SSL pi-embedded-runner.cache.live.test.ts

import fs from "node:fs/promises";
import os from "node:os";
import path from "node:path";
import type { AssistantMessage, Message, Tool } from "@mariozechner/pi-ai";
import { Type } from "typebox";
import { afterAll, beforeAll, describe, expect, it } from "vitest";
import type { OpenClawConfig } from "../config/config.js";
import {
  buildAssistantHistoryTurn as buildTypedAssistantHistoryTurn,
  buildStableCachePrefix,
  completeSimpleWithLiveTimeout,
  computeCacheHitRate,
  extractAssistantText,
  LIVE_CACHE_TEST_ENABLED,
  logLiveCache,
  resolveLiveDirectModel,
  withLiveCacheHeartbeat,
} from "./live-cache-test-support.js";
import { runEmbeddedPiAgent } from "./pi-embedded-runner.js";
import { compactEmbeddedPiSessionDirect } from "./pi-embedded-runner/compact.runtime.js";
import { buildZeroUsage } from "./stream-message-shared.js";

const describeCacheLive = LIVE_CACHE_TEST_ENABLED ? describe : describe.skip;

const OPENAI_TIMEOUT_MS = 120_000;
const ANTHROPIC_TIMEOUT_MS = 120_000;
const OPENAI_SESSION_ID = "live-cache-openai-stable-session";
const ANTHROPIC_SESSION_ID = "live-cache-anthropic-stable-session";
const OPENAI_PREFIX = buildStableCachePrefix("openai");
const ANTHROPIC_PREFIX = buildStableCachePrefix("anthropic");
const OPENAI_STABLE_PREFIX_MIN_CACHE_READ = 4_608;
const OPENAI_STABLE_PREFIX_MIN_HIT_RATE = 0.9;
const OPENAI_TOOL_MIN_CACHE_READ = 4_096;
const OPENAI_TOOL_MIN_HIT_RATE = 0.85;
const OPENAI_IMAGE_MIN_CACHE_READ = 3_840;
const OPENAI_IMAGE_MIN_HIT_RATE = 0.82;
const LIVE_TEST_PNG_URL = new URL(
  "../../apps/android/app/src/main/res/mipmap-xhdpi/ic_launcher.png",
  import.meta.url,
);

type CacheRun = {
  hitRate: number;
  suffix: string;
  text: string;
  usage: AssistantMessage["usage"];
};
type CacheTraceEvent = {
  sessionId?: string;
  stage?: string;
  note?: string;
  options?: {
    previousCacheRead?: number;
    cacheRead?: number;
    changes?: Array<{ code?: string; detail?: string }>;
  };
};
type LiveResolvedModel = Awaited<ReturnType<typeof resolveLiveDirectModel>>;

const NOOP_TOOL: Tool = {
  name: "noop",
  description: "Return ok.",
  parameters: Type.Object({}, { additionalProperties: false }),
};
let liveTestPngBase64 = "";
let liveRunnerRootDir: string | undefined;
let liveCacheTraceFile: string | undefined;
let previousCacheTraceEnv: {
  enabled?: string;
  file?: string;
  messages?: string;
  prompt?: string;
  system?: string;
} | null = null;

type UserContent = Extract<Message, { role: "user" }>["content"];

function makeAssistantHistoryTurn(
  text: string,
  model?: Awaited<ReturnType<typeof resolveLiveDirectModel>>["model"],
): Message {
  return buildTypedAssistantHistoryTurn(text, model);
}

function makeUserHistoryTurn(content: UserContent): Message {
  return {
    role: "user",
    content,
    timestamp: Date.now(),
  };
}

function makeImageUserTurn(text: string): Message {
  if (!liveTestPngBase64) {
    throw new Error("live test PNG not loaded");
  }
  return makeUserHistoryTurn([
    { type: "text", text },
    { type: "image", mimeType: "image/png", data: liveTestPngBase64 },
  ]);
}

function buildRunnerSessionPaths(sessionId: string) {
  if (!liveRunnerRootDir) {
    throw new Error("live runner temp root not initialized");
  }
  return {
    agentDir: liveRunnerRootDir,
    sessionFile: path.join(liveRunnerRootDir, `${sessionId}.jsonl`),
    workspaceDir: path.join(liveRunnerRootDir, `${sessionId}-workspace`),
  };
}

function resolveProviderBaseUrl(model: LiveResolvedModel["model"]): string | undefined {
  const candidate = (model as { baseUrl?: unknown }).baseUrl;
  return typeof candidate === "string" && candidate.trim().length > 0 ? candidate : undefined;
}

async function readCacheTraceEvents(sessionId: string): Promise<CacheTraceEvent[]> {
  if (!liveCacheTraceFile) {
    throw new Error("live cache trace file not initialized");
  }
  const raw = await fs.readFile(liveCacheTraceFile, "utf8").catch(() => "");
  return raw
    .split("\n")
    .map((line) => line.trim())
    .filter(Boolean)
    .map((line) => JSON.parse(line) as CacheTraceEvent)
    .filter((event) => event.sessionId === sessionId);
}

async function expectCacheTraceStages(
  sessionId: string,
  requiredStages: Array<"cache:state" | "cache:result">,
): Promise<void> {
  const events = await readCacheTraceEvents(sessionId);
  const stages = new Set(events.map((event) => event.stage));
  for (const stage of requiredStages) {
    expect(stages.has(stage)).toBe(true);
  }
}

function resolveDefaultProviderBaseUrl(model: LiveResolvedModel["model"]): string {
  if (model.provider === "anthropic") {
    return "https://api.anthropic.com/v1";
  }
  if (model.provider === "openai") {
    return "https://api.openai.com/v1";
  }
  return "https://example.invalid/v1";
}

