feat(agent): Sandbox / CLI Agent (dify.shell) + read-only sandbox file inspector (#36984)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>

api/clients/agent_backend/request_builder.py

@@ -30,6 +30,7 @@ from dify_agent.layers.execution_context import (
     DifyExecutionContextLayerConfig,
 )
 from dify_agent.layers.output import DIFY_OUTPUT_LAYER_TYPE_ID, DifyOutputLayerConfig
+from dify_agent.layers.shell import DIFY_SHELL_LAYER_TYPE_ID, DifyShellLayerConfig
 from dify_agent.protocol import (
     DIFY_AGENT_HISTORY_LAYER_ID,
     DIFY_AGENT_MODEL_LAYER_ID,
@@ -48,6 +49,7 @@ WORKFLOW_USER_PROMPT_LAYER_ID = "workflow_user_prompt"
 AGENT_APP_USER_PROMPT_LAYER_ID = "agent_app_user_prompt"
 DIFY_EXECUTION_CONTEXT_LAYER_ID = "execution_context"
 DIFY_PLUGIN_TOOLS_LAYER_ID = "tools"
+DIFY_SHELL_LAYER_ID = "shell"
 
 # Layer types that hold credentials in their per-run config. These are excluded
 # from the cleanup-replay composition (and from the snapshot that is sent with
@@ -167,6 +169,10 @@ class AgentBackendWorkflowNodeRunInput(BaseModel):
     idempotency_key: str | None = None
     output: AgentBackendOutputConfig | None = None
     tools: DifyPluginToolsLayerConfig | None = None
+    # Inject the sandboxed shell layer (dify.shell). Requires the agent backend
+    # to be wired with a shellctl entrypoint; see configs AGENT_SHELL_ENABLED.
+    include_shell: bool = False
+    shell_config: DifyShellLayerConfig | None = None
     session_snapshot: CompositorSessionSnapshot | None = None
     include_history: bool = True
     suspend_on_exit: bool = True
@@ -199,6 +205,10 @@ class AgentBackendAgentAppRunInput(BaseModel):
     idempotency_key: str | None = None
     output: AgentBackendOutputConfig | None = None
     tools: DifyPluginToolsLayerConfig | None = None
+    # Inject the sandboxed shell layer (dify.shell). Requires the agent backend
+    # to be wired with a shellctl entrypoint; see configs AGENT_SHELL_ENABLED.
+    include_shell: bool = False
+    shell_config: DifyShellLayerConfig | None = None
     session_snapshot: CompositorSessionSnapshot | None = None
     include_history: bool = True
     suspend_on_exit: bool = True
@@ -289,6 +299,18 @@ class AgentBackendRunRequestBuilder:
                 )
             )
 
+        if run_input.include_shell:
+            # Sandboxed bash workspace (dify.shell). The layer declares NoLayerDeps,
+            # so the spec carries no deps; shellctl connection is server-injected.
+            layers.append(
+                RunLayerSpec(
+                    name=DIFY_SHELL_LAYER_ID,
+                    type=DIFY_SHELL_LAYER_TYPE_ID,
+                    metadata=run_input.metadata,
+                    config=run_input.shell_config or DifyShellLayerConfig(),
+                )
+            )
+
         if run_input.output is not None:
             layers.append(
                 RunLayerSpec(
@@ -432,6 +454,18 @@ class AgentBackendRunRequestBuilder:
                 )
             )
 
+        if run_input.include_shell:
+            # Sandboxed bash workspace (dify.shell). The layer declares NoLayerDeps,
+            # so the spec carries no deps; shellctl connection is server-injected.
+            layers.append(
+                RunLayerSpec(
+                    name=DIFY_SHELL_LAYER_ID,
+                    type=DIFY_SHELL_LAYER_TYPE_ID,
+                    metadata=run_input.metadata,
+                    config=run_input.shell_config or DifyShellLayerConfig(),
+                )
+            )
+
         if run_input.output is not None:
             layers.append(
                 RunLayerSpec(

