Files
AutoGPT/autogpt_platform/backend/test/chat/test_security_hooks.py
Zamil Majdy 52b3aebf71 feat(backend/sdk): Claude Agent SDK integration for CoPilot (#12103)
## Summary

Full integration of the **Claude Agent SDK** to replace the existing
one-turn OpenAI-compatible CoPilot implementation with a multi-turn,
tool-using AI agent.

### What changed

**Core SDK Integration** (`chat/sdk/` — new module)
- **`service.py`**: Main orchestrator — spawns Claude Code CLI as a
subprocess per user message, streams responses back via SSE. Handles
conversation history compression, session lifecycle, and error recovery.
- **`response_adapter.py`**: Translates Claude Agent SDK events (text
deltas, tool use, errors, result messages) into the existing CoPilot
`StreamEvent` protocol so the frontend works unchanged.
- **`tool_adapter.py`**: Bridges CoPilot's MCP tools (find_block,
run_block, create_agent, etc.) into the SDK's tool format. Handles
schema conversion and result serialization.
- **`security_hooks.py`**: Pre/Post tool-use hooks that enforce a strict
allowlist of tools, block path traversal, sandbox file operations to
per-session workspace directories, cap sub-agent spawning, and prevent
the model from accessing unauthorized system resources.
- **`transcript.py`**: JSONL transcript I/O utilities for the stateless
`--resume` feature (see below).

**Stateless Multi-Turn Resume** (new)
- Instead of compressing conversation history via LLM on every turn
(lossy and expensive), we capture Claude Code's native JSONL session
transcript via a **Stop hook** callback, persist it in the DB
(`ChatSession.sdkTranscript`), and restore it on the next turn via
`--resume <file>`.
- This preserves full tool call/result context across turns with zero
token overhead for history.
- Feature-flagged via `CLAUDE_AGENT_USE_RESUME` (default: off).
- DB migration: `ALTER TABLE "ChatSession" ADD COLUMN "sdkTranscript"
TEXT`.

**Sandboxed Tool Execution** (`chat/tools/`)
- **`bash_exec.py`**: Sandboxed bash execution using bubblewrap
(`bwrap`) with read-only root filesystem, per-session writable
workspace, resource limits (CPU, memory, file size), and network
isolation.
- **`sandbox.py`**: Shared bubblewrap sandbox infrastructure — generates
`bwrap` command lines with configurable mounts, environment, and
resource constraints.
- **`web_fetch.py`**: URL fetching tool with domain allowlist, size
limits, and content-type filtering.
- **`check_operation_status.py`**: Polling tool for long-running
operations (agent creation, block execution) so the SDK doesn't block
waiting.
- **`find_block.py`** / **`run_block.py`**: Enhanced with category
filtering, optimized response size (removed raw JSON schemas), and
better error handling.

**Security**
- Path traversal prevention: session IDs sanitized, all file ops
confined to workspace dirs, symlink resolution.
- Tool allowlist enforcement via SDK hooks — model cannot call arbitrary
tools.
- Built-in `Bash` tool blocked via `disallowed_tools` to prevent
bypassing sandboxed `bash_exec`.
- Sub-agent (`Task`) spawning capped at configurable limit (default:
10).
- CodeQL-clean path sanitization patterns.

**Streaming & Reconnection**
- SSE stream registry backed by Redis Streams for crash-resilient
reconnection.
- Long-running operation tracking with TTL-based cleanup.
- Atomic message append to prevent race conditions on concurrent writes.