function buildEmbeddedModelDefinition(model: LiveResolvedModel["model"]) {
  const contextWindowCandidate = (model as { contextWindow?: unknown }).contextWindow;
  const maxTokensCandidate = (model as { maxTokens?: unknown }).maxTokens;
  const reasoningCandidate = (model as { reasoning?: unknown }).reasoning;
  const inputCandidate = (model as { input?: unknown }).input;
  const contextWindow =
    typeof contextWindowCandidate === "number" && Number.isFinite(contextWindowCandidate)
      ? Math.max(1, Math.trunc(contextWindowCandidate))
      : 128_000;
  const maxTokens =
    typeof maxTokensCandidate === "number" && Number.isFinite(maxTokensCandidate)
      ? Math.max(1, Math.trunc(maxTokensCandidate))
      : 8_192;
  const input =
    Array.isArray(inputCandidate) &&
    inputCandidate.every((value) => value === "text" || value === "image")
      ? [...inputCandidate]
      : (["text", "image"] as Array<"text" | "image">);
  return {
    id: model.id,
    name: model.id,
    api: resolveEmbeddedModelApi(model),
    reasoning: typeof reasoningCandidate === "boolean" ? reasoningCandidate : false,
    input,
    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
    contextWindow,
    maxTokens,
  };
}

function resolveEmbeddedModelApi(
  model: LiveResolvedModel["model"],
): "anthropic-messages" | "openai-responses" {
  return model.provider === "anthropic" ? "anthropic-messages" : "openai-responses";
}

function normalizeLiveUsage(
  usage:
    | AssistantMessage["usage"]
    | {
        input?: number;
        output?: number;
        cacheRead?: number;
        cacheWrite?: number;
        total?: number;
      }
    | undefined,
): AssistantMessage["usage"] {
  if (!usage) {
    return buildZeroUsage();
  }
  const input = usage.input ?? 0;
  const output = usage.output ?? 0;
  const cacheRead = usage.cacheRead ?? 0;
  const cacheWrite = usage.cacheWrite ?? 0;
  const totalTokens =
    "totalTokens" in usage && typeof usage.totalTokens === "number"
      ? usage.totalTokens
      : "total" in usage && typeof usage.total === "number"
        ? usage.total
        : input + output;
  const cost =
    "cost" in usage && usage.cost
      ? usage.cost
      : { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 };
  return {
    input,
    output,
    cacheRead,
    cacheWrite,
    totalTokens,
    cost,
  };
}

function buildEmbeddedRunnerConfig(
  params: LiveResolvedModel & {
    cacheRetention: "none" | "short" | "long";
    transport?: "sse" | "websocket";
  },
): OpenClawConfig {
  const provider = params.model.provider;
  const modelKey = `${provider}/${params.model.id}`;
  const providerBaseUrl =
    resolveProviderBaseUrl(params.model) ?? resolveDefaultProviderBaseUrl(params.model);
  return {
    models: {
      providers: {
        [provider]: {
          api: resolveEmbeddedModelApi(params.model),
          auth: "api-key",
          apiKey: params.apiKey,
          baseUrl: providerBaseUrl,
          models: [buildEmbeddedModelDefinition(params.model)],
        },
      },
    },
    agents: {
      defaults: {
        models: {
          [modelKey]: {
            params: {
              cacheRetention: params.cacheRetention,
              ...(params.transport ? { transport: params.transport } : {}),
            },
          },
        },
      },
    },
  };
}

function buildEmbeddedCachePrompt(suffix: string, sections = 48): string {
  const lines = [
    `Reply with exactly CACHE-OK ${suffix}.`,
    "Do not add any extra words or punctuation.",
  ];
  for (let index = 0; index < sections; index += 1) {
    lines.push(
      `Embedded cache section ${index + 1}: deterministic prose about prompt stability, session affinity, request shaping, transport continuity, and cache reuse across identical stable prefixes.`,
    );
  }
  return lines.join("\n");
}

function buildNoisyStructuredPromptVariant(text: string): string {
  return `\r\n${text
    .split("\n")
    .map((line) => `${line}  \t`)
    .join("\r\n")}\r\n\r\n`;
}

function extractRunPayloadText(payloads: Array<{ text?: string } | undefined> | undefined): string {
  return (
    payloads
      ?.map((payload) => payload?.text?.trim())
      .filter((text): text is string => Boolean(text))
      .join(" ") ?? ""
  );
}

async function runEmbeddedCacheProbe(params: {
  apiKey: string;
  cacheRetention: "none" | "short" | "long";
  model: LiveResolvedModel["model"];
  prefix: string;
  providerTag: "anthropic" | "openai";
  sessionId: string;
  suffix: string;
  transport?: "sse" | "websocket";
  promptSections?: number;
}): Promise<CacheRun> {
  const sessionPaths = buildRunnerSessionPaths(params.sessionId);
  await fs.mkdir(sessionPaths.workspaceDir, { recursive: true });
  const result = await withLiveCacheHeartbeat(
    runEmbeddedPiAgent({
      sessionId: params.sessionId,
      sessionKey: `live-cache:${params.providerTag}:${params.sessionId}`,
      sessionFile: sessionPaths.sessionFile,
      workspaceDir: sessionPaths.workspaceDir,
      agentDir: sessionPaths.agentDir,
      config: buildEmbeddedRunnerConfig({
        apiKey: params.apiKey,
        cacheRetention: params.cacheRetention,
        model: params.model,
        transport: params.transport,
      }),
      prompt: buildEmbeddedCachePrompt(params.suffix, params.promptSections),
      provider: params.model.provider,
      model: params.model.id,
      timeoutMs: params.providerTag === "openai" ? OPENAI_TIMEOUT_MS : ANTHROPIC_TIMEOUT_MS,
      runId: `${params.sessionId}-${params.suffix}-${params.transport ?? "default"}`,
      extraSystemPrompt: params.prefix,
      disableTools: true,
      cleanupBundleMcpOnRunEnd: true,
    }),
    `${params.providerTag} embedded cache probe ${params.suffix}${params.transport ? ` (${params.transport})` : ""}`,
  );
  const text = extractRunPayloadText(result.payloads);
  expect(text.toLowerCase()).toContain(params.suffix.toLowerCase());
  const usage = normalizeLiveUsage(result.meta.agentMeta?.usage);
  return {
    suffix: params.suffix,
    text,
    usage,
    hitRate: computeCacheHitRate(usage),
  };
}

