@gaberrb/polypus 0.4.6 → 0.4.7

package/dist/index.js

@@ -110,6 +110,7 @@ var en = {
   "cli.opt.mode": "plan | review | bypass (overrides config)",
   "cli.opt.maxSteps": "maximum agent steps",
   "cli.opt.json": "headless mode: emit a single JSON object (steps, tool calls, files changed, usage) instead of the TUI \u2014 use with --mode bypass",
+  "cli.opt.verify": "after the agent finishes, run project checks (typecheck/build/test) and iterate until they pass",
   "cli.arg.swarmTask": "high-level task to split across agents",
   "cli.opt.agents": "comma-separated agent names (default: all configured)",
   "cli.opt.maxSubtasks": "maximum number of parallel subtasks",
@@ -141,6 +142,11 @@ var en = {
   "review.reject": "reject",
   "review.pickHunks": "pick hunks\u2026",
   "review.selectHunks": "Select the hunks to apply (space to toggle, enter to confirm)",
+  "verify.running": "verifying (running project checks)",
+  "verify.noChecks": "no verification checks detected (no package.json scripts) \u2014 skipping",
+  "verify.passed": "verification passed",
+  "verify.failed": "{n} check(s) failed \u2014 handing the output back to the agent (attempt {attempt})",
+  "verify.giveUp": "{n} check(s) still failing after the retry budget \u2014 stopping",
   // repl
   "repl.welcome": "Polypus interactive session.",
   "repl.welcomeHint": " Type /help for commands, /exit to quit.",
@@ -344,6 +350,7 @@ var ptBR = {
   "cli.opt.mode": "plan | review | bypass (sobrescreve a config)",
   "cli.opt.maxSteps": "n\xFAmero m\xE1ximo de passos do agente",
   "cli.opt.json": "modo headless: emite um \xFAnico objeto JSON (passos, tool calls, arquivos alterados, uso) em vez da TUI \u2014 use com --mode bypass",
+  "cli.opt.verify": "ap\xF3s o agente terminar, roda as checagens do projeto (typecheck/build/test) e itera at\xE9 passar",
   "cli.arg.swarmTask": "tarefa de alto n\xEDvel para dividir entre os agentes",
   "cli.opt.agents": "nomes de agentes separados por v\xEDrgula (padr\xE3o: todos)",
   "cli.opt.maxSubtasks": "n\xFAmero m\xE1ximo de subtarefas paralelas",
@@ -373,6 +380,11 @@ var ptBR = {
   "review.reject": "rejeitar",
   "review.pickHunks": "escolher hunks\u2026",
   "review.selectHunks": "Selecione os hunks a aplicar (espa\xE7o alterna, enter confirma)",
+  "verify.running": "verificando (rodando as checagens do projeto)",
+  "verify.noChecks": "nenhuma checagem detectada (sem scripts no package.json) \u2014 pulando",
+  "verify.passed": "verifica\xE7\xE3o passou",
+  "verify.failed": "{n} checagem(ns) falharam \u2014 devolvendo a sa\xEDda ao agente (tentativa {attempt})",
+  "verify.giveUp": "{n} checagem(ns) ainda falhando ap\xF3s o limite de tentativas \u2014 parando",
   "repl.welcome": "Sess\xE3o interativa do Polypus.",
   "repl.welcomeHint": " Digite /help para comandos, /exit para sair.",
   "repl.modeChanged": "modo \u2192 {mode}",
@@ -2232,6 +2244,58 @@ ${blocks.join("\n\n")}`;
   return { task: augmented, injected };
 }
 
