fix: vendor vision benchmark fixtures (#868)

agent/account_usage.py

@@ -1,326 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from datetime import datetime, timezone
-from typing import Any, Optional
-
-import httpx
-
-from agent.anthropic_adapter import _is_oauth_token, resolve_anthropic_token
-from hermes_cli.auth import _read_codex_tokens, resolve_codex_runtime_credentials
-from hermes_cli.runtime_provider import resolve_runtime_provider
-
-
-def _utc_now() -> datetime:
-    return datetime.now(timezone.utc)
-
-
-@dataclass(frozen=True)
-class AccountUsageWindow:
-    label: str
-    used_percent: Optional[float] = None
-    reset_at: Optional[datetime] = None
-    detail: Optional[str] = None
-
-
-@dataclass(frozen=True)
-class AccountUsageSnapshot:
-    provider: str
-    source: str
-    fetched_at: datetime
-    title: str = "Account limits"
-    plan: Optional[str] = None
-    windows: tuple[AccountUsageWindow, ...] = ()
-    details: tuple[str, ...] = ()
-    unavailable_reason: Optional[str] = None
-
-    @property
-    def available(self) -> bool:
-        return bool(self.windows or self.details) and not self.unavailable_reason
-
-
-def _title_case_slug(value: Optional[str]) -> Optional[str]:
-    cleaned = str(value or "").strip()
-    if not cleaned:
-        return None
-    return cleaned.replace("_", " ").replace("-", " ").title()
-
-
-def _parse_dt(value: Any) -> Optional[datetime]:
-    if value in (None, ""):
-        return None
-    if isinstance(value, (int, float)):
-        return datetime.fromtimestamp(float(value), tz=timezone.utc)
-    if isinstance(value, str):
-        text = value.strip()
-        if not text:
-            return None
-        if text.endswith("Z"):
-            text = text[:-1] + "+00:00"
-        try:
-            dt = datetime.fromisoformat(text)
-            return dt if dt.tzinfo else dt.replace(tzinfo=timezone.utc)
-        except ValueError:
-            return None
-    return None
-
-
-def _format_reset(dt: Optional[datetime]) -> str:
-    if not dt:
-        return "unknown"
-    local_dt = dt.astimezone()
-    delta = dt - _utc_now()
-    total_seconds = int(delta.total_seconds())
-    if total_seconds <= 0:
-        return f"now ({local_dt.strftime('%Y-%m-%d %H:%M %Z')})"
-    hours, rem = divmod(total_seconds, 3600)
-    minutes = rem // 60
-    if hours >= 24:
-        days, hours = divmod(hours, 24)
-        rel = f"in {days}d {hours}h"
-    elif hours > 0:
-        rel = f"in {hours}h {minutes}m"
-    else:
-        rel = f"in {minutes}m"
-    return f"{rel} ({local_dt.strftime('%Y-%m-%d %H:%M %Z')})"
-
-
-def render_account_usage_lines(snapshot: Optional[AccountUsageSnapshot], *, markdown: bool = False) -> list[str]:
-    if not snapshot:
-        return []
-    header = f"📈 {'**' if markdown else ''}{snapshot.title}{'**' if markdown else ''}"
-    lines = [header]
-    if snapshot.plan:
-        lines.append(f"Provider: {snapshot.provider} ({snapshot.plan})")
-    else:
-        lines.append(f"Provider: {snapshot.provider}")
-    for window in snapshot.windows:
-        if window.used_percent is None:
-            base = f"{window.label}: unavailable"
-        else:
-            remaining = max(0, round(100 - float(window.used_percent)))
-            used = max(0, round(float(window.used_percent)))
-            base = f"{window.label}: {remaining}% remaining ({used}% used)"
-        if window.reset_at:
-            base += f" • resets {_format_reset(window.reset_at)}"
-        elif window.detail:
-            base += f" • {window.detail}"
-        lines.append(base)
-    for detail in snapshot.details:
-        lines.append(detail)
-    if snapshot.unavailable_reason:
-        lines.append(f"Unavailable: {snapshot.unavailable_reason}")
-    return lines
-
-
-def _resolve_codex_usage_url(base_url: str) -> str:
-    normalized = (base_url or "").strip().rstrip("/")
-    if not normalized:
-        normalized = "https://chatgpt.com/backend-api/codex"
-    if normalized.endswith("/codex"):
-        normalized = normalized[: -len("/codex")]
-    if "/backend-api" in normalized:
-        return normalized + "/wham/usage"
-    return normalized + "/api/codex/usage"
-
-
-def _fetch_codex_account_usage() -> Optional[AccountUsageSnapshot]:
-    creds = resolve_codex_runtime_credentials(refresh_if_expiring=True)
-    token_data = _read_codex_tokens()
-    tokens = token_data.get("tokens") or {}
-    account_id = str(tokens.get("account_id", "") or "").strip() or None
-    headers = {
-        "Authorization": f"Bearer {creds['api_key']}",
-        "Accept": "application/json",
-        "User-Agent": "codex-cli",
-    }
-    if account_id:
-        headers["ChatGPT-Account-Id"] = account_id
-    with httpx.Client(timeout=15.0) as client:
-        response = client.get(_resolve_codex_usage_url(creds.get("base_url", "")), headers=headers)
-        response.raise_for_status()
-    payload = response.json() or {}
-    rate_limit = payload.get("rate_limit") or {}
-    windows: list[AccountUsageWindow] = []
-    for key, label in (("primary_window", "Session"), ("secondary_window", "Weekly")):
-        window = rate_limit.get(key) or {}
-        used = window.get("used_percent")
-        if used is None:
-            continue
-        windows.append(
-            AccountUsageWindow(
-                label=label,
-                used_percent=float(used),
-                reset_at=_parse_dt(window.get("reset_at")),
-            )
-        )
-    details: list[str] = []
-    credits = payload.get("credits") or {}
-    if credits.get("has_credits"):
-        balance = credits.get("balance")
-        if isinstance(balance, (int, float)):
-            details.append(f"Credits balance: ${float(balance):.2f}")
-        elif credits.get("unlimited"):
-            details.append("Credits balance: unlimited")
-    return AccountUsageSnapshot(
-        provider="openai-codex",
-        source="usage_api",
-        fetched_at=_utc_now(),
-        plan=_title_case_slug(payload.get("plan_type")),
