feat(catalog): add Gemma 4 12B/QAT entries and RTX 3050 bandwidth (#4728)

Add official Gemma 4 12B-it plus QAT-INT4/INT8 catalog entries (with their
GGUF sources), QAT quantization support across the quant tables and the
prequantized-prefix list, and the missing RTX 3050 / 3050 Ti memory
bandwidth so speed estimates stop falling back to the generic cuda value.

app.py

@@ -44,7 +44,7 @@ from typing import Dict
 
 from contextlib import asynccontextmanager
 from fastapi import FastAPI, Request, HTTPException
-from fastapi.responses import JSONResponse, FileResponse, HTMLResponse
+from fastapi.responses import JSONResponse, FileResponse
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.staticfiles import StaticFiles
 from starlette.middleware.base import BaseHTTPMiddleware
@@ -65,7 +65,7 @@ from core.exceptions import (
 
 import bcrypt as _bcrypt
 
-from src.app_helpers import abs_join
+from src.app_helpers import abs_join, serve_html_with_nonce
 from src.generated_images import GENERATED_IMAGE_HEADERS, resolve_generated_image_path
 from starlette.responses import RedirectResponse
 
@@ -791,22 +791,14 @@ app.include_router(setup_companion_routes())
 
 # ========= ROUTES (kept in app.py) =========
 
-def _serve_html_with_nonce(request: Request, file_path: str) -> HTMLResponse:
-    """Read an HTML file and inject the CSP nonce into inline <script> tags."""
-    with open(file_path, "r", encoding="utf-8") as f:
-        html = f.read()
-    nonce = getattr(request.state, "csp_nonce", "")
-    html = html.replace("{{CSP_NONCE}}", nonce)
-    return HTMLResponse(html)
-
 @app.get("/")
 async def serve_index(request: Request):
     static_path = abs_join(BASE_DIR, "static/index.html")
     if os.path.exists(static_path):
-        return _serve_html_with_nonce(request, static_path)
+        return serve_html_with_nonce(request, static_path)
     root_path = abs_join(BASE_DIR, "index.html")
     if os.path.exists(root_path):
-        return _serve_html_with_nonce(request, root_path)
+        return serve_html_with_nonce(request, root_path)
     raise HTTPException(404, "index.html not found")
 
 @app.get("/notes")
@@ -848,13 +840,13 @@ async def serve_library(request: Request):
 @app.get("/backgrounds")
 async def serve_backgrounds(request: Request):
     """Sandbox page for prototyping background effects. No auth required."""
-    return _serve_html_with_nonce(request, abs_join(BASE_DIR, "static/backgrounds.html"))
+    return serve_html_with_nonce(request, abs_join(BASE_DIR, "static/backgrounds.html"))
 
 @app.get("/login")
 async def serve_login(request: Request):
     if not AUTH_ENABLED:
         return RedirectResponse(url="/", status_code=302)
-    return _serve_html_with_nonce(request, abs_join(BASE_DIR, "static/login.html"))
+    return serve_html_with_nonce(request, abs_join(BASE_DIR, "static/login.html"))
 
 @app.get("/api/version")
 async def get_version():

services/hwfit/data/hf_models.json

@@ -14059,6 +14059,138 @@
    "vision"
   ]
  },
+ {
+  "name": "google/gemma-4-12B-it",
+  "provider": "Google",
+  "parameter_count": "12.0B",
+  "parameters_raw": 12000000000,
+  "min_ram_gb": 8.5,
+  "recommended_ram_gb": 11.0,
+  "min_vram_gb": 7.5,
+  "quantization": "Q4_K_M",
+  "context_length": 131072,
+  "use_case": "General purpose, multimodal; unsloth/gemma-4-12B-it-GGUF Dynamic variants reduce VRAM from ~7.5 GB to ~5.5 GB",
+  "is_moe": false,
+  "num_experts": null,
+  "active_experts": null,
+  "active_parameters": null,
+  "architecture": "gemma4",
+  "pipeline_tag": "image-text-to-text",
+  "release_date": "2026-04-01",
+  "gguf_sources": [
+   {
+    "repo": "unsloth/gemma-4-12B-it-GGUF",
+    "provider": "unsloth"
+   }
+  ],
+  "capabilities": [
+   "vision"
+  ]
+ },
+ {
+  "name": "google/gemma-4-12B-it-qat-int4",
+  "provider": "Google",
+  "parameter_count": "12.0B",
+  "parameters_raw": 12000000000,
+  "min_ram_gb": 8.0,
+  "recommended_ram_gb": 9.5,
+  "min_vram_gb": 6.5,
+  "quantization": "QAT-INT4",
+  "context_length": 131072,
+  "use_case": "General purpose, multimodal (QAT quantization-aware training — higher quality than post-train INT4; vLLM native; no GGUF)",
+  "is_moe": false,
+  "num_experts": null,
+  "active_experts": null,
+  "active_parameters": null,
+  "architecture": "gemma4",
+  "pipeline_tag": "image-text-to-text",
+  "release_date": "2026-04-01",
+  "gguf_sources": [],
+  "capabilities": [
+   "vision"
+  ]
+ },
+ {
+  "name": "google/gemma-4-12B-it-qat-int8",
+  "provider": "Google",
+  "parameter_count": "12.0B",
+  "parameters_raw": 12000000000,
+  "min_ram_gb": 15.0,
+  "recommended_ram_gb": 20.0,
+  "min_vram_gb": 13.5,
+  "quantization": "QAT-INT8",
+  "context_length": 131072,
+  "use_case": "General purpose, multimodal (QAT INT8 — highest quality, 2x VRAM of QAT-INT4; vLLM native; no GGUF)",
+  "is_moe": false,
+  "num_experts": null,
+  "active_experts": null,
+  "active_parameters": null,
+  "architecture": "gemma4",
+  "pipeline_tag": "image-text-to-text",
+  "release_date": "2026-04-01",
+  "gguf_sources": [],
+  "capabilities": [
+   "vision"
+  ]
+ },
+ {
+  "name": "google/gemma-4-12B-it-qat-q4_0-gguf",
+  "provider": "Google",
+  "parameter_count": "12.0B",
+  "parameters_raw": 12000000000,
+  "min_ram_gb": 8.5,
+  "recommended_ram_gb": 11.0,
+  "min_vram_gb": 7.5,
+  "quantization": "QAT-INT4",
+  "context_length": 262144,
+  "use_case": "General purpose, multimodal (vision + audio); official Google QAT int4 GGUF — near-bf16 quality at int4 size, served on llama.cpp/Ollama with CPU offload",
+  "is_moe": false,
+  "num_experts": null,
+  "active_experts": null,
+  "active_parameters": null,
+  "architecture": "gemma4",
+  "pipeline_tag": "image-text-to-text",
+  "release_date": "2026-04-01",
+  "gguf_sources": [
+   {
+    "repo": "google/gemma-4-12B-it-qat-q4_0-gguf",
+    "provider": "Google",
+    "file": "gemma-4-12b-it-qat-q4_0.gguf"
+   }
+  ],
+  "capabilities": [
+   "vision",
+   "audio"
+  ]
+ },
+ {
+  "name": "google/gemma-4-26B-A4B-it-qat-q4_0-gguf",
+  "provider": "Google",
+  "parameter_count": "25.2B",
+  "parameters_raw": 25200000000,
+  "min_ram_gb": 14.4,
+  "recommended_ram_gb": 18.0,
+  "min_vram_gb": 14.4,
+  "quantization": "QAT-INT4",
+  "context_length": 262144,
+  "use_case": "High-throughput, multimodal MoE (3.8B active); official Google QAT int4 GGUF — near-bf16 quality at int4 size, served on llama.cpp with CPU offload",
+  "is_moe": true,
+  "num_experts": null,
+  "active_experts": null,
+  "active_parameters": 3800000000,
+  "architecture": "gemma4",
+  "pipeline_tag": "image-text-to-text",
+  "release_date": "2026-04-01",
+  "gguf_sources": [
+   {
+    "repo": "google/gemma-4-26B-A4B-it-qat-q4_0-gguf",
+    "provider": "Google"
+   }
+  ],
+  "capabilities": [
+   "vision"
+  ]
+ },
  {
   "name": "google/gemma-4-31B-it",
   "provider": "Google",
@@ -19144,4 +19276,4 @@
   ],
   "_discovered": true
  }
-]
+]

services/hwfit/fit.py

@@ -9,7 +9,7 @@ from services.hwfit.models import (
 GPU_BANDWIDTH = {
     "5090": 1792, "5080": 960, "5070 ti": 896, "5070": 672, "5060 ti": 448, "5060": 256,
     "4090": 1008, "4080 super": 736, "4080": 717, "4070 ti super": 672, "4070 ti": 504, "4070 super": 504, "4070": 504, "4060 ti": 288, "4060": 272,
-    "3090 ti": 1008, "3090": 936, "3080 ti": 912, "3080": 760, "3070 ti": 608, "3070": 448, "3060 ti": 448, "3060": 360,
+    "3090 ti": 1008, "3090": 936, "3080 ti": 912, "3080": 760, "3070 ti": 608, "3070": 448, "3060 ti": 448, "3060": 360, "3050 ti": 192, "3050": 224,
     "2080 ti": 616, "2080 super": 496, "2080": 448, "2070 super": 448, "2070": 448, "2060 super": 448, "2060": 336,
     "1660 ti": 288, "1660 super": 336, "1660": 192, "1650 super": 192, "1650": 128,
     "h100 sxm": 3350, "h100": 2039, "h200": 4800, "a100 sxm": 2039, "a100": 1555,

