diff --git a/AGENTS.md b/AGENTS.md
index 043130a..29ab4ac 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -68,6 +68,12 @@ Spec drift is never "someone else's problem to fix later." It compounds.
 
 ---
 
+## Rule 7: Place helper functions at the end of implementation files
+
+When adding or refactoring helper functions in source files, place them after the main class/function implementations (i.e. helper functions last in the file), unless language constraints require otherwise.
+
+---
+
 ## Verification Checklist
 
 Run this checklist after every non-trivial change:
diff --git a/packages/core/src/agent-session-do.ts b/packages/core/src/agent-session-do.ts
index 7c1bf7d..dbc06c4 100644
--- a/packages/core/src/agent-session-do.ts
+++ b/packages/core/src/agent-session-do.ts
@@ -43,6 +43,8 @@ import type {
 } from "@piccolo/api";
 import type { FinishReason, LanguageModel, LanguageModelUsage, ModelMessage } from "ai";
 import { stepCountIs, streamText } from "ai";
+import { createAiGateway } from "ai-gateway-provider";
+import { createUnified } from "ai-gateway-provider/providers/unified";
 import { agentCompact, splitForCompaction } from "./agent-compact.ts";
 import { toAiSdkTools } from "./agent-tools.ts";
 import type {
@@ -57,7 +59,6 @@ import type {
 import { generateEntryId, isLegacySystemMessageData, parseEntry } from "./db/entry-types.ts";
 import { getEntries, getSession } from "./db/schema.ts";
 import { ExtensionRunner } from "./extension-runner.ts";
-import { createModel } from "./gateway.ts";
 import { Messages } from "./messages.ts";
 import { ObservableImpl } from "./observable-impl.ts";
 import { buildSessionContext, walkToRoot } from "./session/context.ts";
@@ -1196,3 +1197,12 @@ function estimateTokens(messages: ModelMessage[]): number {
   }
   return Math.ceil(chars / 4);
 }
+
+function createModel(env: Env, modelId: string): LanguageModel {
+  const gateway = createAiGateway({
+    accountId: env.CF_ACCOUNT_ID,
+    gateway: env.CF_AI_GATEWAY_NAME,
+    apiKey: env.CF_AI_GATEWAY_TOKEN,
+  });
+  return gateway(createUnified()(modelId)) as LanguageModel;
+}
diff --git a/packages/core/src/gateway.ts b/packages/core/src/gateway.ts
deleted file mode 100644
index 9eca046..0000000
--- a/packages/core/src/gateway.ts
+++ /dev/null
@@ -1,41 +0,0 @@
-/**
- * AI Gateway / model construction for piccolo-core.
- *
- * `createModel(env, modelId)` is the single factory used by `AgentSessionDO`
- * to create a `LanguageModel` for the agent. All LLM calls are routed through
- * the Cloudflare AI Gateway unified endpoint.
- *
- * Model IDs use the "{provider}/{model-id}" format, e.g.
- *   "anthropic/claude-sonnet-4-5"
- *   "openai/gpt-4o"
- *
- * Spec refs:
- *   specs/agent.md §LLM Backend: Cloudflare AI Gateway
- *   specs/core.md  §Bindings (CF_ACCOUNT_ID, CF_AI_GATEWAY_NAME, CF_AI_GATEWAY_TOKEN)
- */
-
-import type { LanguageModel } from "ai";
-import { createAiGateway } from "ai-gateway-provider";
-import { createUnified } from "ai-gateway-provider/providers/unified";
-
-/**
- * Construct a LanguageModel routed through the Cloudflare AI Gateway.
- *
- * Reads gateway credentials from the Worker env:
- *   env.CF_ACCOUNT_ID       — Cloudflare account ID (var)
- *   env.CF_AI_GATEWAY_NAME  — AI Gateway name / slug (var)
- *   env.CF_AI_GATEWAY_TOKEN — AI Gateway API token (secret)
- *
- * @param env       Worker environment bindings
- * @param modelId   "{provider}/{model-id}" string, e.g. "anthropic/claude-sonnet-4-5"
- */
-export function createModel(env: Env, modelId: string): LanguageModel {
-  const gateway = createAiGateway({
-    accountId: env.CF_ACCOUNT_ID,
-    gateway: env.CF_AI_GATEWAY_NAME,
-    apiKey: env.CF_AI_GATEWAY_TOKEN,
-  });
-  // createUnified() returns an OpenAI-compatible provider that accepts
-  // "{provider}/{model-id}" as the model string.
-  return gateway(createUnified()(modelId)) as LanguageModel;
-}
diff --git a/packages/core/test/do/mock-model.ts b/packages/core/test/do/mock-model.ts
index ec37d26..6d82487 100644
--- a/packages/core/test/do/mock-model.ts
+++ b/packages/core/test/do/mock-model.ts
@@ -4,8 +4,8 @@
  * Mirrors packages/agent/test/mock-gateway.ts but is kept separate so
  * packages/agent's test helpers do not need to be part of its public API.
  *
- * The `AgentSessionDO` constructor calls `createModel(env, modelId)` from
- * gateway.ts which uses ai-gateway-provider. Tests bypass the real gateway by
+ * The `AgentSessionDO` constructor calls an internal `createModel(env, modelId)`
+ * helper that uses ai-gateway-provider. Tests bypass the real gateway because
  * overriding `createModel` is not viable inside a Durable Object context.
  *
  * Instead, tests use `runInDurableObject` to reach directly into the DO
diff --git a/specs/core.md b/specs/core.md
index ea365a5..551c88d 100644
--- a/specs/core.md
+++ b/specs/core.md
@@ -377,7 +377,7 @@ async newSession(
    b. Derive `sessionId = this.ctx.id.name` (the DO was named with `idFromName(uuid)`).
    c. Set `state.sessionId`, `state.userId`, `state.modelId`, `state.name` where `state.name = options.name ?? sessionId`.
    d. Persist `sessionId`, `modelId`, and `name` to DO storage for cold-start recovery.
-   e. Reconstruct the `LanguageModel` via `createModel(env, modelId)`.
+   e. Reconstruct the `LanguageModel` via the internal `createModel(env, modelId)` helper in `agent-session-do.ts`.
 2. Return `getSession(userId)` — the DO's own `SessionImpl` RpcTarget.
 
