@garuda-agent/garuda 0.16.1 → 0.17.0

package/dist/cli/index.js

@@ -15,7 +15,7 @@ import {
   runCommand,
   runTurnInTerminal,
   stopMessage
-} from "../chunk-BQTUOO4G.js";
+} from "../chunk-Z66BWTMD.js";
 import {
   PlainRenderer
 } from "../chunk-FARS4R67.js";
@@ -32,24 +32,27 @@ import {
 } from "../chunk-4JSUUDFQ.js";
 import {
   DEFAULT_MAX_TOKENS,
-  FileSessionStore,
   Runtime,
   ToolRegistry,
   auditDirFor,
-  closeOpenToolCalls,
-  createSession,
   defaultTools,
-  parseRecords,
-  popPoint,
-  pushPoint,
   rebuildState,
-  redoPoint,
   runAgent,
   signature
-} from "../chunk-TILFPYWO.js";
+} from "../chunk-2DYMJMRQ.js";
 import "../chunk-G2AX7MGS.js";
 import "../chunk-A2QAIRQL.js";
-import "../chunk-ULAJPMY7.js";
+import {
+  FileSessionStore,
+  closeOpenToolCalls,
+  createSession,
+  isCostIncomplete,
+  parseRecords,
+  pendingChildTokens,
+  popPoint,
+  pushPoint,
+  redoPoint
+} from "../chunk-735BKABA.js";
 import {
   NO_MODEL,
   chooseModel,
@@ -57,9 +60,11 @@ import {
   loadModelsConfig,
   providerKeySource,
   resolveModel
-} from "../chunk-NP6RIC25.js";
-import "../chunk-EIXUSVFX.js";
-import "../chunk-R7I2NBPX.js";
+} from "../chunk-JRVWH27M.js";
+import {
+  loadCatalog
+} from "../chunk-U65DQZDB.js";
+import "../chunk-J4346IDH.js";
 import {
   loadSearchConfig
 } from "../chunk-B25VL4RU.js";
@@ -69,19 +74,20 @@ import {
 import "../chunk-EU7PA45O.js";
 import {
   VERSION
-} from "../chunk-VBZCNMUD.js";
+} from "../chunk-BSK5T2ZG.js";
 import "../chunk-5DEWP3YX.js";
 import "../chunk-ZAE25JM2.js";
 import "../chunk-ATW7BNIG.js";
 import "../chunk-PZS3V6P3.js";
-import "../chunk-OUCV3FCT.js";
+import "../chunk-AOGFYUKG.js";
 import {
   JOB_DENIAL
-} from "../chunk-GXZACJUI.js";
-import "../chunk-WMOLXTBU.js";
-import "../chunk-3EQU2GUX.js";
-import "../chunk-HUIK3WXO.js";
-import "../chunk-KKB5W47S.js";
+} from "../chunk-B4SJY7KT.js";
+import "../chunk-BHZ6G65V.js";
+import "../chunk-P6XWWIEK.js";
+import "../chunk-VTMD5IHD.js";
+import "../chunk-UIRJOQQS.js";
+import "../chunk-4LFJFDCF.js";
 import "../chunk-3NJNVZTK.js";
 
 // src/cli/index.ts
@@ -254,9 +260,41 @@ var NO_PERMISSIONS = {
     throw new Error("Replay runs no commands.");
   },
   deniedByPolicy: () => true,
+  writeScopeDenial: async () => "replay: no writes",
   commandPolicyDenial: () => "Replay runs no commands."
 };
 
