@gaberrb/polypus 0.4.6 → 0.4.8

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/index.js CHANGED
@@ -2,7 +2,7 @@
2
2
 
3
3
  // src/cli/index.ts
4
4
  import { Command } from "commander";
5
- import pc13 from "picocolors";
5
+ import pc14 from "picocolors";
6
6
 
7
7
  // src/cli/commands/add-agent.ts
8
8
  import pc from "picocolors";
@@ -110,6 +110,9 @@ var en = {
110
110
  "cli.opt.mode": "plan | review | bypass (overrides config)",
111
111
  "cli.opt.maxSteps": "maximum agent steps",
112
112
  "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",
113
+ "cli.opt.verify": "after the agent finishes, run project checks (typecheck/build/test) and iterate until they pass",
114
+ "cli.opt.budget": "stop the run when the estimated session cost reaches this USD amount (OpenRouter pricing)",
115
+ "cli.cmd.usage": "Show token/cost analytics aggregated per day",
113
116
  "cli.arg.swarmTask": "high-level task to split across agents",
114
117
  "cli.opt.agents": "comma-separated agent names (default: all configured)",
115
118
  "cli.opt.maxSubtasks": "maximum number of parallel subtasks",
@@ -141,6 +144,18 @@ var en = {
141
144
  "review.reject": "reject",
142
145
  "review.pickHunks": "pick hunks\u2026",
143
146
  "review.selectHunks": "Select the hunks to apply (space to toggle, enter to confirm)",
147
+ "verify.running": "verifying (running project checks)",
148
+ "verify.noChecks": "no verification checks detected (no package.json scripts) \u2014 skipping",
149
+ "verify.passed": "verification passed",
150
+ "verify.failed": "{n} check(s) failed \u2014 handing the output back to the agent (attempt {attempt})",
151
+ "verify.giveUp": "{n} check(s) still failing after the retry budget \u2014 stopping",
152
+ "budget.session": "session spend: {spent} / budget {budget}",
153
+ "budget.hit": "\u25A0 stopped: estimated cost reached the budget of {budget}",
154
+ // usage analytics
155
+ "usage.header": "Usage (tokens / estimated cost) per day:",
156
+ "usage.empty": "No usage recorded yet. Run a task to start tracking.",
157
+ "usage.total": "total",
158
+ "usage.runs": "runs",
144
159
  // repl
145
160
  "repl.welcome": "Polypus interactive session.",
146
161
  "repl.welcomeHint": " Type /help for commands, /exit to quit.",
@@ -344,6 +359,9 @@ var ptBR = {
344
359
  "cli.opt.mode": "plan | review | bypass (sobrescreve a config)",
345
360
  "cli.opt.maxSteps": "n\xFAmero m\xE1ximo de passos do agente",
346
361
  "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",
362
+ "cli.opt.verify": "ap\xF3s o agente terminar, roda as checagens do projeto (typecheck/build/test) e itera at\xE9 passar",
363
+ "cli.opt.budget": "interrompe a execu\xE7\xE3o quando o custo estimado da sess\xE3o atingir este valor em USD (pre\xE7os do OpenRouter)",
364
+ "cli.cmd.usage": "Mostra analytics de tokens/custo agregados por dia",
347
365
  "cli.arg.swarmTask": "tarefa de alto n\xEDvel para dividir entre os agentes",
348
366
  "cli.opt.agents": "nomes de agentes separados por v\xEDrgula (padr\xE3o: todos)",
349
367
  "cli.opt.maxSubtasks": "n\xFAmero m\xE1ximo de subtarefas paralelas",
@@ -373,6 +391,18 @@ var ptBR = {
373
391
  "review.reject": "rejeitar",
374
392
  "review.pickHunks": "escolher hunks\u2026",
375
393
  "review.selectHunks": "Selecione os hunks a aplicar (espa\xE7o alterna, enter confirma)",
394
+ "verify.running": "verificando (rodando as checagens do projeto)",
395
+ "verify.noChecks": "nenhuma checagem detectada (sem scripts no package.json) \u2014 pulando",
396
+ "verify.passed": "verifica\xE7\xE3o passou",
397
+ "verify.failed": "{n} checagem(ns) falharam \u2014 devolvendo a sa\xEDda ao agente (tentativa {attempt})",
398
+ "verify.giveUp": "{n} checagem(ns) ainda falhando ap\xF3s o limite de tentativas \u2014 parando",
399
+ "budget.session": "gasto da sess\xE3o: {spent} / or\xE7amento {budget}",
400
+ "budget.hit": "\u25A0 interrompido: o custo estimado atingiu o or\xE7amento de {budget}",
401
+ // usage analytics
402
+ "usage.header": "Uso (tokens / custo estimado) por dia:",
403
+ "usage.empty": "Nenhum uso registrado ainda. Rode uma tarefa para come\xE7ar a medir.",
404
+ "usage.total": "total",
405
+ "usage.runs": "execu\xE7\xF5es",
376
406
  "repl.welcome": "Sess\xE3o interativa do Polypus.",
377
