feat(ai): commandes @/ & skills, analyse d'images et capacites des modeles (#81)

backend/bookslm.py

@@ -5,9 +5,11 @@ redaction, builds a system prompt with file contents, and caches results
 for repeated queries.
 """
 
+import base64
 import hashlib
 import json
 import logging
+import mimetypes
 import os
 import time
 from pathlib import Path
@@ -21,6 +23,10 @@ logger = logging.getLogger("obsigate.bookslm")
 BOOKSLM_MAX_FILES = int(os.getenv("BOOKSLM_MAX_FILES", "200"))
 BOOKSLM_MAX_TOTAL_CHARS = int(os.getenv("BOOKSLM_MAX_TOTAL_CHARS", "200000"))
 BOOKSLM_MAX_FILE_CHARS = int(os.getenv("BOOKSLM_MAX_FILE_CHARS", "30000"))
+# Maximum size of an image sent to a vision model (bytes, before base64).
+BOOKSLM_MAX_IMAGE_BYTES = int(os.getenv("BOOKSLM_MAX_IMAGE_BYTES", "10000000"))
+
+IMAGE_EXTENSIONS = {".png", ".jpg", ".jpeg", ".gif", ".webp", ".bmp", ".avif"}
 
 # ── Cache ──
 _cache: dict[str, dict[str, Any]] = {}
@@ -260,6 +266,122 @@ def collect_files_context(
     }
 
 
+def collect_adhoc_context(
+    vault_path: Path,
+    files: list[str] | None = None,
+    directories: list[str] | None = None,
+    scope: str = "general",
+) -> dict[str, Any]:
+    """Collect an ad-hoc mix of explicit files and directories.
+
+    Backs the assistant ``@`` command: the user can attach extra files and
+    directories to the current context without changing the base mode. Paths
+    outside the vault are ignored by the underlying collectors.
+
+    Args:
+        vault_path: Absolute path to the vault root.
+        files: Relative file paths to include.
+        directories: Relative directory paths to include (recursive).
+        scope: Context label exposed to the UI.
+
+    Returns:
+        Same shape as :func:`collect_directory_context`.
+    """
+    vault_resolved = vault_path.resolve()
+    collected: list[dict[str, Any]] = []
+    seen: set[str] = set()
+    total_chars = 0
+
+    def _add(entries: list[dict[str, Any]]) -> None:
+        nonlocal total_chars
+        for entry in entries:
+            if len(collected) >= BOOKSLM_MAX_FILES or total_chars >= BOOKSLM_MAX_TOTAL_CHARS:
+                return
+            path = entry.get("path")
+            if not path or path in seen:
+                continue
+            seen.add(path)
+            collected.append(entry)
+            total_chars += len(entry.get("content", ""))
+
+    if files:
+        _add(collect_files_context(vault_resolved, files, scope=scope)["files"])
+    for directory in directories or []:
+        _add(collect_directory_context(vault_resolved, directory)["files"])
+
+    return {
+        "files": collected,
+        "total_chars": total_chars,
+        "file_count": len(collected),
+        "directory_tree": "",
+        "max_total_chars": BOOKSLM_MAX_TOTAL_CHARS,
+        "max_files": BOOKSLM_MAX_FILES,
+        "scope": scope,
+    }
+
+
+def merge_contexts(base: dict[str, Any], extra: dict[str, Any]) -> dict[str, Any]:
+    """Merge two context payloads, de-duplicating files by path."""
+    files: list[dict[str, Any]] = []
+    seen: set[str] = set()
+    total_chars = 0
+    for entry in list(base.get("files", [])) + list(extra.get("files", [])):
+        path = entry.get("path")
+        if not path or path in seen:
+            continue
+        if len(files) >= BOOKSLM_MAX_FILES or total_chars >= BOOKSLM_MAX_TOTAL_CHARS:
+            break
+        seen.add(path)
+        files.append(entry)
+        total_chars += len(entry.get("content", ""))
+
+    tree = base.get("directory_tree", "") or extra.get("directory_tree", "")
+    scope = extra.get("scope") or base.get("scope", "general")
+    return {
+        "files": files,
+        "total_chars": total_chars,
+        "file_count": len(files),
+        "directory_tree": tree,
+        "max_total_chars": BOOKSLM_MAX_TOTAL_CHARS,
+        "max_files": BOOKSLM_MAX_FILES,
+        "scope": scope,
+    }
+
+
+def is_image_path(path: str) -> bool:
+    """True when the path has a supported image extension."""
+    return Path(path or "").suffix.lower() in IMAGE_EXTENSIONS
+
+
+def load_vault_image_data_url(vault_path: Path, rel_path: str) -> str | None:
+    """Read a vault image and return it as a ``data:`` URL for vision models.
+
+    Returns ``None`` when the path is outside the vault, missing, too large,
+    or not a supported image.
+    """
+    if not rel_path or not is_image_path(rel_path):
+        return None
+    vault_resolved = vault_path.resolve()
+    try:
+        target = (vault_resolved / rel_path).resolve()
+        target.relative_to(vault_resolved)
+    except (ValueError, OSError):
+        return None
+    if not target.is_file():
+        return None
+    try:
+        if target.stat().st_size > BOOKSLM_MAX_IMAGE_BYTES:
+            logger.warning("Image too large to send: %s", rel_path)
+            return None
+        raw = target.read_bytes()
+    except OSError as exc:
+        logger.warning("Cannot read image %s: %s", rel_path, exc)
+        return None
+    mime = mimetypes.guess_type(str(target))[0] or "image/png"
+    encoded = base64.b64encode(raw).decode("ascii")
+    return f"data:{mime};base64,{encoded}"
+
+
 def empty_context(scope: str = "general") -> dict[str, Any]:
     """Return an empty context payload (used by the General assistant)."""
     return {