `## File Ownership Map`.
    - **All within allowed set** → proceed to Phase 2.5.
-   - **Any file outside allowed set** → log the violation in `## Issues & Flags` (`"unauthorized file: <path>"`), then spawn a **targeted revert coder agent** with instruction: "Revert ONLY these files to their pre-commit state. Do not touch any other file: [list unauthorized files]". Re-run build + lint to confirm clean state after revert. Update tracker with revert outcome, then proceed to Phase 2.5.
+   - **Any file outside allowed set** → log the violation in `## Issues & Flags` (`"unauthorized file: <path>"`), then spawn a **targeted revert agent** (`subagent_type: general-purpose`, **`model: "haiku"`** — a file-scoped revert is mechanical git plumbing, not code authoring, so it does NOT need `coder`/opus; this is the sanctioned plumbing carve-out to REGISTRY.md's "never haiku" rule) with an **explicit ROLE BOUNDARY**: "Mechanical file-scoped revert only — restore ONLY these files to their pre-commit state, do NOT write, fix, or refactor any code. Do not touch any other file: [list unauthorized files]". Re-run build + lint to confirm clean state after revert. Update tracker with revert outcome, then proceed to Phase 2.5.

package/framework/.claude/skills/new/references/setup.md

@@ -270,7 +270,7 @@
       - **`WT_PATH` set AND `registry.json` has a complete code entry for this slug** (a finished prior run — `buildVerified` recorded) → **resume**: read `path`/`branch`/`port`/`createdAt`/`buildVerified` from that entry, skip to step 6; re-run the baseline as a single background `Bash` (output to `/tmp`) **only if** `buildVerified` is not `true`.
       - **`WT_PATH` set but NO complete registry entry** (a prior attempt interrupted mid-setup — the normal compaction-mid-barrier state: worktree created, build unfinished, entry never written) → it is a half-built orphan with **no card work** (we are still in pre-flight, zero commits). **Reset clean and recreate**: `git -C "$MAIN" worktree remove --force "$WT_PATH"` then `git -C "$MAIN" branch -D "$WT_BRANCH"` (both ignore-if-absent), then proceed to 4b. A pre-flight worktree has nothing to lose, so a clean recreate is always safe — and it sidesteps the fail-loud collision a naive re-spawn would hit.
       This — detection by `git worktree list`, not the lagging registry — is what makes the deferred-flush pre-flight genuinely idempotent across compaction.
-   b. **Spawn ONE background subagent** (Agent tool, **`mode: "bypassPermissions"`** — mandatory per the SKILL.md meta-rules; a background agent that hit a permission prompt with no human present would stall the barrier forever — `run_in_background: true`, `name: "worktree-setup-<FIRST-CARD-ID>"`, a subagent type that can use the Skill tool — `general-purpose`) whose ENTIRE mission is to run `/nw` and return the block below. Briefing:
+   b. **Spawn ONE background subagent** (Agent tool, **`mode: "bypassPermissions"`** — mandatory per the SKILL.md meta-rules; a background agent that hit a permission prompt with no human present would stall the barrier forever — `run_in_background: true`, `name: "worktree-setup-<FIRST-CARD-ID>"`, a subagent type that can use the Skill tool — `general-purpose`, **`model: "haiku"`** — this mission is pure deterministic plumbing (create the worktree, install deps, allocate the dev-server port, write the registry entry) with no reasoning; haiku is sufficient and is faster on the barrier, and it is the sanctioned plumbing carve-out to REGISTRY.md's "never haiku" rule) whose ENTIRE mission is to run `/nw` and return the block below. Briefing:
       ```
       Invoke the worktree-manager skill in `/nw` programmatic mode with:
         { cards: [<all card IDs>], groupParent: <PARENT-ID|null>, slug: "<slug>" }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "baldart",
-  "version": "4.33.1",
+  "version": "4.33.2",
   "description": "Claude Agent Framework - Reusable framework for coordinating AI agents and humans in software projects",
   "bin": {
     "baldart": "./bin/baldart.js"