Files
pikasTech-HWLAB/web/hwlab-cloud-web/scripts/workbench-r1-parity.test.ts
T

126 lines
5.0 KiB
TypeScript

import assert from "node:assert/strict";
import test from "node:test";
import type { AgentRunProvenance, ChatMessage, TraceEvent } from "../src/types/index.ts";
import { canCancelMessage, canReplayTrace, canRetryMessage, messageTraceId, renderSafeMarkdown, traceEventBody, traceEventLabel, traceIdentityText, visibleTraceEvents } from "../src/components/workbench/message-rendering.ts";
test("R1 markdown rendering keeps structure and strips unsafe HTML", () => {
const html = renderSafeMarkdown("# 标题\n\n- 项目\n\n<script>alert('x')</script>\n\n[link](https://example.com)");
assert.match(html, /<h1>标题<\/h1>/u);
assert.match(html, /<li>项目<\/li>/u);
assert.match(html, /href="https:\/\/example.com"/u);
assert.doesNotMatch(html, /<script|alert/u);
});
test("R1 message actions follow per-message status and trace rules", () => {
const running = agentMessage({ status: "running", traceId: "trc_running", retryInput: "retry me" });
assert.equal(messageTraceId(running), "trc_running");
assert.equal(canCancelMessage(running), true);
assert.equal(canRetryMessage(running), true);
assert.equal(canReplayTrace(running), true);
const completed = agentMessage({ status: "completed", traceId: "trc_done", retryInput: "retry me" });
assert.equal(canCancelMessage(completed), false);
assert.equal(canRetryMessage(completed), false);
assert.equal(canReplayTrace(completed), true);
const failed = agentMessage({ status: "failed", traceId: "trc_failed", retryInput: "retry me" });
assert.equal(canCancelMessage(failed), false);
assert.equal(canRetryMessage(failed), true);
assert.equal(canReplayTrace(failed), true);
const user: ChatMessage = { id: "msg_user", role: "user", title: "用户", text: "hi", status: "sent", createdAt: "2026-01-01T00:00:00.000Z", traceId: "trc_user" };
assert.equal(canReplayTrace(user), false);
});
test("R1 trace rendering hides chunk noise without dropping meaningful events", () => {
const events: TraceEvent[] = [
{ label: "assistant:delta", message: "token" },
{ label: "command:completed", command: "npm test", exitCode: 0, elapsedMs: 42 },
{ type: "assistant_message", text: "final", final: true }
];
const compact = visibleTraceEvents(events, true);
assert.equal(compact.length, 2);
assert.equal(traceEventLabel(compact[0]!), "command:completed");
assert.equal(traceEventBody(compact[0]!), "npm test");
assert.equal(traceEventLabel(compact[1]!), "assistant_message");
assert.equal(traceEventBody(compact[1]!), "final");
assert.equal(visibleTraceEvents(events, false).length, 3);
});
test("R1 status summary keeps the restored 23-row floor", () => {
const message = agentMessage({
status: "failed",
traceId: "trc_summary",
conversationId: "cnv_summary",
sessionId: "ses_summary",
threadId: "thr_summary",
runnerTrace: {
traceId: "trc_summary",
status: "failed",
eventCount: 2,
eventsCompacted: false,
fullTraceLoaded: true,
lastEventLabel: "assistant:completed",
waitingFor: "result",
runnerKind: "agentrun",
sessionMode: "long-lived",
agentRun: {
runId: "run_1",
commandId: "cmd_1",
attemptId: "att_1",
runnerId: "runner_1",
jobName: "job_1",
namespace: "hwlab-v03",
terminalStatus: "failed"
}
},
error: { code: "provider_timeout", category: "provider", providerStatus: 504 }
});
const values = [
message.status,
message.runnerTrace?.status,
message.traceId,
message.conversationId,
message.sessionId,
message.threadId,
message.runnerTrace?.runnerKind,
message.runnerTrace?.sessionMode,
message.runnerTrace?.waitingFor,
message.runnerTrace?.lastEventLabel,
String(message.runnerTrace?.eventCount),
String(message.runnerTrace?.eventsCompacted),
message.runnerTrace?.fullTraceLoaded === true ? "loaded" : "unverified",
asAgentRun(message.runnerTrace?.agentRun)?.runId,
asAgentRun(message.runnerTrace?.agentRun)?.commandId,
asAgentRun(message.runnerTrace?.agentRun)?.attemptId,
asAgentRun(message.runnerTrace?.agentRun)?.runnerId,
asAgentRun(message.runnerTrace?.agentRun)?.jobName,
asAgentRun(message.runnerTrace?.agentRun)?.namespace,
asAgentRun(message.runnerTrace?.agentRun)?.terminalStatus,
message.error?.code,
message.error?.category,
String(message.error?.providerStatus)
];
assert.equal(values.length, 23);
assert.equal(values.every((value) => value != null && value !== ""), true);
assert.match(traceIdentityText(message), /traceId=trc_summary/u);
assert.match(traceIdentityText(message), /sessionId=ses_summary/u);
});
function agentMessage(patch: Partial<ChatMessage>): ChatMessage {
return {
id: "msg_agent",
role: "agent",
title: "Code Agent",
text: "done",
status: "completed",
createdAt: "2026-01-01T00:00:00.000Z",
...patch
} as ChatMessage;
}
function asAgentRun(value: unknown): AgentRunProvenance | null {
return value && typeof value === "object" ? value as AgentRunProvenance : null;
}