+// src/nodeFloor.ts
+var NODE_SECURITY = {
+  date: "2026-10-21",
+  advisory: "https://nodejs.org/en/blog/vulnerability/october-2026-security-releases",
+  lines: { 22: "22.23.4", 24: "24.22.1", 26: "26.12.1" }
+};
+function nodeProblem(version = process.versions.node) {
+  const [major = 0] = parse(version);
+  const lines = Object.keys(NODE_SECURITY.lines).map(Number);
+  const floor = NODE_SECURITY.lines[major];
+  const use = `Use Node ${lines.map((l) => NODE_SECURITY.lines[l]).join(", ")} or later.`;
+  if (floor === void 0) {
+    if (major > Math.max(...lines)) return void 0;
+    return `Node ${version} gets no security fixes (Garuda supports ${lines.join(", ")}). ${use}`;
+  }
+  if (compare(parse(version), parse(floor)) < 0) {
+    return `Node ${version} does not have the security fixes of ${NODE_SECURITY.date} (${NODE_SECURITY.advisory}). Update to ${floor} or later.`;
+  }
+  return void 0;
+}
+function parse(version) {
+  return version.replace(/^v/, "").split(/[.-]/).slice(0, 3).map((n) => Number.parseInt(n, 10) || 0);
+}
+function compare(a, b) {
+  for (let i = 0; i < 3; i++) {
+    const d = (a[i] ?? 0) - (b[i] ?? 0);
+    if (d !== 0) return d;
+  }
+  return 0;
+}
+
 // src/cli/banner.ts
 import { homedir } from "os";
 var WORDMARK = [
@@ -516,7 +554,8 @@ function assistantLine(response, model) {
       }),
       stop_reason: response.stopReason,
       stop_sequence: null,
-      usage: apiUsage(response.usage)
+      usage: apiUsage(response.usage),
+      ...response.usageReported === false ? { usage_is_estimated: true } : {}
     },
     parent_tool_use_id: null
   };
@@ -564,7 +603,12 @@ function resultLine(outcome, info) {
     ...subtype === "success" ? { result: info.lastText } : {},
     stop_reason: result?.modelStopReason ?? null,
     total_cost_usd: info.totalCostUsd,
+    ...info.totalCostUnknown === true ? { total_cost_is_unknown: true } : {},
+    ...info.totalCostIncomplete === true ? { total_cost_is_incomplete: true } : {},
     usage: apiUsage(usage),
+    ...(result?.unreportedTokens ?? 0) > 0 ? { usage_is_incomplete: true, unreported_budget_tokens: result?.unreportedTokens } : {},
+    ...info.runCostUsd === void 0 ? { cost_is_unknown: true } : {},
+    ...info.runCostUsd !== void 0 && info.runCostIncomplete === true ? { cost_is_incomplete: true } : {},
     modelUsage: {
       [info.model]: {
         inputTokens: usage.inputTokens,
@@ -729,6 +773,9 @@ function askTerminal(prompt, history, io, completer) {
 
