acb523de86
Add shared parseBooleanValue()/isTruthyEnvValue() and apply across CLI, gateway, memory, and live-test flags for consistent env handling. Introduce route-first fast paths, lazy subcommand registration, and deferred plugin loading to reduce CLI startup overhead. Centralize config validation via ensureConfigReady() and add config caching/deferred shell env fallback for fewer IO passes. Harden logger initialization/imports and add focused tests for argv, boolean parsing, frontmatter, and CLI subcommands.
1063 lines
39 KiB
TypeScript
1063 lines
39 KiB
TypeScript
import { randomBytes, randomUUID } from "node:crypto";
|
|
import fs from "node:fs/promises";
|
|
import { createServer } from "node:net";
|
|
import os from "node:os";
|
|
import path from "node:path";
|
|
|
|
import type { Api, Model } from "@mariozechner/pi-ai";
|
|
import { discoverAuthStorage, discoverModels } from "@mariozechner/pi-coding-agent";
|
|
import { describe, it } from "vitest";
|
|
import { resolveClawdbotAgentDir } from "../agents/agent-paths.js";
|
|
import { resolveAgentWorkspaceDir } from "../agents/agent-scope.js";
|
|
import {
|
|
type AuthProfileStore,
|
|
ensureAuthProfileStore,
|
|
saveAuthProfileStore,
|
|
} from "../agents/auth-profiles.js";
|
|
import {
|
|
collectAnthropicApiKeys,
|
|
isAnthropicBillingError,
|
|
isAnthropicRateLimitError,
|
|
} from "../agents/live-auth-keys.js";
|
|
import { isModernModelRef } from "../agents/live-model-filter.js";
|
|
import { getApiKeyForModel } from "../agents/model-auth.js";
|
|
import { ensureClawdbotModelsJson } from "../agents/models-config.js";
|
|
import { loadConfig } from "../config/config.js";
|
|
import type { ClawdbotConfig, ModelProviderConfig } from "../config/types.js";
|
|
import { isTruthyEnvValue } from "../infra/env.js";
|
|
import { DEFAULT_AGENT_ID } from "../routing/session-key.js";
|
|
import { GATEWAY_CLIENT_MODES, GATEWAY_CLIENT_NAMES } from "../utils/message-channel.js";
|
|
import { GatewayClient } from "./client.js";
|
|
import { renderCatNoncePngBase64 } from "./live-image-probe.js";
|
|
import { startGatewayServer } from "./server.js";
|
|
|
|
const LIVE =
|
|
isTruthyEnvValue(process.env.LIVE) || isTruthyEnvValue(process.env.CLAWDBOT_LIVE_TEST);
|
|
const GATEWAY_LIVE = isTruthyEnvValue(process.env.CLAWDBOT_LIVE_GATEWAY);
|
|
const ZAI_FALLBACK = isTruthyEnvValue(process.env.CLAWDBOT_LIVE_GATEWAY_ZAI_FALLBACK);
|
|
const PROVIDERS = parseFilter(process.env.CLAWDBOT_LIVE_GATEWAY_PROVIDERS);
|
|
const THINKING_LEVEL = "high";
|
|
const THINKING_TAG_RE = /<\s*\/?\s*(?:think(?:ing)?|thought|antthinking)\s*>/i;
|
|
const FINAL_TAG_RE = /<\s*\/?\s*final\s*>/i;
|
|
|
|
const describeLive = LIVE || GATEWAY_LIVE ? describe : describe.skip;
|
|
|
|
function parseFilter(raw?: string): Set<string> | null {
|
|
const trimmed = raw?.trim();
|
|
if (!trimmed || trimmed === "all") return null;
|
|
const ids = trimmed
|
|
.split(",")
|
|
.map((s) => s.trim())
|
|
.filter(Boolean);
|
|
return ids.length ? new Set(ids) : null;
|
|
}
|
|
|
|
function logProgress(message: string): void {
|
|
console.log(`[live] ${message}`);
|
|
}
|
|
|
|
function assertNoReasoningTags(params: {
|
|
text: string;
|
|
model: string;
|
|
phase: string;
|
|
label: string;
|
|
}): void {
|
|
if (!params.text) return;
|
|
if (THINKING_TAG_RE.test(params.text) || FINAL_TAG_RE.test(params.text)) {
|
|
const snippet = params.text.length > 200 ? `${params.text.slice(0, 200)}…` : params.text;
|
|
throw new Error(
|
|
`[${params.label}] reasoning tag leak (${params.model} / ${params.phase}): ${snippet}`,
|
|
);
|
|
}
|
|
}
|
|
|
|
function extractPayloadText(result: unknown): string {
|
|
const record = result as Record<string, unknown>;
|
|
const payloads = Array.isArray(record.payloads) ? record.payloads : [];
|
|
const texts = payloads
|
|
.map((p) => (p && typeof p === "object" ? (p as Record<string, unknown>).text : undefined))
|
|
.filter((t): t is string => typeof t === "string" && t.trim().length > 0);
|
|
return texts.join("\n").trim();
|
|
}
|
|
|
|
function isMeaningful(text: string): boolean {
|
|
if (!text) return false;
|
|
const trimmed = text.trim();
|
|
if (trimmed.toLowerCase() === "ok") return false;
|
|
if (trimmed.length < 60) return false;
|
|
const words = trimmed.split(/\s+/g).filter(Boolean);
|
|
if (words.length < 12) return false;
|
|
return true;
|
|
}
|
|
|
|
function isGoogleModelNotFoundText(text: string): boolean {
|
|
const trimmed = text.trim();
|
|
if (!trimmed) return false;
|
|
if (!/not found/i.test(trimmed)) return false;
|
|
if (/models\/.+ is not found for api version/i.test(trimmed)) return true;
|
|
if (/"status"\s*:\s*"NOT_FOUND"/.test(trimmed)) return true;
|
|
if (/"code"\s*:\s*404/.test(trimmed)) return true;
|
|
return false;
|
|
}
|
|
|
|
function isGoogleishProvider(provider: string): boolean {
|
|
return provider === "google" || provider.startsWith("google-");
|
|
}
|
|
|
|
function isRefreshTokenReused(error: string): boolean {
|
|
return /refresh_token_reused/i.test(error);
|
|
}
|
|
|
|
function isChatGPTUsageLimitErrorMessage(raw: string): boolean {
|
|
const msg = raw.toLowerCase();
|
|
return msg.includes("hit your chatgpt usage limit") && msg.includes("try again in");
|
|
}
|
|
|
|
function isMissingProfileError(error: string): boolean {
|
|
return /no credentials found for profile/i.test(error);
|
|
}
|
|
|
|
function isEmptyStreamText(text: string): boolean {
|
|
return text.includes("request ended without sending any chunks");
|
|
}
|
|
|
|
function randomImageProbeCode(len = 6): string {
|
|
// Chosen to avoid common OCR confusions in our 5x7 bitmap font.
