feat: harden vision benchmark artifacts

Refs #817
2026-04-22 12:20:49 -04:00
parent 23e093fc75
commit 9d05f77a9b
4 changed files with 1112 additions and 252 deletions
--- a/benchmarks/vision_benchmark.py
+++ b/benchmarks/vision_benchmark.py
@@ -22,10 +22,12 @@ import argparse
 import asyncio
 import base64
 import json
+import mimetypes
 import os
 import statistics
 import sys
 import time
+import urllib.request
 from datetime import datetime, timezone
 from pathlib import Path
 from typing import Any, Dict, List, Optional
@@ -41,12 +43,16 @@ MODELS = {
        "model_id": "google/gemma-4-27b-it",
        "display_name": "Gemma 4 27B",
        "provider": "nous",
+        "fallback_provider": "ollama",
+        "fallback_model_id": "gemma4:latest",
        "description": "Google's multimodal Gemma 4 model",
    },
    "gemini3_flash": {
        "model_id": "google/gemini-3-flash-preview",
        "display_name": "Gemini 3 Flash Preview",
        "provider": "openrouter",
+        "fallback_provider": "gemini",
+        "fallback_model_id": "gemini-2.5-flash",
        "description": "Current default vision model",
    },
 }
@@ -84,91 +90,150 @@ async def analyze_with_model(
    """
    import httpx

+    def _load_image_bytes_cached() -> tuple[bytes, str]:
+        nonlocal _image_bytes, _mime_type
+        if _image_bytes is not None:
+            return _image_bytes, _mime_type
+        if image_url.startswith(("http://", "https://")):
+            with urllib.request.urlopen(image_url, timeout=30) as resp:
+                _image_bytes = resp.read()
+                _mime_type = resp.headers.get_content_type() or mimetypes.guess_type(image_url)[0] or "image/png"
+        else:
+            path = Path(image_url).expanduser()
+            _image_bytes = path.read_bytes()
+            _mime_type = mimetypes.guess_type(str(path))[0] or "image/png"
+        return _image_bytes, _mime_type
+
+    def _data_url() -> str:
+        image_bytes, mime_type = _load_image_bytes_cached()
+        return f"data:{mime_type};base64,{base64.b64encode(image_bytes).decode()}"
+
+    def _provider_key(provider: str) -> str:
+        if provider == "openrouter":
+            return os.getenv("OPENROUTER_API_KEY", "")
+        if provider == "nous":
+            return os.getenv("NOUS_API_KEY", "") or os.getenv("NOUS_INFERENCE_API_KEY", "")
+        if provider == "gemini":
+            return os.getenv("GEMINI_API_KEY", "") or os.getenv("GOOGLE_API_KEY", "")
+        return os.getenv(f"{provider.upper()}_API_KEY", "")
+
    provider = model_config["provider"]
    model_id = model_config["model_id"]
+    candidates = [(provider, model_id)]
+    if model_config.get("fallback_provider") and model_config.get("fallback_model_id"):
+        candidates.append((model_config["fallback_provider"], model_config["fallback_model_id"]))

-    # Prepare messages
-    messages = [
-        {
-            "role": "user",
-            "content": [
-                {"type": "text", "text": prompt},
-                {"type": "image_url", "image_url": {"url": image_url}},
-            ],
-        }
-    ]
+    _image_bytes: Optional[bytes] = None
+    _mime_type = "image/png"
+    failures = []

