feat(ai): ecosysteme d'outils phase 2 - recherche a cle, cache/retry, Playwright, crawl, Gitea/GitHub, documents (#92)

.env.example

@@ -74,3 +74,19 @@ DEEPSEEK_MODEL=deepseek-chat
 # Chaîne de repli sans clé (DuckDuckGo puis Bing) si SearXNG ne remonte rien
 # OBSIGATE_WEB_FALLBACK=1
 # OBSIGATE_WEB_TIMEOUT=10
+# Fournisseurs à clé (#92), essayés avant SearXNG — injecter via Infisical en prod
+# OBSIGATE_TAVILY_API_KEY=
+# OBSIGATE_BRAVE_API_KEY=
+# OBSIGATE_SERPAPI_API_KEY=
+# OBSIGATE_EXA_API_KEY=
+# Ordre des fournisseurs (sinon : clés présentes puis SearXNG puis replis)
+# OBSIGATE_WEB_PROVIDERS=brave,searxng
+# Réessais réseau (backoff maison) + cache SQLite des résultats web
+# OBSIGATE_WEB_RETRY=1
+# OBSIGATE_WEB_CACHE_TTL=900          # secondes ; 0 = cache désactivé
+# Rendu dynamique (pages SPA) — dépendance optionnelle :
+#   pip install playwright && playwright install chromium
+# ── Assistant IA — sources connectées (Gitea / GitHub) ──
+# OBSIGATE_GITEA_URL=https://git.example.net
+# OBSIGATE_GITEA_TOKEN=
+# OBSIGATE_GITHUB_TOKEN=

CHANGELOG.md

@@ -6,7 +6,7 @@ Format basé sur [Keep a Changelog](https://keepachangelog.com/fr/1.1.0/),
 et [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 > **En cours de développement** : les changements à venir sont listés dans la section
-> [Unreleased](#unreleased). La dernière version livrée est **2.9.1**.
+> [Unreleased](#unreleased). La dernière version livrée est **2.10.0**.
 
 ---
 
@@ -14,6 +14,35 @@ et [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ---
 
+## [2.10.0] — 2026-09-17
+
+### Ajouté
+
+- **#92 — Assistant IA : écosystème d'outils phase 2 (web étendu, sources connectées, documents)** :
+  - **Recherche web à clé** : fournisseurs optionnels essayés avant SearXNG — Tavily, Brave
+    Search, SerpAPI (Google) et Exa (`OBSIGATE_TAVILY_API_KEY`, `OBSIGATE_BRAVE_API_KEY`,
+    `OBSIGATE_SERPAPI_API_KEY`, `OBSIGATE_EXA_API_KEY`), avec ordre personnalisable via
+    `OBSIGATE_WEB_PROVIDERS`.
+  - **Transverse** : réessais réseau avec backoff maison (`OBSIGATE_WEB_RETRY`) et cache SQLite
+  des résultats web avec TTL (`OBSIGATE_WEB_CACHE_TTL`, 900 s par défaut, `OBSIGATE_WEB_CACHE_PATH`).
+  - **Rendu dynamique** : `fetch_url(render=True)` délègue les pages SPA à un worker Playwright
+    isolé (dépendance optionnelle, dégradation propre si non installée).
+  - **Crawl de site** : `crawl_site` (WRITE, confirmation) capture jusqu'à 20 pages d'un même
+    hôte et enregistre un condensé Markdown dans un vault.
+  - **Sources connectées** : Gitea (`OBSIGATE_GITEA_URL`/`OBSIGATE_GITEA_TOKEN`) et GitHub
+    (`OBSIGATE_GITHUB_TOKEN`) — `git_list_repos`, `git_search_issues`, `git_get_file` (READ,
+    rate-limités, audités). Les drives cloud (Google Drive / OneDrive) restent orientés serveur
+    MCP externe (#79), conformément à la feuille de route.
+  - **Production de documents** : `create_xlsx` (openpyxl), `create_docx` (python-docx),
+    `create_csv`, `create_pdf` (reportlab) — outils WRITE avec confirmation et sauvegarde
+    dans le vault (backup avant écrasement).
+  - Chaque outil : libellé `labels.py` + clés i18n `ai.step.*` FR/EN + tests unitaires mockés
+    (httpx). Dépendances ajoutées : `openpyxl`, `python-docx`, `reportlab`.
+    Fichiers : `backend/tools/{webcache,webrender,connected,crawler,documents,web,schemas,labels}.py`,
+    `backend/services/mutations.py`, `tests/test_{web_cache,web_search_providers,webrender,connected_tools,document_tools,crawler}.py`.
+
+---
+
 ## [2.9.1] — 2026-09-17
 
 ### Corrigé

README.fr.md

@@ -4,7 +4,7 @@
 
 **Porte d'entrée web ultra-léger pour vos vaults Obsidian** — Accédez, naviguez et recherchez dans toutes vos notes Obsidian depuis n'importe quel appareil via une interface web moderne et responsive.
 
-[![Version](https://img.shields.io/badge/Version-2.9.1-blue.svg)]()
+[![Version](https://img.shields.io/badge/Version-2.10.0-blue.svg)]()
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 [![Docker](https://img.shields.io/badge/Docker-Ready-blue.svg)](https://www.docker.com/)
 [![Python](https://img.shields.io/badge/Python-3.11+-green.svg)](https://www.python.org/)
@@ -282,6 +282,12 @@ Un compte **admin** connecté voit une icône 🛡️ dans le header : liste, cr
 | `OBSIGATE_WEBHOOK_ALLOW_PRIVATE` | Autoriser les webhooks vers des adresses privées/boucle | `false` |
 | `OBSIGATE_PDF_MAX_SIZE_MB` | Taille max des PDF extraits (text indexation) | `50` |
 | `OBSIGATE_PDF_EXTRACT_TIMEOUT` | Timeout extraction PDF (secondes) | `30` |
+| `OBSIGATE_TAVILY_API_KEY` / `OBSIGATE_BRAVE_API_KEY` / `OBSIGATE_SERPAPI_API_KEY` / `OBSIGATE_EXA_API_KEY` | Fournisseurs de recherche web à clé (essayés avant SearXNG) | — |
+| `OBSIGATE_WEB_PROVIDERS` | Ordre des fournisseurs de recherche (ex. `brave,searxng`) | — |
+| `OBSIGATE_WEB_RETRY` | Réessais réseau des outils web (backoff maison) | `1` |
+| `OBSIGATE_WEB_CACHE_TTL` | Durée du cache SQLite des résultats web (secondes, `0` = off) | `900` |
+| `OBSIGATE_GITEA_URL` / `OBSIGATE_GITEA_TOKEN` | Source connectée Gitea (outil `git_list_repos`…) | — |
+| `OBSIGATE_GITHUB_TOKEN` | Jeton GitHub (outil `git_list_repos`…) | — |
 
 ### Volume pour la persistance
 
@@ -916,8 +922,8 @@ Ce projet est sous licence **MIT** — voir le fichier [LICENSE](LICENSE) pour l
 
 ## 📝 Changelog
 
-Consultez le [CHANGELOG.md](./CHANGELOG.md) pour l'historique complet de toutes les versions (v1.0.0 → v2.9.1).
+Consultez le [CHANGELOG.md](./CHANGELOG.md) pour l'historique complet de toutes les versions (v1.0.0 → v2.10.0).
 
 ---
 
-*Projet : ObsiGate | Version : 2.9.1 | Dernière mise à jour : Juin 2026*
+*Projet : ObsiGate | Version : 2.10.0 | Dernière mise à jour : Juin 2026*

README.md

@@ -2,7 +2,7 @@
 
 **Ultra-light web gateway for your Obsidian vaults** — Access, browse, and search all your Obsidian notes from any device via a modern, responsive web interface.
 
-[![Version](https://img.shields.io/badge/Version-2.9.1-blue.svg)]()
+[![Version](https://img.shields.io/badge/Version-2.10.0-blue.svg)]()
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 [![Docker](https://img.shields.io/badge/Docker-Ready-blue.svg)](https://www.docker.com/)
 [![Python](https://img.shields.io/badge/Python-3.11+-green.svg)](https://www.python.org/)
@@ -320,6 +320,12 @@ When an **admin** account is logged in, a 🛡️ icon appears in the header. Cl
 | `OBSIGATE_WEBHOOK_ALLOW_PRIVATE` | Allow webhooks to private/loopback addresses | `false` |
 | `OBSIGATE_PDF_MAX_SIZE_MB` | Max PDF size for text extraction | `50` |
 | `OBSIGATE_PDF_EXTRACT_TIMEOUT` | PDF extraction timeout (seconds) | `30` |
+| `OBSIGATE_TAVILY_API_KEY` / `OBSIGATE_BRAVE_API_KEY` / `OBSIGATE_SERPAPI_API_KEY` / `OBSIGATE_EXA_API_KEY` | Keyed web-search providers (tried before SearXNG) | — |
+| `OBSIGATE_WEB_PROVIDERS` | Search provider order (e.g. `brave,searxng`) | — |
+| `OBSIGATE_WEB_RETRY` | Web tools network retries (house-made backoff) | `1` |
+| `OBSIGATE_WEB_CACHE_TTL` | SQLite cache TTL for web results (seconds, `0` = off) | `900` |
+| `OBSIGATE_GITEA_URL` / `OBSIGATE_GITEA_TOKEN` | Gitea connected source (`git_list_repos`…) | — |
+| `OBSIGATE_GITHUB_TOKEN` | GitHub token (`git_list_repos`…) | — |
 
 >All these variables are documented in `.env.example`.
 
@@ -1085,8 +1091,8 @@ This project is licensed under the **MIT License** - see the [LICENSE](LICENSE)
 
 ## 📝 Changelog
 
-See [CHANGELOG.md](./CHANGELOG.md) for the complete version history (v1.0.0 → v2.9.1).
+See [CHANGELOG.md](./CHANGELOG.md) for the complete version history (v1.0.0 → v2.10.0).
 
 ---
 
-*Project: ObsiGate | Version: 2.9.1 | Last updated: May 2026*
+*Project: ObsiGate | Version: 2.10.0 | Last updated: May 2026*

VERSION

@@ -1 +1 @@
-2.9.1
+2.10.0

backend/requirements.txt

@@ -20,3 +20,6 @@ psutil>=5.9
 pywebpush>=2.3.0
 mcp==1.9.4
 sse-starlette==2.1.3
+openpyxl>=3.1
+python-docx>=1.1
+reportlab>=4.0

backend/services/mutations.py

@@ -44,7 +44,7 @@ def _ensure_writable(root: Path) -> None:
         raise ServiceError("Vault is read-only", code="read_only", status=403)
 
 
-def _validate_extension(file_path: Path, *, allow_images: bool = False) -> None:
+def _validate_extension(file_path: Path, *, allow_images: bool = False, allow_docs: bool = False) -> None:
     """Reject unsupported file extensions (400)."""
     from backend.indexer import SUPPORTED_EXTENSIONS
 
@@ -53,6 +53,9 @@ def _validate_extension(file_path: Path, *, allow_images: bool = False) -> None:
     if allow_images:
         from backend.attachment_indexer import IMAGE_EXTENSIONS
         allowed = allowed | IMAGE_EXTENSIONS
+    if allow_docs:
+        # Office documents produced by the AI tool layer (#92).
+        allowed = allowed | {".xlsx", ".docx"}
 
     if ext not in allowed and file_path.name.lower() not in ("dockerfile", "makefile"):
         raise ServiceError(
@@ -715,17 +718,20 @@ def save_raw_file(
     content: bytes,
     *,
     overwrite: bool = True,
+    allow_docs: bool = False,
 ) -> dict[str, Any]:
     """Save a binary or text file to a vault (e.g. from upload / drag-and-drop).
 
     Creates parent directories automatically and safely validates the path.
-    Supports supported text extensions, images and Excalidraw files.
+    Supports supported text extensions, images, Excalidraw files and — with
+    ``allow_docs`` — Office documents (.xlsx/.docx) produced by the AI tools.
 
     Args:
         vault_name: Name of the vault.
         path: Vault-relative path.
         content: Raw bytes to write.
         overwrite: When True, replace existing files (with backup).
+        allow_docs: Also accept .xlsx/.docx extensions (AI document tools).
 