407
  "repl.welcomeHint": " Digite /help para comandos, /exit para sair.",
378
408
  "repl.modeChanged": "modo \u2192 {mode}",
@@ -2084,9 +2114,9 @@ async function runAgent(opts) {
2084
2114
  let failStreak = 0;
2085
2115
  const maxToolRetries = opts.maxToolRetries ?? 3;
2086
2116
  const autoCorrect = opts.autoCorrect ?? true;
2087
- const usage = { promptTokens: 0, completionTokens: 0 };
2117
+ const usage2 = { promptTokens: 0, completionTokens: 0 };
2088
2118
  for (let step = 1; step <= maxSteps; step++) {
2089
- if (opts.signal?.aborted) return { finished: false, reason: "cancelled", steps: step - 1, messages, usage };
2119
+ if (opts.signal?.aborted) return { finished: false, reason: "cancelled", steps: step - 1, messages, usage: usage2 };
2090
2120
  events?.onStep?.(step);
2091
2121
  let response;
2092
2122
  try {
@@ -2097,12 +2127,12 @@ async function runAgent(opts) {
2097
2127
  signal: opts.signal
2098
2128
  });
2099
2129
  } catch (err) {
2100
- if (opts.signal?.aborted) return { finished: false, reason: "cancelled", steps: step, messages, usage };
2130
+ if (opts.signal?.aborted) return { finished: false, reason: "cancelled", steps: step, messages, usage: usage2 };
2101
2131
  throw err;
2102
2132
  }
2103
- usage.promptTokens += response.usage?.promptTokens ?? 0;
2104
- usage.completionTokens += response.usage?.completionTokens ?? 0;
2105
- events?.onUsage?.(usage);
2133
+ usage2.promptTokens += response.usage?.promptTokens ?? 0;
2134
+ usage2.completionTokens += response.usage?.completionTokens ?? 0;
2135
+ events?.onUsage?.(usage2);
2106
2136
  const { toolCalls, text: text2 } = driver.parse(response);
2107
2137
  messages.push(driver.assistantMessage(response, toolCalls));
2108
2138
  if (text2) events?.onAssistantText?.(text2);
@@ -2116,7 +2146,7 @@ async function runAgent(opts) {
2116
2146
  messages.push({ role: "user", content: guidance });
2117
2147
  continue;
2118
2148
  }
2119
- return { finished: false, reason: "stalled", steps: step, messages, usage };
2149
+ return { finished: false, reason: "stalled", steps: step, messages, usage: usage2 };
2120
2150
  }
2121
2151
  const stalled = text2.trim().length === 0 || looksLikeStall(text2);
2122
2152
  if (stalled) {
@@ -2126,17 +2156,17 @@ async function runAgent(opts) {
2126
2156
  messages.push(driver.repromptMessage());
2127
2157
  continue;
2128
2158
  }
2129
- return { finished: false, reason: "stalled", steps: step, messages, usage };
2159
+ return { finished: false, reason: "stalled", steps: step, messages, usage: usage2 };
2130
2160
  }
2131
- return { finished: false, reason: "reply", steps: step, messages, usage };
2161
+ return { finished: false, reason: "reply", steps: step, messages, usage: usage2 };
2132
2162
  }
2133
2163
  consecutiveNoTool = 0;
2134
2164
  for (const call of toolCalls) {
2135
- if (opts.signal?.aborted) return { finished: false, reason: "cancelled", steps: step, messages, usage };
2165
+ if (opts.signal?.aborted) return { finished: false, reason: "cancelled", steps: step, messages, usage: usage2 };
2136
2166
  events?.onToolCall?.(call);
2137
2167
  if (call.name === "finish") {
2138
2168
  const summary = String(call.arguments.summary ?? "").trim();
2139
- return { finished: true, reason: "finished", summary, steps: step, messages, usage };
2169
+ return { finished: true, reason: "finished", summary, steps: step, messages, usage: usage2 };
2140
2170
  }
2141
2171
  const tool = getTool(call.name);
2142
2172
  const result = tool ? await tool.run(call.arguments, ctx) : { ok: false, output: `Unknown tool "${call.name}". Available: ${toolSpecs().map((t2) => t2.name).join(", ")}` };
@@ -2174,12 +2204,12 @@ ${guidance}`;
2174
2204
  failStreak = sig === lastFailSig ? failStreak + 1 : 1;
2175
2205
  lastFailSig = sig;
2176
2206
  if (failStreak >= maxToolRetries) {
2177
- return { finished: false, reason: "stalled", steps: step, messages, usage };
2207
+ return { finished: false, reason: "stalled", steps: step, messages, usage: usage2 };
2178
2208
  }
2179
2209
  }
2180
2210
  }
2181
2211
  }
2182
- return { finished: false, reason: "maxsteps", steps: maxSteps, messages, usage };
2212
+ return { finished: false, reason: "maxsteps", steps: maxSteps, messages, usage: usage2 };
2183
2213
  }
2184
2214
 
