mirror of
https://github.com/Nezumi-2711/9router.git
synced 2026-09-22 13:38:31 +00:00
fix(antigravity): passthrough tab-autocomplete + mark default agent slot mandatory
MODEL_NO_MAP guard never re-routes Antigravity tab-autocomplete (tab_* models) so latency-critical inline completion stays native. Flags gemini-3.5-flash-low (agent/Default) as mandatory in the dashboard. Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
committed by
decolua
co-authored by
Cursor
parent
61d546627e
commit
38b73bfc6b
+11
-1
@@ -56,6 +56,16 @@ const MODEL_PATTERNS = {
|
||||
],
|
||||
};
|
||||
|
||||
// Models that must NEVER be re-routed — always passthrough to the real upstream, even when
|
||||
// the tool's other models are mapped. Antigravity's tab-autocomplete (`tab_jump_flash_lite_preview`,
|
||||
// `tab_flash_lite_preview`, requestType tab/tab_jump) is latency-critical inline completion; routing
|
||||
// it through 9Router to an external chat model makes typing laggy and burns provider quota per
|
||||
// keystroke. Without this guard the broad `flash` pattern in MODEL_PATTERNS hijacks them onto the
|
||||
// flash-agent slot. Verified via MITM dump capture of streamGenerateContent (see AI_JOURNAL).
|
||||
const MODEL_NO_MAP = {
|
||||
antigravity: [/^tab[_-]/i],
|
||||
};
|
||||
|
||||
// URL substrings whose request/response should NOT be dumped to file (telemetry, polling, empty)
|
||||
const LOG_BLACKLIST_URL_PARTS = [
|
||||
"recordCodeAssistMetrics",
|
||||
@@ -74,4 +84,4 @@ function getToolForHost(host) {
|
||||
return null;
|
||||
}
|
||||
|
||||
module.exports = { IS_DEV, LSOF_BIN, TARGET_HOSTS, URL_PATTERNS, MODEL_SYNONYMS, MODEL_PATTERNS, LOG_BLACKLIST_URL_PARTS, getToolForHost };
|
||||
module.exports = { IS_DEV, LSOF_BIN, TARGET_HOSTS, URL_PATTERNS, MODEL_SYNONYMS, MODEL_PATTERNS, MODEL_NO_MAP, LOG_BLACKLIST_URL_PARTS, getToolForHost };
|
||||
|
||||
+9
-1
@@ -7,7 +7,7 @@ const dns = require("dns");
|
||||
const { promisify } = require("util");
|
||||
const { execSync } = require("child_process");
|
||||
const { log, err, dumpRequest, createResponseDumper, clearDumpDir } = require("./logger");
|
||||
const { IS_DEV, LSOF_BIN, TARGET_HOSTS, URL_PATTERNS, MODEL_SYNONYMS, MODEL_PATTERNS, getToolForHost } = require("./config");
|
||||
const { IS_DEV, LSOF_BIN, TARGET_HOSTS, URL_PATTERNS, MODEL_SYNONYMS, MODEL_PATTERNS, MODEL_NO_MAP, getToolForHost } = require("./config");
|
||||
const { DATA_DIR, MITM_DIR } = require("./paths");
|
||||
const { getCertForDomain } = require("./cert/generate");
|
||||
const { getMitmAlias } = require("./dbReader");
|
||||
@@ -330,6 +330,14 @@ const server = https.createServer(sslOptions, async (req, res) => {
|
||||
}
|
||||
|
||||
const model = extractModel(req.url, bodyBuffer);
|
||||
|
||||
// Intentional passthrough: some models must never be re-routed (e.g. Antigravity
|
||||
// tab-autocomplete) so latency-critical inline completion stays native. Silent — this
|
||||
// is by design, not a leak, and fires per keystroke. See MODEL_NO_MAP in config.js.
