fix(provider): tighten deepseek-v4 effort policy

Author: TheGreatAxios

src/agent/model-family-policy.test.ts

@@ -411,39 +411,39 @@ describe("resolveModelFamilyPolicy", () => {
 
     test("orchestrators get the d1 primary residual, no leaf role body", () => {
       const p = v4({ orchestrator: true });
-      expect(p.family).toBe("deepseek-v4");
+      expect(p.family).toBe("deepseek");
       expect(p.promptResidual).toBeDefined();
       expect(p.promptResidual).toContain("DeepSeek V4 Flash");
       expect(p.leafRoleBody).toBeUndefined();
     });
 
     test("coder leaves get the tuned slim coder body, no residual", () => {
       const p = v4({ orchestrator: false, directorId: "coder" });
-      expect(p.family).toBe("deepseek-v4");
+      expect(p.family).toBe("deepseek");
       expect(p.leafRoleBody).toBeDefined();
       expect(p.leafRoleBody).toContain("Coder");
       expect(p.promptResidual).toBeUndefined();
     });
 
     test("reviewer leaves get the tuned slim reviewer body", () => {
       const p = v4({ orchestrator: false, directorId: "reviewer" });
-      expect(p.family).toBe("deepseek-v4");
+      expect(p.family).toBe("deepseek");
       expect(p.leafRoleBody).toBeDefined();
       expect(p.leafRoleBody).toContain("Reviewer");
       expect(p.promptResidual).toBeUndefined();
     });
 
     test("explorer leaves get the g2 leaf residual only", () => {
       const p = v4({ orchestrator: false, directorId: "explorer" });
-      expect(p.family).toBe("deepseek-v4");
+      expect(p.family).toBe("deepseek");
       expect(p.promptResidual).toBeDefined();
       expect(p.promptResidual).toContain("DeepSeek V4 Flash");
       expect(p.leafRoleBody).toBeUndefined();
     });
 
     test("a leaf without a tuned director id gets the default policy, no body/residual", () => {
       const p = v4({ orchestrator: false });
-      expect(p.family).toBe("deepseek-v4");
+      expect(p.family).toBe("deepseek");
       expect(p.leafRoleBody).toBeUndefined();
       expect(p.promptResidual).toBeUndefined();
     });

src/agent/model-family-policy.ts

@@ -299,7 +299,7 @@ export function resolveModelFamilyPolicy(input: {
       // Generic gpt prompts are byte-identical — only astra carries the
       // residual.
       return { family, ...ASTRA_POLICY };
-    case "deepseek-v4": {
+    case "deepseek": {
       // Winning d1+slim config: d1 primary residual on orchestrators, g2 leaf
       // residual on explorer leaves, tuned slim coder/reviewer bodies.
       const directorId = input.directorId;

src/config/index.ts

@@ -60,7 +60,7 @@ import {
 } 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 { isDeepSeekModel } from "../provider/deepseek-v4-effort.js";
 import { isOllamaProviderId, ollamaOpenAIBaseURL } from "../provider/ollama.js";
 import { selectableGoModelIds } from "../provider/model-catalogs.js";
 import {
@@ -290,7 +290,7 @@ export function buildOpenAISource(fields: {
   registerSourceSecret(fields.id, fields.apiKey);
   return {
     id: fields.id,
-    provider: isDeepSeekV4Model(fields.model)
+    provider: isDeepSeekModel(fields.model)
       ? DEEPSEEK_V4_PROVIDER
       : "openai-compatible",
     baseURL: isOllamaProviderId(fields.id)

src/config/inference-sources.test.ts

@@ -814,12 +814,15 @@ describe("deepseek-v4 provider stamping (PR A)", () => {
     );
   }
 
-  test("a deepseek-v4-* model stamps provider 'deepseek-v4'", () => {
+  test("a deepseek-* model stamps provider 'deepseek-v4' (family routing)", () => {
     const source = v4SourceFor("deepseek-v4-pro");
     expect(source?.provider).toBe("deepseek-v4");
+
+    const preV4 = v4SourceFor("deepseek-v3");
+    expect(preV4?.provider).toBe("deepseek-v4");
   });
 
-  test("non-V4 custom models stay provider 'openai-compatible'", () => {
+  test("non-deepseek custom models stay provider 'openai-compatible'", () => {
     const source = v4SourceFor("gpt-5");
     expect(source?.provider).toBe("openai-compatible");
   });

