185 lines
6.8 KiB
TypeScript
185 lines
6.8 KiB
TypeScript
import test from "node:test";
|
|
import assert from "node:assert/strict";
|
|
import fs from "node:fs";
|
|
import os from "node:os";
|
|
import path from "node:path";
|
|
|
|
// Regression for #3825: PR #3399 disabled the <omniModel>-tag history pinning and
|
|
// replaced it with a server-side session pin in combo.ts that is gated on
|
|
// `relayOptions?.sessionId`. Most OpenAI-compatible clients send no session id, so
|
|
// the read/write never fired → combos lost model stickiness → every turn re-ran the
|
|
// strategy (round-robin rotated) → upstream prompt-cache misses, cold high-reasoning
|
|
// starts, intermittent 504s. The fix derives a stable per-conversation key from the
|
|
// request body when no session id is present, so a combo re-pins to the same model
|
|
// across turns of the same conversation — but ONLY when context_cache_protection is on.
|
|
|
|
const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-combo-pin-3825-"));
|
|
const ORIGINAL_DATA_DIR = process.env.DATA_DIR;
|
|
process.env.DATA_DIR = TEST_DATA_DIR;
|
|
|
|
const { handleComboChat } = await import("../../open-sse/services/combo.ts");
|
|
const { clearSessions } = await import("../../open-sse/services/sessionManager.ts");
|
|
const core = await import("../../src/lib/db/core.ts");
|
|
|
|
function createLog() {
|
|
const entries: any[] = [];
|
|
return {
|
|
info: (tag: any, msg: any) => entries.push({ level: "info", tag, msg }),
|
|
warn: (tag: any, msg: any) => entries.push({ level: "warn", tag, msg }),
|
|
error: (tag: any, msg: any) => entries.push({ level: "error", tag, msg }),
|
|
debug: (tag: any, msg: any) => entries.push({ level: "debug", tag, msg }),
|
|
entries,
|
|
};
|
|
}
|
|
|
|
function okResponse(body: any = { choices: [{ message: { content: "ok" } }] }) {
|
|
return new Response(JSON.stringify(body), {
|
|
status: 200,
|
|
headers: { "content-type": "application/json" },
|
|
});
|
|
}
|
|
|
|
function waitForBackgroundWork() {
|
|
return new Promise((resolve) => setTimeout(resolve, 25));
|
|
}
|
|
|
|
// A stable conversation body — the first user message is identical across turns so
|
|
// extractSessionAffinityKey() fingerprints the same conversation each time.
|
|
function conversationBody() {
|
|
return {
|
|
messages: [{ role: "user", content: "Refactor the streaming handler for combos." }],
|
|
};
|
|
}
|
|
|
|
async function cleanupTestDataDir() {
|
|
let lastError;
|
|
for (let attempt = 0; attempt < 5; attempt += 1) {
|
|
try {
|
|
core.resetDbInstance();
|
|
fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
|
|
return;
|
|
} catch (error: any) {
|
|
lastError = error;
|
|
await new Promise((resolve) => setTimeout(resolve, 25));
|
|
}
|
|
}
|
|
if (lastError) throw lastError;
|
|
}
|
|
|
|
test.beforeEach(async () => {
|
|
clearSessions();
|
|
await cleanupTestDataDir();
|
|
fs.mkdirSync(TEST_DATA_DIR, { recursive: true });
|
|
});
|
|
|
|
test.after(async () => {
|
|
clearSessions();
|
|
if (ORIGINAL_DATA_DIR === undefined) {
|
|
delete process.env.DATA_DIR;
|
|
} else {
|
|
process.env.DATA_DIR = ORIGINAL_DATA_DIR;
|
|
}
|
|
await cleanupTestDataDir();
|
|
});
|
|
|
|
test("context_cache_protection ON + NO sessionId: combo re-pins to the same model across turns (#3825)", async () => {
|
|
const calls: string[] = [];
|
|
// Priority strategy: turn 1 deterministically tries model-a first, but model-a returns
|
|
// an empty (quality-failed) response so the combo falls through to model-b and records
|
|
// model-b as the session model (this exercises the real main-loop write site).
