feat(cloud): research agent hardening

products/tasks/backend/models.py

@@ -165,6 +165,7 @@ def create_and_run(
         start_workflow: bool = True,
         posthog_mcp_scopes: PosthogMcpScopes = "full",
         branch: str | None = None,
+        tools_preset: str = "default",
     ) -> "Task":
         from products.tasks.backend.temporal.client import execute_task_processing_workflow
 
@@ -204,6 +205,7 @@ def create_and_run(
                 create_pr=create_pr,
                 slack_thread_context=slack_thread_context,
                 posthog_mcp_scopes=posthog_mcp_scopes,
+                tools_preset=tools_preset,
             )
 
         return task

products/tasks/backend/services/docker_sandbox.py

@@ -563,16 +563,20 @@ def _build_agent_server_command(
         interaction_origin: str | None = None,
         branch: str | None = None,
         mcp_servers_arg: str = "",
+        tools_preset: str | None = None,
     ) -> str:
         env_prefix = (
             f"env POSTHOG_CODE_INTERACTION_ORIGIN={shlex.quote(interaction_origin)} " if interaction_origin else ""
         )
         branch_flag = f" --baseBranch {shlex.quote(branch)}" if branch else ""
+        tools_preset_flag = (
+            f" --toolsPreset {shlex.quote(tools_preset)}" if tools_preset and tools_preset != "default" else ""
+        )
         return (
             f"cd /scripts && "
             f"nohup {env_prefix}./node_modules/.bin/agent-server --port {AGENT_SERVER_PORT} --repositoryPath {shlex.quote(repo_path)} "
             f"--taskId {shlex.quote(task_id)} --runId {shlex.quote(run_id)} --mode {shlex.quote(mode)}"
-            f"{branch_flag}{mcp_servers_arg} "
+            f"{branch_flag}{mcp_servers_arg}{tools_preset_flag} "
             f"> /tmp/agent-server.log 2>&1 &"
         )
 
@@ -596,6 +600,7 @@ def start_agent_server(
         interaction_origin: str | None = None,
         branch: str | None = None,
         mcp_configs: list[McpServerConfig] | None = None,
+        tools_preset: str | None = None,
     ) -> None:
         """Start the agent-server HTTP server in the sandbox.
 
@@ -617,7 +622,7 @@ def start_agent_server(
             mcp_servers_arg = f" --mcpServers {shlex.quote(mcp_json)}"
 
         command = self._build_agent_server_command(
-            repo_path, task_id, run_id, mode, interaction_origin, branch, mcp_servers_arg
+            repo_path, task_id, run_id, mode, interaction_origin, branch, mcp_servers_arg, tools_preset
         )
 
         logger.info(f"Starting agent-server in sandbox {self.id} for {repository}")
@@ -637,7 +642,14 @@ def start_agent_server(
             self.execute("pkill -f agent-server || true", timeout_seconds=5)
 
             command = self._build_agent_server_command(
-                repo_path, task_id, run_id, mode, interaction_origin, branch=None, mcp_servers_arg=mcp_servers_arg
+                repo_path,
+                task_id,
+                run_id,
+                mode,
+                interaction_origin,
+                branch=None,
+                mcp_servers_arg=mcp_servers_arg,
+                tools_preset=tools_preset,
             )
             if self._launch_and_check(command):
                 logger.info(f"Agent-server started on port {self._host_port} (without --baseBranch)")

products/tasks/backend/services/modal_sandbox.py

@@ -445,6 +445,7 @@ def start_agent_server(
         interaction_origin: str | None = None,
         branch: str | None = None,
         mcp_configs: list[McpServerConfig] | None = None,
+        tools_preset: str | None = None,
     ) -> None:
         """Start the agent-server HTTP server in the sandbox.
 
@@ -467,11 +468,14 @@ def start_agent_server(
             f"env POSTHOG_CODE_INTERACTION_ORIGIN={shlex.quote(interaction_origin)} " if interaction_origin else ""
         )
         branch_flag = f" --baseBranch {shlex.quote(branch)}" if branch else ""
+        tools_preset_flag = (
+            f" --toolsPreset {shlex.quote(tools_preset)}" if tools_preset and tools_preset != "default" else ""
+        )
         command = (
             f"cd /scripts && "
             f"nohup {env_prefix}./node_modules/.bin/agent-server --port {AGENT_SERVER_PORT} --repositoryPath {shlex.quote(repo_path)} "
             f"--taskId {shlex.quote(task_id)} --runId {shlex.quote(run_id)} --mode {shlex.quote(mode)}"
-            f"{branch_flag}{mcp_servers_arg} "
+            f"{branch_flag}{mcp_servers_arg}{tools_preset_flag} "
             f"> /tmp/agent-server.log 2>&1 &"
         )
 

products/tasks/backend/services/sandbox.py

@@ -125,6 +125,7 @@ def start_agent_server(
         interaction_origin: str | None = None,
         branch: str | None = None,
         mcp_configs: list[McpServerConfig] | None = None,
+        tools_preset: str | None = None,
     ) -> None:
         """Start the agent-server HTTP server in the sandbox.
 