2185
2215
  // src/core/context/mentions.ts
@@ -2232,6 +2262,205 @@ ${blocks.join("\n\n")}`;
2232
2262
  return { task: augmented, injected };
2233
2263
  }
2234
2264
 
2265
+ // src/core/agent/verify.ts
2266
+ import { exec as exec2 } from "child_process";
2267
+ import { readFile as readFile9 } from "fs/promises";
2268
+ import { resolve as resolve10 } from "path";
2269
+ import { promisify as promisify2 } from "util";
2270
+ var execAsync2 = promisify2(exec2);
2271
+ var MAX_OUTPUT3 = 8e3;
2272
+ var CHECK_SCRIPTS = ["typecheck", "build", "test"];
2273
+ async function detectChecks(workspace) {
2274
+ try {
2275
+ const raw = await readFile9(resolve10(workspace, "package.json"), "utf8");
2276
+ const scripts = JSON.parse(raw).scripts ?? {};
2277
+ return CHECK_SCRIPTS.filter((s) => typeof scripts[s] === "string").map((s) => `npm run ${s}`);
2278
+ } catch {
2279
+ return [];
2280
+ }
2281
+ }
2282
+ async function runChecks(workspace, commands) {
2283
+ const results = [];
2284
+ for (const command of commands) {
2285
+ try {
2286
+ const { stdout: stdout2, stderr } = await execAsync2(command, {
2287
+ cwd: workspace,
2288
+ timeout: 3e5,
2289
+ maxBuffer: 10 * 1024 * 1024,
2290
+ windowsHide: true
2291
+ });
2292
+ results.push({ command, ok: true, output: clamp2(`${stdout2}${stderr ? `
2293
+ ${stderr}` : ""}`.trim()) });
2294
+ } catch (err) {
2295
+ const e = err;
2296
+ const body = `${e.stdout ?? ""}${e.stderr ?? ""}`.trim() || e.message;
2297
+ results.push({ command, ok: false, output: clamp2(body) });
2298
+ }
2299
+ }
2300
+ return results;
2301
+ }
2302
+ function buildVerifyFeedback(failed) {
2303
+ const blocks = failed.map((f) => `$ ${f.command}
2304
+ ${f.output}`).join("\n\n");
2305
+ return [
2306
+ "Verification failed. The following checks did not pass:",
2307
+ "",
2308
+ blocks,
2309
+ "",
2310
+ "Fix the underlying problem in the code (do not disable or skip the checks), then call `finish`."
2311
+ ].join("\n");
2312
+ }
2313
+ function clamp2(s) {
2314
+ return s.length > MAX_OUTPUT3 ? s.slice(-MAX_OUTPUT3) : s;
2315
+ }
2316
+
2317
+ // src/core/agent/usage.ts
2318
+ import { appendFile, mkdir as mkdir3, readFile as readFile10 } from "fs/promises";
2319
+ import { join as join4 } from "path";
2320
+
2321
+ // src/core/providers/openrouter.ts
2322
+ var MODELS_URL = "https://openrouter.ai/api/v1/models";
2323
+ async function listOpenRouterModels(apiKey, timeoutMs = 8e3) {
2324
+ const controller = new AbortController();
2325
+ const timer = setTimeout(() => controller.abort(), timeoutMs);
2326
+ try {
2327
+ const res = await fetch(MODELS_URL, {
2328
+ signal: controller.signal,
2329
+ headers: apiKey ? { authorization: `Bearer ${apiKey}` } : {}
2330
+ });
2331
+ if (!res.ok) throw new Error(`OpenRouter ${res.status}: ${await res.text().catch(() => "")}`);
2332
+ const data = await res.json();
2333
+ return (data.data ?? []).map(normalize).filter((m) => m.id.length > 0);
2334
+ } finally {
2335
+ clearTimeout(timer);
2336
+ }
2337
+ }
2338
+ function normalize(m) {
2339
+ const promptPrice = toPerMillion(m.pricing?.prompt);
2340
+ const completionPrice = toPerMillion(m.pricing?.completion);
2341
+ return {
2342
+ id: m.id ?? "",
2343
+ name: m.name ?? m.id ?? "",
2344
+ promptPrice,
2345
+ completionPrice,
2346
+ contextLength: m.context_length ?? 0,
2347
+ supportsTools: (m.supported_parameters ?? []).includes("tools"),
2348
+ free: promptPrice === 0 && completionPrice === 0
2349
+ };
2350
+ }
2351
+ function toPerMillion(price) {
2352
+ const n = Number(price ?? "0");
2353
+ return Number.isFinite(n) ? n * 1e6 : 0;
2354
+ }
2355
+ function filterModels(models2, f) {
2356
+ const term = f.search?.trim().toLowerCase();
2357
+ let out = models2.filter((m) => {
2358
+ if (term && !m.id.toLowerCase().includes(term) && !m.name.toLowerCase().includes(term)) {
2359
+ return false;
2360
+ }
2361
+ if (f.tools === "tools" && !m.supportsTools) return false;
