@volter/twin-deepseek 0.1.0 → 0.1.2

package/README.md

@@ -54,9 +54,9 @@ capability with a failable negative verify, grounded in
 | usage cache fields | **`prompt_cache_hit_tokens` / `prompt_cache_miss_tokens`** | `prompt_tokens_details.cached_tokens` |
 | reasoning | **`reasoning_content`** on the message and the delta | `reasoning` on o-series responses only |
 | finish reasons | `stop`, `length`, `content_filter`, `tool_calls`, **`insufficient_system_resource`** | no `insufficient_system_resource` |
-| end-user id | **`user_id`** (`^[a-zA-Z0-9_-]+$`, ≤512) | `user` |
+| end-user id | **`user_id`** (`^[a-zA-Z0-9_-]+$`, ≤512); OpenAI's `user` is accepted and ignored (the docs are silent; ruling in `deepseek-twin.ts`, `OPENAI_USER`) | `user` |
 | structured output (chat) | `text`, `json_object` — **no `json_schema` on `/chat/completions`** (its Responses API's `text.format` does take one) | `text`, `json_object`, `json_schema` |
-| not parameters at all | **`n`, `seed`, `logit_bias`, `top_k`, `user`, `max_completion_tokens`, `service_tier`, `parallel_tool_calls`, `functions`** — each a 422 by name | all accepted |
+| not parameters at all | **`n`, `seed`, `logit_bias`, `top_k`, `max_completion_tokens`, `service_tier`, `parallel_tool_calls`, `functions`** — each a 422 by name | all accepted |
 | deprecated-but-accepted | `frequency_penalty`, `presence_penalty` — **no effect, never an error** | still honoured |
 | ignored-but-accepted | `temperature`, `top_p` while thinking is on — **no effect, never an error** | honoured |
 | beta-only | assistant `prefix: true`, `strict: true` tools, FIM `/beta/completions` | n/a |

package/dist/src/deepseek-capabilities.js

@@ -210,12 +210,12 @@ export const DEEPSEEK_CAPABILITIES = [
         const sixteen = await h({ m: 'POST', p: CHAT_PATH, b: CHAT({ stop: Array.from({ length: 16 }, (_, i) => `s${i}`) }) });
         return many.status === 422 && ok(sixteen);
     })),
-    done('deepseek.chat.rejects_openai_only_params', 'chat', "The OpenAI parameters DeepSeek's closed table does not declare (n, seed, logit_bias, top_k, user, max_completion_tokens, service_tier, parallel_tool_calls, functions) are refused BY NAME with 422", 'api', 'core', () => withRoot(async (h) => {
+    done('deepseek.chat.rejects_openai_only_params', 'chat', "The OpenAI parameters DeepSeek's closed table does not declare (n, seed, logit_bias, top_k, max_completion_tokens, service_tier, parallel_tool_calls, functions) are refused BY NAME with 422", 'api', 'core', () => withRoot(async (h) => {
         // THE flagship OpenAI-divergence check. A twin copied from an OpenAI-shaped exemplar serves
         // every one of these happily, which is the inherited-permissiveness false-green
         // ADDING_A_TWIN.md §0 names. Each must be 422 — DeepSeek's "Invalid Parameters" — not 400.
         for (const [key, value] of [
-            ['n', 2], ['seed', 42], ['logit_bias', { '1': 1 }], ['top_k', 5], ['user', 'u1'],
+            ['n', 2], ['seed', 42], ['logit_bias', { '1': 1 }], ['top_k', 5],
             ['max_completion_tokens', 32], ['service_tier', 'flex'], ['parallel_tool_calls', false],
             ['functions', [{ name: 'f' }]], ['store', true], ['metadata', { a: 'b' }],
         ]) {
@@ -285,6 +285,17 @@ export const DEEPSEEK_CAPABILITIES = [
         const atCap = await h({ m: 'POST', p: CHAT_PATH, b: CHAT({ user_id: 'a'.repeat(512) }) });
         return ok(good) && bad.status === 422 && long.status === 422 && ok(atCap);
     })),
+    done('deepseek.chat.openai_user_ignored', 'chat', "OpenAI's `user` (which OpenAI-compatible software such as LibreChat sends) is accepted and has no effect: it is not DeepSeek's `user_id`, and a non-string is 422", 'api', 'common', () => withRoot(async (h) => {
+        const plain = await h({ m: 'POST', p: CHAT_PATH, b: CHAT() });
+        // a value user_id's charset refuses: accepted here, so it is not being read as user_id
+        const withUser = await h({ m: 'POST', p: CHAT_PATH, b: CHAT({ user: 'sam@friends.test' }) });
+        const both = await h({ m: 'POST', p: CHAT_PATH, b: CHAT({ user: 'u1', user_id: 'bad id' }) });
+        const typed = await h({ m: 'POST', p: CHAT_PATH, b: CHAT({ user: 42 }) });
+        const answer = (r) => JSON.stringify([r.body.choices?.[0]?.message?.content, r.body.usage?.prompt_tokens]);
+        return ok(plain) && ok(withUser) && answer(withUser) === answer(plain)
+            && both.status === 422 && msg(both).includes("'user_id'")
+            && typed.status === 422 && msg(typed).includes("'user'");
+    })),
     done('deepseek.chat.message_name', 'chat', 'Participant `name` is accepted on system/user/assistant turns and counted into prompt tokens', 'api', 'niche', () => withRoot(async (h) => {
         const named = await h({ m: 'POST', p: CHAT_PATH, b: { model: MODEL, messages: [{ role: 'user', content: 'hi', name: 'customer_alpha' }] } });
         const plain = await h({ m: 'POST', p: CHAT_PATH, b: { model: MODEL, messages: [{ role: 'user', content: 'hi' }] } });

