# Harness-Free Mode B Amendment Implementation Plan

> **For agentic workers:** Execute this plan task-by-task with
> `superpowers:executing-plans`; use `superpowers:test-driven-development` for the build behavior
> change and `superpowers:verification-before-completion` before reporting results.

**Goal:** Restore Gradle bootstrap and Clean Architecture dependency enforcement without recreating
the absent development harness.

**Architecture:** One strict JSON registry under `src/config/architecture/` owns all 19 leaf
identities, paths, and allowed production project edges. Gradle settings validate and include the
registry fail-closed; the root dependency verification task reads the same file and checks actual
production project dependencies against it.

**Tech Stack:** Gradle Groovy DSL, Groovy `JsonSlurper`, strict JSON, Java 21.

**Working policy:** Human-only git handling. Do not stage, commit, amend, push, or create a PR.

---

### Task 1: Capture the broken bootstrap

**Files:**

- Read: `src/settings.gradle`

- [x] Run `cd src && ./gradlew help --console=plain`.
- [x] Confirm exit 1 is caused by the missing `.harness/project/modules.yaml`, not dependency
  resolution or an unrelated build failure.

### Task 2: Add the Gradle-owned registry

**Files:**

- Create: `src/config/architecture/modules.json`
- Read: each of the 19 leaf-module `build.gradle` files

- [x] Record exactly 19 unique module IDs, Gradle paths, and repository-relative source paths.
- [x] Set `allowed_dependencies` from each leaf's current `api`, `implementation`, `compileOnly`,
  and `runtimeOnly` project dependencies.
- [x] Exclude test/fixture configurations from production policy and keep `sample-portfolio` a
  fixture consumer that no production leaf may depend on.
- [x] Parse the file with Python's strict JSON parser and compare its edges with the checked-in
  leaf build declarations.

### Task 3: Restore Gradle bootstrap and dependency enforcement

**Files:**

- Modify: `src/settings.gradle`
- Modify: `src/build.gradle`

- [x] Make settings load only `config/architecture/modules.json`.
- [x] Fail closed on a missing registry, wrong root/module/field types, empty values, duplicate
  identities or paths, unsafe path shapes, unknown/self dependencies, count drift, or missing
  source directories.
- [x] Include every registered Gradle path and map it to its repository-root-relative source
  directory.
- [x] Make `verifyCleanArchitectureDependencies` read the same registry without a second module
  list.
- [x] Preserve all-leaf coverage and forbidden-edge checks, explicitly reject a production edge
  to `sample-portfolio`, and replace stale error wording with actionable registry guidance.

### Task 4: Align active repository guidance

**Files:**

- Modify: `AGENTS.md`
- Modify: `CLAUDE.md`
- Modify: `README.md`
- Modify: `src/README.md`
- Modify: all 19 nearest leaf-module `CLAUDE.md` files that name the old registry
- Annotate as superseded: the 2026-07-20 harness design and plan

- [x] Point active topology and allowed-edge guidance to
  `src/config/architecture/modules.json`.
- [x] State that focused commands are derived from the owning Gradle path rather than a task
  packet.
- [x] Keep all eight local HARD-STOP meanings, architecture boundaries, human-only git policy,
  verification discipline, and LLM Wiki capture requirements.
- [x] Make the earlier harness documents explicit historical provenance rather than active
  reconstruction instructions.

### Task 5: Verify from a fresh Gradle invocation

**Files:**

- Verify: all changed files

- [ ] Run `cd src && ./gradlew help --console=plain`.
- [ ] Run `cd src && ./gradlew projects --console=plain`.
- [ ] Run `cd src && ./gradlew verifyCleanArchitectureDependencies --console=plain`.
- [ ] Run a deterministic strict-JSON script proving exactly 19 unique IDs/Gradle paths and
  existing source directories.
- [ ] Run a deterministic comparison between registry edges and leaf production project
  dependencies.
- [ ] Run `git diff --check` and `git status --short`.
- [ ] Report exact exits, any unavailable checks, LLM Wiki capture outcome, and remaining risks
  without claiming the broader Phase A/refactor is complete.