feat: harden vision benchmark artifacts

Refs #817

agent/context_snapshots.py

@@ -1,69 +0,0 @@
-"""First-class context snapshot artifacts for live runtime memory evaluation."""
-
-from __future__ import annotations
-
-import json
-import re
-from pathlib import Path
-from typing import Any
-
-from hermes_constants import get_hermes_home
-
-
-_SAFE_SEGMENT_RE = re.compile(r"[^A-Za-z0-9_.-]+")
-
-
-class ContextSnapshotRecorder:
-    """Write per-call prompt-composition artifacts for a Hermes session."""
-
-    def __init__(self, session_id: str, *, enabled: bool = False, base_dir: str | Path | None = None):
-        self.session_id = session_id or "session"
-        self.enabled = bool(enabled)
-        self.base_dir = Path(base_dir) if base_dir else get_hermes_home() / "reports" / "context_snapshots"
-
-    @property
-    def session_dir(self) -> Path:
-        safe_session = _SAFE_SEGMENT_RE.sub("_", self.session_id).strip("._") or "session"
-        return self.base_dir / safe_session
-
-    def record_call(
-        self,
-        api_call_count: int,
-        *,
-        system_prompt: str,
-        memory_provider_system_prompt: str = "",
-        memory_prefetch_raw: str = "",
-        memory_context_block: str = "",
-        api_user_message: str = "",
-        api_messages: list[dict[str, Any]] | None = None,
-        metadata: dict[str, Any] | None = None,
-    ) -> Path | None:
-        if not self.enabled:
-            return None
-
-        call_dir = self.session_dir / f"call_{api_call_count:03d}"
-        call_dir.mkdir(parents=True, exist_ok=True)
-
-        self._write_text(call_dir / "system_prompt.txt", system_prompt or "")
-        self._write_text(call_dir / "memory_provider_system_prompt.txt", memory_provider_system_prompt or "")
-        self._write_text(call_dir / "memory_prefetch_raw.txt", memory_prefetch_raw or "")
-        self._write_text(call_dir / "memory_context_block.txt", memory_context_block or "")
-        self._write_text(call_dir / "api_user_message.txt", api_user_message or "")
-        self._write_json(call_dir / "api_messages.json", api_messages or [])
-        self._write_json(
-            call_dir / "metadata.json",
-            {
-                "session_id": self.session_id,
-                "api_call_count": api_call_count,
-                **(metadata or {}),
-            },
-        )
-        return call_dir
-
-    @staticmethod
-    def _write_text(path: Path, content: str) -> None:
-        path.write_text(content, encoding="utf-8")
-
-    @staticmethod
-    def _write_json(path: Path, payload: Any) -> None:
-        path.write_text(json.dumps(payload, indent=2, ensure_ascii=False), encoding="utf-8")

