test(adapters): pin the outbound thinking contract — frognano #399 exonerated (root cause server-side)

packages/cli/src/adapters/openai-api-format.test.ts

@@ -0,0 +1,89 @@
+/**
+ * OpenAIAPIFormat — outbound-body thinking contract (#399 exoneration).
+ *
+ * #399 measured (07/10, hub + vLLM `myia_vllm-mini-frognano-4b` :5003) that
+ * frognano-4b requests served via the hub produced reasoning from the first
+ * token while the same prompt direct on :5003 never thinks. The suspected
+ * locus was "an untested field of the hub's outbound body flips the model's
+ * thinking template". These pins hold the claudish side of that answer:
+ *
+ * - The outbound body for a thinking-less request on a non-o1/o3 model carries
+ *   NO thinking-shaped field of any spelling. Measured 2026-10-11 by capturing
+ *   the real pipeline's body through an echo endpoint, byte-identical at
+ *   6f732eb9 (main on 07/10, the incident's era) and af52be5f (current): 198
+ *   bytes = model, messages, temperature, stream, stream_options, max_tokens.
+ * - `frognano-4b` matches neither `qwen` nor `alibaba`
+ *   (`matchesModelFamily` = startsWith | "/fam"), so QwenModelDialect — the
+ *   adapter whose passthrough could emit `enable_thinking` — never ran for it.
+ * - The flip no longer reproduces via the hub (2026-10-11): the exact captured
+ *   request (req-1-11725, minimal) AND a 26-tools consumer body (req-1-6396)
+ *   both answer immediately with no reasoning — same body both eras ⇒ the
+ *   change was server-side (:5003 deployment, refs vllm#70/#71).
+ *
+ * The o1/o3 branch (`isReasoningModel`) is the ONLY site allowed to emit
+ * `reasoning_effort`; if a future change starts emitting thinking controls for
+ * every model, these pins fire.
+ */
+
+import { describe, expect, test } from "bun:test";
+import { OpenAIAPIFormat } from "./openai-api-format.js";
+
+/** The exact inbound body of capture req-1-11725 (07/10 19:39:33Z) — minimal. */
+const R11725 = {
+  model: "frognano-4b",
+  max_tokens: 16,
+  messages: [{ role: "user", content: [{ type: "text", text: "Reponds exactement: ok" }] }],
+};
+
+const THINKING_FIELDS = [
+  "enable_thinking",
+  "reasoning_effort",
+  "thinking",
+  "thinking_budget",
+  "chat_template_kwargs",
+];
+
+function builtPayload(modelId: string, claudeRequest: any): any {
+  const adapter = new OpenAIAPIFormat(modelId);
+  // Mirror ComposedHandler's sequence: buildPayload, then the adapter's
+  // prepareRequest post-pass (tool-name encoding — where a dialect hook could
+  // otherwise inject fields after the build).
+  const messages = [{ role: "user", content: "Reponds exactement: ok" }];
+  const payload = adapter.buildPayload(claudeRequest, messages, []);
+  adapter.prepareRequest(payload, claudeRequest, { wireFormat: "openai-sse" as const });
+  return payload;
+}
+
+describe("OpenAIAPIFormat — #399: no thinking field leaves for a non-o1/o3 model", () => {
+  test("the r11725 body builds to exactly the six neutral fields (the 198-byte capture)", () => {
+    const payload = builtPayload("frognano-4b", R11725);
+    expect(Object.keys(payload).sort()).toEqual(
+      ["max_tokens", "messages", "model", "stream", "stream_options", "temperature"].sort()
+    );
+    for (const field of THINKING_FIELDS) {
+      expect(`${field}: ${field in payload}`).toBe(`${field}: false`);
+    }
+  });
+
+  test("an inbound `thinking` block does NOT become reasoning_effort off the o1/o3 lane", () => {
+    // Claude Code routinely sends an enabled thinking block; for frognano the
+    // body must still carry no thinking control — `isReasoningModel()` is the
+    // only gate that may map it, and it must stay o1/o3-only.
+    const payload = builtPayload("frognano-4b", {
+      ...R11725,
+      thinking: { type: "enabled", budget_tokens: 2048 },
+    });
+    for (const field of THINKING_FIELDS) {
+      expect(`${field}: ${field in payload}`).toBe(`${field}: false`);
+    }
+  });
+
+  test("o1 keeps the budget→reasoning_effort mapping (the only emission site stays open)", () => {
+    const payload = builtPayload("o1-mini", {
+      ...R11725,
+      model: "o1-mini",
+      thinking: { type: "enabled", budget_tokens: 8000 },
+    });
+    expect(payload.reasoning_effort).toBe("low");
+  });
+});