Skip to content

Self-improving AI Employees for everyone

Fix vertexCloudTool panel schema + add vertex agent skills

trohitgf1c2c01b8bfe
diff --git a/server/nodes/agent/vertex_cloud_tool/__init__.py b/server/nodes/agent/vertex_cloud_tool/__init__.py
index 2bc3c3c7..4c4c6578 100644
--- a/server/nodes/agent/vertex_cloud_tool/__init__.py
+++ b/server/nodes/agent/vertex_cloud_tool/__init__.py
@@ -7,15 +7,48 @@ runs entirely in Google's cloud, so these nodes are the canvas-visible
trace of that remote activity: one node per distinct tool, pulsed
executing->success after each turn that used it.
-Pure display: default ``_EmptyParams`` / ``_EmptyOutput`` and no
-``@Operation`` (the plugin contract explicitly exempts pure-display
-plugins). It is never scheduled by the executor — nodes wired to an
-agent's ``input-tools`` handle are excluded as sub-nodes.
+Display-only, but it declares a real ``Params`` model: the frontend
+fetches every node type's input schema for the parameter panel, and
+plugins that stay on the default ``_EmptyParams`` are deliberately NOT
+registered into ``NODE_INPUT_MODELS`` (``BaseNode.__init_subclass__``
+passes ``input_model=None``), which logs a "No plugin Params
+registered" warning and leaves the panel schema-less. The fields mirror
+what ``_ops.ensure_cloud_tool_nodes`` persists per minted node. The
+single ``info`` operation just echoes them — the run button is hidden
+and the executor never schedules this node anyway (nodes wired to an
+agent's ``input-tools`` handle are excluded as sub-nodes).
"""
from __future__ import annotations
-from services.plugin import ActionNode
+from typing import Optional
+
+from pydantic import BaseModel, ConfigDict, Field
+
+from services.plugin import ActionNode, NodeContext, Operation, TaskQueue
+
+
+class VertexCloudToolParams(BaseModel):
+ """Set by the minting helper — not meant to be edited by hand."""
+
+ cloud_tool_key: Optional[str] = Field(
+ default=None,
+ description="Stable key of the cloud-side tool (type:... or fn:...).",
+ )
+ label: Optional[str] = Field(
+ default=None,
+ description="Display label of the cloud-side tool.",
+ )
+
+ model_config = ConfigDict(extra="ignore")
+
+
+class VertexCloudToolOutput(BaseModel):
+ cloud_tool_key: Optional[str] = None
+ label: Optional[str] = None
+ message: Optional[str] = None
+
+ model_config = ConfigDict(extra="allow")
class VertexCloudToolNode(ActionNode):
@@ -44,3 +77,23 @@ class VertexCloudToolNode(ActionNode):
},
)
annotations = {"destructive": False, "readonly": True, "open_world": False}
+ task_queue = TaskQueue.REST_API
+
+ Params = VertexCloudToolParams
+ Output = VertexCloudToolOutput
+
+ @Operation("info")
+ async def info_op(
+ self,
+ ctx: NodeContext,
+ params: VertexCloudToolParams,
+ ) -> VertexCloudToolOutput:
+ """Echo the display metadata (this node performs no work)."""
+ return VertexCloudToolOutput(
+ cloud_tool_key=params.cloud_tool_key,
+ label=params.label,
+ message=(
+ "Display node: shows a cloud-side tool used by the Vertex "
+ "managed agent."
+ ),
+ )
diff --git a/server/nodes/visuals.json b/server/nodes/visuals.json
index e391d42a..c49b9982 100644
--- a/server/nodes/visuals.json
+++ b/server/nodes/visuals.json
@@ -343,10 +343,12 @@
"skill": "stripe-skill"
},
"vertex_managed_agent": {
- "icon": "lobehub:vertexai"
+ "icon": "lobehub:vertexai",
+ "skill": "vertex-agent-skill"
},
"vertex_agent_admin": {
- "icon": "lobehub:googlecloud"
+ "icon": "lobehub:googlecloud",
+ "skill": "vertex-agent-admin-skill"
},
"vertexCloudTool": {
"icon": "lobehub:gemini"
diff --git a/server/skills/vertex_agent/vertex-agent-admin-skill/SKILL.md b/server/skills/vertex_agent/vertex-agent-admin-skill/SKILL.md
new file mode 100644
index 00000000..79ae7bc0
--- /dev/null
+++ b/server/skills/vertex_agent/vertex-agent-admin-skill/SKILL.md
@@ -0,0 +1,50 @@
+---
+name: vertex-agent-admin-skill
+description: Create, list, inspect, and delete custom managed agents on the Gemini Enterprise Agent Platform. Custom agents build on the Antigravity base agent with their own system instruction and built-in cloud tools; run them with the Vertex Agent node.
+allowed-tools: "vertex_agent_admin"