 // src/cli/index.ts
 async function main() {
+  const nodeWarning = nodeProblem();
+  if (nodeWarning !== void 0) process.stderr.write(`Warning: ${nodeWarning}
+`);
   const program = new Command().enablePositionalOptions().name("garuda").description("A terminal coding agent. With no task, it starts a chat in the current folder.").version(VERSION).option("-p, --prompt <task>", "run one task and exit").option("-m, --model <id>", "model id (or set GARUDA_MODEL)").option(
     "--subagent-model <id>",
     "model of the explore subagent (or set GARUDA_SUBAGENT_MODEL); default: the main model"
@@ -772,7 +819,7 @@ async function main() {
     "--test-command <cmd>",
     "with --from-git: the test command; {files} takes the test files (default: detected)"
   ).action(async (options) => {
-    const { runEvalCommand } = await import("../evalCommand-RGMZRYDR.js");
+    const { runEvalCommand } = await import("../evalCommand-YVQAWUZE.js");
     process.exitCode = await runEvalCommand({
       ...options,
       model: options.model ?? process.env.GARUDA_MODEL,
@@ -794,30 +841,30 @@ async function main() {
   program.command("night").description(
     "run this project's night queue (jobs from /schedule), up to 3 at a time, then write one digest"
   ).option("--at <HH:MM>", "wait until this local time first (keep the terminal open)").option("--parallel <n>", "jobs at the same time (1 to 10, default 3)").addOption(new Option("--from-launchd").hideHelp()).action(async (options) => {
-    const { nightCommand } = await import("../nightCommand-65MUEV32.js");
+    const { nightCommand } = await import("../nightCommand-OODJNWDT.js");
     process.exitCode = await nightCommand(options, new PlainRenderer());
   });
   program.command("setup").description(
     "choose a provider and model and store its key (~/.garuda/credentials, only you can read it); editors run it to sign in"
   ).option("--show", "show the model and where each key comes from (keys masked)").option("--forget [key-name]", "remove a stored key, or all of them").action(async (options) => {
-    const { setupCommand, terminalIO } = await import("../setupCommand-VA7D5MZK.js");
+    const { setupCommand, terminalIO } = await import("../setupCommand-GBQIVYJC.js");
     process.exitCode = await setupCommand(options, { io: await terminalIO() });
   });
   program.command("acp").description(
     "run as an agent for an editor over the Agent Client Protocol (VS Code with an extension, Zed, JetBrains, Neovim, Emacs); the editor starts it"
   ).argument("[action]", "setup: run garuda setup (the editor's sign-in, ACP Terminal Auth)").option("-m, --model <id>", "model id (or set GARUDA_MODEL)").action(async (action, options) => {
-    const { acpCommand } = await import("../acpCommand-OTAYW7BS.js");
+    const { acpCommand } = await import("../acpCommand-7EY3NHPG.js");
     process.exitCode = await acpCommand({
       ...options,
       ...action === void 0 ? {} : { action }
     });
   });
   const lsp = program.command("lsp").description("show the language servers for diagnostics (TS/JS, Python, Java)").action(async () => {
-    const { lspStatusCommand } = await import("../lspCommand-JRXKRSL7.js");
+    const { lspStatusCommand } = await import("../lspCommand-ZZ6R4PA2.js");
     process.exitCode = await lspStatusCommand();
   });
   lsp.command("install").description("install a pinned language server into ~/.garuda/lsp (needs the network)").argument("<language>", "typescript, python or java").action(async (language) => {
-    const { lspInstallCommand } = await import("../lspCommand-JRXKRSL7.js");
+    const { lspInstallCommand } = await import("../lspCommand-ZZ6R4PA2.js");
     process.exitCode = await lspInstallCommand(language);
   });
   await program.parseAsync();
@@ -829,7 +876,7 @@ async function start(options, program) {
   let job;
   let jobGuard;
   if (options.job !== void 0) {
-    const { prepareJob } = await import("../jobCommand-ACSGSBQ4.js");
+    const { prepareJob } = await import("../jobCommand-FE6XN3YD.js");
     const prepared = await prepareJob(
       mainRoot,
       options.job,
@@ -841,7 +888,7 @@ async function start(options, program) {
     );
     if (typeof prepared === "number") return prepared;
     job = prepared;
-    const { failJobSync } = await import("../job-3NYDMP2M.js");
+    const { failJobSync } = await import("../job-7JOI4YIQ.js");
     const preparedJob = prepared;
     jobGuard = () => {
       if (preparedJob.job.status === "running") {
@@ -883,13 +930,14 @@ async function start(options, program) {
     }
   }
   const searchConfig = await loadSearchConfig();
+  const catalog = await loadCatalog();
   const credentialsWarning = await keepProviderKeys([
     resolved.def.apiKeyEnv,
     sub?.def.apiKeyEnv,
     "ANTHROPIC_API_KEY"
   ]);
   const anthropicKey = providerKeySource("ANTHROPIC_API_KEY");
-  const { loadTeamPolicy } = await import("../policy-EJPLEEXP.js");
+  const { loadTeamPolicy } = await import("../policy-ZJ7RMOFN.js");
   let team;
   try {
     team = await loadTeamPolicy();
@@ -965,7 +1013,9 @@ async function start(options, program) {