2362
+ if (f.tools === "no-tools" && m.supportsTools) return false;
2363
+ if (f.freeOnly && !m.free) return false;
2364
+ if (f.maxPrice !== void 0 && (m.promptPrice < 0 || m.promptPrice > f.maxPrice)) return false;
2365
+ return true;
2366
+ });
2367
+ const key = (m) => m.promptPrice < 0 ? Number.POSITIVE_INFINITY : m.promptPrice;
2368
+ const sort = f.sort ?? "price";
2369
+ out = out.sort((a, b) => {
2370
+ switch (sort) {
2371
+ case "price-desc":
2372
+ return key(b) - key(a);
2373
+ case "context":
2374
+ return b.contextLength - a.contextLength;
2375
+ case "name":
2376
+ return a.id.localeCompare(b.id);
2377
+ default:
2378
+ return key(a) - key(b);
2379
+ }
2380
+ });
2381
+ return out;
2382
+ }
2383
+ function fmtPrice(perMillion) {
2384
+ if (perMillion < 0) return "var";
2385
+ if (perMillion === 0) return "free";
2386
+ if (perMillion < 1) return `$${perMillion.toFixed(2)}`;
2387
+ if (perMillion < 100) return `$${perMillion.toFixed(perMillion % 1 ? 1 : 0)}`;
2388
+ return `$${Math.round(perMillion)}`;
2389
+ }
2390
+ function fmtContext(n) {
2391
+ if (n >= 1e6) return `${Math.round(n / 1e5) / 10}M`;
2392
+ if (n >= 1e3) return `${Math.round(n / 1e3)}k`;
2393
+ return String(n);
2394
+ }
2395
+
2396
+ // src/core/agent/usage.ts
2397
+ var catalogCache;
2398
+ async function resolveModelPricing(agent) {
2399
+ if (agent.provider !== "openrouter") return void 0;
2400
+ try {
2401
+ catalogCache ??= listOpenRouterModels(resolveSecret(agent.apiKey));
2402
+ const models2 = await catalogCache;
2403
+ const m = models2.find((x) => x.id === agent.model);
2404
+ if (!m || m.promptPrice < 0 || m.completionPrice < 0) return void 0;
2405
+ return { promptPrice: m.promptPrice, completionPrice: m.completionPrice };
2406
+ } catch {
2407
+ return void 0;
2408
+ }
2409
+ }
2410
+ function estimateCost(usage2, pricing) {
2411
+ return usage2.promptTokens / 1e6 * pricing.promptPrice + usage2.completionTokens / 1e6 * pricing.completionPrice;
2412
+ }
2413
+ function fmtUsd(n) {
2414
+ if (n <= 0) return "US$0.00";
2415
+ if (n < 0.01) return `US$${n.toFixed(4)}`;
2416
+ return `US$${n.toFixed(2)}`;
2417
+ }
2418
+ function usagePath() {
2419
+ return join4(configDir(), "usage.jsonl");
2420
+ }
2421
+ async function recordUsage(entry) {
2422
+ try {
2423
+ await mkdir3(configDir(), { recursive: true });
2424
+ await appendFile(usagePath(), JSON.stringify(entry) + "\n", "utf8");
2425
+ } catch {
2426
+ }
2427
+ }
2428
+ async function aggregateUsage() {
2429
+ let text2 = "";
2430
+ try {
2431
+ text2 = await readFile10(usagePath(), "utf8");
2432
+ } catch {
2433
+ return { days: [], total: emptyBucket("total") };
2434
+ }
2435
+ const byDay = /* @__PURE__ */ new Map();
2436
+ const total = emptyBucket("total");
2437
+ for (const line of text2.split("\n")) {
2438
+ if (!line.trim()) continue;
2439
+ let e;
2440
+ try {
2441
+ e = JSON.parse(line);
2442
+ } catch {
2443
+ continue;
2444
+ }
2445
+ const date = (e.ts ?? "").slice(0, 10) || "unknown";
2446
+ const bucket = byDay.get(date) ?? emptyBucket(date);
2447
+ accumulate(bucket, e);
2448
+ byDay.set(date, bucket);
2449
+ accumulate(total, e);
2450
+ }
2451
+ const days = [...byDay.values()].sort((a, b) => a.date.localeCompare(b.date));
2452
+ return { days, total };
2453
+ }
2454
+ function emptyBucket(date) {
2455
+ return { date, promptTokens: 0, completionTokens: 0, costUsd: 0, runs: 0 };
2456
+ }
2457
+ function accumulate(bucket, e) {
2458
+ bucket.promptTokens += e.promptTokens ?? 0;
2459
+ bucket.completionTokens += e.completionTokens ?? 0;
2460
+ bucket.costUsd += e.costUsd ?? 0;
2461
+ bucket.runs += 1;
2462
+ }
2463
+
2235
2464
  // src/cli/commands/json-output.ts
2236
2465
  var OUTPUT_PREVIEW = 500;
2237
2466
  function createJsonCollector() {
@@ -2317,81 +2546,6 @@ async function listOllamaModels(host = ollamaHost(), timeoutMs = 2e3) {
2317
2546
  }
2318
2547
  }
2319
2548
 