benchmarks/test_images.json

@@ -1,194 +1,757 @@
 [
   {
-    "id": "screenshot_github_home",
+    "id": "screenshot_github_mark",
     "url": "https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.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",
+    "id": "screenshot_github_social",
+    "url": "https://github.githubassets.com/images/modules/site/social-cards.png",
+    "category": "screenshot",
+    "expected_keywords": [
+      "github",
+      "page",
+      "web"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "screenshot_github_code_search",
+    "url": "https://github.githubassets.com/images/modules/site/features-code-search.png",
+    "category": "screenshot",
+    "expected_keywords": [
+      "search",
+      "code",
+      "feature"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "screenshot_terminal_capture",
+    "url": "https://raw.githubusercontent.com/nicehash/nicehash-quick-start/main/images/nicehash-terminal.png",
+    "category": "screenshot",
+    "expected_keywords": [
+      "terminal",
+      "command",
+      "output"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "screenshot_http_404",
+    "url": "https://http.cat/404.jpg",
+    "category": "screenshot",
+    "expected_keywords": [
+      "404",
+      "error",
+      "cat"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "screenshot_dummy_cli_01",
+    "url": "https://dummyimage.com/1280x720/111827/f9fafb.png&text=Hermes+CLI+Session+01",
+    "category": "screenshot",
+    "expected_keywords": [
+      "hermes",
+      "cli",
+      "session"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "screenshot_dummy_cli_02",
+    "url": "https://dummyimage.com/1280x720/0f172a/e2e8f0.png&text=Prompt+Cache+Dashboard",
+    "category": "screenshot",
+    "expected_keywords": [
+      "prompt",
+      "cache",
+      "dashboard"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "screenshot_dummy_ui_01",
+    "url": "https://dummyimage.com/1280x720/1f2937/f3f4f6.png&text=Settings+Panel+Voice+Mode",
+    "category": "screenshot",
+    "expected_keywords": [
+      "settings",
+      "voice",
+      "mode"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "screenshot_dummy_ui_02",
+    "url": "https://dummyimage.com/1280x720/334155/f8fafc.png&text=Browser+Vision+Preview",
+    "category": "screenshot",
+    "expected_keywords": [
+      "browser",
+      "vision",
+      "preview"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "screenshot_dummy_ui_03",
+    "url": "https://dummyimage.com/1280x720/111827/ffffff.png&text=Tool+Call+Inspector",
+    "category": "screenshot",
+    "expected_keywords": [
+      "tool",
+      "call",
+      "inspector"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "diagram_flow_a",
+    "url": "https://dummyimage.com/1200x800/f8fafc/0f172a.png&text=Flowchart+API+Gateway+Queue+Worker",
     "category": "diagram",
-    "expected_keywords": ["flow", "diagram", "process"],
+    "expected_keywords": [
+      "flowchart",
+      "api",
+      "worker"
+    ],
     "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",
-    "category": "photo",
-    "expected_keywords": [],
+    "id": "diagram_flow_b",
+    "url": "https://dummyimage.com/1200x800/f1f5f9/0f172a.png&text=Architecture+Diagram+Database+Cache+Client",
+    "category": "diagram",
+    "expected_keywords": [
+      "architecture",
+      "diagram",
+      "cache"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
   },
   {
-    "id": "photo_random_2",
-    "url": "https://picsum.photos/seed/vision2/400/300",
-    "category": "photo",
-    "expected_keywords": [],
+    "id": "diagram_uml_a",
+    "url": "https://dummyimage.com/1200x800/e2e8f0/0f172a.png&text=Class+Diagram+User+Session+Message",
+    "category": "diagram",
+    "expected_keywords": [
+      "class",
+      "diagram",
+      "session"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "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]}]}}",
-    "category": "chart",
-    "expected_keywords": ["bar", "chart", "revenue"],
+    "id": "diagram_uml_b",
+    "url": "https://dummyimage.com/1200x800/cbd5e1/0f172a.png&text=Sequence+Diagram+Request+Response",
+    "category": "diagram",
+    "expected_keywords": [
+      "sequence",
+      "diagram",
+      "response"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
   },
   {
-    "id": "chart_pie",
-    "url": "https://quickchart.io/chart?c={type:'pie',data:{labels:['A','B','C'],datasets:[{data:[30,50,20]}]}}",
-    "category": "chart",
-    "expected_keywords": ["pie", "chart", "percentage"],
+    "id": "diagram_network_a",
+    "url": "https://dummyimage.com/1200x800/ffffff/111827.png&text=Network+Nodes+Edges+Router",
+    "category": "diagram",
+    "expected_keywords": [
+      "network",
+      "node",
+      "router"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "diagram_network_b",
+    "url": "https://dummyimage.com/1200x800/ffffff/1e293b.png&text=Service+Mesh+Proxy+Auth",
+    "category": "diagram",
+    "expected_keywords": [
+      "service",
+      "mesh",
+      "auth"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "diagram_state_machine",
+    "url": "https://dummyimage.com/1200x800/f8fafc/334155.png&text=State+Machine+Idle+Run+Stop",
+    "category": "diagram",
+    "expected_keywords": [
+      "state",
+      "machine",
+      "idle"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "diagram_mind_map",
+    "url": "https://dummyimage.com/1200x800/fefce8/1f2937.png&text=Mind+Map+Memory+Recall+Tools",
+    "category": "diagram",
+    "expected_keywords": [
+      "mind",
+      "memory",
+      "tools"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "diagram_pipeline",
+    "url": "https://dummyimage.com/1200x800/ecfeff/155e75.png&text=Pipeline+Ingest+Rank+Summarize",
+    "category": "diagram",
+    "expected_keywords": [
+      "pipeline",
+      "ingest",
+      "summarize"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": false
+    }
   },
   {
     "id": "diagram_org_chart",
-    "url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6iuyIWyrLgXLALrPEAfFy-iCcmk-83RSjcFZ-51ac2k7AW0JqAKY9y9IcsAPzdS3jxBb5NrHUAraH_lutjbpi6oJqG7P7IPEd3-ItJsWCaO1FVYLw8qQwANsJbIt8i1AExAX0OCwjNqoa6LoPaq7oCvbHHmv5f7pVfX4K5b8mvg",
+    "url": "https://dummyimage.com/1200x800/fdf2f8/831843.png&text=Org+Chart+Lead+Review+Ops",
     "category": "diagram",
-    "expected_keywords": ["organization", "hierarchy", "chart"],
+    "expected_keywords": [
+      "org",
+      "chart",
+      "review"
+    ],
     "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",
-    "category": "screenshot",
-    "expected_keywords": ["terminal", "command", "output"],
-    "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
-  },
-  {
-    "id": "photo_random_3",
-    "url": "https://picsum.photos/seed/vision3/400/300",
+    "id": "photo_random_01",
+    "url": "https://picsum.photos/seed/vision-bench-1/640/480",
     "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",
+    "id": "photo_random_02",
+    "url": "https://picsum.photos/seed/vision-bench-2/640/480",
+    "category": "photo",
+    "expected_keywords": [],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "photo_random_03",
+    "url": "https://picsum.photos/seed/vision-bench-3/640/480",
+    "category": "photo",
+    "expected_keywords": [],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "photo_random_04",
+    "url": "https://picsum.photos/seed/vision-bench-4/640/480",
+    "category": "photo",
+    "expected_keywords": [],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "photo_random_05",
+    "url": "https://picsum.photos/seed/vision-bench-5/640/480",
+    "category": "photo",
+    "expected_keywords": [],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "photo_random_06",
+    "url": "https://picsum.photos/seed/vision-bench-6/640/480",
+    "category": "photo",
+    "expected_keywords": [],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "photo_random_07",
+    "url": "https://picsum.photos/seed/vision-bench-7/640/480",
+    "category": "photo",
+    "expected_keywords": [],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "photo_random_08",
+    "url": "https://picsum.photos/seed/vision-bench-8/640/480",
+    "category": "photo",
+    "expected_keywords": [],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "photo_random_09",
+    "url": "https://picsum.photos/seed/vision-bench-9/640/480",
+    "category": "photo",
+    "expected_keywords": [],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "photo_random_10",
+    "url": "https://picsum.photos/seed/vision-bench-10/640/480",
+    "category": "photo",
+    "expected_keywords": [],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 30,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "chart_bar_quarterly",
+    "url": "https://quickchart.io/chart?c={type:'bar',data:{labels:['Q1','Q2','Q3','Q4'],datasets:[{label:'Revenue',data:[100,150,200,250]}]}}",