|
|
// Notably: 0↔8, B↔8, 6↔9, 3↔B, D↔0.
|
|
// Must stay within the glyph set in `src/gateway/live-image-probe.ts`.
|
|
const alphabet = "24567ACEF";
|
|
const bytes = randomBytes(len);
|
|
let out = "";
|
|
for (let i = 0; i < len; i += 1) {
|
|
out += alphabet[bytes[i] % alphabet.length];
|
|
}
|
|
return out;
|
|
}
|
|
|
|
function editDistance(a: string, b: string): number {
|
|
if (a === b) return 0;
|
|
const aLen = a.length;
|
|
const bLen = b.length;
|
|
if (aLen === 0) return bLen;
|
|
if (bLen === 0) return aLen;
|
|
|
|
let prev = Array.from({ length: bLen + 1 }, (_v, idx) => idx);
|
|
let curr = Array.from({ length: bLen + 1 }, () => 0);
|
|
|
|
for (let i = 1; i <= aLen; i += 1) {
|
|
curr[0] = i;
|
|
const aCh = a.charCodeAt(i - 1);
|
|
for (let j = 1; j <= bLen; j += 1) {
|
|
const cost = aCh === b.charCodeAt(j - 1) ? 0 : 1;
|
|
curr[j] = Math.min(
|
|
prev[j] + 1, // delete
|
|
curr[j - 1] + 1, // insert
|
|
prev[j - 1] + cost, // substitute
|
|
);
|
|
}
|
|
[prev, curr] = [curr, prev];
|
|
}
|
|
|
|
return prev[bLen] ?? Number.POSITIVE_INFINITY;
|
|
}
|
|
async function getFreePort(): Promise<number> {
|
|
return await new Promise((resolve, reject) => {
|
|
const srv = createServer();
|
|
srv.on("error", reject);
|
|
srv.listen(0, "127.0.0.1", () => {
|
|
const addr = srv.address();
|
|
if (!addr || typeof addr === "string") {
|
|
srv.close();
|
|
reject(new Error("failed to acquire free port"));
|
|
return;
|
|
}
|
|
const port = addr.port;
|
|
srv.close((err) => {
|
|
if (err) reject(err);
|
|
else resolve(port);
|
|
});
|
|
});
|
|
});
|
|
}
|
|
|
|
async function isPortFree(port: number): Promise<boolean> {
|
|
if (!Number.isFinite(port) || port <= 0 || port > 65535) return false;
|
|
return await new Promise((resolve) => {
|
|
const srv = createServer();
|
|
srv.once("error", () => resolve(false));
|
|
srv.listen(port, "127.0.0.1", () => {
|
|
srv.close(() => resolve(true));
|
|
});
|
|
});
|
|
}
|
|
|
|
async function getFreeGatewayPort(): Promise<number> {
|
|
// Gateway uses derived ports (bridge/browser/canvas). Avoid flaky collisions by
|
|
// ensuring the common derived offsets are free too.
|
|
for (let attempt = 0; attempt < 25; attempt += 1) {
|
|
const port = await getFreePort();
|
|
const candidates = [port, port + 1, port + 2, port + 4];
|
|
const ok = (await Promise.all(candidates.map((candidate) => isPortFree(candidate)))).every(
|
|
Boolean,
|
|
);
|
|
if (ok) return port;
|
|
}
|
|
throw new Error("failed to acquire a free gateway port block");
|
|
}
|
|
|
|
type AgentFinalPayload = {
|
|
status?: unknown;
|
|
result?: unknown;
|
|
};
|
|
|
|
async function connectClient(params: { url: string; token: string }) {
|
|
return await new Promise<GatewayClient>((resolve, reject) => {
|
|
let settled = false;
|
|
const stop = (err?: Error, client?: GatewayClient) => {
|
|
if (settled) return;
|
|
settled = true;
|
|
clearTimeout(timer);
|
|
if (err) reject(err);
|
|
else resolve(client as GatewayClient);
|
|
};
|
|
const client = new GatewayClient({
|
|
url: params.url,
|
|
token: params.token,
|
|
clientName: GATEWAY_CLIENT_NAMES.TEST,
|
|
clientDisplayName: "vitest-live",
|
|
clientVersion: "dev",
|
|
mode: GATEWAY_CLIENT_MODES.TEST,
|
|
onHelloOk: () => stop(undefined, client),
|
|
onConnectError: (err) => stop(err),
|
|
onClose: (code, reason) =>
|
|
stop(new Error(`gateway closed during connect (${code}): ${reason}`)),
|
|
});
|
|
const timer = setTimeout(() => stop(new Error("gateway connect timeout")), 10_000);
|
|
timer.unref();
|
|
client.start();
|
|
});
|
|
}
|
|
|
|
type GatewayModelSuiteParams = {
|
|
label: string;
|
|
cfg: ClawdbotConfig;
|
|
candidates: Array<Model<Api>>;
|
|
extraToolProbes: boolean;
|
|
extraImageProbes: boolean;
|
|
thinkingLevel: string;
|
|
providerOverrides?: Record<string, ModelProviderConfig>;
|
|
};
|
|
|
|
function buildLiveGatewayConfig(params: {
|
|
cfg: ClawdbotConfig;
|
|
candidates: Array<Model<Api>>;
|
|
providerOverrides?: Record<string, ModelProviderConfig>;
|
|
}): ClawdbotConfig {
|
|
const providerOverrides = params.providerOverrides ?? {};
|
|
const lmstudioProvider = params.cfg.models?.providers?.lmstudio;
|
|
const baseProviders = params.cfg.models?.providers ?? {};
|
|
const nextProviders = {
|
|
...baseProviders,
|
|
...(lmstudioProvider
|
|
? {
|
|
lmstudio: {
|
|
...lmstudioProvider,
|
|
api: "openai-completions",
|
|
},
|
|
}
|
|
: {}),
|
|
...providerOverrides,
|
|
};
|
|
const providers = Object.keys(nextProviders).length > 0 ? nextProviders : baseProviders;
|
|
return {
|
|
...params.cfg,
|
|
agents: {
|
|
...params.cfg.agents,
|
|
list: (params.cfg.agents?.list ?? []).map((entry) => ({
|
|
...entry,
|
|
sandbox: { mode: "off" },
|
|
})),
|
|
defaults: {
|
|
...params.cfg.agents?.defaults,
|
|
// Live tests should avoid Docker sandboxing so tool probes can
|
|
// operate on the temporary probe files we create in the host workspace.
