Files
OmniRoute/tests/unit/openai-responses-reasoning-effort.test.ts
VXNCXNX a5a78fd1d8 fix(kiro): preserve GPT-5.6 Max reasoning via Responses (#9163)
Validated in local merge-train (devbox-vm-06-dev002) @ combined-tip (FAST gates green: static + changed tests + vitest — only pre-existing audit.test.ts flake). Evidence: /home/diegosouzapw/dev/proxys/OmniRoute/.claude/worktrees/merge-train-20260805-213228-suite.log
2026-08-05 21:44:02 -03:00

117 lines
4.1 KiB
TypeScript

/**
* Reasoning-effort mapping across the OpenAI Chat <-> Responses request translators.
*
* Chat Completions carries the reasoning hint as top-level `reasoning_effort`; the
* Responses API nests it as `reasoning.effort`. These tests pin both directions so
* the hint survives when a request crosses formats (e.g. a Responses client routed
* to an OpenAI-native Chat Completions upstream).
*
* Ported from upstream PR https://github.com/decolua/9router/pull/1817 (ryanngit).
* Adapted: OmniRoute previously promoted `reasoning.effort` only behind the
* Copilot-client gate (commit 75d9a83c25), which silently dropped the field for
* every other Responses client (OpenCode, Cursor, raw OpenAI Responses, ...).
* This test pins the unconditional promotion of effort while keeping the
* Copilot-only `summary` -> Claude thinking marker behind its existing gate.
*/
import test from "node:test";
import assert from "node:assert/strict";
import {
openaiToOpenAIResponsesRequest,
openaiResponsesToOpenAIRequest,
} from "../../open-sse/translator/request/openai-responses.ts";
import { convertResponsesApiFormat } from "../../open-sse/translator/helpers/responsesApiHelper.ts";
import { buildKiroPayload } from "../../open-sse/translator/request/openai-to-kiro.ts";
function asRecord(value: unknown): Record<string, unknown> {
return value as Record<string, unknown>;
}
test("Responses -> Chat promotes reasoning.effort for non-Copilot clients", () => {
const out = asRecord(
openaiResponsesToOpenAIRequest(
"gpt-test",
{ input: "hello", reasoning: { effort: "high" } },
true,
{} // no _copilotClient marker
)
);
assert.equal(out.reasoning_effort, "high");
assert.equal(out.reasoning, undefined);
});
test("Responses -> Ollama Cloud Chat preserves every advertised reasoning effort", () => {
for (const effort of ["low", "medium", "high"]) {
const out = asRecord(
openaiResponsesToOpenAIRequest(
"ollama-cloud/gpt-oss:20b",
{ input: "hello", reasoning: { effort } },
true,
{ _provider: "ollama-cloud" }
)
);
assert.equal(out.reasoning_effort, effort);
assert.equal(out.reasoning, undefined);
}
});
test("Responses -> Chat preserves reasoning.effort via the helper wrapper", () => {
const out = asRecord(
convertResponsesApiFormat({ input: "hello", reasoning: { effort: "medium" } })
);
assert.equal(out.reasoning_effort, "medium");
assert.equal(out.reasoning, undefined);
});
test("Responses -> Kiro preserves literal Max for GPT-5.6 models", () => {
for (const model of ["gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna"]) {
const converted = asRecord(
convertResponsesApiFormat(
{
model: `kr/${model}`,
input: "hello",
reasoning: { effort: "max" },
},
null,
"kiro"
)
);
assert.equal(converted.reasoning_effort, "max");
const payload = buildKiroPayload(model, converted, false, null);
assert.equal(payload.additionalModelRequestFields?.reasoning?.effort, "max");
assert.equal(payload.additionalModelRequestFields?.output_config, undefined);
assert.equal(payload.additionalModelRequestFields?.thinking, undefined);
assert.equal(payload.additionalModelRequestFields?.max_tokens, undefined);
assert.doesNotMatch(
payload.conversationState.currentMessage.userInputMessage.content,
/<thinking_mode>/
);
}
});
test("Responses -> Chat does not overwrite an explicit reasoning_effort", () => {
const out = asRecord(
openaiResponsesToOpenAIRequest(
"gpt-test",
{ input: "hello", reasoning_effort: "low", reasoning: { effort: "high" } },
true,
{}
)
);
// Explicit Chat-level field wins over the Responses nesting.
assert.equal(out.reasoning_effort, "low");
});
test("Chat -> Responses already wraps reasoning_effort into reasoning.effort", () => {
const out = asRecord(
openaiToOpenAIResponsesRequest(
"gpt-test",
{ messages: [{ role: "user", content: "hi" }], reasoning_effort: "high" },
true,
{}
)
);
assert.deepEqual(out.reasoning, { effort: "high", summary: "auto" });
});