-        windows=tuple(windows),
-        details=tuple(details),
-    )
-
-
-def _fetch_anthropic_account_usage() -> Optional[AccountUsageSnapshot]:
-    token = (resolve_anthropic_token() or "").strip()
-    if not token:
-        return None
-    if not _is_oauth_token(token):
-        return AccountUsageSnapshot(
-            provider="anthropic",
-            source="oauth_usage_api",
-            fetched_at=_utc_now(),
-            unavailable_reason="Anthropic account limits are only available for OAuth-backed Claude accounts.",
-        )
-    headers = {
-        "Authorization": f"Bearer {token}",
-        "Accept": "application/json",
-        "Content-Type": "application/json",
-        "anthropic-beta": "oauth-2025-04-20",
-        "User-Agent": "claude-code/2.1.0",
-    }
-    with httpx.Client(timeout=15.0) as client:
-        response = client.get("https://api.anthropic.com/api/oauth/usage", headers=headers)
-        response.raise_for_status()
-    payload = response.json() or {}
-    windows: list[AccountUsageWindow] = []
-    mapping = (
-        ("five_hour", "Current session"),
-        ("seven_day", "Current week"),
-        ("seven_day_opus", "Opus week"),
-        ("seven_day_sonnet", "Sonnet week"),
-    )
-    for key, label in mapping:
-        window = payload.get(key) or {}
-        util = window.get("utilization")
-        if util is None:
-            continue
-        used = float(util) * 100 if float(util) <= 1 else float(util)
-        windows.append(
-            AccountUsageWindow(
-                label=label,
-                used_percent=used,
-                reset_at=_parse_dt(window.get("resets_at")),
-            )
-        )
-    details: list[str] = []
-    extra = payload.get("extra_usage") or {}
-    if extra.get("is_enabled"):
-        used_credits = extra.get("used_credits")
-        monthly_limit = extra.get("monthly_limit")
-        currency = extra.get("currency") or "USD"
-        if isinstance(used_credits, (int, float)) and isinstance(monthly_limit, (int, float)):
-            details.append(
-                f"Extra usage: {used_credits:.2f} / {monthly_limit:.2f} {currency}"
-            )
-    return AccountUsageSnapshot(
-        provider="anthropic",
-        source="oauth_usage_api",
-        fetched_at=_utc_now(),
-        windows=tuple(windows),
-        details=tuple(details),
-    )
-
-
-def _fetch_openrouter_account_usage(base_url: Optional[str], api_key: Optional[str]) -> Optional[AccountUsageSnapshot]:
-    runtime = resolve_runtime_provider(
-        requested="openrouter",
-        explicit_base_url=base_url,
-        explicit_api_key=api_key,
-    )
-    token = str(runtime.get("api_key", "") or "").strip()
-    if not token:
-        return None
-    normalized = str(runtime.get("base_url", "") or "").rstrip("/")
-    credits_url = f"{normalized}/credits"
-    key_url = f"{normalized}/key"
-    headers = {
-        "Authorization": f"Bearer {token}",
-        "Accept": "application/json",
-    }
-    with httpx.Client(timeout=10.0) as client:
-        credits_resp = client.get(credits_url, headers=headers)
-        credits_resp.raise_for_status()
-        credits = (credits_resp.json() or {}).get("data") or {}
-        try:
-            key_resp = client.get(key_url, headers=headers)
-            key_resp.raise_for_status()
-            key_data = (key_resp.json() or {}).get("data") or {}
-        except Exception:
-            key_data = {}
-    total_credits = float(credits.get("total_credits") or 0.0)
-    total_usage = float(credits.get("total_usage") or 0.0)
-    details = [f"Credits balance: ${max(0.0, total_credits - total_usage):.2f}"]
-    windows: list[AccountUsageWindow] = []
-    limit = key_data.get("limit")
-    limit_remaining = key_data.get("limit_remaining")
-    limit_reset = str(key_data.get("limit_reset") or "").strip()
-    usage = key_data.get("usage")
-    if (
-        isinstance(limit, (int, float))
-        and float(limit) > 0
-        and isinstance(limit_remaining, (int, float))
-        and 0 <= float(limit_remaining) <= float(limit)
-    ):
-        limit_value = float(limit)
-        remaining_value = float(limit_remaining)
-        used_percent = ((limit_value - remaining_value) / limit_value) * 100
-        detail_parts = [f"${remaining_value:.2f} of ${limit_value:.2f} remaining"]
-        if limit_reset:
-            detail_parts.append(f"resets {limit_reset}")
-        windows.append(
-            AccountUsageWindow(
-                label="API key quota",
-                used_percent=used_percent,
-                detail=" • ".join(detail_parts),
-            )
-        )
-    if isinstance(usage, (int, float)):
-        usage_parts = [f"API key usage: ${float(usage):.2f} total"]
-        for value, label in (
-            (key_data.get("usage_daily"), "today"),
-            (key_data.get("usage_weekly"), "this week"),
-            (key_data.get("usage_monthly"), "this month"),
-        ):
-            if isinstance(value, (int, float)) and float(value) > 0:
-                usage_parts.append(f"${float(value):.2f} {label}")
-        details.append(" • ".join(usage_parts))
-    return AccountUsageSnapshot(
-        provider="openrouter",
-        source="credits_api",
-        fetched_at=_utc_now(),
-        windows=tuple(windows),
-        details=tuple(details),
-    )
-
-
-def fetch_account_usage(
-    provider: Optional[str],
-    *,
-    base_url: Optional[str] = None,
-    api_key: Optional[str] = None,
-) -> Optional[AccountUsageSnapshot]:
-    normalized = str(provider or "").strip().lower()
-    if normalized in {"", "auto", "custom"}:
-        return None
-    try:
-        if normalized == "openai-codex":
-            return _fetch_codex_account_usage()
-        if normalized == "anthropic":
-            return _fetch_anthropic_account_usage()
-        if normalized == "openrouter":
-            return _fetch_openrouter_account_usage(base_url, api_key)
-    except Exception:
-        return None
-    return None