|
|
sandbox: { mode: "off" },
|
|
models: Object.fromEntries(params.candidates.map((m) => [`${m.provider}/${m.id}`, {}])),
|
|
},
|
|
},
|
|
models:
|
|
Object.keys(providers).length > 0 ? { ...params.cfg.models, providers } : params.cfg.models,
|
|
};
|
|
}
|
|
|
|
function sanitizeAuthConfig(params: {
|
|
cfg: ClawdbotConfig;
|
|
agentDir: string;
|
|
}): ClawdbotConfig["auth"] | undefined {
|
|
const auth = params.cfg.auth;
|
|
if (!auth) return auth;
|
|
const store = ensureAuthProfileStore(params.agentDir, {
|
|
allowKeychainPrompt: false,
|
|
});
|
|
|
|
let profiles: NonNullable<ClawdbotConfig["auth"]>["profiles"] | undefined;
|
|
if (auth.profiles) {
|
|
profiles = {};
|
|
for (const [profileId, profile] of Object.entries(auth.profiles)) {
|
|
if (!store.profiles[profileId]) continue;
|
|
profiles[profileId] = profile;
|
|
}
|
|
if (Object.keys(profiles).length === 0) profiles = undefined;
|
|
}
|
|
|
|
let order: Record<string, string[]> | undefined;
|
|
if (auth.order) {
|
|
order = {};
|
|
for (const [provider, ids] of Object.entries(auth.order)) {
|
|
const filtered = ids.filter((id) => Boolean(store.profiles[id]));
|
|
if (filtered.length === 0) continue;
|
|
order[provider] = filtered;
|
|
}
|
|
if (Object.keys(order).length === 0) order = undefined;
|
|
}
|
|
|
|
if (!profiles && !order && !auth.cooldowns) return undefined;
|
|
return {
|
|
...auth,
|
|
profiles,
|
|
order,
|
|
};
|
|
}
|
|
|
|
function buildMinimaxProviderOverride(params: {
|
|
cfg: ClawdbotConfig;
|
|
api: "openai-completions" | "anthropic-messages";
|
|
baseUrl: string;
|
|
}): ModelProviderConfig | null {
|
|
const existing = params.cfg.models?.providers?.minimax;
|
|
if (!existing || !Array.isArray(existing.models) || existing.models.length === 0) return null;
|
|
return {
|
|
...existing,
|
|
api: params.api,
|
|
baseUrl: params.baseUrl,
|
|
};
|
|
}
|
|
|
|
async function runGatewayModelSuite(params: GatewayModelSuiteParams) {
|
|
const previous = {
|
|
configPath: process.env.CLAWDBOT_CONFIG_PATH,
|
|
token: process.env.CLAWDBOT_GATEWAY_TOKEN,
|
|
skipChannels: process.env.CLAWDBOT_SKIP_CHANNELS,
|
|
skipGmail: process.env.CLAWDBOT_SKIP_GMAIL_WATCHER,
|
|
skipCron: process.env.CLAWDBOT_SKIP_CRON,
|
|
skipCanvas: process.env.CLAWDBOT_SKIP_CANVAS_HOST,
|
|
agentDir: process.env.CLAWDBOT_AGENT_DIR,
|
|
piAgentDir: process.env.PI_CODING_AGENT_DIR,
|
|
stateDir: process.env.CLAWDBOT_STATE_DIR,
|
|
};
|
|
let tempAgentDir: string | undefined;
|
|
let tempStateDir: string | undefined;
|
|
|
|
process.env.CLAWDBOT_SKIP_CHANNELS = "1";
|
|
process.env.CLAWDBOT_SKIP_GMAIL_WATCHER = "1";
|
|
process.env.CLAWDBOT_SKIP_CRON = "1";
|
|
process.env.CLAWDBOT_SKIP_CANVAS_HOST = "1";
|
|
|
|
const token = `test-${randomUUID()}`;
|
|
process.env.CLAWDBOT_GATEWAY_TOKEN = token;
|
|
const agentId = "dev";
|
|
|
|
const hostAgentDir = resolveClawdbotAgentDir();
|
|
const hostStore = ensureAuthProfileStore(hostAgentDir, {
|
|
allowKeychainPrompt: false,
|
|
});
|
|
const sanitizedStore: AuthProfileStore = {
|
|
version: hostStore.version,
|
|
profiles: { ...hostStore.profiles },
|
|
// Keep selection state so the gateway picks the same known-good profiles
|
|
// as the host (important when some profiles are rate-limited/disabled).