services/hwfit/models.py

@@ -12,6 +12,7 @@ QUANT_BPP = {
     "Q4_K_M": 0.58, "Q4_0": 0.58, "Q3_K_M": 0.48, "Q2_K": 0.37,
     "AWQ-4bit": 0.50, "AWQ-8bit": 1.0,
     "GPTQ-Int4": 0.50, "GPTQ-Int8": 1.0,
+    "QAT-INT4": 0.50, "QAT-INT8": 1.0,
     "mlx-4bit": 0.55, "mlx-8bit": 1.0, "mlx-6bit": 0.75,
     # DeepSeek-V4-style mixed: MoE experts in FP4 (bulk), attention + non-
     # expert dense in FP8, embeddings/LM head in BF16. By weight count the
@@ -30,6 +31,7 @@ QUANT_SPEED_MULT = {
     "Q4_K_M": 1.15, "Q4_0": 1.15, "Q3_K_M": 1.25, "Q2_K": 1.35,
     "AWQ-4bit": 1.2, "AWQ-8bit": 0.85,
     "GPTQ-Int4": 1.2, "GPTQ-Int8": 0.85,
+    "QAT-INT4": 1.15, "QAT-INT8": 0.85,
     "mlx-4bit": 1.15, "mlx-8bit": 0.85, "mlx-6bit": 1.0,
     "FP4-MoE-Mixed": 1.10,  # slightly slower than pure FP4 because of mixed-dtype dispatch
     "FP8-Mixed": 0.85,
@@ -47,6 +49,10 @@ QUANT_QUALITY_PENALTY = {
     # penalty so FP8 wins when both fit. AWQ-4bit stays heavier.
     "AWQ": -1.0, "AWQ-4bit": -4.0, "AWQ-8bit": -1.0,
     "GPTQ": -1.0, "GPTQ-Int4": -4.0, "GPTQ-Int8": -1.0,
+    # Quantization-aware training recovers most of the int4 quality loss, so a
+    # QAT-INT4 build lands far closer to bf16 than a post-training Q4/INT4
+    # (Google reports near-bf16 quality). Penalize it lightly, not like Q4_K_M.
+    "QAT-INT4": -1.0, "QAT-INT8": 0.0,
     "mlx-4bit": -4.0, "mlx-8bit": -0.5, "mlx-6bit": -1.5,
     # DeepSeek-V4 mixed: only MoE experts at FP4 (the rest is FP8/BF16),
     # so the realized quality is much closer to FP8 than to pure FP4 —
@@ -63,6 +69,7 @@ QUANT_BYTES_PER_PARAM = {
     "Q4_K_M": 0.5, "Q4_0": 0.5, "Q3_K_M": 0.375, "Q2_K": 0.25,
     "AWQ-4bit": 0.5, "AWQ-8bit": 1.0,
     "GPTQ-Int4": 0.5, "GPTQ-Int8": 1.0,
+    "QAT-INT4": 0.5, "QAT-INT8": 1.0,
     "mlx-4bit": 0.5, "mlx-8bit": 1.0, "mlx-6bit": 0.75,
     "FP4-MoE-Mixed": 0.55,
     "FP8-Mixed": 1.0,
@@ -74,6 +81,7 @@ PREQUANTIZED_PREFIXES = (
     "AWQ-", "GPTQ-", "mlx-", "FP8", "FP4", "NVFP4", "MXFP4", "NF4",
     "INT4", "INT8", "W4A16", "W8A8", "W8A16",
     "FP4-MoE-Mixed", "FP8-Mixed",
+    "QAT-",
 )
 
 

src/app_helpers.py

@@ -1,6 +1,13 @@
 # src/app_helpers.py
-import os
 import base64
+import logging
+import os
+
+from fastapi import HTTPException
+from fastapi.responses import HTMLResponse
+from starlette.requests import Request
+
+logger = logging.getLogger(__name__)
 
 def read_if_exists(path: str) -> str:
     """Read file if it exists, return empty string otherwise."""
@@ -20,6 +27,28 @@ def abs_join(base_dir: str, rel: str) -> str:
     """Join paths and return absolute path."""
     return os.path.abspath(os.path.join(base_dir, rel))
 
+def serve_html_with_nonce(request: Request, file_path: str) -> HTMLResponse:
+    """Read an app-bundled HTML page and inject the CSP nonce into inline <script> tags.
+
+    Callers pass fixed, server-owned template paths (index/login/backgrounds),
+    never a client-supplied path. So any read failure here — a missing file
+    (broken deployment) or a permission/IO error — is a server fault, not a
+    client "not found": map all of them to a logged 500 so a missing core
+    template surfaces in 5xx alerting instead of hiding behind a 404. If a
+    future caller serves a client-influenced path where 404 is correct, branch
+    that at the call site rather than defaulting this shared helper to 404.
+    """
+    try:
+        with open(file_path, "r", encoding="utf-8") as f:
+            html = f.read()
+    except OSError:
+        logger.exception("Failed to read page %s", file_path)
+        raise HTTPException(500, "Internal server error")
+    nonce = getattr(request.state, "csp_nonce", "")
+    html = html.replace("{{CSP_NONCE}}", nonce)
+    return HTMLResponse(html)
+
+
 def inside_base_dir(base_dir: str, path: str) -> bool:
     """Check if path is inside base directory."""
     if not isinstance(base_dir, str) or not isinstance(path, str):

src/llm_core.py

@@ -345,43 +345,102 @@ def _normalize_ollama_url(url: str) -> str:
     return base.rstrip("/") + "/chat"
 
 
-def _ollama_normalize_tool_messages(messages: List[Dict]) -> List[Dict]:
+def _ollama_normalize_messages(messages: List[Dict]) -> List[Dict]:
     """Adapt Odysseus' canonical OpenAI-style messages to native Ollama /api/chat.
 
-    Odysseus carries assistant tool calls in the OpenAI shape, where
+    Two shape mismatches silently break requests:
-    `function.arguments` is a JSON *string*. Native Ollama expects it to be a
+
-    JSON *object*; given the string it fails the whole request with HTTP 400
+    1. Tool calls: Odysseus carries `function.arguments` as a JSON *string*.
-    "Value looks like object, but can't find closing '}' symbol", which aborts
+       Native Ollama expects a JSON *object* and rejects the string form with
-    every follow-up (tool-result) round. Parse the arguments back into an object
+       HTTP 400 ("Value looks like object, but can't find closing '}' symbol"),
-    here, on a shallow copy, leaving non-tool messages untouched. The opaque
+       aborting every follow-up (tool-result) round. Parse the arguments back
-    Gemini `extra_content` (thought_signature) is dropped — it is meaningless to
+       into an object here, on a shallow copy, leaving non-tool messages
-    Ollama and only matters when the conversation is replayed to Gemini.
+       untouched. The opaque Gemini `extra_content` (thought_signature) is
+       dropped — it is meaningless to Ollama and only matters when the
+       conversation is replayed to Gemini.
+
+    2. Images (issue #4723): Odysseus carries multimodal user content as an
+       OpenAI-style list ``[{type: "text", ...}, {type: "image_url",
+       image_url: {url: "data:image/...;base64,XXX"}}, ...]``. Native Ollama
+       does not accept a list for ``content`` — it wants ``content`` as a
+       string plus a separate ``images`` array of raw base64 strings (no
+       ``data:`` prefix). Without this conversion the image blocks pass
+       through untouched, the vision-capable model never sees the picture,
+       and the user gets "I can't see any image" even though the request
+       succeeded.
     """
     out: List[Dict] = []
     for m in messages or []:
-        tcs = m.get("tool_calls") if isinstance(m, dict) else None
+        if not isinstance(m, dict):
-        if not tcs:
             out.append(m)
             continue
-        new_calls = []
+
-        for tc in tcs:
-            fn = tc.get("function") or {}
-            args = fn.get("arguments")
-            if isinstance(args, str):
-                try:
-                    args = json.loads(args) if args.strip() else {}
-                except (json.JSONDecodeError, TypeError):
-                    args = {}
-            call: Dict = {"function": {"name": fn.get("name", ""), "arguments": args or {}}}
-            if tc.get("id"):
-                call["id"] = tc["id"]
-            new_calls.append(call)
         nm = dict(m)
-        nm["tool_calls"] = new_calls
+
+        # 1. Tool-call argument strings -> objects.
+        tcs = nm.get("tool_calls")
+        if tcs:
+            new_calls = []
+            for tc in tcs:
+                fn = tc.get("function") or {}
+                args = fn.get("arguments")
+                if isinstance(args, str):
+                    try:
+                        args = json.loads(args) if args.strip() else {}
+                    except (json.JSONDecodeError, TypeError):
+                        args = {}
+                call: Dict = {"function": {"name": fn.get("name", ""), "arguments": args or {}}}
+                if tc.get("id"):
+                    call["id"] = tc["id"]
+                new_calls.append(call)
+            nm["tool_calls"] = new_calls
+
+        # 2. Multimodal content list -> native content string + images array.
+        content = nm.get("content")
+        if isinstance(content, list):
+            text_parts: List[str] = []
+            images: List[str] = list(nm.get("images") or [])
+            for block in content:
+                if not isinstance(block, dict):
+                    continue
+                btype = block.get("type")
+                if btype == "text":
+                    t = block.get("text")
+                    if t:
+                        text_parts.append(str(t))
+                elif btype == "image_url":
+                    url = (block.get("image_url") or {}).get("url", "")
+                    if not url:
+                        continue
+                    if url.startswith("data:"):
+                        # Strip the ``data:[...];base64,`` prefix — native
+                        # Ollama wants only the base64 bytes.
+                        _, _, b64 = url.partition(",")
+                        if b64:
+                            images.append(b64)
+                    else:
+                        # Native Ollama images[] is base64-only; it does
+                        # not fetch HTTP URLs.  Skip unsupported schemes
+                        # rather than sending a non-base64 string that the
+                        # model silently ignores.
+                        logger.warning(
+                            "Skipping non-data image_url (Ollama images[] "
+                            "requires base64): %s",
+                            url[:80],
+                        )
+            nm["content"] = "\n".join(text_parts).strip()
+            if images:
+                nm["images"] = images
+
         out.append(nm)
     return out
 
