feat(ai)!: descope ai-runtime — remove vendor clients + builtinCost rate table (ADR-0024)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: dmealing

bun.lock

@@ -11,16 +11,6 @@
       "types": "./dist/index.d.ts",
       "default": "./dist/index.js"
     },
-    "./anthropic": {
-      "bun": "./src/anthropic.ts",
-      "types": "./dist/anthropic.d.ts",
-      "default": "./dist/anthropic.js"
-    },
-    "./openai": {
-      "bun": "./src/openai.ts",
-      "types": "./dist/openai.d.ts",
-      "default": "./dist/openai.js"
-    },
     "./langfuse": {
       "bun": "./src/langfuse.ts",
       "types": "./dist/langfuse.d.ts",
@@ -54,14 +44,10 @@
     "@metaobjectsdev/runtime-ts": "workspace:*"
   },
   "peerDependencies": {
-    "@anthropic-ai/sdk": ">=0.30.0",
-    "openai": ">=4.0.0",
     "langfuse": ">=3.0.0",
     "@opentelemetry/api": ">=1.0.0"
   },
   "peerDependenciesMeta": {
-    "@anthropic-ai/sdk": { "optional": true },
-    "openai": { "optional": true },
     "langfuse": { "optional": true },
     "@opentelemetry/api": { "optional": true }
   },

server/typescript/packages/ai-runtime/package.json

@@ -15,7 +15,7 @@ import {
   type LlmRequest,
   type LlmCompletion,
 } from "./client.js";
-import { builtinCost, type CostFn } from "./cost.js";
+import type { CostFn } from "./cost.js";
 
 export interface RunLlmCallInput {
   /** Discriminator / call identity. */
@@ -54,7 +54,6 @@ export async function runLlmCall(
 ): Promise<RunLlmCallResult> {
   const clock = deps.clock ?? systemClock;
   const ids = deps.ids ?? uuidIds;
-  const cost = deps.cost ?? builtinCost;
 
   const spanId = ids.next();
   const traceId = input.traceId ?? ids.next();
@@ -92,8 +91,8 @@ export async function runLlmCall(
   if (completion?.usage?.inputTokens !== undefined) recInput.inputTokens = completion.usage.inputTokens;
   if (completion?.usage?.outputTokens !== undefined) recInput.outputTokens = completion.usage.outputTokens;
   if (completion?.finishReason !== undefined) recInput.finishReason = completion.finishReason;
-  if (completion !== undefined) {
-    const c = cost(completion.model ?? input.request.model, completion.usage);
+  if (completion !== undefined && deps.cost !== undefined) {
+    const c = deps.cost(completion.model ?? input.request.model, completion.usage);
     if (c !== null) recInput.costMinor = c;
   }
 

server/typescript/packages/ai-runtime/src/anthropic.ts

@@ -2,29 +2,8 @@ import type { LlmUsage } from "./client.js";
 
 /**
  * Maps (model, usage) to a cost in integer USD minor units (cents), per the
- * field.currency wire contract. Returns null when the cost is unknown
- * (unknown model or missing usage) — never throws.
+ * field.currency wire contract. Returns null when unknown — never throws. The
+ * library ships NO rate table (ADR-0024); adopters supply their own CostFn
+ * (from their LLM library's usage + their own rates) via `deps.cost`.
  */
 export type CostFn = (model: string, usage: LlmUsage | undefined) => number | null;
-
-/**
- * Best-effort static rate table: USD dollars per 1,000,000 tokens.
- * Intentionally small — NOT a maintained pricing oracle. Adopters override by
- * passing their own CostFn to callLlm. Public model identifiers only.
- */
-const MODEL_RATES: Record<string, { inputPerM: number; outputPerM: number }> = {
-  "gpt-4o": { inputPerM: 2.5, outputPerM: 10 },
-  "gpt-4o-mini": { inputPerM: 0.15, outputPerM: 0.6 },
-  "claude-3-5-sonnet": { inputPerM: 3, outputPerM: 15 },
-  "claude-3-5-haiku": { inputPerM: 0.8, outputPerM: 4 },
-};
-
-export const builtinCost: CostFn = (model, usage) => {
-  if (usage === undefined) return null;
-  const rate = MODEL_RATES[model];
-  if (rate === undefined) return null;
-  const inTok = usage.inputTokens ?? 0;
-  const outTok = usage.outputTokens ?? 0;
-  const dollars = (inTok * rate.inputPerM + outTok * rate.outputPerM) / 1_000_000;
-  return Math.round(dollars * 100); // → integer cents
-};