|
|
order: hostStore.order ? { ...hostStore.order } : undefined,
|
|
lastGood: hostStore.lastGood ? { ...hostStore.lastGood } : undefined,
|
|
usageStats: hostStore.usageStats ? { ...hostStore.usageStats } : undefined,
|
|
};
|
|
tempStateDir = await fs.mkdtemp(path.join(os.tmpdir(), "clawdbot-live-state-"));
|
|
process.env.CLAWDBOT_STATE_DIR = tempStateDir;
|
|
tempAgentDir = path.join(tempStateDir, "agents", DEFAULT_AGENT_ID, "agent");
|
|
saveAuthProfileStore(sanitizedStore, tempAgentDir);
|
|
const tempSessionAgentDir = path.join(tempStateDir, "agents", agentId, "agent");
|
|
if (tempSessionAgentDir !== tempAgentDir) {
|
|
saveAuthProfileStore(sanitizedStore, tempSessionAgentDir);
|
|
}
|
|
process.env.CLAWDBOT_AGENT_DIR = tempAgentDir;
|
|
process.env.PI_CODING_AGENT_DIR = tempAgentDir;
|
|
|
|
const workspaceDir = resolveAgentWorkspaceDir(params.cfg, agentId);
|
|
await fs.mkdir(workspaceDir, { recursive: true });
|
|
const nonceA = randomUUID();
|
|
const nonceB = randomUUID();
|
|
const toolProbePath = path.join(workspaceDir, `.clawdbot-live-tool-probe.${nonceA}.txt`);
|
|
await fs.writeFile(toolProbePath, `nonceA=${nonceA}\nnonceB=${nonceB}\n`);
|
|
|
|
const agentDir = resolveClawdbotAgentDir();
|
|
const sanitizedCfg: ClawdbotConfig = {
|
|
...params.cfg,
|
|
auth: sanitizeAuthConfig({ cfg: params.cfg, agentDir }),
|
|
};
|
|
const nextCfg = buildLiveGatewayConfig({
|
|
cfg: sanitizedCfg,
|
|
candidates: params.candidates,
|
|
providerOverrides: params.providerOverrides,
|
|
});
|
|
const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), "clawdbot-live-"));
|
|
const tempConfigPath = path.join(tempDir, "clawdbot.json");
|
|
await fs.writeFile(tempConfigPath, `${JSON.stringify(nextCfg, null, 2)}\n`);
|
|
process.env.CLAWDBOT_CONFIG_PATH = tempConfigPath;
|
|
|
|
await ensureClawdbotModelsJson(nextCfg);
|
|
|
|
const port = await getFreeGatewayPort();
|
|
const server = await startGatewayServer(port, {
|
|
bind: "loopback",
|
|
auth: { mode: "token", token },
|
|
controlUiEnabled: false,
|
|
});
|
|
|
|
const client = await connectClient({
|
|
url: `ws://127.0.0.1:${port}`,
|
|
token,
|
|
});
|
|
|
|
try {
|
|
logProgress(
|
|
`[${params.label}] running ${params.candidates.length} models (thinking=${params.thinkingLevel})`,
|
|
);
|
|
const anthropicKeys = collectAnthropicApiKeys();
|
|
if (anthropicKeys.length > 0) {
|
|
process.env.ANTHROPIC_API_KEY = anthropicKeys[0];
|
|
logProgress(`[${params.label}] anthropic keys loaded: ${anthropicKeys.length}`);
|
|
}
|
|
const sessionKey = `agent:${agentId}:${params.label}`;
|
|
const failures: Array<{ model: string; error: string }> = [];
|
|
let skippedCount = 0;
|
|
const total = params.candidates.length;
|
|
|
|
for (const [index, model] of params.candidates.entries()) {
|
|
const modelKey = `${model.provider}/${model.id}`;
|
|
const progressLabel = `[${params.label}] ${index + 1}/${total} ${modelKey}`;
|
|
|
|
const attemptMax =
|
|
model.provider === "anthropic" && anthropicKeys.length > 0 ? anthropicKeys.length : 1;
|
|
|
|
for (let attempt = 0; attempt < attemptMax; attempt += 1) {
|
|
if (model.provider === "anthropic" && anthropicKeys.length > 0) {
|
|
process.env.ANTHROPIC_API_KEY = anthropicKeys[attempt];
|
|
}
|
|
try {
|
|
// Ensure session exists + override model for this run.
|
|
// Reset between models: avoids cross-provider transcript incompatibilities
|
|
// (notably OpenAI Responses requiring reasoning replay for function_call items).
|
|
await client.request<Record<string, unknown>>("sessions.reset", {
|
|
key: sessionKey,
|
|
});
|
|
await client.request<Record<string, unknown>>("sessions.patch", {
|
|
key: sessionKey,
|
|
model: modelKey,
|
|
});
|
|
|
|
logProgress(`${progressLabel}: prompt`);
|
|
const runId = randomUUID();
|
|
const payload = await client.request<AgentFinalPayload>(
|
|
"agent",
|
|
{
|
|
sessionKey,
|
|
idempotencyKey: `idem-${runId}`,
|
|
message:
|
|
"Explain in 2-3 sentences how the JavaScript event loop handles microtasks vs macrotasks. Must mention both words: microtask and macrotask.",
|
|
thinking: params.thinkingLevel,
|
|
deliver: false,
|
|
},
|
|
{ expectFinal: true },
|
|
);
|
|
|
|
if (payload?.status !== "ok") {
|
|
throw new Error(`agent status=${String(payload?.status)}`);
|
|
}
|
|
let text = extractPayloadText(payload?.result);
|
|
if (!text) {
|
|
logProgress(`${progressLabel}: empty response, retrying`);
|
|
const retry = await client.request<AgentFinalPayload>(
|
|
"agent",
|
|
{
|
|
sessionKey,
|
|
idempotencyKey: `idem-${randomUUID()}-retry`,
|
|
message:
|
|
"Explain in 2-3 sentences how the JavaScript event loop handles microtasks vs macrotasks. Must mention both words: microtask and macrotask.",
|
|
thinking: params.thinkingLevel,
|
|
deliver: false,
|
|
},
|
|
{ expectFinal: true },
|
|
);
|
|
if (retry?.status !== "ok") {
|
|
throw new Error(`agent status=${String(retry?.status)}`);
|
|
}
|
|
text = extractPayloadText(retry?.result);
|
|
}
|
|
if (!text && isGoogleishProvider(model.provider)) {
|
|
logProgress(`${progressLabel}: skip (google empty response)`);
|
|
break;
|
|
}
|
|
if (
|
|
isEmptyStreamText(text) &&
|
|
(model.provider === "minimax" || model.provider === "openai-codex")
|
|
) {
|
|
logProgress(`${progressLabel}: skip (${model.provider} empty response)`);
|
|
break;
|
|
}
|
|
if (isGoogleishProvider(model.provider) && isGoogleModelNotFoundText(text)) {
|
|
// Catalog drift: model IDs can disappear or become unavailable on the API.