 
+# Backward-compatible alias for callers/tests that imported the older name
+# (it only handled tool messages originally — issue #4723 broadened scope).
+_ollama_normalize_tool_messages = _ollama_normalize_messages
+
+
 def _build_ollama_payload(
     model: str,
     messages: List[Dict],
@@ -404,7 +463,7 @@ def _build_ollama_payload(
     """
     payload: Dict = {
         "model": model,
-        "messages": _ollama_normalize_tool_messages(messages),
+        "messages": _ollama_normalize_messages(messages),
         "stream": stream,
     }
     options: Dict = {}
@@ -618,6 +677,8 @@ def _detect_provider(url: str) -> str:
     from src.copilot import is_copilot_base
     if is_copilot_base(url):
         return "copilot"
+    if _host_match(url, "mistral.ai"):
+        return "mistral"
     return "openai"
 
 
@@ -906,10 +967,17 @@ def _anthropic_rejects_temperature(model: str) -> bool:
         return False
     return (int(match.group(1)), int(match.group(2))) >= (4, 7)
 
+# Reasoning effort level sent to Mistral thinking-capable models. Mistral's
+# API accepts "high", "medium", "low", "none" — see
+# https://docs.mistral.ai/capabilities/reasoning/. Override via env var
+# ODYSSEUS_MISTRAL_REASONING_EFFORT (e.g. set to "medium" for cheaper chat).
+_MISTRAL_REASONING_EFFORT = os.getenv("ODYSSEUS_MISTRAL_REASONING_EFFORT", "high")
+
 # Models that support structured thinking — may output </think> without opening tag
 _THINKING_MODEL_PATTERNS = (
     "qwen3", "qwq", "deepseek-r1", "deepseek-reasoner", "minimax",
     "m2-reap", "gemma", "stepfun", "step-3", "step3",
+    "magistral", "mistral-small", "mistral-medium",
 )
 
 def _supports_thinking(model: str) -> bool:
@@ -919,6 +987,38 @@ def _supports_thinking(model: str) -> bool:
     m = model.lower()
     return any(p in m for p in _THINKING_MODEL_PATTERNS)
 
+def _normalize_mistral_content(content):
+    """Mistral returns content as a structured array when reasoning is on:
+        [{"type": "thinking", "thinking": [{"type": "text", "text": "..."}], "closed": true},
+         {"type": "text", "text": "...final answer..."}]
+    Convert to (text, thinking) tuple of plain strings. Pass through strings
+    unchanged so non-Mistral OpenAI-compat endpoints are unaffected.
+    """
+    if isinstance(content, str):
+        return content, ""
+    if not isinstance(content, list):
+        return "", ""
+    text_parts = []
+    thinking_parts = []
+    for block in content:
+        if not isinstance(block, dict):
+            continue
+        btype = block.get("type")
+        if btype == "text":
+            t = block.get("text", "")
+            if t:
+                text_parts.append(t)
+        elif btype == "thinking":
+            inner = block.get("thinking", [])
+            if isinstance(inner, list):
+                for tb in inner:
+                    if isinstance(tb, dict) and tb.get("text"):
+                        thinking_parts.append(tb["text"])
+            elif isinstance(inner, str):
+                thinking_parts.append(inner)
+    return "".join(text_parts), "".join(thinking_parts)
+
+
 def _convert_openai_content_to_anthropic(content):
     """Convert OpenAI multimodal content blocks to Anthropic format.
 
@@ -1441,6 +1541,8 @@ def llm_call(url: str, model: str, messages: List[Dict], temperature: float = LL
         if max_tokens and max_tokens > 0:
             tok_key = "max_completion_tokens" if _uses_max_completion_tokens(model) else "max_tokens"
             payload[tok_key] = max_tokens
+        if provider == "mistral" and _supports_thinking(model):
+            payload["reasoning_effort"] = _MISTRAL_REASONING_EFFORT
     try:
         note_model_activity(target_url, model)
         r = httpx_post_kimi_aware(target_url, h, json=payload, timeout=timeout)
@@ -1456,7 +1558,16 @@ def llm_call(url: str, model: str, messages: List[Dict], temperature: float = LL
             response = _parse_ollama_response(data)
         else:
             msg = data["choices"][0]["message"]
-            response = msg.get("content") or msg.get("reasoning_content") or ""
+            content = msg.get("content")
+            if isinstance(content, list):
+                # Mistral structured content — extract thinking + text
+                text_part, thinking_part = _normalize_mistral_content(content)
+                if thinking_part:
+                    response = thinking_part + "\n\n" + (text_part or "")
+                else:
+                    response = text_part or msg.get("reasoning_content") or ""
+            else:
+                response = content or msg.get("reasoning_content") or ""
         _set_cached_response(cache_key, response)
         return response
     except Exception:
@@ -1638,6 +1749,8 @@ async def llm_call_async(
         # Suppress thinking for qwen3/gemma4 on Ollama /v1 — same as stream_llm.
         if _is_ollama_openai_compat_url(url) and _supports_thinking(model):
             payload["think"] = False
+        if provider == "mistral" and _supports_thinking(model):
+            payload["reasoning_effort"] = _MISTRAL_REASONING_EFFORT
         _apply_local_cache_affinity(payload, url, session_id)
 
     if _is_host_dead(target_url):
@@ -1756,6 +1869,12 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
             payload[tok_key] = max_tokens
         if tools:
             payload["tools"] = tools
+        # Mistral thinking-capable models — send reasoning_effort so Mistral
+        # activates thinking mode and returns structured reasoning_content.
+        # Effort level is configurable via ODYSSEUS_MISTRAL_REASONING_EFFORT
+        # (high / medium / low / none); default "high".
+        if provider == "mistral" and _supports_thinking(model):
+            payload["reasoning_effort"] = _MISTRAL_REASONING_EFFORT
         # For Ollama's OpenAI-compat /v1 endpoint with thinking models (qwen3,
         # gemma4, etc.), suppress thinking so tool calls aren't swallowed inside
         # <think> blocks. Ollama /v1 accepts "think": false as a top-level param.
@@ -2134,9 +2253,17 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
                                         # Text content
                                         # Reasoning tokens (VLLM --reasoning-parser, e.g. Qwen3/DeepSeek-R1, Nemotron). vLLM 0.20.2 / NIM emit the field as `reasoning`; older builds use `reasoning_content`. Some OpenAI-compatible Ollama builds use `thinking`.
                                         reasoning = delta.get("reasoning_content") or delta.get("reasoning") or delta.get("thinking") or ""
+                                        content = delta.get("content") or ""
+                                        # Mistral structured content: content is a list of typed blocks
+                                        # ({"type": "thinking", ...}, {"type": "text", ...}). Split into
+                                        # reasoning + text so thinking streams into the thinking panel.
+                                        if isinstance(content, list):
+                                            text_part, thinking_part = _normalize_mistral_content(content)
+                                            if thinking_part:
+                                                reasoning = (reasoning + thinking_part) if reasoning else thinking_part
+                                            content = text_part
                                         if reasoning:
                                             yield _stream_delta_event(reasoning, thinking=True)
-                                        content = delta.get("content") or ""
                                         if content:
                                             content = re.sub(r"<mm:think(\s+[^>]*)?>", r"<think\1>", content, flags=re.IGNORECASE)
                                             content = re.sub(r"</mm:think>", "</think>", content, flags=re.IGNORECASE)