     Returns:
         Dict with ``success``, ``vault``, ``path``, and ``size``.
@@ -733,7 +739,7 @@ def save_raw_file(
     root = get_vault_root(vault_name)
     _ensure_writable(root)
     file_path = resolve_safe_path(root, path)
-    _validate_extension(file_path, allow_images=True)
+    _validate_extension(file_path, allow_images=True, allow_docs=allow_docs)
 
     rel_path = _rel(root, file_path)
 

backend/tools/api.py

@@ -9,6 +9,9 @@ Note: ObsiGate uses implicit namespace packages (no tracked ``__init__.py``,
 which ``.gitignore`` excludes via ``_*.py``), hence this explicit facade.
 """
 
+from backend.tools import connected as _connected  # noqa: F401  (registers connected-source tools)
+from backend.tools import crawler as _crawler  # noqa: F401  (registers the site crawler)
+from backend.tools import documents as _documents  # noqa: F401  (registers document tools)
 from backend.tools import service as _service  # noqa: F401  (registers tools)
 from backend.tools import web as _web  # noqa: F401  (registers web tools)
 from backend.tools.context import (

backend/tools/connected.py

@@ -0,0 +1,239 @@
+"""Connected sources — Gitea & GitHub repositories (phase 2 #92).
+
+The assistant can query the source-hosting platforms the project actually
+uses (ObsiGate is hosted on Gitea): repositories, issues/pull requests and
+repository files. Everything is READ-risk, rate-limited through the shared
+registry and audited.
+
+Configuration (environment — injected by Infisical in production, never
+hard-coded):
+
+* ``OBSIGATE_GITEA_URL``    — base URL of the self-hosted instance (e.g.
+  ``https://git.example.net``); the ``gitea`` provider is only available when
+  this variable is set. Admin-controlled, so the SSRF guard does not apply
+  (unlike user-supplied URLs).
+* ``OBSIGATE_GITEA_TOKEN``  — optional personal access token (private repos).
+* ``OBSIGATE_GITHUB_TOKEN`` — optional token (raises the API rate limits and
+  unlocks private repositories).
+
+Cloud drives (Google Drive / OneDrive) deliberately stay out of the core:
+per the documented roadmap they are best served by an *external MCP server*
+(#79) so the OAuth surface remains outside ObsiGate.
+"""
+
+from __future__ import annotations
+
+import base64
+import binascii
+import logging
+import os
+from typing import Any
+
+import httpx
+
+from backend.tools.context import ToolError, ToolRisk
+from backend.tools.registry import tool
+from backend.tools.schemas import GitGetFileInput, GitProviderInput, GitSearchIssuesInput
+
+logger = logging.getLogger("obsigate.tools.connected")
+
+TIMEOUT = 10.0
+USER_AGENT = "ObsiGateAssistant/1.0 (+self-hosted vault AI)"
+MAX_FILE_BYTES = 300_000
+GITHUB_API = "https://api.github.com"
+
+
+def _provider_base(provider: str) -> tuple[str, str]:
+    """Return (base_url, auth_header_value) for the requested provider."""
+    if provider == "gitea":
+        base = os.environ.get("OBSIGATE_GITEA_URL", "").strip().rstrip("/")
+        if not base:
+            raise ToolError(
+                "Source Gitea non configurée (OBSIGATE_GITEA_URL absente).",
+                code="provider_not_configured",
+            )
+        token = os.environ.get("OBSIGATE_GITEA_TOKEN", "").strip()
+        return base, f"token {token}" if token else ""
+    if provider == "github":
+        token = os.environ.get("OBSIGATE_GITHUB_TOKEN", "").strip()
+        return GITHUB_API, f"Bearer {token}" if token else ""
+    raise ToolError(
+        f"Fournisseur inconnu : {provider} ('gitea' ou 'github')",
+        code="invalid_arguments",
+    )
+
+
+def _headers(auth: str) -> dict[str, str]:
+    headers = {"User-Agent": USER_AGENT, "Accept": "application/json"}
+    if auth:
+        headers["Authorization"] = auth
+    return headers
+
+
+def _request(method: str, url: str, auth: str, **kwargs: Any) -> httpx.Response:
+    try:
+        resp = httpx.request(
+            method, url, headers=_headers(auth), timeout=TIMEOUT, follow_redirects=False,
+            **kwargs,
+        )
+    except httpx.HTTPError as e:
+        logger.warning("connected source request failed %s: %s", url, e)
+        raise ToolError(
+            "Source connectée momentanément indisponible.",
+            code="connected_source_unavailable",
+        ) from e
+    if resp.status_code in (401, 403):
+        raise ToolError(
+            "Accès refusé par la source connectée (jeton manquant ou expiré).",
+            code="permission_denied",
+        )
+    if resp.status_code == 404:
+        raise ToolError("Ressource introuvable sur la source connectée.", code="not_found")
+    resp.raise_for_status()
+    return resp
+
+
+def _normalize_repo(item: dict[str, Any]) -> dict[str, Any]:
+    return {
+        "name": item.get("name") or "",
+        "full_name": item.get("full_name") or "",
+        "url": item.get("html_url") or item.get("clone_url") or "",
+        "description": item.get("description") or "",
+        "updated": item.get("updated_at") or "",
+        "private": bool(item.get("private", False)),
+    }
+
+
+@tool(
+    name="git_list_repos",
+    description=(
+        "List repositories on the connected Gitea instance or GitHub account "
+        "(name, url, description, last update). Use when the user asks about "
+        "their code projects."
+    ),
+    input_model=GitProviderInput,
+    risk=ToolRisk.READ,
+)
+def git_list_repos(ctx, params: GitProviderInput) -> dict[str, Any]:
+    """Query the configured source and return normalized repositories."""
+    base, auth = _provider_base(params.provider)
+    if params.provider == "gitea":
+        url = base + "/api/v1/repos/search"
+        query: dict[str, Any] = {"limit": params.limit}
+        if params.repo:
+            query["q"] = params.repo
+        resp = _request("GET", url, auth, params=query)
+        items = resp.json().get("data") or []
+    else:
+        if params.repo:
+            url = GITHUB_API + f"/repos/{params.repo.strip('/')}"
+            items = [_request("GET", url, auth).json()]
+        else:
+            resp = _request(
+                "GET", GITHUB_API + "/user/repos",
+                auth, params={"per_page": params.limit, "sort": "updated"},
+            )
+            items = resp.json()
+    repos = [_normalize_repo(item) for item in items if isinstance(item, dict)]
+    return {"provider": params.provider, "count": len(repos), "repos": repos}
+
+
+@tool(
+    name="git_search_issues",
+    description=(
+        "Search issues and pull requests on the connected Gitea instance or "
+        "GitHub (title/body keywords, optional repository scope, open/closed)."
+    ),
+    input_model=GitSearchIssuesInput,
+    risk=ToolRisk.READ,
+)
+def git_search_issues(ctx, params: GitSearchIssuesInput) -> dict[str, Any]:
+    """Query issues (and PRs) from the configured source."""
+    base, auth = _provider_base(params.provider)
+    state = params.state if params.state in ("open", "closed") else "open"
+    if params.provider == "gitea":
+        if params.repo:
+            url = base + f"/api/v1/repos/{params.repo.strip('/')}/issues"
+            query: dict[str, Any] = {"state": state, "limit": params.limit, "q": params.query}
+            resp = _request("GET", url, auth, params=query)
+            items = resp.json()
+        else:
+            url = base + "/api/v1/repos/issues/search"
+            resp = _request("GET", url, auth, params={
+                "q": params.query, "state": state, "limit": params.limit,
+            })
+            items = resp.json()
+    else:
+        clause = f"{params.query} is:issue is:{state}"
+        if params.repo:
+            clause += f" repo:{params.repo.strip('/')}"
+        resp = _request(
+            "GET", GITHUB_API + "/search/issues", auth,
+            params={"q": clause, "per_page": params.limit},
+        )
+        items = (resp.json().get("items") or [])
+    issues = [
+        {
+            "id": item.get("number") or item.get("id") or "",
+            "title": (item.get("title") or "")[:300],
+            "url": item.get("html_url") or "",
+            "state": item.get("state") or "",
+            "pull_request": bool(item.get("pull_request")),
+        }
+        for item in (items if isinstance(items, list) else [])
+        if isinstance(item, dict)
+    ]
+    return {
+        "provider": params.provider,
+        "query": params.query,
+        "count": len(issues),
+        "issues": issues,
+    }
+
+
+@tool(
+    name="git_get_file",
+    description=(
+        "Read a file's content from a connected Gitea or GitHub repository "
+        "(source code, docs, config). Text/JSON only, size-capped."
+    ),
+    input_model=GitGetFileInput,
+    risk=ToolRisk.READ,
+)
+def git_get_file(ctx, params: GitGetFileInput) -> dict[str, Any]:
+    """Fetch one repository file and return its decoded text content."""
+    base, auth = _provider_base(params.provider)
+    repo = params.repo.strip("/")
+    path = params.path.strip("/")
+    if not repo or not path:
+        raise ToolError(
+            "'repo' (owner/nom) et 'path' sont obligatoires", code="invalid_arguments"
+        )
+    if params.provider == "gitea":
+        url = base + f"/api/v1/repos/{repo}/contents/{path}"
+    else:
+        url = GITHUB_API + f"/repos/{repo}/contents/{path}"
+    if params.ref:
+        url += f"?ref={params.ref}"
+    resp = _request("GET", url, auth)
+    data = resp.json()
+    encoded = data.get("content") or ""
+    if (data.get("encoding") or "") == "base64" and encoded:
+        try:
+            content = base64.b64decode(encoded).decode("utf-8", errors="replace")
+        except (ValueError, binascii.Error) as e:
+            raise ToolError(
+                "Contenu du fichier illisible (encodage inattendu).",
+                code="file_decode_error",
+            ) from e
+    else:
+        content = encoded
+    truncated = len(content) > MAX_FILE_BYTES
+    return {
+        "provider": params.provider,
+        "repo": repo,
+        "path": data.get("path") or path,
+        "size": data.get("size") or len(content),
+        "content": content[:MAX_FILE_BYTES],
+        "truncated": truncated,
+    }

backend/tools/crawler.py

@@ -0,0 +1,197 @@
+"""Multi-page site crawl — ``crawl_site`` (phase 2 #92, WRITE + confirmation).
+
+The assistant can digest a small public site (documentation, docs portal) and
+store a Markdown summary inside a vault: one section per page, title, URL and
+readable text. The crawl is bounded and same-host only:
+
+* max 20 pages (``max_pages``), same hostname, breadth-first from the entry URL;
+* SSRF guard on every URL (scheme + private-address rejection), size caps;
+* no third-party crawler dependency (scrapy deliberately avoided — a bounded
+  httpx BFS keeps the surface small and the runtime predictable; the task is
+  executed as a single background-style tool run instead of a web request
+  pipeline).
+
+Risk is WRITE: the digest is written into a vault, so the two-step
+confirmation applies (Apply card in the UI, propose/apply over MCP).
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+import time
+from typing import Any
+from urllib.parse import urljoin, urlparse
+
+import httpx
+
+from backend.tools.context import ToolContext, ToolError, ToolRisk
+from backend.tools.registry import tool
+from backend.tools.schemas import CrawlSiteInput
+from backend.tools.web import (
+    USER_AGENT,
+    _assert_public_http_url,
+    _html_to_text,
+    _response_text,
+)
+
+logger = logging.getLogger("obsigate.tools.crawler")
+
+MAX_PAGE_BYTES = 800_000
+MAX_TOTAL_BYTES = 6_000_000
+MAX_TEXT_PER_PAGE = 12_000
+PAGE_TIMEOUT = 10.0
+_LINK_RE = re.compile(r'<a[^>]*href="([^"#]+)"', re.IGNORECASE)
+_TITLE_RE = re.compile(r"<title[^>]*>(.*?)</title>", re.IGNORECASE | re.DOTALL)
+
+
+def _same_host(url: str, host: str) -> bool:
+    return (urlparse(url).hostname or "") == host
+
+
+def _extract_links(raw: str, base_url: str) -> list[str]:
+    import html as html_lib
+
+    links: list[str] = []
+    for match in _LINK_RE.finditer(raw):
+        href = html_lib.unescape(match.group(1)).strip()
+        if not href or href.lower().startswith(("javascript:", "mailto:", "tel:")):
+            continue
+        absolute = urljoin(base_url, href)
+        if absolute.lower().endswith((".png", ".jpg", ".jpeg", ".gif", ".svg", ".webp", ".pdf", ".zip")):
+            continue
+        links.append(absolute.split("#", 1)[0])
+    return links
+
+
+def _fetch_page(url: str) -> tuple[str, str]:
+    """Fetch one page (SSRF-guarded, manual redirects) → (title, text)."""
+    current = _assert_public_http_url(url)
+    resp = None
+    for _hop in range(5):
+        resp = httpx.get(
+            current,
+            headers={"User-Agent": USER_AGENT, "Accept": "text/html,*/*"},
+            timeout=PAGE_TIMEOUT,
+            follow_redirects=False,
+        )
+        if resp.status_code in (301, 302, 303, 307, 308):
+            location = resp.headers.get("location") or ""
+            if not location:
+                break
+            current = _assert_public_http_url(str(httpx.URL(current).join(location)))
+            continue
+        break
+    assert resp is not None
+    resp.raise_for_status()
+    ctype = (resp.headers.get("content-type") or "").lower()
+    if "html" not in ctype and "text" not in ctype:
+        raise ToolError(
+            f"Type de contenu non pris en charge: {ctype.split(';')[0] or 'inconnu'}",
+            code="unsupported_content_type",
+        )
+    raw = (resp.content[:MAX_PAGE_BYTES]).decode(resp.encoding or "utf-8", errors="replace")
+    title_match = _TITLE_RE.search(raw)
+    import html as html_lib
+
+    title = html_lib.unescape(title_match.group(1)).strip()[:300] if title_match else ""
+    return title, _html_to_text(raw)[:MAX_TEXT_PER_PAGE]
+
+
+@tool(
+    name="crawl_site",
+    description=(
+        "Crawl a small public site (same-host only, max 20 pages) starting at "
+        "a URL and save a Markdown digest (title, url, readable text per page) "
+        "into a vault. Use to capture an online documentation for offline use."
+    ),
+    input_model=CrawlSiteInput,
+    risk=ToolRisk.WRITE,
+    requires_vault=True,
+)
+def crawl_site(ctx: ToolContext, params: CrawlSiteInput) -> dict[str, Any]:
+    """Bounded BFS crawl; writes the digest file and returns a summary."""
+    from backend.services.errors import ServiceError
+    from backend.services.mutations import save_raw_file
+
+    start = _assert_public_http_url(params.url.strip())
+    host = urlparse(start).hostname or ""
+    if not host:
+        raise ToolError("URL sans hôte", code="invalid_url")
+
+    queue: list[str] = [start]
+    seen: set[str] = {start}
+    pages: list[dict[str, Any]] = []
+    total_bytes = 0
+    failures: list[str] = []
+
+    while queue and len(pages) < params.max_pages and total_bytes < MAX_TOTAL_BYTES:
+        url = queue.pop(0)
+        try:
+            title, text = _fetch_page(url)
+        except ToolError as e:
+            failures.append(url)
+            logger.warning("crawl_site page failed %s: %s", url, e.code)
+            continue
+        except httpx.HTTPError as e:
+            failures.append(url)
+            logger.warning("crawl_site page failed %s: %s", url, e)
+            continue
+        pages.append({"url": url, "title": title, "text": text})
+        total_bytes += len(text)
+        if len(pages) >= params.max_pages:
+            break
+        try:
+            raw_resp = httpx.get(
+                url, headers={"User-Agent": USER_AGENT}, timeout=PAGE_TIMEOUT,
+                follow_redirects=False,
+            )
+            raw = _response_text(raw_resp)
+        except (httpx.HTTPError, ValueError):
+            continue
+        for link in _extract_links(raw, url):
+            if len(pages) + len(queue) >= params.max_pages:
+                break
+            if link in seen or not _same_host(link, host):
+                continue
+            try:
+                _assert_public_http_url(link)
+            except ToolError:
+                continue
+            seen.add(link)
+            queue.append(link)
+
+    if not pages:
+        raise ToolError(
+            "Aucune page n'a pu être récupérée pour ce site.",
+            code="crawl_failed",
+        )
+
+    lines = [
+        f"# Crawl de {host}",
+        "",
+        f"> {len(pages)} page(s) capturée(s) depuis {start} — {time.strftime('%Y-%m-%d %H:%M')}",
+        "",
+    ]
+    for page in pages:
+        lines.append(f"## {page['title'] or page['url']}")
+        lines.append("")
+        lines.append(f"Source : {page['url']}")
+        lines.append("")
+        lines.append(page["text"])
+        lines.append("")
+    digest = "\n".join(lines).encode("utf-8")
+    try:
+        saved = save_raw_file(
+            params.vault, params.path, digest, overwrite=True, allow_docs=False
+        )
+    except ServiceError as e:
+        raise ToolError(e.message, code=e.code, details=e.details) from e
+    return {
+        "url": start,
+        "vault": params.vault,
+        "path": saved.get("path", params.path),
+        "pages": len(pages),
+        "failed": failures[:10],
+        "size": saved.get("size", len(digest)),
+    }