|
|
// Treat as skip when scanning "all models" for Google.
|
|
logProgress(`${progressLabel}: skip (google model not found)`);
|
|
break;
|
|
}
|
|
assertNoReasoningTags({
|
|
text,
|
|
model: modelKey,
|
|
phase: "prompt",
|
|
label: params.label,
|
|
});
|
|
if (!isMeaningful(text)) {
|
|
if (isGoogleishProvider(model.provider) && /gemini/i.test(model.id)) {
|
|
logProgress(`${progressLabel}: skip (google not meaningful)`);
|
|
break;
|
|
}
|
|
throw new Error(`not meaningful: ${text}`);
|
|
}
|
|
if (!/\bmicro\s*-?\s*tasks?\b/i.test(text) || !/\bmacro\s*-?\s*tasks?\b/i.test(text)) {
|
|
throw new Error(`missing required keywords: ${text}`);
|
|
}
|
|
|
|
// Real tool invocation: force the agent to Read a local file and echo a nonce.
|
|
logProgress(`${progressLabel}: tool-read`);
|
|
const runIdTool = randomUUID();
|
|
const toolProbe = await client.request<AgentFinalPayload>(
|
|
"agent",
|
|
{
|
|
sessionKey,
|
|
idempotencyKey: `idem-${runIdTool}-tool`,
|
|
message:
|
|
"Clawdbot live tool probe (local, safe): " +
|
|
`use the tool named \`read\` (or \`Read\`) with JSON arguments {"path":"${toolProbePath}"}. ` +
|
|
"Then reply with the two nonce values you read (include both).",
|
|
thinking: params.thinkingLevel,
|
|
deliver: false,
|
|
},
|
|
{ expectFinal: true },
|
|
);
|
|
if (toolProbe?.status !== "ok") {
|
|
throw new Error(`tool probe failed: status=${String(toolProbe?.status)}`);
|
|
}
|
|
const toolText = extractPayloadText(toolProbe?.result);
|
|
if (
|
|
isEmptyStreamText(toolText) &&
|
|
(model.provider === "minimax" || model.provider === "openai-codex")
|
|
) {
|
|
logProgress(`${progressLabel}: skip (${model.provider} empty response)`);
|
|
break;
|
|
}
|
|
assertNoReasoningTags({
|
|
text: toolText,
|
|
model: modelKey,
|
|
phase: "tool-read",
|
|
label: params.label,
|
|
});
|
|
if (!toolText.includes(nonceA) || !toolText.includes(nonceB)) {
|
|
throw new Error(`tool probe missing nonce: ${toolText}`);
|
|
}
|
|
|
|
if (params.extraToolProbes) {
|
|
logProgress(`${progressLabel}: tool-exec`);
|
|
const nonceC = randomUUID();
|
|
const toolWritePath = path.join(tempDir, `write-${runIdTool}.txt`);
|
|
|
|
const execReadProbe = await client.request<AgentFinalPayload>(
|
|
"agent",
|
|
{
|
|
sessionKey,
|
|
idempotencyKey: `idem-${runIdTool}-exec-read`,
|
|
message:
|
|
"Clawdbot live tool probe (local, safe): " +
|
|
"use the tool named `exec` (or `Exec`) to run this command: " +
|
|
`mkdir -p "${tempDir}" && printf '%s' '${nonceC}' > "${toolWritePath}". ` +
|
|
`Then use the tool named \`read\` (or \`Read\`) with JSON arguments {"path":"${toolWritePath}"}. ` +
|
|
"Finally reply including the nonce text you read back.",
|
|
thinking: params.thinkingLevel,
|
|
deliver: false,
|
|
},
|
|
{ expectFinal: true },
|
|
);
|
|
if (execReadProbe?.status !== "ok") {
|
|
throw new Error(`exec+read probe failed: status=${String(execReadProbe?.status)}`);
|
|
}
|
|
const execReadText = extractPayloadText(execReadProbe?.result);
|
|
if (
|
|
isEmptyStreamText(execReadText) &&
|
|
(model.provider === "minimax" || model.provider === "openai-codex")
|
|
) {
|
|
logProgress(`${progressLabel}: skip (${model.provider} empty response)`);
|
|
break;
|
|
}
|
|
assertNoReasoningTags({
|
|
text: execReadText,
|
|
model: modelKey,
|
|
phase: "tool-exec",
|
|
label: params.label,
|
|
});
|
|
if (!execReadText.includes(nonceC)) {
|
|
throw new Error(`exec+read probe missing nonce: ${execReadText}`);
|
|
}
|
|
|
|
await fs.rm(toolWritePath, { force: true });
|
|
}
|
|
|
|
if (params.extraImageProbes && model.input?.includes("image")) {
|
|
logProgress(`${progressLabel}: image`);
|
|
// Shorter code => less OCR flake across providers, still tests image attachments end-to-end.
|
|
const imageCode = randomImageProbeCode();
|
|
const imageBase64 = renderCatNoncePngBase64(imageCode);
|
|
const runIdImage = randomUUID();
|
|
|
|
const imageProbe = await client.request<AgentFinalPayload>(
|
|
"agent",
|
|
{
|
|
sessionKey,
|
|
idempotencyKey: `idem-${runIdImage}-image`,
|
|
message:
|
|
"Look at the attached image. Reply with exactly two tokens separated by a single space: " +
|
|
"(1) the animal shown or written in the image, lowercase; " +
|
|
"(2) the code printed in the image, uppercase. No extra text.",
|
|
attachments: [
|
|
{
|
|
mimeType: "image/png",
|
|
fileName: `probe-${runIdImage}.png`,
|
|
content: imageBase64,
|
|
},
|
|
],
|
|
thinking: params.thinkingLevel,
|
|
deliver: false,
|
|
},
|
|
{ expectFinal: true },
|
|
);
|
|
// Best-effort: do not fail the whole live suite on flaky image handling.