           ...r.maxTokens === void 0 ? {} : { maxTokens: r.maxTokens }
         };
       },
-      configured: Object.keys(models.config.models)
+      configured: Object.keys(models.config.models),
+      // The account's model list from `garuda setup` (0.17); no network call here.
+      ...catalog === void 0 ? {} : { catalog }
     },
     ...options.resume === void 0 ? {} : { resume: options.resume },
     onEvent: (event) => events.event(event),
@@ -982,11 +1032,11 @@ async function start(options, program) {
       "A job needs the OS sandbox (Seatbelt or bubblewrap): without it, every command would need an approval, and nobody is there to give it."
     );
     await runtime.close();
-    const { saveJob } = await import("../job-3NYDMP2M.js");
+    const { saveJob } = await import("../job-7JOI4YIQ.js");
     job.job.status = "failed";
     await saveJob(job.job);
     if (options.fromLaunchd === true || job.job.launchd !== void 0) {
-      const { removeAgent, defaultAgentEnv } = await import("../launchd-SSX362VG.js");
+      const { removeAgent, defaultAgentEnv } = await import("../launchd-3A3ONEW5.js");
       await removeAgent(runtime.executor, job.job, defaultAgentEnv());
     }
     return 1;
@@ -1036,11 +1086,19 @@ async function start(options, program) {
     });
     if (output !== void 0) events = output;
     const costBefore = runtime.session?.costUsd ?? 0;
+    const unreportedBefore = runtime.session?.unreportedTokens ?? 0;
+    const incompleteBefore = runtime.session?.costIncomplete ?? false;
+    const pendingBefore = runtime.session === void 0 ? 0 : pendingChildTokens(runtime.session);
     const finish = (outcome2) => {
-      const cost = runtime.session?.costUsd;
+      const session2 = runtime.session;
+      const cost = session2?.costUsd;
+      const runIncomplete = session2 !== void 0 && (session2.unreportedTokens > unreportedBefore || pendingChildTokens(session2) > pendingBefore || session2.costIncomplete && !incompleteBefore);
       output?.finish(outcome2, {
         totalCostUsd: cost ?? 0,
+        ...cost === void 0 ? { totalCostUnknown: true } : {},
+        ...session2 !== void 0 && isCostIncomplete(session2) ? { totalCostIncomplete: true } : {},
         runCostUsd: cost === void 0 ? void 0 : cost - costBefore,
+        ...runIncomplete ? { runCostIncomplete: true } : {},
         contextWindow: runtime.limits.contextWindow,
         maxOutputTokens: resolved.maxTokens ?? DEFAULT_MAX_TOKENS,
         ...outcome2.kind === "done" ? { stopMessage: stopMessage(outcome2.result.stopReason, runtime.limits) } : {}
@@ -1057,7 +1115,7 @@ async function start(options, program) {
       if (resolved2.kind === "prompt") prompt = resolved2.prompt;
     }
     if (job !== void 0) {
-      const { testsBefore } = await import("../jobCommand-ACSGSBQ4.js");
+      const { testsBefore } = await import("../jobCommand-FE6XN3YD.js");
       await testsBefore(job, runtime, renderer);
     }
     const outcome = await runTurnInTerminal(
@@ -1070,7 +1128,7 @@ async function start(options, program) {
     await runtime.close();
     finish(outcome);
     if (job !== void 0) {
-      const { finishJob } = await import("../jobCommand-ACSGSBQ4.js");
+      const { finishJob } = await import("../jobCommand-FE6XN3YD.js");
       await finishJob(job, outcome, runtime, renderer, void 0, {
         fromLaunchd: options.fromLaunchd === true
       });
@@ -1083,7 +1141,7 @@ async function start(options, program) {
   const bannerInfo = {
     version: VERSION,
     model: modelId,
-    sandbox: runtime.executor.isolation === "none" ? "none: each command asks first" : `${runtime.executor.name} \xB7 no network`,
+    sandbox: runtime.executor.isolation === "none" ? "none: each command asks first, unless a rule allows it" : `${runtime.executor.name} \xB7 no network`,
     root,
     extras: runtime.extras(),
     subagents: runtime.subagentsSummary(),
@@ -1156,7 +1214,7 @@ function wantsInk() {
 }
 async function loadInk(renderer) {
   try {
-    return await import("../inkChat-WEDGQMCY.js");
+    return await import("../inkChat-FNUBK43U.js");
   } catch (error) {
     renderer.info(`Plain chat (Ink is not available: ${error.message.split("\n")[0]}).`);
     return void 0;
@@ -1187,9 +1245,9 @@ function recordJobDenial(job, tool, target) {
   job.denied.push({ tool, target: shown });
 }
 async function jobBatchClient(job, resolved, renderer) {
-  const { DeadlineClient } = await import("../deadline-F5RJJ2QW.js");
-  const { nextTime } = await import("../launchd-SSX362VG.js");
-  const { DEFAULT_FINISH_BY, DEFAULT_STEP_LIMIT_MINUTES, SWITCH_BEFORE_MS } = await import("../create-VXIH2GUZ.js");
+  const { DeadlineClient } = await import("../deadline-TP7ZGLT2.js");
+  const { nextTime } = await import("../launchd-3A3ONEW5.js");
+  const { DEFAULT_FINISH_BY, DEFAULT_STEP_LIMIT_MINUTES, SWITCH_BEFORE_MS } = await import("../create-W6BV75QX.js");
   const finishBy = job.job.finishBy ?? DEFAULT_FINISH_BY;
   const stepLimit = job.job.stepLimitMinutes ?? DEFAULT_STEP_LIMIT_MINUTES;
   const switchAt = new Date(nextTime(finishBy).getTime() - SWITCH_BEFORE_MS);

