Add Integration Skill

You are an expert at adding complete integrations to Sim. This skill orchestrates the full process of adding a new service integration.

Overview

Adding an integration involves these steps in order:

1. Research - Read the service's API documentation
2. Create Tools - Build tool configurations for each API operation
3. Create Block - Build the block UI configuration
4. Add Icon - Add the service's brand icon
5. Create Triggers (optional) - If the service supports webhooks
6. Register - Register tools, block, and triggers in their registries
7. Configure Deployment Availability - Wire OAuth client and service-account metadata
8. Generate and Validate the Catalog - Regenerate docs/catalog artifacts and run drift checks

Step 1: Research the API

Before writing any code:

1. Use Context7 to find official documentation: mcp__context7__resolve-library-id, then fetch with mcp__context7__query-docs
2. Or use WebFetch to read API docs directly
3. Identify:
   • Supported authentication grants, who owns the app, and which permissions each operation needs
   • Available operations (CRUD, search, etc.)
   • Required vs optional parameters
   • Response structures

Choose the connection flow before building tools

Read the provider's current authentication documentation first. “OAuth” describes a protocol; it does not imply a browser redirect or a deployment-wide client ID and secret. Compare the supported paths and choose the simplest supported setup for the intended user:

Do not present customer-owned credentials as a way around a provider's approval requirements for a shared public integration. Explain the supported account/app type in the setup docs. Keep existing OAuth connections usable when other users depend on them, unless a migration or removal is explicitly authorized.

For client credentials, extend the existing descriptors and minter registry under apps/sim/lib/credentials/client-credential-accounts/; for token credentials, use token-service-accounts/. Reuse the connect modal, encrypted storage, authorized credential operations, and execution-time token resolver. Do not collect reusable secrets on every block, invent a credential route, or store a short-lived access token as if it were permanent.

Verify the complete lifecycle, including connect-time verification, concurrent executions on different workers, expiry, secret rotation, and permission changes. Some providers invalidate the previous token whenever another is minted. Those providers require shared coordination keyed by the provider application identity, including across duplicate saved credentials; a process-local or per-scope token cache is insufficient. A cache hit must never authenticate a wrong secret or silently grant broader permissions. Bound token responses and retries, prevent credential-bearing redirects, and never include provider response bodies or secrets in errors.

Use explicit permission choices when the supported operations have different access needs. Keep region and permission selections on reconnect unless the user changes them. Verify documented identity endpoints rather than guessing which person a client-credentials token represents.

Pair saved credentials with block and resource selectors

oauth-input is the shared saved-credential picker, including for service accounts that never redirect to OAuth. Give its basic/advanced pair one canonicalParamId: 'oauthCredential', with the correct serviceId, and wire tool OAuth metadata to that same service. Tokens and trusted API origins are hidden execution inputs; the model receives a credential ID, never a secret. Use credentialKind: 'service-account' when only service accounts are supported, including in tool OAuth metadata. Use credentialKind: 'any' on the picker when both browser OAuth and saved service accounts are supported; omitting it defaults the connect action to browser OAuth. Per-connection scope choices belong in the descriptor and encrypted credential, not an all-permissions block requiredScopes array that would reject read-only connections.

When a documented list endpoint makes a resource ID discoverable, pair the saved credential with a dynamic resource selector using the add-selector and validate-selector skills:

• Declare the credential subblock and any parent resource in dependsOn, and give the resource selector and its manual advanced input the same canonical parameter.
• Register browser-safe selector metadata and a server attachment through the shared selector framework. Resolve the credential with the expected provider binding and use a fixed or credential-bound destination; selectors and execution must use the same auth/region policy.
• Exercise switching credentials, parent resources, pagination, expired tokens, denied access, and the advanced environment-reference path. Check that stale choices cannot survive a change of account. Do not fetch provider data or mint tokens in the browser.
• Keep a manual ID path when the provider cannot enumerate a resource. Do not invent an endpoint solely to provide a dropdown.

For an existing integration, inspect persisted workflow serialization as well as the visible form before removing old auth fields. Establish whether existing users need a migration or compatibility path; do not add permanent legacy branches speculatively when removal is authorized. When compatibility is needed, prove that old values survive serialization and execution; hiding a field is not proof. If the user authorizes a production usage check, query only the aggregate usage/auth-shape evidence needed and keep identities, secrets, and local evidence out of commits and PRs.