package/dist/src/deepseek-conformance.js

@@ -195,7 +195,8 @@ const REJECTIONS = [
     { name: 'seed', method: 'POST', path: '/chat/completions', body: { ...CHAT, seed: 42 }, status: 422, why: "'seed' is not a DeepSeek parameter" },
     { name: 'logit_bias', method: 'POST', path: '/chat/completions', body: { ...CHAT, logit_bias: { '1': 1 } }, status: 422, why: "'logit_bias' is not a DeepSeek parameter" },
     { name: 'top_k', method: 'POST', path: '/chat/completions', body: { ...CHAT, top_k: 5 }, status: 422, why: "'top_k' is not a DeepSeek parameter" },
-    { name: 'user', method: 'POST', path: '/chat/completions', body: { ...CHAT, user: 'u1' }, status: 422, why: "DeepSeek's end-user identifier is 'user_id', not OpenAI's 'user'" },
+    // OpenAI's `user` as a string is accepted and ignored (deepseek-twin.ts, OPENAI_USER); only a non-string is refused
+    { name: 'user (not a string)', method: 'POST', path: '/chat/completions', body: { ...CHAT, user: 42 }, status: 422, why: "OpenAI's 'user' is taken and ignored, but only as a string" },
     { name: 'max_completion_tokens', method: 'POST', path: '/chat/completions', body: { ...CHAT, max_completion_tokens: 10 }, status: 422, why: "DeepSeek declares 'max_tokens' only" },
     { name: 'service_tier', method: 'POST', path: '/chat/completions', body: { ...CHAT, service_tier: 'flex' }, status: 422, why: 'DeepSeek has no service tiers' },
     { name: 'parallel_tool_calls', method: 'POST', path: '/chat/completions', body: { ...CHAT, parallel_tool_calls: false }, status: 422, why: 'not in DeepSeek\'s parameter table' },

