Initial commit: hardened DeerFlow factory

Vendored deer-flow upstream (bytedance/deer-flow) plus prompt-injection
hardening:

- New deerflow.security package: content_delimiter, html_cleaner,
  sanitizer (8 layers — invisible chars, control chars, symbols, NFC,
  PUA, tag chars, horizontal whitespace collapse with newline/tab
  preservation, length cap)
- New deerflow.community.searx package: web_search, web_fetch,
  image_search backed by a private SearX instance, every external
  string sanitized and wrapped in <<<EXTERNAL_UNTRUSTED_CONTENT>>>
  delimiters
- All native community web providers (ddg_search, tavily, exa,
  firecrawl, jina_ai, infoquest, image_search) replaced with hard-fail
  stubs that raise NativeWebToolDisabledError at import time, so a
  misconfigured tool.use path fails loud rather than silently falling
  back to unsanitized output
- Native client back-doors (jina_client.py, infoquest_client.py)
  stubbed too
- Native-tool tests quarantined under tests/_disabled_native/
  (collect_ignore_glob via local conftest.py)
- Sanitizer Layer 7 fix: only collapse horizontal whitespace, preserve
  newlines and tabs so list/table structure survives
- Hardened runtime config.yaml references only the searx-backed tools
- Factory overlay (backend/) kept in sync with deer-flow tree as a
  reference / source

See HARDENING.md for the full audit trail and verification steps.
This commit is contained in:
2026-04-12 14:23:57 +02:00
commit 6de0bf9f5b
889 changed files with 173052 additions and 0 deletions