Hard Rule: No Guessed Response Schemas

If the official docs do not clearly show the response JSON shape for an endpoint, you MUST stop and tell the user exactly which outputs are unknown.

• Do NOT guess response field names
• Do NOT infer nested JSON paths from related endpoints
• Do NOT invent output properties just because they seem likely
• Do NOT implement transformResponse against unverified payload shapes

If response schemas are missing or incomplete, do one of the following before proceeding:

1. Ask the user for sample responses
2. Ask the user for test credentials so you can verify the live payload
3. Reduce the scope to only endpoints whose response shapes are documented
4. Leave the tool unimplemented and explicitly report why

Step 2: Create Tools

Directory Structure

apps/sim/tools/{service}/
├── index.ts          # Barrel exports
├── types.ts          # TypeScript interfaces
├── {action1}.ts      # Tool for action 1
├── {action2}.ts      # Tool for action 2
└── ...

Key Patterns

Choose the tool boundary before writing the declaration:

• Use InternalToolConfig.operation for same-process Sim/provider work. Put the handler under apps/sim/lib/internal/{service}/execute-tool.ts and register every ID in apps/sim/lib/internal/tool-operations/registry.server.ts.
• Use ToolConfig.request only for an absolute external HTTP(S) provider endpoint.

Never point a tool at /api/..., construct an absolute URL back to Sim, declare request.internal, add a directExecution property (it fails bun run check:tool-request-boundary), or add an API route merely to reuse code, normalize files, or authorize resources. A real external/browser route and an in-process tool may share the same operation, but neither calls the other. Follow the full transport and handler rules in the add-tools skill.

types.ts:

import type { ToolResponse } from '@/tools/types'

export interface {Service}{Action}Params {
  accessToken: string      // For OAuth services
  // OR
  apiKey: string          // For API key services

  requiredParam: string
  optionalParam?: string
}

export interface {Service}{Action}Response extends ToolResponse {
  output: {
    // Define output structure
  }
}

Declare one response interface per tool, imported by that tool's ToolConfig<Params, Response> (or InternalToolConfig for in-process work). Never add an umbrella {Service}Response union: nothing imports it.

Tool file pattern: an external provider API uses ToolConfig with request (absolute https:// URL, headers, body, transformResponse); same-process Sim work uses InternalToolConfig with operation. Both full templates, param visibility rules, and output typing live in .agents/skills/add-tools/SKILL.md — read it before writing the first tool.

Critical Rules

• visibility: 'hidden' for OAuth tokens
• visibility: 'user-only' for API keys and user credentials
• visibility: 'user-or-llm' for operation parameters
• Always use ?? null for nullable API response fields
• Always use ?? [] for optional array fields
• Set optional: true for outputs that may not exist
• Never output raw JSON dumps - extract meaningful fields
• When using type: 'json' and you know the object shape, define properties with the inner fields so downstream consumers know the structure. Only use bare type: 'json' when the shape is truly dynamic

Resolved Secrets at Model and Persistence Boundaries

Classify every request field (ordinary provider input / AI-consumed text / opaque model bytes / Sim-durable storage) before implementing the tool and apply the shared projection or provenance mechanism only where a concrete Sim {{...}} resolution path reaches a later model or log boundary. Full rules and the required tests are in .agents/skills/add-tools/SKILL.md → "Resolved Secrets and Provenance Boundaries".

Step 3: Create Block

File Location

apps/sim/blocks/blocks/{service}.ts

Follow .agents/skills/add-block/SKILL.md for the block structure, subBlock types, condition/dependsOn/required/mode syntax, outputs, canvasPresentation sentences, and the {Service}BlockMeta export (minimum 7 templates, plus url and skills). Every block declares canvasPresentation; bun run apps/sim/scripts/check-canvas-sentences.ts --block={service} must pass (CI runs check:canvas-sentences --require-coverage).