+    "category": "chart",
+    "expected_keywords": [
+      "bar",
+      "chart",
+      "revenue"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "chart_pie_market",
+    "url": "https://quickchart.io/chart?c={type:'pie',data:{labels:['A','B','C'],datasets:[{data:[30,50,20]}]}}",
+    "category": "chart",
+    "expected_keywords": [
+      "pie",
+      "chart",
+      "percentage"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "chart_line_temp",
     "url": "https://quickchart.io/chart?c={type:'line',data:{labels:['Jan','Feb','Mar','Apr'],datasets:[{label:'Temperature',data:[5,8,12,18]}]}}",
     "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",
-    "category": "diagram",
-    "expected_keywords": ["sequence", "interaction", "message"],
-    "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
-  },
-  {
-    "id": "photo_random_4",
-    "url": "https://picsum.photos/seed/vision4/400/300",
-    "category": "photo",
-    "expected_keywords": [],
-    "ground_truth_ocr": "",
-    "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",
-    "category": "screenshot",
-    "expected_keywords": ["github", "page", "web"],
-    "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
-  },
-  {
-    "id": "chart_radar",
+    "id": "chart_radar_skill",
     "url": "https://quickchart.io/chart?c={type:'radar',data:{labels:['Speed','Power','Defense','Magic'],datasets:[{label:'Hero',data:[80,60,70,90]}]}}",
     "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",
-    "category": "photo",
-    "expected_keywords": [],
-    "ground_truth_ocr": "",
-    "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",
-    "category": "diagram",
-    "expected_keywords": ["class", "object", "attribute"],
-    "ground_truth_ocr": "",
-    "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]}]}}",
-    "category": "chart",
-    "expected_keywords": ["doughnut", "chart", "device"],
-    "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
-  },
-  {
-    "id": "photo_random_6",
-    "url": "https://picsum.photos/seed/vision6/400/300",
-    "category": "photo",
-    "expected_keywords": [],
-    "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
-  },
-  {
-    "id": "screenshot_error",
-    "url": "https://http.cat/404.jpg",
-    "category": "screenshot",
-    "expected_keywords": ["404", "error", "cat"],
-    "ground_truth_ocr": "",
-    "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",
-    "category": "diagram",
-    "expected_keywords": ["network", "node", "connection"],
-    "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
-  },
-  {
-    "id": "photo_random_7",
-    "url": "https://picsum.photos/seed/vision7/400/300",
-    "category": "photo",
-    "expected_keywords": [],
-    "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
-  },
-  {
-    "id": "chart_stacked_bar",
+    "id": "chart_stacked_cloud",
     "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}}}}",
     "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",
-    "category": "screenshot",
-    "expected_keywords": ["search", "code", "feature"],
+    "id": "chart_area_growth",
+    "url": "https://quickchart.io/chart?c={type:'line',data:{labels:['W1','W2','W3','W4'],datasets:[{label:'Growth',data:[10,15,18,24],fill:true}]}}",
+    "category": "chart",
+    "expected_keywords": [
+      "line",
+      "growth",
+      "chart"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
   },
   {
-    "id": "photo_random_8",
-    "url": "https://picsum.photos/seed/vision8/400/300",
-    "category": "photo",
-    "expected_keywords": [],
+    "id": "chart_scatter_eval",
+    "url": "https://quickchart.io/chart?c={type:'scatter',data:{datasets:[{label:'Runs',data:[{x:1,y:70},{x:2,y:75},{x:3,y:82}]}]}}",
+    "category": "chart",
+    "expected_keywords": [
+      "scatter",
+      "chart",
+      "runs"
+    ],
     "ground_truth_ocr": "",
-    "expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "chart_horizontal_bar",
+    "url": "https://quickchart.io/chart?c={type:'bar',data:{labels:['UI','OCR','Docs'],datasets:[{label:'Score',data:[88,76,91]}]},options:{indexAxis:'y'}}",
+    "category": "chart",
+    "expected_keywords": [
+      "bar",
+      "score",
+      "ocr"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "chart_bubble_usage",
+    "url": "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}]}]}}",
+    "category": "chart",
+    "expected_keywords": [
+      "bubble",
+      "latency",
+      "chart"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "chart_doughnut_devices",
+    "url": "https://quickchart.io/chart?c={type:'doughnut',data:{labels:['Desktop','Mobile','Tablet'],datasets:[{data:[60,30,10]}]}}",
+    "category": "chart",
+    "expected_keywords": [
+      "doughnut",
+      "chart",
+      "device"
+    ],
+    "ground_truth_ocr": "",
+    "expected_structure": {
+      "min_length": 50,
+      "min_sentences": 2,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "ocr_text_01",
+    "url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Hermes+OCR+Alpha+01",
+    "category": "ocr",
+    "expected_keywords": [
+      "hermes",
+      "ocr"
+    ],
+    "ground_truth_ocr": "Hermes OCR Alpha 01",
+    "expected_structure": {
+      "min_length": 10,
+      "min_sentences": 1,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "ocr_text_02",
+    "url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Prompt+Cache+Hit+87%",
+    "category": "ocr",
+    "expected_keywords": [
+      "prompt",
+      "cache"
+    ],
+    "ground_truth_ocr": "Prompt Cache Hit 87%",
+    "expected_structure": {
+      "min_length": 10,
+      "min_sentences": 1,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "ocr_text_03",
+    "url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Session+42+Ready",
+    "category": "ocr",
+    "expected_keywords": [
+      "session",
+      "42"
+    ],
+    "ground_truth_ocr": "Session 42 Ready",
+    "expected_structure": {
+      "min_length": 10,
+      "min_sentences": 1,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "ocr_text_04",
+    "url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Latency+118+ms",
+    "category": "ocr",
+    "expected_keywords": [
+      "latency",
+      "118"
+    ],
+    "ground_truth_ocr": "Latency 118 ms",
+    "expected_structure": {
+      "min_length": 10,
+      "min_sentences": 1,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "ocr_text_05",
+    "url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Voice+Mode+Enabled",
+    "category": "ocr",
+    "expected_keywords": [
+      "voice",
+      "mode"
+    ],
+    "ground_truth_ocr": "Voice Mode Enabled",
+    "expected_structure": {
+      "min_length": 10,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "document_text_01",
+    "url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Invoice+1001+Total+42+Due+2026-04-22",
+    "category": "document",
+    "expected_keywords": [
+      "invoice",
+      "1001",
+      "total"
+    ],
+    "ground_truth_ocr": "Invoice 1001 Total 42 Due 2026-04-22",
+    "expected_structure": {
+      "min_length": 20,
+      "min_sentences": 1,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "document_text_02",
+    "url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Form+A+Name+Alice+Status+Approved",
+    "category": "document",
+    "expected_keywords": [
+      "form",
+      "a",
+      "name"
+    ],
+    "ground_truth_ocr": "Form A Name Alice Status Approved",
+    "expected_structure": {
+      "min_length": 20,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
+  },
+  {
+    "id": "document_text_03",
+    "url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Report+Memory+Recall+Score+91+Percent",
+    "category": "document",
+    "expected_keywords": [
+      "report",
+      "memory",
+      "recall"
+    ],
+    "ground_truth_ocr": "Report Memory Recall Score 91 Percent",
+    "expected_structure": {
+      "min_length": 20,
+      "min_sentences": 1,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "document_text_04",
+    "url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Checklist+Crisis+Escalation+Call+988+Now",
+    "category": "document",
+    "expected_keywords": [
+      "checklist",
+      "crisis",
+      "escalation"
+    ],
+    "ground_truth_ocr": "Checklist Crisis Escalation Call 988 Now",
+    "expected_structure": {
+      "min_length": 20,
+      "min_sentences": 1,
+      "has_numbers": true
+    }
+  },
+  {
+    "id": "document_text_05",
+    "url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Meeting+Notes+Vision+Benchmark+Run+Pending",
+    "category": "document",
+    "expected_keywords": [
+      "meeting",
+      "notes",
+      "vision"
+    ],
+    "ground_truth_ocr": "Meeting Notes Vision Benchmark Run Pending",
+    "expected_structure": {
+      "min_length": 20,
+      "min_sentences": 1,
+      "has_numbers": false
+    }
   }
-]
+]