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

@@ -47,12 +47,12 @@ describe("deepseek-v4 adapter effort wiring", () => {
     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("off-ladder effort is rejected (no pass-through to the wire)", () => {
+    expect(() =>
+      bodyFor({
+        providerOptions: { reasoning_effort: "high" },
+      } as InferenceOptions),
+    ).toThrow(/does not support reasoning effort "high"/);
   });
 
   test("unset effort does not inject empty chat_template_kwargs", () => {
@@ -223,4 +223,29 @@ describe("deepseek-v4 adapter byte-identical to base for non-V4", () => {
     const body = JSON.parse(built.body) as Record<string, unknown>;
     expect(body).toEqual(baseBody);
   });
+
+  test("pre-V4 deepseek gets no V4 quirks (family route, V4-variant gate)", () => {
+    const baseAdapter = createOpenAICompatibleAdapter({
+      ...source,
+      model: "deepseek-v3",
+    } as typeof source);
+    const baseBuilt = baseAdapter.buildRequest(messages, "deepseek-v3", {
+      providerOptions: { reasoning_effort: "none" },
+    } as InferenceOptions);
+
+    const adapter = createDeepSeekV4Adapter({
+      ...source,
+      model: "deepseek-v3",
+    } as typeof source);
+    const built = adapter.buildRequest(messages, "deepseek-v3", {
+      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);
+    expect("top_p" in body).toBe(false);
+    expect("temperature" in body).toBe(false);
+    expect("stream_options" in body).toBe(false);
+  });
 });

src/provider/deepseek-v4-adapter.ts

@@ -2,6 +2,7 @@ import { type ProviderAdapter } from "@intx/inference";
 import { createOpenAICompatibleAdapter } from "./openai-compatible-adapter.js";
 import {
   DEEPSEEK_V4_MODEL_CARD,
+  isDeepSeekModel,
   isDeepSeekV4Model,
   mapV4Effort,
   V4_STREAM_OPTIONS,
@@ -13,7 +14,9 @@ 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.
+// routes the whole DeepSeek family through this family adapter and applies the
+// V4 effort/sampling/stream quirks to the built body only for the V4 variant
+// (pre-V4 deepseek stays on the generic openai-compatible body, byte-identical).
 export function createDeepSeekV4Adapter(
   source: AdapterSource,
 ): ProviderAdapter {
@@ -25,12 +28,12 @@ export function createDeepSeekV4Adapter(
     options,
   ) => {
     const req = base.buildRequest(messages, model, options);
-    if (!isDeepSeekV4Model(model)) return req;
+    if (!isDeepSeekModel(model) || !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.
+    // turns thinking off; xhigh/max go out raw with thinking on. Off-ladder
+    // efforts throw inside mapV4Effort (rejected, never passed to the wire).
     if ("reasoning_effort" in body) {
       const effort = body["reasoning_effort"] as Parameters<
         typeof mapV4Effort
@@ -39,7 +42,7 @@ export function createDeepSeekV4Adapter(
       if (wire === null) {
         delete body["reasoning_effort"];
         mergeThinking(body, false);
-      } else if (wire !== undefined) {
+      } else {
         body["reasoning_effort"] = wire;
         mergeThinking(body, true);
       }

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

@@ -2,12 +2,43 @@ import { describe, test, expect } from "bun:test";
 import {
   DEEPSEEK_V4_EFFORTS,
   DEEPSEEK_V4_MODEL_CARD,
+  DEEPSEEK_V4_ROLE_EFFORT,
+  isDeepSeekModel,
   isDeepSeekV4Model,
   mapV4Effort,
   V4_STREAM_OPTIONS,
   v4Thinking,
 } from "./deepseek-v4-effort.js";
 
+describe("isDeepSeekModel", () => {
+  test.each([
+    "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",
+    "deepseek-v3",
+    "deepseek-r1",
+    "deepseek-coder",
+    "deepseek-chat",
+    "deepseek-ai/DeepSeek-Chat",
+  ])("true for deepseek family id %s", (model) => {
+    expect(isDeepSeekModel(model)).toBe(true);
+  });
+  test.each([
+    "gpt-5",
+    "kimi-k2",
+    "claude-sonnet-4.5",
+    "glm-5.3",
+    "grok-4.6",
+    "some-proxy-deepseek-pro",
+  ])("false for non-deepseek id %s", (model) => {
+    expect(isDeepSeekModel(model)).toBe(false);
+  });
+});
+
 describe("isDeepSeekV4Model", () => {
   test.each([
     "deepseek-v4",
@@ -25,6 +56,7 @@ describe("isDeepSeekV4Model", () => {
     "deepseek-r1",
     "deepseek-coder",
     "deepseek-chat",
+    "deepseek-ai/DeepSeek-Chat",
     "gpt-5",
     "kimi-k2",
     "claude-sonnet-4.5",
@@ -42,6 +74,28 @@ describe("DEEPSEEK_V4_EFFORTS", () => {
   });
 });
 