backend/tools/documents.py

@@ -0,0 +1,188 @@
+"""Document-production tools (phase 2 #92) — WRITE, confirmation required.
+
+The assistant can generate real files inside a vault:
+
+* ``create_xlsx`` — spreadsheet (openpyxl);
+* ``create_docx`` — Word document (python-docx);
+* ``create_csv``  — CSV (stdlib);
+* ``create_pdf``  — PDF (reportlab, from markdown-ish content).
+
+Every tool is ``WRITE`` (two-step confirm in the UI / propose-apply over MCP),
+vault-scoped through ``requires_vault`` and saved via the shared mutation
+service (path safety, read-only check, backup on overwrite).
+"""
+
+from __future__ import annotations
+
+import csv as csv_lib
+import io
+import logging
+import re
+from typing import Any
+from xml.sax import saxutils
+
+from backend.services.errors import ServiceError
+from backend.services.mutations import save_raw_file
+from backend.tools.context import ToolContext, ToolError, ToolRisk
+from backend.tools.registry import tool
+from backend.tools.schemas import CsvInput, DocxInput, PdfInput, SpreadsheetInput
+
+logger = logging.getLogger("obsigate.tools.documents")
+
+MAX_PDF_CHARS = 200_000
+MAX_ROWS = 5_000
+
+
+def _save(vault: str, path: str, content: bytes, overwrite: bool) -> dict[str, Any]:
+    """Shared save helper (maps ServiceError to ToolError)."""
+    try:
+        return save_raw_file(vault, path, content, overwrite=overwrite, allow_docs=True)
+    except ServiceError as e:
+        raise ToolError(e.message, code=e.code, details=e.details) from e
+
+
+def _check_rows(rows: list[list[Any]]) -> None:
+    if not rows:
+        raise ToolError("Aucune ligne fournie", code="invalid_arguments")
+    if len(rows) > MAX_ROWS:
+        raise ToolError(
+            f"Trop de lignes ({len(rows)} > {MAX_ROWS})", code="invalid_arguments"
+        )
+
+
+def _check_extension(path: str, expected: str) -> str:
+    """Enforce the document extension; return the normalized path."""
+    path = (path or "").strip()
+    if not path.lower().endswith(expected):
+        raise ToolError(
+            f"Extension attendue : {expected}", code="invalid_arguments"
+        )
+    return path
+
+
+@tool(
+    name="create_xlsx",
+    description=(
+        "Create an .xlsx spreadsheet in a vault from rows of cell values "
+        "(first row = header). Use for tables, budgets, checklists the user "
+        "asked to turn into an Excel file."
+    ),
+    input_model=SpreadsheetInput,
+    risk=ToolRisk.WRITE,
+    requires_vault=True,
+)
+def create_xlsx(ctx: ToolContext, params: SpreadsheetInput) -> dict[str, Any]:
+    """Build the workbook with openpyxl and save it into the vault."""
+    from openpyxl import Workbook
+
+    _check_rows(params.rows)
+    path = _check_extension(params.path, ".xlsx")
+    wb = Workbook()
+    ws = wb.active
+    ws.title = params.sheet_name[:31] or "Feuille1"
+    for row in params.rows:
+        ws.append(list(row))
+    buffer = io.BytesIO()
+    wb.save(buffer)
+    return _save(params.vault, path, buffer.getvalue(), params.overwrite)
+
+
+@tool(
+    name="create_docx",
+    description=(
+        "Create a .docx Word document in a vault from an optional title and "
+        "ordered paragraphs. Use for letters, reports, structured drafts."
+    ),
+    input_model=DocxInput,
+    risk=ToolRisk.WRITE,
+    requires_vault=True,
+)
+def create_docx(ctx: ToolContext, params: DocxInput) -> dict[str, Any]:
+    """Build the document with python-docx and save it into the vault."""
+    from docx import Document
+
+    if not params.paragraphs:
+        raise ToolError("Aucun paragraphe fourni", code="invalid_arguments")
+    path = _check_extension(params.path, ".docx")
+    doc = Document()
+    if params.title.strip():
+        doc.add_heading(params.title.strip(), level=1)
+    for paragraph in params.paragraphs:
+        doc.add_paragraph(paragraph)
+    buffer = io.BytesIO()
+    doc.save(buffer)
+    return _save(params.vault, path, buffer.getvalue(), params.overwrite)
+
+
+@tool(
+    name="create_csv",
+    description=(
+        "Create a .csv file in a vault from rows of cell values (first row = "
+        "header). Use for flat data exports, simple tables."
+    ),
+    input_model=CsvInput,
+    risk=ToolRisk.WRITE,
+    requires_vault=True,
+)
+def create_csv(ctx: ToolContext, params: CsvInput) -> dict[str, Any]:
+    """Serialize the rows and save the CSV into the vault."""
+    _check_rows(params.rows)
+    path = _check_extension(params.path, ".csv")
+    delimiter = params.delimiter if params.delimiter in (",", ";", "\t") else ","
+    buffer = io.StringIO()
+    writer = csv_lib.writer(buffer, delimiter=delimiter, lineterminator="\n")
+    writer.writerows(params.rows)
+    return _save(params.vault, path, buffer.getvalue().encode("utf-8"), params.overwrite)
+
+
+_HEADING_RE = re.compile(r"^(#{1,6})\s+(.*)$")
+
+
+def _markdown_to_flowables(content: str) -> list[tuple[str, str]]:
+    """Split markdown-ish content into (style, text) blocks for reportlab."""
+    blocks: list[tuple[str, str]] = []
+    for raw_line in content.splitlines():
+        line = raw_line.rstrip()
+        if not line.strip():
+            continue
+        heading = _HEADING_RE.match(line)
+        if heading:
+            blocks.append((f"H{min(3, len(heading.group(1)))}", heading.group(2).strip()))
+        else:
+            blocks.append(("P", line.strip()))
+    return blocks
+
+
+@tool(
+    name="create_pdf",
+    description=(
+        "Create a .pdf document in a vault from markdown-ish content "
+        "(# headings + paragraphs). Use for printable deliverables."
+    ),
+    input_model=PdfInput,
+    risk=ToolRisk.WRITE,
+    requires_vault=True,
+)
+def create_pdf(ctx: ToolContext, params: PdfInput) -> dict[str, Any]:
+    """Render the content with reportlab and save the PDF into the vault."""
+    from reportlab.lib.pagesizes import A4
+    from reportlab.lib.styles import getSampleStyleSheet
+    from reportlab.platypus import Paragraph, SimpleDocTemplate, Spacer
+
+    path = _check_extension(params.path, ".pdf")
+    text = params.content[:MAX_PDF_CHARS]
+    styles = getSampleStyleSheet()
+    style_map = {
+        "P": styles["BodyText"],
+        "H1": styles["Heading1"],
+        "H2": styles["Heading2"],
+        "H3": styles["Heading3"],
+    }
+    buffer = io.BytesIO()
+    doc = SimpleDocTemplate(buffer, pagesize=A4, title=params.title[:200])
+    story: list[Any] = [Paragraph(saxutils.escape(params.title[:300]), styles["Title"])]
+    for style, line in _markdown_to_flowables(text):
+        story.append(Spacer(1, 4))
+        story.append(Paragraph(saxutils.escape(line), style_map[style]))
+    doc.build(story)
+    return _save(params.vault, path, buffer.getvalue(), params.overwrite)

backend/tools/labels.py

@@ -47,6 +47,14 @@ _STEP_LABELS: dict[str, tuple[str, str | None]] = {
     "restore_backup": ("backup_restore", "path"),
     "web_search": ("web_search", "query"),
     "fetch_url": ("fetch_url", "url"),
+    "crawl_site": ("crawl", "url"),
+    "git_list_repos": ("git_repos", "provider"),
+    "git_search_issues": ("git_issues", "query"),
+    "git_get_file": ("git_file", "path"),
+    "create_xlsx": ("xlsx_create", "path"),
+    "create_docx": ("docx_create", "path"),
+    "create_csv": ("csv_create", "path"),
+    "create_pdf": ("pdf_create", "path"),
 }
 
 GENERIC_KEY = "generic"

backend/tools/schemas.py

@@ -251,6 +251,90 @@ class FetchUrlInput(BaseModel):
     """Fetch one public web page and return its readable text."""
 
     url: str = Field(..., description="Absolute http(s) URL of a public page")
+    render: bool = Field(
+        False,
+        description="Render JavaScript with the optional Playwright worker (dynamic SPA pages)",
+    )
+
+
+class CrawlSiteInput(BaseModel):
+    """Crawl a small public site (same-host only) and save a digest into a vault."""
+
+    url: str = Field(..., description="Absolute http(s) URL where the crawl starts")
+    vault: str = Field(..., description="Vault name")
+    path: str = Field(..., description="Vault-relative path of the digest file to write (.md)")
+    max_pages: int = Field(5, ge=1, le=20, description="Maximum number of pages to crawl")
+
+
+class GitProviderInput(BaseModel):
+    """Base fields for connected-source tools (Gitea / GitHub)."""
+
+    provider: str = Field(..., description="'gitea' (OBSIGATE_GITEA_URL) or 'github'")
+    repo: str = Field("", description="Optional 'owner/name' repository filter")
+    limit: int = Field(20, ge=1, le=50, description="Maximum number of entries")
+
+
+class GitSearchIssuesInput(BaseModel):
+    """Search issues/pull requests on a connected Gitea or GitHub instance."""
+
+    provider: str = Field(..., description="'gitea' or 'github'")
+    query: str = Field(..., min_length=1, description="Search keywords")
+    repo: str = Field("", description="Optional 'owner/name' scope (empty = instance-wide)")
+    state: str = Field("open", description="'open' or 'closed'")
+    limit: int = Field(10, ge=1, le=20, description="Maximum number of issues")
+
+
+class GitGetFileInput(BaseModel):
+    """Read a file from a connected Gitea or GitHub repository."""
+
+    provider: str = Field(..., description="'gitea' or 'github'")
+    repo: str = Field(..., description="'owner/name' repository")
+    path: str = Field(..., description="Repository-relative file path")
+    ref: str = Field("", description="Optional branch/tag/commit (empty = default branch)")
+
+
+class SpreadsheetInput(BaseModel):
+    """Create an .xlsx spreadsheet in a vault from rows of cells."""
+
+    vault: str = Field(..., description="Vault name")
+    path: str = Field(..., description="Vault-relative path of the file to write (.xlsx)")
+    rows: list[list[str | int | float | bool | None]] = Field(
+        ..., description="Rows of cell values (first row = header)"
+    )
+    sheet_name: str = Field("Feuille1", description="Worksheet name")
+    overwrite: bool = Field(True, description="Replace an existing file (with backup)")
+
+
+class DocxInput(BaseModel):
+    """Create a .docx Word document in a vault from paragraphs."""
+
+    vault: str = Field(..., description="Vault name")
+    path: str = Field(..., description="Vault-relative path of the file to write (.docx)")
+    title: str = Field("", description="Optional document title (heading 1)")
+    paragraphs: list[str] = Field(..., description="Paragraph texts, in order")
+    overwrite: bool = Field(True, description="Replace an existing file (with backup)")
+
+
+class CsvInput(BaseModel):
+    """Create a .csv file in a vault from rows of cells."""
+
+    vault: str = Field(..., description="Vault name")
+    path: str = Field(..., description="Vault-relative path of the file to write (.csv)")
+    rows: list[list[str | int | float | bool | None]] = Field(
+        ..., description="Rows of cell values (first row = header)"
+    )
+    delimiter: str = Field(",", description="Field separator (',' ';' '\\t')")
+    overwrite: bool = Field(True, description="Replace an existing file (with backup)")
+
+
+class PdfInput(BaseModel):
+    """Create a .pdf document in a vault from markdown-ish content."""
+
+    vault: str = Field(..., description="Vault name")
+    path: str = Field(..., description="Vault-relative path of the file to write (.pdf)")
+    title: str = Field("Document", description="Document title")
+    content: str = Field(..., description="Content (headings with #/##, then paragraphs)")
+    overwrite: bool = Field(True, description="Replace an existing file (with backup)")
 
 
 class ToolResult(BaseModel):

backend/tools/web.py

@@ -8,6 +8,16 @@ Phase 1 of the documented web-toolset roadmap:
   answering « je n'ai pas accès à internet ».
 * ``fetch_url``   — retrieve a public web page and return readable text.
 
