@clawpump/claw-agent 0.1.4 → 0.1.6

package/agent/tools/vision_tools.py

@@ -74,35 +74,36 @@ _VISION_DOWNLOAD_TIMEOUT = _resolve_download_timeout()
 _VISION_MAX_DOWNLOAD_BYTES = 50 * 1024 * 1024
 
 
-def _validate_image_url(url: str) -> bool:
-    """
-    Basic validation of image URL format.
-    
-    Args:
-        url (str): The URL to validate
-        
-    Returns:
-        bool: True if URL appears to be valid, False otherwise
-    """
+def _image_url_shape_ok(url: str) -> bool:
+    """HTTP(S) shape check only (scheme, netloc). No DNS."""
     if not url or not isinstance(url, str):
         return False
-
     # Basic HTTP/HTTPS URL check
     if not url.startswith(("http://", "https://")):
         return False
-
     # Parse to ensure we at least have a network location; still allow URLs
     # without file extensions (e.g. CDN endpoints that redirect to images).
     parsed = urlparse(url)
     if not parsed.netloc:
         return False
+    return True
+
 
+def _validate_image_url(url: str) -> bool:
+    """Validate image URL for sync callers and tests (SSRF via sync DNS check)."""
+    if not _image_url_shape_ok(url):
+        return False
     # Block private/internal addresses to prevent SSRF
     from tools.url_safety import is_safe_url
-    if not is_safe_url(url):
-        return False
+    return is_safe_url(url)
 
-    return True
+
+async def _validate_image_url_async(url: str) -> bool:
+    """Validate remote image URL without blocking the event loop on DNS."""
+    if not _image_url_shape_ok(url):
+        return False
+    from tools.url_safety import async_is_safe_url
+    return await async_is_safe_url(url)
 
 
 def _detect_image_mime_type(image_path: Path) -> Optional[str]:
@@ -181,8 +182,8 @@ async def _download_image(image_url: str, destination: Path, max_retries: int =
         """
         if response.is_redirect and response.next_request:
             redirect_url = str(response.next_request.url)
-            from tools.url_safety import is_safe_url
-            if not is_safe_url(redirect_url):
+            from tools.url_safety import async_is_safe_url
+            if not await async_is_safe_url(redirect_url):
                 raise ValueError(
                     f"Blocked redirect to private/internal address: {redirect_url}"
                 )
@@ -326,6 +327,15 @@ _MAX_BASE64_BYTES = 20 * 1024 * 1024
 # whether we resize proactively or reactively.
 _EMBED_TARGET_BYTES = 4 * 1024 * 1024
 
+# Proactive embed dimension cap (px, longest side).  Anthropic enforces an
+# 8000px per-side ceiling INDEPENDENTLY of the 5 MB byte cap — a tall full-page
+# screenshot can be well under 5 MB yet far over 8000px (e.g. 1200×12000 at
+# 0.06 MB), so the byte-only embed check above lets it slip into immutable
+# history un-resized and the session bricks on a non-retryable 400.  We cap at
+# 7900 (headroom under 8000) so the proactive resize shrinks tall small-byte
+# images before they are embedded.
+_EMBED_MAX_DIMENSION = 7900
+
 # Target size when auto-resizing on API failure (5 MB).  After a provider
 # rejects an image, we downscale to this target and retry once.
 _RESIZE_TARGET_BYTES = 5 * 1024 * 1024
@@ -341,21 +351,59 @@ def _is_image_size_error(error: Exception) -> bool:
     ))
 
 
+def _image_exceeds_dimension(image_path: Path, max_dimension: int) -> bool:
+    """True if the image's longest side exceeds ``max_dimension`` px.
+
+    Anthropic enforces an 8000px per-side cap independently of the 5 MB byte
+    cap, so a tall small-byte screenshot can pass every byte check yet trip a
+    non-retryable 400.  Returns False (don't force a resize) when Pillow is
+    unavailable or the file can't be read as an image — the byte-based checks
+    still apply, and we never want a missing soft dependency to break the
+    embed path.
+    """
+    try:
+        from PIL import Image as _PILImage
+        with _PILImage.open(image_path) as _img:
+            return max(_img.size) > max_dimension
+    except Exception:
+        return False
+
+
 def _resize_image_for_vision(image_path: Path, mime_type: Optional[str] = None,
-                              max_base64_bytes: int = _RESIZE_TARGET_BYTES) -> str:
+                              max_base64_bytes: int = _RESIZE_TARGET_BYTES,
+                              max_dimension: Optional[int] = None) -> str:
     """Convert an image to a base64 data URL, auto-resizing if too large.
 