+describe("DEEPSEEK_V4_ROLE_EFFORT", () => {
+  test("coder defaults to max", () => {
+    expect(DEEPSEEK_V4_ROLE_EFFORT.coder).toBe("max");
+  });
+  test("dispatch/planner/reviewer are max and other leaves xhigh", () => {
+    expect(DEEPSEEK_V4_ROLE_EFFORT.dispatch).toBe("max");
+    expect(DEEPSEEK_V4_ROLE_EFFORT.planner).toBe("max");
+    expect(DEEPSEEK_V4_ROLE_EFFORT.reviewer).toBe("max");
+    for (const role of [
+      "explorer",
+      "artist",
+      "qa-lead",
+      "prober",
+      "designer",
+      "shakespeare",
+      "warden",
+    ] as const) {
+      expect(DEEPSEEK_V4_ROLE_EFFORT[role]).toBe("xhigh");
+    }
+  });
+});
+
 describe("mapV4Effort", () => {
   test.each([
     ["xhigh", "xhigh"],
@@ -53,9 +107,11 @@ describe("mapV4Effort", () => {
     expect(mapV4Effort("none")).toBeNull();
   });
   test.each(["low", "medium", "high", "ultra", "minimal"] as const)(
-    "leaves off-ladder %s untouched (undefined)",
+    "rejects off-ladder %s with a clear error (no pass-through)",
     (effort) => {
-      expect(mapV4Effort(effort)).toBeUndefined();
+      expect(() => mapV4Effort(effort)).toThrow(
+        `DeepSeek V4 does not support reasoning effort "${effort}"`,
+      );
     },
   );
 });

src/provider/deepseek-v4-effort.ts

@@ -2,15 +2,15 @@ import type { ReasoningEffort } from "../agent/profile-types.js";
 import type { DirectorId } from "../agent/directors/types.js";
 
 /**
- * Per-role V4 effort defaults. coder is the cheapest spawn: 'low' clamps to
- * the V4 'none' wire rung (thinking off), per the operator's "simplest
- * implementer = none" intent. Every other role takes its rung from the map.
+ * Per-role V4 effort defaults. dispatch/planner/reviewer dispatch and judge —
+ * max. coder defaults to max (operator-confirmed). Every other leaf reasons at
+ * xhigh.
  */
 export const DEEPSEEK_V4_ROLE_EFFORT: Record<DirectorId, ReasoningEffort> = {
   dispatch: "max",
   planner: "max",
   reviewer: "max",
-  coder: "low",
+  coder: "max",
   explorer: "xhigh",
   artist: "xhigh",
   "qa-lead": "xhigh",
@@ -20,13 +20,23 @@ export const DEEPSEEK_V4_ROLE_EFFORT: Record<DirectorId, ReasoningEffort> = {
   warden: "xhigh",
 };
 
