Welington
0f488c7027
Anthropic's API-level refusal (streaming classifier / ToS) ends the stream with stop_reason "refusal", stop_details carrying the reason, zero output tokens and no content blocks. Map refusal to content_filter in both directions, surface stop_details.explanation as message text, and add CLAUDE_STOP.REFUSAL to schema.
77 lines
3.2 KiB
JavaScript
77 lines
3.2 KiB
JavaScript
// A refusal from the Anthropic API (stop_reason "refusal", zero output tokens, no
|
|
// content blocks) must reach an OpenAI-format client as finish_reason
|
|
// "content_filter" carrying Anthropic's explanation — not as a clean, empty "stop".
|
|
// Captured live on 2026-09-20 against claude-opus-5 via a Claude Code OAuth
|
|
// connection: 9Router logged "Model succeeded · OUT 0" and the client saw nothing.
|
|
import { describe, it, expect } from "vitest";
|
|
import { claudeToOpenAIResponse } from "../../open-sse/translator/response/claude-to-openai.js";
|
|
|
|
const EXPLANATION =
|
|
"This request was blocked as it seems to violate Anthropic's Terms of Service restrictions on reverse engineering or duplicating model outputs.";
|
|
|
|
function runStream(events) {
|
|
const state = {};
|
|
const out = [];
|
|
for (const ev of events) {
|
|
const r = claudeToOpenAIResponse(ev, state);
|
|
if (Array.isArray(r)) out.push(...r);
|
|
else if (r) out.push(r);
|
|
}
|
|
return { state, out };
|
|
}
|
|
|
|
const refusalStream = [
|
|
{
|
|
type: "message_start",
|
|
message: {
|
|
id: "msg_refusal", model: "claude-opus-5", role: "assistant", content: [],
|
|
usage: { input_tokens: 637, cache_creation_input_tokens: 206779, cache_read_input_tokens: 0, output_tokens: 0 }
|
|
}
|
|
},
|
|
{
|
|
type: "message_delta",
|
|
delta: {
|
|
stop_reason: "refusal",
|
|
stop_sequence: null,
|
|
stop_details: { type: "refusal", category: "reasoning_extraction", explanation: EXPLANATION }
|
|
},
|
|
usage: { input_tokens: 637, cache_creation_input_tokens: 206779, cache_read_input_tokens: 0, output_tokens: 0 }
|
|
},
|
|
{ type: "message_stop" }
|
|
];
|
|
|
|
describe("claude-to-openai: refusal stop_reason", () => {
|
|
it("finishes with content_filter, not stop", () => {
|
|
const { out } = runStream(refusalStream);
|
|
const finishes = out.map(c => c.choices?.[0]?.finish_reason).filter(Boolean);
|
|
expect(finishes).toEqual(["content_filter"]);
|
|
});
|
|
|
|
it("surfaces Anthropic's explanation as message content", () => {
|
|
const { out } = runStream(refusalStream);
|
|
const text = out.map(c => c.choices?.[0]?.delta?.content || "").join("");
|
|
expect(text).toBe(EXPLANATION);
|
|
});
|
|
|
|
it("keeps usage on the final chunk (prompt tokens were billed)", () => {
|
|
const { out } = runStream(refusalStream);
|
|
const final = out.find(c => c.choices?.[0]?.finish_reason === "content_filter");
|
|
expect(final.usage.prompt_tokens).toBe(637 + 206779);
|
|
expect(final.usage.completion_tokens).toBe(0);
|
|
});
|
|
|
|
it("leaves a normal end_turn untouched", () => {
|
|
const { out } = runStream([
|
|
{ type: "message_start", message: { id: "m", model: "claude-opus-5", role: "assistant", content: [], usage: { input_tokens: 5, output_tokens: 0 } } },
|
|
{ type: "content_block_start", index: 0, content_block: { type: "text", text: "" } },
|
|
{ type: "content_block_delta", index: 0, delta: { type: "text_delta", text: "ok" } },
|
|
{ type: "content_block_stop", index: 0 },
|
|
{ type: "message_delta", delta: { stop_reason: "end_turn", stop_sequence: null, stop_details: null }, usage: { output_tokens: 1 } },
|
|
{ type: "message_stop" }
|
|
]);
|
|
const finishes = out.map(c => c.choices?.[0]?.finish_reason).filter(Boolean);
|
|
expect(finishes).toEqual(["stop"]);
|
|
expect(out.map(c => c.choices?.[0]?.delta?.content || "").join("")).toBe("ok");
|
|
});
|
|
});
|