**Configuration** (`config.py`)
- `use_claude_agent_sdk` — master toggle (default: on)
- `claude_agent_model` — model override for SDK path
- `claude_agent_max_buffer_size` — JSON parsing buffer (10MB)
- `claude_agent_max_subtasks` — sub-agent cap (10)
- `claude_agent_use_resume` — transcript-based resume (default: off)
- `thinking_enabled` — extended thinking for Claude models

**Tests**
- `sdk/response_adapter_test.py` — 366 lines covering all event
translation paths
- `sdk/security_hooks_test.py` — 165 lines covering tool blocking, path
traversal, subtask limits
- `chat/model_test.py` — 214 lines covering session model serialization
- `chat/service_test.py` — Integration tests including multi-turn resume
keyword recall
- `tools/find_block_test.py` / `run_block_test.py` — Extended with new
tool behavior tests

## Test plan
- [x] Unit tests pass (`sdk/response_adapter_test.py`,
`security_hooks_test.py`, `model_test.py`)
- [x] Integration test: multi-turn keyword recall via `--resume`
(`service_test.py::test_sdk_resume_multi_turn`)
- [x] Manual E2E: CoPilot chat sessions with tool calls, bash execution,
and multi-turn context
- [x] Pre-commit hooks pass (ruff, isort, black, pyright, flake8)
- [ ] Staging deployment with `claude_agent_use_resume=false` initially
- [ ] Enable resume in staging, verify transcript capture and recall

<!-- greptile_comment -->

<h2>Greptile Overview</h2>

<details><summary><h3>Greptile Summary</h3></summary>

This PR replaces the existing OpenAI-compatible CoPilot with a full
Claude Agent SDK integration, introducing multi-turn conversations,
stateless resume via JSONL transcripts, and sandboxed tool execution.

**Key changes:**
- **SDK integration** (`chat/sdk/`): spawns Claude Code CLI subprocess
per message, translates events to frontend protocol, bridges MCP tools
- **Stateless resume**: captures JSONL transcripts via Stop hook,
persists in `ChatSession.sdkTranscript`, restores with `--resume`
(feature-flagged, default off)
- **Sandboxed execution**: bubblewrap sandbox for bash commands with
filesystem whitelist, network isolation, resource limits
- **Security hooks**: tool allowlist enforcement, path traversal
prevention, workspace-scoped file operations, sub-agent spawn limits
- **Long-running operations**: delegates `create_agent`/`edit_agent` to
existing stream_registry infrastructure for SSE reconnection
- **Feature flag**: `CHAT_USE_CLAUDE_AGENT_SDK` with LaunchDarkly
support, defaults to enabled

**Security issues found:**
- Path traversal validation has logic errors in `security_hooks.py:82`
(tilde expansion order) and `service.py:266` (redundant `..` check)
- Config validator always prefers env var over explicit `False` value
(`config.py:162`)
- Race condition in `routes.py:323` — message persisted before task
registration, could duplicate on retry
- Resource limits in sandbox may fail silently (`sandbox.py:109`)

**Test coverage is strong** with 366 lines for response adapter, 165 for
security hooks, and integration tests for multi-turn resume.
</details>


<details><summary><h3>Confidence Score: 3/5</h3></summary>

- This PR is generally safe but has critical security issues in path
validation that must be fixed before merge
- Score reflects strong architecture and test coverage offset by real
security vulnerabilities: the tilde expansion bug in `security_hooks.py`
could allow sandbox escape, the race condition could cause message
duplication, and the silent ulimit failures could bypass resource
limits. The bubblewrap sandbox and allowlist enforcement are
well-designed, but the path validation bugs need fixing. The transcript
resume feature is properly feature-flagged. Overall the implementation
is solid but the security issues prevent a higher score.
- Pay close attention to
`backend/api/features/chat/sdk/security_hooks.py` (path traversal
vulnerability), `backend/api/features/chat/routes.py` (race condition),
`backend/api/features/chat/tools/sandbox.py` (silent resource limit
failures), and `backend/api/features/chat/sdk/service.py` (redundant
security check)
</details>


