Merge pull request #1374 from corbitsdev/cl-10243-deepseek-v4-per-family-adapter

feat(provider): add deepseek-v4 per-family adapter

Author: TheGreatAxios

src/config/index.ts

@@ -59,6 +59,8 @@ import {
   GROK_SESSION_ID_OPTION,
 } from "../provider/grok-responses.js";
 import { BIFROST_PROVIDER } from "../provider/bifrost-adapter.js";
+import { DEEPSEEK_V4_PROVIDER } from "../provider/deepseek-v4-adapter.js";
+import { isDeepSeekV4Model } from "../provider/deepseek-v4-effort.js";
 import { isOllamaProviderId, ollamaOpenAIBaseURL } from "../provider/ollama.js";
 import { selectableGoModelIds } from "../provider/model-catalogs.js";
 import {
@@ -265,8 +267,9 @@ export function resolveMcpServers(
 }
 
 // Build the OpenAI-compatible InferenceSource the runtime consumes. `id` is the
-// user-facing name for this source (e.g. "zen"); `provider` is always
-// "openai-compatible" so the inference registry routes it to the right adapter.
+// user-facing name for this source (e.g. "zen"); `provider` routes to the
+// openai-compatible adapter, or the deepseek-v4 per-family adapter for V4
+// models, so the inference registry picks the right adapter.
 export function buildOpenAISource(fields: {
   id: string;
   baseURL: string;
@@ -287,7 +290,9 @@ export function buildOpenAISource(fields: {
   registerSourceSecret(fields.id, fields.apiKey);
   return {
     id: fields.id,
-    provider: "openai-compatible",
+    provider: isDeepSeekV4Model(fields.model)
+      ? DEEPSEEK_V4_PROVIDER
+      : "openai-compatible",
     baseURL: isOllamaProviderId(fields.id)
       ? ollamaOpenAIBaseURL(fields.baseURL)
       : normalizeOpenAICompatibleBaseURL(fields.baseURL),

src/config/inference-sources.test.ts

@@ -794,3 +794,33 @@ describe("session source bundles", () => {
     expect(bundle.defaultSource).toBe("openai");
   });
 });
+
+describe("deepseek-v4 provider stamping (PR A)", () => {
+  function v4SourceFor(model: string) {
+    return buildInferenceSourceForRef(
+      { provider: "ds", model },
+      {
+        sessionId: "s1",
+        catalog: [
+          {
+            name: "ds",
+            baseURL: "https://ds.example/v1",
+            apiKey: "k",
+            models: [model],
+          },
+        ],
+      },
+      undefined,
+    );
+  }
+
+  test("a deepseek-v4-* model stamps provider 'deepseek-v4'", () => {
+    const source = v4SourceFor("deepseek-v4-pro");
+    expect(source?.provider).toBe("deepseek-v4");
+  });
+
+  test("non-V4 custom models stay provider 'openai-compatible'", () => {
+    const source = v4SourceFor("gpt-5");
+    expect(source?.provider).toBe("openai-compatible");
+  });
+});