package/dist/src/deepseek-twin.js

@@ -25,15 +25,17 @@
 //     includes two OpenAI does not use: 402 Insufficient Balance and 422 Invalid Parameters
 //     (api-docs.deepseek.com/quick_start/error_codes). OpenAI 400s a bad parameter; DeepSeek 422s
 //     it. 400 is reserved for "Invalid request body format".
-//   • `n`, `seed`, `logit_bias`, `top_k`, `user`, `max_completion_tokens`, `service_tier`,
+//   • `n`, `seed`, `logit_bias`, `top_k`, `max_completion_tokens`, `service_tier`,
 //     `parallel_tool_calls`, `functions` are NOT DeepSeek parameters. The accepted set is a closed
 //     documented table (api-docs.deepseek.com/api/create-chat-completion), which §6 licenses as a
-//     literal-allowlist oracle — so anything outside CHAT_PARAMS is 422 by name.
+//     literal-allowlist oracle — so anything outside CHAT_PARAMS is 422 by name. OpenAI's `user` is
+//     the one exception, accepted and ignored (OPENAI_USER below).
 //   • `response_format` accepts `text` and `json_object` ONLY. `json_schema` is an OpenAI feature;
 //     `@ai-sdk/deepseek` confirms it by INJECTING the schema into a system message instead of
 //     sending `response_format.json_schema` (src/chat/convert-to-deepseek-chat-messages.ts).
-//   • `user_id`, not `user` — with a regex and a 512-character cap the SDK enforces client-side and
-//     the vendor documents server-side.
+//   • `user_id` is DeepSeek's end-user identifier — with a regex and a 512-character cap the SDK
+//     enforces client-side and the vendor documents server-side. OpenAI's `user` is taken and has
+//     no effect (OPENAI_USER).
 //   • Assistant-prefix completion and strict tool calls exist ONLY under the `/beta` base URL.
 //   • `frequency_penalty` / `presence_penalty` are DEPRECATED but MUST NOT ERROR, and
 //     `temperature` / `top_p` are IGNORED (not rejected) while thinking is enabled: "setting these
@@ -226,13 +228,31 @@ function parseJson(body) {
         return { error: invalidFormat('the request body is not valid JSON') };
     }
 }
+/**
+ * OpenAI's `user`, taken and IGNORED — where DeepSeek's documentation stops and the twin decides
+ * (read 2026-09-28). The reference table (api-docs.deepseek.com/api/create-chat-completion) does not
+ * list `user` (its end-user field is `user_id`) and says nothing of a key it does not list. Against
+ * that silence: DeepSeek states "The DeepSeek API uses an API format compatible with OpenAI/Anthropic"
+ * and invites "the OpenAI/Anthropic SDK or softwares compatible with the OpenAI/Anthropic API"
+ * (api-docs.deepseek.com, "Your First API Call"); its Anthropic-compatible surface answers the fields
+ * it does not use as "Ignored" rather than refusing them (api-docs.deepseek.com/guides/anthropic_api:
+ * metadata's "user_id is supported, others are ignored"); and OpenAI-compatible software sends `user`
+ * to DeepSeek as a matter of course — LibreChat's DeepSeek endpoint, through LangChain's ChatOpenAI,
+ * sends it on every chat (cookbook/librechat, finding: the twin's 422 was retried into a hang), which
+ * could not be DeepSeek's supported configuration if the real API refused it. So the twin accepts a
+ * string `user` and gives it no effect: it is NOT `user_id` (no isolation, no content-safety
+ * identity) and is not validated against `user_id`'s charset. A non-string `user` is 422, as OpenAI's
+ * declared type is a string. The rest of the table stays closed: this ruling is evidence for `user`
+ * alone, and none has been found for `n`, `seed` and the others.
+ */
+const OPENAI_USER = 'user';
 /**
  * THE CLOSED PARAMETER TABLE for `POST /chat/completions`
  * (api-docs.deepseek.com/api/create-chat-completion, read 2026-08-31) plus the two streaming keys
  * and the top-level `reasoning_effort` `@ai-sdk/deepseek` actually sends.
  *
  * §6 licenses a literal allowlist as an ORACLE precisely when the vendor documents a closed set,
- * and this is one — which is what gives `n`, `seed`, `logit_bias`, `top_k`, `user`,
+ * and this is one — which is what gives `n`, `seed`, `logit_bias`, `top_k`,
  * `max_completion_tokens`, `service_tier` and `parallel_tool_calls` a *checkable* refusal instead
  * of the silent acceptance an OpenAI-copied twin gives them.
  *
@@ -244,7 +264,7 @@ function parseJson(body) {
 const CHAT_PARAMS = new Set([
     'messages', 'model', 'thinking', 'max_tokens', 'response_format', 'stop', 'stream',
     'stream_options', 'temperature', 'top_p', 'tools', 'tool_choice', 'logprobs', 'top_logprobs',
-    'user_id', 'frequency_penalty', 'presence_penalty', 'reasoning_effort',
+    'user_id', 'frequency_penalty', 'presence_penalty', 'reasoning_effort', OPENAI_USER,
 ]);
 /** The closed table for `POST /beta/completions` (api-docs.deepseek.com/api/create-completion). */
 const FIM_PARAMS = new Set([
@@ -432,7 +452,11 @@ function validateChat(params, beta) {
             return { error: invalidParameters("'top_logprobs' requires 'logprobs' to be true") };
         topLogprobs = v;
     }
-    // (10) user_id — DeepSeek's end-user identifier. NOT OpenAI's `user` (which CHAT_PARAMS refuses).
+    // (10) user_id — DeepSeek's end-user identifier. NOT OpenAI's `user`, which is taken and ignored
+    //      (OPENAI_USER): only its type is checked.
+    if (params[OPENAI_USER] !== undefined && params[OPENAI_USER] !== null && typeof params[OPENAI_USER] !== 'string') {
+        return { error: invalidParameters(`'${OPENAI_USER}' must be a string`) };
+    }
     let userId;
     if (params.user_id !== undefined && params.user_id !== null) {
         const v = params.user_id;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@volter/twin-deepseek",
-  "version": "0.1.0",
+  "version": "0.1.2",
   "description": "Local DeepSeek twin — a faithful, stateful local DeepSeek Platform API your real `@ai-sdk/deepseek` talks to unmodified. The model is stubbed (deterministic), but the protocol envelope is vendor-faithful and so are DeepSeek's OpenAI-DIVERGENT REFUSALS: 422 for a bad parameter and 402 for a drained balance (not OpenAI's 400), no /v1 path segment, no json_schema/n/seed/logit_bias/user, beta-only prefix completion and strict tools, reasoning_content, and a real KV-cache-split usage ledger. Built on @volter/world-core.",
   "keywords": [
     "twin",
@@ -52,10 +52,10 @@
     "postpack": "node ../../../scripts/publish/prepare-publish.mjs postpack"
   },
   "peerDependencies": {
-    "@volter/world-core": "2.0.0"
+    "@volter/world-core": "2.0.2"
   },
   "devDependencies": {
-    "@volter/world-core": "2.0.0",
+    "@volter/world-core": "2.0.2",
     "@volter/world-tooling": "0.1.0",
     "@ai-sdk/deepseek": "^3.0.37",
     "@types/bun": "^1.2.20",