<details><summary><h3>Sequence Diagram</h3></summary>

```mermaid
sequenceDiagram
    participant Frontend
    participant Routes as routes.py
    participant SDKService as sdk/service.py
    participant ClaudeSDK as Claude Agent SDK CLI
    participant SecurityHooks as security_hooks.py
    participant ToolAdapter as tool_adapter.py
    participant CoPilotTools as tools/*
    participant Sandbox as sandbox.py (bwrap)
    participant DB as Database
    participant Redis as stream_registry

    Frontend->>Routes: POST /chat (user message)
    Routes->>SDKService: stream_chat_completion_sdk()
    
    SDKService->>DB: get_chat_session()
    DB-->>SDKService: session + messages
    
    alt Resume enabled AND transcript exists
        SDKService->>SDKService: validate_transcript()
        SDKService->>SDKService: write_transcript_to_tempfile()
        Note over SDKService: Pass --resume to SDK
    else No resume
        SDKService->>SDKService: _compress_conversation_history()
        Note over SDKService: Inject history into user message
    end
    
    SDKService->>SecurityHooks: create_security_hooks()
    SDKService->>ToolAdapter: create_copilot_mcp_server()
    SDKService->>ClaudeSDK: spawn subprocess with MCP server
    
    loop Streaming Conversation
        ClaudeSDK->>SDKService: AssistantMessage (text/tool_use)
        SDKService->>Frontend: StreamTextDelta / StreamToolInputAvailable
        
        alt Tool Call
            ClaudeSDK->>SecurityHooks: PreToolUse hook
            SecurityHooks->>SecurityHooks: validate path, check allowlist
            alt Tool blocked
                SecurityHooks-->>ClaudeSDK: deny
            else Tool allowed
                SecurityHooks-->>ClaudeSDK: allow
                ClaudeSDK->>ToolAdapter: call MCP tool
                
                alt Long-running tool (create_agent, edit_agent)
                    ToolAdapter->>Redis: register task
                    ToolAdapter->>DB: save OperationPendingResponse
                    ToolAdapter->>ToolAdapter: spawn background task
                    ToolAdapter-->>ClaudeSDK: OperationStartedResponse
                else Regular tool (find_block, bash_exec)
                    ToolAdapter->>CoPilotTools: execute()
                    alt bash_exec
                        CoPilotTools->>Sandbox: run_sandboxed()
                        Sandbox->>Sandbox: build bwrap command
                        Note over Sandbox: Network isolation,<br/>filesystem whitelist,<br/>resource limits
                        Sandbox-->>CoPilotTools: stdout, stderr, exit_code
                    end
                    CoPilotTools-->>ToolAdapter: result
                    ToolAdapter->>ToolAdapter: stash full output
                    ToolAdapter-->>ClaudeSDK: MCP response
                end
                
                SecurityHooks->>SecurityHooks: PostToolUse hook (log)
            end
        end
        
        ClaudeSDK->>SDKService: UserMessage (ToolResultBlock)
        SDKService->>ToolAdapter: pop_pending_tool_output()
        SDKService->>Frontend: StreamToolOutputAvailable
    end
    
    ClaudeSDK->>SecurityHooks: Stop hook
    SecurityHooks->>SDKService: transcript_path callback
    SDKService->>SDKService: read_transcript_file()
    SDKService->>DB: save transcript to session.sdkTranscript
    
    ClaudeSDK->>SDKService: ResultMessage (success)
    SDKService->>Frontend: StreamFinish
    SDKService->>DB: upsert_chat_session()
```
</details>


