Files
OmniRoute/open-sse/services/responsesInputSanitizer.ts

172 lines
6.0 KiB
TypeScript

type JsonRecord = Record<string, unknown>;
type SanitizeResponsesInputOptions = {
dropInternalAssistantMessages?: boolean;
};
const INTERNAL_ASSISTANT_PHASES = new Set(["commentary"]);
const SERVER_ITEM_ID_PREFIX_BY_TYPE: Record<string, string> = {
function_call: "fc_",
message: "msg_",
reasoning: "rs_",
};
const SERVER_ITEM_ID_PATTERN = /^(fc|msg|rs|resp)_/;
function toRecord(value: unknown): JsonRecord | null {
return value && typeof value === "object" && !Array.isArray(value) ? (value as JsonRecord) : null;
}
function isResponsesMessageItem(record: JsonRecord): boolean {
return record.type === "message" || (!record.type && typeof record.role === "string");
}
export function isInternalAssistantMessage(record: JsonRecord): boolean {
if (!isResponsesMessageItem(record)) return false;
if (record.role !== "assistant") return false;
const phase = typeof record.phase === "string" ? record.phase.trim().toLowerCase() : "";
if (!phase) return false;
// Drop only known internal runtime frames. Visible assistant turns such as
// `final` and `final_answer` must survive replay for Codex/OpenCode follow-ups.
return INTERNAL_ASSISTANT_PHASES.has(phase);
}
// OpenAI Responses API enforces two constraints on name fields in input items:
// 1. Max 128 characters
// 2. Must match ^[a-zA-Z0-9_-]+$
// Sanitize after cloning so upstream never sees an invalid name.
function sanitizeFunctionName(name: string): string {
// Replace any character not in [a-zA-Z0-9_-] with underscore, then truncate.
return name.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 128);
}
function sanitizeInputItemId(record: JsonRecord): JsonRecord {
if (typeof record.id !== "string") return record;
const type = typeof record.type === "string" ? record.type : "";
const expectedPrefix = SERVER_ITEM_ID_PREFIX_BY_TYPE[type];
const hasExpectedPrefix = expectedPrefix
? record.id.startsWith(expectedPrefix)
: SERVER_ITEM_ID_PATTERN.test(record.id);
if (hasExpectedPrefix) return record;
const next = { ...record };
delete next.id;
return next;
}
function imageUrlToText(value: unknown): string {
if (typeof value === "string") return value;
const record = toRecord(value);
return typeof record?.url === "string" ? record.url : "";
}
function sanitizeContentPart(part: unknown, role: string): unknown {
const record = toRecord(part);
if (!record) return part;
if (record.type === "image_url") {
const url = imageUrlToText(record.image_url);
// `output_text` is only a legal content-part type on assistant-role OUTPUT
// items. Every other role (user, system, developer, ...) is input-side and
// must use `input_image` -- otherwise the Codex/Responses backend rejects
// the replayed history with "Invalid value: 'output_text'" (#8089).
if (role !== "assistant") {
const next: JsonRecord = { type: "input_image", image_url: url };
const image = toRecord(record.image_url);
if (image?.detail !== undefined) next.detail = image.detail;
return next;
}
return { type: "output_text", text: url ? `[Image: ${url}]` : "[Image]" };
}
if (role === "assistant" && record.type === "input_image") {
const url = imageUrlToText(record.image_url);
return { type: "output_text", text: url ? `[Image: ${url}]` : "[Image]" };
}
return part;
}
function sanitizeMessageContent(record: JsonRecord): JsonRecord {
if (!Array.isArray(record.content)) return record;
const role = typeof record.role === "string" ? record.role.toLowerCase() : "";
const content = record.content.map((part) => sanitizeContentPart(part, role));
return { ...record, content };
}
function sanitizeNestedOutputPart(part: unknown): unknown {
const record = toRecord(part);
if (!record) return part;
// `output` on replayed items is an input-side container. Its content uses
// input content-part types even when the enclosing item originated from an
// assistant/tool response. Converting an image placeholder to output_text
// here makes Codex reject the request with the inverse 400.
if (record.type === "output_text" || record.type === "refusal") {
const next: JsonRecord = { ...record, type: "input_text" };
if (typeof next.text !== "string") {
next.text = typeof record.refusal === "string" ? record.refusal : "";
}
delete next.annotations;
delete next.logprobs;
delete next.obfuscation;
delete next.refusal;
return next;
}
return sanitizeContentPart(part, "user");
}
function sanitizeOutputContent(record: JsonRecord): JsonRecord {
if (!Array.isArray(record.output)) return record;
// Some clients replay previous Responses output items inside the next
// Responses input. In that shape OpenAI validates `input[n].output[m].type`
// against output content part types, so legacy Chat-style `image_url` parts
// must be normalized here too, not only in message.content.
const output = record.output.map(sanitizeNestedOutputPart);
return { ...record, output };
}
function sanitizeInputItem(item: unknown): unknown {
const record = toRecord(item);
if (!record) return item;
let next = sanitizeInputItemId(record);
if (isResponsesMessageItem(next)) {
next = sanitizeMessageContent(next);
}
next = sanitizeOutputContent(next);
if (
(next.type === "function_call" || next.type === "function_call_output") &&
typeof next.name === "string" &&
!/^[a-zA-Z0-9_-]{1,128}$/.test(next.name)
) {
next = { ...next, name: sanitizeFunctionName(next.name) };
}
return next;
}
export function sanitizeResponsesInputItems(
items: readonly unknown[],
clone = true,
options: SanitizeResponsesInputOptions = {}
): unknown[] {
const dropInternalAssistantMessages = options.dropInternalAssistantMessages ?? true;
const sanitized: unknown[] = [];
for (const item of items) {
const record = toRecord(item);
if (dropInternalAssistantMessages && record && isInternalAssistantMessage(record)) {
continue;
}
const cloned = clone ? structuredClone(item) : item;
sanitized.push(sanitizeInputItem(cloned));
}
return sanitized;
}