wehub-resource-sync
88 lines
3.3 KiB
TypeScript
88 lines
3.3 KiB
TypeScript
/**
|
|
* Regression test for #6220 — GitLab Duo executor must feed tool results back into the
|
|
* prompt so the agent continues instead of re-emitting the same tool_call forever.
|
|
*
|
|
* Complements #6051 (PR #6111), which taught the executor to DETECT `<tool>` blocks and
|
|
* emit a valid OpenAI `tool_calls` response. #6220 is the follow-up turn: once the client
|
|
* appends `assistant{tool_calls}` + `tool{result}` and calls again, `buildPrompt()` used
|
|
* to branch only on `system`/`user` and take `userParts.at(-1)`, silently dropping the
|
|
* assistant and tool messages. The reconstructed prompt was byte-identical to turn 1, so
|
|
* the model re-derived the same `<tool>` call → infinite loop.
|
|
*
|
|
* Fix: when the message array carries a tool exchange (an assistant with `tool_calls` or
|
|
* a `tool` result), serialize the full conversation — including the tool result keyed by
|
|
* `tool_call_id` — so the prompt differs from turn 1 and carries the observation.
|
|
*/
|
|
import test from "node:test";
|
|
import assert from "node:assert/strict";
|
|
import { buildPrompt } from "../../open-sse/executors/gitlab.ts";
|
|
|
|
const SYSTEM = { role: "system", content: "You are a helpful assistant." };
|
|
const USER = { role: "user", content: "What's the weather in Paris?" };
|
|
|
|
test("buildPrompt: simple system+user prompt is unchanged (legacy path) [#6220]", () => {
|
|
const prompt = buildPrompt([SYSTEM, USER]);
|
|
assert.match(prompt, /System instructions:/);
|
|
assert.match(prompt, /What's the weather in Paris\?/);
|
|
// No tool exchange → no conversation serialization markers.
|
|
assert.doesNotMatch(prompt, /Tool result/);
|
|
});
|
|
|
|
test("buildPrompt: only the latest user message is used when no tool exchange (legacy) [#6220]", () => {
|
|
const prompt = buildPrompt([
|
|
{ role: "user", content: "first" },
|
|
{ role: "user", content: "second" },
|
|
]);
|
|
assert.equal(prompt, "second");
|
|
});
|
|
|
|
test("buildPrompt: tool result is fed back and prompt differs from turn 1 [#6220]", () => {
|
|
const turn1 = buildPrompt([SYSTEM, USER]);
|
|
|
|
const turn2 = buildPrompt([
|
|
SYSTEM,
|
|
USER,
|
|
{
|
|
role: "assistant",
|
|
content: "",
|
|
tool_calls: [
|
|
{
|
|
id: "call_abc123",
|
|
type: "function",
|
|
function: { name: "get_weather", arguments: '{"city":"Paris"}' },
|
|
},
|
|
],
|
|
},
|
|
{ role: "tool", tool_call_id: "call_abc123", name: "get_weather", content: '{"temp_c":21}' },
|
|
]);
|
|
|
|
// The tool result must be present…
|
|
assert.match(turn2, /temp_c/);
|
|
assert.match(turn2, /21/);
|
|
// …the tool_call_id must key the observation…
|
|
assert.match(turn2, /call_abc123/);
|
|
// …and the prompt must NOT be identical to turn 1 (that identity was the loop).
|
|
assert.notEqual(turn2, turn1);
|
|
});
|
|
|
|
test("buildPrompt: multiple tool results are each keyed by their tool_call_id [#6220]", () => {
|
|
const prompt = buildPrompt([
|
|
USER,
|
|
{
|
|
role: "assistant",
|
|
content: "",
|
|
tool_calls: [
|
|
{ id: "call_1", type: "function", function: { name: "a", arguments: "{}" } },
|
|
{ id: "call_2", type: "function", function: { name: "b", arguments: "{}" } },
|
|
],
|
|
},
|
|
{ role: "tool", tool_call_id: "call_1", name: "a", content: "result-A" },
|
|
{ role: "tool", tool_call_id: "call_2", name: "b", content: "result-B" },
|
|
]);
|
|
|
|
assert.match(prompt, /call_1/);
|
|
assert.match(prompt, /result-A/);
|
|
assert.match(prompt, /call_2/);
|
|
assert.match(prompt, /result-B/);
|
|
});
|