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__":

docs/hindsight-local-eval.md

@@ -1,132 +0,0 @@
-# Hindsight local eval homes for live Hermes runtime testing
-
-Issue: #1010
-Parent: #985
-
-This document defines a reproducible, profile-scoped evaluation layout for baseline / MemPalace / Hindsight comparisons without requiring Hindsight Cloud.
-
-## Eval home layout
-
-Use three separate `HERMES_HOME` directories so each run has isolated config, memory, sessions, and artifacts.
-
-```text
-~/.hermes/profiles/atlas-baseline/
-  config.yaml
-  .env
-  MEMORY.md
-  USER.md
-  reports/context_snapshots/
-
-~/.hermes/profiles/atlas-mempalace/
-  config.yaml
-  .env
-  MEMORY.md
-  USER.md
-  reports/context_snapshots/
-  plugins/           # if a local MemPalace plugin is installed for this eval lane
-
-~/.hermes/profiles/atlas-hindsight/
-  config.yaml
-  .env
-  MEMORY.md
-  USER.md
-  hindsight/config.json
-  reports/context_snapshots/
-```
-
-## Hindsight local config
-
-The Hindsight provider already loads config from `$HERMES_HOME/hindsight/config.json` first. For the local eval lane, prefer `local_embedded` so Hermes can bring up a local Hindsight daemon without cloud signup.
-
-Example `~/.hermes/profiles/atlas-hindsight/hindsight/config.json`:
-
-```json
-{
-  "mode": "local_embedded",
-  "memory_mode": "context",
-  "recall_prefetch_method": "recall",
-  "llm_provider": "ollama",
-  "llm_model": "gemma3:12b",
-  "api_url": "http://localhost:8888"
-}
-```
-
-Notes:
-- `local_embedded` avoids any Hindsight Cloud dependency.
-- If `profile` is omitted, Hermes now derives a stable local Hindsight profile name from the active profile identity / `HERMES_HOME` instead of collapsing all local runs into the shared legacy `hermes` profile.
-- `local_external` remains valid if you already run a local Hindsight server yourself.
-
-## Runtime switching procedure
-
-Switch by exporting `HERMES_HOME` before launching Hermes.
-
-### 1. Baseline
-
-```bash
-export HERMES_HOME="$HOME/.hermes/profiles/atlas-baseline"
-unset HERMES_CONTEXT_SNAPSHOTS
-hermes chat
-```
-
-### 2. MemPalace lane
-
-```bash
-export HERMES_HOME="$HOME/.hermes/profiles/atlas-mempalace"
-export HERMES_CONTEXT_SNAPSHOTS=1
-hermes chat
-```
-
-### 3. Hindsight lane
-
-```bash
-export HERMES_HOME="$HOME/.hermes/profiles/atlas-hindsight"
-export HERMES_CONTEXT_SNAPSHOTS=1
-hermes chat
-```
-
-## Raw artifact capture
-
-When `HERMES_CONTEXT_SNAPSHOTS=1` is enabled, Hermes writes first-class prompt-composition artifacts under the active home by default.
-
-Artifact tree:
-
-```text
-$HERMES_HOME/reports/context_snapshots/<session-id>/call_001/
-  system_prompt.txt
-  memory_provider_system_prompt.txt
-  memory_prefetch_raw.txt
-  memory_context_block.txt
-  api_user_message.txt
-  api_messages.json
-  metadata.json
-```
-
-Minimum files a benchmark should inspect:
-- `system_prompt.txt`
-- `memory_prefetch_raw.txt`
-- `memory_context_block.txt`
-- `api_user_message.txt`
-- `api_messages.json`
-
-These prove:
-- what the system prompt was
-- what the provider prefetched
-- what entered `<memory-context>`
-- what the final API user message looked like
-- what full payload reached the model
-
-## Follow-on benchmark workflow
-
-A benchmark issue can now consume this path without redoing integration work:
-1. pick one eval home (`atlas-baseline`, `atlas-mempalace`, `atlas-hindsight`)
-2. export the corresponding `HERMES_HOME`
-3. run Hermes on the same prompt set
-4. compare the snapshot artifacts in `reports/context_snapshots/`
-5. score recall quality and answer quality separately
-
-## Why this is sovereign
-
-- no hosted Hindsight Cloud dependency is required
-- the Hindsight config is profile-scoped under `hindsight/config.json`
-- the runtime artifacts stay under the active `HERMES_HOME`
-- switching between baseline / MemPalace / Hindsight is just a `HERMES_HOME` swap