Three rules that are easy to get wrong when copying from existing blocks:

• Every remote selectorKey must use the unified server selector path. Apply the add-selector skill: add browser-safe metadata to apps/sim/lib/selectors/manifest.ts, reuse or extract a server-only provider listing primitive, and add a credential- and destination-bound server attachment. Do not add a client provider fetcher, a provider-specific query key, browser token acquisition, or a selector-only API route. The shared context builder sends only active dependsOn values and preserves exact {{KEY}} environment references for server-side resolution.
• Basic/advanced pairs use a canonicalParamId; its constraints are in .claude/rules/sim-integrations.md and the add-block skill → canonicalParamId Pattern.
• Every text-entry subBlock (short-input, long-input, code) and every selector declares a placeholder; an empty box tells the user nothing. Secrets read Enter your {thing} (e.g. Enter your API key), free text names what to type (Enter branch name), and formatted values show the shape (2023-01-01T00:00:00Z, 1 to 1000). An optional field with a server-side default names that default (Defaults to the database region). Dropdowns, switches, and oauth-input do not need one.

Step 4: Add Icon

File Location

apps/sim/components/icons.tsx

Pattern

export function {Service}Icon(props: SVGProps<SVGSVGElement>) {
  return (
    <svg
      {...props}
      viewBox="0 0 24 24"
      fill="none"
      xmlns="http://www.w3.org/2000/svg"
    >
      {/* SVG paths from user-provided SVG */}
    </svg>
  )
}

Getting Icons

Do not search for icons yourself. At the end of implementation, ask the user to paste the service's SVG (usually on its brand/press kit page).

Once the user provides the SVG:

1. Extract the SVG paths/content
2. Create a React component that spreads props
3. Ensure viewBox is preserved from the original SVG

Theme-safety (bare rendering) — REQUIRED

The icon renders both inside its colored bgColor tile AND "bare" (no tile) on a neutral page — e.g. the home Suggested actions list — in both light and dark mode. A monochrome logo whose paths hardcode a single near-white or near-black fill is invisible bare on the matching background (white-on-white in light mode, black-on-black in dark mode).

Rules when adding the SVG:

• Monochrome logos (a single white or black mark): draw the shape with fill='currentColor', not fill='#fff' / fill='#000000'. It then inherits white inside dark tiles, near-black inside light tiles (via getTileIconColorClass), and the theme-aware var(--text-icon) bare — legible everywhere. Do NOT set iconColor for these.
• Multi-color brand logos (their own vivid fills): keep the hardcoded fills. They read on any background. Only set iconColor (a vivid brand hex, never a near-black/near-white tile color) if the bare icon should adopt a brand tint.
• A large white shape with a tiny vivid accent (e.g. a logo where the body is the white negative space) still vanishes bare — convert the body to currentColor.

Verify with bun run check:bare-icons (also runs in CI). It flags purely monochrome hazards; for partial-accent logos, eyeball the suggested-actions list in both light and dark mode.

Step 5: Create Triggers (Optional)

If the service supports webhooks or needs polling, follow .agents/skills/add-trigger/SKILL.md (directory layout, buildTriggerSubBlocks, provider handler, polling handler); then wire triggers.enabled / triggers.available into the block and spread each trigger's getTrigger(id).subBlocks after the tool subBlocks.

Step 6: Register Everything

Tools Registry (apps/sim/tools/registry.ts)

// Add import (alphabetically)
import {
  {service}Action1Tool,
  {service}Action2Tool,
} from '@/tools/{service}'

// Add to tools object (alphabetically)
export const tools: Record<string, ExecutableToolConfig> = {
  // ... existing tools ...
  {service}_action1: {service}Action1Tool,
  {service}_action2: {service}Action2Tool,
}

Then regenerate the generated tool metadata and commit it:

bun run generate:tool-metadata

Client code reads params/outputs from these artifacts rather than importing the registry, so a tool you add, change or remove is invisible to the UI until they are regenerated, and CI fails on stale ones. See .agents/skills/tool-registry-boundary/SKILL.md.

Block Registry (apps/sim/blocks/registry-maps.ts)

The data maps (BLOCK_REGISTRY + BLOCK_META_REGISTRY) live in registry-maps.ts; registry.ts holds only the accessor functions. Add the import and an entry to each map alphabetically:

// Add import (alphabetically)
import { {Service}Block, {Service}BlockMeta } from '@/blocks/blocks/{service}'

// Add to the config map (alphabetically)
export const BLOCK_REGISTRY: Record<string, BlockConfig> = {
  // ... existing blocks ...
  {service}: {Service}Block,
}

// Add to the catalog-meta map (alphabetically)
export const BLOCK_META_REGISTRY: Record<string, BlockMeta> = {
  // ... existing metas ...
  {service}: {Service}BlockMeta,
}