benchmarks/test_images.json

@@ -1,194 +1,354 @@
 [
   {
     "id": "screenshot_github_home",
-    "url": "https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png",
+    "url": "test_images/screenshot_github_home.png",
     "category": "screenshot",
-    "expected_keywords": ["github", "logo", "mark"],
+    "expected_keywords": [
+      "github",
+      "logo",
+      "mark"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "diagram_mermaid_flow",
-    "url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6siSZXVhjQTlgl1nigHg5fRBOzSfebopROCu_cytObSfgLSE1ANOeZWkO2IH5upZxYot8m1hqAdpD_63WRl0xdUG1jdl9kPiOb_EWk2JBtPaiKkF4eVIYgO0EtkW-RSgC4gJ6HJYRG1UNdN0HNVd0Bftjj7X8P92qPj-F8l8T3w",
+    "url": "test_images/diagram_mermaid_flow.png",
     "category": "diagram",
-    "expected_keywords": ["flow", "diagram", "process"],
+    "expected_keywords": [
+      "flow",
+      "diagram",
+      "process"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
   },
   {
     "id": "photo_random_1",
-    "url": "https://picsum.photos/seed/vision1/400/300",
+    "url": "test_images/photo_random_1.png",
     "category": "photo",
     "expected_keywords": [],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "photo_random_2",
-    "url": "https://picsum.photos/seed/vision2/400/300",
+    "url": "test_images/photo_random_2.png",
     "category": "photo",
     "expected_keywords": [],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "chart_simple_bar",
-    "url": "https://quickchart.io/chart?c={type:'bar',data:{labels:['Q1','Q2','Q3','Q4'],datasets:[{label:'Revenue',data:[100,150,200,250]}]}}",
+    "url": "test_images/chart_simple_bar.png",
     "category": "chart",
-    "expected_keywords": ["bar", "chart", "revenue"],
+    "expected_keywords": [
+      "bar",
+      "chart",
+      "revenue"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
   },
   {
     "id": "chart_pie",
-    "url": "https://quickchart.io/chart?c={type:'pie',data:{labels:['A','B','C'],datasets:[{data:[30,50,20]}]}}",
+    "url": "test_images/chart_pie.png",
     "category": "chart",
-    "expected_keywords": ["pie", "chart", "percentage"],
+    "expected_keywords": [
+      "pie",
+      "chart",
+      "percentage"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
   },
   {
     "id": "diagram_org_chart",
-    "url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6iuyIWyrLgXLALrPEAfFy-iCcmk-83RSjcFZ-51ac2k7AW0JqAKY9y9IcsAPzdS3jxBb5NrHUAraH_lutjbpi6oJqG7P7IPEd3-ItJsWCaO1FVYLw8qQwANsJbIt8i1AExAX0OCwjNqoa6LoPaq7oCvbHHmv5f7pVfX4K5b8mvg",
+    "url": "test_images/diagram_org_chart.png",
     "category": "diagram",
-    "expected_keywords": ["organization", "hierarchy", "chart"],
+    "expected_keywords": [
+      "organization",
+      "hierarchy",
+      "chart"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
   },
   {
     "id": "screenshot_terminal",
-    "url": "https://raw.githubusercontent.com/nicehash/nicehash-quick-start/main/images/nicehash-terminal.png",
+    "url": "test_images/screenshot_terminal.png",
     "category": "screenshot",
-    "expected_keywords": ["terminal", "command", "output"],
+    "expected_keywords": [
+      "terminal",
+      "command",
+      "output"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "photo_random_3",
-    "url": "https://picsum.photos/seed/vision3/400/300",
+    "url": "test_images/photo_random_3.png",
     "category": "photo",
     "expected_keywords": [],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "chart_line",
-    "url": "https://quickchart.io/chart?c={type:'line',data:{labels:['Jan','Feb','Mar','Apr'],datasets:[{label:'Temperature',data:[5,8,12,18]}]}}",
+    "url": "test_images/chart_line.png",
     "category": "chart",
-    "expected_keywords": ["line", "chart", "temperature"],
+    "expected_keywords": [
+      "line",
+      "chart",
+      "temperature"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
   },
   {
     "id": "diagram_sequence",
-    "url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6iuyIWyrLgXLALrPEAfFy-iCcmk-83RSjcFZ-51ac2k7AW0JqAKY9y9IcsAPzdS3jxBb5NrHUAraH_lutjbpi6oJqG7P7IPEd3-ItJsWCaO1FVYLw8qQwANsJbIt8i1AExAX0OCwjNqoa6LoPaq7oCvbHHmv5f7pVfX4K5b8mvg",
+    "url": "test_images/diagram_sequence.png",
     "category": "diagram",
-    "expected_keywords": ["sequence", "interaction", "message"],
+    "expected_keywords": [
+      "sequence",
+      "interaction",
+      "message"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
   },
   {
     "id": "photo_random_4",
-    "url": "https://picsum.photos/seed/vision4/400/300",
+    "url": "test_images/photo_random_4.png",
     "category": "photo",
     "expected_keywords": [],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "screenshot_webpage",