server/typescript/packages/ai-runtime/src/call-loop.ts

@@ -12,7 +12,7 @@ export {
   type IdGen,
 } from "./client.js";
 
-export { builtinCost, type CostFn } from "./cost.js";
+export type { CostFn } from "./cost.js";
 
 export {
   callLlm,

server/typescript/packages/ai-runtime/src/cost.ts

@@ -11,7 +11,7 @@ class Capture extends NullRecorder {
 }
 
 describe("callLlm", () => {
-  test("happy path: CALL then persist the base row, captures latency/cost/ids", async () => {
+  test("happy path: CALL then persist the base row, captures latency/ids", async () => {
     // Deterministic seams.
     let t = 1000;
     const clock: Clock = { now: () => (t += 500) };
@@ -40,7 +40,6 @@ describe("callLlm", () => {
     expect(row.callType).toBe("Verdict");
     expect(row.spanId).toBe("id1"); // first id → span
     expect(row.traceId).toBe("id2"); // second id → trace (none supplied)
-    expect(row.costMinor).toBe(75); // builtinCost gpt-4o-mini @1M+1M
     expect(typeof row.latencyMs).toBe("number");
     expect(row.status).toBe("ok");
     expect(row.requestModel).toBe("gpt-4o-mini");
@@ -49,6 +48,42 @@ describe("callLlm", () => {
     expect("voResponse" in row).toBe(false);
   });
 
+  test("injected CostFn is called and costMinor is set", async () => {
+    const client: LlmClient = {
+      async complete(req) {
+        return {
+          body: JSON.stringify({ verdict: "ok" }),
+          model: req.model,
+          usage: { inputTokens: 1_000_000, outputTokens: 1_000_000 },
+        };
+      },
+    };
+    const rec = new Capture();
+    await callLlm(
+      { callType: "X", request: { prompt: "p", model: "m" } },
+      { client, recorder: rec, cost: () => 75 },
+    );
+    expect(rec.rows[0]!.costMinor).toBe(75);
+  });
+
+  test("no CostFn → costMinor is null (no built-in rate table)", async () => {
+    const client: LlmClient = {
+      async complete(req) {
+        return {
+          body: JSON.stringify({ verdict: "ok" }),
+          model: req.model,
+          usage: { inputTokens: 1_000_000, outputTokens: 1_000_000 },
+        };
+      },
+    };
+    const rec = new Capture();
+    await callLlm(
+      { callType: "X", request: { prompt: "p", model: "m" } },
+      { client, recorder: rec },
+    );
+    expect(rec.rows[0]!.costMinor).toBeNull();
+  });
+
   test("supplied traceId is preserved (no new trace id)", async () => {
     let n = 0;
     const ids: IdGen = { next: () => `s${++n}` };
@@ -162,7 +197,8 @@ describe("runLlmCall", () => {
     expect(result.input.requestModel).toBe("gpt-4o-mini");
     expect(typeof result.input.latencyMs).toBe("number");
     expect(typeof result.input.startedAt).toBe("string");
-    expect(result.input.costMinor).toBe(75);
+    // No CostFn injected → costMinor absent on input (no built-in rate table, ADR-0024).
+    expect(result.input.costMinor).toBeUndefined();
     expect(result.input.llmResponseText).toBe(JSON.stringify({ verdict: "ok" }));
   });