+metadata:
+ author: machina
+ version: "1.0"
+ category: agents
+
+---
+
+# Vertex Agent Admin Skill
+
+Lifecycle management for **custom managed agents** on the Gemini
+Enterprise Agent Platform (Agents API). A custom agent is the prebuilt
+Antigravity agent plus your own system instruction and a chosen set of
+built-in cloud tools.
+
+## Node: Vertex Agent Admin
+
+### Operations
+
+| Operation | Purpose | Key fields |
+|---|---|---|
+| `create` | Create a custom agent (long-running: ~2-3 min the first time, seconds after) | `agent_id`, `description`, `system_instruction`, `base_agent`, `tools` |
+| `list` | List custom agents in the project | — |
+| `get` | Inspect one agent's config | `agent_id` |
+| `delete` | Delete a custom agent (destructive, irreversible) | `agent_id` |
+
+### Fields
+
+- `agent_id` — 1-63 chars, lowercase letters / numbers / hyphens, must
+ start with a letter. Example: `research-assistant-v1`.
+- `tools` — built-in cloud tools the agent may use: `code_execution`,
+ `filesystem`, `google_search`, `url_context`.
+- `project_id` — GCP project (auth via gcloud Application Default
+ Credentials). Leave empty to use a stored `AIza` Gemini API key.
+
+### Response
+
+`create` / `get` return the agent config under `agent`; `list` returns
+`agents` + `count`; `delete` returns `deleted: true`.
+
+## Using a custom agent
+
+After `create`, put the `agent_id` into a **Vertex Agent** node's
+`agent` parameter — the interaction node runs it with the same memory
+continuity and tool bridging as the prebuilt agent. Deleting an agent
+does not delete past interactions or sandboxes (those expire on their
+own 7-day TTL).
diff --git a/server/skills/vertex_agent/vertex-agent-skill/SKILL.md b/server/skills/vertex_agent/vertex-agent-skill/SKILL.md
new file mode 100644
index 00000000..a3183607
--- /dev/null
+++ b/server/skills/vertex_agent/vertex-agent-skill/SKILL.md
@@ -0,0 +1,53 @@
+---
+name: vertex-agent-skill
+description: Delegate heavy tasks to a Google cloud-hosted managed agent (Antigravity) with its own sandboxed Linux environment - code execution, file management, Google Search, and URL fetching all run in Google's cloud. Use for multi-step research or compute jobs that should not run locally.
+allowed-tools: "vertex_managed_agent"
+metadata:
+ author: machina
+ version: "1.0"
+ category: agents
+
+---
+
+# Vertex Managed Agent Skill
+
+Delegate work to Google's **managed agent** (the Antigravity harness on
+the Gemini Enterprise Agent Platform). The agent runs entirely in
+Google's cloud inside an isolated Linux sandbox with Python 3.11,
+Node.js, bash, file management, Google Search, and URL fetching.
+
+## Tool: delegate_to_vertex_managed_agent
+
+Non-blocking delegation. Two fields:
+
+| Field | Purpose |
+|---|---|
+| `task` | What the cloud agent should accomplish — its system instruction for the run |
+| `context` | Optional background, inputs, or data the agent needs |
+
+The call returns immediately with `{"status": "delegated", "task_id": ...}`.
+Use `check_delegated_tasks` to poll for the result — do NOT re-delegate
+the same task while one is in flight.
+
+## When to delegate to the cloud agent
+
+- Multi-step research: search the web, fetch pages, synthesize a report.
+- Compute jobs: run Python/Node code, process data, generate files —
+ all inside the cloud sandbox, not on this machine.
+- Long-running work you want isolated from the local workflow.
+
+Prefer LOCAL tools (python_executor, shell, file tools) when the work
+needs files in this workflow's workspace — the cloud sandbox cannot see
+local files unless their content is passed in `context`.
+
+## Behavior notes
+
+- The cloud agent keeps its own multi-turn state: when a simpleMemory
+ node is connected to the Vertex Agent, conversation and sandbox
+ (installed packages, created files) persist across runs for up to
+ 7 days.
+- Cloud-side tool usage (sandbox commands, searches) appears on the
+ canvas as Cloud Tool nodes wired to the Vertex Agent — they are
+ display-only indicators, safe to delete.
+- Task descriptions should be self-contained: the cloud agent cannot
+ ask you follow-up questions mid-run.