fix(hooks): preserve non-text parts in fromHookLLMRequest (#26275)

This commit is contained in:
Sandy Tao
2026-05-04 10:45:52 -07:00
committed by GitHub
parent 790f2cf815
commit 165efa8a38
2 changed files with 329 additions and 17 deletions
+120 -17
View File
@@ -5,8 +5,10 @@
*/
import type {
Content,
GenerateContentResponse,
GenerateContentParameters,
Part,
ToolConfig,
FinishReason,
FunctionCallingConfig,
@@ -100,11 +102,10 @@ function hasTextProperty(value: unknown): value is { text: string } {
}
/**
* Type guard to check if content has role and parts properties
* Type guard to check if a value is a Content object (i.e. has role and parts
* properties). Narrows to Content so callers can access `parts` as Part[].
*/
function isContentWithParts(
content: unknown,
): content is { role: string; parts: unknown } {
function isContentWithParts(content: unknown): content is Content {
return (
typeof content === 'object' &&
content !== null &&
@@ -226,22 +227,124 @@ export class HookTranslatorGenAIv1 extends HookTranslator {
baseRequest?: GenerateContentParameters,
): GenerateContentParameters {
// Convert hook messages back to SDK Content format.
//
// When both hookRequest.messages and baseRequest.contents are present, we
// merge the hook's text edits back into the original contents in place,
// preserving non-text parts (functionCall, functionResponse, inlineData,
// thought, etc.) that toHookLLMRequest filtered out for the simplified
// hook API. Without this merge, a BeforeModel hook that modifies text
// would destroy tool call/response history and cause the model to loop
// (see https://github.com/google-gemini/gemini-cli/issues/25558).
//
// If the hook returned a partial request without messages (e.g. only
// overriding `model`), fall back to the base request's contents so the
// conversation is preserved.
const contents = hookRequest.messages
? hookRequest.messages.map((message) => ({
role: message.role === 'model' ? 'model' : message.role,
parts: [
{
text:
typeof message.content === 'string'
? message.content
: String(message.content),
},
],
}))
: (baseRequest?.contents ?? []);
let contents: GenerateContentParameters['contents'];
if (!hookRequest.messages) {
contents = baseRequest?.contents ?? [];
} else if (baseRequest?.contents) {
// Merge hook messages back into base contents, preserving non-text parts.
const baseContents = Array.isArray(baseRequest.contents)
? baseRequest.contents
: [baseRequest.contents];
// The merged result is uniformly Content[] — ContentListUnion does not
// allow mixing strings (PartUnion) and Content objects in the same
// array, so any string entries from baseContents are normalized to
// Content here.
const merged: Content[] = [];
let messageIndex = 0;
const messageToContent = (
message: LLMRequest['messages'][number],
): Content => ({
role: message.role === 'model' ? 'model' : message.role,
parts: [
{
text:
typeof message.content === 'string'
? message.content
: String(message.content),
},
],
});
for (const content of baseContents) {
// Normalize each baseContents entry into a Content object so the
// merged array is homogeneous.
if (typeof content === 'string') {
// String entries always contributed one message to the hook view.
if (messageIndex < hookRequest.messages.length) {
merged.push(messageToContent(hookRequest.messages[messageIndex++]));
} else {
merged.push({ role: 'user', parts: [{ text: content }] });
}
continue;
}
if (!isContentWithParts(content)) {
// Bare Part object (PartUnion expansion: Content | Part | string).
// toHookLLMRequest does not emit a message for these, so preserve
// them as a single-part Content with a default role.
merged.push({ role: 'user', parts: [content] });
continue;
}
const parts: Part[] = content.parts ?? [];
const hasText = parts.some(hasTextProperty);
const baseContent: Content = { ...content, parts };
if (!hasText) {
// toHookLLMRequest skipped this entry — preserve it untouched so
// tool-call/response history is not lost.
merged.push(baseContent);
continue;
}
// This entry contributed a message — merge the hook's text back in
// and keep any non-text parts in their original order.
if (messageIndex < hookRequest.messages.length) {
const message = hookRequest.messages[messageIndex++];
const newText =
typeof message.content === 'string'
? message.content
: String(message.content);
const nonTextParts = parts.filter(
(p): p is Part => !hasTextProperty(p),
);
merged.push({
...baseContent,
role: message.role === 'model' ? 'model' : message.role,
parts: [{ text: newText }, ...nonTextParts],
});
} else {
merged.push(baseContent);
}
}
// Append any remaining hook messages beyond baseContents (the hook may
// have added new turns).
while (messageIndex < hookRequest.messages.length) {
merged.push(messageToContent(hookRequest.messages[messageIndex++]));
}
contents = merged;
} else {
// No baseRequest contents to merge against — fall back to text-only.
contents = hookRequest.messages.map((message) => ({
role: message.role === 'model' ? 'model' : message.role,
parts: [
{
text:
typeof message.content === 'string'
? message.content
: String(message.content),
},
],
}));
}
// Build the result with proper typing.
// Use nullish coalescing so a hook that only sets `model` still works --