@capagents/forgeqe 1.1.2 → 1.1.4

package/src/monitor/run.ts

@@ -3,7 +3,7 @@ import path from "node:path"
 
 import type { Card } from "../card.ts"
 import { loadCases, type TestCase } from "../cases/index.ts"
-import type { AppConfig } from "../config.ts"
+import type { AppConfig, BrowserProvider } from "../config.ts"
 import { counts, suiteUsage } from "../suite/report.ts"
 import { runSuite, type SuiteEvent, type SuiteOptions, type SuiteResult } from "../suite/run.ts"
 import { formatDuration, formatTokens } from "../tui/theme.ts"
@@ -28,6 +28,7 @@ export type MonitorOptions = {
   config: AppConfig
   /** Test case sources; default monitor.checks. */
   checks: string[]
+  browserProvider?: BrowserProvider
   headed?: boolean
   url?: string
   /** Browser skills for every check (empty: none); unset: the engine picks once per check for the monitor's lifetime. */
@@ -67,6 +68,7 @@ export async function monitorRound(opts: MonitorOptions): Promise<RoundResult> {
     suite = await runSuite({
       config: { ...config, suite: { ...config.suite, outDir: m.outDir, code: false } },
       cases,
+      browserProvider: opts.browserProvider,
       parallel: Math.min(m.parallel, cases.length),
       writeBack: m.writeBack,
       replay: m.replay,

package/src/run.ts

@@ -5,7 +5,7 @@ import path from "node:path"
 import { assertGoalUrlsAllowed, hostAllowed, normalizeHttpUrl } from "./guard.ts"
 import { missingKey } from "./llm.ts"
 import type { AgentEvent, ReplayMode, RunResult } from "./types.ts"
-import type { AppConfig } from "./config.ts"
+import { browserConfigFor, type AppConfig, type BrowserProvider } from "./config.ts"
 import { runEngine } from "./engine/bridge.ts"
 import { MOCK_ENGINE } from "./engine/mock.ts"
 import { ensureEngine } from "./engine/setup.ts"
@@ -23,6 +23,7 @@ import { knowledgeBlock, readKnowledge, relevantKnowledge } from "./knowledge.ts
 export type ExecuteOptions = {
   config: AppConfig
   goal: string
+  browserProvider?: BrowserProvider
   url?: string
   headed?: boolean
   maxSteps?: number
@@ -148,6 +149,7 @@ export async function executeRun(opts: ExecuteOptions): Promise<{ result: RunRes
 
   const dir = opts.outDir || runDirectory(opts.config.export.outDir)
   const maxSteps = opts.maxSteps || opts.config.browser.maxSteps
+  const browser = browserConfigFor(opts.config, opts.browserProvider)
   const t0 = Date.now()
   const withReport = (out: { result: RunResult; files: string[] }) => {
     out.result.cost = costOf(out.result.usage, opts.config.llm.price)
@@ -179,6 +181,11 @@ export async function executeRun(opts: ExecuteOptions): Promise<{ result: RunRes
 
   const missing = missingKey(opts.config.llm)
   if (missing) throw new Error(missing)
+  // A remote grid's keys are checked here, before the engine starts, rather than failing inside it.
+  if (browser.remote) {
+    const unset = [browser.remote.usernameEnv, browser.remote.accessKeyEnv].filter((name) => !process.env[name]?.trim())
+    if (unset.length) throw new Error(`${browser.provider}: set ${unset.join(" and ")} in the environment, or save them with "forgeqe auth login ${browser.provider}".`)
+  }
   const log = (text: string) => opts.onEvent?.({ type: "log", level: "info", text })
   const python = await ensureEngine(log)
   // The browser agent's added instructions: forgeqe.yaml's, prompts/browser-extra.md, and the product knowledge that matches the goal.
@@ -216,7 +223,7 @@ export async function executeRun(opts: ExecuteOptions): Promise<{ result: RunRes
       headed: Boolean(opts.headed) || !opts.config.browser.headless,
       maxSteps,
       allowedHosts: allowed,
-      browser: { provider: opts.config.browser.provider, remote: opts.config.browser.remote },
+      browser,
       extendSystemMessage: extend || undefined,
       memory,
       env,

package/src/suite/run.ts

@@ -6,7 +6,7 @@ import { costOf } from "../cases/types.ts"
 import { addUsage, NO_USAGE } from "../usage.ts"
 import type { AgentEvent, CodeResult, ReplayMode, RunResult } from "../types.ts"
 import { codeSkipReason, runCodeAgent } from "../code/index.ts"
-import type { AppConfig } from "../config.ts"
+import type { AppConfig, BrowserProvider } from "../config.ts"
 import { fileBugs } from "../cases/bugs.ts"
 import { classifyFailure, FAILURE_LABEL } from "../cases/failure.ts"
 import { associateAdoAutomation, readMappings } from "../code/traceability.ts"
@@ -43,6 +43,7 @@ export type SuiteOptions = {
   cases: TestCase[]
   /** Repo for the coding agent; unset: browser runs only. */
   repo?: string
+  browserProvider?: BrowserProvider
   parallel?: number
   outDir?: string
   writeBack?: boolean
@@ -104,6 +105,7 @@ export async function runSuite(opts: SuiteOptions): Promise<SuiteResult> {
         const { result } = await execute({
           config,
           goal: goalFor(tc),
+          browserProvider: opts.browserProvider,
           url: tc.startUrl || opts.url,
           headed: opts.headed,
           maxSteps: opts.maxSteps,

package/testmu.example.yaml

@@ -0,0 +1,13 @@
+provider: testmu
+remote:
+  browser: Chrome
+  browser_version: latest
+  platform: Windows 11
+  username_env: LT_USERNAME
+  access_key_env: LT_ACCESS_KEY
+  project: ForgeQE
+  build: nightly
+  name: checkout smoke
+  local: false
+  extra:                 # merged into LT:Options, so keys go straight here
+    console: true