package/dist/{create-VXIH2GUZ.js → create-W6BV75QX.js}

@@ -3,7 +3,7 @@ import {
   jobBase,
   linkCandidates,
   worktreeDir
-} from "./chunk-44QN5SWZ.js";
+} from "./chunk-55CFEEIN.js";
 import {
   agentLabel,
   defaultAgentEnv,
@@ -12,18 +12,17 @@ import {
   nextTime,
   nightAgentTime,
   nightSpec
-} from "./chunk-USCUXCAU.js";
+} from "./chunk-5B5HDTUR.js";
 import {
   jobPrompt
-} from "./chunk-YQTGMCU6.js";
-import "./chunk-OUCV3FCT.js";
+} from "./chunk-3MRQJKUO.js";
+import "./chunk-AOGFYUKG.js";
 import {
   JobGitError
-} from "./chunk-TUUEBHG2.js";
-import "./chunk-GXZACJUI.js";
-import "./chunk-WMOLXTBU.js";
-import "./chunk-3EQU2GUX.js";
-import "./chunk-HUIK3WXO.js";
+} from "./chunk-6XHON3Q6.js";
+import "./chunk-B4SJY7KT.js";
+import "./chunk-BHZ6G65V.js";
+import "./chunk-P6XWWIEK.js";
 import {
   DEFAULT_JOB_MAX_STEPS,
   JOBS_DIR,
@@ -32,7 +31,11 @@ import {
   planPermissions,
   planTitle,
   saveJob
-} from "./chunk-KKB5W47S.js";
+} from "./chunk-VTMD5IHD.js";
+import {
+  proposeWriteScope
+} from "./chunk-UIRJOQQS.js";
+import "./chunk-4LFJFDCF.js";
 import "./chunk-3NJNVZTK.js";
 
 // src/jobs/create.ts
@@ -62,6 +65,7 @@ async function createJob(options) {
     throw error;
   }
   const { rules, problems } = planPermissions(options.plan.plan);
+  const proposed = proposeWriteScope(rules, options.root);
   const id = newJobId(options.now);
   const links = linkCandidates(options.root);
   const job = {
@@ -77,6 +81,7 @@ async function createJob(options) {
     planSession: options.plan.sessionId,
     model: options.modelId,
     allow: rules,
+    ...proposed.scope === void 0 ? {} : { writeScope: proposed.scope },
     onUnapproved: "deny-and-continue",
     ...options.at === void 0 ? {} : { at: options.at },
     maxSteps: options.maxSteps ?? DEFAULT_JOB_MAX_STEPS,
@@ -98,6 +103,7 @@ async function createJob(options) {
     ...problems.length === 0 ? [] : ["", "Not rules, left out:", ...problems.map((p) => `  ${p}`)],
     "",
     "Any other call that would ask is denied at run time; the job goes on and reports it.",
+    proposed.scope === void 0 ? `Write scope: the whole worktree (${proposed.reason}). Commands may write anywhere in it.` : `Write scope: commands (also the test command) and edits may change only ${proposed.scope.length === 0 ? "nothing in the worktree" : proposed.scope.join(", ")} (plus temp folders, caches and the links). The OS sandbox enforces it for commands; no command runs outside the sandbox. A test that writes outside the scope fails.`,
     job.network === void 0 ? "Network: none for commands." : `Network: ${job.network.join(", ")}, through Garuda's proxy; other hosts are denied and reported.`,
     "",
     options.test === void 0 ? 'Proof of work: no test command found (set "test" in the job file); a principal-engineer review of the diff (one model call).' : `Proof of work: \`${options.test}\` runs in the sandbox before and after the job; a principal-engineer review of the diff (one model call).`,
@@ -144,7 +150,7 @@ async function createJob(options) {
   await saveJob(job);
   const lines = [
     `Created job ${id}: ${jobPath(options.root, id)}`,
-    'You can edit its approval list ("allow") in that file before the run.'
+    'You can edit its approval list ("allow") and its write scope ("writeScope") in that file before the run.'
   ];
   const platform = options.launchd?.platform ?? process.platform;
   if (options.launchd !== void 0 && options.at !== void 0 && platform === "darwin") {
@@ -221,4 +227,4 @@ export {
   createJob,
   switchTime
 };
-//# sourceMappingURL=create-VXIH2GUZ.js.map
+//# sourceMappingURL=create-W6BV75QX.js.map