     Tries Pillow first to progressively downscale oversized images.  If Pillow
     is not installed or resizing still exceeds the limit, falls back to the raw
     bytes and lets the caller handle the size check.
 
+    Args:
+        max_dimension: If set, images whose longest side exceeds this pixel
+            count are forcibly downscaled even if they're under the byte
+            budget.  Anthropic enforces an 8000 px per-side cap independently
+            of the 5 MB byte cap.
+
     Returns the base64 data URL string.
     """
     # Quick file-size estimate: base64 expands by ~4/3, plus data URL header.
     # Skip the expensive full-read + encode if Pillow can resize directly.
     file_size = image_path.stat().st_size
     estimated_b64 = (file_size * 4) // 3 + 100  # ~header overhead
-    if estimated_b64 <= max_base64_bytes:
+    needs_resize_for_bytes = estimated_b64 > max_base64_bytes
+
+    # Check pixel dimensions even if bytes are fine.
+    needs_resize_for_dims = False
+    if max_dimension is not None:
+        try:
+            from PIL import Image as _PILQuick
+            with _PILQuick.open(image_path) as _quick_img:
+                if max(_quick_img.size) > max_dimension:
+                    needs_resize_for_dims = True
+        except Exception:
+            pass  # can't check; Pillow path below will handle or skip
+
+    if not needs_resize_for_bytes and not needs_resize_for_dims:
         # Small enough — just encode directly.
         data_url = _image_to_base64_data_url(image_path, mime_type=mime_type)
         if len(data_url) <= max_base64_bytes:
@@ -368,14 +416,28 @@ def _resize_image_for_vision(image_path: Path, mime_type: Optional[str] = None,
         from PIL import Image
         import io as _io
     except ImportError:
-        logger.info("Pillow not installed — cannot auto-resize oversized image")
-        if data_url is None:
-            data_url = _image_to_base64_data_url(image_path, mime_type=mime_type)
-        return data_url  # caller will raise the size error
+        # Pillow is a lazy-installable soft dependency. Try a best-effort
+        # install (respects security.allow_lazy_installs; no-op if disabled or
+        # offline), then re-import. If it still isn't importable, fall back to
+        # the raw bytes and let the caller raise the size error.
+        try:
+            from tools.lazy_deps import ensure as _ensure_dep
+            # prompt=False: never raise a blocking input() prompt mid-session.
+            # Under the interactive CLI prompt_toolkit owns stdin, so a bare
+            # input() deadlocks the terminal (#40490). The install is already
+            # gated by security.allow_lazy_installs, so reaching here is opt-in.
+            _ensure_dep("tool.vision", prompt=False)
+            from PIL import Image
+            import io as _io
+        except Exception:
+            logger.info("Pillow not installed — cannot auto-resize oversized image")
+            if data_url is None:
+                data_url = _image_to_base64_data_url(image_path, mime_type=mime_type)
+            return data_url  # caller will raise the size error
 
-    logger.info("Image file is %.1f MB (estimated base64 %.1f MB, limit %.1f MB), auto-resizing...",
+    logger.info("Image file is %.1f MB (estimated base64 %.1f MB, limit %.1f MB, max_dimension=%s), auto-resizing...",
                 file_size / (1024 * 1024), estimated_b64 / (1024 * 1024),
-                max_base64_bytes / (1024 * 1024))
+                max_base64_bytes / (1024 * 1024), max_dimension)
 
     mime = mime_type or _determine_mime_type(image_path)
     # Choose output format: JPEG for photos (smaller), PNG for transparency
@@ -393,13 +455,20 @@ def _resize_image_for_vision(image_path: Path, mime_type: Optional[str] = None,
     if pil_format == "JPEG" and img.mode in {"RGBA", "P"}:
         img = img.convert("RGB")
 