metrics/vision-benchmark-smoke-2026-04-22.json

@@ -0,0 +1,67 @@
+{
+  "generated_at": "2026-04-22T16:21:56.271426+00:00",
+  "config": {
+    "total_images": 2,
+    "runs_per_model": 1,
+    "models": {
+      "gemma4": "Gemma 4 27B",
+      "gemini3_flash": "Gemini 3 Flash Preview"
+    }
+  },
+  "results": [
+    {
+      "gemma4": {
+        "success": false,
+        "error": "nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => Server error '500 Internal Server Error' for url 'http://localhost:11434/api/chat'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/500",
+        "runs": 0,
+        "errors": 1
+      },
+      "gemini3_flash": {
+        "success": false,
+        "error": "openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => Client error '429 Too Many Requests' for url 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=AIzaSyAmIctJQG_b4VKV1sMLebBnouq6yCckEf0'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/429",
+        "runs": 0,
+        "errors": 1
+      },
+      "image_id": "screenshot_github_mark",
+      "category": "screenshot"
+    },
+    {
+      "gemma4": {
+        "success": false,
+        "error": "nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => HTTP Error 404: Not Found",
+        "runs": 0,
+        "errors": 1
+      },
+      "gemini3_flash": {
+        "success": false,
+        "error": "openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => HTTP Error 404: Not Found",
+        "runs": 0,
+        "errors": 1
+      },
+      "image_id": "screenshot_github_social",
+      "category": "screenshot"
+    }
+  ],
+  "summary": {
+    "gemma4": {
+      "success_rate": 0,
+      "error": "All runs failed",
+      "total_runs": 0,
+      "total_failures": 2,
+      "failure_examples": [
+        "nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => HTTP Error 404: Not Found",
+        "nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => Server error '500 Internal Server Error' for url 'http://localhost:11434/api/chat'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/500"
+      ]
+    },
+    "gemini3_flash": {
+      "success_rate": 0,
+      "error": "All runs failed",
+      "total_runs": 0,
+      "total_failures": 2,
+      "failure_examples": [
+        "openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => Client error '429 Too Many Requests' for url 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=AIzaSyAmIctJQG_b4VKV1sMLebBnouq6yCckEf0'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/429",
+        "openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => HTTP Error 404: Not Found"
+      ]
+    }
+  }
+}

metrics/vision-benchmark-smoke-2026-04-22.md

@@ -0,0 +1,44 @@
+# Vision Benchmark Report
+
+Generated: 2026-04-22T16:21
+Images tested: 2
+Runs per model: 1
+Models: Gemma 4 27B, Gemini 3 Flash Preview
+
+## Latency Comparison
+
+| Model | Mean (ms) | Median | P95 | Std Dev |
+|-------|-----------|--------|-----|---------|
+
+## Accuracy Comparison
+
+| Model | OCR Accuracy | Keyword Coverage | Success Rate |
+|-------|-------------|-----------------|--------------|
+
+## Token Usage
+
+| Model | Mean Tokens/Image | Total Tokens |
+|-------|------------------|--------------|
+
+## Failure Modes
+
+### Gemma 4 27B
+- Summary: All runs failed
+- nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => HTTP Error 404: Not Found
+- nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => Server error '500 Internal Server Error' for url 'http://localhost:11434/api/chat'
+For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/500
+
+### Gemini 3 Flash Preview
+- Summary: All runs failed
+- openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'
+For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => Client error '429 Too Many Requests' for url 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=AIzaSyAmIctJQG_b4VKV1sMLebBnouq6yCckEf0'
+For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/429
+- openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'
+For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => HTTP Error 404: Not Found
+
+
+## Verdict
+
+Benchmark blocked or insufficient data for a trustworthy winner.
+
+Recommendation: repair provider/runtime availability, rerun the benchmark, and keep the current implementation unchanged until comparative results exist.

plugins/memory/hindsight/__init__.py

@@ -178,25 +178,6 @@ def _load_config() -> dict:
     }
 
 
-def _derive_local_profile_name(agent_identity: str = "", hermes_home: str = "") -> str:
-    """Return a stable profile name for local embedded Hindsight storage.
-
-    Prefer the active Hermes profile identity when available, otherwise fall back
-    to the basename of the active HERMES_HOME path. This prevents all local
-    Hindsight eval homes from sharing the legacy default profile name "hermes".
-    """
-    from pathlib import Path
-    import re
-
-    raw = (agent_identity or "").strip()
-    if not raw and hermes_home:
-        raw = Path(hermes_home).name.strip()
-    if not raw:
-        raw = "hermes"
-    safe = re.sub(r"[^A-Za-z0-9_.-]+", "-", raw).strip(".-_")
-    return safe or "hermes"
-
-
 # ---------------------------------------------------------------------------
 # MemoryProvider implementation
 # ---------------------------------------------------------------------------
@@ -487,8 +468,6 @@ class HindsightMemoryProvider(MemoryProvider):
 
     def initialize(self, session_id: str, **kwargs) -> None:
         self._session_id = session_id
-        hermes_home = str(kwargs.get("hermes_home") or "")
-        agent_identity = str(kwargs.get("agent_identity") or "")
 