package/dist/{deadline-F5RJJ2QW.js → deadline-TP7ZGLT2.js}

@@ -41,7 +41,8 @@ var DeadlineClient = class {
     try {
       const events = [];
       for await (const event of this.primary.stream(request, {
-        signal: AbortSignal.any(signals)
+        signal: AbortSignal.any(signals),
+        ...options?.meter === void 0 ? {} : { meter: options.meter }
       })) {
         events.push(event);
       }
@@ -49,6 +50,7 @@ var DeadlineClient = class {
       yield* events;
     } catch (error) {
       if (options?.signal?.aborted === true || !timer.signal.aborted) throw error;
+      options?.meter?.retry();
       if (stepOnly) {
         this.calls.slow++;
         this.fallbackClient ??= await this.fallback();
@@ -83,4 +85,4 @@ export {
   DeadlineReached,
   StepTooSlow
 };
-//# sourceMappingURL=deadline-F5RJJ2QW.js.map
+//# sourceMappingURL=deadline-TP7ZGLT2.js.map

package/dist/{evalCommand-RGMZRYDR.js → evalCommand-YVQAWUZE.js}

@@ -5,28 +5,30 @@ import {
 import {
   CODE_INDEX_MODES,
   DEFAULT_CODE_INDEX_MODE,
-  FileSessionStore,
   Runtime,
-  detectProfiles,
-  newSessionId
-} from "./chunk-TILFPYWO.js";
+  detectProfiles
+} from "./chunk-2DYMJMRQ.js";
 import "./chunk-G2AX7MGS.js";
 import "./chunk-A2QAIRQL.js";
-import "./chunk-ULAJPMY7.js";
+import {
+  FileSessionStore,
+  isCostIncomplete,
+  newSessionId
+} from "./chunk-735BKABA.js";
 import {
   NO_MODEL,
   chooseModel,
   keepProviderKeys,
   loadModelsConfig,
   resolveModel
-} from "./chunk-NP6RIC25.js";
-import "./chunk-EIXUSVFX.js";
+} from "./chunk-JRVWH27M.js";
+import "./chunk-U65DQZDB.js";
 import {
   totalTokens
-} from "./chunk-R7I2NBPX.js";
+} from "./chunk-J4346IDH.js";
 import "./chunk-B25VL4RU.js";
 import "./chunk-23FXBNBB.js";
-import "./chunk-VBZCNMUD.js";
+import "./chunk-BSK5T2ZG.js";
 import "./chunk-5DEWP3YX.js";
 import "./chunk-ZAE25JM2.js";
 import "./chunk-ATW7BNIG.js";
@@ -34,17 +36,17 @@ import "./chunk-PZS3V6P3.js";
 import {
   LINK_CANDIDATES,
   cleanArgs
-} from "./chunk-44QN5SWZ.js";
+} from "./chunk-55CFEEIN.js";
 import {
   MANIFEST,
   TEST_PATH,
   detectTestCommand,
   tail
-} from "./chunk-OUCV3FCT.js";
+} from "./chunk-AOGFYUKG.js";
 import {
   git,
   shellWord
-} from "./chunk-TUUEBHG2.js";
+} from "./chunk-6XHON3Q6.js";
 import {
   DEFAULT_ENV_ALLOWLIST,
   EXECUTOR_NAMES,
@@ -52,10 +54,11 @@ import {
   createExecutor,
   expandAllowlist,
   parseSettings
-} from "./chunk-GXZACJUI.js";
-import "./chunk-WMOLXTBU.js";
-import "./chunk-3EQU2GUX.js";
-import "./chunk-HUIK3WXO.js";
+} from "./chunk-B4SJY7KT.js";
+import "./chunk-BHZ6G65V.js";
+import "./chunk-P6XWWIEK.js";
+import "./chunk-UIRJOQQS.js";
+import "./chunk-4LFJFDCF.js";
 import "./chunk-3NJNVZTK.js";
 