static/js/calendar.js

@@ -5,6 +5,7 @@
 import uiModule from './ui.js';
 import spinnerModule from './spinner.js';
 import * as Modals from './modalManager.js';
+import { topPortalZ } from './toolWindowZOrder.js';
 import { makeWindowDraggable } from './windowDrag.js';
 import { attachColorPicker } from './colorPicker.js';
 import { bindMenuDismiss } from './escMenuStack.js';
@@ -470,7 +471,7 @@ function _showEventMoreMenu(ev, anchor) {
   dropdown.className = 'cal-event-dropdown';
   let closeMenu = () => dropdown.remove();
   const rect = anchor.getBoundingClientRect();
-  dropdown.style.cssText = `position:fixed;z-index:10001;min-width:180px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:4px;font-size:12px;top:${rect.bottom + 4}px;left:0px;visibility:hidden;`;
+  dropdown.style.cssText = `position:fixed;z-index:${topPortalZ()};min-width:180px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:4px;font-size:12px;top:${rect.bottom + 4}px;left:0px;visibility:hidden;`;
 
   const _item = (icon, label, onClick, danger) => {
     const it = document.createElement('div');

static/js/chatRenderer.js

@@ -407,8 +407,44 @@ function _openVisionEditor(att, userMsgEl) {
 
 // Tool call syntax patterns to strip from displayed text
 const TOOL_CALL_RE = /\[TOOL_CALL\][\s\S]*?\[\/TOOL_CALL\]/gi;
-// Only strip fenced tool-call blocks that look like structured invocations, not regular code examples
+// Strip fenced tool-call blocks that look like structured invocations, not
-const EXEC_FENCE_RE = /```(?:web_search|read_file|write_file|create_document|edit_document|update_document)\s*\n[\s\S]*?```/gi;
+// regular code examples. The tool tags are NOT hard-coded here — they are the
+// backend's authoritative TOOL_TAGS set, fetched once from GET /api/tools and
+// built into EXEC_FENCE_RE at load. TOOL_TAGS (src/agent_tools/__init__.py) is
+// thus the single source: the live-strip list can never drift from the backend
+// or miss a future tool (#3993). bash/python are carved out on purpose — they
+// are languages a user may legitimately have asked the model to show, not tool
+// invocations.
+//
+// Until the fetch resolves, EXEC_FENCE_RE stays null and exec fences aren't
+// stripped — normally a sub-second window before the first stream. If the fetch
+// fails it stays null for the rest of the session (logged below), so live exec
+// fences won't be stripped until reload. Either way the backend already strips
+// persisted history (src/tool_parsing.py builds the same regex from TOOL_TAGS),
+// so a reload always renders clean.
+let EXEC_FENCE_RE = null;
+const EXEC_FENCE_NON_TOOL = new Set(['bash', 'python']);
+
+async function loadExecFenceRegex() {
+  try {
+    const res = await fetch('/api/tools', { credentials: 'same-origin' });
+    const data = await res.json();
+    const tags = (data.tools || [])
+      .map((t) => t.id)
+      .filter((id) => id && !EXEC_FENCE_NON_TOOL.has(id));
+    if (tags.length) {
+      EXEC_FENCE_RE = new RegExp(
+        '```(?:' + tags.join('|') + ')\\s*\\n[\\s\\S]*?```', 'gi'
+      );
+    }
+  } catch (err) {
+    // Surface the failure rather than swallowing it: EXEC_FENCE_RE stays null,
+    // so this session won't strip live exec fences until reload (persisted path
+    // stays clean regardless).
+    console.warn('chatRenderer: /api/tools fetch failed; live exec-fence stripping disabled until reload', err);
+  }
+}
+loadExecFenceRegex();
 // XML-style tool calls: <minimax:tool_call>, <tool_call>, <function_call>, bare <invoke>
 const XML_TOOL_CALL_RE = /<(?:[\w]+:)?(?:tool_call|function_call)>[\s\S]*?<\/(?:[\w]+:)?(?:tool_call|function_call)>/gi;
 const XML_INVOKE_RE = /<invoke\s+name=['"][^'"]*['"]>[\s\S]*?<\/invoke>/gi;
@@ -853,7 +889,7 @@ export function roleTimestamp(when) {
  */
 export function stripToolBlocks(text) {
   let cleaned = text.replace(TOOL_CALL_RE, '');
-  cleaned = cleaned.replace(EXEC_FENCE_RE, '');
+  if (EXEC_FENCE_RE) cleaned = cleaned.replace(EXEC_FENCE_RE, '');
   cleaned = cleaned.replace(DSML_TOOL_RE, '');
   cleaned = cleaned.replace(DSML_STRAY_RE, '');
   cleaned = cleaned.replace(XML_TOOL_CALL_RE, '');

static/js/cookbook.js

@@ -34,6 +34,7 @@ import {
 } from './cookbookServe.js';
 
 import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
+import { topPortalZ } from './toolWindowZOrder.js';
 
 const STORAGE_KEY = 'cookbook-presets';
 const LAST_STATE_KEY = 'cookbook-last-state';
@@ -1529,7 +1530,7 @@ async function _fetchDependencies() {
       const minW = 150;
       let left = Math.min(rect.right - minW, window.innerWidth - minW - 8);
       left = Math.max(8, left);
-      dropdown.style.cssText = `position:fixed;display:block;z-index:10001;top:${rect.bottom + 6}px;left:${left}px;right:auto;min-width:${minW}px;max-width:calc(100vw - 16px);background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:10px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:6px;font-size:11px;`;
+      dropdown.style.cssText = `position:fixed;display:block;z-index:${topPortalZ()};top:${rect.bottom + 6}px;left:${left}px;right:auto;min-width:${minW}px;max-width:calc(100vw - 16px);background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:10px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:6px;font-size:11px;`;
       const upIco = '<svg width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M21 2v6h-6"/><path d="M3 12a9 9 0 0 1 15-6.7L21 8"/><path d="M3 22v-6h6"/><path d="M21 12a9 9 0 0 1-15 6.7L3 16"/></svg>';
       const it = document.createElement('div');
       it.className = 'dropdown-item-compact';

static/js/cookbookServe.js

@@ -11,6 +11,7 @@ import { modelColor } from './chatRenderer.js';
 import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
 import { openCookbookDependencies } from './cookbook-diagnosis.js';
 import { _hwfitCache } from './cookbook-hwfit.js';
+import { topPortalZ } from './toolWindowZOrder.js';
 
 // Shared state/functions injected by init()
 let _envState;
@@ -1019,7 +1020,7 @@ function _rerenderCachedModels() {
       cancelDiv.addEventListener('click', () => { closeDropdown(); });
       dropdown.appendChild(cancelDiv);
       const rect = btn.getBoundingClientRect();
-      dropdown.style.cssText = `position:fixed;z-index:10001;visibility:hidden;top:0;right:${window.innerWidth-rect.right}px;background:var(--panel);border:1px solid var(--border);border-radius:8px;padding:4px;box-shadow:0 8px 24px rgba(0,0,0,0.3);font-size:12px;`;
+      dropdown.style.cssText = `position:fixed;z-index:${topPortalZ()};visibility:hidden;top:0;right:${window.innerWidth-rect.right}px;background:var(--panel);border:1px solid var(--border);border-radius:8px;padding:4px;box-shadow:0 8px 24px rgba(0,0,0,0.3);font-size:12px;`;
       document.body.appendChild(dropdown);
       // Clamp into the VISIBLE area (visualViewport, not innerHeight — they differ
       // on mobile under the dynamic toolbar). Flip above the button if there's no
@@ -2166,7 +2167,7 @@ function _rerenderCachedModels() {
         // Cap width/height to the viewport and start hidden — we clamp the final
         // position after mount (below) using the menu's real measured size, so it
         // can't run off-screen on a narrow mobile viewport.
-        dropdown.style.cssText = `position:fixed;display:block;visibility:hidden;z-index:10001;top:0;left:0;right:auto;min-width:${minW}px;max-width:calc(100vw - 16px);max-height:calc(100vh - 24px);overflow-y:auto;box-sizing:border-box;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:10px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:6px;font-size:11px;`;
+        dropdown.style.cssText = `position:fixed;display:block;visibility:hidden;z-index:${topPortalZ()};top:0;left:0;right:auto;min-width:${minW}px;max-width:calc(100vw - 16px);max-height:calc(100vh - 24px);overflow-y:auto;box-sizing:border-box;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:10px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:6px;font-size:11px;`;
 
         if (!modelSlots.length) {
           const empty = document.createElement('div');