-    "url": "https://github.githubassets.com/images/modules/site/social-cards.png",
+    "url": "test_images/screenshot_webpage.png",
     "category": "screenshot",
-    "expected_keywords": ["github", "page", "web"],
+    "expected_keywords": [
+      "github",
+      "page",
+      "web"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "chart_radar",
-    "url": "https://quickchart.io/chart?c={type:'radar',data:{labels:['Speed','Power','Defense','Magic'],datasets:[{label:'Hero',data:[80,60,70,90]}]}}",
+    "url": "test_images/chart_radar.png",
     "category": "chart",
-    "expected_keywords": ["radar", "chart", "skill"],
+    "expected_keywords": [
+      "radar",
+      "chart",
+      "skill"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
   },
   {
     "id": "photo_random_5",
-    "url": "https://picsum.photos/seed/vision5/400/300",
+    "url": "test_images/photo_random_5.png",
     "category": "photo",
     "expected_keywords": [],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "diagram_class",
-    "url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6iuyIWyrLgXLALrPEAfFy-iCcmk-83RSjcFZ-51ac2k7AW0JqAKY9y9IcsAPzdS3jxBb5NrHUAraH_lutjbpi6oJqG7P7IPEd3-ItJsWCaO1FVYLw8qQwANsJbIt8i1AExAX0OCwjNqoa6LoPaq7oCvbHHmv5f7pVfX4K5b8mvg",
+    "url": "test_images/diagram_class.png",
     "category": "diagram",
-    "expected_keywords": ["class", "object", "attribute"],
+    "expected_keywords": [
+      "class",
+      "object",
+      "attribute"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
   },
   {
     "id": "chart_doughnut",
-    "url": "https://quickchart.io/chart?c={type:'doughnut',data:{labels:['Desktop','Mobile','Tablet'],datasets:[{data:[60,30,10]}]}}",
+    "url": "test_images/chart_doughnut.png",
     "category": "chart",
-    "expected_keywords": ["doughnut", "chart", "device"],
+    "expected_keywords": [
+      "doughnut",
+      "chart",
+      "device"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
   },
   {
     "id": "photo_random_6",
-    "url": "https://picsum.photos/seed/vision6/400/300",
+    "url": "test_images/photo_random_6.png",
     "category": "photo",
     "expected_keywords": [],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "screenshot_error",
-    "url": "https://http.cat/404.jpg",
+    "url": "test_images/screenshot_error.png",
     "category": "screenshot",
-    "expected_keywords": ["404", "error", "cat"],
+    "expected_keywords": [
+      "404",
+      "error",
+      "cat"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": true}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": true
+    }
   },
   {
     "id": "diagram_network",
-    "url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6iuyIWyrLgXLALrPEAfFy-iCcmk-83RSjcFZ-51ac2k7AW0JqAKY9y9IcsAPzdS3jxBb5NrHUAraH_lutjbpi6oJqG7P7IPEd3-ItJsWCaO1FVYLw8qQwANsJbIt8i1AExAX0OCwjNqoa6LoPaq7oCvbHHmv5f7pVfX4K5b8mvg",
+    "url": "test_images/diagram_network.png",
     "category": "diagram",
-    "expected_keywords": ["network", "node", "connection"],
+    "expected_keywords": [
+      "network",
+      "node",
+      "connection"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
   },
   {
     "id": "photo_random_7",
-    "url": "https://picsum.photos/seed/vision7/400/300",
+    "url": "test_images/photo_random_7.png",
     "category": "photo",
     "expected_keywords": [],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "chart_stacked_bar",
-    "url": "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}}}}",
+    "url": "test_images/chart_stacked_bar.png",
     "category": "chart",
-    "expected_keywords": ["stacked", "bar", "chart"],
+    "expected_keywords": [
+      "stacked",
+      "bar",
+      "chart"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
   },
   {
     "id": "screenshot_dashboard",
-    "url": "https://github.githubassets.com/images/modules/site/features-code-search.png",
+    "url": "test_images/screenshot_dashboard.png",
     "category": "screenshot",
-    "expected_keywords": ["search", "code", "feature"],
+    "expected_keywords": [
+      "search",
+      "code",
+      "feature"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   },
   {
     "id": "photo_random_8",
-    "url": "https://picsum.photos/seed/vision8/400/300",
+    "url": "test_images/photo_random_8.png",
     "category": "photo",
     "expected_keywords": [],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   }
 ]

benchmarks/vision_benchmark.py

@@ -11,17 +11,19 @@ Usage:
 
     # Single image test
     python benchmarks/vision_benchmark.py --url https://example.com/image.png
+    python benchmarks/vision_benchmark.py --url benchmarks/test_images/photo_random_1.png
 
