agex 0.2.3 → 0.2.4

package/dist/{chunk-EN7TTESF.js → chunk-CY6NF2GO.js}

@@ -103,7 +103,7 @@ function formatMessage(level, prefix, message) {
 function createLogger(prefix = "", level) {
   const minLevel = level ?? getDefaultLevel();
   const minOrder = LEVEL_ORDER[minLevel];
-  function log12(logLevel, message, ...args) {
+  function log13(logLevel, message, ...args) {
     if (LEVEL_ORDER[logLevel] < minOrder)
       return;
     const formatted = formatMessage(logLevel, prefix, message);
@@ -122,10 +122,10 @@ function createLogger(prefix = "", level) {
     }
   }
   return {
-    debug: (message, ...args) => log12("debug", message, ...args),
-    info: (message, ...args) => log12("info", message, ...args),
-    warn: (message, ...args) => log12("warn", message, ...args),
-    error: (message, ...args) => log12("error", message, ...args),
+    debug: (message, ...args) => log13("debug", message, ...args),
+    info: (message, ...args) => log13("info", message, ...args),
+    warn: (message, ...args) => log13("warn", message, ...args),
+    error: (message, ...args) => log13("error", message, ...args),
     child: (childPrefix) => createLogger(prefix ? `${prefix}:${childPrefix}` : childPrefix, level)
   };
 }
@@ -636,8 +636,8 @@ function getErrorMap() {
 
 // ../../node_modules/.pnpm/zod@3.25.76/node_modules/zod/v3/helpers/parseUtil.js
 var makeIssue = (params) => {
-  const { data, path: path7, errorMaps, issueData } = params;
-  const fullPath = [...path7, ...issueData.path || []];
+  const { data, path: path8, errorMaps, issueData } = params;
+  const fullPath = [...path8, ...issueData.path || []];
   const fullIssue = {
     ...issueData,
     path: fullPath
@@ -753,11 +753,11 @@ var errorUtil;
 
 // ../../node_modules/.pnpm/zod@3.25.76/node_modules/zod/v3/types.js
 var ParseInputLazyPath = class {
-  constructor(parent, value, path7, key) {
+  constructor(parent, value, path8, key) {
     this._cachedPath = [];
     this.parent = parent;
     this.data = value;
-    this._path = path7;
+    this._path = path8;
     this._key = key;
   }
   get path() {
@@ -5200,11 +5200,11 @@ async function runWithStreaming({ commandSpec, request, agent, env, parser, time
       process.stderr.write(chunk);
     }
   });
-  const exitCode = await new Promise((resolve6) => {
+  const exitCode = await new Promise((resolve7) => {
     child.on("close", (code) => {
       if (timeoutHandle)
         clearTimeout(timeoutHandle);
-      resolve6(timedOut ? DEFAULTS.TIMEOUT_EXIT_CODE : code ?? 1);
+      resolve7(timedOut ? DEFAULTS.TIMEOUT_EXIT_CODE : code ?? 1);
     });
   });
   if (lineProcessor) {
@@ -6118,6 +6118,193 @@ var reviewRequestSchema = external_exports.object({
   timeoutMs: external_exports.number().positive().optional()
 });
 