src/provider/deepseek-v4-adapter.test.ts

@@ -0,0 +1,226 @@
+import { describe, test, expect } from "bun:test";
+import type { ConversationTurn, InferenceOptions } from "@intx/types/runtime";
+import {
+  createDeepSeekV4Adapter,
+  DEEPSEEK_V4_PROVIDER,
+} from "./deepseek-v4-adapter.js";
+import { createOpenAICompatibleAdapter } from "./openai-compatible-adapter.js";
+
+const source = {
+  id: "test",
+  provider: DEEPSEEK_V4_PROVIDER,
+  baseURL: "https://example.test",
+  apiKey: "sk-test",
+  model: "deepseek-v4-pro",
+} as unknown as Parameters<typeof createDeepSeekV4Adapter>[0];
+
+const messages: ConversationTurn[] = [
+  {
+    role: "user",
+    content: [{ type: "text", text: "hi" }],
+  } as unknown as ConversationTurn,
+];
+
+function bodyFor(options: InferenceOptions): Record<string, unknown> {
+  const adapter = createDeepSeekV4Adapter(source);
+  const built = adapter.buildRequest(messages, "deepseek-v4-pro", options);
+  return JSON.parse(built.body) as Record<string, unknown>;
+}
+
+describe("deepseek-v4 adapter effort wiring", () => {
+  test.each(["xhigh", "max"] as const)(
+    "wires V4 effort %s through raw and turns thinking on",
+    (effort) => {
+      const body = bodyFor({
+        providerOptions: { reasoning_effort: effort },
+      } as InferenceOptions);
+      expect(body["reasoning_effort"]).toBe(effort);
+      expect(body["chat_template_kwargs"]).toEqual({ thinking: true });
+    },
+  );
+
+  test("none → chat_template_kwargs.thinking false and NO reasoning_effort", () => {
+    const body = bodyFor({
+      providerOptions: { reasoning_effort: "none" },
+    } as InferenceOptions);
+    expect(body["chat_template_kwargs"]).toEqual({ thinking: false });
+    expect("reasoning_effort" in body).toBe(false);
+  });
+
+  test("off-ladder effort is left untouched (no high→xhigh coercion)", () => {
+    const body = bodyFor({
+      providerOptions: { reasoning_effort: "high" },
+    } as InferenceOptions);
+    expect(body["reasoning_effort"]).toBe("high");
+    expect("chat_template_kwargs" in body).toBe(false);
+  });
+
+  test("unset effort does not inject empty chat_template_kwargs", () => {
+    const body = bodyFor({} as InferenceOptions);
+    expect("chat_template_kwargs" in body).toBe(false);
+    expect(body["reasoning_effort"]).toBeUndefined();
+  });
+});
+
+describe("deepseek-v4 adapter request settings", () => {
+  test("stream_options include_usage is true", () => {
+    const body = bodyFor({} as InferenceOptions);
+    expect(body["stream_options"]).toEqual({ include_usage: true });
+  });
+
+  test("top_p 0.95 and temperature 1.0 defaulted when absent", () => {
+    const body = bodyFor({} as InferenceOptions);
+    expect(body["top_p"]).toBe(0.95);
+    expect(body["temperature"]).toBe(1.0);
+  });
+
+  test("top_p and temperature preserved when providerOptions set them", () => {
+    const body = bodyFor({
+      providerOptions: { temperature: 0.7, top_p: 0.9 },
+    } as InferenceOptions);
+    expect(body["top_p"]).toBe(0.9);
+    expect(body["temperature"]).toBe(0.7);
+  });
+
+  test("provider-set stream_options is preserved (absent-only)", () => {
+    const body = bodyFor({
+      providerOptions: { stream_options: { include_usage: false } },
+    } as InferenceOptions);
+    expect(body["stream_options"]).toEqual({ include_usage: false });
+  });
+
+  test("chat_template_kwargs merge preserves existing keys", () => {
+    const body = bodyFor({
+      providerOptions: {
+        reasoning_effort: "xhigh",
+        chat_template_kwargs: { some: "existing" },
+      },
+    } as InferenceOptions);
+    expect(body["chat_template_kwargs"]).toEqual({
+      thinking: true,
+      some: "existing",
+    });
+  });
+
+  test("provider-set thinking is preserved (absent-only thinking)", () => {
+    const noneThenTrue = bodyFor({
+      providerOptions: {
+        reasoning_effort: "none",
+        chat_template_kwargs: { thinking: true },
+      },
+    } as InferenceOptions);
