@integrity-labs/agt-cli 0.28.922 → 0.28.924

package/dist/{claude-pair-runtime-NNHCDXLG.js → claude-pair-runtime-SY5VCTCE.js}

@@ -100,7 +100,7 @@ async function spawnPairSession(session) {
     return { ok: true };
   } catch {
   }
-  const { resolveClaudeBinary } = await import("./persistent-session-OQRYZ3YP.js");
+  const { resolveClaudeBinary } = await import("./persistent-session-UEAJJ72P.js");
   const claudeBin = resolveClaudeBinary();
   const pairEnv = {
     ...process.env,
@@ -373,4 +373,4 @@ export {
   startClaudePair,
   submitClaudePairCode
 };
-//# sourceMappingURL=claude-pair-runtime-NNHCDXLG.js.map
+//# sourceMappingURL=claude-pair-runtime-SY5VCTCE.js.map

package/dist/lib/manager-worker.js

@@ -63,7 +63,7 @@ import {
   safeWriteJsonAtomic,
   setConfigHash,
   tripClass
-} from "../chunk-JTWIYGY2.js";
+} from "../chunk-FAG5UIWZ.js";
 import {
   getProjectDir as getProjectDir2,
   getReadyTasks,
@@ -136,7 +136,7 @@ import {
   takeZombieDetection,
   toOpencodeModel,
   writeEgressAllowlist
-} from "../chunk-NGJFE75J.js";
+} from "../chunk-BLEV5PAO.js";
 import {
   ACCOUNT_ENFORCEMENT_MARKER_FILENAME,
   AnchorSessionClient,
@@ -231,7 +231,7 @@ import {
   subagentActivityAgeSeconds,
   sumTranscriptUsageInWindow,
   transcriptActivityAgeSeconds
-} from "../chunk-ZOPP3ZC7.js";
+} from "../chunk-VHO7KAB7.js";
 import {
   reapOrphanChannelMcps
 } from "../chunk-XWVM4KPK.js";
@@ -3536,6 +3536,51 @@ async function maybeReconcileWorkflowRunTokens(args) {
 // src/lib/conversation-evaluator.ts
 import { readdirSync as readdirSync3, readFileSync as readFileSync12, statSync as statSync3 } from "fs";
 import { join as join17 } from "path";
+
+// src/lib/detect-meta-egress.ts
+var REPOST_PREAMBLE_WINDOW = 200;
+var REPOST_ANNOUNCE = /\b(?:re-?(?:post|send)(?:ing)?\s+(?:it|this|that)\b|(?:posting|sending)\s+(?:it|this|that)\s+(?:now|again)\b|re-?posting\s+now\b)/i;
+var REPOST_FAILURE = /\b(?:it|this|that)\s+(?:did(?:n['’]t| not)|could(?:n['’]t| not)|wasn['’]t|failed to)\s+(?:parse|send|post|go through|land|deliver|come through)\b|\b(?:did(?:n['’]t| not)|could(?:n['’]t| not)|failed to)\s+(?:send|post|deliver)\s+(?:it|this|that)\b/i;
+var REPOST_PRIOR = /\b(?:drafted|dialled|dialed|typed|composed|wrote)\s+(?:this|that|it)\s+(?:a moment ago|moments ago|just now|earlier)\b/i;
+function hasRepostPreamble(message) {
+  const head = message.slice(0, REPOST_PREAMBLE_WINDOW);
+  return REPOST_ANNOUNCE.test(head) && (REPOST_FAILURE.test(head) || REPOST_PRIOR.test(head));
+}
+var STATUS_NARRATION = [
+  // "Assessment delivered to Samuel:" - a ledger line: the noun OPENS a sentence
+  // (or parenthetical) and the recipient is a capitalised name. Case-sensitive
+  // on purpose, so "Summary delivered to the board" and "Your answer delivered
+  // to the customer" are not matched. \p{Lu} (not [A-Z]) so a non-ASCII
+  // recipient ("Élodie") is still a capitalised name.
+  new RegExp("(?:^|[.!?]\\s+|\\(\\s*)(?:[Aa]ssessment|[Ss]ummary|[Rr]eport|[Aa]nalysis|[Rr]esponse|[Rr]eply|[Uu]pdate|[Aa]nswer)\\s+delivered\\s+to\\s+\\p{Lu}", "u"),
+  // third-person "on his/her/their go" (waiting on the user's go-ahead, narrated)
+  /\bon (?:his|her|their) go\b(?![-\w])/i
+];
+var LOOP_TICK = [
+  // "Standing down - no change." / "Standing down, no action needed"
+  /\bstanding down\s*[—–,:;.-]?\s*no (?:change|changes|action needed|new work)\b/i,
+  // A message that is ONLY "Standing down."