static/js/documentLibrary.js

@@ -4,6 +4,7 @@
  * Extracted from document.js to reduce file size.
  */
 
+import { topPortalZ } from './toolWindowZOrder.js';
 import uiModule from './ui.js';
 import sessionModule from './sessions.js';
 import spinnerModule from './spinner.js';
@@ -227,7 +228,7 @@ let _libraryArchivedView = false;   // Documents tab showing archived docs?
     dd.style.right = (window.innerWidth - rect.right) + 'px';
     dd.style.top = (rect.bottom + 2) + 'px';
     dd.style.display = 'block';
-    dd.style.zIndex = '100000';
+    dd.style.zIndex = String(topPortalZ());
     requestAnimationFrame(() => {
       const mr = dd.getBoundingClientRect();
       if (mr.bottom > window.innerHeight - 8) {
@@ -629,7 +630,7 @@ let _libraryArchivedView = false;   // Documents tab showing archived docs?
           const rect = menuBtn.getBoundingClientRect();
           document.body.appendChild(dropdown);
           dropdown.dataset.owner = doc.id;
-          dropdown.style.cssText = 'position:fixed;z-index:10000;min-width:0;width:max-content;padding:4px;background:var(--panel);border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);backdrop-filter:blur(12px);font-size:12px;display:block;';
+          dropdown.style.cssText = `position:fixed;z-index:${topPortalZ()};min-width:0;width:max-content;padding:4px;background:var(--panel);border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);backdrop-filter:blur(12px);font-size:12px;display:block;`;
           dropdown.style.top = (rect.bottom + 4) + 'px';
           dropdown.style.left = 'auto';
           dropdown.style.right = (window.innerWidth - rect.right) + 'px';

static/js/emailLibrary.js

@@ -8,6 +8,7 @@ import { styledConfirm, showToast, emptyStateIcon } from './ui.js';
 import { folderDisplayName, sortedFolders } from './emailInbox.js';
 import settingsModule from './settings.js';
 import * as Modals from './modalManager.js';
+import { topPortalZ } from './toolWindowZOrder.js';
 import { makeWindowDraggable } from './windowDrag.js';
 import {
   _esc, _escLinkify, _extractName, _parseTurnMeta,
@@ -5512,7 +5513,7 @@ function _showReaderMoreMenu(em, card, reader, anchor) {
   dropdown._anchor = anchor;
   anchor.classList.add('reader-more-active');
   const rect = anchor.getBoundingClientRect();
-  dropdown.style.cssText = `position:fixed;z-index:10001;min-width:180px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:4px;font-size:12px;top:${rect.bottom + 4}px;right:${window.innerWidth - rect.right}px;`;
+  dropdown.style.cssText = `position:fixed;z-index:${topPortalZ()};min-width:180px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:4px;font-size:12px;top:${rect.bottom + 4}px;right:${window.innerWidth - rect.right}px;`;
 
   const _icon = (svg) => `<span class="dropdown-icon">${svg}</span>`;
   const _unreadIcon = '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><circle cx="12" cy="12" r="10"/><circle cx="12" cy="12" r="3" fill="currentColor"/></svg>';
@@ -5749,7 +5750,7 @@ function _showCardMenu(em, anchor) {
   const dropdown = document.createElement('div');
   dropdown.className = 'email-card-dropdown';
   const rect = anchor.getBoundingClientRect();
-  dropdown.style.cssText = `position:fixed;z-index:10001;min-width:140px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:4px;font-size:12px;top:${rect.bottom + 4}px;right:${window.innerWidth - rect.right}px;`;
+  dropdown.style.cssText = `position:fixed;z-index:${topPortalZ()};min-width:140px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:4px;font-size:12px;top:${rect.bottom + 4}px;right:${window.innerWidth - rect.right}px;`;
 
   const _icon = (svg) => `<span class="dropdown-icon">${svg}</span>`;
   const _replyIcon = '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><polyline points="9 17 4 12 9 7"/><path d="M20 18v-2a4 4 0 0 0-4-4H4"/></svg>';
@@ -5940,7 +5941,7 @@ function _showBulkActionsMenu(anchor) {
   const dropdown = document.createElement('div');
   dropdown.className = 'email-card-dropdown email-bulk-menu';
   const rect = anchor.getBoundingClientRect();
-  dropdown.style.cssText = `position:fixed;z-index:10001;min-width:160px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:4px;font-size:12px;top:${rect.bottom + 4}px;left:${rect.left}px;`;
+  dropdown.style.cssText = `position:fixed;z-index:${topPortalZ()};min-width:160px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:8px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:4px;font-size:12px;top:${rect.bottom + 4}px;left:${rect.left}px;`;
   const _readIco = '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M22 2 11 13"/><path d="m22 2-7 20-4-9-9-4 20-7z"/></svg>';
   const _unreadIco = '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><circle cx="12" cy="12" r="10"/><circle cx="12" cy="12" r="3" fill="currentColor"/></svg>';
   const _doneIco = '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round"><polyline points="20 6 9 17 4 12"/></svg>';
@@ -6221,7 +6222,7 @@ function _showAiReplyChoice(btn, em, data) {
     `max-height:${window.innerHeight - 16}px`,
     'overflow:auto',
     'box-sizing:border-box',
-    'z-index:10060',
+    `z-index:${topPortalZ()}`,
     'display:flex',
     'gap:6px',
     'padding:6px',

static/js/emojiPicker.js

@@ -8,6 +8,8 @@
  * faces (😂, 👍, 😎) have no text form and are intentionally excluded.
  */
 
+import { topPortalZ } from './toolWindowZOrder.js';
+
 // Each entry: [char, label, svgPath OR svg]
 // SVG icons matching Lucide style (24x24 viewBox, 2 stroke)
 const I = (path) => `<svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">${path}</svg>`;
@@ -158,7 +160,7 @@ function togglePicker(anchor, target) {
   _pickerEl.style.position = 'fixed';
   _pickerEl.style.top = (rect.bottom + 4) + 'px';
   _pickerEl.style.left = rect.left + 'px';
-  _pickerEl.style.zIndex = '10000';
+  _pickerEl.style.zIndex = String(topPortalZ());
 
   requestAnimationFrame(() => {
     const pr = _pickerEl.getBoundingClientRect();

static/js/gallery.js

@@ -7,6 +7,7 @@ import { openEditor, closeEditor, isEditorOpen } from './galleryEditor.js';
 import spinnerModule from './spinner.js';
 import { makeWindowDraggable } from './windowDrag.js';
 import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
+import { topPortalZ } from './toolWindowZOrder.js';
 
 const API_BASE = window.location.origin;
 let _open = false;
@@ -2524,7 +2525,7 @@ export function openGallery() {
     const left = Math.min(rect.left, window.innerWidth - 200);
     // Inline the standard dropdown look so it renders correctly even where the
     // `.dropdown` rule is scoped out (e.g. hover-only media queries on mobile).
-    dropdown.style.cssText = `position:fixed;display:block;z-index:10001;top:${rect.bottom + 6}px;left:${Math.max(8, left)}px;right:auto;min-width:180px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:10px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:6px;font-size:11px;`;
+    dropdown.style.cssText = `position:fixed;display:block;z-index:${topPortalZ()};top:${rect.bottom + 6}px;left:${Math.max(8, left)}px;right:auto;min-width:180px;background:var(--panel,var(--bg));border:1px solid var(--border);border-radius:10px;box-shadow:0 8px 24px rgba(0,0,0,0.3);padding:6px;font-size:11px;`;
     const _favIco = '<svg width="14" height="14" viewBox="0 0 24 24" fill="currentColor"><path d="M12 21s-6.7-4.35-9.33-8.04C.9 10.3 1.4 6.9 4.1 5.6c1.9-.9 4 .03 5 1.7 1-1.67 3.1-2.6 5-1.7 2.7 1.3 3.2 4.7 1.43 7.36C18.7 16.65 12 21 12 21z"/></svg>';
     const _tagIco = '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M20.59 13.41 13.42 20.58a2 2 0 0 1-2.83 0L2 12V2h10l8.59 8.59a2 2 0 0 1 0 2.82z"/><line x1="7" y1="7" x2="7.01" y2="7"/></svg>';
     const _dlIco = '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M21 15v4a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2v-4"/><polyline points="7 10 12 15 17 10"/><line x1="12" y1="15" x2="12" y2="3"/></svg>';

static/js/memory.js

@@ -6,6 +6,7 @@ import sessionModule from './sessions.js';
 import spinnerModule from './spinner.js';
 import { makeWindowDraggable } from './windowDrag.js';
 import { snapModalToZone } from './tileManager.js';
+import { topPortalZ } from './toolWindowZOrder.js';
 
 var escapeHtml = uiModule.esc;
 