+Phase 2 (#92) additions:
+
+* keyed providers — Tavily, Brave Search, SerpAPI and Exa are used first when
+  their API key is configured (env, injected by Infisical in production);
+* SQLite cache — search/fetch results are cached with a TTL
+  (:mod:`backend.tools.webcache`);
+* retry with backoff — transient network errors get one extra attempt;
+* dynamic rendering — ``fetch_url(render=True)`` uses an isolated Playwright
+  worker (optional dependency, graceful degradation).
+
 All are READ-risk tools (no confirmation), rate-limited through the shared
 registry, SSRF-guarded (scheme + private-address rejection), and size-capped.
 
@@ -16,6 +26,14 @@ Configuration (environment):
 * ``OBSIGATE_WEB_TIMEOUT``   — seconds, default 10
 * ``OBSIGATE_WEB_FALLBACK``  — ``0``/``false`` disables the keyless HTML
   fallbacks (SearXNG only), default enabled
+* ``OBSIGATE_TAVILY_API_KEY`` / ``OBSIGATE_BRAVE_API_KEY`` /
+  ``OBSIGATE_SERPAPI_API_KEY`` / ``OBSIGATE_EXA_API_KEY`` — optional keyed
+  providers, tried before SearXNG when set
+* ``OBSIGATE_WEB_PROVIDERS`` — optional comma-separated provider order
+  (e.g. ``brave,searxng``); keyed providers without a key are skipped
+* ``OBSIGATE_WEB_RETRY``     — extra attempts for transient network errors
+  (default 1)
+* ``OBSIGATE_WEB_CACHE_TTL`` — cache TTL seconds, ``0`` disables (default 900)
 """
 
 from __future__ import annotations
@@ -28,12 +46,14 @@ import logging
 import os
 import re
 import socket
+import time
 from collections.abc import Callable
 from typing import Any
 from urllib.parse import parse_qs, urlparse
 
 import httpx
 
+from backend.tools import webcache
 from backend.tools.context import ToolError, ToolRisk, ToolScope
 from backend.tools.registry import tool
 from backend.tools.schemas import FetchUrlInput, WebSearchInput
@@ -48,6 +68,7 @@ WEB_FALLBACK_ENABLED = os.environ.get("OBSIGATE_WEB_FALLBACK", "1").strip().lowe
     "no",
     "off",
 }
+WEB_RETRY_ATTEMPTS = int(os.environ.get("OBSIGATE_WEB_RETRY", "1"))
 USER_AGENT = "ObsiGateAssistant/1.0 (+self-hosted vault AI)"
 # Search engines reject non-browser agents on their public HTML endpoints.
 BROWSER_UA = (
@@ -163,6 +184,115 @@ def _result(
     }
 
 
+def _with_retry(call: Callable[[], Any]) -> Any:
+    """Run *call* with one extra attempt on transient network errors.
+
+    House-made backoff (the roadmap's « tenacity ou boucle maison »): DNS
+    blips and rate-limit hiccups are the common failure mode, and a single
+    retry keeps the fallback chain from being consumed too early.
+    """
+    for attempt in range(1 + max(0, WEB_RETRY_ATTEMPTS)):
+        try:
+            return call()
+        except httpx.TransportError:
+            if attempt >= max(0, WEB_RETRY_ATTEMPTS):
+                raise
+            time.sleep(0.2 * (attempt + 1))
+    raise RuntimeError("unreachable")  # pragma: no cover
+
+
+def _env_key(name: str) -> str:
+    """Read an API key from the environment (Infisical-injected in prod)."""
+    return os.environ.get(name, "").strip()
+
+
+def _search_tavily(query: str, params: WebSearchInput) -> tuple[list[dict[str, Any]], list[str]]:
+    """Tavily Search API (agent-oriented results, key required)."""
+    resp = httpx.post(
+        "https://api.tavily.com/search",
+        json={
+            "api_key": _env_key("OBSIGATE_TAVILY_API_KEY"),
+            "query": query,
+            "max_results": params.max_results,
+            "search_depth": "basic",
+            "include_answer": False,
+        },
+        headers={"User-Agent": USER_AGENT},
+        timeout=WEB_TIMEOUT,
+    )
+    resp.raise_for_status()
+    data = resp.json()
+    return [
+        _result(item.get("title") or "", item.get("url") or "", item.get("content") or "")
+        for item in (data.get("results") or [])
+    ], []
+
+
+def _search_brave(query: str, params: WebSearchInput) -> tuple[list[dict[str, Any]], list[str]]:
+    """Brave Search API (key required)."""
+    resp = httpx.get(
+        "https://api.search.brave.com/res/v1/web/search",
+        params={"q": query, "count": params.max_results, "safesearch": "moderate"},
+        headers={
+            "X-Subscription-Id": _env_key("OBSIGATE_BRAVE_API_KEY"),
+            "Accept": "application/json",
+            "User-Agent": USER_AGENT,
+        },
+        timeout=WEB_TIMEOUT,
+    )
+    resp.raise_for_status()
+    data = resp.json()
+    return [
+        _result(item.get("title") or "", item.get("url") or "", item.get("description") or "")
+        for item in ((data.get("web") or {}).get("results") or [])
+    ], []
+
+
+def _search_serpapi(query: str, params: WebSearchInput) -> tuple[list[dict[str, Any]], list[str]]:
+    """SerpAPI (Google SERP, key required)."""
+    resp = httpx.get(
+        "https://serpapi.com/search",
+        params={"q": query, "api_key": _env_key("OBSIGATE_SERPAPI_API_KEY"),
+                "num": params.max_results},
+        headers={"User-Agent": USER_AGENT},
+        timeout=WEB_TIMEOUT,
+    )
+    resp.raise_for_status()
+    data = resp.json()
+    return [
+        _result(item.get("title") or "", item.get("link") or "", item.get("snippet") or "")
+        for item in (data.get("organic_results") or [])
+    ], []
+
+
+def _search_exa(query: str, params: WebSearchInput) -> tuple[list[dict[str, Any]], list[str]]:
+    """Exa neural search (key required)."""
+    resp = httpx.post(
+        "https://api.exa.ai/search",
+        json={"query": query, "numResults": params.max_results},
+        headers={
+            "x-api-key": _env_key("OBSIGATE_EXA_API_KEY"),
+            "User-Agent": USER_AGENT,
+        },
+        timeout=WEB_TIMEOUT,
+    )
+    resp.raise_for_status()
+    data = resp.json()
+    return [
+        _result(item.get("title") or "", item.get("url") or "", (item.get("text") or "")[:600])
+        for item in (data.get("results") or [])
+    ], []
+
+
+# Keyed providers: name -> (implementation, API key env var)
+_KEYED_PROVIDERS: dict[str, tuple[_Provider, str]] = {
+    "tavily": (_search_tavily, "OBSIGATE_TAVILY_API_KEY"),
+    "brave": (_search_brave, "OBSIGATE_BRAVE_API_KEY"),
+    "serpapi": (_search_serpapi, "OBSIGATE_SERPAPI_API_KEY"),
+    "exa": (_search_exa, "OBSIGATE_EXA_API_KEY"),
+}
+
+
 def _search_searxng(
     query: str, params: WebSearchInput
 ) -> tuple[list[dict[str, Any]], list[str]]:
@@ -288,8 +418,22 @@ _Provider = Callable[[str, WebSearchInput], "tuple[list[dict[str, Any]], list[st
 
 
 def _provider_chain() -> list[tuple[str, _Provider]]:
-    """Ordered providers: self-hosted meta-search first, then keyless fallbacks."""
-    chain: list[tuple[str, _Provider]] = [("searxng", _search_searxng)]
+    """Ordered providers: keyed APIs first, then self-hosted, then keyless.
+
+    ``OBSIGATE_WEB_PROVIDERS`` (comma-separated) overrides the default order;
+    unknown names are ignored and keyed providers without their key are skipped.
+    """
+    chain: list[tuple[str, _Provider]] = []
+    configured = [
+        name.strip().lower()
+        for name in os.environ.get("OBSIGATE_WEB_PROVIDERS", "").split(",")
+        if name.strip()
+    ]
+    for name in configured or list(_KEYED_PROVIDERS):
+        entry = _KEYED_PROVIDERS.get(name)
+        if entry and _env_key(entry[1]):
+            chain.append((name, entry[0]))
+    chain.append(("searxng", _search_searxng))
     if WEB_FALLBACK_ENABLED:
         chain.append(("duckduckgo", _search_duckduckgo))
         chain.append(("bing", _search_bing))
@@ -313,6 +457,17 @@ def web_search(ctx, params: WebSearchInput) -> dict[str, Any]:
     if not query:
         raise ToolError("Requête vide", code="invalid_arguments")
 
+    key = webcache.cache_key("search", {
+        "q": query,
+        "max_results": params.max_results,
+        "category": params.category,
+        "language": params.language,
+        "page": params.page,
+    })
+    cached = webcache.cache_get(key)
+    if cached is not None:
+        return {**cached, "cached": True}
+
     attempts: list[str] = []
     unresponsive: list[str] = []
     reachable = False
@@ -320,8 +475,12 @@ def web_search(ctx, params: WebSearchInput) -> dict[str, Any]:
 
     for name, provider in _provider_chain():
         attempts.append(name)
+
+        def _attempt(p: _Provider = provider) -> tuple[list[dict[str, Any]], list[str]]:
+            return p(query, params)
+
         try:
-            results, engines = provider(query, params)
+            results, engines = _with_retry(_attempt)
         except (httpx.HTTPError, ValueError, AttributeError) as e:
             logger.warning("web_search provider %s failed: %s", name, e)
             last_error = e
@@ -338,6 +497,7 @@ def web_search(ctx, params: WebSearchInput) -> dict[str, Any]:
             }
             if unresponsive:
                 payload["unresponsive_engines"] = unresponsive[:8]
+            webcache.cache_set(key, payload)
             return payload
 
     if not reachable:
@@ -378,6 +538,20 @@ def web_search(ctx, params: WebSearchInput) -> dict[str, Any]:
 def fetch_url(ctx, params: FetchUrlInput) -> dict[str, Any]:
     """Retrieve one page, guard against SSRF, and extract its text."""
     url = _assert_public_http_url(params.url.strip())
+    key = webcache.cache_key("fetch", {"url": url, "render": params.render})
+    cached = webcache.cache_get(key)
+    if cached is not None:
+        return {**cached, "cached": True}
+
+    if params.render:
+        # Dynamic pages (SPA/React): delegated to the isolated Playwright
+        # worker; the browser dependency stays optional (graceful error).
+        from backend.tools.webrender import render_page
+
+        payload = render_page(url)
+        webcache.cache_set(key, payload)
+        return payload
+
     try:
         # Follow redirects manually so every hop is re-checked against the
         # private-address SSRF guard (a public page can redirect to 127.0.0.1).
@@ -414,10 +588,12 @@ def fetch_url(ctx, params: FetchUrlInput) -> dict[str, Any]:
     title_match = re.search(r"<title[^>]*>(.*?)</title>", raw, re.IGNORECASE | re.DOTALL)
     title = html_lib.unescape(title_match.group(1)).strip()[:300] if title_match else ""
     text = _html_to_text(raw)[:MAX_TEXT_CHARS]
-    return {
+    payload = {
         "url": str(resp.url),
         "status": resp.status_code,
         "title": title,
         "text": text,
         "truncated": len(raw) > MAX_TEXT_CHARS,
     }
+    webcache.cache_set(key, payload)
+    return payload