-    # Route to provider
-    if provider == "openrouter":
-        api_url = "https://openrouter.ai/api/v1/chat/completions"
-        api_key = os.getenv("OPENROUTER_API_KEY", "")
-    elif provider == "nous":
-        api_url = "https://inference.nousresearch.com/v1/chat/completions"
-        api_key = os.getenv("NOUS_API_KEY", "") or os.getenv("NOUS_INFERENCE_API_KEY", "")
-    else:
-        api_url = os.getenv(f"{provider.upper()}_API_URL", "")
-        api_key = os.getenv(f"{provider.upper()}_API_KEY", "")
+    for candidate_provider, candidate_model in candidates:
+        api_key = _provider_key(candidate_provider)
+        start = time.perf_counter()
+        try:
+            if candidate_provider in {"openrouter", "nous"}:
+                api_url = (
+                    "https://openrouter.ai/api/v1/chat/completions"
+                    if candidate_provider == "openrouter"
+                    else "https://inference.nousresearch.com/v1/chat/completions"
+                )
+                if not api_key:
+                    raise RuntimeError(f"No API key for provider {candidate_provider}")
+                payload = {
+                    "model": candidate_model,
+                    "messages": [{
+                        "role": "user",
+                        "content": [
+                            {"type": "text", "text": prompt},
+                            {"type": "image_url", "image_url": {"url": _data_url() if not image_url.startswith(("http://", "https://")) else image_url}},
+                        ],
+                    }],
+                    "max_tokens": 2000,
+                    "temperature": 0.1,
+                }
+                headers = {
+                    "Authorization": f"Bearer {api_key}",
+                    "Content-Type": "application/json",
+                }
+                async with httpx.AsyncClient(timeout=timeout) as client:
+                    resp = await client.post(api_url, json=payload, headers=headers)
+                    resp.raise_for_status()
+                    data = resp.json()
+                analysis = data.get("choices", [{}])[0].get("message", {}).get("content", "")
+                usage = data.get("usage", {})
+                tokens = {
+                    "prompt_tokens": usage.get("prompt_tokens", 0),
+                    "completion_tokens": usage.get("completion_tokens", 0),
+                    "total_tokens": usage.get("total_tokens", 0),
+                }
+            elif candidate_provider == "gemini":
+                if not api_key:
+                    raise RuntimeError("No API key for provider gemini")
+                image_bytes, mime_type = _load_image_bytes_cached()
+                api_url = f"https://generativelanguage.googleapis.com/v1beta/models/{candidate_model}:generateContent?key={api_key}"
+                payload = {
+                    "contents": [{"parts": [
+                        {"text": prompt},
+                        {"inline_data": {"mime_type": mime_type, "data": base64.b64encode(image_bytes).decode()}},
+                    ]}],
+                    "generationConfig": {"temperature": 0.1, "maxOutputTokens": 2000},
+                }
+                async with httpx.AsyncClient(timeout=timeout) as client:
+                    resp = await client.post(api_url, json=payload)
+                    resp.raise_for_status()
+                    data = resp.json()
+                parts = data.get("candidates", [{}])[0].get("content", {}).get("parts", [])
+                analysis = "\n".join(part.get("text", "") for part in parts if isinstance(part, dict) and part.get("text"))
+                usage = data.get("usageMetadata", {})
+                tokens = {
+                    "prompt_tokens": usage.get("promptTokenCount", 0),
+                    "completion_tokens": usage.get("candidatesTokenCount", 0),
+                    "total_tokens": usage.get("totalTokenCount", 0),
+                }
+            elif candidate_provider == "ollama":
+                image_bytes, _ = _load_image_bytes_cached()
+                payload = {
+                    "model": candidate_model,
+                    "stream": False,
+                    "messages": [{"role": "user", "content": prompt, "images": [base64.b64encode(image_bytes).decode()]}],
+                    "options": {"temperature": 0.1},
+                }
+                async with httpx.AsyncClient(timeout=timeout) as client:
+                    resp = await client.post("http://localhost:11434/api/chat", json=payload)
+                    resp.raise_for_status()
+                    data = resp.json()
+                analysis = data.get("message", {}).get("content", "")
+                tokens = {
+                    "prompt_tokens": data.get("prompt_eval_count", 0),
+                    "completion_tokens": data.get("eval_count", 0),
+                    "total_tokens": (data.get("prompt_eval_count", 0) or 0) + (data.get("eval_count", 0) or 0),
+                }
+            else:
+                raise RuntimeError(f"Unsupported provider {candidate_provider}")

-    if not api_key:
-        return {
-            "analysis": "",
-            "latency_ms": 0,
-            "tokens": {},
-            "success": False,
-            "error": f"No API key for provider {provider}",
-        }
+            latency_ms = (time.perf_counter() - start) * 1000
+            return {
+                "analysis": analysis,
+                "latency_ms": round(latency_ms, 1),
+                "tokens": tokens,
+                "success": True,
+                "error": "",
+                "provider_used": candidate_provider,
+                "model_used": candidate_model,
+            }
+        except Exception as e:
+            failures.append(f"{candidate_provider}:{candidate_model} => {e}")