-    # Strategy: halve dimensions until base64 fits, up to 4 rounds.
+    # Strategy: halve dimensions until both base64 fits AND pixel dimensions
+    # are within limits, up to 4 rounds.
     # For JPEG, also try reducing quality at each size step.
     # For PNG, quality is irrelevant — only dimension reduction helps.
     quality_steps = (85, 70, 50) if pil_format == "JPEG" else (None,)
     prev_dims = (img.width, img.height)
     candidate = None  # will be set on first loop iteration
 
+    def _dims_ok(w: int, h: int) -> bool:
+        """True if both pixel dimensions are within the limit."""
+        if max_dimension is None:
+            return True
+        return max(w, h) <= max_dimension
+
     for attempt in range(5):
         if attempt > 0:
             # Proportional scaling: halve the longer side and scale the
@@ -430,7 +499,7 @@ def _resize_image_for_vision(image_path: Path, mime_type: Optional[str] = None,
             img.save(buf, **save_kwargs)
             encoded = base64.b64encode(buf.getvalue()).decode("ascii")
             candidate = f"data:{out_mime};base64,{encoded}"
-            if len(candidate) <= max_base64_bytes:
+            if len(candidate) <= max_base64_bytes and _dims_ok(img.width, img.height):
                 logger.info("Auto-resized image fits: %.1f MB (quality=%s, %dx%d)",
                             len(candidate) / (1024 * 1024), q,
                             img.width, img.height)
@@ -476,7 +545,8 @@ def _supports_media_in_tool_results(provider: str, model: str) -> bool:
         results. Older Gemini does NOT.
 
     For unknown / legacy providers we conservatively return False — the
-    caller falls back to the legacy aux-LLM text path.
+    caller falls back to the legacy aux-LLM text path.  The check is relaxed
+    when the provider's ``ProviderProfile`` declares ``supports_vision=True``.
     """
     if not isinstance(provider, str):
         return False
@@ -513,6 +583,17 @@ def _supports_media_in_tool_results(provider: str, model: str) -> bool:
             return True
         return False
 
+    # Check the provider's registered profile for the supports_vision flag.
+    # This covers vision-capable providers like xiaomi, minimax, etc. that
+    # aren't in the hardcoded list above.
+    try:
+        from providers import get_provider_profile
+        profile = get_provider_profile(p)
+        if profile is not None and profile.supports_vision:
+            return True
+    except Exception:
+        pass
+
     # Other vision-capable provider stacks. Conservative default: False.
     # Add explicit entries here as we verify each provider's tool-result
     # multimodal support empirically.
@@ -640,7 +721,7 @@ async def _vision_analyze_native(
         if local_path.is_file():
             temp_image_path = local_path
             should_cleanup = False
-        elif _validate_image_url(image_url):
+        elif await _validate_image_url_async(image_url):
             blocked = check_website_access(image_url)
             if blocked:
                 return tool_error(blocked["message"], success=False)
@@ -669,15 +750,19 @@ async def _vision_analyze_native(
 
         # Proactive embed cap: this image gets baked into conversation
         # history and re-sent on every subsequent turn.  Anthropic rejects
-        # any single base64 image over 5 MB with a 400, and because history
-        # is immutable, an oversized embed permanently wedges the session —
-        # retries can't clear bytes that are already in the request.  Resize
-        # DOWN to the embed target (4 MB, headroom under 5 MB) whenever the
-        # payload exceeds it, not just at the 20 MB hard ceiling.
-        if len(image_data_url) > _EMBED_TARGET_BYTES:
+        # any single base64 image over 5 MB OR over 8000px per side with a
+        # 400, and because history is immutable, an oversized embed
+        # permanently wedges the session — retries can't clear bytes (or
+        # pixels) that are already in the request.  Resize DOWN to the embed
+        # target (4 MB / 7900px, headroom under both ceilings) whenever the
+        # payload exceeds either limit, not just at the 20 MB hard ceiling.
+        _over_bytes = len(image_data_url) > _EMBED_TARGET_BYTES
+        _over_dims = _image_exceeds_dimension(temp_image_path, _EMBED_MAX_DIMENSION)
+        if _over_bytes or _over_dims:
             image_data_url = _resize_image_for_vision(
                 temp_image_path, mime_type=detected_mime_type,
                 max_base64_bytes=_EMBED_TARGET_BYTES,
+                max_dimension=_EMBED_MAX_DIMENSION,
             )
             # If even resizing can't get under the absolute hard ceiling,
             # there's nothing more we can do — reject rather than embed a
@@ -790,7 +875,7 @@ async def vision_analyze_tool(
             logger.info("Using local image file: %s", image_url)
             temp_image_path = local_path
             should_cleanup = False  # Don't delete cached/local files
-        elif _validate_image_url(image_url):
+        elif await _validate_image_url_async(image_url):
             # Remote URL -- download to a temporary location
             blocked = check_website_access(image_url)
             if blocked:
@@ -1185,8 +1270,8 @@ async def _download_video(video_url: str, destination: Path, max_retries: int =
     async def _ssrf_redirect_guard(response):
         if response.is_redirect and response.next_request:
             redirect_url = str(response.next_request.url)
-            from tools.url_safety import is_safe_url
-            if not is_safe_url(redirect_url):
+            from tools.url_safety import async_is_safe_url
+            if not await async_is_safe_url(redirect_url):
                 raise ValueError(
                     f"Blocked redirect to private/internal address: {redirect_url}"
                 )
@@ -1292,7 +1377,7 @@ async def video_analyze_tool(
             logger.info("Using local video file: %s", video_url)
             temp_video_path = local_path
             should_cleanup = False
-        elif _validate_image_url(video_url):
+        elif await _validate_image_url_async(video_url):
             blocked = check_website_access(video_url)
             if blocked:
                 raise PermissionError(blocked["message"])