Trigger Registry (apps/sim/triggers/registry.ts) - If triggers exist

// Add import (alphabetically)
import {
  {service}EventATrigger,
  {service}EventBTrigger,
  {service}WebhookTrigger,
} from '@/triggers/{service}'

// Add to TRIGGER_REGISTRY (alphabetically)
export const TRIGGER_REGISTRY: TriggerRegistry = {
  // ... existing triggers ...
  {service}_event_a: {service}EventATrigger,
  {service}_event_b: {service}EventBTrigger,
  {service}_webhook: {service}WebhookTrigger,
}

Step 7: Configure Deployment Availability

Do this for every integration that uses the shared credential picker. Only a connection that depends on deployment-wide OAuth client fields needs an OAuth client capability; a customer-owned service account must remain available without those fields.

The block's oauth-input.serviceId is the canonical link between the generated integration catalog, the OAuth service configuration, deployment availability, and the setup CLI.

1. Set authMode: AuthMode.OAuth for integrations using browser OAuth or customer-owned OAuth client credentials. Token-only service accounts can retain AuthMode.ApiKey with the shared picker, as Coda does; oauth-input alone does not determine the authentication protocol. Register a new token-only service ID and block type in tokenCredentialIntegrationTypes in packages/deployment-config/src/integration-availability.ts so availability and integration policy recognize the saved credential path. For OAuth integrations, this value lets the catalog discover the connection flow instead of routing "Add to Sim" to chat. Ensure the block has exactly one distinct OAuth serviceId matching the canonical service in apps/sim/lib/oauth/oauth.ts. The canonical service's authType selects browser OAuth or the service-account modal. Verify the resulting catalog and block connection actions for the chosen authentication method.
2. For browser OAuth, confirm resolveOAuthClientCapabilityId(serviceId) resolves to the intended provider entry in OAUTH_CLIENT_CAPABILITIES in packages/deployment-config/src/env-capabilities.ts. Google and Microsoft service IDs deliberately share provider-level capabilities.
3. For a new browser OAuth provider, add the required client fields to OAUTH_CLIENT_CAPABILITIES, add every referenced field to the env schema in apps/sim/lib/core/config/env.ts, and add the matching text or secret entries to OAUTH_CLIENT_SETUP_FIELDS in packages/sim-setup/src/capability-config.ts. Do not create integration-specific setup logic or infer secret fields from naming; the CLI mapping is exhaustively checked against the runtime fields.
4. If the canonical OAuth service has serviceAccountProviderId, run bun run generate:deployment-config to refresh packages/deployment-config/src/service-account-providers.generated.ts; never hand-edit the generated provider-ID map. In packages/deployment-config/src/service-account-metadata.ts, use:
   • no deploymentRequirement when the service-account path works independently of OAuth client fields;
   • 'oauth-client' when it requires the same deployment OAuth client fields;
   • 'preview-gated' when availability is controlled by the service-account preview block. For a service-account-only default, set the canonical service's authType: 'service_account' and serviceAccountProviderId. Verify both block availability and the connect modal with no deployment OAuth credentials configured. An existing browser OAuth path may remain for legacy credentials without becoming a prerequisite for the new path.

Never add a permissive fallback for missing capability metadata. A browser OAuth connection without a resolvable capability must fail validation; an independent service account uses its own metadata.

Step 8: Generate and Validate the Catalog

Run bun run generate:tool-metadata, bun run scripts/generate-docs.ts, bun run generate:deployment-config, then bun run check:audits (see the validate-integration skill → Regenerate Derived Artifacts for the full list and what each check verifies).

The docs generator creates apps/docs/content/docs/integrations/{service}.mdx — one page per service carrying the block's Actions and, if it has one, its Triggers section. Never hand-edit generated pages; the only editable region is the {/* MANUAL-CONTENT */} block (see scripts/README.md).

Every generated integration page carries a hand-written intro directly under <BlockInfoCard />. The generator preserves it across regenerations, so write it once after the first generate:

{/* MANUAL-CONTENT-START:intro */}
[{Service}](https://service.com/) is {one sentence on what the service is}.

With the {Service} block, you can:

- **{Capability}**: {what the operations in this group do}
- **{Capability}**: {...}

{How to connect: which credential to create and where, if it is not OAuth.}

In Sim, the {Service} block lets your agents {concrete workflow uses}.
{/* MANUAL-CONTENT-END */}

Group the bullets by what the user gets done, not one bullet per tool. Only describe operations the block actually ships. Follow .claude/rules/constitution.md for voice. Re-run bun run scripts/generate-docs.ts afterwards and confirm the section survived unchanged.