async function compactLiveCacheSession(params: {
  apiKey: string;
  cacheRetention: "none" | "short" | "long";
  model: LiveResolvedModel["model"];
  providerTag: "anthropic" | "openai";
  sessionId: string;
}) {
  const sessionPaths = buildRunnerSessionPaths(params.sessionId);
  await fs.mkdir(sessionPaths.workspaceDir, { recursive: true });
  return await withLiveCacheHeartbeat(
    compactEmbeddedPiSessionDirect({
      sessionId: params.sessionId,
      sessionKey: `live-cache:${params.providerTag}:${params.sessionId}`,
      sessionFile: sessionPaths.sessionFile,
      workspaceDir: sessionPaths.workspaceDir,
      agentDir: sessionPaths.agentDir,
      config: buildEmbeddedRunnerConfig({
        apiKey: params.apiKey,
        cacheRetention: params.cacheRetention,
        model: params.model,
      }),
      provider: params.model.provider,
      model: params.model.id,
      force: true,
      trigger: "manual",
      runId: `${params.sessionId}-compact`,
      tokenBudget: 512,
    }),
    `${params.providerTag} embedded compaction ${params.sessionId}`,
  );
}

function extractFirstToolCall(message: AssistantMessage) {
  return message.content.find((block) => block.type === "toolCall");
}

function buildToolResultMessage(
  toolCallId: string,
  toolName = "noop",
  text = "ok",
): Extract<Message, { role: "toolResult" }> {
  return {
    role: "toolResult",
    toolCallId,
    toolName,
    content: [{ type: "text", text }],
    isError: false,
    timestamp: Date.now(),
  };
}

async function runToolOnlyTurn(params: {
  apiKey: string;
  cacheRetention: "none" | "short" | "long";
  model: Awaited<ReturnType<typeof resolveLiveDirectModel>>["model"];
  providerTag: "anthropic" | "openai";
  sessionId: string;
  systemPrompt: string;
  tool: Tool;
}) {
  let prompt = `Call the tool \`${params.tool.name}\` with {}. IMPORTANT: respond ONLY with the tool call and no other text.`;
  let response = await completeSimpleWithLiveTimeout(
    params.model,
    {
      systemPrompt: params.systemPrompt,
      messages: [
        {
          role: "user",
          content: prompt,
          timestamp: Date.now(),
        },
      ],
      tools: [params.tool],
    },
    {
      apiKey: params.apiKey,
      cacheRetention: params.cacheRetention,
      sessionId: params.sessionId,
      maxTokens: 128,
      temperature: 0,
      ...(params.providerTag === "openai" ? { reasoning: "none" as unknown as never } : {}),
    },
    `${params.providerTag} tool-only turn`,
    params.providerTag === "openai" ? OPENAI_TIMEOUT_MS : ANTHROPIC_TIMEOUT_MS,
  );

  let toolCall = extractFirstToolCall(response);
  let text = extractAssistantText(response);
  for (let attempt = 0; attempt < 2 && (!toolCall || text.length > 0); attempt += 1) {
    prompt = `Return only a tool call for \`${params.tool.name}\` with {}. No text.`;
    response = await completeSimpleWithLiveTimeout(
      params.model,
      {
        systemPrompt: params.systemPrompt,
        messages: [
          {
            role: "user",
            content: prompt,
            timestamp: Date.now(),
          },
        ],
        tools: [params.tool],
      },
      {
        apiKey: params.apiKey,
        cacheRetention: params.cacheRetention,
        sessionId: params.sessionId,
        maxTokens: 128,
        temperature: 0,
        ...(params.providerTag === "openai" ? { reasoning: "none" as unknown as never } : {}),
      },
      `${params.providerTag} tool-only retry ${attempt + 1}`,
      params.providerTag === "openai" ? OPENAI_TIMEOUT_MS : ANTHROPIC_TIMEOUT_MS,
    );
    toolCall = extractFirstToolCall(response);
    text = extractAssistantText(response);
  }

  expect(toolCall).toBeTruthy();
  expect(text.length).toBe(0);
  if (!toolCall || toolCall.type !== "toolCall") {
    throw new Error("expected tool call");
  }