@@ -865,7 +866,13 @@ export function renderMemoryList() {
         dropdown.style.top = rect.bottom + 2 + 'px';
         dropdown.style.right = (window.innerWidth - rect.right) + 'px';
         dropdown.style.left = 'auto';
-        dropdown.style.zIndex = '10001';
+        // Portaled to <body>, so it must outrank the Brain modal it belongs to.
+        // Tool modals get a monotonically increasing z-index from modalManager's
+        // bring-to-front counter, which climbs unbounded over a long session —
+        // once it passed the old hardcoded 10001 the menu rendered behind the
+        // panel (#4720). topPortalZ() derives the value from the live tool-window
+        // stack so the menu always sits just above, however high it has climbed.
+        dropdown.style.zIndex = String(topPortalZ());
         dropdown.style.display = 'block';
         document.body.appendChild(dropdown);
         // Keep on-screen (mobile): flip above the button if it overflows the

static/js/notes.js

@@ -10,7 +10,7 @@ import { attachColorPicker } from './colorPicker.js';
 import { makeWindowDraggable } from './windowDrag.js';
 import { snapModalToZone } from './tileManager.js';
 import { applyEdgeDock, clearDockSide } from './modalSnap.js';
-import { topToolWindowZ } from './toolWindowZOrder.js';
+import { topToolWindowZ, topPortalZ } from './toolWindowZOrder.js';
 import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
 
 const API_BASE = window.location.origin;
@@ -4335,7 +4335,7 @@ function _openNoteCornerMenu(btn) {
   const mh = menu.offsetHeight || 96;
   const below = window.innerHeight - r.bottom;
   const top = (below < mh + 8 && r.top > mh + 8) ? (r.top - mh - 4) : (r.bottom + 4);
-  menu.style.cssText += `position:fixed;z-index:11000;top:${Math.round(top)}px;left:${Math.round(left)}px;`;
+  menu.style.cssText += `position:fixed;z-index:${topPortalZ()};top:${Math.round(top)}px;left:${Math.round(left)}px;`;
   const close = bindMenuDismiss(menu, () => { menu.remove(); });
   menu.querySelector('[data-act="copy"]').addEventListener('click', () => { close(); _copyNote(id, btn); });
   menu.querySelector('[data-act="agent"]').addEventListener('click', () => { close(); _agentSolveNote(id); });
@@ -4349,7 +4349,7 @@ function _positionNoteMenu(menu, btn, width = 196) {
   const mh = menu.offsetHeight || 112;
   const below = window.innerHeight - r.bottom;
   const top = (below < mh + 8 && r.top > mh + 8) ? (r.top - mh - 4) : (r.bottom + 4);
-  menu.style.cssText += `position:fixed;z-index:11000;top:${Math.round(top)}px;left:${Math.round(left)}px;min-width:${width}px;`;
+  menu.style.cssText += `position:fixed;z-index:${topPortalZ()};top:${Math.round(top)}px;left:${Math.round(left)}px;min-width:${width}px;`;
   const close = (ev) => {
     if (ev && menu.contains(ev.target)) return;
     menu.remove();

static/js/toolWindowZOrder.js

@@ -27,3 +27,20 @@ export function nextToolWindowZ(options = {}) {
   if (Number.isFinite(currentZ) && currentZ > top) return currentZ;
   return top + 1;
 }
+
+// Dock chips pinned by the minimized-dock drag interactions reach z 10030
+// (free-drag) / 10020 (mobile rest) — see modalManager.js. A body-portaled
+// dropdown has to clear those too, not just the open tool-window stack, so this
+// floor keeps it above a chip even when no modal is currently raised.
+const DOCK_OVERLAY_FLOOR = 10030;
+
+// The z a body-portaled dropdown/menu needs so it always sits just above every
+// open tool window (and the dock chips) right now. Tool modals get a
+// monotonically increasing z from the bring-to-front counter (modalManager),
+// which climbs unbounded over a long session — so the hardcoded `z-index: 10001`
+// these dropdowns historically used eventually rendered them BEHIND their own
+// modal (#4720). Derive the value from the live stack instead, sharing the same
+// single source of truth as nextToolWindowZ().
+export function topPortalZ(options = {}) {
+  return Math.max(topToolWindowZ(options), DOCK_OVERLAY_FLOOR) + 1;
+}

tests/test_hwfit_gemma4_12b.py

@@ -0,0 +1,51 @@
+from services.hwfit.fit import rank_models
+from services.hwfit.models import get_models, is_prequantized
+
+
+def _8gb_vram_system():
+    return {
+        "has_gpu": True,
+        "backend": "cuda",
+        "gpu_name": "NVIDIA GeForce RTX 4060",
+        "gpu_vram_gb": 8.0,
+        "gpu_count": 1,
+        "available_ram_gb": 32.0,
+        "total_ram_gb": 32.0,
+    }
+
+
+def test_gemma4_12b_in_catalog():
+    catalog = {m["name"]: m for m in get_models()}
+    assert "google/gemma-4-12B-it" in catalog, "gemma-4-12B-it missing from catalog"
+
+
+def test_gemma4_12b_has_gguf_source():
+    catalog = {m["name"]: m for m in get_models()}
+    entry = catalog["google/gemma-4-12B-it"]
+    assert entry.get("gguf_sources"), "gemma-4-12B-it has no gguf_sources"
+    repos = [s["repo"] for s in entry["gguf_sources"]]
+    assert "unsloth/gemma-4-12B-it-GGUF" in repos
+
+
+def test_gemma4_12b_rank_models_returns_it_for_8gb_vram():
+    results = rank_models(_8gb_vram_system(), search="gemma-4-12B-it", limit=20)
+    names = [r["name"] for r in results]
+    assert "google/gemma-4-12B-it" in names, "rank_models did not return gemma-4-12B-it for 8 GB VRAM"
+
+
+def test_gemma4_12b_qat_entries_in_catalog():
+    catalog = {m["name"]: m for m in get_models()}
+    assert "google/gemma-4-12B-it-qat-int4" in catalog
+    assert "google/gemma-4-12B-it-qat-int8" in catalog
+
+
+def test_gemma4_12b_qat_entries_are_prequantized():
+    catalog = {m["name"]: m for m in get_models()}
+    assert is_prequantized(catalog["google/gemma-4-12B-it-qat-int4"])
+    assert is_prequantized(catalog["google/gemma-4-12B-it-qat-int8"])
+
+
+def test_gemma4_12b_qat_entries_have_no_gguf():
+    catalog = {m["name"]: m for m in get_models()}
+    assert catalog["google/gemma-4-12B-it-qat-int4"]["gguf_sources"] == []
+    assert catalog["google/gemma-4-12B-it-qat-int8"]["gguf_sources"] == []

tests/test_live_strip_email_tool_fences.py

@@ -0,0 +1,119 @@
+"""Regression test for #3993 — live chat leaves executed tool fences visible.
+
+The backend strips every fenced tool block (``src/tool_parsing.py`` builds its
+regex from the full ``TOOL_TAGS`` set), so a reloaded session renders cleanly.
+The live frontend path uses its own regex, ``EXEC_FENCE_RE`` in
+``static/js/chatRenderer.js``.
+
+Originally that regex came from a hand-maintained subset, so any executable tool
+not in it — and every *future* tool added to ``TOOL_TAGS`` — left its executed
+fence lingering as a raw code block in the live bubble until reload. The fix
+makes ``TOOL_TAGS`` the single source: ``chatRenderer.js`` no longer hard-codes a
+tool list at all. It fetches the backend's authoritative set once from
+``GET /api/tools`` (which serves ``sorted(TOOL_TAGS)``) and builds
+``EXEC_FENCE_RE`` from it at load, minus ``bash``/``python`` (legitimate code
+examples a user may have asked the model to show). There is no second list to
+drift.
+
+``chatRenderer.js`` pulls browser globals and can't be imported under node, so
+the behavioral tests exercise an equivalent Python regex built straight from the
+backend ``TOOL_TAGS`` — the same source the live regex now derives from — and
+source-level guards assert the frontend keeps no hard-coded list.
+"""
+import re
+from pathlib import Path
+
+_SRC = Path("static/js/chatRenderer.js")
+_TOOLS_SRC = Path("src/agent_tools/__init__.py")
+_ROUTES_SRC = Path("routes/model_routes.py")
+
+# Deliberately NOT stripped: legitimate code-example languages, not tool
+# invocations. Must match the carve-out in chatRenderer.js.
+_NON_STRIPPED = {"bash", "python"}
+
+
+def _tool_tags() -> set[str]:
+    """Extract the backend TOOL_TAGS set from src/agent_tools/__init__.py (source-level)."""
+    source = _TOOLS_SRC.read_text(encoding="utf-8")
+    m = re.search(r"TOOL_TAGS\s*=\s*\{(?P<body>.*?)\}", source, re.DOTALL)
+    assert m, "TOOL_TAGS literal not found in src/agent_tools/__init__.py"
+    return set(re.findall(r'"([a-z_]+)"', m.group("body")))
+
+
+def _exec_fence_regex() -> re.Pattern:
+    """Rebuild EXEC_FENCE_RE's behavior from the same source the live regex now
+    derives from: the backend TOOL_TAGS (served via /api/tools) minus bash/python."""
+    tags = _tool_tags() - _NON_STRIPPED
+    assert tags, "TOOL_TAGS is empty"
+    return re.compile(r"```(?:" + "|".join(sorted(tags)) + r")\s*\n[\s\S]*?```", re.IGNORECASE)
+
+
+def test_strips_executed_email_tool_fences():
+    rx = _exec_fence_regex()
+    # The exact shape the reporter observed lingering in the live bubble.
+    text = 'Here are emails\n\n```list_emails\n{"max_results":10}\n```'
+    assert rx.sub("", text).strip() == "Here are emails"
+
+
+def test_strips_every_named_email_tool_fence():
+    rx = _exec_fence_regex()
+    email_tools = [
+        "list_email_accounts", "send_email", "list_emails", "read_email",
+        "reply_to_email", "bulk_email", "archive_email", "delete_email",
+        "mark_email_read",
+    ]
+    for tool in email_tools:
+        fence = f"```{tool}\n{{}}\n```"
+        assert rx.sub("", fence).strip() == "", f"{tool} fence not stripped"
+
+
+def test_preserves_existing_web_search_stripping():
+    rx = _exec_fence_regex()
+    fence = '```web_search\n{"q":"x"}\n```'
+    assert rx.sub("", fence).strip() == ""
+
+
+def test_does_not_strip_bash_or_python_code_examples():
+    """bash/python fences are deliberately excluded — they are legitimate code
+    examples a user may have asked the model to show, not tool invocations."""
+    rx = _exec_fence_regex()
+    for lang in sorted(_NON_STRIPPED):
+        example = f"```{lang}\nls -la\n```"
+        assert rx.sub("", example) == example, f"{lang} example wrongly stripped"
+
+
+def test_frontend_keeps_no_hardcoded_tool_list():
+    """Root-cause guard for #3993: chatRenderer.js must NOT reintroduce a
+    hand-maintained tool list. A hard-coded mirror of TOOL_TAGS silently drifts
+    when a new tool is added — leaving its executed fence in the live bubble
+    until reload. The live regex must instead be built from the backend's
+    authoritative set fetched at runtime."""
+    source = _SRC.read_text(encoding="utf-8")
+    assert "EXEC_TOOL_TAGS" not in source, (
+        "chatRenderer.js reintroduced a hard-coded EXEC_TOOL_TAGS list; the "
+        "live-strip tags must come from GET /api/tools so TOOL_TAGS stays the "
+        "single source (#3993)."
+    )
+    assert "/api/tools" in source, (
+        "chatRenderer.js must fetch the tool set from /api/tools to build "
+        "EXEC_FENCE_RE."
+    )
+    # The bash/python carve-out must survive the move to the runtime list.
+    m = re.search(r"EXEC_FENCE_NON_TOOL\s*=\s*new Set\(\[(?P<body>.*?)\]\)", source, re.DOTALL)
+    assert m, "bash/python carve-out (EXEC_FENCE_NON_TOOL) not found in chatRenderer.js"
+    carve_out = set(re.findall(r"['\"]([a-z_]+)['\"]", m.group("body")))
+    assert carve_out == _NON_STRIPPED, (
+        f"EXEC_FENCE_NON_TOOL must carve out exactly {sorted(_NON_STRIPPED)}, "
+        f"got {sorted(carve_out)}"
+    )
+
+
+def test_api_tools_endpoint_serves_full_tool_tags():
+    """The frontend's single source is GET /api/tools. Guard that the endpoint
+    serves the complete TOOL_TAGS set (sorted) — if it ever served a subset, the
+    live-strip list would silently shrink with no second list to catch it."""
+    source = _ROUTES_SRC.read_text(encoding="utf-8")
+    assert re.search(r"for\s+tag\s+in\s+sorted\(\s*TOOL_TAGS\s*\)", source), (
+        "GET /api/tools must iterate sorted(TOOL_TAGS) so the frontend's "
+        "EXEC_FENCE_RE covers every executable tool (#3993)."
+    )