+// ../prove-pr/dist/prove-pr.js
+import * as fs5 from "fs";
+import * as path7 from "path";
+
+// ../prove-pr/dist/prompt.js
+function buildProvePrPrompt(gitContext, hypothesesPlan, outputDir, url) {
+  const urlInstruction = url ? `
+Dev server URL: ${url}` : "";
+  return `You are proving that code changes work correctly by gathering factual evidence.
+
+GIT CONTEXT:
+${gitContext}
+
+HYPOTHESES TO PROVE:
+${hypothesesPlan}
+${urlInstruction}
+
+YOUR TASK:
+For each hypothesis (H1, H2, etc.), gather evidence to prove or disprove it.
+
+HELPER SCRIPTS:
+- ./ab-open <url>         # Open URL with viewport, recording, and effects
+- ./ab-screenshot <name>  # Save screenshot to output dir
+- ./ab-close              # Stop recording and close browser
+- ./ab-fx                 # Re-inject effects after navigation
+
+AGENT-BROWSER COMMANDS:
+- agent-browser snapshot -i          # Get interactive elements with refs (@e1, @e2)
+- agent-browser click @e1            # Click element
+- agent-browser fill @e2 "text"      # Fill input
+- agent-browser get text @e1         # Get element text
+- agent-browser eval "<js>"          # Run JavaScript
+- agent-browser wait @e1             # Wait for element
+
+EVIDENCE GATHERING:
+- For visual/UI changes: Use ./ab-open and ./ab-screenshot
+- For behavior changes: Run the code, check logs, execute commands
+- For API changes: Make requests and verify responses
+- Use whatever combination makes sense for each hypothesis
+
+INSTRUCTIONS:
+1. Work through each hypothesis systematically
+2. Take screenshots at key moments (use descriptive names like "H1-login-form.png")
+3. Run commands and capture output when relevant
+4. Document what you observe
+
+OUTPUT FORMAT:
+For each hypothesis, output:
+[H1_VERDICT:PASS|FAIL|SKIP] - Brief explanation
+Evidence: <list what you captured>
+
+After all hypotheses, output:
+[PROVE_PR_COMPLETE]
+Summary: X passed, Y failed, Z skipped`;
+}
+
+// ../prove-pr/dist/verdict-parser.js
+function parseHypothesesVerdicts(output) {
+  const verdicts = [];
+  const regex = /\[H(\d+)_VERDICT:(PASS|FAIL|SKIP)\]\s*-?\s*([^\n]*)/gi;
+  for (const match of output.matchAll(regex)) {
+    const id = `H${match[1]}`;
+    verdicts.push({
+      id,
+      hypothesis: "",
+      verdict: match[2].toLowerCase(),
+      reason: match[3]?.trim() || "",
+      evidence: []
+    });
+  }
+  return verdicts;
+}
+
+// ../prove-pr/dist/report-builder.js
+function generateReport(result) {
+  const lines = [
+    "# Prove PR Report",
+    "",
+    `**Base:** ${result.baseRef}`,
+    `**Duration:** ${(result.durationMs / 1e3).toFixed(1)}s`,
+    `**Results:** ${result.summary.pass} passed, ${result.summary.fail} failed, ${result.summary.skip} skipped`,
+    "",
+    "## Hypothesis Verdicts",
+    ""
+  ];
+  for (const h of result.hypotheses) {
+    const icon = h.verdict === "pass" ? "\u2705" : h.verdict === "fail" ? "\u274C" : "\u23ED\uFE0F";
+    lines.push(`### ${h.id} ${icon} ${h.verdict.toUpperCase()}`);
+    lines.push(`> ${h.reason}`);
+    lines.push("");
+  }
+  return lines.join("\n");
+}
+
+// ../prove-pr/dist/prove-pr.js
+var log12 = createLogger("prove-pr");
+function createOutputDir2(baseDir) {
+  const ts = timestamp();
+  const dir = path7.resolve(baseDir, `prove-pr_${ts}`);
+  fs5.mkdirSync(dir, { recursive: true });
+  return dir;
+}
+async function provePr(request) {
+  const baseRef = request.baseRef || detectDefaultBranch();
+  const outputDir = createOutputDir2(request.outputDir);
+  const { width, height } = request.viewport;
+  log12.info(`proving PR: base=${baseRef} agent=${request.agent} hypotheses=${request.hypothesesCount}`);
+  const startTime = Date.now();
+  const gitContext = await buildGitContext(baseRef, true);
+  log12.debug("generating hypotheses...");
+  const planPrompt = buildPlanPrompt(gitContext, request.hypothesesCount, request.hint);
+  const planResult = await runAgent({
+    agent: request.agent,
+    model: request.model,
+    prompt: planPrompt,
+    streamOutput: true,
+    timeoutMs: request.timeoutMs
+  });
+  const hypothesesPlan = planResult.json.stdout;
+  log12.debug("proving hypotheses with evidence...");
+  const browserConfig = {
+    output: {
+      dir: outputDir,
+      video: true,
+      screenshots: true,
+      viewport: { width, height },
+      showCursor: true,
+      effects: true
+    }
+  };
+  const workspaceFiles = createWorkspaceFiles(browserConfig);
+  const provePrompt = buildProvePrPrompt(gitContext, hypothesesPlan, outputDir, request.url);
+  try {
+    const proveResult = await runAgent({
+      agent: request.agent,
+      model: request.model,
+      prompt: provePrompt,
+      approveMcps: true,
+      browser: true,
+      streamOutput: true,
+      approvalPolicy: "never",
+      workspacePath: workspaceFiles.workspacePath,
+      timeoutMs: request.timeoutMs
+    });
+    const durationMs = Date.now() - startTime;
+    const verdicts = parseHypothesesVerdicts(proveResult.json.stdout);
+    const summary = {
+      pass: verdicts.filter((v) => v.verdict === "pass").length,
+      fail: verdicts.filter((v) => v.verdict === "fail").length,
+      skip: verdicts.filter((v) => v.verdict === "skip").length
+    };
+    const result = {
+      baseRef,
+      hypotheses: verdicts,
+      summary,
+      outputDir,
+      durationMs
+    };
+    const report = generateReport(result);
+    fs5.writeFileSync(path7.join(outputDir, "report.md"), report);
+    fs5.writeFileSync(path7.join(outputDir, "result.json"), JSON.stringify(result, null, 2));
+    fs5.writeFileSync(path7.join(outputDir, "hypotheses.txt"), hypothesesPlan);
+    fs5.writeFileSync(path7.join(outputDir, "evidence.txt"), proveResult.json.stdout);
+    log12.info(`prove-pr complete: ${summary.pass} passed, ${summary.fail} failed, ${summary.skip} skipped`);
+    return result;
+  } finally {
+    workspaceFiles.cleanup();
+  }
+}
+
+// ../prove-pr/dist/validation.js
+var viewportSchema2 = external_exports.object({
+  width: external_exports.number().int().positive(),
+  height: external_exports.number().int().positive()
+});
+var provePrRequestSchema = external_exports.object({
+  baseRef: external_exports.string().optional(),
+  agent: external_exports.enum(AGENT_NAMES),
+  url: external_exports.string().url("Invalid URL").optional(),
+  outputDir: external_exports.string().min(1, "Output directory cannot be empty"),
+  hypothesesCount: external_exports.number().int().positive("Hypotheses count must be a positive integer"),
+  model: external_exports.string().optional(),
+  viewport: viewportSchema2,
+  hint: external_exports.string().optional(),
+  timeoutMs: external_exports.number().positive("Timeout must be positive").optional()
+});
+
 export {
   AgexError,
   isAgexError,
@@ -6130,5 +6317,6 @@ export {
   runAgent,
   prove,
   proveOptionsSchema,
-  runReview
+  runReview,
+  provePr
 };