     # Generate test report
     python benchmarks/vision_benchmark.py --images benchmarks/test_images.json --output benchmarks/vision_results.json
 
-Test image dataset: benchmarks/test_images.json (50-100 diverse images)
+Test image dataset: benchmarks/test_images.json (committed local fixtures under benchmarks/test_images/)
 """
 
 import argparse
 import asyncio
 import base64
 import json
+import mimetypes
 import os
 import statistics
 import sys
@@ -67,6 +69,28 @@ EVAL_PROMPTS = {
 # ---------------------------------------------------------------------------
 
 
+def _is_remote_image_source(image_source: str) -> bool:
+    return image_source.startswith(("http://", "https://", "data:", "file://"))
+
+
+def _image_source_to_payload_url(image_source: str) -> str:
+    """Convert local image paths into data URLs; keep remote URLs unchanged."""
+    if image_source.startswith(("http://", "https://", "data:")):
+        return image_source
+
+    resolved = image_source[len("file://"):] if image_source.startswith("file://") else image_source
+    local_path = Path(os.path.expanduser(resolved)).resolve()
+    if not local_path.is_file():
+        return image_source
+
+    mime_type, _ = mimetypes.guess_type(str(local_path))
+    if not mime_type:
+        mime_type = "application/octet-stream"
+
+    encoded = base64.b64encode(local_path.read_bytes()).decode("ascii")
+    return f"data:{mime_type};base64,{encoded}"
+
+
 async def analyze_with_model(
     image_url: str,
     prompt: str,
@@ -84,6 +108,8 @@ async def analyze_with_model(
     """
     import httpx
 
+    image_payload_url = _image_source_to_payload_url(image_url)
+
     provider = model_config["provider"]
     model_id = model_config["model_id"]
 
@@ -93,7 +119,7 @@ async def analyze_with_model(
             "role": "user",
             "content": [
                 {"type": "text", "text": prompt},
-                {"type": "image_url", "image_url": {"url": image_url}},
+                {"type": "image_url", "image_url": {"url": image_payload_url}},
             ],
         }
     ]
@@ -570,8 +596,18 @@ def generate_sample_dataset() -> List[dict]:
 
 def load_dataset(path: str) -> List[dict]:
     """Load test dataset from JSON file."""
-    with open(path) as f:
-        return json.load(f)
+    dataset_path = Path(path).resolve()
+    with open(dataset_path) as f:
+        dataset = json.load(f)
+
+    base_dir = dataset_path.parent
+    for image in dataset:
+        image_url = image.get("url")
+        if not image_url or _is_remote_image_source(image_url):
+            continue
+        image["url"] = str((base_dir / image_url).resolve())
+
+    return dataset
 
 
 # ---------------------------------------------------------------------------
@@ -582,7 +618,7 @@ def load_dataset(path: str) -> List[dict]:
 async def main():
     parser = argparse.ArgumentParser(description="Vision Benchmark Suite (Issue #817)")
     parser.add_argument("--images", help="Path to test images JSON file")
-    parser.add_argument("--url", help="Single image URL to test")
+    parser.add_argument("--url", help="Single image URL or local file path 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("--runs", type=int, default=1, help="Runs per model per image")

cli.py

@@ -13,7 +13,6 @@ Usage:
     python cli.py --list-tools             # List available tools and exit
 """
 
-import concurrent.futures
 import logging
 import os
 import shutil
@@ -64,7 +63,6 @@ from agent.usage_pricing import (
     format_duration_compact,
     format_token_count_compact,
 )
-from agent.account_usage import fetch_account_usage, render_account_usage_lines
 from hermes_cli.banner import _format_context_length, format_banner_version_label
 
 _COMMAND_SPINNER_FRAMES = ("⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏")
@@ -6473,29 +6471,6 @@ class HermesCLI:
         if cost_result.status == "unknown":
             print(f"  Note:             Pricing unknown for {agent.model}")
 
-        # Account limits -- fetched off-thread with a hard timeout so slow
-        # provider APIs don't hang the prompt.
-        provider = getattr(agent, "provider", None) or getattr(self, "provider", None)
-        base_url = getattr(agent, "base_url", None) or getattr(self, "base_url", None)
-        api_key = getattr(agent, "api_key", None) or getattr(self, "api_key", None)
-        account_snapshot = None
-        if provider:
-            with concurrent.futures.ThreadPoolExecutor(max_workers=1) as _pool:
-                try:
-                    account_snapshot = _pool.submit(
-                        fetch_account_usage,
-                        provider,
-                        base_url=base_url,
-                        api_key=api_key,
-                    ).result(timeout=10.0)
-                except (concurrent.futures.TimeoutError, Exception):
-                    account_snapshot = None
-        account_lines = [f"  {line}" for line in render_account_usage_lines(account_snapshot)]
-        if account_lines:
-            print()
-            for line in account_lines:
-                print(line)
-
         if self.verbose:
             logging.getLogger().setLevel(logging.DEBUG)
             for noisy in ('openai', 'openai._base_client', 'httpx', 'httpcore', 'asyncio', 'hpack', 'grpc', 'modal'):

gateway/run.py

@@ -28,8 +28,6 @@ from pathlib import Path
 from datetime import datetime
 from typing import Dict, Optional, Any, List
 