package/agent/tools/voice_mode.py

@@ -75,6 +75,7 @@ def _termux_api_app_installed() -> bool:
             text=True,
             timeout=5,
             check=False,
+            stdin=subprocess.DEVNULL,
         )
         return "package:com.termux.api" in (result.stdout or "")
     except Exception:
@@ -388,7 +389,7 @@ class TermuxAudioRecorder:
             "-c", str(CHANNELS),
         ]
         try:
-            subprocess.run(command, capture_output=True, text=True, timeout=15, check=True)
+            subprocess.run(command, capture_output=True, text=True, timeout=15, check=True, stdin=subprocess.DEVNULL)
         except subprocess.CalledProcessError as e:
             details = (e.stderr or e.stdout or str(e)).strip()
             raise RuntimeError(f"Termux microphone start failed: {details}") from e
@@ -405,7 +406,7 @@ class TermuxAudioRecorder:
         mic_cmd = _termux_microphone_command()
         if not mic_cmd:
             return
-        subprocess.run([mic_cmd, "-q"], capture_output=True, text=True, timeout=15, check=False)
+        subprocess.run([mic_cmd, "-q"], capture_output=True, text=True, timeout=15, check=False, stdin=subprocess.DEVNULL)
 
     def stop(self) -> Optional[str]:
         with self._lock:
@@ -1095,7 +1096,7 @@ def play_audio_file(file_path: str) -> bool:
         exe = shutil.which(cmd[0])
         if exe:
             try:
-                proc = subprocess.Popen(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+                proc = subprocess.Popen(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, stdin=subprocess.DEVNULL)
                 with _playback_lock:
                     _active_playback = proc
                 proc.wait(timeout=300)

package/agent/tools/web_tools.py

@@ -102,7 +102,7 @@ from tools.tool_backend_helpers import (  # noqa: F401
     nous_tool_gateway_unavailable_message,
     prefers_gateway,
 )
-from tools.url_safety import is_safe_url
+from tools.url_safety import async_is_safe_url, normalize_url_for_request
 import sys
 
 logger = logging.getLogger(__name__)
@@ -110,9 +110,28 @@ logger = logging.getLogger(__name__)
 
 # ─── Backend Selection ────────────────────────────────────────────────────────
 
+def _env_value(name: str) -> str:
+    """Resolve ``name`` via Hermes config-aware env, falling back to process env.
+
+    Mirrors the SearXNG provider's ``_searxng_url()`` so that values set
+    through Hermes' config/.env layer (``hermes config set``, ``hermes tools``)
+    are honored here too — not just raw process-env exports. Without this,
+    a config-only ``SEARXNG_URL`` (or any provider key) leaves the backend
+    auto-detect cascade and ``check_web_api_key()`` blind to it. See #34290.
+    """
+    try:
+        from hermes_cli.config import get_env_value
+
+        val = get_env_value(name)
+    except Exception:
+        val = None
+    if val is None:
+        val = os.getenv(name, "")
+    return (val or "").strip()
+
+
 def _has_env(name: str) -> bool:
-    val = os.getenv(name)
-    return bool(val and val.strip())
+    return bool(_env_value(name))
 
 def _load_web_config() -> dict:
     """Load the ``web:`` section from ~/.hermes/config.yaml."""
@@ -134,15 +153,18 @@ def _get_backend() -> str:
         return configured
 