2320
- // src/core/providers/openrouter.ts
2321
- var MODELS_URL = "https://openrouter.ai/api/v1/models";
2322
- async function listOpenRouterModels(apiKey, timeoutMs = 8e3) {
2323
- const controller = new AbortController();
2324
- const timer = setTimeout(() => controller.abort(), timeoutMs);
2325
- try {
2326
- const res = await fetch(MODELS_URL, {
2327
- signal: controller.signal,
2328
- headers: apiKey ? { authorization: `Bearer ${apiKey}` } : {}
2329
- });
2330
- if (!res.ok) throw new Error(`OpenRouter ${res.status}: ${await res.text().catch(() => "")}`);
2331
- const data = await res.json();
2332
- return (data.data ?? []).map(normalize).filter((m) => m.id.length > 0);
2333
- } finally {
2334
- clearTimeout(timer);
2335
- }
2336
- }
2337
- function normalize(m) {
2338
- const promptPrice = toPerMillion(m.pricing?.prompt);
2339
- const completionPrice = toPerMillion(m.pricing?.completion);
2340
- return {
2341
- id: m.id ?? "",
2342
- name: m.name ?? m.id ?? "",
2343
- promptPrice,
2344
- completionPrice,
2345
- contextLength: m.context_length ?? 0,
2346
- supportsTools: (m.supported_parameters ?? []).includes("tools"),
2347
- free: promptPrice === 0 && completionPrice === 0
2348
- };
2349
- }
2350
- function toPerMillion(price) {
2351
- const n = Number(price ?? "0");
2352
- return Number.isFinite(n) ? n * 1e6 : 0;
2353
- }
2354
- function filterModels(models2, f) {
2355
- const term = f.search?.trim().toLowerCase();
2356
- let out = models2.filter((m) => {
2357
- if (term && !m.id.toLowerCase().includes(term) && !m.name.toLowerCase().includes(term)) {
2358
- return false;
2359
- }
2360
- if (f.tools === "tools" && !m.supportsTools) return false;
2361
- if (f.tools === "no-tools" && m.supportsTools) return false;
2362
- if (f.freeOnly && !m.free) return false;
2363
- if (f.maxPrice !== void 0 && (m.promptPrice < 0 || m.promptPrice > f.maxPrice)) return false;
2364
- return true;
2365
- });
2366
- const key = (m) => m.promptPrice < 0 ? Number.POSITIVE_INFINITY : m.promptPrice;
2367
- const sort = f.sort ?? "price";
2368
- out = out.sort((a, b) => {
2369
- switch (sort) {
2370
- case "price-desc":
2371
- return key(b) - key(a);
2372
- case "context":
2373
- return b.contextLength - a.contextLength;
2374
- case "name":
2375
- return a.id.localeCompare(b.id);
2376
- default:
2377
- return key(a) - key(b);
2378
- }
2379
- });
2380
- return out;
2381
- }
2382
- function fmtPrice(perMillion) {
2383
- if (perMillion < 0) return "var";
2384
- if (perMillion === 0) return "free";
2385
- if (perMillion < 1) return `$${perMillion.toFixed(2)}`;
2386
- if (perMillion < 100) return `$${perMillion.toFixed(perMillion % 1 ? 1 : 0)}`;
2387
- return `$${Math.round(perMillion)}`;
2388
- }
2389
- function fmtContext(n) {
2390
- if (n >= 1e6) return `${Math.round(n / 1e5) / 10}M`;
2391
- if (n >= 1e3) return `${Math.round(n / 1e3)}k`;
2392
- return String(n);
2393
- }
2394
-
2395
2549
  // src/ui/wizard.ts
2396
2550
  function bail(value) {
2397
2551
  if (p.isCancel(value)) {
@@ -2924,10 +3078,10 @@ async function readLineTTY(prompt) {
2924
3078
  stdin.resume();
2925
3079
  stdin.on("data", onData);
2926
3080
  try {
2927
- const line = await new Promise((resolve11) => {
2928
- rl.question(prompt).then(resolve11, () => resolve11(null));
2929
- rl.on("SIGINT", () => resolve11(null));
2930
- rl.on("close", () => resolve11(null));
3081
+ const line = await new Promise((resolve12) => {
3082
+ rl.question(prompt).then(resolve12, () => resolve12(null));
3083
+ rl.on("SIGINT", () => resolve12(null));
3084
+ rl.on("close", () => resolve12(null));
2931
3085
  });
2932
3086
  return line === null ? null : store.expand(line);
2933
3087
  } finally {
@@ -3072,7 +3226,7 @@ import pc7 from "picocolors";
3072
3226
  // src/core/git/worktree.ts
3073
3227
  import { mkdtemp } from "fs/promises";
3074
3228
  import { tmpdir } from "os";
3075
- import { join as join4 } from "path";
3229
+ import { join as join5 } from "path";
3076
3230
  import { simpleGit } from "simple-git";
3077
3231
  async function ensureRepo(workspace) {
3078
3232
  const git = simpleGit(workspace);
@@ -3093,7 +3247,7 @@ async function identityArgs(git) {
3093
3247
  }
3094
3248
  async function createWorktree(git, label) {
3095
3249
  const branch = `polypus/${label}-${Date.now().toString(36)}`;
3096
- const path = await mkdtemp(join4(tmpdir(), "polypus-wt-"));
3250
+ const path = await mkdtemp(join5(tmpdir(), "polypus-wt-"));
3097
3251
  await git.raw(["worktree", "add", "-b", branch, path, "HEAD"]);
3098
3252
  return { path, branch };
3099
3253
  }
@@ -3521,6 +3675,7 @@ var Spinner = class {
3521
3675
  };
3522
3676
 