+// src/core/agent/verify.ts
+import { exec as exec2 } from "child_process";
+import { readFile as readFile9 } from "fs/promises";
+import { resolve as resolve10 } from "path";
+import { promisify as promisify2 } from "util";
+var execAsync2 = promisify2(exec2);
+var MAX_OUTPUT3 = 8e3;
+var CHECK_SCRIPTS = ["typecheck", "build", "test"];
+async function detectChecks(workspace) {
+  try {
+    const raw = await readFile9(resolve10(workspace, "package.json"), "utf8");
+    const scripts = JSON.parse(raw).scripts ?? {};
+    return CHECK_SCRIPTS.filter((s) => typeof scripts[s] === "string").map((s) => `npm run ${s}`);
+  } catch {
+    return [];
+  }
+}
+async function runChecks(workspace, commands) {
+  const results = [];
+  for (const command of commands) {
+    try {
+      const { stdout: stdout2, stderr } = await execAsync2(command, {
+        cwd: workspace,
+        timeout: 3e5,
+        maxBuffer: 10 * 1024 * 1024,
+        windowsHide: true
+      });
+      results.push({ command, ok: true, output: clamp2(`${stdout2}${stderr ? `
+${stderr}` : ""}`.trim()) });
+    } catch (err) {
+      const e = err;
+      const body = `${e.stdout ?? ""}${e.stderr ?? ""}`.trim() || e.message;
+      results.push({ command, ok: false, output: clamp2(body) });
+    }
+  }
+  return results;
+}
+function buildVerifyFeedback(failed) {
+  const blocks = failed.map((f) => `$ ${f.command}
+${f.output}`).join("\n\n");
+  return [
+    "Verification failed. The following checks did not pass:",
+    "",
+    blocks,
+    "",
+    "Fix the underlying problem in the code (do not disable or skip the checks), then call `finish`."
+  ].join("\n");
+}
+function clamp2(s) {
+  return s.length > MAX_OUTPUT3 ? s.slice(-MAX_OUTPUT3) : s;
+}
+
 // src/cli/commands/json-output.ts
 var OUTPUT_PREVIEW = 500;
 function createJsonCollector() {
@@ -2924,10 +2988,10 @@ async function readLineTTY(prompt) {
   stdin.resume();
   stdin.on("data", onData);
   try {
-    const line = await new Promise((resolve11) => {
-      rl.question(prompt).then(resolve11, () => resolve11(null));
-      rl.on("SIGINT", () => resolve11(null));
-      rl.on("close", () => resolve11(null));
+    const line = await new Promise((resolve12) => {
+      rl.question(prompt).then(resolve12, () => resolve12(null));
+      rl.on("SIGINT", () => resolve12(null));
+      rl.on("close", () => resolve12(null));
     });
     return line === null ? null : store.expand(line);
   } finally {
@@ -3521,6 +3585,7 @@ var Spinner = class {
 };
 
 // src/cli/commands/run.ts
+var MAX_VERIFY_FIXES = 3;
 async function run(task, opts) {
   let config = await loadConfig();
   const agentConfig = resolveAgent(config, opts.agent);
@@ -3555,7 +3620,7 @@ async function run(task, opts) {
         )
       );
     }
-    await executeTask(task, resolved2, workspace, session, opts.json ?? false);
+    await executeTask(task, resolved2, workspace, session, opts.json ?? false, opts.verify ?? false);
     return;
   }
   const resolved = createProvider(agentConfig);
@@ -3578,7 +3643,7 @@ async function run(task, opts) {
   };
   await startRepl(ctx);
 }
-async function executeTask(task, resolved, workspace, session, json = false) {
+async function executeTask(task, resolved, workspace, session, json = false, verify = false) {
   const mention = await resolveMentions(task, {
     workspace,
     allow: session.allow,
@@ -3605,25 +3670,29 @@ async function executeTask(task, resolved, workspace, session, json = false) {
       return ok;
     }
   });
+  const runOnce = (taskText) => runAgent({
+    task: taskText,
+    workspace,
+    agent: resolved,
+    permissions,
+    promptContext: { workspace, mode: session.mode, allow: session.allow },
+    history: session.history,
+    maxSteps: session.maxSteps,
+    signal: controller.signal,
+    events: collector ? collector.events : renderEvents(spinner3)
+  });
   if (!json) spinner3.start(t("ui.thinking"));
   let result;
   try {
-    result = await runAgent({
-      task,
-      workspace,
-      agent: resolved,
-      permissions,
-      promptContext: { workspace, mode: session.mode, allow: session.allow },
-      history: session.history,
-      maxSteps: session.maxSteps,
-      signal: controller.signal,
-      events: collector ? collector.events : renderEvents(spinner3)
-    });
+    result = await runOnce(task);
+    session.history = result.messages;
+    if (verify && result.reason === "finished" && !controller.signal.aborted) {
+      result = await runVerification(runOnce, workspace, session, spinner3, json, controller.signal, result);
+    }
   } finally {
     spinner3.stop();
     cancel2.dispose();
   }
-  session.history = result.messages;
   if (collector) {
     process.stdout.write(JSON.stringify(collector.build(result)) + "\n");
     return;
@@ -3702,6 +3771,36 @@ async function confirmAction(req) {
   if (p2.isCancel(answer)) return false;
   return answer === true;
 }
+async function runVerification(runOnce, workspace, session, spinner3, json, signal, initial) {
+  const checks = await detectChecks(workspace);
+  if (checks.length === 0) {
+    if (!json) console.log(pc8.dim(t("verify.noChecks")));
+    return initial;
+  }
+  let result = initial;
+  for (let fix = 0; ; fix++) {
+    if (signal.aborted) return result;
+    if (!json) spinner3.start(t("verify.running"));
+    const results = await runChecks(workspace, checks);
+    spinner3.stop();
+    const failed = results.filter((r) => !r.ok);
+    if (failed.length === 0) {
+      if (!json) console.log(pc8.green("\u2713 " + t("verify.passed")));
+      return result;
+    }
+    if (fix >= MAX_VERIFY_FIXES) {
+      if (!json) console.log(pc8.yellow("\u26A0 " + t("verify.giveUp", { n: failed.length })));
+      return result;
+    }
+    if (!json) {
+      console.log(pc8.yellow("\u2717 " + t("verify.failed", { n: failed.length, attempt: fix + 1 })));
+    }
+    if (!json) spinner3.start(t("ui.thinking"));
+    result = await runOnce(buildVerifyFeedback(failed));
+    spinner3.stop();
+    session.history = result.messages;
+  }
+}
 function renderDiff(hunks) {
   for (const h of hunks) {
     console.log(pc8.cyan(`@@ -${h.oldStart + 1},${h.oldCount} +${h.newStart + 1},${h.newCount} @@`));
@@ -4100,9 +4199,9 @@ async function resolveOpenRouterKey() {
 }
 