tests/test_llm_core_mistral_content.py

@@ -0,0 +1,156 @@
+"""Tests for _normalize_mistral_content() — Mistral's structured content parser.
+
+Mistral's chat completions API returns content as a typed array when reasoning
+is enabled, instead of the plain string most OpenAI-compat servers use:
+
+    "content": [
+      {"type": "thinking", "thinking": [{"type": "text", "text": "..."}], "closed": true},
+      {"type": "text", "text": "..."}
+    ]
+
+_normalize_mistral_content() splits that into (text, thinking) plain strings.
+The function is called from three sites:
+  - llm_call (sync, non-streaming response parser)
+  - llm_call_async (async, non-streaming response parser)
+  - stream_llm (streaming delta parser)
+
+These tests pin the contract: string passthrough, the array shape, and the
+edge cases (empty, garbage, missing fields) so a refactor doesn't silently
+drop thinking content or break non-Mistral providers.
+"""
+from src.llm_core import _normalize_mistral_content
+
+
+def test_string_passthrough_returns_text_with_empty_thinking():
+    """Plain string content (the common case) passes through unchanged."""
+    text, thinking = _normalize_mistral_content("hello world")
+    assert text == "hello world"
+    assert thinking == ""
+
+
+def test_empty_string_passthrough():
+    text, thinking = _normalize_mistral_content("")
+    assert text == ""
+    assert thinking == ""
+
+
+def test_array_with_thinking_and_text_blocks():
+    """Mistral's documented format: thinking block + text block."""
+    content = [
+        {
+            "type": "thinking",
+            "thinking": [{"type": "text", "text": "Let me work through this..."}],
+            "closed": True,
+        },
+        {"type": "text", "text": "The answer is 42."},
+    ]
+    text, thinking = _normalize_mistral_content(content)
+    assert text == "The answer is 42."
+    assert thinking == "Let me work through this..."
+
+
+def test_array_with_only_thinking_block():
+    """Streaming deltas often contain only a thinking fragment (no text block yet)."""
+    content = [
+        {
+            "type": "thinking",
+            "thinking": [{"type": "text", "text": "Okay, let's"}],
+            "closed": True,
+        }
+    ]
+    text, thinking = _normalize_mistral_content(content)
+    assert text == ""
+    assert thinking == "Okay, let's"
+
+
+def test_array_with_only_text_block():
+    """Final answer delta — only the text block, no thinking."""
+    content = [{"type": "text", "text": "Final answer."}]
+    text, thinking = _normalize_mistral_content(content)
+    assert text == "Final answer."
+    assert thinking == ""
+
+
+def test_array_concatenates_multiple_text_blocks():
+    """Multiple text blocks are concatenated in order."""
+    content = [
+        {"type": "text", "text": "part 1 "},
+        {"type": "text", "text": "part 2"},
+    ]
+    text, thinking = _normalize_mistral_content(content)
+    assert text == "part 1 part 2"
+
+
+def test_array_concatenates_multiple_thinking_fragments():
+    """Multiple thinking sub-blocks are concatenated in order."""
+    content = [
+        {
+            "type": "thinking",
+            "thinking": [
+                {"type": "text", "text": "first "},
+                {"type": "text", "text": "second"},
+            ],
+            "closed": True,
+        }
+    ]
+    text, thinking = _normalize_mistral_content(content)
+    assert text == ""
+    assert thinking == "first second"
+
+
+def test_empty_array_returns_empty_strings():
+    text, thinking = _normalize_mistral_content([])
+    assert text == ""
+    assert thinking == ""
+
+
+def test_array_with_garbage_entries_skips_them():
+    """Non-dict entries, missing type, missing text — all silently skipped."""
+    content = [
+        "not a dict",
+        None,
+        {"type": "unknown_type", "text": "should be ignored"},
+        {"type": "text"},  # missing text key
+        {"type": "thinking"},  # missing thinking key
+        {"type": "text", "text": "valid text"},
+    ]
+    text, thinking = _normalize_mistral_content(content)
+    assert text == "valid text"
+    assert thinking == ""
+
+
+def test_none_returns_empty_strings():
+    """Defensive: None content (server bug or schema drift) doesn't crash."""
+    text, thinking = _normalize_mistral_content(None)
+    assert text == ""
+    assert thinking == ""
+
+
+def test_int_returns_empty_strings():
+    """Defensive: wrong-typed content doesn't crash."""
+    text, thinking = _normalize_mistral_content(42)
+    assert text == ""
+    assert thinking == ""
+
+
+def test_thinking_block_with_string_inner():
+    """Some Mistral API versions may use a string instead of an array for
+    the inner 'thinking' field. Accept both shapes."""
+    content = [
+        {"type": "thinking", "thinking": "inline string thinking"},
+        {"type": "text", "text": "answer"},
+    ]
+    text, thinking = _normalize_mistral_content(content)
+    assert text == "answer"
+    assert thinking == "inline string thinking"
+
+
+def test_thinking_block_with_empty_text_field():
+    """Empty text fields don't pollute the output."""
+    content = [
+        {"type": "thinking", "thinking": [{"type": "text", "text": ""}], "closed": True},
+        {"type": "text", "text": ""},
+    ]
+    text, thinking = _normalize_mistral_content(content)
+    assert text == ""
+    assert thinking == ""