backend/tools/webcache.py

@@ -0,0 +1,138 @@
+"""SQLite cache for web tool results (search results, fetched pages).
+
+Phase 2 of the web-toolset roadmap (« Transverse »): repeated web searches and
+page fetches (common in agent loops, where the model re-reads a source) must
+not hammer the providers. Results are cached in a dedicated SQLite table with
+a TTL; the cache is best-effort — any error silently disables it so a broken
+database file never takes the assistant down.
+
+Configuration (environment):
+* ``OBSIGATE_DATA_DIR``        — base data directory (default ``data``)
+* ``OBSIGATE_WEB_CACHE_PATH``  — explicit cache file override
+* ``OBSIGATE_WEB_CACHE_TTL``   — seconds, ``0`` disables the cache (default 900)
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import os
+import sqlite3
+import threading
+import time
+from pathlib import Path
+from typing import Any
+
+logger = logging.getLogger("obsigate.tools.webcache")
+
+DEFAULT_TTL_SECONDS = 900
+_schema_ready = False
+_write_lock = threading.Lock()
+
+
+def ttl_seconds() -> float:
+    """Configured TTL in seconds (``0`` = cache disabled)."""
+    return float(os.environ.get("OBSIGATE_WEB_CACHE_TTL", str(DEFAULT_TTL_SECONDS)))
+
+
+def _cache_path() -> Path:
+    override = os.environ.get("OBSIGATE_WEB_CACHE_PATH", "").strip()
+    if override:
+        return Path(override)
+    return Path(os.environ.get("OBSIGATE_DATA_DIR", "data")) / "web_cache.sqlite3"
+
+
+def _connect() -> sqlite3.Connection:
+    """Open (and lazily create) the cache database."""
+    global _schema_ready
+    path = _cache_path()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    conn = sqlite3.connect(path, timeout=5, check_same_thread=False)
+    if not _schema_ready:
+        conn.execute(
+            "CREATE TABLE IF NOT EXISTS web_cache ("
+            "key TEXT PRIMARY KEY, value TEXT NOT NULL, created REAL NOT NULL)"
+        )
+        conn.commit()
+        _schema_ready = True
+    return conn
+
+
+def cache_key(prefix: str, payload: dict[str, Any]) -> str:
+    """Deterministic cache key from a prefix and the normalized arguments."""
+    raw = json.dumps(payload, ensure_ascii=False, sort_keys=True, default=str)
+    digest = hashlib.sha256(raw.encode("utf-8")).hexdigest()[:32]
+    return f"{prefix}:{digest}"
+
+
+def cache_get(key: str) -> Any | None:
+    """Return the cached payload for *key*, or ``None`` (miss/expiry/disabled)."""
+    if ttl_seconds() <= 0:
+        return None
+    try:
+        conn = _connect()
+        row = conn.execute(
+            "SELECT value, created FROM web_cache WHERE key = ?", (key,)
+        ).fetchone()
+        conn.close()
+    except sqlite3.Error as e:
+        logger.warning("web cache read failed (%s): %s", key, e)
+        return None
+    if row is None:
+        return None
+    value, created = row
+    if time.time() - float(created) > ttl_seconds():
+        return None
+    try:
+        return json.loads(value)
+    except (ValueError, TypeError):
+        return None
+
+
+def cache_set(key: str, value: Any) -> None:
+    """Store *value* under *key* (best effort, never raises)."""
+    if ttl_seconds() <= 0:
+        return
+    try:
+        with _write_lock:
+            conn = _connect()
+            conn.execute(
+                "INSERT INTO web_cache (key, value, created) VALUES (?, ?, ?) "
+                "ON CONFLICT(key) DO UPDATE SET value = excluded.value, created = excluded.created",
+                (key, json.dumps(value, ensure_ascii=False, default=str), time.time()),
+            )
+            conn.commit()
+            conn.close()
+    except sqlite3.Error as e:
+        logger.warning("web cache write failed (%s): %s", key, e)
+
+
+def purge_expired() -> int:
+    """Delete expired rows; return the number of removed entries (maintenance)."""
+    try:
+        conn = _connect()
+        cursor = conn.execute(
+            "DELETE FROM web_cache WHERE created < ?", (time.time() - ttl_seconds(),)
+        )
+        conn.commit()
+        deleted = cursor.rowcount
+        conn.close()
+        return int(deleted)
+    except sqlite3.Error as e:
+        logger.warning("web cache purge failed: %s", e)
+        return 0
+
+
+def clear_cache() -> int:
+    """Drop every cached entry (tests / admin); returns the number of rows."""
+    try:
+        conn = _connect()
+        cursor = conn.execute("DELETE FROM web_cache")
+        conn.commit()
+        deleted = cursor.rowcount
+        conn.close()
+        return int(deleted)
+    except sqlite3.Error as e:
+        logger.warning("web cache clear failed: %s", e)
+        return 0

backend/tools/webrender.py

@@ -0,0 +1,100 @@
+"""Dynamic page rendering (Playwright) — ``fetch_url(render=True)``.
+
+Static pages are fetched with httpx inside :mod:`backend.tools.web`. Dynamic
+pages (SPA/React, JS-loaded content) need a real browser engine; this module
+runs one Playwright call inside a dedicated worker thread so browser
+crashes/timeouts never take over the tool layer, and the heavyweight
+dependency stays optional:
+
+* not installed → ``ToolError(code="playwright_unavailable")`` with a clear
+  message (the assistant explains the limitation instead of hanging);
+* installed → ``pip install playwright && playwright install chromium``.
+
+The SSRF guard (scheme + private-address rejection) is applied before the
+browser navigates. Note: unlike the httpx path, internal redirects performed
+by the browser engine are not re-checked hop by hop.
+"""
+
+from __future__ import annotations
+
+import html as html_lib
+import logging
+import re
+from concurrent.futures import ThreadPoolExecutor
+from typing import Any
+
+from backend.tools.context import ToolError
+from backend.tools.web import (
+    MAX_TEXT_CHARS,
+    USER_AGENT,
+    _assert_public_http_url,
+    _html_to_text,
+)
+
+logger = logging.getLogger("obsigate.tools.webrender")
+
+# One worker: browser automation is serialized on purpose (one Chromium at a
+# time keeps memory predictable on small hosts).
+_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="obsigate-playwright")
+GOTO_TIMEOUT_MS = 20_000
+
+
+def _playwright_available() -> bool:
+    try:
+        import playwright  # noqa: F401
+    except ImportError:
+        return False
+    return True
+
+
+def _render_in_worker(url: str) -> dict[str, Any]:
+    """Synchronous Playwright render — runs in the dedicated worker thread."""
+    from playwright.sync_api import sync_playwright
+
+    status = 0
+    with sync_playwright() as p:
+        browser = p.chromium.launch(headless=True)
+        try:
+            page = browser.new_page(user_agent=USER_AGENT)
+            response = page.goto(url, wait_until="networkidle", timeout=GOTO_TIMEOUT_MS)
+            if response is not None:
+                status = response.status
+            raw = page.content()
+            title = html_lib.unescape(page.title() or "").strip()
+            text = _html_to_text(raw)[:MAX_TEXT_CHARS]
+        finally:
+            browser.close()
+    title = re.sub(r"\s+", " ", title)[:300]
+    return {
+        "url": url,
+        "status": status,
+        "title": title,
+        "text": text,
+        "rendered": True,
+        "truncated": len(raw) > MAX_TEXT_CHARS,
+    }
+
+
+def render_page(url: str) -> dict[str, Any]:
+    """Render *url* (JavaScript included) and return readable text.
+
+    Raises:
+        ToolError: ``playwright_unavailable`` when the optional dependency is
+            missing, ``render_unavailable`` when the render itself failed.
+    """
+    _assert_public_http_url(url)
+    if not _playwright_available():
+        raise ToolError(
+            "Rendu dynamique indisponible : Playwright n'est pas installé "
+            "(pip install playwright && playwright install chromium).",
+            code="playwright_unavailable",
+        )
+    try:
+        return _executor.submit(_render_in_worker, url).result(timeout=GOTO_TIMEOUT_MS / 1000 + 40)
+    except ToolError:
+        raise
+    except Exception as e:
+        logger.warning("render_page failed for %s: %s", url, e)
+        raise ToolError(
+            "Le rendu dynamique de la page a échoué.", code="render_unavailable"
+        ) from e

desktop/Cargo.lock

@@ -2626,7 +2626,7 @@ dependencies = [
 
 [[package]]
 name = "obsigate-desktop"
-version = "2.9.1"
+version = "2.10.0"
 dependencies = [
  "chrono",
  "env_logger",

desktop/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "obsigate-desktop"
-version = "2.9.1"
+version = "2.10.0"
 description = "ObsiGate Desktop — Porte d'entrée native pour vos vaults Obsidian"
 authors = ["Bruno Charest"]
 edition = "2021"

desktop/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://raw.githubusercontent.com/nicedoc/obsigate/main/desktop/tauri.conf.schema.json",
   "productName": "ObsiGate",
-  "version": "2.9.1",
+  "version": "2.10.0",
   "identifier": "com.obsigate.desktop",
   "build": {
     "frontendDist": "../frontend",

docs/ROADMAP.md

@@ -1,6 +1,6 @@
 # ObsiGate — Roadmap
 
-> **Version :** 2.9.1 | **Dernière mise à jour :** 2026-09-17
+> **Version :** 2.10.0 | **Dernière mise à jour :** 2026-09-17
 > **Ce fichier ne contient que le travail à venir** (🔵 En cours + ⚪ Backlog) et un index compact
 > vers les fonctionnalités livrées.
 > - **Méthode de livraison à appliquer pour toute tâche : [DELIVERY_WORKFLOW.md](./DELIVERY_WORKFLOW.md)**
@@ -78,23 +78,6 @@
   - [x] Personnalisation (clé à molette) : ajouter / supprimer / réordonner les commandes
   - [x] i18n FR/EN + tests frontend (helpers purs) + E2E mobile
 
-### 92. Assistant IA — Écosystème d'outils (phase 2 : web étendu, sources connectées, documents)
-
-- **Effort :** 3-5 jours | **Impact :** 🟠 | **Zone :** backend (`backend/tools/`)
-- **Dépend de :** #91 (registre + section « steps » + `web_search`/`fetch_url` livrés)
-- **Description :** étendre le catalogue d'outils de l'assistant au-delà du vault, en
-  suivant la feuille de route technique détaillée :
-  [features/ai-tools-roadmap.md](./features/ai-tools-roadmap.md) (frameworks évalués,
-  bibliothèques par catégorie, transverse retry/cache/secrets/async).
-- **Sous-tâches :**
-  - [x] `web_search` : chaîne de repli sans clé (SearXNG → DuckDuckGo → Bing, `OBSIGATE_WEB_FALLBACK`) — BUG-051
-  - [ ] `web_search` : fournisseurs optionnels à clé (Tavily, Brave, SerpAPI, Exa)
-  - [ ] `fetch_url` : pages dynamiques via Playwright (worker isolé) ; crawl multi-pages Scrapy en tâche de fond
-  - [ ] Sources connectées : Gitea/GitHub (priorité haute) puis Google Drive / OneDrive (OAuth2 `authlib`)
-  - [ ] Production de documents : conversion, tableurs, PDF/Word (outils WRITE + confirmation)
-  - [ ] Transverse : `tenacity` (backoff), cache SQLite des résultats web avec TTL, secrets via Infisical
-  - [ ] Chaque outil : libellé `labels.py` + clés i18n `ai.step.*` FR/EN + tests (httpx mocké)
-
 ---
 
 ## ⚪ Backlog — Sécurité, architecture & performance (P0/P1)
@@ -201,6 +184,7 @@
 | 101 | Forge — Assistant IA partagé (bouton AI Panel = assistant, fournisseur/modèle configuré, autocomplétion) + plein écran Forge/Editer | 2.8.0 | [features/forge-assistant.md](./features/forge-assistant.md) |
 | BUG-057 | Assistant IA — bouton « Ajouter » fonctionnel dans l'éditeur Forge (en plus d'« Editer ») | 2.9.0 | [archive](./archive/COMPLETED_v1-v2.md) |
 | 102 | Assistant IA — bouton « Ajouter la section » par bloc de code (insertion du bloc seul) | 2.9.0 | [archive](./archive/COMPLETED_v1-v2.md) |
+| 92 | Assistant IA — Écosystème d'outils phase 2 (recherche à clé, cache/retry, Playwright, crawl, Gitea/GitHub, documents XLSX/DOCX/CSV/PDF) | 2.10.0 | [features/ai-tools-roadmap.md](./features/ai-tools-roadmap.md) |
 
 ---
 
@@ -208,12 +192,11 @@
 
 | Priorité | Items | Effort total estimé |
 |---|---|---|
-| ✅ Complété | #1 → #59, #61–72, #74–76, #78–84, #88–93, #94–100 | ~114 jours réalisés |
+| ✅ Complété | #1 → #59, #61–72, #74–76, #78–84, #88–93, #94–100, #102, #92 | ~114 jours réalisés |
 | 🔵 P2 restant | #77 Desktop : signature de code (non retenue), 6 tests E2E **manuels** ([protocole](./DESKTOP_E2E_CHECKLIST.md)) | ~0,5-1 jour |
 | ⚪ P4 restant | #73 Sync (6-8j) | 6-8 jours |
-| ⚪ P2 restant | #92 Assistant IA — écosystème d'outils phase 2 (web étendu, sources connectées, documents) | 3-5 jours |
 | ⚪ P0/P1 restant | #85-87 Refonte architecturale, performance, CI/CD (issues BUG-035 → BUG-040) | ~15-23 jours |
-| **Total restant** | **10 items + finitions** | **~30-47 jours** |
+| **Total restant** | **7 items + finitions** | **~27-42 jours** |
 
 ---
 

docs/features/ai-tools-roadmap.md

@@ -1,6 +1,6 @@
 # #92 — Assistant IA — Écosystème d'outils : feuille de route technique
 
-> **Statut :** ⚪ Backlog (phase 1 livrée dans #91)
+> **Statut :** ✅ livré (phase 2, version 2.10.0) — phase 1 livrée dans #91
 > **Effort estimé :** 3-5 jours pour la phase 2 | **Impact :** 🟠
 > **Références :** [Roadmap](../ROADMAP.md) · [Outils & MCP #79](./ai-tools-mcp.md) ·
 > [Fenêtre de discussion #91](./ai-assistant-conversation-ux.md) · [Changelog](../../CHANGELOG.md)
@@ -36,6 +36,17 @@ réécriture de la boucle n'est nécessaire.
 