|
|
// (We still keep prompt + tool probes as hard checks.)
|
|
if (imageProbe?.status !== "ok") {
|
|
logProgress(`${progressLabel}: image skip (status=${String(imageProbe?.status)})`);
|
|
} else {
|
|
const imageText = extractPayloadText(imageProbe?.result);
|
|
if (
|
|
isEmptyStreamText(imageText) &&
|
|
(model.provider === "minimax" || model.provider === "openai-codex")
|
|
) {
|
|
logProgress(`${progressLabel}: image skip (${model.provider} empty response)`);
|
|
} else {
|
|
assertNoReasoningTags({
|
|
text: imageText,
|
|
model: modelKey,
|
|
phase: "image",
|
|
label: params.label,
|
|
});
|
|
if (!/\bcat\b/i.test(imageText)) {
|
|
logProgress(`${progressLabel}: image skip (missing 'cat')`);
|
|
} else {
|
|
const candidates = imageText.toUpperCase().match(/[A-Z0-9]{6,20}/g) ?? [];
|
|
const bestDistance = candidates.reduce((best, cand) => {
|
|
if (Math.abs(cand.length - imageCode.length) > 2) return best;
|
|
return Math.min(best, editDistance(cand, imageCode));
|
|
}, Number.POSITIVE_INFINITY);
|
|
// OCR / image-read flake: allow a small edit distance, but still require the "cat" token above.
|
|
if (!(bestDistance <= 3)) {
|
|
logProgress(`${progressLabel}: image skip (code mismatch)`);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
// Regression: tool-call-only turn followed by a user message (OpenAI responses bug class).
|
|
if (
|
|
(model.provider === "openai" && model.api === "openai-responses") ||
|
|
(model.provider === "openai-codex" && model.api === "openai-codex-responses")
|
|
) {
|
|
logProgress(`${progressLabel}: tool-only regression`);
|
|
const runId2 = randomUUID();
|
|
const first = await client.request<AgentFinalPayload>(
|
|
"agent",
|
|
{
|
|
sessionKey,
|
|
idempotencyKey: `idem-${runId2}-1`,
|
|
message: `Call the tool named \`read\` (or \`Read\`) on "${toolProbePath}". Do not write any other text.`,
|
|
thinking: params.thinkingLevel,
|
|
deliver: false,
|
|
},
|
|
{ expectFinal: true },
|
|
);
|
|
if (first?.status !== "ok") {
|
|
throw new Error(`tool-only turn failed: status=${String(first?.status)}`);
|
|
}
|
|
const firstText = extractPayloadText(first?.result);
|
|
assertNoReasoningTags({
|
|
text: firstText,
|
|
model: modelKey,
|
|
phase: "tool-only",
|
|
label: params.label,
|
|
});
|
|
|
|
const second = await client.request<AgentFinalPayload>(
|
|
"agent",
|
|
{
|
|
sessionKey,
|
|
idempotencyKey: `idem-${runId2}-2`,
|
|
message: `Now answer: what are the values of nonceA and nonceB in "${toolProbePath}"? Reply with exactly: ${nonceA} ${nonceB}.`,
|
|
thinking: params.thinkingLevel,
|
|
deliver: false,
|
|
},
|
|
{ expectFinal: true },
|
|
);
|
|
if (second?.status !== "ok") {
|
|
throw new Error(`post-tool message failed: status=${String(second?.status)}`);
|
|
}
|
|
const reply = extractPayloadText(second?.result);
|
|
assertNoReasoningTags({
|
|
text: reply,
|
|
model: modelKey,
|
|
phase: "tool-only-followup",
|
|
label: params.label,
|
|
});
|
|
if (!reply.includes(nonceA) || !reply.includes(nonceB)) {
|
|
throw new Error(`unexpected reply: ${reply}`);
|
|
}
|
|
}
|
|
|
|
logProgress(`${progressLabel}: done`);
|
|
break;
|
|
} catch (err) {
|
|
const message = String(err);
|
|
if (
|
|
model.provider === "anthropic" &&
|
|
isAnthropicRateLimitError(message) &&
|
|
attempt + 1 < attemptMax
|
|
) {
|
|
logProgress(`${progressLabel}: rate limit, retrying with next key`);
|
|
continue;
|
|
}
|
|
if (model.provider === "anthropic" && isAnthropicBillingError(message)) {
|
|
if (attempt + 1 < attemptMax) {
|
|
logProgress(`${progressLabel}: billing issue, retrying with next key`);
|
|
continue;
|
|
}
|
|
logProgress(`${progressLabel}: skip (anthropic billing)`);
|
|
break;
|
|
}
|
|
if (
|
|
model.provider === "anthropic" &&
|
|
isEmptyStreamText(message) &&
|
|
attempt + 1 < attemptMax
|
|
) {
|
|
logProgress(`${progressLabel}: empty response, retrying with next key`);
|
|
continue;
|
|
}
|
|
if (model.provider === "anthropic" && isEmptyStreamText(message)) {
|
|
skippedCount += 1;
|
|
logProgress(`${progressLabel}: skip (anthropic empty response)`);
|
|
break;
|
|
}
|
|
// OpenAI Codex refresh tokens can become single-use; skip instead of failing all live tests.
