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

backend/bookslm_routes.py

@@ -15,12 +15,17 @@ from backend.auth.middleware import check_vault_access, require_auth
 from backend.bookslm import (
     build_general_system_prompt,
     build_system_prompt,
+    collect_adhoc_context,
     collect_directory_context,
     collect_files_context,
     empty_context,
+    load_vault_image_data_url,
+    merge_contexts,
 )
 from backend.indexer import get_vault_data, index
+from backend.model_capabilities import model_supports_vision
 from backend.schemas import BooksLMContextResponse
+from backend.skills import get_skill_prompt
 from backend.tools.api import ToolContext, ToolMode
 
 logger = logging.getLogger("obsigate.bookslm_routes")
@@ -40,6 +45,14 @@ class BooksLMContextRequest(BaseModel):
         default_factory=list,
         description="Relative file paths to use as context (documents mode)",
     )
+    extra_files: list[str] = Field(
+        default_factory=list,
+        description="Ad-hoc files added with the '@' command (any mode)",
+    )
+    extra_directories: list[str] = Field(
+        default_factory=list,
+        description="Ad-hoc directories added with the '@' command (any mode)",
+    )
 
 
 class BooksLMChatRequest(BaseModel):
@@ -50,7 +63,24 @@ class BooksLMChatRequest(BaseModel):
         default_factory=list,
         description="Relative file paths to use as context (documents mode)",
     )
+    extra_files: list[str] = Field(
+        default_factory=list,
+        description="Ad-hoc files added with the '@' command (any mode)",
+    )
+    extra_directories: list[str] = Field(
+        default_factory=list,
+        description="Ad-hoc directories added with the '@' command (any mode)",
+    )
     message: str = Field(description="User message")
+    images: list[dict[str, Any]] = Field(
+        default_factory=list,
+        description="Images for vision models. Each item: {data, mime_type} (pasted) "
+                    "or {path} (vault-relative file).",
+    )
+    skill: str | None = Field(
+        default=None,
+        description="Skill id selected with the '/' command; its prompt is added to the system prompt.",
+    )
     conversation_history: list[dict[str, str]] = Field(
         default_factory=list,
         description="Previous conversation turns [{role, content}]",
@@ -93,15 +123,35 @@ def _resolve_vault_path(vault: str | None, current_user):
     return vault, Path(vault_data["path"])
 
 
-def _build_context(mode: str, vault_path: Path | None, directory: str, context_files: list[str]):
-    """Collect the context payload for the requested mode."""
+def _build_context(
+    mode: str,
+    vault_path: Path | None,
+    directory: str,
+    context_files: list[str],
+    extra_files: list[str] | None = None,
+    extra_directories: list[str] | None = None,
+):
+    """Collect the context payload for the requested mode.
+
+    Ad-hoc files/directories (``@`` command) are merged on top of the base
+    context. In General mode, attaching ad-hoc files promotes the effective
+    scope to ``documents`` so the prompt actually includes their content.
+    """
     if mode == "general":
-        return empty_context("general")
-    if mode == "documents":
+        base = empty_context("general")
+    elif mode == "documents":
         ctx = collect_files_context(vault_path, context_files, scope="documents")  # type: ignore[arg-type]
         # No open document could be read → degrade gracefully to General.
-        return ctx if ctx["file_count"] else empty_context("general")
-    return collect_directory_context(vault_path, directory)  # type: ignore[arg-type]
+        base = ctx if ctx["file_count"] else empty_context("general")
+    else:
+        base = collect_directory_context(vault_path, directory)  # type: ignore[arg-type]
+
+    if (extra_files or extra_directories) and vault_path is not None:
+        adhoc_scope = "documents" if base.get("scope") == "general" else base.get("scope", "directory")
+        adhoc = collect_adhoc_context(vault_path, extra_files, extra_directories, scope=adhoc_scope)
+        if adhoc["file_count"]:
+            base = merge_contexts(base, adhoc)
+    return base
 
 
 def _resolve_system_prompt(req, current_user) -> str:
@@ -114,16 +164,85 @@ def _resolve_system_prompt(req, current_user) -> str:
     if mode != "general":
         _, vault_path = _resolve_vault_path(req.vault, current_user)
 