3523
3677
  // src/cli/commands/run.ts
3678
+ var MAX_VERIFY_FIXES = 3;
3524
3679
  async function run(task, opts) {
3525
3680
  let config = await loadConfig();
3526
3681
  const agentConfig = resolveAgent(config, opts.agent);
@@ -3532,7 +3687,9 @@ async function run(task, opts) {
3532
3687
  deny: config.permissions.deny,
3533
3688
  allowedCommands: config.permissions.allowedCommands,
3534
3689
  maxSteps: opts.maxSteps ? Number(opts.maxSteps) : void 0,
3535
- history: []
3690
+ history: [],
3691
+ budget: opts.budget ? Number(opts.budget) : void 0,
3692
+ costUsd: 0
3536
3693
  };
3537
3694
  const runTask = async (taskText) => {
3538
3695
  const active = resolveAgent(config, session.agentName);
@@ -3555,7 +3712,10 @@ async function run(task, opts) {
3555
3712
  )
3556
3713
  );
3557
3714
  }
3558
- await executeTask(task, resolved2, workspace, session, opts.json ?? false);
3715
+ await executeTask(task, resolved2, workspace, session, opts.json ?? false, opts.verify ?? false);
3716
+ if (session.budget !== void 0 && !opts.json) {
3717
+ console.log(pc8.dim(t("budget.session", { spent: fmtUsd(session.costUsd), budget: fmtUsd(session.budget) })));
3718
+ }
3559
3719
  return;
3560
3720
  }
3561
3721
  const resolved = createProvider(agentConfig);
@@ -3578,7 +3738,7 @@ async function run(task, opts) {
3578
3738
  };
3579
3739
  await startRepl(ctx);
3580
3740
  }
3581
- async function executeTask(task, resolved, workspace, session, json = false) {
3741
+ async function executeTask(task, resolved, workspace, session, json = false, verify = false) {
3582
3742
  const mention = await resolveMentions(task, {
3583
3743
  workspace,
3584
3744
  allow: session.allow,
@@ -3592,6 +3752,21 @@ async function executeTask(task, resolved, workspace, session, json = false) {
3592
3752
  const controller = new AbortController();
3593
3753
  const cancel2 = listenForCancel(controller);
3594
3754
  const collector = json ? createJsonCollector() : void 0;
3755
+ const pricing = await resolveModelPricing(resolved.config);
3756
+ let budgetHit = false;
3757
+ const baseEvents = collector ? collector.events : renderEvents(spinner3);
3758
+ const events = {
3759
+ ...baseEvents,
3760
+ onUsage(u) {
3761
+ baseEvents.onUsage?.(u);
3762
+ if (session.budget !== void 0 && pricing && !controller.signal.aborted) {
3763
+ if (session.costUsd + estimateCost(u, pricing) >= session.budget) {
3764
+ budgetHit = true;
3765
+ controller.abort();
3766
+ }
3767
+ }
3768
+ }
3769
+ };
3595
3770
  const permissions = new PermissionEngine({
3596
3771
  mode: session.mode,
3597
3772
  policy: { workspace, allow: session.allow, deny: session.deny },
@@ -3605,29 +3780,47 @@ async function executeTask(task, resolved, workspace, session, json = false) {
3605
3780
  return ok;
3606
3781
  }
3607
3782
  });
3783
+ const runOnce = (taskText) => runAgent({
3784
+ task: taskText,
3785
+ workspace,
3786
+ agent: resolved,
3787
+ permissions,
3788
+ promptContext: { workspace, mode: session.mode, allow: session.allow },
3789
+ history: session.history,
3790
+ maxSteps: session.maxSteps,
3791
+ signal: controller.signal,
3792
+ events
3793
+ });
3608
3794
  if (!json) spinner3.start(t("ui.thinking"));
3609
3795
  let result;
3610
3796
  try {
3611
- result = await runAgent({
3612
- task,
3613
- workspace,
3614
- agent: resolved,
3615
- permissions,
3616
- promptContext: { workspace, mode: session.mode, allow: session.allow },
3617
- history: session.history,
3618
- maxSteps: session.maxSteps,
3619
- signal: controller.signal,
3620
- events: collector ? collector.events : renderEvents(spinner3)
3621
- });
3797
+ result = await runOnce(task);
3798
+ session.history = result.messages;
3799
+ if (verify && result.reason === "finished" && !controller.signal.aborted) {
3800
+ result = await runVerification(runOnce, workspace, session, spinner3, json, controller.signal, result);
3801
+ }
3622
3802
  } finally {
3623
3803
  spinner3.stop();
3624
3804
  cancel2.dispose();
3625
3805
  }
3626
- session.history = result.messages;
3806
+ const runCost = pricing ? estimateCost(result.usage, pricing) : 0;
3807
+ session.costUsd += runCost;
3808
+ await recordUsage({
3809
+ ts: (/* @__PURE__ */ new Date()).toISOString(),
3810
+ agent: resolved.config.name,
3811
+ provider: resolved.config.provider,
3812
+ model: resolved.config.model,
3813
+ promptTokens: result.usage.promptTokens,
3814
+ completionTokens: result.usage.completionTokens,
3815
+ costUsd: runCost
3816
+ });
3627
3817
  if (collector) {
3628
3818
  process.stdout.write(JSON.stringify(collector.build(result)) + "\n");
3629
3819
  return;
3630
3820
  }
3821
+ if (budgetHit) {
3822
+ console.log(pc8.yellow("\n" + t("budget.hit", { budget: fmtUsd(session.budget ?? 0) })));
3823
+ }
3631
3824
  if (result.reason === "finished") {
3632
3825
  console.log(pc8.green("\n" + t("run.done", { steps: result.steps })) + (result.summary ? ` ${result.summary}` : ""));
3633
3826
  } else if (result.reason === "cancelled") {
@@ -3637,15 +3830,13 @@ async function executeTask(task, resolved, workspace, session, json = false) {
3637
3830
  }
3638
3831
  if (result.usage.promptTokens || result.usage.completionTokens) {
3639
3832
  const total = result.usage.promptTokens + result.usage.completionTokens;
3640
- console.log(
3641
- pc8.dim(
3642
- "\u21B3 " + t("ui.tokens", {
3643
- total: fmtTokens(total),
3644
- in: fmtTokens(result.usage.promptTokens),
3645
- out: fmtTokens(result.usage.completionTokens)
3646
- })
3647
- )
3648
- );
3833
+ const tokensLine = t("ui.tokens", {
3834
+ total: fmtTokens(total),
3835
+ in: fmtTokens(result.usage.promptTokens),
3836
+ out: fmtTokens(result.usage.completionTokens)
3837
+ });
3838
+ const cost = pricing ? ` \xB7 ~${fmtUsd(runCost)}` : "";
3839
+ console.log(pc8.dim("\u21B3 " + tokensLine + cost));
3649
3840
  }
3650
3841
  }
3651
3842
  function fmtTokens(n) {
@@ -3702,6 +3893,36 @@ async function confirmAction(req) {
3702
3893
  if (p2.isCancel(answer)) return false;
3703
3894
  return answer === true;
3704
3895
  }
3896
+ async function runVerification(runOnce, workspace, session, spinner3, json, signal, initial) {
3897
+ const checks = await detectChecks(workspace);
3898
+ if (checks.length === 0) {
3899
+ if (!json) console.log(pc8.dim(t("verify.noChecks")));
3900
+ return initial;
3901
+ }
3902
+ let result = initial;
3903
+ for (let fix = 0; ; fix++) {
3904
+ if (signal.aborted) return result;
3905
+ if (!json) spinner3.start(t("verify.running"));
3906
+ const results = await runChecks(workspace, checks);
3907
+ spinner3.stop();
3908
+ const failed = results.filter((r) => !r.ok);
3909
+ if (failed.length === 0) {
3910
+ if (!json) console.log(pc8.green("\u2713 " + t("verify.passed")));
3911
+ return result;
3912
+ }
3913
+ if (fix >= MAX_VERIFY_FIXES) {
3914
+ if (!json) console.log(pc8.yellow("\u26A0 " + t("verify.giveUp", { n: failed.length })));
3915
+ return result;
3916
+ }
3917
+ if (!json) {
3918
+ console.log(pc8.yellow("\u2717 " + t("verify.failed", { n: failed.length, attempt: fix + 1 })));
3919
+ }
3920
+ if (!json) spinner3.start(t("ui.thinking"));
3921
+ result = await runOnce(buildVerifyFeedback(failed));
3922
+ spinner3.stop();
3923
+ session.history = result.messages;
3924
+ }
3925
+ }
3705
3926
  function renderDiff(hunks) {
3706
3927
  for (const h of hunks) {
3707
3928
  console.log(pc8.cyan(`@@ -${h.oldStart + 1},${h.oldCount} +${h.newStart + 1},${h.newCount} @@`));
@@ -3717,8 +3938,8 @@ function renderEvents(spinner3) {
3717
3938
  onStep() {
3718
3939
  spinner3.start(t("ui.thinking"));
3719
3940
  },
3720
- onUsage(usage) {
3721
- const total = usage.promptTokens + usage.completionTokens;
3941
+ onUsage(usage2) {
3942
+ const total = usage2.promptTokens + usage2.completionTokens;
3722
3943
  if (total > 0) spinner3.setSuffix(t("ui.tokensShort", { total: fmtTokens(total) }));
3723
3944
  },
3724
3945
  onAssistantText(text2) {
@@ -3756,8 +3977,8 @@ async function setup() {
3756
3977
  import pc9 from "picocolors";
3757
3978
 
