Merge PR #8968: fix(ai): hoist assistant message interleaved in responses tool batch (@roboomp)
This commit is contained in:
can1357
@@ -7,6 +7,7 @@
|
||||
- Fixed `/compact` (and automatic compaction) resurrecting pre-`/clear` conversation turns: `prepareCompaction` now honors the latest `reset_boundary`, so a compaction after an in-place `/clear` only summarizes messages created after the reset ([#8718](https://github.com/can1357/oh-my-pi/issues/8718)).
|
||||
- Hardened compaction summarization against prompt injection: conversation history and previous summaries are now treated as untrusted, and embedded `<conversation>`/`<previous-summary>` boundary tags are neutralized before prompt assembly ([#8727](https://github.com/can1357/oh-my-pi/pull/8727) by [@koopmannleon19977-cmyk](https://github.com/koopmannleon19977-cmyk)).
|
||||
- Compaction summarization input is now bounded to the summary model's context (windowed fold for oversized spans) and deterministic context-overflow 400s are no longer retried up to the full retry budget; artifact ids containing `503` no longer misclassify hard 400s as transient.
|
||||
- Fixed remote compaction mirroring the #8789 Responses shape: `buildOpenAiNativeHistory` now hoists an assistant `message` wedged between a tool-call batch and its outputs ahead of the batch, so compaction requests to strict opencode-go gateways match the canonical `message(s) → calls → outputs` order ([#8789](https://github.com/can1357/oh-my-pi/issues/8789)).
|
||||
|
||||
## [17.3.5] - 2026-08-16
|
||||
|
||||
|
||||
@@ -22,7 +22,11 @@ import {
|
||||
createOpenAICodexCompatibilityMetadata,
|
||||
getCodexAttestationHeader,
|
||||
} from "@oh-my-pi/pi-ai/providers/openai-codex-responses";
|
||||
import { parseAzureDeploymentNameMap, parseTextSignature } from "@oh-my-pi/pi-ai/providers/openai-shared";
|
||||
import {
|
||||
hoistInterleavedResponsesToolBatchMessages,
|
||||
parseAzureDeploymentNameMap,
|
||||
parseTextSignature,
|
||||
} from "@oh-my-pi/pi-ai/providers/openai-shared";
|
||||
import { transformMessages } from "@oh-my-pi/pi-ai/providers/transform-messages";
|
||||
import type {
|
||||
Api,
|
||||
@@ -740,7 +744,7 @@ export function buildOpenAiNativeHistory(
|
||||
msgIndex++;
|
||||
}
|
||||
|
||||
return stripOpenAIResponsesOutputOnlyStatusesForReplay(input);
|
||||
return stripOpenAIResponsesOutputOnlyStatusesForReplay(hoistInterleavedResponsesToolBatchMessages(input));
|
||||
}
|
||||
|
||||
// ============================================================================
|
||||
|
||||
@@ -365,6 +365,45 @@ function toolResultFor(callId: string, custom = false): ToolResultMessage {
|
||||
};
|
||||
}
|
||||
|
||||
describe("buildOpenAiNativeHistory interleaved assistant message (#8789)", () => {
|
||||
test("hoists a trailing text block before its tool-call batch", () => {
|
||||
// deepseek-v4-flash on opencode-go streamed [thinking, 2 tool calls,
|
||||
// trailing "</thinking" text]; the compaction history builder must not
|
||||
// wedge the demoted text between the calls and their outputs.