+    expect(noneThenTrue["chat_template_kwargs"]).toEqual({ thinking: true });
+
+    const xhighThenFalse = bodyFor({
+      providerOptions: {
+        reasoning_effort: "xhigh",
+        chat_template_kwargs: { thinking: false },
+      },
+    } as InferenceOptions);
+    expect(xhighThenFalse["chat_template_kwargs"]).toEqual({ thinking: false });
+  });
+});
+
+describe("deepseek-v4 adapter inherits base quirks", () => {
+  const withThinking: ConversationTurn[] = [
+    { role: "user", content: [{ type: "text", text: "hi" }] },
+    {
+      role: "assistant",
+      model: "deepseek-v4",
+      content: [
+        { type: "thinking", thinking: "ponder" },
+        { type: "text", text: "hello" },
+      ],
+    },
+    { role: "user", content: [{ type: "text", text: "again" }] },
+  ] as unknown as ConversationTurn[];
+
+  test("keeps reasoning_content on replayed assistant turns (inherited base)", () => {
+    const adapter = createDeepSeekV4Adapter(source);
+    const built = adapter.buildRequest(
+      withThinking,
+      "deepseek-v4",
+      {} as InferenceOptions,
+    );
+    const assistant = (
+      JSON.parse(built.body) as { messages: Record<string, unknown>[] }
+    ).messages.find((m) => m["role"] === "assistant");
+    expect(assistant?.["reasoning_content"]).toBe("ponder");
+  });
+
+  test("vendor-qualified V4 id keeps reasoning_content on replays (org/ prefix)", () => {
+    const adapter = createDeepSeekV4Adapter(source);
+    const built = adapter.buildRequest(
+      withThinking,
+      "deepseek-ai/DeepSeek-V4-Flash-0731",
+      {
+        providerOptions: { reasoning_effort: "xhigh" },
+      } as InferenceOptions,
+    );
+    const body = JSON.parse(built.body) as {
+      messages: Record<string, unknown>[];
+      reasoning_effort?: unknown;
+    };
+    const assistant = body.messages.find((m) => m["role"] === "assistant");
+    expect(assistant?.["reasoning_content"]).toBe("ponder");
+    expect(body.reasoning_effort).toBe("xhigh");
+  });
+
+  test("non-V4 select a generic body build (base short-circuits, no V4 fields)", () => {
+    const adapter = createDeepSeekV4Adapter({
+      ...source,
+      model: "gpt-5",
+    } as typeof source);
+    const built = adapter.buildRequest(
+      messages,
+      "gpt-5",
+      {} as InferenceOptions,
+    );
+    const body = JSON.parse(built.body) as Record<string, unknown>;
+    expect("top_p" in body).toBe(false);
+    expect("temperature" in body).toBe(false);
+    expect("stream_options" in body).toBe(false);
+    expect("chat_template_kwargs" in body).toBe(false);
+    expect(body["model"]).toBe("gpt-5");
+  });
+
+  test("keeps reasoning_content for non-DeepSeek models", () => {
+    const adapter = createDeepSeekV4Adapter({
+      ...source,
+      model: "kimi-k2",
+    } as typeof source);
+    const built = adapter.buildRequest(
+      withThinking,
+      "kimi-k2",
+      {} as InferenceOptions,
+    );
+    const assistant = (
+      JSON.parse(built.body) as { messages: Record<string, unknown>[] }
+    ).messages.find((m) => m["role"] === "assistant");
+    expect(assistant?.["reasoning_content"]).toBe("ponder");
+  });
+});
+
+describe("deepseek-v4 adapter byte-identical to base for non-V4", () => {
+  test("non-V4 body matches the patched generic base byte-for-byte", () => {
+    const baseAdapter = createOpenAICompatibleAdapter({
+      ...source,
+      model: "gpt-5",
+    } as typeof source);
+    const baseBuilt = baseAdapter.buildRequest(messages, "gpt-5", {
+      providerOptions: { reasoning_effort: "none" },
+    } as InferenceOptions);
+
+    const adapter = createDeepSeekV4Adapter({
+      ...source,
+      model: "gpt-5",
+    } as typeof source);
+    const built = adapter.buildRequest(messages, "gpt-5", {
+      providerOptions: { reasoning_effort: "none" },
+    } as InferenceOptions);
+
+    const baseBody = JSON.parse(baseBuilt.body) as Record<string, unknown>;
+    const body = JSON.parse(built.body) as Record<string, unknown>;
+    expect(body).toEqual(baseBody);
+  });
+});

