fix(openai-chat): preserve empty thinking replay

src/adapters/openai-chat/messages.ts

@@ -224,7 +224,7 @@ export function messagesToChatFormat(parsed: OcxParsedRequest, provider: OcxProv
         let reasoningContent = thinkingParts.map(p => p.thinking).join("");
         if (
           reasoningContent.length === 0
-          && toolCalls.length > 0
+          && (toolCalls.length > 0 || thinkingParts.length > 0)
           && modelInList(provider.preserveReasoningContentModels, parsed.modelId)
         ) {
           const cached = toolCalls
@@ -235,11 +235,11 @@ export function messagesToChatFormat(parsed: OcxParsedRequest, provider: OcxProv
           if (cached.length > 0) {
             reasoningContent = [...new Set(cached)].join("\n");
           } else if (modelInList(provider.requiresReasoningPlaceholderModels ?? provider.preserveReasoningContentModels, parsed.modelId)) {
-            // Fallback (extends #950, closes #1193): the replay cache is
+            // Fallback (extends #950 and #1193; fixes #5421): the replay cache is
             // bounded (64 entries / 256 KiB / 1 h TTL) and always misses on
-            // long sessions, and some tool rounds carry no recorded reasoning
-            // at all. DeepSeek thinking mode rejects ANY tool_call assistant
-            // message missing reasoning_content with HTTP 400, so inject a
+            // long sessions, and some thinking/tool rounds carry no recorded
+            // reasoning at all. DeepSeek thinking mode rejects replay without
+            // reasoning_content with HTTP 400, so inject a
             // minimal placeholder rather than emit a bare continuation the
             // upstream will reject. Scoped to requiresReasoningPlaceholderModels
             // (defaulting to the preserve list): preserve-listed providers with

structure/providers/chat-compat.md

@@ -365,6 +365,10 @@ the desktop thinking band shows the "Thinking…" placeholder, and raw text appe
 which only fits native OpenAI providers that author real summaries. Diagnosis and codex-rs
 grouping evidence: `devlog/_fin/260709_native_response_pattern/`.
 
+For models that require a reasoning placeholder, a preserved thinking-only assistant turn with no
+plaintext receives that placeholder even when it has no tool call. Otherwise the Chat serializer
+drops the turn and strict DeepSeek continuations can reject the following request (#5421).
+
 The process-local raw-reasoning fallback is fail-closed unless a request has an explicit client
 thread plus an exact provider destination, wire adapter, final model, and physical credential
 identity. API-key material is represented only by a process-keyed HMAC; OAuth replay is bound to the

tests/providers/deepseek-reasoning-replay-gaps.test.ts

@@ -2,6 +2,7 @@ import { afterEach, beforeEach, describe, expect, test } from "bun:test";
 import { createOpenAIChatAdapter } from "../../src/adapters/openai-chat";
 import { buildResponseJSON } from "../../src/bridge";
 import { parseRequest } from "../../src/responses/parser";
+import { encodeReasoningEnvelope } from "../../src/responses/reasoning-envelope";
 import {
   clearReasoningReplayCacheForTests,
   peekReasoningForCall as peekReasoningForCallRaw,
@@ -129,6 +130,29 @@ describe("issue #950 — tool-call reasoning replay invariant (openai-chat wire)
     expect(assistant!["reasoning_content"]).toBe(REASONING);
   });
 
+  test("GAP F (issue #5421): signed thinking-only turn without tools gets a placeholder", () => {
+    const { messages } = wireFor([
+      {
+        type: "reasoning",
+        id: "rs_empty",
+        summary: [],
+        encrypted_content: encodeReasoningEnvelope({ sig: "opaque-signature" }),
+      },
+      {
+        type: "agent_message",
+        author: "parent",
+        recipient: "child",
+        content: [{ type: "input_text", text: "return OK" }],
+      },
+    ]);
+
+    const assistantIndex = messages.findIndex(message => message.role === "assistant");
+    const taskIndex = messages.findIndex(message => message.role === "user" && message.content === "return OK");
+    expect(assistantIndex).toBeGreaterThanOrEqual(0);
+    expect(assistantIndex).toBeLessThan(taskIndex);
+    expect(messages[assistantIndex]!["reasoning_content"]).toBe(" ");
+  });
+
   test("GAP A: reasoning item arriving AFTER its function_call is attached to its turn", () => {
     // Reconstructed histories (resume/retry/synthetic) may order the reasoning
     // item after the call it belongs to. The parser used to clear the pending
@@ -287,6 +311,23 @@ describe("issue #950 — tool-call reasoning replay invariant (openai-chat wire)
     const miss = toolCallAssistant(missResult.wire.messages);
     expect(miss).toBeDefined();
     expect(miss!["reasoning_content"]).toBeUndefined();
+    // Signed thinking without plaintext also stays omitted for an explicit
+    // placeholder opt-out, even though the parser preserves the thinking turn.
+    const thinkingOnly = minimaxWire([
+      {
+        type: "reasoning",
+        id: "rs_minimax_empty",
+        summary: [],
+        encrypted_content: encodeReasoningEnvelope({ sig: "opaque-signature" }),
+      },
+      {
+        type: "agent_message",
+        author: "parent",
+        recipient: "child",
+        content: [{ type: "input_text", text: "return OK" }],
+      },
+    ]).wire.messages;
+    expect(thinkingOnly.some(message => message.role === "assistant")).toBeFalse();
     // Cache hit on the same path: the recorded reasoning still replays.
     rememberReasoningForCall("call_1", REASONING, missResult.replayScope);
     const hit = toolCallAssistant(minimaxWire([userMessage(), functionCallOutputItem()]).wire.messages);