|
||||
const model = makeOpenAiModel({
|
||||
id: "deepseek-v4-flash",
|
||||
provider: "opencode-go",
|
||||
baseUrl: "https://opencode.ai/zen/go/v1",
|
||||
});
|
||||
const assistant: AssistantMessage = {
|
||||
role: "assistant",
|
||||
content: [
|
||||
{ type: "thinking", thinking: "planning" },
|
||||
{ type: "toolCall", id: "call_a|fc_call_a", name: "read", arguments: { path: "a" } },
|
||||
{ type: "toolCall", id: "call_b|fc_call_b", name: "read", arguments: { path: "b" } },
|
||||
{ type: "text", text: "<think>\n</thinking\n</think>" },
|
||||
],
|
||||
timestamp: Date.now(),
|
||||
provider: "opencode-go",
|
||||
model: "deepseek-v4-flash",
|
||||
api: "openai-responses",
|
||||
usage: ZERO_USAGE,
|
||||
stopReason: "toolUse",
|
||||
};
|
||||
|
||||
const items = buildOpenAiNativeHistory([assistant, toolResultFor("call_a"), toolResultFor("call_b")], model);
|
||||
|
||||
expect(items.map(item => item.type)).toEqual([
|
||||
"message",
|
||||
"function_call",
|
||||
"function_call",
|
||||
"function_call_output",
|
||||
"function_call_output",
|
||||
]);
|
||||
expect(JSON.stringify(items[0]?.content)).toContain("</thinking");
|
||||
});
|
||||
});
|
||||
|
||||
describe("buildOpenAiNativeHistory call-id tracking", () => {
|
||||
test("registers function_call ids carried in providerPayload so later tool results are emitted", () => {
|
||||
const items = buildOpenAiNativeHistory(
|
||||
|
||||
@@ -19,6 +19,7 @@
|
||||
- Cloud Code Assist Gemini 3.6/3.7 Flash requests at `minimal` now send `thinkingLevel: LOW` on the aliased `-low` SKU instead of `MINIMAL`, which the API rejects with HTTP 400.
|
||||
- Answer Cursor `interaction_query` permission gates (hosted web search, Exa, unnamed field-9 WebFetch) so the Run RPC continues instead of sitting silent until the 300s idle watchdog.
|
||||
- Fixed provider tool calls arriving with flattened array argument paths (e.g. Gemini's `questions[0].id`) being stripped and rejected by argument validation; well-formed flattened paths are now rebuilt into the nested arrays the tool schema expects ([#8886](https://github.com/can1357/oh-my-pi/issues/8886)).
|
||||
- Fixed opencode-go (Console Go) rejecting Responses turns with `400 No tool output found for tool call …` (naming a random call of the batch on each retry) when a model streamed a trailing text/thinking block after its tool calls: `buildResponsesInput` emitted that block as an assistant `message` item wedged between the `function_call` batch and its `function_call_output` items. Such interleaved messages are now hoisted ahead of their call batch (canonical `message(s) → calls → outputs`), which the strict gateway validator accepts; content is unchanged ([#8789](https://github.com/can1357/oh-my-pi/issues/8789)).
|
||||
|
||||
## [17.3.7] - 2026-08-17
|
||||
|
||||
|
||||
@@ -1522,6 +1522,77 @@ export function repairOrphanResponsesToolCalls(input: ResponseInput): ResponseIn
|
||||
return repaired;
|
||||
}
|
||||
|
||||
type ResponsesBatchItemKind = "call" | "output" | "assistant-message" | "other";
|
||||
|
||||
/** Classify a Responses input item for tool-call/output batch normalization. */
|
||||
function classifyResponsesBatchItem(item: object): ResponsesBatchItemKind {
|
||||
const type = "type" in item ? item.type : undefined;
|
||||
if (responsesToolCallKind(type) !== undefined) return "call";
|
||||
if (responsesToolOutputKind(type) !== undefined) return "output";
|
||||
const role = "role" in item ? item.role : undefined;
|
||||
if (type === "message" && role === "assistant") return "assistant-message";
|
||||
return "other";
|
||||
}
|
||||
|
||||
/**
|
||||
* Relocate assistant `message` items wedged inside a tool-call → tool-output
|
||||
* batch to before the batch, yielding canonical `message(s) → calls → outputs`
|
||||
* order. Idempotent; returns the same array reference when nothing moves.
|
||||
*
|
||||
* OpenAI's Responses API pairs tool outputs by `call_id` and tolerates any item
|
||||
* order, but stricter gateways (notably opencode-go's "Console Go") reject a
|
||||
* shape where an assistant message interrupts a `function_call` →
|
||||
* `function_call_output` run, 400ing with `No tool output found for tool call …`
|
||||
* (naming a random call of the batch on each retry). This arises whenever a
|
||||
* model streams a trailing text / demoted-thinking block *after* its tool calls:
|
||||
* the block-encode path preserves stream order, emitting the message between the
|
||||
* calls and the outputs appended afterward. Moving the already-model-owned
|
||||
* message ahead of its call batch keeps content identical while satisfying the
|
||||
* strict validator. See #8789.
|
||||
*/
|
||||
export function hoistInterleavedResponsesToolBatchMessages<T extends object>(items: readonly T[]): T[] {
|
||||
const moved = new Set<number>();
|
||||
const insertBefore = new Map<number, number[]>();
|
||||
for (let index = 0; index < items.length; index++) {
|
||||
if (classifyResponsesBatchItem(items[index]) !== "output") continue;
|
||||
// Only anchor on the first output of a run.