3758
3979
  // src/core/scaffold/init.ts
3759
- import { mkdir as mkdir3, writeFile as writeFile4, access } from "fs/promises";
3760
- import { dirname as dirname3, join as join5 } from "path";
3980
+ import { mkdir as mkdir4, writeFile as writeFile4, access } from "fs/promises";
3981
+ import { dirname as dirname3, join as join6 } from "path";
3761
3982
 
3762
3983
  // src/core/scaffold/templates.ts
3763
3984
  function polyTemplates(locale) {
@@ -4000,12 +4221,12 @@ async function scaffoldPoly(workspace, opts) {
4000
4221
  const skipped = [];
4001
4222
  for (const [rel, content] of Object.entries(templates)) {
4002
4223
  const display = `.poly/${rel}`;
4003
- const abs = join5(workspace, ".poly", ...rel.split("/"));
4224
+ const abs = join6(workspace, ".poly", ...rel.split("/"));
4004
4225
  if (!opts.force && await exists(abs)) {
4005
4226
  skipped.push(display);
4006
4227
  continue;
4007
4228
  }
4008
- await mkdir3(dirname3(abs), { recursive: true });
4229
+ await mkdir4(dirname3(abs), { recursive: true });
4009
4230
  await writeFile4(abs, content, "utf8");
4010
4231
  created.push(display);
4011
4232
  }
@@ -4099,11 +4320,32 @@ async function resolveOpenRouterKey() {
4099
4320
  }
4100
4321
  }
4101
4322
 
4323
+ // src/cli/commands/usage.ts
4324
+ import pc11 from "picocolors";
4325
+ async function usage() {
4326
+ const { days, total } = await aggregateUsage();
4327
+ if (days.length === 0) {
4328
+ console.log(pc11.yellow(t("usage.empty")));
4329
+ return;
4330
+ }
4331
+ console.log(pc11.bold(t("usage.header")));
4332
+ for (const d of days) console.log(" " + formatRow(d));
4333
+ console.log(pc11.dim(" " + "\u2500".repeat(40)));
4334
+ console.log(" " + pc11.bold(formatRow({ ...total, date: t("usage.total") })));
4335
+ }
4336
+ function formatRow(b) {
4337
+ const tokens2 = fmtTokens2(b.promptTokens + b.completionTokens);
4338
+ return `${b.date.padEnd(12)} ${tokens2.padStart(7)} tok ${fmtUsd(b.costUsd).padStart(10)} (${b.runs} ${t("usage.runs")})`;
4339
+ }
4340
+ function fmtTokens2(n) {
4341
+ return n >= 1e3 ? `${(n / 1e3).toFixed(1)}k` : String(n);
4342
+ }
4343
+
4102
4344
  // src/cli/commands/prd.ts
4103
- import { writeFile as writeFile5, readFile as readFile9 } from "fs/promises";
4345
+ import { writeFile as writeFile5, readFile as readFile11 } from "fs/promises";
4104
4346
  import { execFile } from "child_process";
4105
- import { promisify as promisify2 } from "util";
4106
- import pc11 from "picocolors";
4347
+ import { promisify as promisify3 } from "util";
4348
+ import pc12 from "picocolors";
4107
4349
 
4108
4350
  // src/core/agent/prd.ts
4109
4351
  var SYSTEM = [
@@ -4192,13 +4434,13 @@ async function withRetry(fn, opts = {}) {
4192
4434
 
4193
4435
  // src/cli/commands/cli-io.ts
4194
4436
  import { readFileSync, existsSync as existsSync2 } from "fs";
4195
- import { resolve as resolve10 } from "path";
4437
+ import { resolve as resolve11 } from "path";
4196
4438
  var GUIDE_MAX = 12e3;
4197
4439
  function readProjectGuide(files) {
4198
4440
  const parts = [];
4199
4441
  for (const file of files) {
4200
4442
  try {
4201
- const path = resolve10(process.cwd(), file);
4443
+ const path = resolve11(process.cwd(), file);
4202
4444
  if (existsSync2(path)) parts.push(`# ${file}
4203
4445
  ${readFileSync(path, "utf8").trim()}`);
4204
4446
  } catch {
@@ -4224,7 +4466,7 @@ function stripBom(s) {
4224
4466
  }
4225
4467
 
4226
4468
  // src/cli/commands/prd.ts
4227
- var exec2 = promisify2(execFile);
4469
+ var exec3 = promisify3(execFile);
4228
4470
  async function prd(issueRef, opts) {
4229
4471
  const issue = await loadIssue(issueRef, opts.input);
4230
4472
  const { provider } = resolveFreeProvider(opts.model ?? DEFAULT_PRD_MODEL);
@@ -4232,18 +4474,18 @@ async function prd(issueRef, opts) {
4232
4474
  const markdown = await withRetry(() => generatePrd(issue, provider, guide));
4233
4475
  if (opts.out) {
4234
4476
  await writeFile5(opts.out, markdown + "\n", "utf8");
4235
- console.error(pc11.green(t("prd.wrote", { path: opts.out })));
4477
+ console.error(pc12.green(t("prd.wrote", { path: opts.out })));
4236
4478
  } else {
4237
4479
  process.stdout.write(markdown + "\n");
4238
4480
  }
4239
4481
  }
4240
4482
  async function loadIssue(issueRef, input) {
4241
4483
  if (input) {
4242
- const raw = input === "-" ? await readStdin() : await readFile9(input, "utf8");
4484
+ const raw = input === "-" ? await readStdin() : await readFile11(input, "utf8");
4243
4485
  return normalize2(JSON.parse(stripBom(raw)));
4244
4486
  }
4245
4487
  const num = numericRef(issueRef);
4246
- const { stdout: stdout2 } = await exec2("gh", ["issue", "view", num, "--json", "number,title,body,comments"]);
4488
+ const { stdout: stdout2 } = await exec3("gh", ["issue", "view", num, "--json", "number,title,body,comments"]);
4247
4489
  const data = normalize2(JSON.parse(stdout2));
4248
4490
  data.number ??= Number(num);
4249
4491
  return data;
@@ -4258,10 +4500,10 @@ function normalize2(raw) {
4258
4500
  }
4259
4501
 