tests/test_ollama_multimodal.py

@@ -0,0 +1,109 @@
+"""Regression tests for Ollama-native multimodal image routing (issue #4723).
+
+Odysseus builds user messages in OpenAI style::
+
+    {"role": "user", "content": [
+        {"type": "text", "text": "..."},
+        {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAA"}},
+    ]}
+
+Native Ollama ``/api/chat`` does **not** accept a list for ``content``. It
+expects ``content`` to be a string and images carried separately on
+``images`` (a list of raw base64 strings, no ``data:`` prefix). Without
+this conversion the image block silently never reaches the vision model —
+the model reports "I can't see the image" even though it is vision-capable
+and the request succeeded.
+"""
+from src import llm_core
+
+
+def _multimodal_msg():
+    return {
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "What is in this picture?"},
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}},
+            {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,BBBB"}},
+        ],
+    }
+
+
+def test_ollama_payload_converts_openai_image_blocks_to_native_images_array():
+    payload = llm_core._build_ollama_payload(
+        "gemma4:e4b", [_multimodal_msg()], temperature=0.0, max_tokens=0,
+    )
+    msg = payload["messages"][0]
+    # Content must be a string, not a list — native Ollama rejects lists.
+    assert isinstance(msg["content"], str)
+    assert "What is in this picture?" in msg["content"]
+    # Base64 data extracted into the native images array (no data: prefix).
+    assert msg["images"] == ["AAAA", "BBBB"]
+
+
+def test_ollama_payload_skips_http_image_url():
+    """Non-data-URI image_url values are skipped with a warning because
+    native Ollama images[] accepts base64 only."""
+    msg = {
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "Look"},
+            {"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
+        ],
+    }
+    payload = llm_core._build_ollama_payload("gemma4:e4b", [msg], temperature=0.0, max_tokens=0)
+    out = payload["messages"][0]
+    assert out["content"] == "Look"
+    # HTTP URL is NOT added to images — Ollama cannot fetch it.
+    assert "images" not in out
+
+
+def test_ollama_payload_preserves_native_images_array():
+    """If the caller already used Ollama's native shape, leave it alone."""
+    msg = {
+        "role": "user",
+        "content": "Describe",
+        "images": ["XXXX"],
+    }
+    payload = llm_core._build_ollama_payload("gemma4:e4b", [msg], temperature=0.0, max_tokens=0)
+    out = payload["messages"][0]
+    assert out["content"] == "Describe"
+    assert out["images"] == ["XXXX"]
+
+
+def test_ollama_payload_merges_native_and_openai_images():
+    """A message that carries both native ``images`` and OpenAI ``image_url``
+    blocks (e.g. assembled by different code paths) must produce one combined
+    list rather than drop either half."""
+    msg = {
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "Hi"},
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,OPENAI"}},
+        ],
+        "images": ["NATIVE"],
+    }
+    payload = llm_core._build_ollama_payload("gemma4:e4b", [msg], temperature=0.0, max_tokens=0)
+    out = payload["messages"][0]
+    assert out["content"] == "Hi"
+    assert out["images"] == ["NATIVE", "OPENAI"]
+
+
+def test_ollama_payload_text_only_message_untouched():
+    msgs = [{"role": "user", "content": "hello"}]
+    payload = llm_core._build_ollama_payload("gemma4:e4b", msgs, temperature=0.0, max_tokens=0)
+    assert payload["messages"][0] == {"role": "user", "content": "hello"}
+
+
+def test_ollama_payload_string_content_with_only_image_block():
+    """A message whose content list has only image_url blocks (no text part)
+    still yields a non-empty content string so native Ollama accepts it."""
+    msg = {
+        "role": "user",
+        "content": [
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,QQ=="}},
+        ],
+    }
+    payload = llm_core._build_ollama_payload("gemma4:e4b", [msg], temperature=0.0, max_tokens=0)
+    out = payload["messages"][0]
+    assert isinstance(out["content"], str)
+    assert out["images"] == ["QQ=="]

tests/test_portal_dropdown_z_js.py

@@ -0,0 +1,89 @@
+"""Node-driven regression coverage for body-portaled dropdown z-order.
+
+Tool-modal z climbs unbounded via modalManager's bring-to-front counter, so the
+old hardcoded `z-index: 10001` shared by ~16 body-portaled dropdowns eventually
+rendered them BEHIND their own modal in a long session (#4720). topPortalZ()
+replaces every one of those literals with a value derived from the live
+tool-window stack. These tests pin that it always clears both the modal stack
+and the dock-chip floor, without importing the browser-heavy UI modules.
+"""
+
+import json
+import shutil
+import subprocess
+import textwrap
+from pathlib import Path
+
+import pytest
+
+
+ROOT = Path(__file__).resolve().parents[1]
+HELPER = ROOT / "static" / "js" / "toolWindowZOrder.js"
+pytestmark = pytest.mark.skipif(not shutil.which("node"), reason="node binary not on PATH")
+
+
+def _node_eval(source: str):
+    proc = subprocess.run(
+        ["node", "--input-type=module"],
+        input=source,
+        cwd=ROOT,
+        capture_output=True,
+        text=True,
+        timeout=30,
+    )
+    assert proc.returncode == 0, proc.stderr
+    return json.loads(proc.stdout.strip())
+
+
+def test_portal_z_clears_dock_chip_floor_when_no_modal_is_open():
+    # No tool window raised → topToolWindowZ floors at 250, but a portaled
+    # dropdown must still clear the dock chips pinned up to 10030, so it lands
+    # just above that floor.
+    values = _node_eval(
+        textwrap.dedent(
+            f"""
+            import {{ topPortalZ }} from '{HELPER.as_uri()}';
+            const root = {{ querySelectorAll() {{ return []; }} }};
+            console.log(JSON.stringify({{ z: topPortalZ({{ root, getStyle: () => ({{}}) }}) }}));
+            """
+        )
+    )
+
+    assert values == {"z": 10031}
+
+
+def test_portal_z_sits_above_a_modal_whose_counter_has_climbed_past_10001():
+    # The #4720 scenario: a long session bumped the owning modal's bring-to-front
+    # z to 99999. A hardcoded 10001 dropdown rendered BEHIND it; topPortalZ must
+    # land one above the live modal z.
+    values = _node_eval(
+        textwrap.dedent(
+            f"""
+            import {{ topPortalZ }} from '{HELPER.as_uri()}';
+            const cls = (...names) => ({{ contains: (name) => names.includes(name) }});
+            const modal = {{ id: 'memory-modal', classList: cls(), style: {{ zIndex: '99999' }} }};
+            const root = {{ querySelectorAll() {{ return [modal]; }} }};
+            console.log(JSON.stringify({{ z: topPortalZ({{ root, getStyle: (el) => el.style }}) }}));
+            """
+        )
+    )
+
+    assert values == {"z": 100000}
+
+
+def test_portal_z_uses_chip_floor_when_the_open_modal_sits_below_it():
+    # A modal raised to 5000 is still below the dock-chip floor, so the floor
+    # (10030) wins and the dropdown lands at 10031 — never below a pinned chip.
+    values = _node_eval(
+        textwrap.dedent(
+            f"""
+            import {{ topPortalZ }} from '{HELPER.as_uri()}';
+            const cls = (...names) => ({{ contains: (name) => names.includes(name) }});
+            const modal = {{ id: 'cookbook-modal', classList: cls(), style: {{ zIndex: '5000' }} }};
+            const root = {{ querySelectorAll() {{ return [modal]; }} }};
+            console.log(JSON.stringify({{ z: topPortalZ({{ root, getStyle: (el) => el.style }}) }}));
+            """
+        )
+    )
+
+    assert values == {"z": 10031}

tests/test_serve_html_with_nonce.py

@@ -0,0 +1,52 @@
+"""Behavior tests for src.app_helpers.serve_html_with_nonce.
+
+Every caller of this helper serves a fixed, app-bundled template
+(index/login/backgrounds), never a client-supplied path. So a read failure —
+a missing file (broken deployment) or a permission/IO error — is a server
+fault, not a client "not found", and must surface as a logged 500 rather than
+hiding behind a 404 where 5xx alerting can't see it. These tests lock that
+intent (raised in the PR #4637 review).
+"""
+import types
+
+import pytest
+
+pytest.importorskip("fastapi")
+pytest.importorskip("starlette.responses")
+from fastapi import HTTPException
+
+from src.app_helpers import serve_html_with_nonce
+
+
+def _request_with_nonce(nonce: str = ""):
+    """Minimal stand-in for a Starlette Request: only request.state.csp_nonce is read."""
+    return types.SimpleNamespace(state=types.SimpleNamespace(csp_nonce=nonce))
+
+
+def test_missing_fixed_template_returns_500_not_404(tmp_path):
+    missing = tmp_path / "does_not_exist.html"
+    with pytest.raises(HTTPException) as exc_info:
+        serve_html_with_nonce(_request_with_nonce(), str(missing))
+    assert exc_info.value.status_code == 500
+    # Generic detail — no OS error string or absolute path leaked to the client.
+    assert exc_info.value.detail == "Internal server error"
+
+
+def test_unreadable_template_returns_500(tmp_path):
+    # A directory at the path makes open() raise an OSError subtype
+    # (IsADirectoryError on POSIX, PermissionError on Windows) — same branch.
+    a_dir = tmp_path / "a_dir.html"
+    a_dir.mkdir()
+    with pytest.raises(HTTPException) as exc_info:
+        serve_html_with_nonce(_request_with_nonce(), str(a_dir))
+    assert exc_info.value.status_code == 500
+
+
+def test_readable_template_injects_nonce(tmp_path):
+    page = tmp_path / "page.html"
+    page.write_text('<script nonce="{{CSP_NONCE}}">x</script>', encoding="utf-8")
+    resp = serve_html_with_nonce(_request_with_nonce("nonce-abc"), str(page))
+    assert resp.status_code == 200
+    body = resp.body.decode("utf-8")
+    assert "nonce-abc" in body
+    assert "{{CSP_NONCE}}" not in body