 // src/cli/evalCommand.ts
@@ -604,7 +607,9 @@ async function runEvalTask(task4, options) {
   const session = runtime.session;
   if (session !== void 0) {
     result.tokens = totalTokens(session.usage);
+    if (session.unreportedTokens > 0) result.unreportedTokens = session.unreportedTokens;
     result.costUsd = session.costUsd;
+    if (isCostIncomplete(session)) result.costIncomplete = true;
     result.cacheReadTokens = session.usage.cacheReadTokens;
     if (options.outDir !== void 0) {
       mkdirSync2(options.outDir, { recursive: true });
@@ -684,26 +689,41 @@ async function changedFiles(root, before) {
 function isErrorRun(result) {
   return result.stopReason === "error";
 }
+function summarize(results) {
+  const attempted = results.length;
+  const errors = results.filter(isErrorRun).length;
+  const scored = attempted - errors;
+  const passed = results.filter((r) => r.passed).length;
+  return {
+    attempted,
+    errors,
+    scored,
+    passed,
+    ...scored === 0 ? {} : { scoredRate: passed / scored },
+    ...attempted === 0 ? {} : { attemptedRate: passed / attempted }
+  };
+}
+var percent = (rate) => rate === void 0 ? "\u2013" : `${Math.round(rate * 100)}%`;
 function formatReport(results) {
   const rows = results.map(
     (r) => [
       r.passed ? "PASS" : isErrorRun(r) ? "ERR " : "FAIL",
       r.id.padEnd(18),
       `${String(r.steps).padStart(3)} steps`,
-      `${(r.tokens / 1e3).toFixed(1).padStart(7)}k tok`,
-      r.costUsd === void 0 ? "   cost ?" : `$${r.costUsd.toFixed(4).padStart(7)}`,
+      (r.unreportedTokens ?? 0) > 0 ? `tok ? (${r.unreportedTokens} reserved)` : `${(r.tokens / 1e3).toFixed(1).padStart(7)}k tok`,
+      r.costUsd === void 0 ? "   cost ?" : `${r.costIncomplete === true ? "\u2265" : ""}$${r.costUsd.toFixed(4).padStart(7)}`,
       `${(r.durationMs / 1e3).toFixed(0).padStart(4)} s`,
       r.stopReason
     ].join("  ")
   );
-  const passed = results.filter((r) => r.passed).length;
-  const errors = results.filter(isErrorRun).length;
-  const cost = results.every((r) => r.costUsd !== void 0) ? `$${results.reduce((s, r) => s + (r.costUsd ?? 0), 0).toFixed(4)}` : "unknown";
+  const summary = summarize(results);
+  const known = `$${results.reduce((s, r) => s + (r.costUsd ?? 0), 0).toFixed(4)}`;
+  const cost = !results.every((r) => r.costUsd !== void 0) ? "unknown" : results.some((r) => r.costIncomplete === true) ? `\u2265 ${known} (incomplete)` : known;
   const steps = results.reduce((s, r) => s + r.steps, 0);
-  const counted = `${passed}/${results.length - errors} passed${errors > 0 ? ` \xB7 ${errors} error run(s) not counted` : ""}`;
+  const counted = `${summary.passed}/${summary.scored} scored runs passed (${percent(summary.scoredRate)}) \xB7 ${summary.passed}/${summary.attempted} of all attempted runs (${percent(summary.attemptedRate)})` + (summary.errors > 0 ? ` \xB7 ${summary.errors} error run(s)` : "");
   const tokens = results.reduce((s, r) => s + r.tokens, 0);
   const cached = results.reduce((s, r) => s + (r.cacheReadTokens ?? 0), 0);
-  const share = tokens === 0 ? "" : ` \xB7 ${Math.round(cached / tokens * 100)}% of tokens from the cache`;
+  const share = results.some((r) => (r.unreportedTokens ?? 0) > 0) ? " \xB7 cache share unknown" : tokens === 0 ? "" : ` \xB7 ${Math.round(cached / tokens * 100)}% of tokens from the cache`;
   const lines4 = [...rows, "", `${counted} \xB7 ${steps} steps \xB7 cost ${cost}${share}`];
