fix(response): strip internal reasoning placeholder from all reasoning fields (#9853)

copyOpenAICompatibleReasoningFields only stripped the sentinel
(NON_ANTHROPIC_THINKING_PLACEHOLDER = "(prior reasoning summary
unavailable)") from reasoning_content and reasoning. Non-standard
reasoning fields (reasoning_text, thinking, thought) and
reasoning_details items passed through raw, leaking the internal
replay sentinel to clients on providers that use those fields
(e.g. Venice), where the model echo surfaces as a bogus thought block
and can degrade into empty turns.

Strip the sentinel from every forwarded reasoning field, including
per-item text/content inside reasoning_details; drop items/fields that
strip to nothing while preserving non-text details such as
reasoning.encrypted.

Fixes #9765
Refs #8081, #9606

Co-authored-by: safeer <asafeer1994@gmail.com>

open-sse/utils/reasoningFields.ts

@@ -62,10 +62,38 @@ export function hasAnyReasoningSignal(value: unknown): boolean {
   );
 }
 
+const STRIPPABLE_REASONING_FIELDS = [
+  "reasoning_content",
+  "reasoning",
+  "reasoning_text",
+  "thinking",
+  "thought",
+] as const;
+
+/**
+ * Strip the internal replay placeholder from a single string reasoning field,
+ * deleting the field when nothing meaningful remains. Returns true only when a
+ * present string field was fully stripped to empty (absent/non-string fields
+ * return false so callers can distinguish "removed" from "never had text").
+ */
+function stripPlaceholderFromField(target: JsonRecord, field: string): boolean {
+  const value = target[field];
+  if (typeof value !== "string") return false;
+  const stripped = stripInternalReasoningPlaceholder(value);
+  if (stripped === "") {
+    delete target[field];
+    return true;
+  }
+  if (stripped !== value) target[field] = stripped;
+  return false;
+}
+
 export function copyOpenAICompatibleReasoningFields(source: JsonRecord, target: JsonRecord) {
   if (source.reasoning_content !== undefined) target.reasoning_content = source.reasoning_content;
   if (source.reasoning !== undefined) target.reasoning = source.reasoning;
   if (source.reasoning_text !== undefined) target.reasoning_text = source.reasoning_text;
+  if (source.thinking !== undefined) target.thinking = source.thinking;
+  if (source.thought !== undefined) target.thought = source.thought;
   if (Array.isArray(source.reasoning_details)) target.reasoning_details = source.reasoning_details;
   if (!getReadableReasoningValue(target)) {
     const mirrored = getUnsupportedReasoningValue(source);
@@ -73,15 +101,31 @@ export function copyOpenAICompatibleReasoningFields(source: JsonRecord, target:
   }
   // ponytail: the internal replay placeholder is request scaffolding, never
   // real reasoning — models echo it and it poisons client history + the cache
-  // (#8081 echo). Strip it from anything we forward to the client.
-  if (typeof target.reasoning_content === "string") {
-    const stripped = stripInternalReasoningPlaceholder(target.reasoning_content);
-    if (stripped === "") delete target.reasoning_content;
-    else if (stripped !== target.reasoning_content) target.reasoning_content = stripped;
+  // (#8081 echo). Strip it from anything we forward to the client, including
+  // non-standard reasoning fields (reasoning_text / thinking / thought) and
+  // reasoning_details items that non-OpenAI-compatible upstreams (e.g.
+  // Venice) use (#9765 uncovered path).
+  for (const field of STRIPPABLE_REASONING_FIELDS) {
+    stripPlaceholderFromField(target, field);
   }
-  if (typeof target.reasoning === "string") {
-    const stripped = stripInternalReasoningPlaceholder(target.reasoning);
-    if (stripped === "") delete target.reasoning;
-    else if (stripped !== target.reasoning) target.reasoning = stripped;
+  if (Array.isArray(target.reasoning_details)) {
+    const cleaned: unknown[] = [];
+    for (const detail of target.reasoning_details) {
+      const record = asReasoningRecord(detail);
+      const next: JsonRecord = { ...record };
+      // Track whether the item originally carried text/content at all so
+      // non-text details (e.g. `reasoning.encrypted` carrying only `data`)
+      // survive untouched.
+      const hadText = typeof next.text === "string";
+      const hadContent = typeof next.content === "string";
+      stripPlaceholderFromField(next, "text");
+      stripPlaceholderFromField(next, "content");
+      const textGone = next.text === undefined;
+      const contentGone = next.content === undefined;
+      if ((hadText || hadContent) && textGone && contentGone) continue;
+      cleaned.push(next);
+    }
+    if (cleaned.length === 0) delete target.reasoning_details;
+    else target.reasoning_details = cleaned;
   }
 }