  return {
    prompt,
    response,
    toolCall,
  };
}

async function runOpenAiToolCacheProbe(params: {
  apiKey: string;
  model: Awaited<ReturnType<typeof resolveLiveDirectModel>>["model"];
  sessionId: string;
  suffix: string;
}): Promise<CacheRun> {
  const toolTurn = await runToolOnlyTurn({
    apiKey: params.apiKey,
    cacheRetention: "short",
    model: params.model,
    providerTag: "openai",
    sessionId: params.sessionId,
    systemPrompt: OPENAI_PREFIX,
    tool: NOOP_TOOL,
  });
  const response = await completeSimpleWithLiveTimeout(
    params.model,
    {
      systemPrompt: OPENAI_PREFIX,
      messages: [
        {
          role: "user",
          content: toolTurn.prompt,
          timestamp: Date.now(),
        },
        toolTurn.response,
        buildToolResultMessage(toolTurn.toolCall.id, NOOP_TOOL.name, "ok"),
        makeAssistantHistoryTurn("TOOL HISTORY ACKNOWLEDGED", params.model),
        makeUserHistoryTurn("Keep the tool output stable in history."),
        makeAssistantHistoryTurn("TOOL HISTORY PRESERVED", params.model),
        {
          role: "user",
          content: `Reply with exactly CACHE-OK ${params.suffix}.`,
          timestamp: Date.now(),
        },
      ],
      tools: [NOOP_TOOL],
    },
    {
      apiKey: params.apiKey,
      cacheRetention: "short",
      sessionId: params.sessionId,
      maxTokens: 64,
      temperature: 0,
      reasoning: "none" as unknown as never,
    },
    `openai cache probe ${params.suffix}`,
    OPENAI_TIMEOUT_MS,
  );
  const text = extractAssistantText(response);
  expect(text.toLowerCase()).toContain(params.suffix.toLowerCase());
  return {
    suffix: params.suffix,
    text,
    usage: response.usage,
    hitRate: computeCacheHitRate(response.usage),
  };
}

async function runOpenAiCacheProbe(params: {
  apiKey: string;
  model: Awaited<ReturnType<typeof resolveLiveDirectModel>>["model"];
  sessionId: string;
  suffix: string;
}): Promise<CacheRun> {
  const response = await completeSimpleWithLiveTimeout(
    params.model,
    {
      systemPrompt: OPENAI_PREFIX,
      messages: [
        {
          role: "user",
          content: `Reply with exactly CACHE-OK ${params.suffix}.`,
          timestamp: Date.now(),
        },
      ],
    },
    {
      apiKey: params.apiKey,
      cacheRetention: "short",
      sessionId: params.sessionId,
      maxTokens: 32,
      temperature: 0,
    },
    `openai cache probe ${params.suffix}`,
    OPENAI_TIMEOUT_MS,
  );
  const text = extractAssistantText(response);
  expect(text.toLowerCase()).toContain(params.suffix.toLowerCase());
  return {
    suffix: params.suffix,
    text,
    usage: response.usage,
    hitRate: computeCacheHitRate(response.usage),
  };
}

async function runOpenAiImageCacheProbe(params: {
  apiKey: string;
  model: Awaited<ReturnType<typeof resolveLiveDirectModel>>["model"];
  sessionId: string;
  suffix: string;
}): Promise<CacheRun> {
  const response = await completeSimpleWithLiveTimeout(
    params.model,
    {
      systemPrompt: OPENAI_PREFIX,
      messages: [
        makeImageUserTurn(
          "An image is attached. Ignore image semantics but keep the bytes in history.",
        ),
        makeAssistantHistoryTurn("IMAGE HISTORY ACKNOWLEDGED", params.model),
        makeUserHistoryTurn("Keep the earlier image turn stable in context."),
        makeAssistantHistoryTurn("IMAGE HISTORY PRESERVED", params.model),
        makeUserHistoryTurn(`Reply with exactly CACHE-OK ${params.suffix}.`),
      ],
    },
    {
      apiKey: params.apiKey,
      cacheRetention: "short",
      sessionId: params.sessionId,
      maxTokens: 64,
      temperature: 0,
      reasoning: "none" as unknown as never,
    },
    `openai image cache probe ${params.suffix}`,
    OPENAI_TIMEOUT_MS,
  );
  const text = extractAssistantText(response);
  expect(text.toLowerCase()).toContain(params.suffix.toLowerCase());
  return {
    suffix: params.suffix,
    text,
    usage: response.usage,
    hitRate: computeCacheHitRate(response.usage),
  };
}

async function runAnthropicCacheProbe(params: {
  apiKey: string;
  model: Awaited<ReturnType<typeof resolveLiveDirectModel>>["model"];
  sessionId: string;
  suffix: string;
  cacheRetention: "none" | "short" | "long";
}): Promise<CacheRun> {
  const response = await completeSimpleWithLiveTimeout(
    params.model,
    {
      systemPrompt: ANTHROPIC_PREFIX,
      messages: [
        {
          role: "user",
          content: `Reply with exactly CACHE-OK ${params.suffix}.`,
          timestamp: Date.now(),
        },
      ],
    },
    {
      apiKey: params.apiKey,
      cacheRetention: params.cacheRetention,
      sessionId: params.sessionId,
      maxTokens: 32,
      temperature: 0,
    },
    `anthropic cache probe ${params.suffix} (${params.cacheRetention})`,
    ANTHROPIC_TIMEOUT_MS,
  );
  const text = extractAssistantText(response);
  expect(text.toLowerCase()).toContain(params.suffix.toLowerCase());
  return {
    suffix: params.suffix,
    text,
    usage: response.usage,
    hitRate: computeCacheHitRate(response.usage),
  };
}