   const ids = [...new Set(results.map((r) => r.id))];
   if (ids.length < results.length) lines4.push("", "Mean per task:", ...meanRows(results, ids));
@@ -714,7 +734,8 @@ function meanRows(results, ids) {
     const all = results.filter((r) => r.id === id);
     const runs = all.filter((r) => !isErrorRun(r));
     const errors = all.length - runs.length;
-    const note = errors > 0 ? `  (${errors} error run(s) left out)` : "";
+    const passesAll = all.filter((r) => r.passed).length;
+    const note = errors > 0 ? `  (${passesAll}/${all.length} of all attempts; ${errors} error run(s))` : "";
     if (runs.length === 0) return `${"0/0".padEnd(4)}  ${id.padEnd(18)}${note}`;
     const mean = (f) => runs.reduce((s, r) => s + f(r), 0) / runs.length;
     const passes = runs.filter((r) => r.passed).length;
@@ -722,8 +743,8 @@ function meanRows(results, ids) {
       `${passes}/${runs.length}`.padEnd(4),
       id.padEnd(18),
       `${mean((r) => r.steps).toFixed(1).padStart(5)} steps`,
-      `${(mean((r) => r.tokens) / 1e3).toFixed(1).padStart(7)}k tok`,
-      runs.every((r) => r.costUsd !== void 0) ? `$${mean((r) => r.costUsd ?? 0).toFixed(4).padStart(7)}` : "   cost ?"
+      runs.some((r) => (r.unreportedTokens ?? 0) > 0) ? "    tok ?" : `${(mean((r) => r.tokens) / 1e3).toFixed(1).padStart(7)}k tok`,
+      runs.every((r) => r.costUsd !== void 0) ? `${runs.some((r) => r.costIncomplete === true) ? "\u2265" : ""}$${mean((r) => r.costUsd ?? 0).toFixed(4).padStart(7)}` : "   cost ?"
     ].join("  ") + note;
   });
 }
@@ -2542,7 +2563,7 @@ async function runEvalCommand(options) {
       return 1;
     }
   }
-  const { assertModelAllowedByPolicy, loadTeamPolicy } = await import("./policy-EJPLEEXP.js");
+  const { assertModelAllowedByPolicy, loadTeamPolicy } = await import("./policy-ZJ7RMOFN.js");
   let team;
   try {
     team = await loadTeamPolicy();
@@ -2677,7 +2698,7 @@ ${where}, in scratch folders.
 Wall time: ${Math.round(wallMs / 1e3)} s${parallel > 1 ? ` (${parallel} tasks at a time)` : ""}.`;
   writeFileSync(
     join5(outDir, "report.json"),
-    `${JSON.stringify({ model: modelId, suite, repeat, codeIndex: index, subagents: subagentsMode, subagentModel: sub?.spec, todo: todoMode, lsp: lspMode, keepThinking: thinkingMode, format: formatMode, network: network ?? "off", batch: batchMode, parallel, wallMs, executor: executor.name, results }, null, 2)}
+    `${JSON.stringify({ model: modelId, suite, repeat, codeIndex: index, subagents: subagentsMode, subagentModel: sub?.spec, todo: todoMode, lsp: lspMode, keepThinking: thinkingMode, format: formatMode, network: network ?? "off", batch: batchMode, parallel, wallMs, executor: executor.name, summary: summarize(results), results }, null, 2)}
 `
   );
   process.stdout.write(`
@@ -2809,4 +2830,4 @@ export {
   BATCH_TASK_TIMEOUT_MS,
   runEvalCommand
 };
-//# sourceMappingURL=evalCommand-RGMZRYDR.js.map
+//# sourceMappingURL=evalCommand-YVQAWUZE.js.map

package/dist/{export-HN3CT7AV.js → export-SJ3DUP6W.js}

@@ -11,7 +11,7 @@ import {
   PathOutsideRootError,
   displayPath,
   resolveInRoot
-} from "./chunk-WMOLXTBU.js";
+} from "./chunk-4LFJFDCF.js";
 
 // src/cli/export.ts
 import { mkdir, writeFile } from "fs/promises";
@@ -155,4 +155,4 @@ export {
   sessionMarkdown,
   writeExport
 };
-//# sourceMappingURL=export-HN3CT7AV.js.map
+//# sourceMappingURL=export-SJ3DUP6W.js.map