-    headers = {
-        "Authorization": f"Bearer {api_key}",
-        "Content-Type": "application/json",
+    return {
+        "analysis": "",
+        "latency_ms": 0,
+        "tokens": {},
+        "success": False,
+        "error": " | ".join(failures) if failures else "No runs",
+        "provider_used": candidates[-1][0] if candidates else provider,
+        "model_used": candidates[-1][1] if candidates else model_id,
    }

-    payload = {
-        "model": model_id,
-        "messages": messages,
-        "max_tokens": 2000,
-        "temperature": 0.1,
-    }
-
-    start = time.perf_counter()
-    try:
-        async with httpx.AsyncClient(timeout=timeout) as client:
-            resp = await client.post(api_url, json=payload, headers=headers)
-            resp.raise_for_status()
-            data = resp.json()
-
-        latency_ms = (time.perf_counter() - start) * 1000
-
-        analysis = ""
-        choices = data.get("choices", [])
-        if choices:
-            msg = choices[0].get("message", {})
-            analysis = msg.get("content", "")
-
-        usage = data.get("usage", {})
-        tokens = {
-            "prompt_tokens": usage.get("prompt_tokens", 0),
-            "completion_tokens": usage.get("completion_tokens", 0),
-            "total_tokens": usage.get("total_tokens", 0),
-        }
-
-        return {
-            "analysis": analysis,
-            "latency_ms": round(latency_ms, 1),
-            "tokens": tokens,
-            "success": True,
-            "error": "",
-        }
-
-    except Exception as e:
-        return {
-            "analysis": "",
-            "latency_ms": round((time.perf_counter() - start) * 1000, 1),
-            "tokens": {},
-            "success": False,
-            "error": str(e),
-        }
-

 # ---------------------------------------------------------------------------
 # Evaluation metrics
@@ -398,7 +463,13 @@ def aggregate_results(results: List[dict], models: dict) -> dict:
        failed = [r[model_name] for r in results if not r[model_name]["success"]]

        if not model_results:
-            summary[model_name] = {"success_rate": 0, "error": "All runs failed"}
+            summary[model_name] = {
+                "success_rate": 0,
+                "error": "All runs failed",
+                "total_runs": 0,
+                "total_failures": len(failed),
+                "failure_examples": sorted({f.get("error", "unknown failure") for f in failed})[:3],
+            }
            continue

        latencies = [r["avg_latency_ms"] for r in model_results]
@@ -410,6 +481,7 @@ def aggregate_results(results: List[dict], models: dict) -> dict:
            "success_rate": round(len(model_results) / (len(model_results) + len(failed)), 4),
            "total_runs": len(model_results),
            "total_failures": len(failed),
+            "failure_examples": sorted({f.get("error", "unknown failure") for f in failed})[:3],
            "latency": {
                "mean_ms": round(statistics.mean(latencies), 1),
                "median_ms": round(statistics.median(latencies), 1),
@@ -495,6 +567,23 @@ def to_markdown(report: dict) -> str:
                f"| {mname} | {tok['mean_total']:.0f} | {tok['total_used']} |"
            )

+    lines += ["", "## Failure Modes", ""]
+    had_failures = False
+    for mkey, mname in config["models"].items():
+        model_summary = summary.get(mkey, {})
+        failure_examples = model_summary.get("failure_examples", [])
+        if not failure_examples and not model_summary.get("error"):
+            continue
+        had_failures = True
+        lines.append(f"### {mname}")
+        if model_summary.get("error"):
+            lines.append(f"- Summary: {model_summary['error']}")
+        for err in failure_examples:
+            lines.append(f"- {err}")
+        lines.append("")
+    if not had_failures:
+        lines.append("- No provider/runtime failures recorded.")
+
    # Verdict
    lines += ["", "## Verdict", ""]

@@ -516,8 +605,12 @@ def to_markdown(report: dict) -> str:

    if best_model:
        lines.append(f"**Best overall: {best_model}** (composite score: {best_score:.1%})")
+        lines.append("")
+        lines.append("Recommendation: keep the best-performing Gemma/Gemini lane from this run and only switch if repeated runs disagree.")
    else:
-        lines.append("No clear winner — insufficient data.")
+        lines.append("Benchmark blocked or insufficient data for a trustworthy winner.")
+        lines.append("")
+        lines.append("Recommendation: repair provider/runtime availability, rerun the benchmark, and keep the current implementation unchanged until comparative results exist.")

    return "\n".join(lines)

@@ -528,44 +621,124 @@ def to_markdown(report: dict) -> str:


 def generate_sample_dataset() -> List[dict]:
-    """Generate a sample test dataset with diverse public images.
+    """Generate a larger benchmark dataset aligned with issue #817.