|
||||
if (model && (MODEL_NO_MAP[tool] || []).some((re) => re.test(model))) {
|
||||
return passthrough(req, res, bodyBuffer);
|
||||
}
|
||||
|
||||
const mappedModel = getMappedModel(tool, model);
|
||||
if (!mappedModel) {
|
||||
return passthrough(req, res, bodyBuffer);
|
||||
|
||||
@@ -10,7 +10,12 @@ export const MITM_TOOLS = {
|
||||
mitmDomain: "daily-cloudcode-pa.googleapis.com",
|
||||
modelAliases: ["gemini-3-flash-agent", "gemini-3.5-flash-low", "gemini-3.5-flash-extra-low", "gemini-3.1-pro-low", "gemini-pro-agent", "claude-sonnet-4-6", "claude-opus-4-6-thinking", "gpt-oss-120b-medium", "gemini-3-flash"],
|
||||
defaultModels: [
|
||||
{ id: "gemini-3.5-flash-low", name: "Gemini 3.5 Flash (Medium) / Default", alias: "gemini-3.5-flash-low" },
|
||||
// `mandatory: true` on the out-of-box agent/Default model — verified via MITM dump capture:
|
||||
// Antigravity's agent loop sends `gemini-3.5-flash-low` (requestType agent/checkpoint) by
|
||||
// default. The other slots only appear when the user explicitly picks that model, so they
|
||||
// stay optional. (Tab-autocomplete uses `tab_*` models that are never re-routed — see
|
||||
// MODEL_NO_MAP in src/mitm/config.js.)
|
||||
{ id: "gemini-3.5-flash-low", name: "Gemini 3.5 Flash (Medium) / Default", alias: "gemini-3.5-flash-low", mandatory: true },
|
||||
{ id: "gemini-3-flash-agent", name: "Gemini 3.5 Flash (High)", alias: "gemini-3-flash-agent" },
|
||||
{ id: "gemini-3.5-flash-extra-low", name: "Gemini 3.5 Flash (Low)", alias: "gemini-3.5-flash-extra-low" },
|
||||
{ id: "gemini-3.1-pro-low", name: "Gemini 3.1 Pro (Low)", alias: "gemini-3.1-pro-low" },
|
||||
|
||||
@@ -0,0 +1,40 @@
|
||||
import { describe, expect, it } from "vitest";
|
||||
import { createRequire } from "module";
|
||||
import { MITM_TOOLS } from "../../src/shared/constants/cliTools.js";
|
||||
|
||||
// config.js is the CJS MITM bundle module (dependency-isolated for the runtime copy).
|
||||
const require = createRequire(import.meta.url);
|
||||
const { MODEL_NO_MAP } = require("../../src/mitm/config.js");
|
||||
|
||||
// All assertions below are grounded in a live MITM dump capture of Antigravity's
|
||||
// streamGenerateContent requests (see AI_JOURNAL): the agent loop sends
|
||||
// `gemini-3.5-flash-low`, tab-autocomplete sends `tab_jump_flash_lite_preview` /
|
||||
// `tab_flash_lite_preview`.
|
||||
describe("Antigravity MITM model handling", () => {
|
||||
const ag = MITM_TOOLS.antigravity;
|
||||
|
||||
it("flags the out-of-box agent/Default model mandatory", () => {
|
||||
expect(ag.defaultModels.find((m) => m.id === "gemini-3.5-flash-low")?.mandatory).toBe(true);
|
||||
});
|
||||
|
||||
it("leaves models not proven auto-sent optional", () => {
|
||||
for (const id of ["gemini-3-flash-agent", "gemini-3.1-pro-low", "claude-sonnet-4-6", "gpt-oss-120b-medium"]) {
|
||||
expect(ag.defaultModels.find((m) => m.id === id)?.mandatory).toBeFalsy();
|
||||
}
|
||||
});
|
||||
|
||||
// Tab-autocomplete is latency-critical inline completion — it must passthrough natively,
|
||||
// never get re-routed onto a chat-model mapping by the broad `flash` pattern.
|
||||
it.each(["tab_jump_flash_lite_preview", "tab_flash_lite_preview"])(
|
||||
"excludes tab-autocomplete model '%s' from re-routing",
|
||||
(id) => {
|
||||
expect((MODEL_NO_MAP.antigravity || []).some((re) => re.test(id))).toBe(true);
|
||||
}
|
||||
);
|
||||
|
||||
it("does not exclude real agent models from re-routing", () => {
|
||||
for (const id of ["gemini-3.5-flash-low", "gemini-3-flash-agent", "claude-sonnet-4-6"]) {
|
||||
expect((MODEL_NO_MAP.antigravity || []).some((re) => re.test(id))).toBe(false);
|
||||
}
|
||||
});
|
||||
});
|
||||
Reference in New Issue
Block a user