The docs generator refreshes packages/deployment-config/src/integrations.json, and the deployment config generator projects service-account provider IDs from that catalog plus the canonical OAuth registry. The checks compare both committed projections with their sources. Review the generated diff and keep only intentional changes.

V2 Integration Pattern

If creating V2 versions (API-aligned outputs):

1. V2 Tools - Add _v2 suffix, version 2.0.0, flat outputs

2. V2 Block - Add _v2 type, use createVersionedToolSelector

3. V1 Block - Add (Legacy) to name, set hideFromToolbar: true, and add sunset: { status: 'legacy', replacedBy: '{service}_v2' } — check-block-registry fails a legacy block with no replacedBy, and the amber legacy badge plus its click-to-upgrade action read from that field.

   Only add replacedBy once the target is GA. The same check also fails when the target is unregistered, itself sunset, or still preview: true. If v2 is preview-gated, leave v1 alone until GA and drop preview in the same commit that adds the sunset — splitting them breaks the build in between.

4. Registry - Register both versions

// In registry
{service}: {Service}Block,        // V1 (legacy, hidden)
{service}_v2: {Service}V2Block,   // V2 (visible)

Complete Checklist

Tools

• Created tools/{service}/ directory
• Created types.ts with all interfaces
• Created tool file for each operation
• Chose exactly one boundary per tool: registered InternalToolConfig.operation or absolute external HTTP(S) ToolConfig.request
• No tool points to /api/..., constructs a URL back to Sim, declares request.internal or a directExecution property (fails bun run check:tool-request-boundary), or has an HTTP fallback for an in-process operation
• All params have correct visibility
• All nullable fields use ?? null
• All optional outputs have optional: true
• Created index.ts barrel export
• Registered all tools in tools/registry.ts
• Ran bun run generate:tool-metadata and committed the regenerated artifacts
• Classified every model-visible, opaque, Sim-durable, and internal-execution request field
• Added shared model-input projection or private provenance only where required; ordinary external resource locators and control inputs retain their request semantics
• Confirmed ordinary third-party tool results are not generically sanitized
• Added provenance compatibility and fail-closed boundary tests where applicable
• bun run check:tool-request-boundary passes
• Internal-operation registry completeness test passes for every operation-backed tool

Block

• Created blocks/blocks/{service}.ts
• Set integrationType to the correct IntegrationType enum value
• {Service}BlockMeta.tags lists every applicable IntegrationTag (tags live on the meta, not the block)
• Defined operation dropdown with all operations
• Added the saved-credential picker with the supported credentialKind; browser OAuth scopes use getScopesForService('{service}'), while variable service-account permissions stay on the credential
• Added conditional fields per operation
• Every short-input, long-input, code, and selector subBlock has a placeholder
• Set up dependsOn for cascading selectors
• Every remote selectorKey exists in the shared manifest and has one server attachment with trusted credential provider binding and a fixed, credential-bound, or explicitly reviewed user-controlled destination policy
• No selector provider logic, credential resolution, or provider route call runs in the browser
• Configured tools.access with all tool IDs
• Configured tools.config.tool selector
• Defined outputs matching tool outputs
• Registered block + meta in blocks/registry-maps.ts (BLOCK_REGISTRY / BLOCK_META_REGISTRY)
• If triggers: set triggers.enabled and triggers.available
• If triggers: spread trigger subBlocks with getTrigger()
• Exported {Service}BlockMeta with at least 7 templates
• canvasPresentation.sentences covers every operation; bun run apps/sim/scripts/check-canvas-sentences.ts --block={service} passes
• {Service}BlockMeta also sets url (verified external homepage) and skills (grounded in tools.access, sourced from real use cases) — see add-block → BlockMeta

Authentication and saved credentials

• Compared documented authorization code, client credentials, token, and key-based methods
• Chosen app ownership and approval requirements match the intended user
• Saved credential picker, tools, and resource selectors share the same provider/region binding
• Connect verification, expiry, concurrent workers, secret rotation, and scope changes are sound
• Existing usage and the migration/removal decision are established; any required compatibility is verified through serialization and execution
• New connection and reconnect flows verified in the running UI

