Write Custom Backend

Write Custom Backend

Contract

By the end of this skill the consumer has:

• A server route that exports a POST handler returned by createPilotHandler.
• A clear model argument in one of the three supported shapes (string, LanguageModel instance, thunk).
• A system prompt (server-owned; always prepended before client-derived instructions).
• A maxSteps chosen deliberately, not accepted blindly.
• (Optionally) getProviderOptions for per-request provider tuning (cache hints, thinking budgets).

Iron Law: server-owned system prompt always wins

The handler composes the final system prompt as [options.system, body.system, body.context] (see composeSystemPrompt in packages/agentickit/src/server/handler.ts lines 613-628). options.system is server-controlled and comes first; client-derived sections (.pilot/ skills, registered state) are appended. If you need guardrails the client cannot tamper with (tone, safety instructions, tenant isolation), put them in options.system. A compromised client can inject body.system but cannot shadow options.system.

Phases

Phase 1: pick a runtime shape

createPilotHandler returns (request: Request) => Promise<Response>, so it works anywhere the Web Fetch API is available:

• Next.js App Router (tested): export const POST = createPilotHandler({...}).
• Bun: pass into Bun.serve({ fetch: POST }).
• Cloudflare Workers: export as export default { fetch: POST }.
• Hono: app.post("/api/pilot", (c) => POST(c.req.raw)).

Phase 2: write the minimal route

// app/api/pilot/route.ts
import { createPilotHandler } from "@hec-ovi/agentickit/server";

export const POST = createPilotHandler({});

Auto-detects a provider from env. See skills/choose-provider/SKILL.md for the priority order.

Phase 3: the full options surface

import { createPilotHandler } from "@hec-ovi/agentickit/server";

export const POST = createPilotHandler({
  model: "anthropic/claude-sonnet-4-5",
  system: [
    "You are the support copilot for a kanban app.",
    "Always confirm destructive actions before invoking them.",
    "Never expose internal card IDs; reference cards by title.",
  ].join(" "),
  maxSteps: 5,
  getProviderOptions: () => ({
    anthropic: { cacheControl: { type: "ephemeral" } },
  }),
});

The verified options shape (from CreatePilotHandlerOptions in server/handler.ts lines 224-275):

interface CreatePilotHandlerOptions {
  system?: string;
  model?: ModelSpec;                                // see Phase 4
  getProviderOptions?: () => Record<string, unknown>;
  maxSteps?: number;                                // default 5
}

Phase 4: the three ModelSpec shapes

From the exported type (line 115):

type ModelSpec = string | LanguageModel | (() => LanguageModel | Promise<LanguageModel>);

Shape 1: string.

createPilotHandler({ model: "openai/gpt-4o" });
createPilotHandler({ model: "openrouter/qwen/qwen3-coder:free" });
createPilotHandler({ model: "auto" });   // synonym for omitting model

Resolution order (see planResolution in lines 532-585):

1. Direct provider key (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.) + matching @ai-sdk/* peer package installed → direct adapter.
2. No direct key but AI_GATEWAY_API_KEY / VERCEL_OIDC_TOKEN set → string handed to streamText verbatim; the Vercel AI Gateway resolves it.
3. Neither → the handler factory throws at creation time with a clear error naming the missing env var and package.

Supported prefixes (line 29): openai, anthropic, groq, openrouter, google, mistral. Anything else throws "unsupported model prefix" at handler creation.

Shape 2: LanguageModel instance.

import { createOllama } from "ai-sdk-ollama";
const ollama = createOllama();
export const POST = createPilotHandler({ model: ollama("llama3.3") });

Detected via isLanguageModelInstance (lines 380-392). Prefix validation is skipped. Use this for Ollama, Azure, Bedrock, or any custom adapter.

Shape 3: thunk.

createPilotHandler({
  model: async () => {
    const token = await refreshToken();
    return customAdapter(token, "model-id");
  },
});

Called exactly once at handler creation (lines 710-722). The resolved value must be a LanguageModel instance; strings are rejected with a clear error. Useful for async auth exchanges at startup.

Phase 5: per-request model overrides

The client can pass <Pilot model="openai/gpt-4o-mini"> and the string is forwarded in the request body. The server re-validates the prefix against the same allow-list (lines 841-853), so the client cannot inject an arbitrary string.

Overrides are only honored when options.model is a string (lines 854-860); if the handler uses a LanguageModel instance or thunk, the client override returns a 400 unsupported_provider. This is deliberate: the instance / thunk paths don't have a resolver plumbed.

Phase 6: maxSteps

The handler passes stopWhen: stepCountIs(options.maxSteps ?? 5) to streamText (line 907). A "step" is one model call plus any tool calls it emits. Five is enough for call → result → follow-up → polish; raise it to 10+ if your app has chained tools that legitimately need more round-trips. Lower it to cap cost.

Phase 7: getProviderOptions

Called for every request (line 888-891). Returns provider-specific options forwarded verbatim to streamText({ providerOptions }). Use for:

• Anthropic cache control.
• OpenAI reasoning / thinking budgets.
• Groq tool-call retry settings.
• Per-provider temperature overrides.

Do NOT return API keys here; they live in env vars. The option's type is intentionally loose (Record<string, unknown>) to avoid leaking AI SDK internal types through the public API.

Phase 8: the error envelope

All non-streaming errors return a narrow JSON envelope (PilotErrorBody, lines 283-287):

{ "error": "human-readable message", "code": "invalid_request" }

Codes (narrow, client-matchable):

• invalid_request: 400, body didn't parse.
• unsupported_provider: 400, model prefix not allowed.
• internal_error: 500, catchall.
• method_not_allowed: 405, not a POST.

Stack traces never leak (line 925: console.error server-side, sanitized message to client).

Phase 9: CORS

The handler emits permissive CORS headers by default (line 367-372). Consumers who need tighter policy wrap the handler in their own middleware. Don't modify the response headers from inside getProviderOptions (they're applied after streamText returns).

Anti-Patterns

• Putting the API key in getProviderOptions. Env vars, always.
• Returning new Response() from inside getProviderOptions. That hook feeds streamText, not the response pipeline.
• Parsing the request body yourself to inject state. The handler's Zod schema (lines 325-359) validates exactly what useChat sends; add a custom preprocessor by wrapping the handler in middleware, not by patching internals.
• Setting maxSteps: 100. At that point a runaway loop costs real money. If you need more steps, examine whether the chain is the right tool. A multi-step LLM loop is rarely the cheapest or most reliable path.
• Assuming streamText is called on the server for every tool. Client tools (declared via usePilotAction) stream back to the browser. The server wraps them with dynamicTool and a throwing execute that signals "this is client-side only" (lines 668-672).

Output Format

After configuring, report:

• The runtime (Next.js / Bun / Workers / Hono).
• The model shape (string / instance / thunk) and the specific value.
• Whether system / maxSteps / getProviderOptions are set and why.
• A one-sentence description of the error-handling contract the consumer should expect.

Tools Used

• Edit the server route file.
• Read packages/agentickit/src/server/handler.ts to verify the exact option shape and provider registry.