View File

@@ -0,0 +1,190 @@
# RFC: Extract Shared Skill Installer and Upload Manager into Harness
## 1. Problem
Gateway (`app/gateway/routers/skills.py`, `uploads.py`) and Client (`deerflow/client.py`) each independently implement the same business logic:
### Skill Installation
| Logic | Gateway (`skills.py`) | Client (`client.py`) |
|-------|----------------------|---------------------|
| Zip safety check | `_is_unsafe_zip_member()` | Inline `Path(info.filename).is_absolute()` |
| Symlink filtering | `_is_symlink_member()` | `p.is_symlink()` post-extraction delete |
| Zip bomb defence | `total_size += info.file_size` (declared) | `total_size > 100MB` (declared) |
| macOS metadata filter | `_should_ignore_archive_entry()` | None |
| Frontmatter validation | `_validate_skill_frontmatter()` | `_validate_skill_frontmatter()` |
| Duplicate detection | `HTTPException(409)` | `ValueError` |
**Two implementations, inconsistent behaviour**: Gateway streams writes and tracks real decompressed size; Client sums declared `file_size`. Gateway skips symlinks during extraction; Client extracts everything then walks and deletes symlinks.
### Upload Management
| Logic | Gateway (`uploads.py`) | Client (`client.py`) |
|-------|----------------------|---------------------|
| Directory access | `get_uploads_dir()` + `mkdir` | `_get_uploads_dir()` + `mkdir` |
| Filename safety | Inline `Path(f).name` + manual checks | No checks, uses `src_path.name` directly |
| Duplicate handling | None (overwrites) | None (overwrites) |
| Listing | Inline `iterdir()` | Inline `os.scandir()` |
| Deletion | Inline `unlink()` + traversal check | Inline `unlink()` + traversal check |
| Path traversal | `resolve().relative_to()` | `resolve().relative_to()` |
**The same traversal check is written twice** — any security fix must be applied to both locations.
## 2. Design Principles
### Dependency Direction
```
app.gateway.routers.skills ──┐
app.gateway.routers.uploads ──┤── calls ──→ deerflow.skills.installer
deerflow.client ──┘ deerflow.uploads.manager
```
- Shared modules live in the harness layer (`deerflow.*`), pure business logic, no FastAPI dependency
- Gateway handles HTTP adaptation (`UploadFile` → bytes, exceptions → `HTTPException`)
- Client handles local adaptation (`Path` → copy, exceptions → Python exceptions)
- Satisfies `test_harness_boundary.py` constraint: harness never imports app
### Exception Strategy
| Shared Layer Exception | Gateway Maps To | Client |
|----------------------|-----------------|--------|
| `FileNotFoundError` | `HTTPException(404)` | Propagates |
| `ValueError` | `HTTPException(400)` | Propagates |
| `SkillAlreadyExistsError` | `HTTPException(409)` | Propagates |
| `PermissionError` | `HTTPException(403)` | Propagates |
Replaces stringly-typed routing (`"already exists" in str(e)`) with typed exception matching (`SkillAlreadyExistsError`).
## 3. New Modules
### 3.1 `deerflow.skills.installer`
```python
# Safety checks
is_unsafe_zip_member(info: ZipInfo) -> bool # Absolute path / .. traversal
is_symlink_member(info: ZipInfo) -> bool # Unix symlink detection
should_ignore_archive_entry(path: Path) -> bool # __MACOSX / dotfiles
# Extraction
safe_extract_skill_archive(zip_ref, dest_path, max_total_size=512MB)
# Streaming write, accumulates real bytes (vs declared file_size)
# Dual traversal check: member-level + resolve-level
# Directory resolution
resolve_skill_dir_from_archive(temp_path: Path) -> Path
# Auto-enters single directory, filters macOS metadata
# Install entry point
install_skill_from_archive(zip_path, *, skills_root=None) -> dict
# is_file() pre-check before extension validation
# SkillAlreadyExistsError replaces ValueError
# Exception
class SkillAlreadyExistsError(ValueError)
```
### 3.2 `deerflow.uploads.manager`
```python
# Directory management
get_uploads_dir(thread_id: str) -> Path # Pure path, no side effects
ensure_uploads_dir(thread_id: str) -> Path # Creates directory (for write paths)
# Filename safety
normalize_filename(filename: str) -> str
# Path.name extraction + rejects ".." / "." / backslash / >255 bytes
deduplicate_filename(name: str, seen: set) -> str
# _N suffix increment for dedup, mutates seen in place
# Path safety
validate_path_traversal(path: Path, base: Path) -> None
# resolve().relative_to(), raises PermissionError on failure
# File operations
list_files_in_dir(directory: Path) -> dict
# scandir with stat inside context (no re-stat)
# follow_symlinks=False to prevent metadata leakage
# Non-existent directory returns empty list
delete_file_safe(base_dir: Path, filename: str) -> dict
# Validates traversal first, then unlinks
# URL helpers
upload_artifact_url(thread_id, filename) -> str # Percent-encoded for HTTP safety
upload_virtual_path(filename) -> str # Sandbox-internal path
enrich_file_listing(result, thread_id) -> dict # Adds URLs, stringifies sizes
```
## 4. Changes
### 4.1 Gateway Slimming
**`app/gateway/routers/skills.py`**:
- Remove `_is_unsafe_zip_member`, `_is_symlink_member`, `_safe_extract_skill_archive`, `_should_ignore_archive_entry`, `_resolve_skill_dir_from_archive_root` (~80 lines)
- `install_skill` route becomes a single call to `install_skill_from_archive(path)`
- Exception mapping: `SkillAlreadyExistsError → 409`, `ValueError → 400`, `FileNotFoundError → 404`
**`app/gateway/routers/uploads.py`**:
- Remove inline `get_uploads_dir` (replaced by `ensure_uploads_dir`/`get_uploads_dir`)
- `upload_files` uses `normalize_filename()` instead of inline safety checks
- `list_uploaded_files` uses `list_files_in_dir()` + enrichment
- `delete_uploaded_file` uses `delete_file_safe()` + companion markdown cleanup
### 4.2 Client Slimming
**`deerflow/client.py`**:
- Remove `_get_uploads_dir` static method
- Remove ~50 lines of inline zip handling in `install_skill`
- `install_skill` delegates to `install_skill_from_archive()`
- `upload_files` uses `deduplicate_filename()` + `ensure_uploads_dir()`
- `list_uploads` uses `get_uploads_dir()` + `list_files_in_dir()`
- `delete_upload` uses `get_uploads_dir()` + `delete_file_safe()`
- `update_mcp_config` / `update_skill` now reset `_agent_config_key = None`
### 4.3 Read/Write Path Separation
| Operation | Function | Creates dir? |
|-----------|----------|:------------:|
| upload (write) | `ensure_uploads_dir()` | Yes |
| list (read) | `get_uploads_dir()` | No |
| delete (read) | `get_uploads_dir()` | No |
Read paths no longer have `mkdir` side effects — non-existent directories return empty lists.
## 5. Security Improvements
| Improvement | Before | After |
|-------------|--------|-------|
| Zip bomb detection | Sum of declared `file_size` | Streaming write, accumulates real bytes |
| Symlink handling | Gateway skips / Client deletes post-extract | Unified skip + log |
| Traversal check | Member-level only | Member-level + `resolve().is_relative_to()` |
| Filename backslash | Gateway checks / Client doesn't | Unified rejection |
| Filename length | No check | Reject > 255 bytes (OS limit) |
| thread_id validation | None | Reject unsafe filesystem characters |
| Listing symlink leak | `follow_symlinks=True` (default) | `follow_symlinks=False` |
| 409 status routing | `"already exists" in str(e)` | `SkillAlreadyExistsError` type match |
| Artifact URL encoding | Raw filename in URL | `urllib.parse.quote()` |
## 6. Alternatives Considered
| Alternative | Why Not |
|-------------|---------|
| Keep logic in Gateway, Client calls Gateway via HTTP | Adds network dependency to embedded Client; defeats the purpose of `DeerFlowClient` as an in-process API |
| Abstract base class with Gateway/Client subclasses | Over-engineered for what are pure functions; no polymorphism needed |
| Move everything into `client.py` and have Gateway import it | Violates harness/app boundary — Client is in harness, but Gateway-specific models (Pydantic response types) should stay in app layer |
| Merge Gateway and Client into one module | They serve different consumers (HTTP vs in-process) with different adaptation needs |
## 7. Breaking Changes
**None.** All public APIs (Gateway HTTP endpoints, `DeerFlowClient` methods) retain their existing signatures and return formats. The `SkillAlreadyExistsError` is a subclass of `ValueError`, so existing `except ValueError` handlers still catch it.
## 8. Tests
| Module | Test File | Count |
|--------|-----------|:-----:|
| `skills.installer` | `tests/test_skills_installer.py` | 22 |
| `uploads.manager` | `tests/test_uploads_manager.py` | 20 |
| `client` hardening | `tests/test_client.py` (new cases) | ~40 |
| `client` e2e | `tests/test_client_e2e.py` (new file) | ~20 |
Coverage: unsafe zip / symlink / zip bomb / frontmatter / duplicate / extension / macOS filter / normalize / deduplicate / traversal / list / delete / agent invalidation / upload lifecycle / thread isolation / URL encoding / config pollution.