|
|
const combo = {
|
|
name: "sticky-priority",
|
|
strategy: "priority",
|
|
context_cache_protection: true,
|
|
models: ["model-a", "model-b"],
|
|
config: { maxRetries: 0 },
|
|
};
|
|
|
|
const handleSingleModel = async (_body: any, modelStr: string) => {
|
|
calls.push(modelStr);
|
|
if (modelStr === "model-a") {
|
|
return okResponse({ choices: [{ message: { content: "" } }] }); // empty → fall through
|
|
}
|
|
return okResponse({ choices: [{ message: { content: "fallback ok" } }] });
|
|
};
|
|
|
|
// Two sequential turns of the SAME conversation, NO relayOptions.sessionId.
|
|
for (let turn = 0; turn < 2; turn += 1) {
|
|
const result = await handleComboChat({
|
|
body: conversationBody(),
|
|
combo,
|
|
handleSingleModel: handleSingleModel as any,
|
|
isModelAvailable: async () => true,
|
|
log: createLog(),
|
|
settings: null,
|
|
relayOptions: null as any,
|
|
allCombos: null,
|
|
});
|
|
assert.equal(result.ok, true);
|
|
// Let the best-effort session-model write settle before the next turn reads it.
|
|
await waitForBackgroundWork();
|
|
}
|
|
|
|
// Turn 1 records model-b (the model that actually succeeded). With the fix, turn 2 reads
|
|
// that pin via the derived sessionless key and dispatches model-b DIRECTLY, skipping
|
|
// model-a entirely → calls === [a, b, b]. Without the fix the derived key is never used,
|
|
// so turn 2 re-runs priority from the top → calls === [a, b, a, b].
|
|
assert.deepEqual(
|
|
calls,
|
|
["model-a", "model-b", "model-b"],
|
|
"turn 2 must re-pin to model-b (turn 1's selected model) and skip model-a"
|
|
);
|
|
});
|
|
|
|
test("context_cache_protection OFF + NO sessionId: no pin, strategy re-runs each turn, no <omniModel> tag (#3825 scope guard)", async () => {
|
|
const calls: string[] = [];
|
|
const forwardedContents: string[] = [];
|
|
// Identical priority+failover setup as the positive case, but with the toggle OFF.
|
|
const combo = {
|
|
name: "unprotected-priority",
|
|
strategy: "priority",
|
|
// context_cache_protection intentionally omitted (falsy) — #3399 behavior preserved.
|
|
models: ["model-a", "model-b"],
|
|
config: { maxRetries: 0 },
|
|
};
|
|
|
|
const handleSingleModel = async (forwardedBody: any, modelStr: string) => {
|
|
calls.push(modelStr);
|
|
const msgs = Array.isArray(forwardedBody?.messages) ? forwardedBody.messages : [];
|
|
for (const m of msgs) {
|
|
if (typeof m?.content === "string") forwardedContents.push(m.content);
|
|
}
|
|
if (modelStr === "model-a") {
|
|
return okResponse({ choices: [{ message: { content: "" } }] });
|
|
}
|
|
return okResponse({ choices: [{ message: { content: "fallback ok" } }] });
|
|
};
|
|
|
|
for (let turn = 0; turn < 2; turn += 1) {
|
|
const result = await handleComboChat({
|
|
body: conversationBody(),
|
|
combo,
|
|
handleSingleModel: handleSingleModel as any,
|
|
isModelAvailable: async () => true,
|
|
log: createLog(),
|
|
settings: null,
|
|
relayOptions: null as any,
|
|
allCombos: null,
|
|
});
|
|
assert.equal(result.ok, true);
|
|
await waitForBackgroundWork();
|
|
}
|
|
|
|
// Toggle OFF → no pin is recorded or read → turn 2 re-runs priority from the top,
|
|
// re-trying model-a before falling through to model-b → [a, b, a, b].
|
|
assert.deepEqual(
|
|
calls,
|
|
["model-a", "model-b", "model-a", "model-b"],
|
|
"with protection OFF the combo must re-run strategy each turn (no sessionless pin)"
|
|
);
|
|
assert.ok(
|
|
forwardedContents.every((c) => !c.includes("<omniModel>")),
|
|
"no <omniModel> tag may be injected into the forwarded body (#454/#3399)"
|
|
);
|
|
});
|