4260
4502
  // src/cli/commands/review.ts
4261
- import { writeFile as writeFile6, readFile as readFile10 } from "fs/promises";
4503
+ import { writeFile as writeFile6, readFile as readFile12 } from "fs/promises";
4262
4504
  import { execFile as execFile2 } from "child_process";
4263
- import { promisify as promisify3 } from "util";
4264
- import pc12 from "picocolors";
4505
+ import { promisify as promisify4 } from "util";
4506
+ import pc13 from "picocolors";
4265
4507
 
4266
4508
  // src/core/agent/review.ts
4267
4509
  var MAX_DIFF_CHARS = Number(process.env.POLYPUS_MAX_DIFF_CHARS) || 6e4;
@@ -4318,7 +4560,7 @@ ${projectGuide}`
4318
4560
  }
4319
4561
 
4320
4562
  // src/cli/commands/review.ts
4321
- var exec3 = promisify3(execFile2);
4563
+ var exec4 = promisify4(execFile2);
4322
4564
  async function review(prRef, opts) {
4323
4565
  const num = opts.input ? prRef.replace(/^#/, "") : numericRef(prRef);
4324
4566
  const diff = await loadDiff(num, opts.input);
@@ -4328,25 +4570,25 @@ async function review(prRef, opts) {
4328
4570
  const markdown = await withRetry(() => reviewDiff(diff, meta, provider, guide));
4329
4571
  if (opts.out) {
4330
4572
  await writeFile6(opts.out, markdown + "\n", "utf8");
4331
- console.error(pc12.green(t("review.wrote", { path: opts.out })));
4573
+ console.error(pc13.green(t("review.wrote", { path: opts.out })));
4332
4574
  } else {
4333
4575
  process.stdout.write(markdown + "\n");
4334
4576
  }
4335
4577
  }
4336
4578
  async function loadDiff(num, input) {
4337
- if (input) return input === "-" ? readStdin() : readFile10(input, "utf8");
4338
- const { stdout: stdout2 } = await exec3("gh", ["pr", "diff", num]);
4579
+ if (input) return input === "-" ? readStdin() : readFile12(input, "utf8");
4580
+ const { stdout: stdout2 } = await exec4("gh", ["pr", "diff", num]);
4339
4581
  return stdout2;
4340
4582
  }
4341
4583
  async function loadMeta(num, input) {
4342
4584
  if (input) return { number: Number(num) || void 0, title: `PR ${num}`, body: "" };
4343
- const { stdout: stdout2 } = await exec3("gh", ["pr", "view", num, "--json", "number,title,body"]);
4585
+ const { stdout: stdout2 } = await exec4("gh", ["pr", "view", num, "--json", "number,title,body"]);
4344
4586
  const raw = JSON.parse(stdout2);
4345
4587
  return { number: raw.number, title: raw.title ?? "", body: raw.body ?? "" };
4346
4588
  }
4347
4589
 
4348
4590
  // src/cli/index.ts
4349
- import { join as join6 } from "path";
4591
+ import { join as join7 } from "path";
4350
4592
 
4351
4593
  // src/core/config/dotenv.ts
4352
4594
  import { existsSync as existsSync3, readFileSync as readFileSync2 } from "fs";
@@ -4379,7 +4621,7 @@ async function launchInteractive() {
4379
4621
  const config = await loadConfig();
4380
4622
  if (config.agents.length === 0) {
4381
4623
  console.log(banner());
4382
- console.log(" " + pc13.yellow(t("welcome.firstRun")) + "\n");
4624
+ console.log(" " + pc14.yellow(t("welcome.firstRun")) + "\n");
4383
4625
  await setup();
4384
4626
  }
4385
4627
  await run(void 0, {});
@@ -4406,20 +4648,21 @@ function buildProgram() {
4406
4648
  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));
4407
4649
  program.command("remove-agent").argument("<name>", t("cli.arg.removeAgentName")).description(t("cli.cmd.removeAgent")).action((name) => removeAgent(name));
4408
4650
  program.command("list-agents").alias("agents").description(t("cli.cmd.listAgents")).action(() => listAgents());
4409
- 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));
4651
+ 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")).option("--budget <usd>", t("cli.opt.budget")).description(t("cli.cmd.run")).action((task, opts) => run(task, opts));
4410
4652
  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));
4411
4653
  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));
4654
+ program.command("usage").description(t("cli.cmd.usage")).action(() => usage());
4412
4655
  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));
4413
4656
  program.command("review").argument("<pr>", t("cli.arg.reviewPr")).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.review")).action((pr, opts) => review(pr, opts));
4414
4657
  return program;
4415
4658
  }
4416
4659
  async function main() {
4417
4660
  try {
4418
- loadDotenv([join6(configDir(), ".env"), join6(process.cwd(), ".env")]);
4661
+ loadDotenv([join7(configDir(), ".env"), join7(process.cwd(), ".env")]);
4419
4662
  await resolveLocale();
4420
4663
  await buildProgram().parseAsync(process.argv);
4421
4664
  } catch (err) {
4422
- console.error(pc13.red(`\u2717 ${err.message}`));
4665
+ console.error(pc14.red(`\u2717 ${err.message}`));
4423
4666
  process.exitCode = 1;
4424
4667
  }
4425
4668
  }