|
||||
if (index > 0 && classifyResponsesBatchItem(items[index - 1]) === "output") continue;
|
||||
// Walk back over the batch body (calls interleaved with assistant messages).
|
||||
let start = index;
|
||||
let sawCall = false;
|
||||
const messageIndexes: number[] = [];
|
||||
while (start > 0) {
|
||||
const kind = classifyResponsesBatchItem(items[start - 1]);
|
||||
if (kind === "call") {
|
||||
sawCall = true;
|
||||
} else if (kind === "assistant-message") {
|
||||
messageIndexes.push(start - 1);
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
start -= 1;
|
||||
}
|
||||
// Nothing to hoist unless a message actually sits among the calls.
|
||||
if (!sawCall || messageIndexes.length === 0) continue;
|
||||
messageIndexes.reverse();
|
||||
const target = insertBefore.get(start) ?? [];
|
||||
for (const messageIndex of messageIndexes) {
|
||||
moved.add(messageIndex);
|
||||
target.push(messageIndex);
|
||||
}
|
||||
insertBefore.set(start, target);
|
||||
}
|
||||
if (moved.size === 0) return items.slice();
|
||||
const result: T[] = [];
|
||||
for (let index = 0; index < items.length; index++) {
|
||||
const pending = insertBefore.get(index);
|
||||
if (pending) for (const messageIndex of pending) result.push(items[messageIndex]);
|
||||
if (moved.has(index)) continue;
|
||||
result.push(items[index]);
|
||||
}
|
||||
return result;
|
||||
}
|
||||
|
||||
/**
|
||||
* Some Responses backends (notably GitHub Copilot) reject the OpenAI image
|
||||
* `detail: "original"` value with a 400. When the model does not advertise
|
||||
@@ -1923,7 +1994,8 @@ export function buildResponsesInput<TApi extends Api>(options: BuildResponsesInp
|
||||
msgIndex++;
|
||||
}
|
||||
|
||||
const withRepairedOutputs = options.repairOrphanOutputs ? repairOrphanResponsesToolOutputs(messages) : messages;
|
||||
const hoisted = hoistInterleavedResponsesToolBatchMessages(messages);
|
||||
const withRepairedOutputs = options.repairOrphanOutputs ? repairOrphanResponsesToolOutputs(hoisted) : hoisted;
|
||||
const withRepairedCalls = repairOrphanResponsesToolCalls(withRepairedOutputs);
|
||||
return stripUnpairedOpenAIResponsesComputerReasoningIdsForReplay(withRepairedCalls);
|
||||
}
|
||||
|
||||
@@ -0,0 +1,142 @@
|
||||
import { describe, expect, it } from "bun:test";
|
||||
import type { ResponseInput } from "@oh-my-pi/pi-ai/providers/openai-responses-wire";
|
||||
import { buildResponsesInput } from "@oh-my-pi/pi-ai/providers/openai-shared";
|
||||
import type { AssistantMessage, Context, ToolResultMessage } from "@oh-my-pi/pi-ai/types";
|
||||
import { buildModel } from "@oh-my-pi/pi-catalog/build";
|
||||
|
||||
// deepseek-v4-flash on opencode-go (Console Go) — the gateway from #8789 that
|
||||
// rejects an assistant message interleaved between a function_call batch and
|
||||
// its function_call_output items.
|
||||
const model = buildModel({
|
||||
id: "deepseek-v4-flash",
|
||||
name: "DeepSeek V4 Flash",
|
||||
api: "openai-responses",
|
||||
provider: "opencode-go",
|
||||
baseUrl: "https://opencode.ai/zen/go/v1",
|
||||
reasoning: true,
|
||||
input: ["text"],
|
||||
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
|
||||
contextWindow: 128_000,
|
||||
maxTokens: 16_000,
|
||||
});
|
||||
|
||||
const zeroUsage = {
|
||||
input: 0,
|
||||
output: 0,
|
||||
cacheRead: 0,
|
||||
cacheWrite: 0,
|
||||
totalTokens: 0,
|
||||
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
|
||||
};
|
||||
|
||||
function wireType(item: ResponseInput[number]): string {
|
||||
if ("type" in item && typeof item.type === "string") return item.type;
|
||||
if ("role" in item && typeof item.role === "string") return `message:${item.role}`;
|
||||
return "unknown";
|
||||
}
|
||||
|
||||
function toolResult(callId: string, text: string): ToolResultMessage {
|
||||
return {
|
||||
role: "toolResult",
|
||||
toolCallId: callId,
|
||||
toolName: "read",
|
||||
content: [{ type: "text", text }],
|
||||
isError: false,
|
||||
timestamp: 2,
|
||||
};
|
||||
}
|
||||
|
||||
describe("buildResponsesInput #8789 interleaved assistant message", () => {
|
||||
it("hoists a trailing text block before its tool-call batch", () => {
|
||||
// Model streamed [thinking, 3 tool calls, trailing "</thinking" text].