+  /^standing down[.!]?$/i,
+  // "All clear, no pending work."
+  /\ball clear\s*[,—–:;-]?\s*no pending work\b/i
+];
+function detectMetaEgress(agentMessages) {
+  const kinds = /* @__PURE__ */ new Set();
+  for (const raw of agentMessages) {
+    if (!raw) continue;
+    const m = raw.trim();
+    if (!m) continue;
+    if (hasRepostPreamble(m)) kinds.add("repost_preamble");
+    if (STATUS_NARRATION.some((re) => re.test(m))) kinds.add("status_narration");
+    if (LOOP_TICK.some((re) => re.test(m))) kinds.add("loop_tick");
+  }
+  return {
+    meta_leak: kinds.size > 0,
+    meta_leak_kinds: [...kinds].sort()
+  };
+}
+
+// src/lib/conversation-evaluator.ts
 var MIN_CHECK_INTERVAL_MS4 = 5 * 6e4;
 var TRANSCRIPT_MTIME_WINDOW_MS3 = 7 * 24 * 60 * 60 * 1e3;
 var WINDOW_PAD_MS = 5 * 6e4;
@@ -3908,6 +3953,9 @@ async function maybeEvaluateConversations(args) {
       continue;
     }
     backendSucceeded = true;
+    const flags = detectMetaEgress(
+      turns.filter((t) => t.role === "assistant").map((t) => t.text)
+    );
     try {
       await api2.post("/host/conversations/evaluation", {
         agent_id: agentId,
@@ -3916,7 +3964,8 @@ async function maybeEvaluateConversations(args) {
         verdict: verdict.verdict,
         summary: verdict.summary,
         failure_category: verdict.failureCategory,
-        model: backend.model
+        model: backend.model,
+        eval_flags: flags
       });
       log2(
         `[conversation-eval] ${codeName}: ${conv.conversation_id.slice(0, 8)} \u2192 ${verdict.verdict} (${verdict.score})`
@@ -13138,7 +13187,7 @@ var pendingDayRolloverReset = /* @__PURE__ */ new Set();
 var dayRolloverInboundHold = /* @__PURE__ */ new Map();
 var INBOUND_HOLD_EPISODE_GAP_MS = 12e4;
 async function channelInboundActivityAgeSecondsFor(codeName) {
-  const { newestPendingInboundActivityMtimeMs } = await import("../responsiveness-probe-RBWP6HJE.js");
+  const { newestPendingInboundActivityMtimeMs } = await import("../responsiveness-probe-VXJKA6RF.js");
   const newest = newestPendingInboundActivityMtimeMs(dirname11(paneLogPath(codeName)));
   if (newest === null) return null;
   return Math.max(0, Math.floor((Date.now() - newest) / 1e3));
@@ -13823,7 +13872,7 @@ var agentRestartTimezoneInputs = /* @__PURE__ */ new Map();
 var lastVersionCheckAt = 0;
 var VERSION_CHECK_INTERVAL_MS = 5 * 60 * 1e3;
 var lastResponsivenessProbeAt = 0;
-var agtCliVersion = true ? "0.28.922" : "dev";
+var agtCliVersion = true ? "0.28.924" : "dev";
 function resolveBrewPath(execFileSync2) {
   try {
     const out = execFileSync2("which", ["brew"], { timeout: 5e3 }).toString().trim();
@@ -15413,7 +15462,7 @@ function flushRestartedAgentDiagnostics(hostId, codeNames) {
   if (codeNames.length === 0) return;
   void (async () => {
     try {
-      const { collectDiagnostics } = await import("../persistent-session-OQRYZ3YP.js");
+      const { collectDiagnostics } = await import("../persistent-session-UEAJJ72P.js");
       await api.post("/host/heartbeat", {
         host_id: hostId,
         agent_diagnostics: collectDiagnostics(codeNames, quarantineEntriesFor, claudeMdSizeFor, spawnOutcomeForDiagnostics, pidPressureFor, forwardedToolsFor)
@@ -15555,7 +15604,7 @@ async function pollCycleInner() {
     }
     try {
       const { detectHostSecurity } = await import("../host-security-6PDFG7F5.js");
-      const { collectDiagnostics } = await import("../persistent-session-OQRYZ3YP.js");
+      const { collectDiagnostics } = await import("../persistent-session-UEAJJ72P.js");
       const diagCodeNames = [...agentState.persistentSessionAgents];
       const agentDiagnostics = diagCodeNames.length > 0 ? collectDiagnostics(diagCodeNames, quarantineEntriesFor, claudeMdSizeFor, spawnOutcomeForDiagnostics, pidPressureFor, forwardedToolsFor) : void 0;
       let tailscaleHostname;
@@ -15730,7 +15779,7 @@ async function pollCycleInner() {
         collectPanelessActivityProbes,