-from agent.account_usage import fetch_account_usage, render_account_usage_lines
-
 # ---------------------------------------------------------------------------
 # SSL certificate auto-detection for NixOS and other non-standard systems.
 # Must run BEFORE any HTTP library (discord, aiohttp, etc.) is imported.
@@ -6483,38 +6481,6 @@ class GatewayRunner:
                     if cached:
                         agent = cached[0]
 
-        # Resolve provider/base_url/api_key for the account-usage fetch.
-        # Prefer the live agent; fall back to persisted billing data on the
-        # SessionDB row so `/usage` still returns account info between turns
-        # when no agent is resident.
-        provider = getattr(agent, "provider", None) if agent and agent is not _AGENT_PENDING_SENTINEL else None
-        base_url = getattr(agent, "base_url", None) if agent and agent is not _AGENT_PENDING_SENTINEL else None
-        api_key = getattr(agent, "api_key", None) if agent and agent is not _AGENT_PENDING_SENTINEL else None
-        if not provider and getattr(self, "_session_db", None) is not None:
-            try:
-                _entry_for_billing = self.session_store.get_or_create_session(source)
-                persisted = self._session_db.get_session(_entry_for_billing.session_id) or {}
-            except Exception:
-                persisted = {}
-            provider = provider or persisted.get("billing_provider")
-            base_url = base_url or persisted.get("billing_base_url")
-
-        # Fetch account usage off the event loop so slow provider APIs don't
-        # block the gateway. Failures are non-fatal -- account_lines stays [].
-        account_lines: list[str] = []
-        if provider:
-            try:
-                account_snapshot = await asyncio.to_thread(
-                    fetch_account_usage,
-                    provider,
-                    base_url=base_url,
-                    api_key=api_key,
-                )
-            except Exception:
-                account_snapshot = None
-            if account_snapshot:
-                account_lines = render_account_usage_lines(account_snapshot, markdown=True)
-
         if agent and hasattr(agent, "session_total_tokens") and agent.session_api_calls > 0:
             lines = []
 
@@ -6572,10 +6538,6 @@ class GatewayRunner:
             if ctx.compression_count:
                 lines.append(f"Compressions: {ctx.compression_count}")
 
-            if account_lines:
-                lines.append("")
-                lines.extend(account_lines)
-
             return "\n".join(lines)
 
         # No agent at all -- check session history for a rough count
@@ -6585,18 +6547,12 @@ class GatewayRunner:
             from agent.model_metadata import estimate_messages_tokens_rough
             msgs = [m for m in history if m.get("role") in ("user", "assistant") and m.get("content")]
             approx = estimate_messages_tokens_rough(msgs)
-            lines = [
-                "📊 **Session Info**",
-                f"Messages: {len(msgs)}",
-                f"Estimated context: ~{approx:,} tokens",
-                "_(Detailed usage available after the first agent response)_",
-            ]
-            if account_lines:
-                lines.append("")
-                lines.extend(account_lines)
-            return "\n".join(lines)
-        if account_lines:
-            return "\n".join(account_lines)
+            return (
+                f"📊 **Session Info**\n"
+                f"Messages: {len(msgs)}\n"
+                f"Estimated context: ~{approx:,} tokens\n"
+                f"_(Detailed usage available after the first agent response)_"
+            )
         return "No usage data available for this session."
 
     async def _handle_insights_command(self, event: MessageEvent) -> str:

tests/gateway/test_usage_command.py

@@ -175,79 +175,3 @@ class TestUsageCachedAgent:
             result = await runner._handle_usage_command(event)
 
         assert "Cost: included" in result
-
-
-class TestUsageAccountSection:
-    """Account-limits section appended to /usage output."""
-
-    @pytest.mark.asyncio
-    async def test_usage_command_includes_account_section(self, monkeypatch):
-        agent = _make_mock_agent(provider="openai-codex")
-        agent.base_url = "https://chatgpt.com/backend-api/codex"
-        agent.api_key = "unused"
-        runner = _make_runner(SK, cached_agent=agent)
-        event = MagicMock()
-
-        monkeypatch.setattr(
-            "gateway.run.fetch_account_usage",
-            lambda provider, base_url=None, api_key=None: object(),
-        )
-        monkeypatch.setattr(
-            "gateway.run.render_account_usage_lines",
-            lambda snapshot, markdown=False: [
-                "📈 **Account limits**",
-                "Provider: openai-codex (Pro)",
-                "Session: 85% remaining (15% used)",
-            ],
-        )
-        with patch("agent.rate_limit_tracker.format_rate_limit_compact", return_value="RPM: 50/60"), \
-             patch("agent.usage_pricing.estimate_usage_cost") as mock_cost:
-            mock_cost.return_value = MagicMock(amount_usd=None, status="included")
-            result = await runner._handle_usage_command(event)
-
-        assert "📊 **Session Token Usage**" in result
-        assert "📈 **Account limits**" in result
-        assert "Provider: openai-codex (Pro)" in result
-
-    @pytest.mark.asyncio
-    async def test_usage_command_uses_persisted_provider_when_agent_not_running(self, monkeypatch):
-        runner = _make_runner(SK)
-        runner._session_db = MagicMock()
-        runner._session_db.get_session.return_value = {
-            "billing_provider": "openai-codex",
-            "billing_base_url": "https://chatgpt.com/backend-api/codex",
-        }
-        session_entry = MagicMock()
-        session_entry.session_id = "sess-1"
-        runner.session_store.get_or_create_session.return_value = session_entry
-        runner.session_store.load_transcript.return_value = [
-            {"role": "user", "content": "earlier"},
-        ]
-
-        calls = {}
-
-        async def _fake_to_thread(fn, *args, **kwargs):
-            calls["args"] = args
-            calls["kwargs"] = kwargs
-            return fn(*args, **kwargs)
-
-        monkeypatch.setattr("gateway.run.asyncio.to_thread", _fake_to_thread)
-        monkeypatch.setattr(
-            "gateway.run.fetch_account_usage",
-            lambda provider, base_url=None, api_key=None: object(),
-        )
-        monkeypatch.setattr(
-            "gateway.run.render_account_usage_lines",
-            lambda snapshot, markdown=False: [
-                "📈 **Account limits**",
-                "Provider: openai-codex (Pro)",
-            ],
-        )
-
-        event = MagicMock()
-        result = await runner._handle_usage_command(event)
-
-        assert calls["args"] == ("openai-codex",)
-        assert calls["kwargs"]["base_url"] == "https://chatgpt.com/backend-api/codex"
-        assert "📊 **Session Info**" in result
-        assert "📈 **Account limits**" in result