products/tasks/backend/temporal/client.py

@@ -41,6 +41,7 @@ async def execute_task_processing_workflow_async(
     slack_thread_context: Optional[Any] = None,
     skip_user_check: bool = False,
     posthog_mcp_scopes: PosthogMcpScopes = "read_only",
+    tools_preset: str = "default",
 ) -> None:
     """
     Start the task processing workflow asynchronously. Fire-and-forget.
@@ -101,6 +102,7 @@ async def execute_task_processing_workflow_async(
             create_pr=create_pr,
             slack_thread_context=slack_context_dict,
             posthog_mcp_scopes=posthog_mcp_scopes,
+            tools_preset=tools_preset,
         )
 
         logger.info(
@@ -141,6 +143,7 @@ def execute_task_processing_workflow(
     slack_thread_context: Optional["SlackThreadContext"] = None,
     skip_user_check: bool = False,
     posthog_mcp_scopes: PosthogMcpScopes = "read_only",
+    tools_preset: str = "default",
 ) -> None:
     """
     Start the task processing workflow synchronously. Fire-and-forget.
@@ -200,6 +203,7 @@ def execute_task_processing_workflow(
             create_pr=create_pr,
             slack_thread_context=slack_context_dict,
             posthog_mcp_scopes=posthog_mcp_scopes,
+            tools_preset=tools_preset,
         )
 
         logger.info(

products/tasks/backend/temporal/process_task/activities/start_agent_server.py

@@ -25,6 +25,7 @@ class StartAgentServerInput:
     sandbox_url: str
     sandbox_connect_token: str | None = None
     posthog_mcp_scopes: PosthogMcpScopes = "read_only"
+    tools_preset: str = "default"
 
 
 @dataclass
@@ -84,6 +85,7 @@ def start_agent_server(input: StartAgentServerInput) -> StartAgentServerOutput:
                 interaction_origin=ctx.interaction_origin,
                 branch=ctx.branch,
                 mcp_configs=mcp_configs or None,
+                tools_preset=input.tools_preset if input.tools_preset != "default" else None,
             )
         except Exception as e:
             raise SandboxExecutionError(

products/tasks/backend/temporal/process_task/workflow.py

@@ -45,6 +45,7 @@ class ProcessTaskInput:
     create_pr: bool = True
     slack_thread_context: Optional[dict[str, Any]] = None
     posthog_mcp_scopes: PosthogMcpScopes = "read_only"
+    tools_preset: str = "default"
 
 
 @dataclass
@@ -65,6 +66,7 @@ def __init__(self) -> None:
         self._context: Optional[TaskProcessingContext] = None
         self._slack_thread_context: Optional[dict[str, Any]] = None
         self._posthog_mcp_scopes: PosthogMcpScopes = "read_only"
+        self._tools_preset: str = "default"
         self._task_completed: bool = False
         self._completion_status: str = "completed"
         self._completion_error: Optional[str] = None
@@ -85,6 +87,7 @@ def parse_inputs(inputs: list[str]) -> ProcessTaskInput:
             create_pr=loaded.get("create_pr", True),
             slack_thread_context=loaded.get("slack_thread_context"),
             posthog_mcp_scopes=loaded.get("posthog_mcp_scopes", "read_only"),
+            tools_preset=loaded.get("tools_preset", "default"),
         )
 
     @temporalio.workflow.run
@@ -98,6 +101,10 @@ async def run(self, input: ProcessTaskInput) -> ProcessTaskOutput:
         try:
             self._context = await self._get_task_processing_context(input)
             self._posthog_mcp_scopes = input.posthog_mcp_scopes
+            self._tools_preset = input.tools_preset
+            # Force read-only MCP scopes for research_background_agent
+            if self._tools_preset == "research_background_agent":
+                self._posthog_mcp_scopes = "read_only"
             await self._update_task_run_status("in_progress")
 
             await self._track_workflow_event(
@@ -280,6 +287,7 @@ async def _start_agent_server(self, sandbox_output: GetSandboxForRepositoryOutpu
                 sandbox_url=sandbox_output.sandbox_url,
                 sandbox_connect_token=sandbox_output.connect_token,
                 posthog_mcp_scopes=self._posthog_mcp_scopes,
+                tools_preset=self._tools_preset,
             ),
             start_to_close_timeout=timedelta(minutes=5),
             retry_policy=RetryPolicy(maximum_attempts=3),