     # Fallback for manual / legacy config — pick the highest-priority
-    # available backend. Firecrawl also counts as available when the managed
-    # tool gateway is configured for Nous subscribers.
-    # Free-tier backends (searxng / brave-free / ddgs) trail the paid ones so
-    # existing paid setups are unaffected.
+    # available backend. Explicit user credentials (TAVILY_API_KEY etc.)
+    # beat the managed-tool-gateway probe so a deliberate setup is not
+    # pre-empted by a Nous OAuth token whose subscription tier may not
+    # actually grant web-search access (the gateway then fails at runtime
+    # with "no subscription" and the tool returns an error to the agent
+    # without falling back). Free-tier backends trail the paid ones.
     backend_candidates = (
-        ("firecrawl", _has_env("FIRECRAWL_API_KEY") or _has_env("FIRECRAWL_API_URL") or _is_tool_gateway_ready()),
-        ("parallel", _has_env("PARALLEL_API_KEY")),
         ("tavily", _has_env("TAVILY_API_KEY")),
         ("exa", _has_env("EXA_API_KEY")),
+        ("parallel", _has_env("PARALLEL_API_KEY")),
+        ("firecrawl", _has_env("FIRECRAWL_API_KEY") or _has_env("FIRECRAWL_API_URL")),
+        ("firecrawl", _is_tool_gateway_ready()),
         ("searxng", _has_env("SEARXNG_URL")),
         ("brave-free", _has_env("BRAVE_SEARCH_API_KEY")),
         ("ddgs", _ddgs_package_importable()),
@@ -902,17 +924,25 @@ async def web_extract_tool(
     # URL-decode first so percent-encoded secrets (%73k- = sk-) are caught.
     from agent.redact import _PREFIX_RE
     from urllib.parse import unquote
+    normalized_urls: List[str] = []
     for _url in urls:
-        if _PREFIX_RE.search(_url) or _PREFIX_RE.search(unquote(_url)):
+        normalized_url = normalize_url_for_request(_url)
+        if (
+            _PREFIX_RE.search(_url)
+            or _PREFIX_RE.search(unquote(_url))
+            or _PREFIX_RE.search(normalized_url)
+            or _PREFIX_RE.search(unquote(normalized_url))
+        ):
             return json.dumps({
                 "success": False,
                 "error": "Blocked: URL contains what appears to be an API key or token. "
                          "Secrets must not be sent in URLs.",
             })
+        normalized_urls.append(normalized_url)
 
     debug_call_data = {
         "parameters": {
-            "urls": urls,
+            "urls": normalized_urls,
             "format": format,
             "use_llm_processing": use_llm_processing,
             "model": model,
@@ -928,13 +958,13 @@ async def web_extract_tool(
     }
     
     try:
-        logger.info("Extracting content from %d URL(s)", len(urls))
+        logger.info("Extracting content from %d URL(s)", len(normalized_urls))
 
         # ── SSRF protection — filter out private/internal URLs before any backend ──
         safe_urls = []
         ssrf_blocked: List[Dict[str, Any]] = []
-        for url in urls:
-            if not is_safe_url(url):
+        for url in normalized_urls:
+            if not await async_is_safe_url(url):
                 ssrf_blocked.append({
                     "url": url, "title": "", "content": "",
                     "error": "Blocked: URL targets a private or internal network address",
@@ -1196,7 +1226,7 @@ if __name__ == "__main__":
         elif backend == "tavily":
             print("   Using Tavily API (https://tavily.com)")
         elif backend == "searxng":
-            print(f"   Using SearXNG (search only): {os.getenv('SEARXNG_URL', '').strip()}")
+            print(f"   Using SearXNG (search only): {_env_value('SEARXNG_URL')}")
         elif backend == "brave-free":
             print("   Using Brave Search free tier (search only)")
         elif backend == "ddgs":