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>
This commit is contained in:
Diego Rodrigues de Sa e Souza
2026-08-09 09:52:02 -03:00
committed by GitHub
parent 4fe0fffb31
commit 87c145a2be
2 changed files with 169 additions and 9 deletions

View File

@@ -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" }]);
});