 // src/cli/commands/prd.ts
-import { writeFile as writeFile5, readFile as readFile9 } from "fs/promises";
+import { writeFile as writeFile5, readFile as readFile10 } from "fs/promises";
 import { execFile } from "child_process";
-import { promisify as promisify2 } from "util";
+import { promisify as promisify3 } from "util";
 import pc11 from "picocolors";
 
 // src/core/agent/prd.ts
@@ -4192,13 +4291,13 @@ async function withRetry(fn, opts = {}) {
 
 // src/cli/commands/cli-io.ts
 import { readFileSync, existsSync as existsSync2 } from "fs";
-import { resolve as resolve10 } from "path";
+import { resolve as resolve11 } from "path";
 var GUIDE_MAX = 12e3;
 function readProjectGuide(files) {
   const parts = [];
   for (const file of files) {
     try {
-      const path = resolve10(process.cwd(), file);
+      const path = resolve11(process.cwd(), file);
       if (existsSync2(path)) parts.push(`# ${file}
 ${readFileSync(path, "utf8").trim()}`);
     } catch {
@@ -4224,7 +4323,7 @@ function stripBom(s) {
 }
 
 // src/cli/commands/prd.ts
-var exec2 = promisify2(execFile);
+var exec3 = promisify3(execFile);
 async function prd(issueRef, opts) {
   const issue = await loadIssue(issueRef, opts.input);
   const { provider } = resolveFreeProvider(opts.model ?? DEFAULT_PRD_MODEL);
@@ -4239,11 +4338,11 @@ async function prd(issueRef, opts) {
 }
 async function loadIssue(issueRef, input) {
   if (input) {
-    const raw = input === "-" ? await readStdin() : await readFile9(input, "utf8");
+    const raw = input === "-" ? await readStdin() : await readFile10(input, "utf8");
     return normalize2(JSON.parse(stripBom(raw)));
   }
   const num = numericRef(issueRef);
-  const { stdout: stdout2 } = await exec2("gh", ["issue", "view", num, "--json", "number,title,body,comments"]);
+  const { stdout: stdout2 } = await exec3("gh", ["issue", "view", num, "--json", "number,title,body,comments"]);
   const data = normalize2(JSON.parse(stdout2));
   data.number ??= Number(num);
   return data;
@@ -4258,9 +4357,9 @@ function normalize2(raw) {
 }
 