-    Returns list of test image definitions.
+    Returns 50+ images across screenshots, diagrams, photos, OCR, charts,
+    and document-like images so the harness matches the issue contract.
    """
-    return [
-        # Screenshots
-        {
-            "id": "screenshot_github",
-            "url": "https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png",
+    dataset: List[dict] = []
+
+    screenshots = [
+        ("github_mark", "https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png", ["github", "logo", "mark"]),
+        ("github_social", "https://github.githubassets.com/images/modules/site/social-cards.png", ["github", "page", "web"]),
+        ("github_code_search", "https://github.githubassets.com/images/modules/site/features-code-search.png", ["search", "code", "feature"]),
+        ("terminal_capture", "https://raw.githubusercontent.com/nicehash/nicehash-quick-start/main/images/nicehash-terminal.png", ["terminal", "command", "output"]),
+        ("http_404", "https://http.cat/404.jpg", ["404", "error", "cat"]),
+        ("dummy_cli_01", "https://dummyimage.com/1280x720/111827/f9fafb.png&text=Hermes+CLI+Session+01", ["hermes", "cli", "session"]),
+        ("dummy_cli_02", "https://dummyimage.com/1280x720/0f172a/e2e8f0.png&text=Prompt+Cache+Dashboard", ["prompt", "cache", "dashboard"]),
+        ("dummy_ui_01", "https://dummyimage.com/1280x720/1f2937/f3f4f6.png&text=Settings+Panel+Voice+Mode", ["settings", "voice", "mode"]),
+        ("dummy_ui_02", "https://dummyimage.com/1280x720/334155/f8fafc.png&text=Browser+Vision+Preview", ["browser", "vision", "preview"]),
+        ("dummy_ui_03", "https://dummyimage.com/1280x720/111827/ffffff.png&text=Tool+Call+Inspector", ["tool", "call", "inspector"]),
+    ]
+    for ident, url, keywords in screenshots:
+        dataset.append({
+            "id": f"screenshot_{ident}",
+            "url": url,
            "category": "screenshot",
-            "expected_keywords": ["github", "logo", "octocat"],
-            "expected_structure": {"min_length": 50, "min_sentences": 2},
-        },
-        # Diagrams
-        {
-            "id": "diagram_architecture",
-            "url": "https://mermaid.ink/img/pako:eNp9kMtOwzAQRX_F8hKpJbhJFVJBi1QJiMWCG8eZNsGJLdlOiqIid5RdufiHnZRA7GbuzJwZe4ZGH2SCBPYUwgxoQKvJnCR2YY0F5YBdJJkD4uX0oXB6PnF3U4zCWcWdW3FqOwGvCKkBmHKSTB2gJeRrLTeJLfJdJKkBGYf9P1sTNdUXVJqY3YNJK7xLVwR0mxJFU6rCgEKnhSGIL2Eq8BdEERAX0OGwEiVQ1R0MaNFR8QfqKxmHigbX8VLjDz_Q0L8Wc_qPxDw",
+            "expected_keywords": keywords,
+            "ground_truth_ocr": "",
+            "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": False},
+        })
+
+    diagrams = [
+        ("flow_a", "https://dummyimage.com/1200x800/f8fafc/0f172a.png&text=Flowchart+API+Gateway+Queue+Worker", ["flowchart", "api", "worker"]),
+        ("flow_b", "https://dummyimage.com/1200x800/f1f5f9/0f172a.png&text=Architecture+Diagram+Database+Cache+Client", ["architecture", "diagram", "cache"]),
+        ("uml_a", "https://dummyimage.com/1200x800/e2e8f0/0f172a.png&text=Class+Diagram+User+Session+Message", ["class", "diagram", "session"]),
+        ("uml_b", "https://dummyimage.com/1200x800/cbd5e1/0f172a.png&text=Sequence+Diagram+Request+Response", ["sequence", "diagram", "response"]),
+        ("network_a", "https://dummyimage.com/1200x800/ffffff/111827.png&text=Network+Nodes+Edges+Router", ["network", "node", "router"]),
+        ("network_b", "https://dummyimage.com/1200x800/ffffff/1e293b.png&text=Service+Mesh+Proxy+Auth", ["service", "mesh", "auth"]),
+        ("state_machine", "https://dummyimage.com/1200x800/f8fafc/334155.png&text=State+Machine+Idle+Run+Stop", ["state", "machine", "idle"]),
+        ("mind_map", "https://dummyimage.com/1200x800/fefce8/1f2937.png&text=Mind+Map+Memory+Recall+Tools", ["mind", "memory", "tools"]),
+        ("pipeline", "https://dummyimage.com/1200x800/ecfeff/155e75.png&text=Pipeline+Ingest+Rank+Summarize", ["pipeline", "ingest", "summarize"]),
+        ("org_chart", "https://dummyimage.com/1200x800/fdf2f8/831843.png&text=Org+Chart+Lead+Review+Ops", ["org", "chart", "review"]),
+    ]
+    for ident, url, keywords in diagrams:
+        dataset.append({
+            "id": f"diagram_{ident}",
+            "url": url,
            "category": "diagram",
-            "expected_keywords": ["architecture", "component", "service"],
-            "expected_structure": {"min_length": 100, "min_sentences": 3},
-        },
-        # Photos
-        {
-            "id": "photo_nature",
-            "url": "https://picsum.photos/seed/bench1/400/300",
+            "expected_keywords": keywords,
+            "ground_truth_ocr": "",
+            "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": False},