Browser OAuth Scopes (if authorization-code flow is supported)

• Defined scopes in lib/oauth/oauth.ts under OAUTH_PROVIDERS
• Added scope descriptions in SCOPE_DESCRIPTIONS within lib/oauth/utils.ts
• Used getCanonicalScopesForProvider() in lib/auth/connectors/providers.ts (never hardcode)
• Used getScopesForService() for the browser OAuth permissions the block needs (never hardcode)
• A picker that also accepts service accounts does not require broader scopes than every supported connection needs; per-connection service-account permissions are validated by the descriptor/minter

Deployment Availability (if using the saved-credential picker)

• Block declares exactly one distinct oauth-input.serviceId
• Browser OAuth resolves to the intended OAUTH_CLIENT_CAPABILITIES entry; independent service accounts work without deployment OAuth fields
• Every new OAuth capability field exists in apps/sim/lib/core/config/env.ts
• Runtime OAuth fields live in OAUTH_CLIENT_CAPABILITIES; matching CLI input modes live in the exhaustively checked OAUTH_CLIENT_SETUP_FIELDS
• If serviceAccountProviderId is configured, SERVICE_ACCOUNT_METADATA_BY_OAUTH_SERVICE_ID has the matching projection and deployment requirement

Icon

• Asked user to provide SVG
• Added icon to components/icons.tsx
• Icon spreads props correctly
• Monochrome marks use fill='currentColor' (not hardcoded white/black) so the icon renders bare in light AND dark mode — verified with bun run check:bare-icons

Triggers (if service supports webhooks)

• Created triggers/{service}/ directory
• Created utils.ts with options, instructions, and extra fields helpers
• Primary trigger uses includeDropdown: true
• Secondary triggers do NOT have includeDropdown
• All triggers use buildTriggerSubBlocks helper
• Created index.ts barrel export
• Registered all triggers in triggers/registry.ts

Docs and deployment metadata

• Ran bun run scripts/generate-docs.ts
• Ran bun run generate:deployment-config for OAuth or service-account changes
• Verified docs file created
• Wrote the {/* MANUAL-CONTENT-START:intro */} section under <BlockInfoCard /> and confirmed it survives a regenerate
• Reviewed and committed the generated packages/deployment-config/src/integrations.json change
• bun run check:integration-catalog passes
• bun run check:docs passes — CI fails on stale generated docs, so commit the full generator output, including catch-up regeneration for pages another PR left stale (never revert it as "unrelated drift")
• bun run check:deployment-config passes

Final Validation (Required)

• Read every tool file and cross-referenced inputs/outputs against the API docs
• Verified block subBlocks cover all required tool params with correct conditions
• Verified block outputs match what the tools actually return
• Verified tools.config.params correctly maps and coerces all param types
• Verified every tool output and transformResponse path against documented or live-verified JSON responses
• If any response schema remained unknown, explicitly told the user instead of guessing
• {Service}BlockMeta exported with at least 7 templates, each having icon, title, prompt, modules, category, and tags

File Handling

When your integration handles file uploads or downloads, follow these patterns to work with UserFile objects consistently.

What is a UserFile?

UserFile (apps/sim/executor/types.ts) is the standard file representation in Sim — id, name, an access url (not guaranteed presigned — remoteUrl is the short-lived signed one, set only for providers that fetch by URL), size, MIME type, storage key, and optional inline base64 / provider file handles. Read file bytes through the documented upload helpers, never by fetching url directly. Read the interface rather than relying on a copy here.

File Input Pattern (Uploads)

File authorization, normalization, storage reads, provider upload, and response mapping belong in a registered in-process operation. Do not create an internal API route for file tools.

1. Block SubBlocks for File Input

Use the basic/advanced mode pattern:

// Basic mode: File upload UI
{
  id: 'uploadFile',
  title: 'File',
  type: 'file-upload',
  canonicalParamId: 'file',  // Maps to 'file' param
  placeholder: 'Upload file',
  mode: 'basic',
  multiple: false,
  required: true,
  condition: { field: 'operation', value: 'upload' },
},
// Advanced mode: Reference from previous block
{
  id: 'fileRef',
  title: 'File',
  type: 'short-input',
  canonicalParamId: 'file',  // Same canonical param
  placeholder: 'Reference file (e.g., {{file_block.output}})',
  mode: 'advanced',
  required: true,
  condition: { field: 'operation', value: 'upload' },
},

Critical: canonicalParamId must NOT match the id of a subblock outside its canonical group.