api/clients/agent_backend/workspace_files_client.py

@@ -0,0 +1,135 @@
+"""API-side client for the agent backend's read-only workspace file endpoints.
+
+The agent backend exposes ``/workspaces/{session_id}/files{,/preview,/download}``
+to inspect a shell-layer sandbox workspace. This thin synchronous client proxies
+those reads for the console FS inspector and normalizes transport/HTTP failures
+into the API backend's ``AgentBackendError`` boundary, preserving the backend's
+status code and ``{code, message}`` detail so the controller can relay them.
+"""
+
+from __future__ import annotations
+
+import base64
+import binascii
+from dataclasses import dataclass
+from typing import Literal
+
+import httpx
+from pydantic import BaseModel
+
+from clients.agent_backend.errors import AgentBackendHTTPError, AgentBackendTransportError
+
+_DEFAULT_TIMEOUT_SECONDS = 30.0
+
+
+class WorkspaceFileEntry(BaseModel):
+    """One entry in a workspace directory listing."""
+
+    name: str
+    type: Literal["file", "dir", "symlink"]
+    size: int
+    mtime: int
+
+
+class WorkspaceListResult(BaseModel):
+    """Directory listing of a workspace path."""
+
+    path: str
+    entries: list[WorkspaceFileEntry]
+    truncated: bool
+
+
+class WorkspacePreviewResult(BaseModel):
+    """Inline preview of a workspace file."""
+
+    path: str
+    size: int
+    truncated: bool
+    binary: bool
+    text: str | None = None
+
+
+@dataclass(frozen=True, slots=True)
+class WorkspaceDownloadResult:
+    """Decoded bytes of a workspace file for download."""
+
+    path: str
+    size: int
+    truncated: bool
+    content: bytes
+
+
+class WorkspaceFilesBackendClient:
+    """Synchronous proxy to the agent backend workspace file endpoints."""
+
+    def __init__(
+        self,
+        base_url: str,
+        *,
+        timeout: float = _DEFAULT_TIMEOUT_SECONDS,
+        transport: httpx.BaseTransport | None = None,
+    ) -> None:
+        self._base_url = base_url.rstrip("/")
+        self._timeout = timeout
+        self._transport = transport
+
+    def list_files(self, session_id: str, path: str) -> WorkspaceListResult:
+        data = self._get(f"/workspaces/{session_id}/files", params={"path": path})
+        return WorkspaceListResult.model_validate(data)
+
+    def preview(self, session_id: str, path: str) -> WorkspacePreviewResult:
+        data = self._get(f"/workspaces/{session_id}/files/preview", params={"path": path})
+        return WorkspacePreviewResult.model_validate(data)
+
+    def download(self, session_id: str, path: str) -> WorkspaceDownloadResult:
+        data = self._get(f"/workspaces/{session_id}/files/download", params={"path": path})
+        encoded = data.get("content_base64")
+        if not isinstance(encoded, str):
+            raise AgentBackendHTTPError("agent backend download response missing content", status_code=502, detail=data)
+        try:
+            content = base64.b64decode(encoded, validate=True)
+        except (binascii.Error, ValueError) as exc:
+            raise AgentBackendHTTPError(
+                "agent backend returned undecodable download content", status_code=502, detail=str(exc)
+            ) from exc
+        size = data.get("size")
+        return WorkspaceDownloadResult(
+            path=str(data.get("path", path)),
+            size=int(size) if isinstance(size, (int, float)) else len(content),
+            truncated=bool(data.get("truncated")),
+            content=content,
+        )
+
+    def _get(self, route: str, *, params: dict[str, str]) -> dict[str, object]:
+        url = f"{self._base_url}{route}"
+        try:
+            with httpx.Client(timeout=self._timeout, transport=self._transport, trust_env=False) as client:
+                response = client.get(url, params=params)
+        except httpx.HTTPError as exc:
+            raise AgentBackendTransportError(f"failed to reach agent backend workspace endpoint: {exc}") from exc
+        if response.status_code >= 400:
+            detail: object
+            try:
+                detail = response.json().get("detail", response.text)
+            except ValueError:
+                detail = response.text
+            raise AgentBackendHTTPError(
+                f"agent backend workspace request failed ({response.status_code})",
+                status_code=response.status_code,
+                detail=detail,
+            )
+        body = response.json()
+        if not isinstance(body, dict):
+            raise AgentBackendHTTPError(
+                "agent backend workspace response was not an object", status_code=502, detail=body
+            )
+        return body
+
+
+__all__ = [
+    "WorkspaceDownloadResult",
+    "WorkspaceFileEntry",
+    "WorkspaceFilesBackendClient",
+    "WorkspaceListResult",
+    "WorkspacePreviewResult",
+]