From 7f0404bf82b311d201bcb1e06ec70dbf1874f861 Mon Sep 17 00:00:00 2001 From: Diego Rodrigues de Sa e Souza Date: Tue, 18 Aug 2026 10:50:15 -0300 Subject: [PATCH] fix(open-sse): stop concurrent requests colliding on dedup hash for non-OpenAI formats (#10438) * fix(open-sse): stop concurrent requests colliding on dedup hash for non-OpenAI formats computeRequestHash() in requestDedup.ts projected the prompt content from body.messages only. The dedup site in chatCore.ts hashes the *translated* (target-format) request body, and non-OpenAI target formats don't carry a messages field: Gemini-translated bodies use `contents`, Responses-API bodies use `input`. So for those formats messages was always undefined, every prompt hashed to the same null-backed value for a given model, and concurrent requests with different prompts joined the same in-flight promise -- the second caller silently received the first caller's response verbatim (#10249). Fix: project body.messages ?? body.contents ?? body.input ?? null instead of only body.messages, keeping the rest of the canonical hash projection unchanged. Genuinely identical concurrent requests still dedupe (the intended perf behavior); different prompts under Gemini/Responses-API target formats no longer collide. Regression test: tests/unit/request-dedup-10249.test.ts reproduces the two collision scenarios from the plan-file (Gemini `contents`, Responses-API `input`), confirms the OpenAI `messages` case was already correct, and asserts identical-request dedup keeps working. Verified RED (byte-identical hashes 0b24fd88.../dc16d5b7... pre-fix) -> GREEN (distinct hashes, dedup preserved) against this exact diff. * fix(open-sse): cover nested translator shapes + system fields in dedup hash (#10438) computeRequestHash() only read top-level body.messages ?? body.contents ?? body.input, but several translated request shapes nest their prompt content: the Antigravity Cloud Code envelope under request.contents, and Kiro under conversationState.currentMessage.userInputMessage.content (plus conversationState.history). Two different concurrent prompts to those targets could hash identically and share/leak a response between callers. Adds extractPromptContent()/extractSystemContent() helpers covering every prompt-bearing shape produced by open-sse/translator/request/*.ts (OpenAI/Cursor messages, Claude messages+system, Gemini contents+ systemInstruction, Responses input+instructions, Antigravity and Kiro nesting), and folds system/instructions/systemInstruction into the canonical hash so two requests with the same user message but a different system prompt no longer collide either. --------- Co-authored-by: adevwithpurpose --- .../fixes/10249-dedup-hash-collision.md | 1 + open-sse/services/requestDedup.ts | 98 ++++++++- tests/unit/request-dedup-10249.test.ts | 207 ++++++++++++++++++ 3 files changed, 304 insertions(+), 2 deletions(-) create mode 100644 changelog.d/fixes/10249-dedup-hash-collision.md create mode 100644 tests/unit/request-dedup-10249.test.ts diff --git a/changelog.d/fixes/10249-dedup-hash-collision.md b/changelog.d/fixes/10249-dedup-hash-collision.md new file mode 100644 index 0000000000..f118196dfe --- /dev/null +++ b/changelog.d/fixes/10249-dedup-hash-collision.md @@ -0,0 +1 @@ +- fix(open-sse): stop concurrent requests colliding on the same dedup hash for non-OpenAI target formats (#10249) diff --git a/open-sse/services/requestDedup.ts b/open-sse/services/requestDedup.ts index 5ccde19529..1a39216197 100644 --- a/open-sse/services/requestDedup.ts +++ b/open-sse/services/requestDedup.ts @@ -32,16 +32,110 @@ export interface DedupResult { const inflight = new Map>(); +function asRecord(value: unknown): Record | null { + return value !== null && typeof value === "object" && !Array.isArray(value) + ? (value as Record) + : null; +} + +/** + * Extract the prompt-bearing content from a (possibly translated) request body. + * + * The prompt content lives under different keys depending on the target + * provider format the body has already been translated to: + * - OpenAI-style bodies (`open-sse/translator/request/*-to-openai.ts`, + * `openai-to-cursor.ts`): `messages` + * - Gemini-translated bodies (`openai-to-gemini.ts`, + * `claude-to-gemini.ts`): `contents` + * - Responses-API-translated bodies (`openai-responses/toResponses.ts`): + * `input` + * - Antigravity-translated bodies (`openai-to-gemini.ts` + * `openaiToAntigravityRequest` / `wrapInCloudCodeEnvelope`): nested under + * `request.contents` (a Cloud Code envelope wrapper) + * - Kiro-translated bodies (`openai-to-kiro.ts` `buildKiroPayload`): nested + * under `conversationState.currentMessage.userInputMessage.content` (the + * current turn) plus `conversationState.history` (prior turns) + * + * Falling back to only `messages` made every non-OpenAI-format body hash the + * prompt as `null`, colliding different prompts onto the same dedup hash + * (#10249). The Antigravity/Kiro nesting was still missed by the flat + * `messages ?? contents ?? input` fallback chain, so different prompts + * targeting those two providers still collided (#10438). + */ +function extractPromptContent(body: Record): unknown { + if (body.messages !