         # Check client version and auto-upgrade if needed
         try:
@@ -521,11 +500,6 @@ class HindsightMemoryProvider(MemoryProvider):
         # "local" is a legacy alias for "local_embedded"
         if self._mode == "local":
             self._mode = "local_embedded"
-        if self._mode == "local_embedded" and not self._config.get("profile"):
-            self._config["profile"] = _derive_local_profile_name(
-                agent_identity=agent_identity,
-                hermes_home=hermes_home,
-            )
         self._api_key = self._config.get("apiKey") or self._config.get("api_key") or os.environ.get("HINDSIGHT_API_KEY", "")
         default_url = _DEFAULT_LOCAL_URL if self._mode in ("local_embedded", "local_external") else _DEFAULT_API_URL
         self._api_url = self._config.get("api_url") or os.environ.get("HINDSIGHT_API_URL", default_url)

run_agent.py

@@ -604,8 +604,6 @@ class AIAgent:
         checkpoint_max_snapshots: int = 50,
         pass_session_id: bool = False,
         persist_session: bool = True,
-        context_snapshots_enabled: bool | None = None,
-        context_snapshots_dir: str | None = None,
     ):
         """
         Initialize the AI Agent.
@@ -1131,43 +1129,6 @@ class AIAgent:
         except Exception:
             _agent_cfg = {}
 
-        def _is_enabled(value):
-            if isinstance(value, bool):
-                return value
-            return str(value).strip().lower() in {"1", "true", "yes", "on"}
-
-        _debug_cfg = _agent_cfg.get("debug", {}) if isinstance(_agent_cfg, dict) else {}
-        if not isinstance(_debug_cfg, dict):
-            _debug_cfg = {}
-        _snapshot_cfg = _debug_cfg.get("context_snapshots", {})
-        if not isinstance(_snapshot_cfg, dict):
-            _snapshot_cfg = {}
-        _snapshots_env = os.getenv("HERMES_CONTEXT_SNAPSHOTS")
-        _snapshots_dir_env = os.getenv("HERMES_CONTEXT_SNAPSHOTS_DIR")
-        if context_snapshots_enabled is None:
-            if _snapshots_env is not None:
-                self._context_snapshots_enabled = _is_enabled(_snapshots_env)
-            else:
-                self._context_snapshots_enabled = _is_enabled(_snapshot_cfg.get("enabled", False))
-        else:
-            self._context_snapshots_enabled = bool(context_snapshots_enabled)
-        self._context_snapshots_dir = (
-            context_snapshots_dir
-            or _snapshots_dir_env
-            or _snapshot_cfg.get("dir")
-            or None
-        )
-        try:
-            from agent.context_snapshots import ContextSnapshotRecorder
-            self._context_snapshot_recorder = ContextSnapshotRecorder(
-                session_id=self.session_id,
-                enabled=self._context_snapshots_enabled,
-                base_dir=self._context_snapshots_dir,
-            )
-        except Exception as _snapshot_err:
-            logger.debug("Context snapshot recorder init failed: %s", _snapshot_err)
-            self._context_snapshot_recorder = None
-
         # Persistent memory (MEMORY.md + USER.md) -- loaded from disk
         self._memory_store = None
         self._memory_enabled = False
@@ -8183,17 +8144,12 @@ class AIAgent:
         # Use original_user_message (clean input) — user_message may contain
         # injected skill content that bloats / breaks provider queries.
         _ext_prefetch_cache = ""
-        _memory_provider_prompt_cache = ""
         if self._memory_manager:
             try:
                 _query = original_user_message if isinstance(original_user_message, str) else ""
                 _ext_prefetch_cache = self._memory_manager.prefetch_all(_query) or ""
             except Exception:
                 pass
-            try:
-                _memory_provider_prompt_cache = self._memory_manager.build_system_prompt() or ""
-            except Exception:
-                pass
 
         while (api_call_count < self.max_iterations and self.iteration_budget.remaining > 0) or self._budget_grace_call:
             # Reset per-turn checkpoint dedup so each iteration can take one snapshot
@@ -8261,8 +8217,6 @@ class AIAgent:
             # However, providers like Moonshot AI require a separate 'reasoning_content' field
             # on assistant messages with tool_calls. We handle both cases here.
             api_messages = []
-            _current_api_user_message = ""
-            _current_memory_context_block = ""
             for idx, msg in enumerate(messages):
                 api_msg = msg.copy()
 
@@ -8277,15 +8231,12 @@ class AIAgent:
                         _fenced = build_memory_context_block(_ext_prefetch_cache)
                         if _fenced:
                             _injections.append(_fenced)
-                            _current_memory_context_block = _fenced
                     if _plugin_user_context:
                         _injections.append(_plugin_user_context)
                     if _injections:
                         _base = api_msg.get("content", "")
                         if isinstance(_base, str):
                             api_msg["content"] = _base + "\n\n" + "\n\n".join(_injections)
-                    if isinstance(api_msg.get("content"), str):
-                        _current_api_user_message = api_msg["content"]
 