async function runAnthropicToolCacheProbe(params: {
  apiKey: string;
  model: Awaited<ReturnType<typeof resolveLiveDirectModel>>["model"];
  sessionId: string;
  suffix: string;
  cacheRetention: "none" | "short" | "long";
}): Promise<CacheRun> {
  const toolTurn = await runToolOnlyTurn({
    apiKey: params.apiKey,
    cacheRetention: params.cacheRetention,
    model: params.model,
    providerTag: "anthropic",
    sessionId: params.sessionId,
    systemPrompt: ANTHROPIC_PREFIX,
    tool: NOOP_TOOL,
  });
  const response = await completeSimpleWithLiveTimeout(
    params.model,
    {
      systemPrompt: ANTHROPIC_PREFIX,
      messages: [
        {
          role: "user",
          content: toolTurn.prompt,
          timestamp: Date.now(),
        },
        toolTurn.response,
        buildToolResultMessage(toolTurn.toolCall.id, NOOP_TOOL.name, "ok"),
        makeAssistantHistoryTurn("TOOL HISTORY ACKNOWLEDGED", params.model),
        makeUserHistoryTurn("Keep the tool output stable in history."),
        makeAssistantHistoryTurn("TOOL HISTORY PRESERVED", params.model),
        {
          role: "user",
          content: `Reply with exactly CACHE-OK ${params.suffix}.`,
          timestamp: Date.now(),
        },
      ],
      tools: [NOOP_TOOL],
    },
    {
      apiKey: params.apiKey,
      cacheRetention: params.cacheRetention,
      sessionId: params.sessionId,
      maxTokens: 64,
      temperature: 0,
    },
    `anthropic cache probe ${params.suffix} (${params.cacheRetention})`,
    ANTHROPIC_TIMEOUT_MS,
  );
  const text = extractAssistantText(response);
  expect(text.toLowerCase()).toContain(params.suffix.toLowerCase());
  return {
    suffix: params.suffix,
    text,
    usage: response.usage,
    hitRate: computeCacheHitRate(response.usage),
  };
}

async function runAnthropicImageCacheProbe(params: {
  apiKey: string;
  model: Awaited<ReturnType<typeof resolveLiveDirectModel>>["model"];
  sessionId: string;
  suffix: string;
  cacheRetention: "none" | "short" | "long";
}): Promise<CacheRun> {
  const response = await completeSimpleWithLiveTimeout(
    params.model,
    {
      systemPrompt: ANTHROPIC_PREFIX,
      messages: [
        makeImageUserTurn(
          "An image is attached. Ignore image semantics but keep the bytes in history.",
        ),
        makeAssistantHistoryTurn("IMAGE HISTORY ACKNOWLEDGED", params.model),
        makeUserHistoryTurn("Keep the earlier image turn stable in context."),
        makeAssistantHistoryTurn("IMAGE HISTORY PRESERVED", params.model),
        makeUserHistoryTurn(`Reply with exactly CACHE-OK ${params.suffix}.`),
      ],
    },
    {
      apiKey: params.apiKey,
      cacheRetention: params.cacheRetention,
      sessionId: params.sessionId,
      maxTokens: 64,
      temperature: 0,
    },
    `anthropic image cache probe ${params.suffix} (${params.cacheRetention})`,
    ANTHROPIC_TIMEOUT_MS,
  );
  const text = extractAssistantText(response);
  expect(text.toLowerCase()).toContain(params.suffix.toLowerCase());
  return {
    suffix: params.suffix,
    text,
    usage: response.usage,
    hitRate: computeCacheHitRate(response.usage),
  };
}

describeCacheLive("pi embedded runner prompt caching (live)", () => {
  beforeAll(async () => {
    liveRunnerRootDir = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-live-cache-"));
    liveCacheTraceFile = path.join(liveRunnerRootDir, "cache-trace.jsonl");
    liveTestPngBase64 = (await fs.readFile(LIVE_TEST_PNG_URL)).toString("base64");
    previousCacheTraceEnv = {
      enabled: process.env.OPENCLAW_CACHE_TRACE,
      file: process.env.OPENCLAW_CACHE_TRACE_FILE,
      messages: process.env.OPENCLAW_CACHE_TRACE_MESSAGES,
      prompt: process.env.OPENCLAW_CACHE_TRACE_PROMPT,
      system: process.env.OPENCLAW_CACHE_TRACE_SYSTEM,
    };
    process.env.OPENCLAW_CACHE_TRACE = "1";
    process.env.OPENCLAW_CACHE_TRACE_FILE = liveCacheTraceFile;
    process.env.OPENCLAW_CACHE_TRACE_MESSAGES = "0";
    process.env.OPENCLAW_CACHE_TRACE_PROMPT = "0";
    process.env.OPENCLAW_CACHE_TRACE_SYSTEM = "0";
  }, 120_000);

  afterAll(async () => {
    if (previousCacheTraceEnv) {
      const restore = (
        key:
          | "OPENCLAW_CACHE_TRACE"
          | "OPENCLAW_CACHE_TRACE_FILE"
          | "OPENCLAW_CACHE_TRACE_MESSAGES"
          | "OPENCLAW_CACHE_TRACE_PROMPT"
          | "OPENCLAW_CACHE_TRACE_SYSTEM",
        value: string | undefined,
      ) => {
        if (value === undefined) {
          delete process.env[key];
        } else {
          process.env[key] = value;
        }
      };
      restore("OPENCLAW_CACHE_TRACE", previousCacheTraceEnv.enabled);
      restore("OPENCLAW_CACHE_TRACE_FILE", previousCacheTraceEnv.file);
      restore("OPENCLAW_CACHE_TRACE_MESSAGES", previousCacheTraceEnv.messages);
      restore("OPENCLAW_CACHE_TRACE_PROMPT", previousCacheTraceEnv.prompt);
      restore("OPENCLAW_CACHE_TRACE_SYSTEM", previousCacheTraceEnv.system);
    }
    previousCacheTraceEnv = null;
    liveCacheTraceFile = undefined;
    if (liveRunnerRootDir) {
      await fs.rm(liveRunnerRootDir, { recursive: true, force: true });
    }
    liveRunnerRootDir = undefined;
  });

  describe("openai", () => {
    let fixture: Awaited<ReturnType<typeof resolveLiveDirectModel>>;