== undefined) return body.messages; + if (body.contents !== undefined) return body.contents; + if (body.input !== undefined) return body.input; + + // Antigravity Cloud Code envelope: { request: { contents, ... } } + const request = asRecord(body.request); + if (request && request.contents !== undefined) { + return request.contents; + } + + // Kiro conversationState envelope: + // { conversationState: { currentMessage: { userInputMessage: { content } }, history } } + const conversationState = asRecord(body.conversationState); + if (conversationState) { + const currentMessage = asRecord(conversationState.currentMessage); + const userInputMessage = asRecord(currentMessage?.userInputMessage); + if (userInputMessage || conversationState.history !== undefined) { + return { + content: userInputMessage?.content ?? null, + history: conversationState.history ?? null, + }; + } + } + + return null; +} + +/** + * Extract the system/instruction content that shapes generation but is not + * carried in the message list itself. Two requests with the same user + * message but a different system prompt must hash differently — omitting + * this field let them collide. + * + * - Claude-translated bodies (`openai-to-claude.ts`): `system` + * - Responses-API-translated bodies (`openai-responses/toResponses.ts`): + * `instructions` + * - Gemini-translated bodies (`openai-to-gemini.ts`, `claude-to-gemini.ts`): + * `systemInstruction` + * - Antigravity-translated bodies: nested under `request.systemInstruction` + * (note: the client system prompt is folded into `request.contents[0]` + * instead per #9030, so this is usually the constant Antigravity + * default — it is still included for completeness/future-proofing) + */ +function extractSystemContent(body: Record): unknown { + if (body.system !== undefined) return body.system; + if (body.instructions !== undefined) return body.instructions; + if (body.systemInstruction !== undefined) return body.systemInstruction; + + const request = asRecord(body.request); + if (request && request.systemInstruction !== undefined) { + return request.systemInstruction; + } + + return null; +} + /** * Compute a deterministic hash for a request body. - * Includes: model, messages, temperature, tools, tool_choice, max_tokens, response_format + * Includes: model, messages/prompt content, system/instructions, temperature, + * tools, tool_choice, max_tokens, response_format * Excludes: stream, user, metadata (don't affect LLM output) + * + * `computeRequestHash` is called post-translation (`chatCore.ts`, on + * `translatedBody`), so the body shape here is whatever the target provider + * format produced — see `extractPromptContent`/`extractSystemContent` for the + * full list of shapes this must cover (#10249, #10438). */ export function computeRequestHash(requestBody: unknown): string { const body = requestBody as Record; const canonical = { model: body.model ?? null, - messages: body.messages ?? null, + messages: extractPromptContent(body), + system: extractSystemContent(body), temperature: typeof body.temperature === "number" ? body.temperature : 1.0, tools: body.tools ?? null, tool_choice: body.tool_choice ?? null, diff --git a/tests/unit/request-dedup-10249.test.ts b/tests/unit/request-dedup-10249.test.ts new file mode 100644 index 0000000000..3ce236dbd3 --- /dev/null +++ b/tests/unit/request-dedup-10249.test.ts @@ -0,0 +1,207 @@ +import { test } from "node:test"; +import assert from "node:assert/strict"; +import { computeRequestHash, deduplicate, clearInflight } from "../../open-sse/services/requestDedup.ts"; + +// Regression tests for #10249: the dedup hash used to read only `body.messages`, +// so translated (target-format) bodies that carry the prompt under a different +// key (`contents` for Gemini, `input` for the Responses API) always hashed the +// prompt as `null`. Concurrent requests with