adversarial-drawing-reviewer

name	adversarial-drawing-reviewer
description	Launches an independent adversarial drawing-review sub-agent to evaluate SVG outputs for builder usefulness before success can be claimed.
version	1.3.2
allowed-tools	Read, Write, Bash, Vision
tags	["validation","svg","drawings","red-team","quality-gate"]
metadata	{"garden_structure_designer":{"category":"validation","consumes":["outputs/*.svg","context/staging/structure.json"],"produces":["context/staging/drawing-red-team-report.json","outputs/drawing-red-team-report.md"],"validators":["plugins/garden-structure-designer/scripts/hardcode_guard.py","plugins/garden-structure-designer/scripts/visual_svg_smoke_test.py","plugins/garden-structure-designer/scripts/svg_validator.py","plugins/garden-structure-designer/scripts/drawing_content_validator.py"],"repair_axes":["svg_placeholder_geometry","svg_missing_dimensions","svg_missing_component_details","svg_not_builder_meaningful","visual_drift_or_regression","static_hardcode_violation"]}}

Adversarial Drawing Reviewer Skill

This skill exists to prevent green-check theater.

The drawing generator cannot certify its own success.

This skill runs a strict multi-layered gate (Static Scan, Browser Headless rendering, Computer Vision topology check, and Perceptual Difference audit) of every generated drawing before the package can be marked PASS.

Required Behavior

1. Static Invariant Guard: Scan the codebase using hardcode_guard.py to ensure no literal scales or silent fallbacks exist.

2. PNG co-generation (token-efficient): render_drawings.py produces outputs/visual-smoke/*.png inline alongside each SVG using qlmanage (macOS) or Playwright fallback — no separate browser session required. Inspect the pre-generated PNGs directly with Read/Vision tools:

   • outputs/visual-smoke/drawing-isometric-view.png
   • outputs/visual-smoke/drawing-perspective-view.png
   • outputs/visual-smoke/blueprint-*.png

   Only re-run visual_svg_smoke_test.py if PNGs are missing or stale relative to the SVGs. Playwright is a last-resort fallback, not the primary rendering path.

3. Computer Vision Topology Audit: Execute OpenCV-driven feature line detections on the PNGs using Canny edge filtering and Hough line transforms to verify posts, beams, rafters, and hub densities against the expected model configuration.

4. Perceptual Image Difference: Compute structural pHash Perceptual Hashing and Mean Absolute Difference (MAD) against baselines/reference images to detect visual drift.

5. Multimodal Visual Inspection (Carpenter's Eye Test): Open and visually inspect the PNG renders (specifically outputs/visual-smoke/drawing-isometric-view.png and drawing-perspective-view.png) alongside the reference baseline image outputs/high-resolution-image/hexagon_pergola_render.png using vision capabilities. Systematically verify every visible joint and member against this checklist:

   • Hub Apex Connections:
     • Rafter ends must terminate flat and flush against the hub vertical faces.
     • Rafters must not visually slice through one another or overlap in the apex region.
     • The hub block must be subtle and integrated, not oversized or projecting as a chimney block above the rafter tops.
     • There must be absolutely no floating, disconnected primitive blocks hovering in space near the hub or rafters.
   • Post-to-Beam Seats:
     • Beam support ledgers must sit squarely and flush on post tops/notches with zero visual gaps or floating elements.
   • Knee Braces:

Mandatory Failure Rule

If drawings are blank, tiny, childish, symbolic, or have visual errors (such as rafters colliding or the central hub sticking up above the roof line), this skill must return FAIL or PARTIAL.

Do not allow PASS merely because XML validation passes.

Authoritative Commands

1. Static & Visual Smoke Gate Command

Runs the AST scan guard, Playwright browser rendering, OpenCV topology, and pHash/MAD baseline verification:

# Verify no static scale hardcodes exist
python3 plugins/garden-structure-designer/scripts/hardcode_guard.py

# Render and audit browser sheets against expected topology
python3 plugins/garden-structure-designer/scripts/visual_svg_smoke_test.py \
  --structure context/staging/structure.json \
  --svg-dir outputs \
  --out-dir outputs/visual-smoke \
  --report-json context/staging/visual-smoke-report.json \
  --report-md outputs/visual-smoke-report.md \
  --fail-on-regression

2. Traditional Red-Team Content Command

Enumerates all SVGs, runs the content and scale checks, and builds the composite red-team summary:

python3 plugins/garden-structure-designer/scripts/run_drawing_red_team.py \
  --svg-dir outputs \
  --model context/staging/structure.json \
  --report-dir context/staging \
  --md-dir outputs

Obstacle Escalation

When the red-team gate fails, the reviewer must distinguish between:

• drawings are bad;
• reports are stale;
• paths are inconsistent;
• required artifacts are missing;
• validators are misconfigured.

The reviewer must not only say "FAIL." The reviewer must identify the likely class of failure and recommend the next control-plane action.

Required output fields in drawing-red-team-report.json:

{
  "failure_classification": [
    "RENDERER_PLACEHOLDER_OUTPUT",
    "PATH_SPLIT",
    "STALE_REPORT"
  ],
  "recommended_next_action": "FIX_PATHS_FIRST|RERUN_REPORTS_FIRST|FIX_RENDERER_FIRST|BLOCK_AND_REPORT"
}

This writes:

• context/staging/drawing-content-report.json — per-sheet content detail
• context/staging/drawing-red-team-report.json — gate verdict + may_claim_success
• outputs/drawing-red-team-report.md — human-readable summary

Diagnostic Commands (optional supplemental detail)

Individual validator runs for debugging a specific sheet:

python3 plugins/garden-structure-designer/scripts/svg_validator.py \
  outputs/<sheet>.svg \
  context/staging/structure.json

python3 plugins/garden-structure-designer/scripts/drawing_content_validator.py \
  outputs/<sheet>.svg \
  context/staging/structure.json \
  --json-output context/staging/drawing-content-report.json \
  --append

Then perform adversarial review using the drawing-red-team-agent rubric.

Final Gate

The orchestrator must read:

context/staging/drawing-red-team-report.json

If:

"may_claim_success": false

then the package status must not be PASS.

Allowed statuses are:

PARTIAL
BLOCKED
DRAFT ONLY

The job is not done until this skill approves the drawings.