Consistency CheckSAFE

Turn Claude Code into a full game dev studio — 49 AI agents, 72 workflow skills, and a complete coordination system mirroring real studio hierarchy.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: consistency-check
description: "Scan GDDs against the entity registry for cross-document conflicts. Grep-first approach targets conflicting sections, different stats."
argument-hint: "[full | since-last-review | entity:<name> | item:<name>]"
user-invocable: true
allowed-tools: Read, Glob, Grep, Write, Edit, Bash, AskUserQuestion, Bash(bash "*/.claude/skills/consistency-check/../../hooks/yaml-helper.sh" resolve_config *)
model: sonnet
---

!`bash "${CLAUDE_SKILL_DIR}/../../hooks/yaml-helper.sh" resolve_config --keys automation,workflow,system_overrides`



# Consistency Check

Detects cross-document inconsistencies by comparing all GDDs against the
entity registry (`design/registry/entities.yaml`). Uses a grep-first approach:
reads the registry once, then targets only the GDD sections that mention
registered names — no full document reads unless a conflict needs investigation.

**This skill is the write-time safety net.** It catches what `/design-system`'s
per-section checks may have missed and what `/review-all-gdds`'s holistic review
catches too late.

**When to run:**
- After writing each new GDD (before moving to the next system)
- Before `/review-all-gdds` (so that skill starts with a clean baseline)
- Before `/create-architecture` (inconsistencies poison downstream ADRs)
- On demand: `/consistency-check entity:[name]` to check one entity specifically

**Output:** Conflict report + optional registry corrections

---

Every `AskUserQuestion` call follows `.claude/docs/automation-modes.md`
(collaborative asks always · guided major-only · autonomous logs and proceeds;
`automation_always_ask` categories always prompt).

**`workflow`** (see `.claude/docs/workflow-modes.md`), applied per GDD at its
**effective** tier — the `system_overrides` row for that GDD's system (its
filename stem) when the block lists one, else the project value:
- `full` — full entity-registry cross-check against all GDD sections.
- `standard` — cross-check against the required sections only. The report's
  Scope line says so, and each value stated for a registered name outside those
  sections is listed as not checked at this tier (Phase 3 applies this).
- `minimal` — not meaningful (no GDDs to check).

## Phase 1: Parse Arguments and Load Registry

**Modes:**
- No argument / `full` — check all registered entries against all GDDs
- `since-last-review` — check only GDDs modified since the last review report
- `entity:<name>` — check one specific entity across all GDDs
- `item:<name>` — check one specific item across all GDDs

**Load the registry:**

```
Read path="design/registry/entities.yaml"
```

If the file does not exist or has no entries:
> "Entity registry is empty. Run `/design-system` to write GDDs — the registry
> is populated automatically after each GDD is completed. Nothing to check yet."

Verdict: **NOT ASSESSED — entity registry empty**. Stop: nothing was compared,
so there is no PASS to give and nothing to write.

Build four lookup tables from the registry:
- **entity_map**: `{ name → { source, attributes, referenced_by } }`
- **item_map**: `{ name → { source, value_gold, weight, ... } }`
- **formula_map**: `{ name → { source, variables, output_range } }`
- **constant_map**: `{ name → { source, value, unit } }`

Count total registered entries. Report:
```
Registry loaded: [N] entities, [N] items, [N] formulas, [N] constants
Scope: [full | since-last-review | entity:name]
```

For `entity:<name>` or `item:<name>`, if the registry has no entry by that name
the verdict is **NOT ASSESSED — `<name>` is not in the registry**: list the
closest registered names and stop. Checking an unregistered name compares
nothing, and that is not a PASS.

---

## Phase 2: Locate In-Scope GDDs

```
Glob pattern="design/gdd/*.md"
```

Exclude the docs that live in `design/gdd/` but are not system GDDs:
`game-concept.md`, `systems-index.md`, `game-pillars.md`, `gameplay-tags.md`,
`entity-registry.md`, `fixture-swap-ledger.md`, `sound-bible.md` and any
`gdd-cross-review-*.md` (the report `/review-all-gdds` writes here).

For `since-last-review` mode:
```bash
git log --name-only --pretty=format: -- design/gdd/ | grep "\.md$" | sort -u
```
Limit to GDDs modified since the most recent `design/gdd/gdd-cross-review-*.md`
file's creation date.

Report the in-scope GDD list before scanning. **If the list is empty** — no system
GDDs exist, or `since-last-review` found none changed — the verdict is
**NOT ASSESSED — no GDDs in scope**: say which filter produced zero, and stop.
A scan of nothing finds no conflicts, and that is not a PASS.

---

## Phase 3: Grep-First Conflict Scan

For each registered entry, grep every in-scope GDD for the entry's name.
Do NOT do full reads — extract only the matching lines and their immediate
context (-C 3 lines).

This is the core optimization: instead of reading 10 GDDs × 400 lines each
(4,000 lines), you grep 50 entity names × 10 GDDs (50 targeted searches,
each returning ~10 lines on a hit).

**In a GDD whose effective tier is `standard`, compare only hits under a
required section.** The effective tier is per GDD: `system_overrides.<stem>`
replaces the project tier for that GDD in either direction (a `full` row makes
every section compared). Grep each such GDD once for `^## ` to get its headings
and line numbers, and file every name hit under the nearest heading above it:
only hits under Overview, Detailed Rules / Detailed Design, Formulas, Edge
Cases, Dependencies or Acceptance Criteria are compared — plus Tuning Knobs when
`project.yaml` sets `workflow_overrides.tuning_knobs: true` (read that flag
once; it only ever adds the section). Any other hit (Tuning Knobs without the
flag, Player Fantasy, any other section) raises no finding and is not logged;
list each one that states a value (name, GDD, section) on the report's
`Not checked at this tier` line.

### 3a: Entity Scan

For each entity in entity_map:

```
Grep pattern="[entity_name]" glob="design/gdd/*.md" output_mode="content" -C 

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.