    beforeAll(async () => {
      fixture = await resolveLiveDirectModel({
        provider: "openai",
        api: "openai-responses",
        envVar: "OPENCLAW_LIVE_OPENAI_CACHE_MODEL",
        preferredModelIds: ["gpt-5.5", "gpt-5.4-mini", "gpt-5.4"],
      });
      logLiveCache(`openai model=${fixture.model.provider}/${fixture.model.id}`);
    }, 120_000);

    it(
      "hits the expected OpenAI cache plateau on repeated stable prefixes",
      async () => {
        const warmup = await runOpenAiCacheProbe({
          ...fixture,
          sessionId: OPENAI_SESSION_ID,
          suffix: "warmup",
        });
        logLiveCache(
          `openai warmup cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );

        const hitRuns = [
          await runOpenAiCacheProbe({
            ...fixture,
            sessionId: OPENAI_SESSION_ID,
            suffix: "hit-a",
          }),
          await runOpenAiCacheProbe({
            ...fixture,
            sessionId: OPENAI_SESSION_ID,
            suffix: "hit-b",
          }),
        ];

        const bestHit = hitRuns.reduce((best, candidate) =>
          (candidate.usage.cacheRead ?? 0) > (best.usage.cacheRead ?? 0) ? candidate : best,
        );
        logLiveCache(
          `openai stable-prefix plateau suffix=${bestHit.suffix} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThanOrEqual(
          OPENAI_STABLE_PREFIX_MIN_CACHE_READ,
        );
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(OPENAI_STABLE_PREFIX_MIN_HIT_RATE);
      },
      6 * 60_000,
    );

    it(
      "keeps the expected OpenAI cache plateau across tool-call followup turns",
      async () => {
        const warmup = await runOpenAiToolCacheProbe({
          ...fixture,
          sessionId: `${OPENAI_SESSION_ID}-tool`,
          suffix: "tool-warmup",
        });
        logLiveCache(
          `openai tool warmup cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );

        const hitA = await runOpenAiToolCacheProbe({
          ...fixture,
          sessionId: `${OPENAI_SESSION_ID}-tool`,
          suffix: "tool-hit-a",
        });
        const hitB = await runOpenAiToolCacheProbe({
          ...fixture,
          sessionId: `${OPENAI_SESSION_ID}-tool`,
          suffix: "tool-hit-b",
        });
        const bestHit = (hitA.usage.cacheRead ?? 0) >= (hitB.usage.cacheRead ?? 0) ? hitA : hitB;
        logLiveCache(
          `openai tool plateau suffix=${bestHit.suffix} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThanOrEqual(OPENAI_TOOL_MIN_CACHE_READ);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(OPENAI_TOOL_MIN_HIT_RATE);
      },
      8 * 60_000,
    );

    it(
      "keeps the expected OpenAI cache plateau across image-heavy followup turns",
      async () => {
        const warmup = await runOpenAiImageCacheProbe({
          ...fixture,
          sessionId: `${OPENAI_SESSION_ID}-image`,
          suffix: "image-warmup",
        });
        logLiveCache(
          `openai image warmup cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );

        const hitA = await runOpenAiImageCacheProbe({
          ...fixture,
          sessionId: `${OPENAI_SESSION_ID}-image`,
          suffix: "image-hit-a",
        });
        const hitB = await runOpenAiImageCacheProbe({
          ...fixture,
          sessionId: `${OPENAI_SESSION_ID}-image`,
          suffix: "image-hit-b",
        });
        const bestHit = (hitA.usage.cacheRead ?? 0) >= (hitB.usage.cacheRead ?? 0) ? hitA : hitB;
        logLiveCache(
          `openai image plateau suffix=${bestHit.suffix} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThanOrEqual(OPENAI_IMAGE_MIN_CACHE_READ);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(OPENAI_IMAGE_MIN_HIT_RATE);
      },
      6 * 60_000,
    );

    it(
      "keeps high cache-read rates across repeated embedded-runner turns",
      async () => {
        const sessionId = `${OPENAI_SESSION_ID}-embedded`;
        const warmup = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: OPENAI_PREFIX,
          providerTag: "openai",
          sessionId,
          suffix: "embedded-warmup",
        });
        logLiveCache(
          `openai embedded warmup cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );

        const hitA = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: OPENAI_PREFIX,
          providerTag: "openai",
          sessionId,
          suffix: "embedded-hit-a",
        });
        const hitB = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: OPENAI_PREFIX,
          providerTag: "openai",
          sessionId,
          suffix: "embedded-hit-b",
        });
        const bestHit = (hitA.usage.cacheRead ?? 0) >= (hitB.usage.cacheRead ?? 0) ? hitA : hitB;
        logLiveCache(
          `openai embedded best-hit suffix=${bestHit.suffix} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThan(1_024);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(0.4);
        await expectCacheTraceStages(sessionId, ["cache:state", "cache:result"]);
      },
      8 * 60_000,
    );

    it(
      "keeps high cache-read rates when the same embedded session flips from websocket to sse",
      async () => {
        const sessionId = `${OPENAI_SESSION_ID}-transport-flip`;
        const warmup = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: OPENAI_PREFIX,
          providerTag: "openai",
          sessionId,
          suffix: "ws-warmup",
          transport: "websocket",
        });
        logLiveCache(
          `openai transport warmup cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );

        const hitA = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: OPENAI_PREFIX,
          providerTag: "openai",
          sessionId,
          suffix: "sse-hit-a",
          transport: "sse",
        });
        const hitB = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: OPENAI_PREFIX,
          providerTag: "openai",
          sessionId,
          suffix: "sse-hit-b",
          transport: "sse",
        });
        const bestHit = (hitA.usage.cacheRead ?? 0) >= (hitB.usage.cacheRead ?? 0) ? hitA : hitB;
        logLiveCache(
          `openai transport-flip best-hit suffix=${bestHit.suffix} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThan(1_024);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(0.35);
        await expectCacheTraceStages(sessionId, ["cache:state", "cache:result"]);
      },
      8 * 60_000,
    );