|
||||
const assistant: AssistantMessage = {
|
||||
role: "assistant",
|
||||
content: [
|
||||
{ type: "thinking", thinking: "let me call some tools" },
|
||||
{ type: "toolCall", id: "call_a", name: "read", arguments: { path: "a" } },
|
||||
{ type: "toolCall", id: "call_b", name: "read", arguments: { path: "b" } },
|
||||
{ type: "toolCall", id: "call_c", name: "read", arguments: { path: "c" } },
|
||||
{ type: "text", text: "<think>\n</thinking\n</think>" },
|
||||
],
|
||||
api: model.api,
|
||||
provider: model.provider,
|
||||
model: model.id,
|
||||
usage: zeroUsage,
|
||||
stopReason: "toolUse",
|
||||
timestamp: 1,
|
||||
};
|
||||
const context: Context = {
|
||||
messages: [
|
||||
assistant,
|
||||
toolResult("call_a", "out a"),
|
||||
toolResult("call_b", "out b"),
|
||||
toolResult("call_c", "out c"),
|
||||
{ role: "user", content: "continue", timestamp: 5 },
|
||||
],
|
||||
};
|
||||
|
||||
const items = buildResponsesInput({
|
||||
model,
|
||||
context,
|
||||
strictResponsesPairing: false,
|
||||
supportsImageDetailOriginal: false,
|
||||
});
|
||||
|
||||
expect(items.map(wireType)).toEqual([
|
||||
"message",
|
||||
"function_call",
|
||||
"function_call",
|
||||
"function_call",
|
||||
"function_call_output",
|
||||
"function_call_output",
|
||||
"function_call_output",
|
||||
"message:user",
|
||||
]);
|
||||
|
||||
// The demoted-thinking text is preserved verbatim as the hoisted message.
|
||||
const hoisted = items[0];
|
||||
expect(JSON.stringify(hoisted)).toContain("</thinking");
|
||||
|
||||
// Invariant: no assistant message sits between a call and an output.
|
||||
const types = items.map(wireType);
|
||||
const lastCall = types.lastIndexOf("function_call");
|
||||
const firstOutput = types.indexOf("function_call_output");
|
||||
expect(lastCall).toBeLessThan(firstOutput);
|
||||
expect(types.slice(lastCall + 1, firstOutput)).toEqual([]);
|
||||
});
|
||||
|
||||
it("leaves an already-canonical message → calls → outputs turn unchanged", () => {
|
||||
const assistant: AssistantMessage = {
|
||||
role: "assistant",
|
||||
content: [
|
||||
{ type: "text", text: "calling read on two files" },
|
||||
{ type: "toolCall", id: "call_a", name: "read", arguments: { path: "a" } },
|
||||
{ type: "toolCall", id: "call_b", name: "read", arguments: { path: "b" } },
|
||||
],
|
||||
api: model.api,
|
||||
provider: model.provider,
|
||||
model: model.id,
|
||||
usage: zeroUsage,
|
||||
stopReason: "toolUse",
|
||||
timestamp: 1,
|
||||
};
|
||||
const context: Context = {
|
||||
messages: [assistant, toolResult("call_a", "out a"), toolResult("call_b", "out b")],
|
||||
};
|
||||
|
||||
const items = buildResponsesInput({
|
||||
model,
|
||||
context,
|
||||
strictResponsesPairing: false,
|
||||
supportsImageDetailOriginal: false,
|
||||
});
|
||||
|
||||
expect(items.map(wireType)).toEqual([
|
||||
"message",
|
||||
"function_call",
|
||||
"function_call",
|
||||
"function_call_output",
|
||||
"function_call_output",
|
||||
]);
|
||||
});
|
||||
});
|
||||
Reference in New Issue
Block a user