tests/test_account_usage.py

@@ -1,203 +0,0 @@
-from datetime import datetime, timezone
-
-from agent.account_usage import (
-    AccountUsageSnapshot,
-    AccountUsageWindow,
-    fetch_account_usage,
-    render_account_usage_lines,
-)
-
-
-class _Response:
-    def __init__(self, payload, status_code=200):
-        self._payload = payload
-        self.status_code = status_code
-
-    def raise_for_status(self):
-        if self.status_code >= 400:
-            raise RuntimeError(f"HTTP {self.status_code}")
-
-    def json(self):
-        return self._payload
-
-
-class _Client:
-    def __init__(self, payload):
-        self._payload = payload
-
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc, tb):
-        return False
-
-    def get(self, url, headers=None):
-        return _Response(self._payload)
-
-
-class _RoutingClient:
-    def __init__(self, payloads):
-        self._payloads = payloads
-
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc, tb):
-        return False
-
-    def get(self, url, headers=None):
-        return _Response(self._payloads[url])
-
-
-def test_fetch_account_usage_codex(monkeypatch):
-    monkeypatch.setattr(
-        "agent.account_usage.resolve_codex_runtime_credentials",
-        lambda refresh_if_expiring=True: {
-            "provider": "openai-codex",
-            "base_url": "https://chatgpt.com/backend-api/codex",
-            "api_key": "***",
-        },
-    )
-    monkeypatch.setattr(
-        "agent.account_usage._read_codex_tokens",
-        lambda: {"tokens": {"account_id": "acct_123"}},
-    )
-    monkeypatch.setattr(
-        "agent.account_usage.httpx.Client",
-        lambda timeout=15.0: _Client(
-            {
-                "plan_type": "pro",
-                "rate_limit": {
-                    "primary_window": {
-                        "used_percent": 15,
-                        "reset_at": 1_900_000_000,
-                        "limit_window_seconds": 18000,
-                    },
-                    "secondary_window": {
-                        "used_percent": 40,
-                        "reset_at": 1_900_500_000,
-                        "limit_window_seconds": 604800,
-                    },
-                },
-                "credits": {"has_credits": True, "balance": 12.5},
-            }
-        ),
-    )
-
-    snapshot = fetch_account_usage("openai-codex")
-
-    assert snapshot is not None
-    assert snapshot.plan == "Pro"
-    assert len(snapshot.windows) == 2
-    assert snapshot.windows[0].label == "Session"
-    assert snapshot.windows[0].used_percent == 15.0
-    assert snapshot.windows[0].reset_at == datetime.fromtimestamp(1_900_000_000, tz=timezone.utc)
-    assert "Credits balance: $12.50" in snapshot.details
-
-
-def test_render_account_usage_lines_includes_reset_and_provider():
-    snapshot = AccountUsageSnapshot(
-        provider="openai-codex",
-        source="usage_api",
-        fetched_at=datetime.now(timezone.utc),
-        plan="Pro",
-        windows=(
-            AccountUsageWindow(
-                label="Session",
-                used_percent=25,
-                reset_at=datetime.now(timezone.utc),
-            ),
-        ),
-        details=("Credits balance: $9.99",),
-    )
-    lines = render_account_usage_lines(snapshot)
-
-    assert lines[0] == "📈 Account limits"
-    assert "openai-codex (Pro)" in lines[1]
-    assert "Session: 75% remaining (25% used)" in lines[2]
-    assert "Credits balance: $9.99" in lines[3]
-
-
-def test_fetch_account_usage_openrouter_uses_limit_remaining_and_ignores_deprecated_rate_limit(monkeypatch):
-    monkeypatch.setattr(
-        "agent.account_usage.resolve_runtime_provider",
-        lambda requested, explicit_base_url=None, explicit_api_key=None: {
-            "provider": "openrouter",
-            "base_url": "https://openrouter.ai/api/v1",
-            "api_key": "***",
-        },
-    )
-    monkeypatch.setattr(
-        "agent.account_usage.httpx.Client",
-        lambda timeout=10.0: _RoutingClient(
-            {
-                "https://openrouter.ai/api/v1/credits": {
-                    "data": {"total_credits": 300.0, "total_usage": 10.92}
-                },
-                "https://openrouter.ai/api/v1/key": {
-                    "data": {
-                        "limit": 100.0,
-                        "limit_remaining": 70.0,
-                        "limit_reset": "monthly",
-                        "usage": 12.5,
-                        "usage_daily": 0.5,
-                        "usage_weekly": 2.0,
-                        "usage_monthly": 8.0,
-                        "rate_limit": {"requests": -1, "interval": "10s"},
-                    }
-                },
-            }
-        ),
-    )
-
-    snapshot = fetch_account_usage("openrouter")
-
-    assert snapshot is not None
-    assert snapshot.windows == (
-        AccountUsageWindow(
-            label="API key quota",
-            used_percent=30.0,
-            detail="$70.00 of $100.00 remaining • resets monthly",
-        ),
-    )
-    assert "Credits balance: $289.08" in snapshot.details
-    assert "API key usage: $12.50 total • $0.50 today • $2.00 this week • $8.00 this month" in snapshot.details
-    assert all("-1 requests / 10s" not in line for line in render_account_usage_lines(snapshot))
-
-
-def test_fetch_account_usage_openrouter_omits_quota_window_when_key_has_no_limit(monkeypatch):
-    monkeypatch.setattr(
-        "agent.account_usage.resolve_runtime_provider",
-        lambda requested, explicit_base_url=None, explicit_api_key=None: {
-            "provider": "openrouter",
-            "base_url": "https://openrouter.ai/api/v1",
-            "api_key": "***",
-        },
-    )
-    monkeypatch.setattr(
-        "agent.account_usage.httpx.Client",
-        lambda timeout=10.0: _RoutingClient(
-            {
-                "https://openrouter.ai/api/v1/credits": {
-                    "data": {"total_credits": 100.0, "total_usage": 25.5}
-                },
-                "https://openrouter.ai/api/v1/key": {
-                    "data": {
-                        "limit": None,
-                        "limit_remaining": None,
-                        "usage": 25.5,
-                        "usage_daily": 1.25,
-                        "usage_weekly": 4.5,
-                        "usage_monthly": 18.0,
-                    }
-                },
-            }
-        ),
-    )
-
-    snapshot = fetch_account_usage("openrouter")
-
-    assert snapshot is not None
-    assert snapshot.windows == ()
-    assert "Credits balance: $74.50" in snapshot.details
-    assert "API key usage: $25.50 total • $1.25 today • $4.50 this week • $18.00 this month" in snapshot.details

