feat(agent-toolkit): agents subgraph tools expansion

docs/superpowers/specs/2026-05-13-agent-tools-expansion-design.md

@@ -0,0 +1,280 @@
+# Agent Tools Expansion — Design Spec
+
+**Date:** 2026-05-13  
+**Package:** `packages/agent-toolkit`  
+**Location:** `src/core/tools/platform-api-tools/agents-tools/`
+
+---
+
+## Problem
+
+The existing agent tools (`get_agent`, `create_agent`, `delete_agent`) cover only basic CRUD. The agents subgraph exposes a much richer surface — skills, triggers, state management, knowledge/resource access — that agents cannot yet use via the toolkit.
+
+The core challenge is discoverability: consumers never know available skill IDs, trigger `block_reference_id` values, or the required `field_values` shape for a given trigger type. The catalog queries exist precisely to solve this, and the tool descriptions must encode the lookup-first workflow so agents know to consult them before acting.
+
+---
+
+## Scope
+
+6 new tools covering all remaining agents subgraph operations:
+
+| Tool | Type | Operations |
+|------|------|-----------|
+| `get_agent_catalog` | READ | `agent_skills_catalog`, `agent_triggers_catalog` |
+| `manage_agent_triggers` | WRITE | `agent_active_triggers` (list), `add_trigger_to_agent`, `remove_trigger_from_agent` |
+| `manage_agent_skills` | WRITE | `add_skill_to_agent`, `remove_skill_from_agent` |
+| `update_agent` | WRITE | `update_agent` |
+| `manage_agent_state` | WRITE | `activate_agent`, `deactivate_agent`, `run_agent` |
+| `manage_agent_knowledge` | WRITE | `agent_knowledge` (list), `add_agent_resource_access`, `remove_agent_resource_access`, `update_agent_resource_access` |
+
+Out of scope: `get_agent` is unchanged — the subgraph will be enhanced in a future iteration to include more fields (active triggers, knowledge) directly on the Agent type.
+
+---
+
+## File Structure
+
+Each tool is co-located with its own `.graphql.dev.ts` file. The `shared/` directory keeps only the `AgentFields` fragment (used by the existing get/create/delete tools).
+
+```
+agents-tools/
+├── shared/
+│   └── agents.graphql.dev.ts                        ← AgentFields fragment only (unchanged)
+├── get-agent-catalog/
+│   ├── get-agent-catalog-tool.ts
+│   └── get-agent-catalog.graphql.dev.ts
+├── manage-agent-triggers/
+│   ├── manage-agent-triggers-tool.ts
+│   └── manage-agent-triggers.graphql.dev.ts
+├── manage-agent-skills/
+│   ├── manage-agent-skills-tool.ts
+│   └── manage-agent-skills.graphql.dev.ts
+├── update-agent/
+│   ├── update-agent-tool.ts
+│   └── update-agent.graphql.dev.ts
+├── manage-agent-state/
+│   ├── manage-agent-state-tool.ts
+│   └── manage-agent-state.graphql.dev.ts
+└── manage-agent-knowledge/
+    ├── manage-agent-knowledge-tool.ts
+    └── manage-agent-knowledge.graphql.dev.ts
+```
+
+---
+
+## Tool Designs
+
+### `get_agent_catalog` (READ)
+
+**Purpose:** Account-wide discovery of available trigger types and skills. Always call this before adding a trigger or skill to an agent.
+
+**Input schema:**
+```typescript
+{
+  type: z.enum(['triggers', 'skills']),
+  // For triggers only — fetches specific entries much faster than full catalog
+  block_reference_ids: z.array(z.string()).optional()
+}
+```
+
+**Behavior:**
+- `type: 'triggers'` → calls `agent_triggers_catalog`. Returns entries with `block_reference_id`, `name`, `description`, `field_schemas`, `required_fields`. `field_schemas` describes the shape of `field_values` required when adding (e.g. `{ board_id: <ID> }`). `required_fields` lists fields the user must supply.
+- `type: 'skills'` → calls `agent_skills_catalog`. Returns entries with `id`, `name`, `description`.
+
+**Description preamble:**
+> Call this tool first before adding a trigger or skill to an agent. For triggers: inspect `field_schemas` and `required_fields` to know what to collect from the user (e.g. board_id, column_id) before calling `manage_agent_triggers`. For skills: use the returned `id` to call `manage_agent_skills`.
+
+---
+
+### `manage_agent_triggers` (WRITE)
+
+**Purpose:** List, add, and remove triggers on a specific agent.
+
+**Input schema (discriminated by `action`):**
+```typescript
+{
+  action: z.enum(['list', 'add', 'remove']),
+  agent_id: z.string(),
+  // add only
+  block_reference_id: z.string().optional(),
+  field_values: z.record(z.unknown()).optional(),
+  // remove only
+  node_id: z.string().optional()