+        })
+
+    for idx in range(1, 11):
+        dataset.append({
+            "id": f"photo_random_{idx:02d}",
+            "url": f"https://picsum.photos/seed/vision-bench-{idx}/640/480",
            "category": "photo",
            "expected_keywords": [],
-            "expected_structure": {"min_length": 30, "min_sentences": 1},
-        },
-        # Charts
-        {
-            "id": "chart_bar",
-            "url": "https://quickchart.io/chart?c={type:'bar',data:{labels:['Q1','Q2','Q3','Q4'],datasets:[{label:'Users',data:[50,60,70,80]}]}}",
-            "category": "chart",
-            "expected_keywords": ["bar", "chart", "data"],
-            "expected_structure": {"min_length": 50, "min_sentences": 2},
-        },
+            "ground_truth_ocr": "",
+            "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": False},
+        })
+
+    charts = [
+        ("bar_quarterly", "https://quickchart.io/chart?c={type:'bar',data:{labels:['Q1','Q2','Q3','Q4'],datasets:[{label:'Revenue',data:[100,150,200,250]}]}}", ["bar", "chart", "revenue"]),
+        ("pie_market", "https://quickchart.io/chart?c={type:'pie',data:{labels:['A','B','C'],datasets:[{data:[30,50,20]}]}}", ["pie", "chart", "percentage"]),
+        ("line_temp", "https://quickchart.io/chart?c={type:'line',data:{labels:['Jan','Feb','Mar','Apr'],datasets:[{label:'Temperature',data:[5,8,12,18]}]}}", ["line", "chart", "temperature"]),
+        ("radar_skill", "https://quickchart.io/chart?c={type:'radar',data:{labels:['Speed','Power','Defense','Magic'],datasets:[{label:'Hero',data:[80,60,70,90]}]}}", ["radar", "chart", "skill"]),
+        ("stacked_cloud", "https://quickchart.io/chart?c={type:'bar',data:{labels:['2022','2023','2024'],datasets:[{label:'Cloud',data:[100,150,200]},{label:'On-prem',data:[200,180,160]}]},options:{scales:{x:{stacked:true},y:{stacked:true}}}}", ["stacked", "bar", "chart"]),
+        ("area_growth", "https://quickchart.io/chart?c={type:'line',data:{labels:['W1','W2','W3','W4'],datasets:[{label:'Growth',data:[10,15,18,24],fill:true}]}}", ["line", "growth", "chart"]),
+        ("scatter_eval", "https://quickchart.io/chart?c={type:'scatter',data:{datasets:[{label:'Runs',data:[{x:1,y:70},{x:2,y:75},{x:3,y:82}]}]}}", ["scatter", "chart", "runs"]),
+        ("horizontal_bar", "https://quickchart.io/chart?c={type:'bar',data:{labels:['UI','OCR','Docs'],datasets:[{label:'Score',data:[88,76,91]}]},options:{indexAxis:'y'}}", ["bar", "score", "ocr"]),
+        ("bubble_usage", "https://quickchart.io/chart?c={type:'bubble',data:{datasets:[{label:'Latency',data:[{x:1,y:120,r:8},{x:2,y:95,r:6},{x:3,y:180,r:10}]}]}}", ["bubble", "latency", "chart"]),
+        ("doughnut_devices", "https://quickchart.io/chart?c={type:'doughnut',data:{labels:['Desktop','Mobile','Tablet'],datasets:[{data:[60,30,10]}]}}", ["doughnut", "chart", "device"]),
    ]
+    for ident, url, keywords in charts:
+        dataset.append({
+            "id": f"chart_{ident}",
+            "url": url,
+            "category": "chart",
+            "expected_keywords": keywords,
+            "ground_truth_ocr": "",
+            "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": True},
+        })
+
+    ocr_texts = [
+        "Hermes OCR Alpha 01",
+        "Prompt Cache Hit 87%",
+        "Session 42 Ready",
+        "Latency 118 ms",
+        "Voice Mode Enabled",
+    ]
+    for idx, text in enumerate(ocr_texts, start=1):
+        dataset.append({
+            "id": f"ocr_text_{idx:02d}",
+            "url": f"https://dummyimage.com/1200x320/ffffff/000000.png&text={text.replace(' ', '+')}",
+            "category": "ocr",
+            "expected_keywords": text.lower().split()[:2],
+            "ground_truth_ocr": text,
+            "expected_structure": {"min_length": 10, "min_sentences": 1, "has_numbers": any(ch.isdigit() for ch in text)},
+        })
+
+    documents = [
+        "Invoice 1001 Total 42 Due 2026-04-22",
+        "Form A Name Alice Status Approved",
+        "Report Memory Recall Score 91 Percent",
+        "Checklist Crisis Escalation Call 988 Now",
+        "Meeting Notes Vision Benchmark Run Pending",
+    ]
+    for idx, text in enumerate(documents, start=1):
+        dataset.append({
+            "id": f"document_text_{idx:02d}",
+            "url": f"https://dummyimage.com/1400x900/f8fafc/0f172a.png&text={text.replace(' ', '+')}",
+            "category": "document",
+            "expected_keywords": text.lower().split()[:3],
+            "ground_truth_ocr": text,
+            "expected_structure": {"min_length": 20, "min_sentences": 1, "has_numbers": any(ch.isdigit() for ch in text)},
+        })
+
+    return dataset