tests/test_vision_benchmark.py

@@ -11,12 +11,14 @@ import pytest
 sys.path.insert(0, str(Path(__file__).parent.parent / "benchmarks"))
 
 from vision_benchmark import (
+    analyze_with_model,
     compute_ocr_accuracy,
     compute_description_completeness,
     compute_structural_accuracy,
     aggregate_results,
     to_markdown,
     generate_sample_dataset,
+    load_dataset,
     MODELS,
     EVAL_PROMPTS,
 )
@@ -197,6 +199,71 @@ class TestMarkdown:
 
 
 class TestDataset:
+    def test_repo_dataset_uses_local_image_paths(self):
+        dataset_path = Path(__file__).parent.parent / "benchmarks" / "test_images.json"
+        dataset = json.loads(dataset_path.read_text())
+
+        assert dataset, "benchmark dataset should not be empty"
+        assert all(not entry["url"].startswith(("http://", "https://")) for entry in dataset)
+
+    def test_load_dataset_resolves_relative_local_paths(self, tmp_path):
+        images_dir = tmp_path / "images"
+        images_dir.mkdir()
+        image_path = images_dir / "sample.png"
+        image_path.write_bytes(b"png-bytes")
+
+        dataset_path = tmp_path / "dataset.json"
+        dataset_path.write_text(json.dumps([
+            {
+                "id": "sample",
+                "url": "images/sample.png",
+                "category": "photo",
+                "expected_keywords": [],
+                "expected_structure": {"min_length": 30, "min_sentences": 1},
+            }
+        ]))
+
+        loaded = load_dataset(str(dataset_path))
+
+        assert loaded[0]["url"] == str(image_path.resolve())
+
+    @pytest.mark.asyncio
+    async def test_analyze_with_model_encodes_local_file_as_data_url(self, tmp_path, monkeypatch):
+        image_path = tmp_path / "tiny.png"
+        image_path.write_bytes(
+            bytes.fromhex(
+                "89504E470D0A1A0A"
+                "0000000D49484452000000010000000108060000001F15C489"
+                "0000000D49444154789C6360000002000154A24F5D00000000"
+                "49454E44AE426082"
+            )
+        )
+
+        fake_response = MagicMock()
+        fake_response.raise_for_status.return_value = None
+        fake_response.json.return_value = {
+            "choices": [{"message": {"content": "Looks like a tiny image."}}],
+            "usage": {"prompt_tokens": 1, "completion_tokens": 2, "total_tokens": 3},
+        }
+
+        fake_client = MagicMock()
+        fake_client.post = AsyncMock(return_value=fake_response)
+        fake_ctx = MagicMock()
+        fake_ctx.__aenter__ = AsyncMock(return_value=fake_client)
+        fake_ctx.__aexit__ = AsyncMock(return_value=None)
+
+        monkeypatch.setenv("OPENROUTER_API_KEY", "test-key")
+        with patch("httpx.AsyncClient", return_value=fake_ctx):
+            result = await analyze_with_model(
+                str(image_path),
+                "Describe this image",
+                {"provider": "openrouter", "model_id": "fake/model"},
+            )
+
+        assert result["success"] is True
+        sent_url = fake_client.post.await_args.kwargs["json"]["messages"][0]["content"][1]["image_url"]["url"]
+        assert sent_url.startswith("data:image/png;base64,")
+
     def test_sample_dataset_has_entries(self):
         dataset = generate_sample_dataset()
         assert len(dataset) >= 4