src/provider/deepseek-v4-adapter.ts

@@ -0,0 +1,71 @@
+import { type ProviderAdapter } from "@intx/inference";
+import { createOpenAICompatibleAdapter } from "./openai-compatible-adapter.js";
+import {
+  DEEPSEEK_V4_MODEL_CARD,
+  isDeepSeekV4Model,
+  mapV4Effort,
+  V4_STREAM_OPTIONS,
+} from "./deepseek-v4-effort.js";
+
+export const DEEPSEEK_V4_PROVIDER = "deepseek-v4";
+
+type AdapterSource = Parameters<typeof createOpenAICompatibleAdapter>[0];
+
+// Wraps the patched openai-compatible adapter so V4 requests inherit the
+// providerOptions merge, NIM null-delta patch, and Accept header; this layer
+// only applies the V4 effort/sampling/stream quirks to the built body.
+export function createDeepSeekV4Adapter(
+  source: AdapterSource,
+): ProviderAdapter {
+  const base = createOpenAICompatibleAdapter(source);
+
+  const buildRequest: ProviderAdapter["buildRequest"] = (
+    messages,
+    model,
+    options,
+  ) => {
+    const req = base.buildRequest(messages, model, options);
+    if (!isDeepSeekV4Model(model)) return req;
+    const body = JSON.parse(req.body) as Record<string, unknown>;
+
+    // V4 effort ladder (none/xhigh/max): none drops reasoning_effort and
+    // turns thinking off; xhigh/max go out raw with thinking on; off-ladder
+    // or absent efforts are left untouched.
+    if ("reasoning_effort" in body) {
+      const effort = body["reasoning_effort"] as Parameters<
+        typeof mapV4Effort
+      >[0];
+      const wire = mapV4Effort(effort);
+      if (wire === null) {
+        delete body["reasoning_effort"];
+        mergeThinking(body, false);
+      } else if (wire !== undefined) {
+        body["reasoning_effort"] = wire;
+        mergeThinking(body, true);
+      }
+    }
+
+    if (body["top_p"] === undefined)
+      body["top_p"] = DEEPSEEK_V4_MODEL_CARD.topP;
+    if (body["temperature"] === undefined)
+      body["temperature"] = DEEPSEEK_V4_MODEL_CARD.temperature;
+    body["stream_options"] ??= V4_STREAM_OPTIONS.stream_options;
+
+    return { ...req, body: JSON.stringify(body) };
+  };
+
+  return { ...base, buildRequest };
+}
+
+// chat_template_kwargs clone with an absent-only thinking toggle (never mutate
+// the caller's providerOptions; provider-set thinking wins).
+function mergeThinking(body: Record<string, unknown>, thinking: boolean): void {
+  const existing =
+    body["chat_template_kwargs"] !== null &&
+    typeof body["chat_template_kwargs"] === "object"
+      ? (body["chat_template_kwargs"] as Record<string, unknown>)
+      : {};
+  const merged = { ...existing };
+  merged.thinking ??= thinking;
+  Object.assign(body, { chat_template_kwargs: merged });
+}

src/provider/deepseek-v4-effort.test.ts

@@ -13,13 +13,18 @@ describe("isDeepSeekV4Model", () => {
     "deepseek-v4",
     "deepseek-v4-pro",
     "deepseek-v4-flash",
+    "deepseek-v4.1-flash",
     "deepseek-v4-flash-vision-exp",
+    "deepseek-ai/DeepSeek-V4-Flash-0731",
+    "deepseek-ai/deepseek-v4-flash",
   ])("true for V4 id %s", (model) => {
     expect(isDeepSeekV4Model(model)).toBe(true);
   });
   test.each([
     "deepseek-v3",
     "deepseek-r1",
+    "deepseek-coder",
+    "deepseek-chat",
     "gpt-5",
     "kimi-k2",
     "claude-sonnet-4.5",