-    context = _build_context(mode, vault_path, req.directory, req.context_files)
+    context = _build_context(
+        mode,
+        vault_path,
+        req.directory,
+        req.context_files,
+        getattr(req, "extra_files", None),
+        getattr(req, "extra_directories", None),
+    )
     effective_mode = context.get("scope", mode)
 
     if effective_mode == "general":
-        return build_general_system_prompt(list(index.keys()))
-    if effective_mode == "documents":
-        return build_system_prompt(context, scope="documents")
-    if context["file_count"] == 0:
+        prompt = build_general_system_prompt(list(index.keys()))
+    elif effective_mode == "documents":
+        prompt = build_system_prompt(context, scope="documents")
+    elif context["file_count"] == 0:
         raise HTTPException(status_code=404, detail="Aucun fichier markdown trouvé dans ce dossier")
-    return build_system_prompt(context, scope="directory")
+    else:
+        prompt = build_system_prompt(context, scope="directory")
+
+    skill_id = getattr(req, "skill", None)
+    if skill_id:
+        skill_prompt = get_skill_prompt(skill_id, current_user)
+        if skill_prompt:
+            prompt += "\n\n## Skill actif\n" + skill_prompt
+    return prompt
+
+
+def _resolve_optional_vault_path(req, current_user) -> Path | None:
+    """Best-effort vault path resolution (never raises)."""
+    if not getattr(req, "vault", None):
+        return None
+    try:
+        _, vault_path = _resolve_vault_path(req.vault, current_user)
+        return vault_path
+    except HTTPException:
+        return None
+
+
+def _build_user_content(req, vault_path: Path | None):
+    """Build the user message content, attaching images when present.
+
+    Returns a plain string when there is no image, otherwise an OpenAI-style
+    multimodal content array (which ``ai_chat`` adapts for Gemini).
+    """
+    images = getattr(req, "images", None) or []
+    if not images:
+        return req.message
+
+    parts: list[dict[str, Any]] = [{"type": "text", "text": req.message}]
+    for image in images:
+        if not isinstance(image, dict):
+            continue
+        data_url: str | None = None
+        if image.get("data"):
+            mime = image.get("mime_type") or "image/png"
+            data_url = f"data:{mime};base64,{image['data']}"
+        elif image.get("path") and vault_path is not None:
+            data_url = load_vault_image_data_url(vault_path, str(image["path"]))
+        if data_url:
+            parts.append({"type": "image_url", "image_url": {"url": data_url}})
+    return parts
+
+
+def _validate_vision_support(req) -> None:
+    """Reject image requests when the selected model cannot analyse images."""
+    if not (getattr(req, "images", None)):
+        return
+    provider = _resolve_provider_name(req.provider)
+    if not provider:
+        return
+    from backend.ai import PROVIDERS
+
+    model = req.model or PROVIDERS.get(provider, {}).get("model", "")
+    if not model_supports_vision(provider, model):
+        raise HTTPException(
+            status_code=400,
+            detail=f"Le modèle '{model or provider}' ne supporte pas l'analyse d'images. "
+                   "Choisissez un modèle compatible vision.",
+        )
 
 
 def _resolve_provider_name(requested: str | None) -> str | None:
@@ -157,10 +276,15 @@ async def api_bookslm_context(
     """
     mode = _normalize_mode(req.mode)
     if mode == "general":
-        return empty_context("general")
-
-    _, vault_path = _resolve_vault_path(req.vault, current_user)
-    return _build_context(mode, vault_path, req.directory, req.context_files)
+        if not (req.extra_files or req.extra_directories):
+            return empty_context("general")
+        vault_path = _resolve_optional_vault_path(req, current_user)
+    else:
+        _, vault_path = _resolve_vault_path(req.vault, current_user)
+    return _build_context(
+        mode, vault_path, req.directory, req.context_files,
+        req.extra_files, req.extra_directories,
+    )
 
 
 @router.post(
@@ -178,7 +302,9 @@ async def api_bookslm_chat(
     documents or general app knowledge), then streams the provider's answer
     as Server-Sent Events.
     """
+    _validate_vision_support(req)
     system_prompt = _resolve_system_prompt(req, current_user)
+    vault_path = _resolve_optional_vault_path(req, current_user)
 
     messages: list[dict[str, Any]] = [{"role": "system", "content": system_prompt}]
     for turn in req.conversation_history:
@@ -186,7 +312,7 @@ async def api_bookslm_chat(
         content = turn.get("content", "")
         if role in ("user", "assistant") and content:
             messages.append({"role": role, "content": content})
-    messages.append({"role": "user", "content": req.message})
+    messages.append({"role": "user", "content": _build_user_content(req, vault_path)})
 
     async def generate_sse():
         try:
@@ -246,7 +372,9 @@ async def api_bookslm_agent(
     and the conversation snapshot; the client resumes by echoing them back in
     ``confirm`` / ``confirm_messages``.
     """
+    _validate_vision_support(req)
     system_prompt = _resolve_system_prompt(req, current_user)
+    vault_path = _resolve_optional_vault_path(req, current_user)
 
     messages: list[dict] = [{"role": "system", "content": system_prompt}]
     for turn in req.conversation_history:
@@ -254,7 +382,7 @@ async def api_bookslm_agent(
         content = turn.get("content", "")
         if role in ("user", "assistant") and content:
             messages.append({"role": role, "content": content})
-    messages.append({"role": "user", "content": req.message})
+    messages.append({"role": "user", "content": _build_user_content(req, vault_path)})
 
     ctx = ToolContext(user=current_user, mode=ToolMode.IN_APP)