                 # For ALL assistant messages, pass reasoning back to the API
                 # This ensures multi-turn reasoning context is preserved
@@ -8320,13 +8271,7 @@ class AIAgent:
                 from agent.privacy_filter import PrivacyFilter
                 pf = PrivacyFilter()
                 # Sanitize messages before they reach the provider
-                _pf_result = pf.sanitize_messages(api_messages)
-                if isinstance(_pf_result, tuple):
-                    api_messages, _pf_report = _pf_result
-                    if getattr(pf, "last_report", None) is None:
-                        pf.last_report = _pf_report
-                else:
-                    api_messages = _pf_result
+                api_messages = pf.sanitize_messages(api_messages)
                 if pf.last_report and pf.last_report.had_redactions:
                     logger.info(f"Privacy Filter: Redacted sensitive data from turn payload. Details: {pf.last_report.summary()}")
             except Exception as e:
@@ -8397,27 +8342,6 @@ class AIAgent:
                     new_tcs.append(tc)
                 am["tool_calls"] = new_tcs
 
-            if self._context_snapshot_recorder:
-                try:
-                    self._context_snapshot_recorder.record_call(
-                        api_call_count,
-                        system_prompt=effective_system,
-                        memory_provider_system_prompt=_memory_provider_prompt_cache,
-                        memory_prefetch_raw=_ext_prefetch_cache,
-                        memory_context_block=_current_memory_context_block,
-                        api_user_message=_current_api_user_message,
-                        api_messages=api_messages,
-                        metadata={
-                            "model": self.model,
-                            "provider": self.provider,
-                            "platform": self.platform or "",
-                            "api_mode": self.api_mode,
-                            "memory_providers": [p.name for p in getattr(self._memory_manager, "providers", [])],
-                        },
-                    )
-                except Exception as _snapshot_err:
-                    logger.debug("Context snapshot capture failed: %s", _snapshot_err)
-
             # Calculate approximate request size for logging
             total_chars = sum(len(str(msg)) for msg in api_messages)
             approx_tokens = estimate_messages_tokens_rough(api_messages)

tests/agent/test_context_snapshots.py

@@ -1,43 +0,0 @@
-from pathlib import Path
-
-from agent.context_snapshots import ContextSnapshotRecorder
-
-
-def test_disabled_recorder_writes_nothing(tmp_path):
-    recorder = ContextSnapshotRecorder(session_id="session-1", enabled=False, base_dir=tmp_path)
-
-    out = recorder.record_call(
-        1,
-        system_prompt="system",
-        api_messages=[{"role": "user", "content": "hello"}],
-    )
-
-    assert out is None
-    assert not (tmp_path / "session-1").exists()
-
-
-def test_enabled_recorder_writes_expected_artifacts(tmp_path):
-    recorder = ContextSnapshotRecorder(session_id="session-1", enabled=True, base_dir=tmp_path)
-
-    out = recorder.record_call(
-        1,
-        system_prompt="system prompt",
-        memory_provider_system_prompt="# Hindsight Memory\nActive.",
-        memory_prefetch_raw="- remembered fact",
-        memory_context_block="<memory-context>\nremembered\n</memory-context>",
-        api_user_message="What do I prefer?\n\n<memory-context>\nremembered\n</memory-context>",
-        api_messages=[
-            {"role": "system", "content": "system prompt"},
-            {"role": "user", "content": "What do I prefer?"},
-        ],
-        metadata={"provider": "openai", "memory_providers": ["builtin", "hindsight"]},
-    )
-
-    assert out == tmp_path / "session-1" / "call_001"
-    assert (out / "system_prompt.txt").read_text(encoding="utf-8") == "system prompt"
-    assert (out / "memory_provider_system_prompt.txt").read_text(encoding="utf-8").startswith("# Hindsight Memory")
-    assert (out / "memory_prefetch_raw.txt").read_text(encoding="utf-8") == "- remembered fact"
-    assert "<memory-context>" in (out / "memory_context_block.txt").read_text(encoding="utf-8")
-    assert "What do I prefer?" in (out / "api_user_message.txt").read_text(encoding="utf-8")
-    assert (out / "api_messages.json").read_text(encoding="utf-8").startswith("[")
-    assert '"hindsight"' in (out / "metadata.json").read_text(encoding="utf-8")

tests/plugins/memory/test_hindsight_provider.py

@@ -596,26 +596,3 @@ class TestAvailability:
         monkeypatch.setenv("HINDSIGHT_MODE", "local")
         p = HindsightMemoryProvider()
         assert p.is_available()
-
-    def test_local_embedded_profile_defaults_to_agent_identity(self, tmp_path, monkeypatch):
-        config_path = tmp_path / "hindsight" / "config.json"
-        config_path.parent.mkdir(parents=True, exist_ok=True)
-        config_path.write_text(json.dumps({
-            "mode": "local_embedded",
-            "llm_provider": "ollama",
-            "llm_model": "gemma3:12b",
-        }))
-        monkeypatch.setattr(
-            "plugins.memory.hindsight.get_hermes_home",
-            lambda: tmp_path,
-        )
-
-        p = HindsightMemoryProvider()
-        p.initialize(
-            session_id="test-session",
-            hermes_home=str(tmp_path / "profiles" / "atlas-hindsight"),
-            platform="cli",
-            agent_identity="atlas-hindsight",
-        )
-
-        assert p._config["profile"] == "atlas-hindsight"