 ## 3. Phase 2 — catégories à implémenter
 
+> **Livré (2.10.0, #92).** Récapitulatif des décisions finales :
+
+| Catégorie | Décision livrée |
+|---|---|
+| Recherche web étendue | Tavily, Brave, SerpAPI, Exa à clé (`OBSIGATE_*_API_KEY`), essayés avant SearXNG ; ordre via `OBSIGATE_WEB_PROVIDERS` |
+| Lecture de pages | `fetch_url(render=True)` → worker Playwright isolé (`backend/tools/webrender.py`), dépendance optionnelle + erreur explicite |
+| Crawl multi-pages | `crawl_site` (WRITE + confirmation) : BFS httpx borné (≤ 20 pages, même hôte, SSRF sur chaque URL) → condensé Markdown dans le vault. Scrapy écarté (dépendance lourde inutile à cette échelle) |
+| Sources connectées | Gitea + GitHub (`git_list_repos`, `git_search_issues`, `git_get_file`) via env/Infisical ; drives cloud (Drive/OneDrive) orientés serveur MCP externe (#79) |
+| Production de documents | `create_xlsx`, `create_docx`, `create_csv`, `create_pdf` — WRITE + confirmation, écrit via `save_raw_file(allow_docs=True)` (path safety + backup) |
+| Transverse | Cache SQLite (`webcache.py`, TTL `OBSIGATE_WEB_CACHE_TTL`), retry backoff maison (`OBSIGATE_WEB_RETRY`), secrets par env (Infisical-compatible) |
+
 ### 3.1 Recherche web étendue (`web_search`)
 - **Fallback sans clé — ✅ livré (BUG-051)** : chaîne de fournisseurs dans
   `backend/tools/web.py` — SearXNG auto-hébergé (`OBSIGATE_SEARXNG_URL`) puis, si

frontend/locales/en.json

@@ -1822,6 +1822,14 @@
   "ai.step.vaults": "Listed the vaults",
   "ai.step.fetch_url": "Opened a web page: {value}",
   "ai.step.web_search": "Searched the web: {value}",
+  "ai.step.crawl": "Crawled a site: {value}",
+  "ai.step.git_repos": "Listed repositories ({value})",
+  "ai.step.git_issues": "Searched issues: {value}",
+  "ai.step.git_file": "Read a repo file: {value}",
+  "ai.step.xlsx_create": "Spreadsheet proposed: {value}",
+  "ai.step.docx_create": "Word document proposed: {value}",
+  "ai.step.csv_create": "CSV file proposed: {value}",
+  "ai.step.pdf_create": "PDF document proposed: {value}",
   "bookslm.copied": "Copied to clipboard",
   "bookslm.error": "AI service error",
   "bookslm.regenerate": "Regenerate",

frontend/locales/fr.json

@@ -1822,6 +1822,14 @@
   "ai.step.vaults": "Liste des vaults consultée",
   "ai.step.fetch_url": "Page web consultée : {value}",
   "ai.step.web_search": "Recherche sur le web : {value}",
+  "ai.step.crawl": "Site exploré : {value}",
+  "ai.step.git_repos": "Dépôts listés ({value})",
+  "ai.step.git_issues": "Issues recherchées : {value}",
+  "ai.step.git_file": "Fichier de dépôt lu : {value}",
+  "ai.step.xlsx_create": "Tableur proposé : {value}",
+  "ai.step.docx_create": "Document Word proposé : {value}",
+  "ai.step.csv_create": "Fichier CSV proposé : {value}",
+  "ai.step.pdf_create": "Document PDF proposé : {value}",
   "bookslm.copied": "Réponse copiée dans le presse-papiers",
   "bookslm.error": "Erreur du service AI",
   "bookslm.regenerate": "Régénérer",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "obsigate",
-  "version": "2.9.1",
+  "version": "2.10.0",
   "description": "**Porte d'entrée web ultra-léger pour vos vaults Obsidian** — Accédez, naviguez et recherchez dans toutes vos notes Obsidian depuis n'importe quel appareil via une interface web moderne et responsive.",
   "main": "patch.js",
   "directories": {

tests/conftest.py

@@ -22,6 +22,23 @@ def _reset_tool_ratelimit():
     ratelimit.reset()
 
 
+@pytest.fixture(autouse=True)
+def _disable_web_cache():
+    """Web cache off by default: tests stay hermetic (no cross-test hits).
+
+    tests/test_web_cache.py re-enables it explicitly with a tmp path.
+    """
+    from backend.tools import webcache
+
+    saved_path = os.environ.get("OBSIGATE_WEB_CACHE_PATH")
+    os.environ["OBSIGATE_WEB_CACHE_TTL"] = "0"
+    yield
+    if saved_path is None:
+        os.environ.pop("OBSIGATE_WEB_CACHE_PATH", None)
+    else:
+        os.environ["OBSIGATE_WEB_CACHE_PATH"] = saved_path
+
+
 @pytest.fixture(autouse=True)
 def _clean_env():
     """Ensure no vault env vars leak between tests — but preserve test vault config."""

tests/test_connected_tools.py

@@ -0,0 +1,230 @@
+"""Unit tests for the connected sources (#92): Gitea & GitHub tools.
+
+All HTTP calls are mocked (httpx.request monkeypatched) — the CI never talks
+to a real Gitea/GitHub instance.
+"""
+
+import base64
+from typing import Any
+
+import pytest
+
+import backend.tools.connected as connected
+from backend.tools.context import ToolContext, ToolError, ToolMode
+from backend.tools.registry import get_tool
+
+
+class FakeResponse:
+    def __init__(self, json_data: Any = None, status_code: int = 200):
+        self._json = json_data
+        self.status_code = status_code
+
+    def json(self):
+        return self._json
+
+    def raise_for_status(self):
+        if self.status_code >= 400:
+            import httpx
+
+            raise httpx.HTTPStatusError("boom", request=None, response=self)  # type: ignore[arg-type]
+
+
+def _ctx() -> ToolContext:
+    return ToolContext(user={"username": "tester", "vaults": []}, mode=ToolMode.IN_APP)
+
+
+@pytest.fixture
+def gitea_env(monkeypatch):
+    monkeypatch.setenv("OBSIGATE_GITEA_URL", "https://git.example.net")
+    monkeypatch.setenv("OBSIGATE_GITEA_TOKEN", "tok-gitea")
+
+
+@pytest.fixture
+def github_env(monkeypatch):
+    monkeypatch.setenv("OBSIGATE_GITHUB_TOKEN", "tok-gh")
+
+
+def _patch_request(monkeypatch, handler):
+    def fake_request(method, url, headers=None, timeout=None, follow_redirects=False, **kw):
+        captured = {"method": method, "url": str(url), "headers": headers or {},
+                    "params": kw.get("params")}
+        return handler(captured)
+
+    monkeypatch.setattr(connected.httpx, "request", fake_request)
+
+
+class TestRegistration:
+    @pytest.mark.parametrize("name", ["git_list_repos", "git_search_issues", "git_get_file"])
+    def test_tools_registered_read(self, name):
+        from backend.tools.context import ToolRisk
+
+        spec = get_tool(name)
+        assert spec is not None
+        assert spec.risk == ToolRisk.READ
+        assert "gitea" in spec.description or "github" in spec.description.lower()
+
+
+class TestConfiguration:
+    def test_gitea_requires_base_url(self, monkeypatch):
+        monkeypatch.delenv("OBSIGATE_GITEA_URL", raising=False)
+        with pytest.raises(ToolError) as ei:
+            connected._provider_base("gitea")
+        assert ei.value.code == "provider_not_configured"
+
+    def test_unknown_provider_rejected(self):
+        with pytest.raises(ToolError) as ei:
+            connected._provider_base("gitlab")
+        assert ei.value.code == "invalid_arguments"
+
+    def test_gitea_token_sent_as_header(self, gitea_env, monkeypatch):
+        captured = {}
+
+        def handler(captured_req):
+            captured.update(captured_req)
+            return FakeResponse(json_data={"data": []})
+
+        _patch_request(monkeypatch, handler)
+        connected.git_list_repos(_ctx(), connected.GitProviderInput(provider="gitea"))
+        assert captured["headers"]["Authorization"] == "token tok-gitea"
+
+
+class TestListRepos:
+    def test_gitea_search_endpoint(self, gitea_env, monkeypatch):
+        captured = {}
+
+        def handler(captured_req):
+            captured.update(captured_req)
+            return FakeResponse(json_data={"data": [
+                {"name": "ObsiGate", "full_name": "bruno/ObsiGate",
+                 "html_url": "https://git.example.net/bruno/ObsiGate",
+                 "description": "vault gateway", "updated_at": "2026-09-01",
+                 "private": False},
+            ]})
+
+        _patch_request(monkeypatch, handler)
+        out = connected.git_list_repos(_ctx(), connected.GitProviderInput(provider="gitea"))
+        assert "/api/v1/repos/search" in captured["url"]
+        assert out["repos"][0]["name"] == "ObsiGate"
+        assert out["count"] == 1
+
+    def test_github_single_repo(self, github_env, monkeypatch):
+        captured = {}
+
+        def handler(captured_req):
+            captured.update(captured_req)
+            return FakeResponse(json_data={
+                "name": "ObsiGate", "full_name": "bruno/ObsiGate",
+                "html_url": "https://github.com/bruno/ObsiGate",
+                "description": "", "updated_at": "2026-09-02", "private": True,
+            })
+
+        _patch_request(monkeypatch, handler)
+        out = connected.git_list_repos(_ctx(), connected.GitProviderInput(
+            provider="github", repo="bruno/ObsiGate"))
+        assert captured["url"].endswith("/repos/bruno/ObsiGate")
+        assert out["repos"][0]["full_name"] == "bruno/ObsiGate"
+        assert out["repos"][0]["private"] is True
+
+
+class TestSearchIssues:
+    def test_gitea_scoped_to_repo(self, gitea_env, monkeypatch):
+        captured = {}
+
+        def handler(captured_req):
+            captured.update(captured_req)
+            return FakeResponse(json_data=[
+                {"number": 12, "title": "Bug affichage", "html_url": "https://x/12",
+                 "state": "open"},
+            ])
+
+        _patch_request(monkeypatch, handler)
+        out = connected.git_search_issues(_ctx(), connected.GitSearchIssuesInput(
+            provider="gitea", query="affichage", repo="bruno/ObsiGate"))
+        assert "/repos/bruno/ObsiGate/issues" in captured["url"]
+        assert out["issues"][0]["id"] == 12
+        assert out["issues"][0]["pull_request"] is False
+
+    def test_github_search_syntax(self, github_env, monkeypatch):
+        captured = {}
+
+        def handler(captured_req):
+            captured.update(captured_req)
+            return FakeResponse(json_data={"items": [
+                {"number": 5, "title": "Crash on save", "html_url": "https://gh/5",
+                 "state": "open", "pull_request": {"url": "x"}},
+            ]})
+
+        _patch_request(monkeypatch, handler)
+        out = connected.git_search_issues(_ctx(), connected.GitSearchIssuesInput(
+            provider="github", query="crash", repo="bruno/ObsiGate", state="open"))
+        assert "/search/issues" in captured["url"]
+        assert "repo:bruno/ObsiGate" in captured["params"]["q"]
+        assert out["issues"][0]["pull_request"] is True
+
+
+class TestGetFile:
+    def test_gitea_base64_content_decoded(self, gitea_env, monkeypatch):
+        payload = base64.b64encode("# Readme\n\nBonjour".encode()).decode()
+
+        def handler(_captured):
+            return FakeResponse(json_data={
+                "path": "README.md", "size": 15, "encoding": "base64", "content": payload,
+            })
+
+        _patch_request(monkeypatch, handler)
+        out = connected.git_get_file(_ctx(), connected.GitGetFileInput(
+            provider="gitea", repo="bruno/ObsiGate", path="README.md"))
+        assert "Bonjour" in out["content"]
+        assert out["truncated"] is False
+
+    def test_github_ref_parameter(self, github_env, monkeypatch):
+        captured = {}
+
+        def handler(captured_req):
+            captured.update(captured_req)
+            return FakeResponse(json_data={
+                "path": "a.md", "size": 1, "encoding": "base64",
+                "content": base64.b64encode(b"x").decode(),
+            })
+
+        _patch_request(monkeypatch, handler)
+        connected.git_get_file(_ctx(), connected.GitGetFileInput(
+            provider="github", repo="o/r", path="a.md", ref="v2.9.0"))
+        assert captured["url"].endswith("?ref=v2.9.0")
+
+    def test_missing_repo_or_path_rejected(self, gitea_env):
+        with pytest.raises(ToolError) as ei:
+            connected.git_get_file(_ctx(), connected.GitGetFileInput(
+                provider="gitea", repo="", path="a.md"))
+        assert ei.value.code == "invalid_arguments"
+
+
+class TestErrors:
+    def test_404_maps_to_not_found(self, gitea_env, monkeypatch):
+        def handler(_captured):
+            return FakeResponse(json_data={}, status_code=404)
+
+        _patch_request(monkeypatch, handler)
+        with pytest.raises(ToolError) as ei:
+            connected.git_list_repos(_ctx(), connected.GitProviderInput(provider="gitea"))
+        assert ei.value.code == "not_found"
+
+    def test_401_maps_to_permission_denied(self, gitea_env, monkeypatch):
+        def handler(_captured):
+            return FakeResponse(json_data={}, status_code=401)
+
+        _patch_request(monkeypatch, handler)
+        with pytest.raises(ToolError) as ei:
+            connected.git_list_repos(_ctx(), connected.GitProviderInput(provider="gitea"))
+        assert ei.value.code == "permission_denied"
+
+    def test_network_error_maps_to_tool_error(self, gitea_env, monkeypatch):
+        import httpx
+
+        def fake_request(*a, **kw):
+            raise httpx.ConnectError("down")
+
+        monkeypatch.setattr(connected.httpx, "request", fake_request)
+        with pytest.raises(ToolError) as ei:
+            connected.git_list_repos(_ctx(), connected.GitProviderInput(provider="gitea"))
+        assert ei.value.code == "connected_source_unavailable"