different prompts then collided on +// the same dedup hash, joined the same in-flight promise, and the second caller +// silently received the first caller's response. + +test("Gemini-format translated bodies with different prompts must NOT collide on dedup hash", async () => { + clearInflight(); + const bodyA = { + contents: [{ role: "user", parts: [{ text: "Summarize the Q3 financial report attached." }] }], + temperature: 0, + }; + const bodyB = { + contents: [{ role: "user", parts: [{ text: "Extract every invoice number from the attached PDF." }] }], + temperature: 0, + }; + const hashA = computeRequestHash({ ...bodyA, model: "gemini/gemini-2.5-flash", stream: false }); + const hashB = computeRequestHash({ ...bodyB, model: "gemini/gemini-2.5-flash", stream: false }); + assert.notEqual(hashA, hashB, "Different prompts must have different dedup hashes"); + + const [resA, resB] = await Promise.all([ + deduplicate(hashA, async () => "RESPONSE_A"), + deduplicate(hashB, async () => "RESPONSE_B"), + ]); + assert.equal(resA.result, "RESPONSE_A"); + assert.equal(resB.result, "RESPONSE_B"); + assert.equal(resB.wasDeduplicated, false); +}); + +test("Responses-API input-format translated bodies with different prompts must NOT collide", async () => { + clearInflight(); + const bodyA = { + input: [{ role: "user", content: [{ type: "input_text", text: "What is the capital of France?" }] }], + temperature: 0, + }; + const bodyB = { + input: [{ role: "user", content: [{ type: "input_text", text: "Explain quantum entanglement." }] }], + temperature: 0, + }; + const hashA = computeRequestHash({ ...bodyA, model: "openai/gpt-4.1", stream: false }); + const hashB = computeRequestHash({ ...bodyB, model: "openai/gpt-4.1", stream: false }); + assert.notEqual(hashA, hashB, "Different prompts must have different dedup hashes"); + + const [resA, resB] = await Promise.all([ + deduplicate(hashA, async () => "RESPONSE_A"), + deduplicate(hashB, async () => "RESPONSE_B"), + ]); + assert.equal(resA.result, "RESPONSE_A"); + assert.equal(resB.result, "RESPONSE_B"); + assert.equal(resB.wasDeduplicated, false); +}); + +test("Sanity: OpenAI-format bodies with different prompts DO get distinct hashes (unchanged behavior)", () => { + const bodyA = { messages: [{ role: "user", content: "Hello there" }], temperature: 0 }; + const bodyB = { messages: [{ role: "user", content: "Goodbye now" }], temperature: 0 }; + const hashA = computeRequestHash({ ...bodyA, model: "openai/gpt-4.1", stream: false }); + const hashB = computeRequestHash({ ...bodyB, model: "openai/gpt-4.1", stream: false }); + assert.notEqual(hashA, hashB); +}); + +// Regression tests for #10438: the flat `messages ?? contents ?? input` +// fallback chain from #10249 still missed the NESTED prompt shapes that +// `openai-to-gemini.ts::wrapInCloudCodeEnvelope` (Antigravity) and +// `openai-to-kiro.ts::buildKiroPayload` (Kiro) actually produce, and never +// looked at the system/instruction fields (`system` for Claude, `instructions` +// for the Responses API, `systemInstruction` for Gemini) at all — two +// requests with the same user message but a different system prompt hashed +// identically. + +test("Antigravity Cloud Code envelope bodies with different prompts must NOT collide on dedup hash", async () => { + clearInflight(); + const buildEnvelope = (text: string) => ({ + project: "proj-123", + requestId: "req-abc", + request: { + sessionId: "sess-1", + contents: [{ role: "user", parts: [{ text }] }], + systemInstruction: { role: "system", parts: [{ text: "You are Antigravity." }] }, + generationConfig: { maxOutputTokens: 8192 }, + }, + model: "gemini-3-pro", + userAgent: "antigravity/1.0", + requestType: "agent", + }); + const bodyA = buildEnvelope("Summarize the Q3 financial report attached."); + const bodyB = buildEnvelope("Extract every invoice number from the attached PDF."); + const hashA = computeRequestHash({ ...bodyA, model: "antigravity/gemini-3-pro", stream: false }); + const hashB = computeRequestHash({ ...bodyB, model: "antigravity/gemini-3-pro", stream: false }); + assert.notEqual(hashA, hashB, "Different prompts must have different dedup hashes"); + + const [resA, resB] = await Promise.all([ + deduplicate(hashA, async () => "RESPONSE_A"), + deduplicate(hashB, async () => "RESPONSE_B"), + ]); + assert.equal(resA.result, "RESPONSE_A"); + assert.equal(resB.result, "RESPONSE_B"); + assert.equal(resB.wasDeduplicated, false); +}); + +test("Kiro conversationState bodies with different prompts must NOT collide on dedup hash", async () => { + clearInflight(); + const buildPayload = (content: string) => ({ + conversationState: { + chatTriggerType: "MANUAL", + conversationId: "conv-1", + currentMessage: { + userInputMessage: { + content, + modelId: "kiro-claude-sonnet", + origin: "AI_EDITOR", + }, + }, + history: [], + }, + }); + const bodyA = buildPayload("[Context: Current time is 2026-08-17]\n\nWhat is the capital of France?"); + const bodyB = buildPayload("[Context: Current time is 2026-08-17]\n\nExplain quantum entanglement."); + const hashA = computeRequestHash({ ...bodyA, model: "kiro/claude-sonnet-4.5", stream: false }); + const hashB = computeRequestHash({ ...bodyB, model: "kiro/claude-sonnet-4.5", stream: false }); + assert.notEqual(hashA, hashB, "Different prompts must have different dedup hashes"); + + const [resA, resB] = await Promise.all([ + deduplicate(hashA, async () => "RESPONSE_A"), + deduplicate(hashB, async () => "RESPONSE_B"), + ]); + assert.equal(resA.result, "RESPONSE_A"); + assert.equal(resB.result, "RESPONSE_B"); + assert.equal(resB.wasDeduplicated, false); +}); + +test("Claude-translated bodies with the same messages but different `system` prompts must NOT collide", () => { + const bodyA = { + messages: [{ role: "user", content: [{ type: "text", text: "Hello" }] }], + system: [{ type: "text", text: "You are a pirate. Speak like one." }], + temperature: 0, + }; + const bodyB = { + messages: [{ role: "user", content: [{ type: "text", text: "Hello" }] }], + system: [{ type: "text", text: "You are a formal legal assistant." }], + temperature: 0, + }; + const hashA = computeRequestHash({ ...bodyA, model: "anthropic/claude-sonnet-4.5", stream: false }); + const hashB = computeRequestHash({ ...bodyB, model: "anthropic/claude-sonnet-4.5", stream: false }); + assert.notEqual(hashA, hashB, "Same messages with a different system prompt must hash differently"); +}); + +test("Responses-API-translated bodies with the same input but different `instructions` must NOT collide", () => { + const bodyA = { + input: [{ type: "message", role: "user", content: [{ type: "input_text", text: "Hello" }] }], + instructions: "You are a pirate. Speak like one.", + }; + const bodyB = { + input: [{ type: "message", role: "user", content: [{ type: "input_text", text: "Hello" }] }], + instructions: "You are a formal legal assistant.", + }; + const hashA = computeRequestHash({ ...bodyA, model: "openai/gpt-5", stream: false }); + const hashB = computeRequestHash({ ...bodyB, model: "openai/gpt-5", stream: false }); + assert.notEqual(hashA, hashB, "Same input with different instructions must hash differently"); +}); + +test("Gemini-translated bodies with the same contents but different `systemInstruction` must NOT collide", () => { + const bodyA = { + contents: [{ role: "user", parts: [{ text: "Hello" }] }], + systemInstruction: { role: "system", parts: [{ text: "You are a pirate. Speak like one." }] }, + temperature: 0, + }; + const bodyB = { + contents: [{ role: "user", parts: [{ text: "Hello" }] }], + systemInstruction: { role: "system", parts: [{ text: "You are a formal legal assistant." }] }, + temperature: 0, + }; + const hashA = computeRequestHash({ ...bodyA, model: "gemini/gemini-2.5-flash", stream: false }); + const hashB = computeRequestHash({ ...bodyB, model: "gemini/gemini-2.5-flash", stream: false }); + assert.notEqual(hashA, hashB, "Same contents with different systemInstruction must hash differently"); +}); + +test("Genuinely identical requests still hash identically and get deduplicated (perf feature preserved)", async () => { + clearInflight(); + const body = { + contents: [{ role: "user", parts: [{ text: "Same prompt text every time" }] }], + temperature: 0, + }; + const hash1 = computeRequestHash({ ...body, model: "gemini/gemini-2.5-flash", stream: false }); + const hash2 = computeRequestHash({ ...body, model: "gemini/gemini-2.5-flash", stream: false }); + assert.equal(hash1, hash2, "Identical bodies must still produce the same hash"); + + let callCount = 0; + const slowFn = async () => { + callCount += 1; + await new Promise((resolve) => setTimeout(resolve, 20)); + return "SHARED_RESPONSE"; + }; + + const [resA, resB] = await Promise.all([ + deduplicate(hash1, slowFn), + deduplicate(hash2, slowFn), + ]); + assert.equal(resA.result, "SHARED_RESPONSE"); + assert.equal(resB.result, "SHARED_RESPONSE"); + assert.equal(callCount, 1, "Identical concurrent requests must share a single upstream call"); + assert.equal(resA.wasDeduplicated === true || resB.wasDeduplicated === true, true); +});