fix(skills): prevent create-workflow from eagerly checking integrations

Simplify MCP sandbox instructions to avoid prompting the model to fetch
integration guides and check connections during the design phase. Templates,
integration gate, and skill guide table removed from MCP context (already
in skill guides). Added "stay in design mode" guard to create-workflow skill.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: cevian

packages/core/src/cli/mcp/sandbox-server.ts

@@ -59,45 +59,20 @@ const CRAYON_CONTEXT = `## Crayon Workflow Engine
 
 You have access to crayon MCP tools for building and running AI-native workflows.
 
-### Available MCP Tools
-
-**Workflow execution:**
-- \`list_workflows\` — discover compiled workflows
-- \`run_workflow\` — execute a workflow with JSON input
-- \`run_node\` — execute a single node (for testing)
-
-**Run history:**
-- \`list_runs\` — list past executions
-- \`get_run\` — get details of a specific run
-- \`get_trace\` — full execution trace with step-by-step output
-
-**Integrations:**
-- \`list_integrations\` — list available OAuth integrations (Salesforce, Slack, etc.)
-- \`get_connection_info\` — get credentials for a specific integration + node
-
-**Scaffolding:**
-- \`create_app\` — scaffold a new crayon project
-- \`create_database\` — set up a database
-- \`setup_app_schema\` — initialize crayon tables in existing DB
-
-**Skill guides:**
-- \`get_skill_guide\` — detailed procedural guides for workflow development
-
 ### Workflow Development Pipeline
 
-1. **Design** — create \`src/crayon/workflows/<name>.ts\` with a \`description\` field that captures the flow (task ordering, conditions, loops)
-2. **Create stubs** — for each task, create node stubs in \`src/crayon/nodes/<name>.ts\` or agent stubs in \`src/crayon/agents/<name>.ts\` + \`<name>.md\`
-3. **Refine** — add typed Zod schemas, tools, implementation details to each node/agent
-4. **Compile** — update the workflow's \`run()\` method from embedded descriptions
+1. **Design** (\`create-workflow\`) — write workflow + node stubs with \`description\` fields capturing intent
+2. **Refine** (\`refine-node\`) — add typed Zod schemas, tools, SDK setup, check connections
+3. **Compile** (\`compile-workflow\`) — regenerate the workflow's \`run()\` method from descriptions
+
+**Each phase has a skill guide.** Call \`get_skill_guide\` with the phase name (e.g., \`create-workflow\`) before starting. The guides contain all templates, connection gates, and step-by-step instructions. Follow them — do not improvise steps from other phases.
 
 ### Key Patterns
 
-- **Description-driven:** The \`description\` field in workflows and nodes is the source of truth. It drives code generation.
-- **Node types:** \`Node.create()\` for deterministic functions, \`Agent.create()\` for AI reasoning (uses Vercel AI SDK)
-- **Agent specs:** Each agent has a colocated \`.md\` file with system prompt, guidelines, and output format
-- **Integrations:** Declare in \`integrations: ["salesforce", "openai"]\` array. Credentials fetched at runtime via \`ctx.getConnection()\`.
-- **Draft first, ask later:** Make your best guess and let the user correct, rather than interrogating upfront.
-- **Run it yourself:** When the user wants to test, use \`run_workflow\` / \`run_node\` tools directly.
+- **Description-driven:** The \`description\` field in workflows and nodes is the source of truth
+- **Node types:** \`Node.create()\` for deterministic functions, \`Agent.create()\` for AI reasoning
+- **Draft first, ask later:** Make your best guess and let the user correct
+- **Run it yourself:** When the user wants to test, use \`run_workflow\` / \`run_node\` tools directly
 