|
|
if (model.provider === "openai-codex" && isRefreshTokenReused(message)) {
|
|
logProgress(`${progressLabel}: skip (codex refresh token reused)`);
|
|
break;
|
|
}
|
|
if (model.provider === "openai-codex" && isChatGPTUsageLimitErrorMessage(message)) {
|
|
logProgress(`${progressLabel}: skip (chatgpt usage limit)`);
|
|
break;
|
|
}
|
|
if (isMissingProfileError(message)) {
|
|
skippedCount += 1;
|
|
logProgress(`${progressLabel}: skip (missing auth profile)`);
|
|
break;
|
|
}
|
|
if (params.label.startsWith("minimax-")) {
|
|
skippedCount += 1;
|
|
logProgress(`${progressLabel}: skip (minimax endpoint error)`);
|
|
break;
|
|
}
|
|
logProgress(`${progressLabel}: failed`);
|
|
failures.push({ model: modelKey, error: message });
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
|
|
if (failures.length > 0) {
|
|
const preview = failures
|
|
.slice(0, 20)
|
|
.map((f) => `- ${f.model}: ${f.error}`)
|
|
.join("\n");
|
|
throw new Error(`gateway live model failures (${failures.length}):\n${preview}`);
|
|
}
|
|
if (skippedCount === total) {
|
|
logProgress(`[${params.label}] skipped all models (missing profiles)`);
|
|
}
|
|
} finally {
|
|
client.stop();
|
|
await server.close({ reason: "live test complete" });
|
|
await fs.rm(toolProbePath, { force: true });
|
|
await fs.rm(tempDir, { recursive: true, force: true });
|
|
if (tempAgentDir) {
|
|
await fs.rm(tempAgentDir, { recursive: true, force: true });
|
|
}
|
|
if (tempStateDir) {
|
|
await fs.rm(tempStateDir, { recursive: true, force: true });
|
|
}
|
|
|
|
process.env.CLAWDBOT_CONFIG_PATH = previous.configPath;
|
|
process.env.CLAWDBOT_GATEWAY_TOKEN = previous.token;
|
|
process.env.CLAWDBOT_SKIP_CHANNELS = previous.skipChannels;
|
|
process.env.CLAWDBOT_SKIP_GMAIL_WATCHER = previous.skipGmail;
|
|
process.env.CLAWDBOT_SKIP_CRON = previous.skipCron;
|
|
process.env.CLAWDBOT_SKIP_CANVAS_HOST = previous.skipCanvas;
|
|
process.env.CLAWDBOT_AGENT_DIR = previous.agentDir;
|
|
process.env.PI_CODING_AGENT_DIR = previous.piAgentDir;
|
|
process.env.CLAWDBOT_STATE_DIR = previous.stateDir;
|
|
}
|
|
}
|
|
|
|
describeLive("gateway live (dev agent, profile keys)", () => {
|
|
it(
|
|
"runs meaningful prompts across models with available keys",
|
|
async () => {
|
|
const cfg = loadConfig();
|
|
await ensureClawdbotModelsJson(cfg);
|
|
|
|
const agentDir = resolveClawdbotAgentDir();
|
|
const authStore = ensureAuthProfileStore(agentDir, {
|
|
allowKeychainPrompt: false,
|
|
});
|
|
const authStorage = discoverAuthStorage(agentDir);
|
|
const modelRegistry = discoverModels(authStorage, agentDir);
|
|
const all = modelRegistry.getAll() as Array<Model<Api>>;
|
|
|
|
const rawModels = process.env.CLAWDBOT_LIVE_GATEWAY_MODELS?.trim();
|
|
const useModern = !rawModels || rawModels === "modern" || rawModels === "all";
|
|
const useExplicit = Boolean(rawModels) && !useModern;
|
|
const filter = useExplicit ? parseFilter(rawModels) : null;
|
|
const wanted = filter
|
|
? all.filter((m) => filter.has(`${m.provider}/${m.id}`))
|
|
: all.filter((m) => isModernModelRef({ provider: m.provider, id: m.id }));
|
|
|
|
const candidates: Array<Model<Api>> = [];
|
|
for (const model of wanted) {
|
|
if (PROVIDERS && !PROVIDERS.has(model.provider)) continue;
|
|
try {
|
|
// eslint-disable-next-line no-await-in-loop
|
|
const apiKeyInfo = await getApiKeyForModel({
|
|
model,
|
|
cfg,
|
|
store: authStore,
|
|
agentDir,
|
|
});
|
|
if (!apiKeyInfo.source.startsWith("profile:")) {
|
|
continue;
|
|
}
|
|
candidates.push(model);
|
|
} catch {
|
|
// no creds; skip
|
|
}
|
|
}
|
|
|
|
if (candidates.length === 0) {
|
|
logProgress("[all-models] no API keys found; skipping");
|
|
return;
|
|
}
|
|
logProgress(`[all-models] selection=${useExplicit ? "explicit" : "modern"}`);
|
|
const imageCandidates = candidates.filter((m) => m.input?.includes("image"));
|
|
if (imageCandidates.length === 0) {
|
|
logProgress("[all-models] no image-capable models selected; image probe will be skipped");
|
|
}
|
|
await runGatewayModelSuite({
|
|
label: "all-models",
|
|
cfg,
|
|
candidates,
|
|
extraToolProbes: true,
|
|
extraImageProbes: true,
|
|
thinkingLevel: THINKING_LEVEL,
|
|
});
|
|
|
|
const minimaxCandidates = candidates.filter((model) => model.provider === "minimax");
|
|
if (minimaxCandidates.length === 0) {
|
|
logProgress("[minimax] no candidates with keys; skipping dual endpoint probes");
|
|
return;
|
|
}
|
|
|
|
const minimaxAnthropic = buildMinimaxProviderOverride({
|
|
cfg,
|
|
api: "anthropic-messages",
|
|
baseUrl: "https://api.minimax.io/anthropic",
|
|
});
|
|
if (minimaxAnthropic) {
|
|
await runGatewayModelSuite({
|
|
label: "minimax-anthropic",
|
|
cfg,
|
|
candidates: minimaxCandidates,
|
|
extraToolProbes: true,
|
|
extraImageProbes: true,
|
|
thinkingLevel: THINKING_LEVEL,
|
|
providerOverrides: { minimax: minimaxAnthropic },
|
|
});
|
|
} else {
|
|
logProgress("[minimax-anthropic] missing minimax provider config; skipping");