-// DeepSeek V4 effort translator shared by the adapter and the effort picker.
-// V4 ships under several ids across two catalogs — deepseek-v4-pro/-flash/
-// -flash-vision-exp — and under a vendor-qualified host id like
-// deepseek-ai/DeepSeek-V4-Flash-0731. Match the `deepseek-v4` segment at the
-// start or after a `/`, so both bare and qualified ids resolve while pre-V4
-// ids (deepseek-coder, deepseek-chat) do not. This mirrors the grok/kimi/astra
-// leaf predicates, which also accept an org-qualified segment.
+// DeepSeek family predicate shared by the adapter router and the family policy.
+// The family ships under several ids across two catalogs — deepseek-v4-pro/
+// -flash/-flash-vision-exp and deepseek-chat/deepseek-coder/deepseek-v3 — and
+// under a vendor-qualified host id like deepseek-ai/DeepSeek-V4-Flash-0731.
+// Match the `deepseek` segment at the start or after a `/`, so both bare and
+// qualified ids resolve across the whole family (including future versions).
+// This mirrors the grok/kimi/astra leaf predicates, which also accept an
+// org-qualified segment.
+export function isDeepSeekModel(model: string): boolean {
+  return /(^|\/)deepseek/i.test(model.trim());
+}
+
+// V4-variant detection: distinguishes the V4 model line from pre-V4 DeepSeek
+// (deepseek-chat, deepseek-coder, deepseek-v3). Only where the wire semantics
+// genuinely differ between V4 and pre-V4 deepseek — the V4 adapter's model-card
+// and effort quirks, the reasoning_content keep-vs-strip behavior, and the
+// none/xhigh/max effort ladder — do callers gate on this variant predicate.
 export function isDeepSeekV4Model(model: string): boolean {
   return /(^|\/)deepseek-v4/i.test(model.trim());
 }
@@ -43,15 +53,17 @@ export const DEEPSEEK_V4_EFFORTS: readonly ReasoningEffort[] = [
 /**
  * Translate a V4 effort to its raw wire value: `xhigh`/`max` pass through
  * verbatim, `none` maps to null (the caller drops `reasoning_effort`, since
- * `reasoning_effort:"none"` is illegal on the wire). Off-ladder values return
- * undefined — the caller leaves `reasoning_effort` untouched rather than
- * coercing it onto a V4 rung.
+ * `reasoning_effort:"none"` is illegal on the wire). The V4 ladder is strictly
+ * none/xhigh/max: any off-ladder effort arriving on a manual/exec/config path
+ * is rejected with a clear error rather than coerced or passed through, so no
+ * off-ladder value can reach the wire for a V4 model.
  */
-export function mapV4Effort(
-  effort: ReasoningEffort,
-): "xhigh" | "max" | null | undefined {
+export function mapV4Effort(effort: ReasoningEffort): "xhigh" | "max" | null {
   if (effort === "none") return null;
-  return effort === "xhigh" || effort === "max" ? effort : undefined;
+  if (effort === "xhigh" || effort === "max") return effort;
+  throw new Error(
+    `DeepSeek V4 does not support reasoning effort "${effort}" (supported: none, xhigh, max).`,
+  );
 }
 
 /**

src/provider/openai-compatible-adapter.ts

@@ -1,7 +1,7 @@
 import { type BuiltRequest, type ProviderAdapter } from "@intx/inference";
 import { createOpenAIAdapter } from "@intx/inference/providers";
 import { normalizeNullDeltaFields } from "./null-delta-fields.js";
-import { isDeepSeekV4Model } from "./deepseek-v4-effort.js";
+import { isDeepSeekModel, isDeepSeekV4Model } from "./deepseek-v4-effort.js";
 
 // The stock OpenAI adapter builds the request body from a fixed set of fields
 // (max_tokens, temperature, tools, messages, response_format) and ignores
@@ -51,7 +51,7 @@ export function createOpenAICompatibleAdapter(
     // keeps each turn's reasoning on replay (handled by its own adapter), so only
     // pre-V4 DeepSeek strips it here.
     const m = model.toLowerCase().trim();
-    const isDeepSeek = m.includes("deepseek");
+    const isDeepSeek = isDeepSeekModel(m);
     const stripReasoning = isDeepSeek && !isDeepSeekV4Model(m);
     needsDeepSeekPatch = isDeepSeek;
     if (!hasProviderOptions && !stripReasoning) return ensureAccept(built);

src/subagent/agent-fleet-effort.test.ts

@@ -390,15 +390,14 @@ describe("deepseek-v4 per-role effort policy (CL-10294)", () => {
     return worker.provider.reasoningEffort as ReasoningEffort;
   }
 
-  test("planner/reviewer max, coder cheapest, other leaves xhigh", async () => {
-    // V4's native ladder is ["none","xhigh","max"]. coder's low clamps to the
-    // cheapest wire rung (none, thinking off) per the operator's "simplest
-    // implementer = none" intent; planner/reviewer resolve max. dispatch is the
-    // primary session (not a spawned worker) and already defaults to max for
+  test("planner/reviewer/coder max, other leaves xhigh", async () => {
+    // V4's native ladder is ["none","xhigh","max"]. coder defaults to max
+    // (operator-confirmed, PR #1377); planner/reviewer resolve max. dispatch is
+    // the primary session (not a spawned worker) and already defaults to max for
     // V4 via defaultEffortForModel.
     expect(await captureEffort("planner")).toBe("max");
     expect(await captureEffort("reviewer")).toBe("max");
-    expect(await captureEffort("coder")).toBe("none");
+    expect(await captureEffort("coder")).toBe("max");
     expect(await captureEffort("explorer")).toBe("xhigh");
     expect(await captureEffort("artist")).toBe("xhigh");
     expect(await captureEffort("qa-lead")).toBe("xhigh");