         getResponsivenessIntervalMs,
         occupancyQualificationClassifications
-      } = await import("../responsiveness-probe-RBWP6HJE.js");
+      } = await import("../responsiveness-probe-VXJKA6RF.js");
       const probeIntervalMs = getResponsivenessIntervalMs();
       if (now - lastResponsivenessProbeAt > probeIntervalMs) {
         const probeCodeNames = [...agentState.persistentSessionAgents];
@@ -15847,7 +15896,7 @@ async function pollCycleInner() {
           collectResponsivenessProbes,
           livePendingInboundOldestAgeSeconds,
           parkPendingInbound
-        } = await import("../responsiveness-probe-RBWP6HJE.js");
+        } = await import("../responsiveness-probe-VXJKA6RF.js");
         const { getProjectDir: wedgeProjectDir } = await import("../scheduler-engine-NDP36U7O.js");
         const wedgeNow = /* @__PURE__ */ new Date();
         const liveAgents = agentState.persistentSessionAgents;
@@ -16003,7 +16052,7 @@ async function pollCycleInner() {
     }
     try {
       const { scrapeMcpFailedBannerCount } = await import("../pane-mcp-banner-scraper-JA437JIB.js");
-      const { probeSessionAuth } = await import("../session-auth-dead-4KUPTD7M.js");
+      const { probeSessionAuth } = await import("../session-auth-dead-XOP4XYBD.js");
       const observations = [];
       const pendingCacheCommits = [];
       const modelApiErrorReportingOn = hostFlagStore().getBoolean("model-api-error-reporting");
@@ -20267,7 +20316,7 @@ async function handleRestartDoorbell(agentId, requestedAt, restartReason) {
       try {
         const freshId = rotateSessionForWedge(codeName);
         consecutiveWedgeCycles.delete(codeName);
-        const { parkPendingInbound: parkForRestart } = await import("../responsiveness-probe-RBWP6HJE.js");
+        const { parkPendingInbound: parkForRestart } = await import("../responsiveness-probe-VXJKA6RF.js");
         const { parked, deadLettered } = parkForRestart(codeName, /* @__PURE__ */ new Date());
         const parkNote = parked > 0 ? `, ${parked} inbound parked` : "";
         const deadNote = deadLettered > 0 ? `, ${deadLettered} undeliverable dead-lettered` : "";
@@ -20310,7 +20359,7 @@ async function handleRestartDoorbell(agentId, requestedAt, restartReason) {
     void api.post("/host/restart-ack", { host_id: hostId, agent_id: agentId, restart_requested_at: requestedAt }).catch((err) => log(`[restart-lane] ack failed for '${codeName}': ${err.message}`));
     void (async () => {
       try {
-        const { collectDiagnostics } = await import("../persistent-session-OQRYZ3YP.js");
+        const { collectDiagnostics } = await import("../persistent-session-UEAJJ72P.js");
         await api.post("/host/heartbeat", {
           host_id: hostId,
           agent_diagnostics: collectDiagnostics([codeName], quarantineEntriesFor, claudeMdSizeFor, spawnOutcomeForDiagnostics, pidPressureFor, forwardedToolsFor)
@@ -20361,7 +20410,7 @@ async function respawnAgentAfterMcpStop(codeName, reason) {
     }
     try {
       const hostId = await getHostId();
-      const { collectDiagnostics } = await import("../persistent-session-OQRYZ3YP.js");
+      const { collectDiagnostics } = await import("../persistent-session-UEAJJ72P.js");
       await api.post("/host/heartbeat", {
         host_id: hostId,
         agent_diagnostics: collectDiagnostics([codeName], quarantineEntriesFor, claudeMdSizeFor, spawnOutcomeForDiagnostics, pidPressureFor, forwardedToolsFor)
@@ -21003,7 +21052,7 @@ async function processClaudePairSessions(agents) {
     killPairSession,
     pairTmuxSession,
     finalizeClaudePairOnboarding
-  } = await import("../claude-pair-runtime-NNHCDXLG.js");
+  } = await import("../claude-pair-runtime-SY5VCTCE.js");
   for (const pairId of pendingResp.cancelled_pair_ids ?? []) {
     log(`[claude-pair] sweeping orphan tmux session for pair ${pairId.slice(0, 8)}`);
     const killed = await killPairSession(pairTmuxSession(pairId));