 The D1 `sessions` row is **not** written here — it is written lazily on the first persisted entry append (see §Lazy session creation).
@@ -871,4 +871,4 @@ Uses `generateText` (non-streaming) with the same `LanguageModel` as the DO. Cal
 
 ### LLM Backend
 
-All LLM calls go through the **Cloudflare AI Gateway** unified endpoint. `createModel(env, modelId)` in `gateway.ts` constructs the `LanguageModel` via `ai-gateway-provider`. Models are addressed as `{provider}/{model-id}` (e.g. `anthropic/claude-sonnet-4-5`). `piccolo-core` constructs the model; the `Agent` class has no knowledge of how it was built.
+All LLM calls go through the **Cloudflare AI Gateway** unified endpoint. `createModel(env, modelId)` in `agent-session-do.ts` constructs the `LanguageModel` via `ai-gateway-provider`. Models are addressed as `{provider}/{model-id}` (e.g. `anthropic/claude-sonnet-4-5`). `piccolo-core` constructs the model; the `Agent` class has no knowledge of how it was built.
diff --git a/specs/implementation_plan.md b/specs/implementation_plan.md
index 87b27e8..ec593e5 100644
--- a/specs/implementation_plan.md
+++ b/specs/implementation_plan.md
@@ -265,7 +265,7 @@ The Durable Object that owns a live session. Wires `piccolo-agent` to persistenc
 | `packages/core/src/do/agent-session.ts` | `AgentSessionDO` class — full DO implementation |
 | `packages/core/src/do/compaction.ts` | `compact()` helper; creates `CompactionEntry`, calls `agentCompact()` |
 | `packages/core/src/do/retry.ts` | `checkRetry()` with exponential backoff, transient/overflow classification |
-| `packages/core/src/do/gateway.ts` | `createModel()` — wraps `ai-gateway-provider` with env bindings |
+| `packages/core/src/agent-session-do.ts` | Internal `createModel()` helper — wraps `ai-gateway-provider` with env bindings |
 | `packages/core/src/do/system-prompt.ts` | `buildBasePrompt(agentName)` — renders the base prompt template |
 | `packages/core/src/do/stubs.ts` | `ExtensionRunnerStub`, `SystemPromptAssemblerStub` — no-op stubs for steps 6–7 |
 | `packages/core/src/do/types-internal.ts` | `SystemPromptAddition`, `MODEL_CATALOG`, `resolveModel()` |
@@ -514,7 +514,7 @@ The Worker that serves the Cap'n Web RPC endpoint and proxies to `piccolo-core`.
 
 ### 10.1 Implementation Notes
 
-**Status:** Complete. 51 tests pass (383 total across all packages), coverage above thresholds. `pnpm biome check .` passes with 1 pre-existing `noNonNullAssertion` warning. `pnpm -r exec tsc --noEmit` has one pre-existing error in `packages/core/src/gateway.ts` (unrelated to step 10).
+**Status:** Complete. 51 tests pass (383 total across all packages), coverage above thresholds. `pnpm biome check .` passes with 1 pre-existing `noNonNullAssertion` warning. `pnpm -r exec tsc --noEmit` had one historical pre-existing error in the former `packages/core/src/gateway.ts` (unrelated to step 10).
 
 **Post-review corrections (all changes fully spec-synced):**
 1. `ITextUI.show*()` → `ITextUI.get*()` — direction fix: tool provides rendering text to gateway (pull), not gateway pushes to tool.