tests/test_crawler.py

@@ -0,0 +1,135 @@
+"""Unit tests for the bounded site crawler (#92): crawl_site (WRITE + confirmation)."""
+
+from typing import Any
+
+import pytest
+
+import backend.tools.crawler as crawler
+from backend.tools.api import ToolConfirmationRequired, ToolContext, ToolError, call_tool
+
+
+class FakeResponse:
+    def __init__(self, content: bytes = b"", status_code: int = 200,
+                 headers: dict | None = None):
+        self.content = content
+        self.status_code = status_code
+        self.headers = headers or {"content-type": "text/html; charset=utf-8"}
+        self.encoding = "utf-8"
+
+    def raise_for_status(self):
+        pass
+
+
+def _ctx() -> ToolContext:
+    return ToolContext(
+        user={"username": "tester", "role": "admin", "vaults": ["*"]},
+        audit_enabled=False,
+    )
+
+
+@pytest.fixture
+def vault(tmp_path, monkeypatch):
+    vault_dir = tmp_path / "Vault"
+    vault_dir.mkdir()
+    monkeypatch.setitem(__import__("backend.indexer", fromlist=["index"]).index,
+                        "Vault", {"name": "Vault", "path": str(vault_dir), "config": {}})
+    return vault_dir
+
+
+PAGE_A = (
+    b"<html><head><title>Docs</title></head><body>"
+    b"<p>Bienvenue sur la documentation.</p>"
+    b'<a href="/page-b">Suite</a><a href="https://other.dev/x">ext</a>'
+    b"</body></html>"
+)
+PAGE_B = (
+    b"<html><head><title>Page B</title></head><body><p>Details ici.</p></body></html>"
+)
+
+
+@pytest.fixture
+def local_urls(monkeypatch):
+    """Skip the DNS-based SSRF guard: test hosts are fake, HTTP is mocked."""
+    monkeypatch.setattr(crawler, "_assert_public_http_url", lambda url: url)
+
+
+@pytest.fixture
+def two_pages(monkeypatch, local_urls):
+    def fake_get(url, **kw):
+        url = str(url)
+        if url.endswith("/page-b"):
+            return FakeResponse(content=PAGE_B)
+        return FakeResponse(content=PAGE_A)
+
+    monkeypatch.setattr(crawler.httpx, "get", fake_get)
+
+
+class TestConfirmation:
+    def test_requires_confirmation(self, vault, two_pages):
+        with pytest.raises(ToolConfirmationRequired):
+            call_tool("crawl_site", _ctx(), {
+                "url": "https://docs.example.dev/start",
+                "vault": "Vault", "path": "Crawls/docs.md",
+            })
+
+
+class TestCrawl:
+    def test_saves_same_host_pages(self, vault, two_pages):
+        out = call_tool("crawl_site", _ctx(), {
+            "url": "https://docs.example.dev/start",
+            "vault": "Vault", "path": "Crawls/docs.md",
+        }, confirm=True)
+        assert out.ok and out.data["pages"] == 2
+        digest = (vault / "Crawls" / "docs.md").read_text(encoding="utf-8")
+        assert "# Crawl de docs.example.dev" in digest
+        assert "Bienvenue sur la documentation." in digest
+        assert "Details ici." in digest
+        assert "other.dev" not in digest
+
+    def test_max_pages_bound(self, vault, monkeypatch, local_urls):
+        # A link farm: every page links to a new page — cap at max_pages.
+        def fake_get(url, **kw):
+            url = str(url)
+            n = int(url.rsplit("/", 1)[-1] or 0)
+            return FakeResponse(
+                content=f"<html><head><title>P{n}</title></head><body>"
+                        f"<p>page {n}</p><a href=\"/{n + 1}\">next</a></body></html>".encode())
+
+        monkeypatch.setattr(crawler.httpx, "get", fake_get)
+        out = call_tool("crawl_site", _ctx(), {
+            "url": "https://farm.example.dev/0",
+            "vault": "Vault", "path": "farm.md", "max_pages": 3,
+        }, confirm=True)
+        assert out.ok and out.data["pages"] == 3
+
+    def test_no_pages_recovered(self, vault, monkeypatch, local_urls):
+        def dead_get(*a, **kw):
+            raise crawler.httpx.ConnectError("down")
+
+        monkeypatch.setattr(crawler.httpx, "get", dead_get)
+        with pytest.raises(ToolError) as ei:
+            call_tool("crawl_site", _ctx(), {
+                "url": "https://dead.example.dev/", "vault": "Vault", "path": "x.md",
+            }, confirm=True)
+        assert ei.value.code == "crawl_failed"
+
+    def test_internal_url_rejected(self, vault):
+        with pytest.raises(ToolError) as ei:
+            call_tool("crawl_site", _ctx(), {
+                "url": "http://127.0.0.1:8080/", "vault": "Vault", "path": "x.md",
+            }, confirm=True)
+        assert ei.value.code in ("ssrf_blocked", "dns_error")
+
+    def test_binary_content_skipped(self, vault, monkeypatch, local_urls):
+        def fake_get(url, **kw):
+            url = str(url)
+            if url.endswith("/x.pdf"):
+                return FakeResponse(content=b"%PDF-1.4", headers={"content-type": "application/pdf"})
+            return FakeResponse(content=PAGE_A)
+
+        monkeypatch.setattr(crawler.httpx, "get", fake_get)
+        out = call_tool("crawl_site", _ctx(), {
+            "url": "https://docs.example.dev/start",
+            "vault": "Vault", "path": "docs.md", "max_pages": 5,
+        }, confirm=True)
+        assert out.ok and out.data["pages"] >= 1

tests/test_document_tools.py

@@ -0,0 +1,160 @@
+"""Unit tests for the document-production tools (#92): create_xlsx, create_docx,
+create_csv, create_pdf — WRITE risk, confirmation gating, vault persistence."""
+
+import csv
+import io
+from pathlib import Path
+
+import pytest
+
+from backend.tools.api import (
+    ToolConfirmationRequired,
+    ToolContext,
+    ToolError,
+    call_tool,
+    get_tool,
+)
+from backend.tools.context import ToolRisk
+
+
+@pytest.fixture
+def vault(tmp_path, monkeypatch):
+    """A minimal configured vault (index entry patched, no full build)."""
+    vault_dir = tmp_path / "Vault"
+    vault_dir.mkdir()
+    monkeypatch.setitem(__import__("backend.indexer", fromlist=["index"]).index,
+                        "Vault", {"name": "Vault", "path": str(vault_dir), "config": {}})
+    return vault_dir
+
+
+def _ctx() -> ToolContext:
+    return ToolContext(
+        user={"username": "tester", "role": "admin", "vaults": ["*"]},
+        audit_enabled=False,
+    )
+
+
+class TestRegistry:
+    @pytest.mark.parametrize("name", ["create_xlsx", "create_docx", "create_csv", "create_pdf"])
+    def test_write_risk_and_confirmation(self, name):
+        spec = get_tool(name)
+        assert spec is not None
+        assert spec.risk == ToolRisk.WRITE
+        assert spec.requires_confirmation is True
+
+
+class TestConfirmationGating:
+    def test_csv_requires_confirmation(self, vault):
+        with pytest.raises(ToolConfirmationRequired):
+            call_tool("create_csv", _ctx(), {
+                "vault": "Vault", "path": "data.csv",
+                "rows": [["a", "b"], [1, 2]],
+            })
+
+    def test_pdf_requires_confirmation(self, vault):
+        with pytest.raises(ToolConfirmationRequired):
+            call_tool("create_pdf", _ctx(), {
+                "vault": "Vault", "path": "doc.pdf", "title": "T", "content": "# H\npara",
+            })
+
+
+class TestCreateCsv:
+    def test_creates_file_in_vault(self, vault):
+        out = call_tool("create_csv", _ctx(), {
+            "vault": "Vault", "path": "Exports/data.csv",
+            "rows": [["nom", "score"], ["alice", 12], ["bob", 9.5]],
+        }, confirm=True)
+        assert out.ok and out.data["success"] is True
+        path = vault / "Exports" / "data.csv"
+        assert path.exists()
+        rows = list(csv.reader(io.StringIO(path.read_text(encoding="utf-8"))))
+        assert rows[0] == ["nom", "score"]
+        assert rows[2] == ["bob", "9.5"]
+
+    def test_semicolon_delimiter(self, vault):
+        call_tool("create_csv", _ctx(), {
+            "vault": "Vault", "path": "d.csv", "delimiter": ";",
+            "rows": [["a", "b"], [1, 2]],
+        }, confirm=True)
+        content = (vault / "d.csv").read_text(encoding="utf-8")
+        assert "a;b" in content
+
+    def test_wrong_extension_rejected(self, vault):
+        with pytest.raises(ToolError) as ei:
+            call_tool("create_csv", _ctx(), {
+                "vault": "Vault", "path": "d.txt", "rows": [["a"], [1]],
+            }, confirm=True)
+        assert ei.value.code == "invalid_arguments"
+
+    def test_empty_rows_rejected(self, vault):
+        with pytest.raises(ToolError):
+            call_tool("create_csv", _ctx(), {
+                "vault": "Vault", "path": "d.csv", "rows": [],
+            }, confirm=True)
+
+
+class TestCreateXlsx:
+    def test_creates_readable_workbook(self, vault):
+        call_tool("create_xlsx", _ctx(), {
+            "vault": "Vault", "path": "Rapports/budget.xlsx",
+            "rows": [["item", "cout"], ["serveur", 1200], ["licence", 300]],
+            "sheet_name": "Budget",
+        }, confirm=True)
+        from openpyxl import load_workbook
+
+        wb = load_workbook(vault / "Rapports" / "budget.xlsx")
+        ws = wb.active
+        assert ws.title == "Budget"
+        assert ws.cell(row=1, column=1).value == "item"
+        assert ws.cell(row=2, column=2).value == 1200
+
+    def test_wrong_extension_rejected(self, vault):
+        with pytest.raises(ToolError):
+            call_tool("create_xlsx", _ctx(), {
+                "vault": "Vault", "path": "b.docx", "rows": [["a"], [1]],
+            }, confirm=True)
+
+
+class TestCreateDocx:
+    def test_creates_readable_document(self, vault):
+        call_tool("create_docx", _ctx(), {
+            "vault": "Vault", "path": "rapport.docx",
+            "title": "Rapport hebdo", "paragraphs": ["Premier point.", "Second point."],
+        }, confirm=True)
+        import docx as docx_lib
+
+        doc = docx_lib.Document(str(vault / "rapport.docx"))
+        texts = [p.text for p in doc.paragraphs]
+        assert "Rapport hebdo" in texts
+        assert "Second point." in texts
+
+    def test_no_paragraphs_rejected(self, vault):
+        with pytest.raises(ToolError):
+            call_tool("create_docx", _ctx(), {
+                "vault": "Vault", "path": "r.docx", "paragraphs": [],
+            }, confirm=True)
+
+
+class TestCreatePdf:
+    def test_creates_valid_pdf(self, vault):
+        call_tool("create_pdf", _ctx(), {
+            "vault": "Vault", "path": "docs/archi.pdf",
+            "title": "Architecture", "content": "# Titre\n\nUn paragraphe.\n## Sous-titre\nAutre texte.",
+        }, confirm=True)
+        raw = (vault / "docs" / "archi.pdf").read_bytes()
+        assert raw.startswith(b"%PDF")
+
+    def test_long_content_truncated(self, vault):
+        call_tool("create_pdf", _ctx(), {
+            "vault": "Vault", "path": "big.pdf", "title": "T", "content": "x" * 500_000,
+        }, confirm=True)
+        assert (vault / "big.pdf").exists()
+
+
+class TestVaultSafety:
+    def test_path_outside_vault_rejected(self, vault):
+        with pytest.raises(ToolError) as ei:
+            call_tool("create_csv", _ctx(), {
+                "vault": "Vault", "path": "../outside.csv", "rows": [["a"], [1]],
+            }, confirm=True)
+        assert ei.value.code in ("path_outside_vault", "invalid_arguments", "tool_execution_error")