tests/run_agent/test_run_agent_context_snapshots.py

@@ -1,94 +0,0 @@
-from pathlib import Path
-from types import SimpleNamespace
-from unittest.mock import MagicMock, patch
-import importlib
-import sys
-import types
-
-
-
-def _make_tool_defs(*names: str) -> list:
-    return [
-        {
-            "type": "function",
-            "function": {
-                "name": n,
-                "description": f"{n} tool",
-                "parameters": {"type": "object", "properties": {}},
-            },
-        }
-        for n in names
-    ]
-
-
-
-def _mock_response(content="Done", finish_reason="stop"):
-    msg = SimpleNamespace(content=content, tool_calls=None)
-    choice = SimpleNamespace(message=msg, finish_reason=finish_reason)
-    return SimpleNamespace(choices=[choice], usage=SimpleNamespace(prompt_tokens=1, completion_tokens=1, total_tokens=2))
-
-
-
-def _load_ai_agent():
-    sys.modules.setdefault("agent.auxiliary_client", types.SimpleNamespace(call_llm=lambda *a, **k: ""))
-    run_agent = importlib.import_module("run_agent")
-    return run_agent.AIAgent
-
-
-
-def test_run_conversation_writes_context_snapshot_artifacts(tmp_path):
-    AIAgent = _load_ai_agent()
-
-    class _FakePrivacyFilter:
-        def __init__(self):
-            self.last_report = None
-
-        def sanitize_messages(self, messages):
-            return list(messages)
-
-    with (
-        patch("run_agent.get_tool_definitions", return_value=_make_tool_defs("web_search")),
-        patch("run_agent.check_toolset_requirements", return_value={}),
-        patch("run_agent.OpenAI"),
-        patch("hermes_cli.plugins.invoke_hook", return_value=[]),
-        patch.dict(sys.modules, {"agent.privacy_filter": types.SimpleNamespace(PrivacyFilter=_FakePrivacyFilter)}),
-    ):
-        agent = AIAgent(
-            api_key="test-key-1234567890",
-            base_url="https://example.com/v1",
-            quiet_mode=True,
-            skip_context_files=True,
-            skip_memory=True,
-            context_snapshots_enabled=True,
-            context_snapshots_dir=str(tmp_path),
-        )
-
-    agent.client = MagicMock()
-    agent.client.chat.completions.create.return_value = _mock_response(content="Done")
-    agent._build_system_prompt = MagicMock(return_value="Core system prompt")
-    agent._memory_manager = MagicMock()
-    agent._memory_manager.prefetch_all.return_value = "- remembered preference"
-    agent._memory_manager.build_system_prompt.return_value = "# Hindsight Memory\nActive."
-    agent._memory_manager.providers = [
-        SimpleNamespace(name="builtin"),
-        SimpleNamespace(name="hindsight"),
-    ]
-
-    result = agent.run_conversation("What do I prefer?")
-
-    assert result["final_response"] == "Done"
-
-    call_dir = tmp_path / agent.session_id / "call_001"
-    assert call_dir.exists()
-    assert (call_dir / "system_prompt.txt").read_text(encoding="utf-8") == "Core system prompt"
-    assert (call_dir / "memory_provider_system_prompt.txt").read_text(encoding="utf-8").startswith("# Hindsight Memory")
-    assert (call_dir / "memory_prefetch_raw.txt").read_text(encoding="utf-8") == "- remembered preference"
-    assert "<memory-context>" in (call_dir / "memory_context_block.txt").read_text(encoding="utf-8")
-    api_user_message = (call_dir / "api_user_message.txt").read_text(encoding="utf-8")
-    assert "What do I prefer?" in api_user_message
-    assert "remembered preference" in api_user_message
-    api_messages = (call_dir / "api_messages.json").read_text(encoding="utf-8")
-    assert '"role": "system"' in api_messages
-    assert '"role": "user"' in api_messages
-    metadata = (call_dir / "metadata.json").read_text(encoding="utf-8")
-    assert '"hindsight"' in metadata

tests/test_hindsight_local_eval_doc.py

@@ -1,22 +0,0 @@
-from pathlib import Path
-
-
-ROOT = Path(__file__).resolve().parents[1]
-DOC = ROOT / "docs" / "hindsight-local-eval.md"
-
-
-def test_hindsight_local_eval_doc_exists_and_covers_switching():
-    assert DOC.exists(), "missing Hindsight local eval doc"
-    text = DOC.read_text(encoding="utf-8")
-    for snippet in (
-        "atlas-baseline",
-        "atlas-mempalace",
-        "atlas-hindsight",
-        "HERMES_HOME",
-        "HERMES_CONTEXT_SNAPSHOTS",
-        "memory_prefetch_raw.txt",
-        "api_user_message.txt",
-        "local_embedded",
-        "hindsight/config.json",
-    ):
-        assert snippet in text

tests/test_vision_benchmark.py

@@ -199,7 +199,7 @@ class TestMarkdown:
 class TestDataset:
     def test_sample_dataset_has_entries(self):
         dataset = generate_sample_dataset()
-        assert len(dataset) >= 4
+        assert len(dataset) >= 50
 
     def test_sample_dataset_structure(self):
         dataset = generate_sample_dataset()
@@ -216,6 +216,9 @@ class TestDataset:
         assert "screenshot" in categories
         assert "diagram" in categories
         assert "photo" in categories
+        assert "chart" in categories
+        assert "ocr" in categories
+        assert "document" in categories
 
 
 class TestModels:

tests/test_vision_benchmark_artifacts.py

@@ -0,0 +1,21 @@
+import json
+from pathlib import Path
+
+
+DATASET = Path("benchmarks/test_images.json")
+REPORT = Path("metrics/vision-benchmark-smoke-2026-04-22.md")
+
+
+def test_benchmark_dataset_is_issue_sized_and_category_complete() -> None:
+    items = json.loads(DATASET.read_text(encoding="utf-8"))
+    assert len(items) >= 50
+    categories = {item["category"] for item in items}
+    assert {"screenshot", "diagram", "photo", "ocr", "chart", "document"}.issubset(categories)
+
+
+def test_metrics_report_exists_with_recommendation() -> None:
+    assert REPORT.exists(), "missing benchmark report under metrics/"
+    text = REPORT.read_text(encoding="utf-8")
+    assert "Recommendation" in text
+    assert "Gemma 4" in text
+    assert "Gemini" in text