tests/unit/reasoning-fields-placeholder-strip.test.ts

@@ -0,0 +1,116 @@
+/**
+ * tests/unit/reasoning-fields-placeholder-strip.test.ts
+ *
+ * copyOpenAICompatibleReasoningFields() must never forward the internal
+ * reasoning-replay placeholder (NON_ANTHROPIC_THINKING_PLACEHOLDER =
+ * "(prior reasoning summary unavailable)") to clients — it is request
+ * scaffolding, and models echo it as their own reasoning (#8081, #9765).
+ * Previously only reasoning_content / reasoning were stripped; non-standard
+ * fields (reasoning_text, thinking, thought) and reasoning_details items
+ * passed through raw, leaking the sentinel on providers that use them
+ * (e.g. Venice).
+ */
+import test from "node:test";
+import assert from "node:assert/strict";
+import { NON_ANTHROPIC_THINKING_PLACEHOLDER } from "../../open-sse/utils/reasoningPlaceholder.ts";
+import { copyOpenAICompatibleReasoningFields } from "../../open-sse/utils/reasoningFields.ts";
+
+function copy(source: Record<string, unknown>): Record<string, unknown> {
+  const target: Record<string, unknown> = {};
+  copyOpenAICompatibleReasoningFields(source, target);
+  return target;
+}
+
+test("real reasoning_content is preserved verbatim", () => {
+  const target = copy({ reasoning_content: "Let me think carefully." });
+  assert.equal(target.reasoning_content, "Let me think carefully.");
+});
+
+test("reasoning_content that is exactly the placeholder is dropped", () => {
+  const target = copy({ reasoning_content: NON_ANTHROPIC_THINKING_PLACEHOLDER });
+  assert.equal("reasoning_content" in target, false);
+});
+
+test("reasoning alias that is exactly the placeholder is dropped", () => {
+  const target = copy({ reasoning: NON_ANTHROPIC_THINKING_PLACEHOLDER });
+  assert.equal("reasoning" in target, false);
+});
+
+test("reasoning_text that is exactly the placeholder is dropped (Venice path, #9765)", () => {
+  const target = copy({ reasoning_text: NON_ANTHROPIC_THINKING_PLACEHOLDER });
+  assert.equal("reasoning_text" in target, false);
+});
+
+test("thinking that is exactly the placeholder is dropped", () => {
+  const target = copy({ thinking: NON_ANTHROPIC_THINKING_PLACEHOLDER });
+  assert.equal("thinking" in target, false);
+});
+
+test("thought that is exactly the placeholder is dropped", () => {
+  const target = copy({ thought: NON_ANTHROPIC_THINKING_PLACEHOLDER });
+  assert.equal("thought" in target, false);
+});
+
+test("placeholder embedded in otherwise real reasoning_text is stripped in place", () => {
+  const target = copy({
+    reasoning_text: `First thought. ${NON_ANTHROPIC_THINKING_PLACEHOLDER} Second thought.`,
+  });
+  assert.equal(target.reasoning_text, "First thought.  Second thought.");
+});
+
+test("no mirrored reasoning_content is emitted when the only signal is the placeholder", () => {
+  const target = copy({ reasoning_text: NON_ANTHROPIC_THINKING_PLACEHOLDER });
+  assert.equal("reasoning_content" in target, false);
+  assert.equal("reasoning_text" in target, false);
+});
+
+test("all-placeholder reasoning_details are dropped entirely", () => {
+  const target = copy({
+    reasoning_details: [
+      { type: "reasoning.text", text: NON_ANTHROPIC_THINKING_PLACEHOLDER },
+      { type: "thinking", content: `  ${NON_ANTHROPIC_THINKING_PLACEHOLDER}  ` },
+    ],
+  });
+  assert.equal("reasoning_details" in target, false);
+  assert.equal("reasoning_content" in target, false);
+});
+
+test("mixed reasoning_details keep real text and drop only placeholder items", () => {
+  const target = copy({
+    reasoning_details: [
+      { type: "reasoning.text", text: "real first step " },
+      { type: "thinking", content: NON_ANTHROPIC_THINKING_PLACEHOLDER },
+      { type: "reasoning.text", text: "real second step" },
+    ],
+  });
+  assert.deepEqual(target.reasoning_details, [
+    { type: "reasoning.text", text: "real first step " },
+    { type: "reasoning.text", text: "real second step" },
+  ]);
+});
+
+test("placeholder inside a reasoning_details text item is stripped in place", () => {
+  const target = copy({
+    reasoning_details: [
+      { type: "reasoning.text", text: `real ${NON_ANTHROPIC_THINKING_PLACEHOLDER} tail` },
+    ],
+  });
+  assert.deepEqual(target.reasoning_details, [{ type: "reasoning.text", text: "real  tail" }]);
+});
+
+test("real reasoning_details still mirror into reasoning_content for readable clients", () => {
+  const target = copy({
+    reasoning_details: [{ type: "reasoning.text", text: "real reasoning here" }],
+  });
+  assert.equal(target.reasoning_content, "real reasoning here");
+  assert.deepEqual(target.reasoning_details, [
+    { type: "reasoning.text", text: "real reasoning here" },
+  ]);
+});
+
+test("non-text reasoning_details (e.g. reasoning.encrypted) survive untouched", () => {
+  const target = copy({
+    reasoning_details: [{ type: "reasoning.encrypted", data: "sig" }],
+  });
+  assert.deepEqual(target.reasoning_details, [{ type: "reasoning.encrypted", data: "sig" }]);
+});