package/dist/cli.js

@@ -7,11 +7,12 @@ import {
   isAgexError,
   prove,
   proveOptionsSchema,
+  provePr,
   runAgent,
   runReview,
   validate,
   viewportStringSchema
-} from "./chunk-EN7TTESF.js";
+} from "./chunk-CY6NF2GO.js";
 
 // src/cli.ts
 import { defineCommand as defineCommand6, runMain } from "citty";
@@ -255,7 +256,6 @@ var prove_default = defineCommand2({
 
 // src/commands/prove-pr.ts
 import { defineCommand as defineCommand3 } from "citty";
-import { provePr } from "agex-prove-pr";
 var prove_pr_default = defineCommand3({
   meta: {
     name: "prove-pr",

package/dist/index.js

@@ -2,12 +2,10 @@ import {
   AgexError,
   isAgexError,
   prove,
+  provePr,
   runAgent,
   runReview
-} from "./chunk-EN7TTESF.js";
-
-// src/index.ts
-import { provePr } from "agex-prove-pr";
+} from "./chunk-CY6NF2GO.js";
 export {
   AgexError,
   isAgexError,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agex",
-  "version": "0.2.3",
+  "version": "0.2.4",
   "type": "module",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -13,19 +13,19 @@
   ],
   "dependencies": {
     "citty": "^0.2.0",
-    "execa": "^9.6.1",
-    "agex-agent": "0.1.0",
-    "agex-browse": "0.1.0",
-    "agex-prove-pr": "0.1.0",
-    "agex-review": "0.1.0",
-    "agex-prove": "0.1.0",
-    "agex-core": "0.1.0"
+    "execa": "^9.6.1"
   },
   "devDependencies": {
     "@types/node": "^22.19.7",
     "tsup": "^8.4.0",
     "typescript": "^5.9.3",
-    "vitest": "^2.1.9"
+    "vitest": "^2.1.9",
+    "agex-agent": "0.1.0",
+    "agex-core": "0.1.0",
+    "agex-browse": "0.1.0",
+    "agex-prove-pr": "0.1.0",
+    "agex-review": "0.1.0",
+    "agex-prove": "0.1.0"
   },
   "scripts": {
     "build": "tsc -p tsconfig.json",