 ### File Locations
 
@@ -106,100 +81,7 @@ You have access to crayon MCP tools for building and running AI-native workflows
 | Workflows | \`src/crayon/workflows/*.ts\` |
 | Nodes | \`src/crayon/nodes/*.ts\` |
 | Agents | \`src/crayon/agents/*.ts\` + \`*.md\` (colocated spec) |
-| Agent tools | \`src/crayon/tools/*.ts\` |
-| Integrations | \`src/crayon/integrations/*.ts\` |
-
-### Workflow Scaffold Template
-
-\`\`\`typescript
-import { z } from "zod";
-import { Workflow } from "runcrayon";
-
-const InputSchema = z.object({ /* ... */ });
-type Input = z.infer<typeof InputSchema>;
-const OutputSchema = z.object({ /* ... */ });
-type Output = z.infer<typeof OutputSchema>;
-
-export const myWorkflow = Workflow.create({
-  name: "my-workflow",
-  version: 1,
-  description: \\\`
-Summary of what the workflow does.
-
-## Tasks
-
-### 1. Task Name
-**Node:** \\\\\\\`node-name\\\\\\\` (node|agent)
-\\\`,
-  inputSchema: InputSchema,
-  outputSchema: OutputSchema,
-  async run(ctx, inputs: Input): Promise<Output> {
-    const result = await ctx.run(myNode, { /* inputs */ });
-    return { /* outputs */ };
-  },
-});
-\`\`\`
-
-### Node Stub Template
-
-\`\`\`typescript
-import { z } from "zod";
-import { Node } from "runcrayon";
-
-export const myNode = Node.create({
-  name: "my-node",
-  description: \\\`What this node does.
-**Input Description:** what it needs
-**Output Description:** what it produces\\\`,
-  inputSchema: z.object({}),
-  outputSchema: z.object({}),
-  async execute(ctx, input) {
-    return {};
-  },
-});
-\`\`\`
-
-### Agent Stub Template
-
-\`\`\`typescript
-import { z } from "zod";
-import { Agent } from "runcrayon";
-import { fileURLToPath } from "url";
-import path from "path";
-const __dirname = path.dirname(fileURLToPath(import.meta.url));
-
-export const myAgent = Agent.create({
-  name: "my-agent",
-  integrations: ["openai"],
-  description: \\\`What this agent does.
-**Input Description:** what it needs
-**Output Description:** what it produces\\\`,
-  inputSchema: z.object({}),
-  outputSchema: z.object({}),
-  tools: {},
-  specPath: path.resolve(__dirname, "./my-agent.md"),
-});
-\`\`\`
-
-### Integration Gate
-
-Before implementing nodes that use external services, call \`get_connection_info\` to verify the connection exists. If it fails, tell the user to set up the connection in the Dev UI first.
-
-### Skill Guides
-
-Detailed procedural guides are available for complex tasks. Call \`get_skill_guide\` to load them:
-
-| Guide | When to fetch |
-|-------|---------------|
-| \`create-workflow\` | Designing a new workflow from scratch |
-| \`compile-workflow\` | Updating a workflow's run() from descriptions |
-| \`refine-node\` | Adding schemas, tools, implementation to nodes |
-| \`integrations\` | Index of integration setup guides |
-| \`integrations/salesforce\` | Salesforce GraphQL integration setup |
-| \`integrations/postgres\` | PostgreSQL query integration setup |
-| \`integrations/unlisted\` | Custom integration for unlisted systems |
-
-**Always fetch the relevant guide before starting a complex workflow task.** The guides contain critical steps, connection gates, and templates that ensure correct implementation.`;
+| Integrations | \`src/crayon/integrations/*.ts\` |`;
 
 /**
  * Start the sandbox MCP server in stdio mode.

skills/create-workflow/SKILL.md

@@ -38,6 +38,8 @@ Then stop. Do not proceed to any subsequent steps.
 
 Once the tool succeeds, the returned integration IDs are what nodes declare in their `integrations: [...]` arrays.
 
+**IMPORTANT — stay in design mode:** Do NOT read integration-specific guides (e.g., `integrations/salesforce`), call `get_connection_info`, or research SDKs/libraries during this phase. You only need the integration IDs for node stub `integrations: [...]` arrays. All integration research and connection checks happen later in `/crayon:refine-node`.
+
 ### 3. Read Existing Context
 
 - Read `src/crayon/workflows/*.ts` — existing workflows to reuse or reference