<sub>Last reviewed commit: 28c1121</sub>

<!-- greptile_other_comments_section -->

<!-- /greptile_comment -->

---------

Co-authored-by: Swifty <craigswift13@gmail.com>
2026-02-13 15:49:03 +00:00

134 lines
4.7 KiB
Python

"""Tests for SDK security hooks — workspace paths, tool access, and deny messages.
These are pure unit tests with no external dependencies (no SDK, no DB, no server).
They validate that the security hooks correctly block unauthorized paths,
tool access, and dangerous input patterns.
Note: Bash command validation was removed — the SDK built-in Bash tool is not in
allowed_tools, and the bash_exec MCP tool has kernel-level network isolation
(unshare --net) making command-level parsing unnecessary.
"""
from backend.api.features.chat.sdk.security_hooks import (
_validate_tool_access,
_validate_workspace_path,
)
SDK_CWD = "/tmp/copilot-test-session"
def _is_denied(result: dict) -> bool:
hook = result.get("hookSpecificOutput", {})
return hook.get("permissionDecision") == "deny"
def _reason(result: dict) -> str:
return result.get("hookSpecificOutput", {}).get("permissionDecisionReason", "")
# ============================================================
# Workspace path validation (Read, Write, Edit, etc.)
# ============================================================
class TestWorkspacePathValidation:
def test_path_in_workspace(self):
result = _validate_workspace_path(
"Read", {"file_path": f"{SDK_CWD}/file.txt"}, SDK_CWD
)
assert not _is_denied(result)
def test_path_outside_workspace(self):
result = _validate_workspace_path("Read", {"file_path": "/etc/passwd"}, SDK_CWD)
assert _is_denied(result)
def test_tool_results_allowed(self):
result = _validate_workspace_path(
"Read",
{"file_path": "~/.claude/projects/abc/tool-results/out.txt"},
SDK_CWD,
)
assert not _is_denied(result)
def test_claude_settings_blocked(self):
result = _validate_workspace_path(
"Read", {"file_path": "~/.claude/settings.json"}, SDK_CWD
)
assert _is_denied(result)
def test_claude_projects_without_tool_results(self):
result = _validate_workspace_path(
"Read", {"file_path": "~/.claude/projects/abc/credentials.json"}, SDK_CWD
)
assert _is_denied(result)
def test_no_path_allowed(self):
"""Glob/Grep without path defaults to cwd — should be allowed."""
result = _validate_workspace_path("Grep", {"pattern": "foo"}, SDK_CWD)
assert not _is_denied(result)
def test_path_traversal_with_dotdot(self):
result = _validate_workspace_path(
"Read", {"file_path": f"{SDK_CWD}/../../../etc/passwd"}, SDK_CWD
)
assert _is_denied(result)
# ============================================================
# Tool access validation
# ============================================================
class TestToolAccessValidation:
def test_blocked_tools(self):
for tool in ("bash", "shell", "exec", "terminal", "command"):
result = _validate_tool_access(tool, {})
assert _is_denied(result), f"Tool '{tool}' should be blocked"
def test_bash_builtin_blocked(self):
"""SDK built-in Bash (capital) is blocked as defence-in-depth."""
result = _validate_tool_access("Bash", {"command": "echo hello"}, SDK_CWD)
assert _is_denied(result)
assert "Bash" in _reason(result)
def test_workspace_tools_delegate(self):
result = _validate_tool_access(
"Read", {"file_path": f"{SDK_CWD}/file.txt"}, SDK_CWD
)
assert not _is_denied(result)
def test_dangerous_pattern_blocked(self):
result = _validate_tool_access("SomeUnknownTool", {"data": "sudo rm -rf /"})
assert _is_denied(result)
def test_safe_unknown_tool_allowed(self):
result = _validate_tool_access("SomeSafeTool", {"data": "hello world"})
assert not _is_denied(result)
# ============================================================
# Deny message quality (ntindle feedback)
# ============================================================
class TestDenyMessageClarity:
"""Deny messages must include [SECURITY] and 'cannot be bypassed'
so the model knows the restriction is enforced, not a suggestion."""
def test_blocked_tool_message(self):
reason = _reason(_validate_tool_access("bash", {}))
assert "[SECURITY]" in reason
assert "cannot be bypassed" in reason
def test_bash_builtin_blocked_message(self):
reason = _reason(_validate_tool_access("Bash", {"command": "echo hello"}))
assert "[SECURITY]" in reason
assert "cannot be bypassed" in reason
def test_workspace_path_message(self):
reason = _reason(
_validate_workspace_path("Read", {"file_path": "/etc/passwd"}, SDK_CWD)
)
assert "[SECURITY]" in reason
assert "cannot be bypassed" in reason