fix: vendor vision benchmark fixtures (#868)

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 +1,354 @@
 [
   {
     "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")

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

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

@@ -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