|
|
}
|
|
},
|
|
20 * 60 * 1000,
|
|
);
|
|
|
|
it("z.ai fallback handles anthropic tool history", async () => {
|
|
if (!ZAI_FALLBACK) return;
|
|
const previous = {
|
|
configPath: process.env.CLAWDBOT_CONFIG_PATH,
|
|
token: process.env.CLAWDBOT_GATEWAY_TOKEN,
|
|
skipChannels: process.env.CLAWDBOT_SKIP_CHANNELS,
|
|
skipGmail: process.env.CLAWDBOT_SKIP_GMAIL_WATCHER,
|
|
skipCron: process.env.CLAWDBOT_SKIP_CRON,
|
|
skipCanvas: process.env.CLAWDBOT_SKIP_CANVAS_HOST,
|
|
};
|
|
|
|
process.env.CLAWDBOT_SKIP_CHANNELS = "1";
|
|
process.env.CLAWDBOT_SKIP_GMAIL_WATCHER = "1";
|
|
process.env.CLAWDBOT_SKIP_CRON = "1";
|
|
process.env.CLAWDBOT_SKIP_CANVAS_HOST = "1";
|
|
|
|
const token = `test-${randomUUID()}`;
|
|
process.env.CLAWDBOT_GATEWAY_TOKEN = token;
|
|
|
|
const cfg = loadConfig();
|
|
await ensureClawdbotModelsJson(cfg);
|
|
|
|
const agentDir = resolveClawdbotAgentDir();
|
|
const authStorage = discoverAuthStorage(agentDir);
|
|
const modelRegistry = discoverModels(authStorage, agentDir);
|
|
const anthropic = modelRegistry.find("anthropic", "claude-opus-4-5") as Model<Api> | null;
|
|
const zai = modelRegistry.find("zai", "glm-4.7") as Model<Api> | null;
|
|
|
|
if (!anthropic || !zai) return;
|
|
try {
|
|
await getApiKeyForModel({ model: anthropic, cfg });
|
|
await getApiKeyForModel({ model: zai, cfg });
|
|
} catch {
|
|
return;
|
|
}
|
|
|
|
const agentId = "dev";
|
|
const workspaceDir = resolveAgentWorkspaceDir(cfg, agentId);
|
|
await fs.mkdir(workspaceDir, { recursive: true });
|
|
const nonceA = randomUUID();
|
|
const nonceB = randomUUID();
|
|
const toolProbePath = path.join(workspaceDir, `.clawdbot-live-zai-fallback.${nonceA}.txt`);
|
|
await fs.writeFile(toolProbePath, `nonceA=${nonceA}\nnonceB=${nonceB}\n`);
|
|
|
|
const port = await getFreeGatewayPort();
|
|
const server = await startGatewayServer({
|
|
configPath: cfg.__meta?.path,
|
|
port,
|
|
token,
|
|
});
|
|
|
|
const client = await connectClient({
|
|
url: `ws://127.0.0.1:${port}`,
|
|
token,
|
|
});
|
|
|
|
try {
|
|
const sessionKey = `agent:${agentId}:live-zai-fallback`;
|
|
|
|
await client.request<Record<string, unknown>>("sessions.patch", {
|
|
key: sessionKey,
|
|
model: "anthropic/claude-opus-4-5",
|
|
});
|
|
await client.request<Record<string, unknown>>("sessions.reset", {
|
|
key: sessionKey,
|
|
});
|
|
|
|
const runId = randomUUID();
|
|
const toolProbe = await client.request<AgentFinalPayload>(
|
|
"agent",
|
|
{
|
|
sessionKey,
|
|
idempotencyKey: `idem-${runId}-tool`,
|
|
message:
|
|
`Call the tool named \`read\` (or \`Read\` if \`read\` is unavailable) with JSON arguments {"path":"${toolProbePath}"}. ` +
|
|
`Then reply with exactly: ${nonceA} ${nonceB}. No extra text.`,
|
|
thinking: THINKING_LEVEL,
|
|
deliver: false,
|
|
},
|
|
{ expectFinal: true },
|
|
);
|
|
if (toolProbe?.status !== "ok") {
|
|
throw new Error(`anthropic tool probe failed: status=${String(toolProbe?.status)}`);
|
|
}
|
|
const toolText = extractPayloadText(toolProbe?.result);
|
|
assertNoReasoningTags({
|
|
text: toolText,
|
|
model: "anthropic/claude-opus-4-5",
|
|
phase: "zai-fallback-tool",
|
|
label: "zai-fallback",
|
|
});
|
|
if (!toolText.includes(nonceA) || !toolText.includes(nonceB)) {
|
|
throw new Error(`anthropic tool probe missing nonce: ${toolText}`);
|
|
}
|
|
|
|
await client.request<Record<string, unknown>>("sessions.patch", {
|
|
key: sessionKey,
|
|
model: "zai/glm-4.7",
|
|
});
|
|
|
|
const followupId = randomUUID();
|
|
const followup = await client.request<AgentFinalPayload>(
|
|
"agent",
|
|
{
|
|
sessionKey,
|
|
idempotencyKey: `idem-${followupId}-followup`,
|
|
message:
|
|
`What are the values of nonceA and nonceB in "${toolProbePath}"? ` +
|
|
`Reply with exactly: ${nonceA} ${nonceB}.`,
|
|
thinking: THINKING_LEVEL,
|
|
deliver: false,
|
|
},
|
|
{ expectFinal: true },
|
|
);
|
|
if (followup?.status !== "ok") {
|
|
throw new Error(`zai followup failed: status=${String(followup?.status)}`);
|
|
}
|
|
const followupText = extractPayloadText(followup?.result);
|
|
assertNoReasoningTags({
|
|
text: followupText,
|
|
model: "zai/glm-4.7",
|
|
phase: "zai-fallback-followup",
|
|
label: "zai-fallback",
|
|
});
|
|
if (!followupText.includes(nonceA) || !followupText.includes(nonceB)) {
|
|
throw new Error(`zai followup missing nonce: ${followupText}`);
|
|
}
|
|
} finally {
|
|
client.stop();
|
|
await server.close({ reason: "live test complete" });
|
|
await fs.rm(toolProbePath, { force: true });
|
|
|
|
process.env.CLAWDBOT_CONFIG_PATH = previous.configPath;
|
|
process.env.CLAWDBOT_GATEWAY_TOKEN = previous.token;
|
|
process.env.CLAWDBOT_SKIP_CHANNELS = previous.skipChannels;
|
|
process.env.CLAWDBOT_SKIP_GMAIL_WATCHER = previous.skipGmail;
|
|
process.env.CLAWDBOT_SKIP_CRON = previous.skipCron;
|
|
process.env.CLAWDBOT_SKIP_CANVAS_HOST = previous.skipCanvas;
|
|
}
|
|
}, 180_000);
|
|
});
|