2. Normalize File Input in Block Config

tools.config.tool selects the tool before variable resolution and must not mutate or coerce input. Use tools.config.params, which runs after variable resolution, to normalize all file variants:

import { normalizeFileInput } from '@/blocks/utils'

tools: {
  config: {
    tool: (params) => `{service}_${params.operation}`,
    params: (params) => {
      // Serialization collapses the basic/advanced pair into the canonical `file` key.
      const normalizedFile = normalizeFileInput(params.file, { single: true })
      return normalizedFile ? { file: normalizedFile } : {}
    },
  },
}

3. Define and register the in-process operation

export const {service}UploadTool: InternalToolConfig<Params, Response> = {
  id: '{service}_upload',
  // ...
  params: {
    file: { type: 'file', required: false, visibility: 'user-or-llm' },
  },
  operation: {
    input: (params) => ({
      accessToken: params.accessToken,
      file: params.file,
    }),
  },
}

Implement apps/sim/lib/internal/{service}/execute-tool.ts and keep the file/provider work in typed operations beside it. The handler validates request.input, derives storage authority only from trusted request.context, authorizes every stored file before reading bytes, forwards request.signal, enforces declared and actual byte caps, and returns the canonical tool response. Register {service}_upload in apps/sim/lib/internal/tool-operations/registry.server.ts; the sweep in apps/sim/tools/request-transport.test.ts fails a forgotten registration (registry.server.test.ts checks registered ids are canonical with loadable handlers). For anything more, run the test-audit gate. There is no HTTP fallback.

File Output Pattern (Downloads)

Declare a file / file[] output on the tool. For a raw binary endpoint, set request.responseType: 'binary' and return output.file = { name, mimeType, data: buffer, size } from transformResponse(response, params?, context?). The executor's FileToolProcessor stores it and replaces it with a UserFile; tools never call it.

In an operation handler, return createInternalToolFileResult / createInternalToolFilesResult from lib/internal/tool-operations/file-result.ts — never base64 JSON. See the add-tools skill → File Downloads and Generated Files.

Key Helpers Reference

Advanced Mode for Optional Fields

Optional fields that are rarely used should be set to mode: 'advanced' so they don't clutter the basic UI. Examples: pagination tokens, time range filters, sort order, max results, reply settings.

WandConfig for Complex Inputs

Use wandConfig for fields that are hard to fill out manually:

• Timestamps: Use generationType: 'timestamp' to inject current date context into the AI prompt
• JSON arrays: Use generationType: 'json-object' for structured data
• Complex queries: Use a descriptive prompt explaining the expected format

{
  id: 'startTime',
  title: 'Start Time',
  type: 'short-input',
  mode: 'advanced',
  wandConfig: {
    enabled: true,
    prompt: 'Generate an ISO 8601 timestamp. Return ONLY the timestamp string.',
    generationType: 'timestamp',
  },
}

OAuth Scopes (Centralized System)

Scopes are maintained in a single source of truth and reused everywhere:

1. Define scopes in lib/oauth/oauth.ts under OAUTH_PROVIDERS[provider].services[service].scopes
2. Add descriptions in SCOPE_DESCRIPTIONS within lib/oauth/utils.ts for the OAuth modal UI
3. Reference in lib/auth/connectors/providers.ts (buildConnectorProviders) using getCanonicalScopesForProvider(providerId) from @/lib/oauth/utils
4. Reference in blocks using getScopesForService(serviceId) from @/lib/oauth/utils

Never hardcode scope arrays in the Better Auth connector providers or block requiredScopes. Always import from the centralized source.

// In lib/auth/connectors/providers.ts (Better Auth connector providers)
scopes: getCanonicalScopesForProvider('{service}'),

// In block credential sub-block
requiredScopes: getScopesForService('{service}'),

Common Gotchas

1. OAuth serviceId must match - The serviceId in oauth-input must match the OAuth provider configuration
2. DependsOn clears options - When an active dependency changes, the shared selector facade refetches with an opaque query revision; dependency values and references never enter query keys
3. Never pass Buffer directly to fetch - Convert to new Uint8Array(buffer) for TypeScript compatibility
4. Legacy fileContent params - Only an existing tool that already accepted base64 fileContent keeps that hidden param; new tools take file only