+}
+```
+
+**Behavior:**
+- `list` → calls `agent_active_triggers(agent_id)`. Returns active triggers with `node_id`, `block_reference_id`, `name`, `description`, `field_summary`.
+- `add` → calls `add_trigger_to_agent`. Requires `block_reference_id` (from catalog) and optional `field_values` (shape determined by `field_schemas` in the catalog entry).
+- `remove` → calls `remove_trigger_from_agent`. Requires `node_id` (from `action: list`).
+
+**Description — add workflow:**
+> To add: first call `get_agent_catalog` with `type: triggers` to find the right entry by name/description. Inspect `field_schemas` and `required_fields` to know what information to collect from the user (e.g. which board, which column). Only then call this tool with `action: add`.
+
+**Description — remove workflow:**
+> To remove: first call this tool with `action: list` to see the active triggers by name and `field_summary`. Match the trigger the user described, get its `node_id`, then call this tool with `action: remove`.
+
+---
+
+### `manage_agent_skills` (WRITE)
+
+**Purpose:** Attach and detach skills from an agent.
+
+**Input schema:**
+```typescript
+{
+  action: z.enum(['add', 'remove']),
+  agent_id: z.string(),
+  skill_id: z.string()
+}
+```
+
+**Behavior:**
+- `add` → calls `add_skill_to_agent`
+- `remove` → calls `remove_skill_from_agent`
+
+**Description preamble:**
+> Always call `get_agent_catalog` with `type: skills` first to discover available skills and resolve the correct `skill_id` — never guess or invent a skill ID.
+
+---
+
+### `update_agent` (WRITE)
+
+**Purpose:** Update an agent's profile or execution plan. Creates a new draft internally and publishes in one call.
+
+**Input schema:**
+```typescript
+{
+  id: z.string(),
+  name: z.string().optional(),
+  role: z.string().optional(),
+  role_description: z.string().optional(),
+  plan: z.string().optional(),       // markdown
+  agent_model: z.string().optional() // discourage unless user explicitly named a model
+}
+```
+
+**Behavior:** Calls `update_agent`. All profile fields are optional — only provided fields are changed. Mirrors the same `agent_model` guidance as `create_agent` (strongly discourage setting it; omit unless user explicitly named a model).
+
+---
+
+### `manage_agent_state` (WRITE)
+
+**Purpose:** Activate, deactivate, or manually trigger a run for an agent.
+
+**Input schema:**
+```typescript
+{
+  action: z.enum(['activate', 'deactivate', 'run']),
+  agent_id: z.string(),
+  // deactivate only
+  inactive_reason: z.enum(['DEACTIVATED_BY_USER', 'ACCOUNT_LEVEL_BLOCKING']).optional()
+}
+```
+
+**Behavior:**
+- `activate` → calls `activate_agent`. Agent transitions to ACTIVE and can receive triggers.
+- `deactivate` → calls `deactivate_agent`. `inactive_reason` defaults to `DEACTIVATED_BY_USER` if omitted.
+- `run` → calls `run_agent`. Fire-and-forget async operation. Returns `trigger_uuid` for downstream correlation. Success means the run was enqueued, not that it completed.
+
+---
+
+### `manage_agent_knowledge` (WRITE)
+
+**Purpose:** List, grant, update, and revoke an agent's access to monday.com boards and docs.
+
+**Input schema (discriminated by `action`):**
+```typescript
+{
+  action: z.enum(['list', 'add', 'update', 'remove']),
+  agent_id: z.string(),
+  // add, update, remove
+  resource_id: z.string().optional(),
+  scope_type: z.enum(['BOARD', 'DOC']).optional(),
+  // add, update
+  permission_type: z.enum(['READ', 'READ_WRITE']).optional()
+}
+```
+
+**Behavior:**
+- `list` → calls `agent_knowledge(agent_id)`. Returns current resources (boards/docs) with `resource_id`, `scope_type`, `permission_type`, plus uploaded files.
+- `add` → calls `add_agent_resource_access`. Grants agent access to a board or doc with the specified permission level.
+- `update` → calls `update_agent_resource_access`. Changes the permission level on an already-granted resource.
+- `remove` → calls `remove_agent_resource_access`. Revokes access entirely.
+
+**Description preamble:**
+> Call with `action: list` first to see the agent's current resource access before adding, updating, or removing. For `add`/`update`, `permission_type: READ` allows the agent to read the resource; `READ_WRITE` also allows writing.
+
+---
+
+## Key Design Decisions
+