package/src/deepseek-capabilities.ts

@@ -225,12 +225,12 @@ export const DEEPSEEK_CAPABILITIES: CapabilitySpec[] = [
     return many.status === 422 && ok(sixteen);
   })),
 
-  done('deepseek.chat.rejects_openai_only_params', 'chat', "The OpenAI parameters DeepSeek's closed table does not declare (n, seed, logit_bias, top_k, user, max_completion_tokens, service_tier, parallel_tool_calls, functions) are refused BY NAME with 422", 'api', 'core', () => withRoot(async (h) => {
+  done('deepseek.chat.rejects_openai_only_params', 'chat', "The OpenAI parameters DeepSeek's closed table does not declare (n, seed, logit_bias, top_k, max_completion_tokens, service_tier, parallel_tool_calls, functions) are refused BY NAME with 422", 'api', 'core', () => withRoot(async (h) => {
     // THE flagship OpenAI-divergence check. A twin copied from an OpenAI-shaped exemplar serves
     // every one of these happily, which is the inherited-permissiveness false-green
     // ADDING_A_TWIN.md §0 names. Each must be 422 — DeepSeek's "Invalid Parameters" — not 400.
     for (const [key, value] of [
-      ['n', 2], ['seed', 42], ['logit_bias', { '1': 1 }], ['top_k', 5], ['user', 'u1'],
+      ['n', 2], ['seed', 42], ['logit_bias', { '1': 1 }], ['top_k', 5],
       ['max_completion_tokens', 32], ['service_tier', 'flex'], ['parallel_tool_calls', false],
       ['functions', [{ name: 'f' }]], ['store', true], ['metadata', { a: 'b' }],
     ] as Array<[string, unknown]>) {
@@ -297,6 +297,19 @@ export const DEEPSEEK_CAPABILITIES: CapabilitySpec[] = [
     return ok(good) && bad.status === 422 && long.status === 422 && ok(atCap);
   })),
 
+  done('deepseek.chat.openai_user_ignored', 'chat', "OpenAI's `user` (which OpenAI-compatible software such as LibreChat sends) is accepted and has no effect: it is not DeepSeek's `user_id`, and a non-string is 422", 'api', 'common', () => withRoot(async (h) => {
+    const plain = await h({ m: 'POST', p: CHAT_PATH, b: CHAT() });
+    // a value user_id's charset refuses: accepted here, so it is not being read as user_id
+    const withUser = await h({ m: 'POST', p: CHAT_PATH, b: CHAT({ user: 'sam@friends.test' }) });
+    const both = await h({ m: 'POST', p: CHAT_PATH, b: CHAT({ user: 'u1', user_id: 'bad id' }) });
+    const typed = await h({ m: 'POST', p: CHAT_PATH, b: CHAT({ user: 42 }) });
+    type Chat = { choices?: Array<{ message?: { content?: string } }>; usage?: { prompt_tokens?: number } };
+    const answer = (r: { body: unknown }) => JSON.stringify([(r.body as Chat).choices?.[0]?.message?.content, (r.body as Chat).usage?.prompt_tokens]);
+    return ok(plain) && ok(withUser) && answer(withUser) === answer(plain)
+      && both.status === 422 && msg(both).includes("'user_id'")
+      && typed.status === 422 && msg(typed).includes("'user'");
+  })),
+
   done('deepseek.chat.message_name', 'chat', 'Participant `name` is accepted on system/user/assistant turns and counted into prompt tokens', 'api', 'niche', () => withRoot(async (h) => {
     const named = await h({ m: 'POST', p: CHAT_PATH, b: { model: MODEL, messages: [{ role: 'user', content: 'hi', name: 'customer_alpha' }] } });
     const plain = await h({ m: 'POST', p: CHAT_PATH, b: { model: MODEL, messages: [{ role: 'user', content: 'hi' }] } });