    it(
      "keeps cache reuse when structured system context only changes by whitespace and line endings",
      async () => {
        const sessionId = `${OPENAI_SESSION_ID}-structured-normalization`;
        const warmup = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: OPENAI_PREFIX,
          providerTag: "openai",
          sessionId,
          suffix: "structured-warmup",
        });
        logLiveCache(
          `openai structured warmup cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );

        const noisyPrefix = buildNoisyStructuredPromptVariant(OPENAI_PREFIX);
        const hitA = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: noisyPrefix,
          providerTag: "openai",
          sessionId,
          suffix: "structured-hit-a",
        });
        const hitB = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: noisyPrefix,
          providerTag: "openai",
          sessionId,
          suffix: "structured-hit-b",
        });
        const bestHit = (hitA.usage.cacheRead ?? 0) >= (hitB.usage.cacheRead ?? 0) ? hitA : hitB;
        logLiveCache(
          `openai structured best-hit suffix=${bestHit.suffix} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThan(1_024);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(0.35);
        await expectCacheTraceStages(sessionId, ["cache:state", "cache:result"]);
      },
      8 * 60_000,
    );
  });

  describe("anthropic", () => {
    let fixture: Awaited<ReturnType<typeof resolveLiveDirectModel>>;

    beforeAll(async () => {
      fixture = await resolveLiveDirectModel({
        provider: "anthropic",
        api: "anthropic-messages",
        envVar: "OPENCLAW_LIVE_ANTHROPIC_CACHE_MODEL",
        preferredModelIds: ["claude-sonnet-4-6", "claude-sonnet-4-6", "claude-haiku-3-5"],
      });
      logLiveCache(`anthropic model=${fixture.model.provider}/${fixture.model.id}`);
    }, 120_000);

    it(
      "writes cache on warmup and reads it back on repeated stable prefixes",
      async () => {
        const warmup = await runAnthropicCacheProbe({
          ...fixture,
          sessionId: ANTHROPIC_SESSION_ID,
          suffix: "warmup",
          cacheRetention: "short",
        });
        logLiveCache(
          `anthropic warmup cacheWrite=${warmup.usage.cacheWrite} cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );
        expect(warmup.usage.cacheWrite ?? 0).toBeGreaterThan(0);

        const hitRuns = [
          await runAnthropicCacheProbe({
            ...fixture,
            sessionId: ANTHROPIC_SESSION_ID,
            suffix: "hit-a",
            cacheRetention: "short",
          }),
          await runAnthropicCacheProbe({
            ...fixture,
            sessionId: ANTHROPIC_SESSION_ID,
            suffix: "hit-b",
            cacheRetention: "short",
          }),
        ];

        const bestHit = hitRuns.reduce((best, candidate) =>
          (candidate.usage.cacheRead ?? 0) > (best.usage.cacheRead ?? 0) ? candidate : best,
        );
        logLiveCache(
          `anthropic best-hit suffix=${bestHit.suffix} cacheWrite=${bestHit.usage.cacheWrite} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThan(1_024);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(0.7);
      },
      6 * 60_000,
    );

    it(
      "keeps high cache-read rates across tool-call followup turns",
      async () => {
        const warmup = await runAnthropicToolCacheProbe({
          ...fixture,
          sessionId: `${ANTHROPIC_SESSION_ID}-tool`,
          suffix: "tool-warmup",
          cacheRetention: "short",
        });
        logLiveCache(
          `anthropic tool warmup cacheWrite=${warmup.usage.cacheWrite} cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );
        expect(warmup.usage.cacheWrite ?? 0).toBeGreaterThan(0);

        const hitA = await runAnthropicToolCacheProbe({
          ...fixture,
          sessionId: `${ANTHROPIC_SESSION_ID}-tool`,
          suffix: "tool-hit-a",
          cacheRetention: "short",
        });
        const hitB = await runAnthropicToolCacheProbe({
          ...fixture,
          sessionId: `${ANTHROPIC_SESSION_ID}-tool`,
          suffix: "tool-hit-b",
          cacheRetention: "short",
        });
        const bestHit = (hitA.usage.cacheRead ?? 0) >= (hitB.usage.cacheRead ?? 0) ? hitA : hitB;
        logLiveCache(
          `anthropic tool best-hit suffix=${bestHit.suffix} cacheWrite=${bestHit.usage.cacheWrite} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThan(1_024);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(0.7);
      },
      8 * 60_000,
    );

    it(
      "keeps high cache-read rates across image-heavy followup turns",
      async () => {
        const warmup = await runAnthropicImageCacheProbe({
          ...fixture,
          sessionId: `${ANTHROPIC_SESSION_ID}-image`,
          suffix: "image-warmup",
          cacheRetention: "short",
        });
        logLiveCache(
          `anthropic image warmup cacheWrite=${warmup.usage.cacheWrite} cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );

        const hitA = await runAnthropicImageCacheProbe({
          ...fixture,
          sessionId: `${ANTHROPIC_SESSION_ID}-image`,
          suffix: "image-hit-a",
          cacheRetention: "short",
        });
        const hitB = await runAnthropicImageCacheProbe({
          ...fixture,
          sessionId: `${ANTHROPIC_SESSION_ID}-image`,
          suffix: "image-hit-b",
          cacheRetention: "short",
        });
        const bestHit = (hitA.usage.cacheRead ?? 0) >= (hitB.usage.cacheRead ?? 0) ? hitA : hitB;
        logLiveCache(
          `anthropic image best-hit suffix=${bestHit.suffix} cacheWrite=${bestHit.usage.cacheWrite} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThan(1_024);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(0.6);
      },
      6 * 60_000,
    );