+**Catalog-first descriptions:** Tool descriptions are the primary mechanism for guiding agent behavior. Every WRITE tool that requires an ID from a catalog or list operation explicitly names the prerequisite call. This is more reliable than trying to enforce it in code.
+
+**`field_values` as record:** `add_trigger_to_agent` takes `field_values: JSON` in the schema. The Zod type is `z.record(z.unknown())` — flexible enough to accommodate any trigger type's shape. The `field_schemas` from the catalog entry describes the expected shape at runtime.
+
+**`manage_agent_triggers` and `manage_agent_knowledge` are typed WRITE:** Both include a `list` action that is read-only in effect, but the tools are typed WRITE because their primary purpose is mutation and the list actions exist specifically to support the write workflow (get `node_id` before remove, inspect resources before update). In `readOnlyMode`, WRITE tools are filtered out — this means `list` is unavailable in read-only configurations, which is an accepted trade-off. If this becomes a problem in practice, the list operations can be split into `get_agent` in a future iteration when the subgraph enhances the Agent type to include active triggers and knowledge directly.
+
+---
+
+## GraphQL Operations (per tool)
+
+### `get-agent-catalog.graphql.dev.ts`
+- `query getAgentTriggersCatalog($block_reference_ids: [ID!])`
+- `query getAgentSkillsCatalog`
+
+### `manage-agent-triggers.graphql.dev.ts`
+- `query getAgentActiveTriggers($agent_id: ID!)`
+- `mutation addTriggerToAgent($agent_id: ID!, $block_reference_id: ID!, $field_values: JSON)`
+- `mutation removeTriggerFromAgent($agent_id: ID!, $node_id: ID!)`
+
+### `manage-agent-skills.graphql.dev.ts`
+- `mutation addSkillToAgent($agent_id: ID!, $skill_id: ID!)`
+- `mutation removeSkillFromAgent($agent_id: ID!, $skill_id: ID!)`
+
+### `update-agent.graphql.dev.ts`
+- `mutation updateAgent($id: ID!, $input: UpdateAgentInput!)`
+
+### `manage-agent-state.graphql.dev.ts`
+- `mutation activateAgent($id: ID!)`
+- `mutation deactivateAgent($id: ID!, $inactive_reason: InactiveReason)`
+- `mutation runAgent($id: ID!)`
+
+### `manage-agent-knowledge.graphql.dev.ts`
+- `query getAgentKnowledge($agent_id: ID!)`
+- `mutation addAgentResourceAccess($id: ID!, $resource_id: ID!, $scope_type: KnowledgeScope!, $permission_type: KnowledgePermission!)`
+- `mutation removeAgentResourceAccess($id: ID!, $resource_id: ID!, $scope_type: KnowledgeScope!)`
+- `mutation updateAgentResourceAccess($id: ID!, $resource_id: ID!, $scope_type: KnowledgeScope!, $permission_type: KnowledgePermission!)`
+
+---
+
+## Registration
+
+In `src/core/tools/platform-api-tools/index.ts`:
+- 6 new imports
+- 6 additions to `allGraphqlApiTools` array (grouped under existing agents comment)
+- 6 new `export *` lines
+
+---
+
+## Error Handling
+
+All tools use `rethrowWithContext(error, '<operation name>')` in a try/catch, consistent with existing agent tools.
+
+---
+
+## Testing
+
+One `.test.ts` per tool, co-located, using `createMockApiClient()`. Each test file covers:
+- Happy path for each action
+- Validation errors (missing required fields for wrong action)
+- API error propagation via `rethrowWithContext`
+
+---
+
+## Post-Implementation
+
+After all tool files are written:
+1. Run `npm run fetch:generate dev` to regenerate dev types from new GraphQL operations
+2. Run `npm test` to verify all tests pass
+3. Run `npm run build` to verify compilation
+4. Bump version in `package.json`

packages/agent-toolkit/CHANGELOG.md

@@ -1,5 +1,16 @@
 # Changelog
 
+## 5.11.0
+
+### Add agent management tools
+
+Four new tools enabling agents to create and manage monday.com platform agents end-to-end:
+
+- `manage_agent` — full lifecycle management: create (AI or manual mode), get, update, delete, activate, deactivate, and manually run an agent
+- `manage_agent_capabilities` — attach and detach triggers (when an agent runs) and skills (what it can do) on a specific agent
+- `manage_agent_knowledge` — grant, update, or revoke an agent's access to boards and docs
+- `agent_catalog` — browse the account-wide catalog of available trigger types and skills; create new custom skills
+
 ## 5.7.1
 
 ### form_questions_editor — fix ConditionOperator and remove existing_column_id

packages/agent-toolkit/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mondaydotcomorg/agent-toolkit",
-  "version": "5.10.3",
+  "version": "5.11.0",
 
 
   "description": "monday.com agent toolkit",