package/src/deepseek-conformance.ts

@@ -251,7 +251,8 @@ const REJECTIONS: Array<{ name: string; method: string; path: string; body: unkn
   { name: 'seed', method: 'POST', path: '/chat/completions', body: { ...CHAT, seed: 42 }, status: 422, why: "'seed' is not a DeepSeek parameter" },
   { name: 'logit_bias', method: 'POST', path: '/chat/completions', body: { ...CHAT, logit_bias: { '1': 1 } }, status: 422, why: "'logit_bias' is not a DeepSeek parameter" },
   { name: 'top_k', method: 'POST', path: '/chat/completions', body: { ...CHAT, top_k: 5 }, status: 422, why: "'top_k' is not a DeepSeek parameter" },
-  { name: 'user', method: 'POST', path: '/chat/completions', body: { ...CHAT, user: 'u1' }, status: 422, why: "DeepSeek's end-user identifier is 'user_id', not OpenAI's 'user'" },
+  // OpenAI's `user` as a string is accepted and ignored (deepseek-twin.ts, OPENAI_USER); only a non-string is refused
+  { name: 'user (not a string)', method: 'POST', path: '/chat/completions', body: { ...CHAT, user: 42 }, status: 422, why: "OpenAI's 'user' is taken and ignored, but only as a string" },
   { name: 'max_completion_tokens', method: 'POST', path: '/chat/completions', body: { ...CHAT, max_completion_tokens: 10 }, status: 422, why: "DeepSeek declares 'max_tokens' only" },
   { name: 'service_tier', method: 'POST', path: '/chat/completions', body: { ...CHAT, service_tier: 'flex' }, status: 422, why: 'DeepSeek has no service tiers' },
   { name: 'parallel_tool_calls', method: 'POST', path: '/chat/completions', body: { ...CHAT, parallel_tool_calls: false }, status: 422, why: 'not in DeepSeek\'s parameter table' },

package/src/deepseek-twin.ts

@@ -25,15 +25,17 @@
 //     includes two OpenAI does not use: 402 Insufficient Balance and 422 Invalid Parameters
 //     (api-docs.deepseek.com/quick_start/error_codes). OpenAI 400s a bad parameter; DeepSeek 422s
 //     it. 400 is reserved for "Invalid request body format".
-//   • `n`, `seed`, `logit_bias`, `top_k`, `user`, `max_completion_tokens`, `service_tier`,
+//   • `n`, `seed`, `logit_bias`, `top_k`, `max_completion_tokens`, `service_tier`,
 //     `parallel_tool_calls`, `functions` are NOT DeepSeek parameters. The accepted set is a closed
 //     documented table (api-docs.deepseek.com/api/create-chat-completion), which §6 licenses as a
-//     literal-allowlist oracle — so anything outside CHAT_PARAMS is 422 by name.
+//     literal-allowlist oracle — so anything outside CHAT_PARAMS is 422 by name. OpenAI's `user` is
+//     the one exception, accepted and ignored (OPENAI_USER below).
 //   • `response_format` accepts `text` and `json_object` ONLY. `json_schema` is an OpenAI feature;
 //     `@ai-sdk/deepseek` confirms it by INJECTING the schema into a system message instead of
 //     sending `response_format.json_schema` (src/chat/convert-to-deepseek-chat-messages.ts).
-//   • `user_id`, not `user` — with a regex and a 512-character cap the SDK enforces client-side and
-//     the vendor documents server-side.
+//   • `user_id` is DeepSeek's end-user identifier — with a regex and a 512-character cap the SDK
+//     enforces client-side and the vendor documents server-side. OpenAI's `user` is taken and has
+//     no effect (OPENAI_USER).
 //   • Assistant-prefix completion and strict tool calls exist ONLY under the `/beta` base URL.
 //   • `frequency_penalty` / `presence_penalty` are DEPRECATED but MUST NOT ERROR, and
 //     `temperature` / `top_p` are IGNORED (not rejected) while thinking is enabled: "setting these
@@ -283,13 +285,32 @@ type ToolChoice = 'auto' | 'none' | 'required' | { name: string };
 type ResponseFormat = 'text' | 'json_object';
 type ReasoningEffort = 'low' | 'high' | 'max';
 