 // src/cli/commands/review.ts
-import { writeFile as writeFile6, readFile as readFile10 } from "fs/promises";
+import { writeFile as writeFile6, readFile as readFile11 } from "fs/promises";
 import { execFile as execFile2 } from "child_process";
-import { promisify as promisify3 } from "util";
+import { promisify as promisify4 } from "util";
 import pc12 from "picocolors";
 
 // src/core/agent/review.ts
@@ -4318,7 +4417,7 @@ ${projectGuide}`
 }
 
 // src/cli/commands/review.ts
-var exec3 = promisify3(execFile2);
+var exec4 = promisify4(execFile2);
 async function review(prRef, opts) {
   const num = opts.input ? prRef.replace(/^#/, "") : numericRef(prRef);
   const diff = await loadDiff(num, opts.input);
@@ -4334,13 +4433,13 @@ async function review(prRef, opts) {
   }
 }
 async function loadDiff(num, input) {
-  if (input) return input === "-" ? readStdin() : readFile10(input, "utf8");
-  const { stdout: stdout2 } = await exec3("gh", ["pr", "diff", num]);
+  if (input) return input === "-" ? readStdin() : readFile11(input, "utf8");
+  const { stdout: stdout2 } = await exec4("gh", ["pr", "diff", num]);
   return stdout2;
 }
 async function loadMeta(num, input) {
   if (input) return { number: Number(num) || void 0, title: `PR ${num}`, body: "" };
-  const { stdout: stdout2 } = await exec3("gh", ["pr", "view", num, "--json", "number,title,body"]);
+  const { stdout: stdout2 } = await exec4("gh", ["pr", "view", num, "--json", "number,title,body"]);
   const raw = JSON.parse(stdout2);
   return { number: raw.number, title: raw.title ?? "", body: raw.body ?? "" };
 }
@@ -4406,7 +4505,7 @@ function buildProgram() {
   program.command("add-agent").argument("<name>", t("cli.arg.addAgentName")).requiredOption("--provider <provider>", t("cli.opt.provider")).requiredOption("--model <model>", t("cli.opt.model")).option("--api-key <key>", t("cli.opt.apiKey")).option("--base-url <url>", t("cli.opt.baseUrl")).option("--tool-mode <mode>", t("cli.opt.toolMode"), "auto").option("--set-default", t("cli.opt.setDefault")).description(t("cli.cmd.addAgent")).action((name, opts) => addAgent(name, opts));
   program.command("remove-agent").argument("<name>", t("cli.arg.removeAgentName")).description(t("cli.cmd.removeAgent")).action((name) => removeAgent(name));
   program.command("list-agents").alias("agents").description(t("cli.cmd.listAgents")).action(() => listAgents());
-  program.command("run").argument("[task]", t("cli.arg.runTask")).option("--agent <name>", t("cli.opt.agent")).option("--mode <mode>", t("cli.opt.mode")).option("--max-steps <n>", t("cli.opt.maxSteps")).option("--json", t("cli.opt.json")).description(t("cli.cmd.run")).action((task, opts) => run(task, opts));
+  program.command("run").argument("[task]", t("cli.arg.runTask")).option("--agent <name>", t("cli.opt.agent")).option("--mode <mode>", t("cli.opt.mode")).option("--max-steps <n>", t("cli.opt.maxSteps")).option("--json", t("cli.opt.json")).option("--verify", t("cli.opt.verify")).description(t("cli.cmd.run")).action((task, opts) => run(task, opts));
   program.command("swarm").argument("<task>", t("cli.arg.swarmTask")).option("--agents <names>", t("cli.opt.agents")).option("--max-subtasks <n>", t("cli.opt.maxSubtasks")).description(t("cli.cmd.swarm")).action((task, opts) => swarm(task, opts));
   program.command("models").option("--search <text>", t("cli.opt.search")).option("--tools", t("cli.opt.toolsOnly")).option("--free", t("cli.opt.free")).option("--max-price <usd>", t("cli.opt.maxPrice")).option("--sort <order>", t("cli.opt.sort")).option("--limit <n>", t("cli.opt.limit")).description(t("cli.cmd.models")).action((opts) => models(opts));
   program.command("prd").argument("<issue>", t("cli.arg.prdIssue")).option("--out <file>", t("cli.opt.out")).option("--model <model>", t("cli.opt.model")).option("--input <file>", t("cli.opt.input")).description(t("cli.cmd.prd")).action((issue, opts) => prd(issue, opts));