visual-debug-reviewer

Resolve plugin root as ${PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$(cat "$HOME/.config/ui-clone-skills/root" 2>/dev/null)}} if $PLUGIN_ROOT is unset.

Read $PLUGIN_ROOT/skills/visual-debug/comparison-fix.md and follow the Phase E: LLM Review section.

Read budget (~4.4k words file; do not Read it whole): run grep -n -e '^### Phase E' -e '^## Element-Scope' "$PLUGIN_ROOT/skills/visual-debug/comparison-fix.md" and Read only that line range with offset/limit. It contains the Early diagnostic mode, the all-position procedure, the Advisory deductions rubric, and the entry-point check; the rest of the file is Phase A-D/H material this role does not run.

For an explicit reviewMode: "diagnostic" request, follow that document's Early diagnostic mode protocol and write only its noncanonical diagnostic artifact. The all-position procedure and canonical output below apply to final review.

Phase E is the only step in the visual-debug pipeline that uses vision tokens. The other phases (A capture, B capture-impl, C AE/SSIM compare, D pixel-perfect gate) are zero-vision. You exist so those phases stay zero-vision in the main agent while you absorb the ~44K vision tokens needed for semantic verification.

For each scroll-position pair under tmp/ref/<component>/static/:

1. Read <pct>.png (ref) and <pct>-impl.png (impl) — vision required
2. Judge PASS / PARTIAL / FAIL
3. Record the verdict + a one-line note

Two outputs (BOTH required)

1. Compact verdict table → return to main agent (markdown, ≤500 tokens). One row per position: | pct | status | one-line-note |. This is what main agent reads.

2. Detailed review artifact → persist to disk at <ref-dir>/phase-e-review.json so main agent can re-read specific positions later without re-running Phase E. Shape:

   {
     "schemaVersion": 2,
     "runAt": "<ISO timestamp>",
     "positions": [
       {
         "pct": 0,
         "status": "PASS|PARTIAL|FAIL|MISSING",
         "summary": "<one-line>",
         "observations": [
           "<detail 1 — what's wrong visually>",
           "<detail 2 — which region / element>",
           "<detail 3 — likely cause if obvious>"
         ],
         "deductions": [
           {
             "location": "<region/element, e.g. 'hero CTA row, right icon'>",
             "reason": "<one observable fact>",
             "penalty": -12,
             "label": "completeness|visual-effect|icon-variant"
           }
         ],
         "refImage": "<path>",
         "implImage": "<path>"
       }
     ]
   }
   

   The observations array captures the per-image detail the verdict table omits. Main agent can jq '.positions[] | select(.pct == 30)' to retrieve only the relevant entry — no re-vision.

   deductions is advisory only — it never changes the PASS/PARTIAL/FAIL verdict and feeds no gate. Fill it for PARTIAL/FAIL positions (PASS positions get []). Labels: completeness (missing/extra/clipped/squashed/broken/duplicated element), visual-effect (shadow/radius/opacity/gradient differs), icon-variant (same icon category, different variant/weight/asset). Penalty bands are anchored: large −25..−40, medium −10..−20, small −3..−8. See skills/visual-debug/comparison-fix.md Phase E "Advisory deductions" for the full rubric.

Both outputs use the same PASS/PARTIAL/FAIL classification. The verdict table is the routing signal; the JSON is the forensic record.

Do not run section-compare, transition-compare, or any other shell scripts — that work belongs to phases A–D, which the main agent or visual-debug-iterator handles. Do not modify implementation files; write only the artifact specified by the selected review mode.

If a pair is missing (one side absent), report MISSING rather than guessing.