+/**
+ * OpenAI's `user`, taken and IGNORED — where DeepSeek's documentation stops and the twin decides
+ * (read 2026-09-28). The reference table (api-docs.deepseek.com/api/create-chat-completion) does not
+ * list `user` (its end-user field is `user_id`) and says nothing of a key it does not list. Against
+ * that silence: DeepSeek states "The DeepSeek API uses an API format compatible with OpenAI/Anthropic"
+ * and invites "the OpenAI/Anthropic SDK or softwares compatible with the OpenAI/Anthropic API"
+ * (api-docs.deepseek.com, "Your First API Call"); its Anthropic-compatible surface answers the fields
+ * it does not use as "Ignored" rather than refusing them (api-docs.deepseek.com/guides/anthropic_api:
+ * metadata's "user_id is supported, others are ignored"); and OpenAI-compatible software sends `user`
+ * to DeepSeek as a matter of course — LibreChat's DeepSeek endpoint, through LangChain's ChatOpenAI,
+ * sends it on every chat (cookbook/librechat, finding: the twin's 422 was retried into a hang), which
+ * could not be DeepSeek's supported configuration if the real API refused it. So the twin accepts a
+ * string `user` and gives it no effect: it is NOT `user_id` (no isolation, no content-safety
+ * identity) and is not validated against `user_id`'s charset. A non-string `user` is 422, as OpenAI's
+ * declared type is a string. The rest of the table stays closed: this ruling is evidence for `user`
+ * alone, and none has been found for `n`, `seed` and the others.
+ */
+const OPENAI_USER = 'user';
+
 /**
  * THE CLOSED PARAMETER TABLE for `POST /chat/completions`
  * (api-docs.deepseek.com/api/create-chat-completion, read 2026-08-31) plus the two streaming keys
  * and the top-level `reasoning_effort` `@ai-sdk/deepseek` actually sends.
  *
  * §6 licenses a literal allowlist as an ORACLE precisely when the vendor documents a closed set,
- * and this is one — which is what gives `n`, `seed`, `logit_bias`, `top_k`, `user`,
+ * and this is one — which is what gives `n`, `seed`, `logit_bias`, `top_k`,
  * `max_completion_tokens`, `service_tier` and `parallel_tool_calls` a *checkable* refusal instead
  * of the silent acceptance an OpenAI-copied twin gives them.
  *
@@ -301,7 +322,7 @@ type ReasoningEffort = 'low' | 'high' | 'max';
 const CHAT_PARAMS = new Set([
   'messages', 'model', 'thinking', 'max_tokens', 'response_format', 'stop', 'stream',
   'stream_options', 'temperature', 'top_p', 'tools', 'tool_choice', 'logprobs', 'top_logprobs',
-  'user_id', 'frequency_penalty', 'presence_penalty', 'reasoning_effort',
+  'user_id', 'frequency_penalty', 'presence_penalty', 'reasoning_effort', OPENAI_USER,
 ]);
 
 /** The closed table for `POST /beta/completions` (api-docs.deepseek.com/api/create-completion). */
@@ -497,7 +518,11 @@ function validateChat(params: Record<string, unknown>, beta: boolean): { args: C
     topLogprobs = v;
   }
 
-  // (10) user_id — DeepSeek's end-user identifier. NOT OpenAI's `user` (which CHAT_PARAMS refuses).
+  // (10) user_id — DeepSeek's end-user identifier. NOT OpenAI's `user`, which is taken and ignored
+  //      (OPENAI_USER): only its type is checked.
+  if (params[OPENAI_USER] !== undefined && params[OPENAI_USER] !== null && typeof params[OPENAI_USER] !== 'string') {
+    return { error: invalidParameters(`'${OPENAI_USER}' must be a string`) };
+  }
   let userId: string | undefined;
   if (params.user_id !== undefined && params.user_id !== null) {
     const v = params.user_id;