    it(
      "does not report meaningful cache activity when retention is disabled",
      async () => {
        const disabled = await runAnthropicCacheProbe({
          ...fixture,
          sessionId: `${ANTHROPIC_SESSION_ID}-disabled`,
          suffix: "no-cache",
          cacheRetention: "none",
        });
        logLiveCache(
          `anthropic none cacheWrite=${disabled.usage.cacheWrite} cacheRead=${disabled.usage.cacheRead} input=${disabled.usage.input}`,
        );

        expect(disabled.usage.cacheRead ?? 0).toBeLessThanOrEqual(32);
        expect(disabled.usage.cacheWrite ?? 0).toBeLessThanOrEqual(32);
      },
      3 * 60_000,
    );

    it(
      "keeps high cache-read rates across repeated embedded-runner turns",
      async () => {
        const sessionId = `${ANTHROPIC_SESSION_ID}-embedded`;
        const warmup = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: ANTHROPIC_PREFIX,
          providerTag: "anthropic",
          sessionId,
          suffix: "embedded-warmup",
        });
        logLiveCache(
          `anthropic embedded warmup cacheWrite=${warmup.usage.cacheWrite} cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );
        expect(warmup.usage.cacheWrite ?? 0).toBeGreaterThan(0);

        const hitA = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: ANTHROPIC_PREFIX,
          providerTag: "anthropic",
          sessionId,
          suffix: "embedded-hit-a",
        });
        const hitB = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: ANTHROPIC_PREFIX,
          providerTag: "anthropic",
          sessionId,
          suffix: "embedded-hit-b",
        });
        const bestHit = (hitA.usage.cacheRead ?? 0) >= (hitB.usage.cacheRead ?? 0) ? hitA : hitB;
        logLiveCache(
          `anthropic embedded best-hit suffix=${bestHit.suffix} cacheWrite=${bestHit.usage.cacheWrite} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThan(1_024);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(0.4);
        await expectCacheTraceStages(sessionId, ["cache:state", "cache:result"]);
      },
      8 * 60_000,
    );

    it(
      "preserves cache-safe shaping across compaction followup turns",
      async () => {
        const sessionId = `${ANTHROPIC_SESSION_ID}-compaction`;
        await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: ANTHROPIC_PREFIX,
          providerTag: "anthropic",
          sessionId,
          suffix: "compact-prime-a",
          promptSections: 96,
        });
        await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: ANTHROPIC_PREFIX,
          providerTag: "anthropic",
          sessionId,
          suffix: "compact-prime-b",
          promptSections: 96,
        });

        const compacted = await compactLiveCacheSession({
          ...fixture,
          cacheRetention: "short",
          providerTag: "anthropic",
          sessionId,
        });
        logLiveCache(
          `anthropic compaction ok=${compacted.ok} compacted=${compacted.compacted} reason=${compacted.reason ?? "none"}`,
        );
        expect(compacted.ok).toBe(true);
        expect(compacted.compacted).toBe(true);

        const followup = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: ANTHROPIC_PREFIX,
          providerTag: "anthropic",
          sessionId,
          suffix: "compact-hit",
        });
        logLiveCache(
          `anthropic compaction followup cacheWrite=${followup.usage.cacheWrite} cacheRead=${followup.usage.cacheRead} input=${followup.usage.input} rate=${followup.hitRate.toFixed(3)}`,
        );

        expect(followup.usage.cacheRead ?? 0).toBeGreaterThan(1_024);
        expect(followup.hitRate).toBeGreaterThanOrEqual(0.3);
        await expectCacheTraceStages(sessionId, ["cache:state", "cache:result"]);
      },
      10 * 60_000,
    );

    it(
      "keeps cache reuse when structured system context only changes by whitespace and line endings",
      async () => {
        const sessionId = `${ANTHROPIC_SESSION_ID}-structured-normalization`;
        const warmup = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: ANTHROPIC_PREFIX,
          providerTag: "anthropic",
          sessionId,
          suffix: "structured-warmup",
        });
        logLiveCache(
          `anthropic structured warmup cacheWrite=${warmup.usage.cacheWrite} cacheRead=${warmup.usage.cacheRead} input=${warmup.usage.input} rate=${warmup.hitRate.toFixed(3)}`,
        );
        expect(warmup.usage.cacheWrite ?? 0).toBeGreaterThan(0);

        const noisyPrefix = buildNoisyStructuredPromptVariant(ANTHROPIC_PREFIX);
        const hitA = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: noisyPrefix,
          providerTag: "anthropic",
          sessionId,
          suffix: "structured-hit-a",
        });
        const hitB = await runEmbeddedCacheProbe({
          ...fixture,
          cacheRetention: "short",
          prefix: noisyPrefix,
          providerTag: "anthropic",
          sessionId,
          suffix: "structured-hit-b",
        });
        const bestHit = (hitA.usage.cacheRead ?? 0) >= (hitB.usage.cacheRead ?? 0) ? hitA : hitB;
        logLiveCache(
          `anthropic structured best-hit suffix=${bestHit.suffix} cacheWrite=${bestHit.usage.cacheWrite} cacheRead=${bestHit.usage.cacheRead} input=${bestHit.usage.input} rate=${bestHit.hitRate.toFixed(3)}`,
        );

        expect(bestHit.usage.cacheRead ?? 0).toBeGreaterThan(1_024);
        expect(bestHit.hitRate).toBeGreaterThanOrEqual(0.35);
        await expectCacheTraceStages(sessionId, ["cache:state", "cache:result"]);
      },
      8 * 60_000,
    );
  });
});