 def load_dataset(path: str) -> List[dict]:
@@ -585,7 +758,9 @@ async def main():
    parser.add_argument("--url", help="Single image URL to test")
    parser.add_argument("--category", default="photo", help="Category for single URL")
    parser.add_argument("--output", default=None, help="Output JSON file")
+    parser.add_argument("--markdown-output", default=None, help="Optional markdown report output path")
    parser.add_argument("--runs", type=int, default=1, help="Runs per model per image")
+    parser.add_argument("--limit", type=int, default=0, help="Limit to the first N images for smoke runs")
    parser.add_argument("--models", nargs="+", default=None,
                        help="Models to test (default: all)")
    parser.add_argument("--markdown", action="store_true", help="Output markdown report")
@@ -617,9 +792,14 @@ async def main():
        print("ERROR: Provide --images or --url")
        sys.exit(1)

+    if args.limit and args.limit > 0:
+        images = images[:args.limit]
+
    # Run benchmark
    report = await run_benchmark_suite(images, selected, args.runs)

+    markdown_report = to_markdown(report)
+
    # Output
    if args.output:
        os.makedirs(os.path.dirname(args.output) or ".", exist_ok=True)
@@ -627,8 +807,14 @@ async def main():
            json.dump(report, f, indent=2)
        print(f"\nResults saved to {args.output}")

+    if args.markdown_output:
+        os.makedirs(os.path.dirname(args.markdown_output) or ".", exist_ok=True)
+        with open(args.markdown_output, "w", encoding="utf-8") as f:
+            f.write(markdown_report)
+        print(f"Markdown report saved to {args.markdown_output}")
+
    if args.markdown or not args.output:
-        print("\n" + to_markdown(report))
+        print("\n" + markdown_report)


 if __name__ == "__main__":