tests/test_web_cache.py

@@ -0,0 +1,82 @@
+"""Unit tests for the SQLite web cache (backend.tools.webcache, #92)."""
+
+import pytest
+
+from backend.tools import webcache
+
+
+@pytest.fixture
+def cache_enabled(tmp_path, monkeypatch):
+    """Enable the cache against an isolated file with a short TTL."""
+    monkeypatch.setenv("OBSIGATE_WEB_CACHE_PATH", str(tmp_path / "cache.sqlite3"))
+    monkeypatch.setenv("OBSIGATE_WEB_CACHE_TTL", "60")
+    webcache._schema_ready = False
+    yield
+    webcache._schema_ready = False
+
+
+class TestCacheKey:
+    def test_deterministic_and_payload_sensitive(self):
+        a = webcache.cache_key("search", {"q": "pizza", "page": 1})
+        b = webcache.cache_key("search", {"page": 1, "q": "pizza"})
+        c = webcache.cache_key("search", {"q": "pasta", "page": 1})
+        d = webcache.cache_key("fetch", {"q": "pizza", "page": 1})
+        assert a == b
+        assert a != c
+        assert a != d
+
+
+class TestCacheRoundTrip:
+    def test_set_get_roundtrip(self, cache_enabled):
+        key = webcache.cache_key("search", {"q": "x"})
+        webcache.cache_set(key, {"results": [1, 2, 3], "provider": "tavily"})
+        assert webcache.cache_get(key) == {"results": [1, 2, 3], "provider": "tavily"}
+
+    def test_miss_returns_none(self, cache_enabled):
+        assert webcache.cache_get("search:unknown") is None
+
+    def test_overwrite_updates_value(self, cache_enabled):
+        key = webcache.cache_key("search", {"q": "x"})
+        webcache.cache_set(key, {"v": 1})
+        webcache.cache_set(key, {"v": 2})
+        assert webcache.cache_get(key) == {"v": 2}
+
+    def test_ttl_expiry(self, tmp_path, monkeypatch):
+        monkeypatch.setenv("OBSIGATE_WEB_CACHE_PATH", str(tmp_path / "cache.sqlite3"))
+        monkeypatch.setenv("OBSIGATE_WEB_CACHE_TTL", "0.05")
+        webcache._schema_ready = False
+        key = webcache.cache_key("search", {"q": "x"})
+        webcache.cache_set(key, {"v": 1})
+        assert webcache.cache_get(key) == {"v": 1}
+        import time
+
+        time.sleep(0.15)
+        assert webcache.cache_get(key) is None
+
+    def test_disabled_when_ttl_zero(self, cache_enabled, monkeypatch):
+        monkeypatch.setenv("OBSIGATE_WEB_CACHE_TTL", "0")
+        key = webcache.cache_key("search", {"q": "x"})
+        webcache.cache_set(key, {"v": 1})
+        assert webcache.cache_get(key) is None
+
+    def test_purge_expired(self, cache_enabled, monkeypatch):
+        key = webcache.cache_key("search", {"q": "x"})
+        webcache.cache_set(key, {"v": 1})
+        monkeypatch.setenv("OBSIGATE_WEB_CACHE_TTL", "0.05")
+        import time
+
+        time.sleep(0.15)
+        assert webcache.purge_expired() >= 1
+        assert webcache.cache_get(key) is None
+
+    def test_corrupt_db_degrades_silently(self, tmp_path, monkeypatch):
+        # A directory as cache file breaks sqlite3.connect: the cache must
+        # disable itself instead of breaking the tools.
+        monkeypatch.setenv("OBSIGATE_WEB_CACHE_PATH", str(tmp_path))
+        monkeypatch.setenv("OBSIGATE_WEB_CACHE_TTL", "60")
+        webcache._schema_ready = False
+        try:
+            assert webcache.cache_get("search:x") is None
+            webcache.cache_set("search:x", {"v": 1})
+        finally:
+            webcache._schema_ready = False

tests/test_web_search_providers.py

@@ -0,0 +1,155 @@
+"""Unit tests for the keyed web-search providers (#92): Tavily, Brave,
+SerpAPI, Exa — plus provider ordering and the transient retry."""
+
+import httpx
+import pytest
+
+import backend.tools.web as web
+from backend.tools.context import ToolContext, ToolError, ToolMode
+
+
+class FakeResponse:
+    def __init__(self, json_data=None, status_code=200, content=b""):
+        self._json = json_data
+        self.status_code = status_code
+        self.content = content
+        self.encoding = "utf-8"
+        self.headers = {}
+
+    def json(self):
+        return self._json
+
+    def raise_for_status(self):
+        if self.status_code >= 400:
+            raise web.httpx.HTTPStatusError("boom", request=None, response=self)  # type: ignore[arg-type]
+
+
+def _ctx() -> ToolContext:
+    return ToolContext(user={"username": "tester", "vaults": []}, mode=ToolMode.IN_APP)
+
+
+def _no_fallback(monkeypatch):
+    """Limit the chain to the provider under test (no searxng/ddg/bing noise)."""
+    monkeypatch.setattr(web, "WEB_FALLBACK_ENABLED", False)
+    monkeypatch.setattr(web, "SEARXNG_URL", "http://searxng.invalid")
+    monkeypatch.setattr(web.httpx, "get", lambda *a, **kw: (_ for _ in ()).throw(
+        httpx.ConnectError("offline")))
+    monkeypatch.setattr(web.httpx, "post", lambda *a, **kw: (_ for _ in ()).throw(
+        httpx.ConnectError("offline")))
+
+
+class TestKeyedProviderParsers:
+    def test_tavily_maps_results(self, monkeypatch):
+        captured = {}
+
+        def fake_post(url, json=None, **kw):
+            captured["url"] = url
+            captured["payload"] = json
+            return FakeResponse(json_data={"results": [
+                {"title": "T", "url": "https://a.dev", "content": "c" * 800},
+            ]})
+
+        monkeypatch.setattr(web.httpx, "post", fake_post)
+        results, engines = web._search_tavily("q", web.WebSearchInput(query="q", max_results=3))
+        assert results[0]["title"] == "T"
+        assert len(results[0]["snippet"]) <= 600
+        assert engines == []
+        assert captured["payload"]["api_key"] == ""
+        assert captured["payload"]["max_results"] == 3
+
+    def test_brave_maps_results(self, monkeypatch):
+        captured = {}
+
+        def fake_get(url, params=None, headers=None, **kw):
+            captured["url"] = url
+            captured["headers"] = headers
+            return FakeResponse(json_data={"web": {"results": [
+                {"title": "B", "url": "https://b.dev", "description": "d"},
+            ]}})
+
+        monkeypatch.setattr(web.httpx, "get", fake_get)
+        results, _ = web._search_brave("q", web.WebSearchInput(query="q"))
+        assert results[0]["title"] == "B"
+        assert "api.search.brave.com" in str(captured["url"])
+
+    def test_serpapi_maps_results(self, monkeypatch):
+        monkeypatch.setattr(web.httpx, "get", lambda *a, **kw: FakeResponse(json_data={
+            "organic_results": [{"title": "S", "link": "https://s.dev", "snippet": "sn"}],
+        }))
+        results, _ = web._search_serpapi("q", web.WebSearchInput(query="q"))
+        assert results[0]["url"] == "https://s.dev"
+
+    def test_exa_maps_results(self, monkeypatch):
+        monkeypatch.setattr(web.httpx, "post", lambda *a, **kw: FakeResponse(json_data={
+            "results": [{"title": "E", "url": "https://e.dev", "text": "t" * 900}],
+        }))
+        results, _ = web._search_exa("q", web.WebSearchInput(query="q"))
+        assert results[0]["title"] == "E"
+        assert len(results[0]["snippet"]) <= 600
+
+
+class TestProviderChain:
+    def test_no_key_falls_back_to_searxng(self, monkeypatch):
+        for var in ("OBSIGATE_TAVILY_API_KEY", "OBSIGATE_BRAVE_API_KEY",
+                    "OBSIGATE_SERPAPI_API_KEY", "OBSIGATE_EXA_API_KEY"):
+            monkeypatch.delenv(var, raising=False)
+        chain = [name for name, _ in web._provider_chain()]
+        assert chain[0] == "searxng"
+        assert "tavily" not in chain
+
+    def test_keyed_provider_used_first_when_key_set(self, monkeypatch):
+        monkeypatch.setenv("OBSIGATE_TAVILY_API_KEY", "k")
+        chain = [name for name, _ in web._provider_chain()]
+        assert chain[0] == "tavily"
+        assert "brave" not in chain  # no key → skipped
+
+    def test_explicit_order_env(self, monkeypatch):
+        monkeypatch.setenv("OBSIGATE_TAVILY_API_KEY", "k")
+        monkeypatch.setenv("OBSIGATE_EXA_API_KEY", "k")
+        monkeypatch.setenv("OBSIGATE_WEB_PROVIDERS", "exa,unknown,tavily")
+        chain = [name for name, _ in web._provider_chain()]
+        assert chain[:2] == ["exa", "tavily"]
+
+    def test_search_uses_keyed_provider_first(self, monkeypatch):
+        _no_fallback(monkeypatch)
+        monkeypatch.setenv("OBSIGATE_BRAVE_API_KEY", "k")
+        monkeypatch.setattr(web.httpx, "get", lambda *a, **kw: FakeResponse(json_data={
+            "web": {"results": [{"title": "B", "url": "https://b.dev", "description": "d"}]},
+        }))
+        out = web.web_search(_ctx(), web.WebSearchInput(query="q"))
+        assert out["provider"] == "brave"
+        assert out["count"] == 1
+
+
+class TestRetry:
+    def test_transient_error_retried_then_succeeds(self, monkeypatch):
+        monkeypatch.setattr(web, "WEB_RETRY_ATTEMPTS", 1)
+        monkeypatch.setattr(web, "_provider_chain", lambda: [("searxng", web._search_searxng)])
+        calls = {"n": 0}
+
+        def flaky_get(*a, **kw):
+            calls["n"] += 1
+            if calls["n"] == 1:
+                raise httpx.ConnectError("blip")
+            return FakeResponse(json_data={"results": [
+                {"title": "A", "url": "https://a.dev", "content": "x"}]})
+
+        monkeypatch.setattr(web.httpx, "get", flaky_get)
+        out = web.web_search(_ctx(), web.WebSearchInput(query="q"))
+        assert out["provider"] == "searxng"
+        assert calls["n"] == 2
+
+    def test_persistent_error_not_retried_forever(self, monkeypatch):
+        monkeypatch.setattr(web, "WEB_RETRY_ATTEMPTS", 1)
+        monkeypatch.setattr(web, "_provider_chain", lambda: [("searxng", web._search_searxng)])
+        calls = {"n": 0}
+
+        def dead_get(*a, **kw):
+            calls["n"] += 1
+            raise httpx.ConnectError("down")
+
+        monkeypatch.setattr(web.httpx, "get", dead_get)
+        with pytest.raises(ToolError) as ei:
+            web.web_search(_ctx(), web.WebSearchInput(query="q"))
+        assert ei.value.code == "web_search_unavailable"
+        assert calls["n"] == 2  # initial + 1 retry, per provider

tests/test_webrender.py

@@ -0,0 +1,73 @@
+"""Unit tests for the dynamic rendering path (#92): fetch_url(render=True)."""
+
+import pytest
+
+import backend.tools.web as web
+from backend.tools import webrender
+from backend.tools.context import ToolContext, ToolError, ToolMode
+from backend.tools.registry import get_tool
+
+
+def _ctx() -> ToolContext:
+    return ToolContext(user={"username": "tester", "vaults": []}, mode=ToolMode.IN_APP)
+
+
+class TestRegistration:
+    def test_render_param_exposed_in_schema(self):
+        spec = get_tool("fetch_url")
+        assert spec is not None
+        assert "render" in spec.input_model.model_fields
+
+
+class TestRenderUnavailable:
+    def test_missing_playwright_clear_error(self, monkeypatch):
+        monkeypatch.setattr(webrender, "_playwright_available", lambda: False)
+        with pytest.raises(ToolError) as ei:
+            web.fetch_url(_ctx(), web.FetchUrlInput(
+                url="https://example.com/spa", render=True))
+        assert ei.value.code == "playwright_unavailable"
+
+    def test_ssrf_guard_applied_before_render(self, monkeypatch):
+        monkeypatch.setattr(webrender, "_playwright_available", lambda: True)
+        with pytest.raises(ToolError) as ei:
+            web.fetch_url(_ctx(), web.FetchUrlInput(
+                url="http://127.0.0.1:9222/devtools", render=True))
+        assert ei.value.code in ("ssrf_blocked", "dns_error")
+
+
+class TestRenderSuccess:
+    def test_fetch_url_delegates_to_worker(self, monkeypatch):
+        captured = {}
+
+        def fake_render(url):
+            captured["url"] = url
+            return {"url": url, "status": 200, "title": "SPA",
+                    "text": "dynamic content", "rendered": True, "truncated": False}
+
+        monkeypatch.setattr(webrender, "render_page", fake_render)
+        out = web.fetch_url(_ctx(), web.FetchUrlInput(
+            url="https://example.com/spa", render=True))
+        assert captured["url"] == "https://example.com/spa"
+        assert out["rendered"] is True
+        assert "dynamic content" in out["text"]
+
+    def test_worker_failure_maps_to_tool_error(self, monkeypatch):
+        monkeypatch.setattr(webrender, "_playwright_available", lambda: True)
+
+        def boom(url):
+            raise RuntimeError("chromium crashed")
+
+        # The executor re-raises the worker exception on .result(); render_page
+        # must wrap it into a ToolError instead of leaking a bare exception.
+        monkeypatch.setattr(webrender, "_render_in_worker", boom)
+        with pytest.raises(ToolError) as ei:
+            web.fetch_url(_ctx(), web.FetchUrlInput(
+                url="https://example.com/spa", render=True))
+        assert ei.value.code == "render_unavailable"
+
+
+class TestMarkdownExtraction:
+    def test_html_to_text_reused(self):
+        text = webrender._html_to_text("<html><body><p>hello</p><script>x()</script></body></html>")
+        assert "hello" in text
+        assert "x()" not in text