@integrity-labs/agt-cli 0.28.184 → 0.28.186

package/dist/{chunk-RLDZLSML.js → chunk-ORO4YLTN.js}

@@ -3,7 +3,7 @@ import {
   formatMissingVar,
   isClaudeFastMode,
   probeMcpEnvSubstitution
-} from "./chunk-K2N44EUC.js";
+} from "./chunk-ERHHYBVS.js";
 import {
   reapOrphanChannelMcps
 } from "./chunk-XWVM4KPK.js";
@@ -1588,4 +1588,4 @@ export {
   stopAllSessionsAndWait,
   getProjectDir
 };
-//# sourceMappingURL=chunk-RLDZLSML.js.map
+//# sourceMappingURL=chunk-ORO4YLTN.js.map

package/dist/{claude-pair-runtime-COEJFJJX.js → claude-pair-runtime-VBTV4KZO.js}

@@ -100,7 +100,7 @@ async function spawnPairSession(session) {
     return { ok: true };
   } catch {
   }
-  const { resolveClaudeBinary } = await import("./persistent-session-KBZFQ7SX.js");
+  const { resolveClaudeBinary } = await import("./persistent-session-SNLNY2VB.js");
   const claudeBin = resolveClaudeBinary();
   const pairEnv = {
     ...process.env,
@@ -373,4 +373,4 @@ export {
   startClaudePair,
   submitClaudePairCode
 };
-//# sourceMappingURL=claude-pair-runtime-COEJFJJX.js.map
+//# sourceMappingURL=claude-pair-runtime-VBTV4KZO.js.map

package/dist/lib/manager-worker.js

@@ -28,7 +28,7 @@ import {
   requireHost,
   safeWriteJsonAtomic,
   setConfigHash
-} from "../chunk-YYIBKSSD.js";
+} from "../chunk-3UMHVNGB.js";
 import {
   getProjectDir as getProjectDir2,
   getReadyTasks,
@@ -70,7 +70,7 @@ import {
   takeZombieDetection,
   transcriptActivityAgeSeconds,
   writeEgressAllowlist
-} from "../chunk-RLDZLSML.js";
+} from "../chunk-ORO4YLTN.js";
 import {
   CONVERSATION_FAILURE_CATEGORIES,
   DEFAULT_FRAMEWORK,
@@ -111,7 +111,7 @@ import {
   resolveChannels,
   resolveDmTarget,
   sumTranscriptUsageInWindow
-} from "../chunk-K2N44EUC.js";
+} from "../chunk-ERHHYBVS.js";
 import {
   parsePsRows,
   reapOrphanChannelMcps
@@ -2500,8 +2500,14 @@ ${transcript}
 
 Score the agent's success:
 - score: integer 0-100 (0 = ignored/unhelpful/wrong, 100 = fully resolved the user's need)
-- verdict: "success" (need clearly met), "partial" (some help but incomplete/ambiguous), or "failure" (did not help / made it worse)
+- verdict: "success" (need clearly met), "partial" (some help but genuinely incomplete), or "failure" (did not help / made it worse)
 - summary: ONE short sentence (max 140 chars) explaining the score. Do NOT quote sensitive user content.
+
+How to choose the verdict (read carefully \u2014 most conversations are short and one-sided):
+- A complete, correct, self-contained answer to the user's question is a "success" EVEN IF the user never replied to confirm. End-users rarely send a closing "thanks". Do NOT downgrade a good answer to "partial" merely because the transcript ends after the agent's reply or because there is no explicit confirmation. Absence of confirmation is NOT ambiguity.
+- Use "partial" ONLY when the agent's own response is genuinely incomplete: it asked the user a clarifying question that was never answered, gave a hedged/half answer, promised a follow-up that is not present, or addressed only part of a multi-part request.
+- Use "failure" when the agent ignored the user, was wrong, or made things worse.
+- If the transcript looks truncated or cut off mid-exchange (a reconstruction artifact, not the agent's fault), judge only what the agent actually delivered; do not invent incompleteness from a missing tail.
 - failure_category: ONLY when verdict is "failure", the single best-fitting reason from this exact set: ${buildFailureCategoryPromptLines()}. Use null when verdict is "success" or "partial".
 
 Respond with ONLY a JSON object, no other text:
@@ -6596,7 +6602,7 @@ var agentRestartTimezoneInputs = /* @__PURE__ */ new Map();
 var lastVersionCheckAt = 0;
 var VERSION_CHECK_INTERVAL_MS = 5 * 60 * 1e3;
 var lastResponsivenessProbeAt = 0;
-var agtCliVersion = true ? "0.28.184" : "dev";
+var agtCliVersion = true ? "0.28.186" : "dev";
 function resolveBrewPath(execFileSync4) {
   try {
     const out = execFileSync4("which", ["brew"], { timeout: 5e3 }).toString().trim();
@@ -7489,7 +7495,7 @@ async function pollCycle() {
     }
     try {
       const { detectHostSecurity } = await import("../host-security-6PDFG7F5.js");
-      const { collectDiagnostics } = await import("../persistent-session-KBZFQ7SX.js");
+      const { collectDiagnostics } = await import("../persistent-session-SNLNY2VB.js");
       const diagCodeNames = [...agentState.persistentSessionAgents];
       const agentDiagnostics = diagCodeNames.length > 0 ? collectDiagnostics(diagCodeNames) : void 0;
       let tailscaleHostname;
@@ -7637,7 +7643,7 @@ async function pollCycle() {
       const {
         collectResponsivenessProbes,
         getResponsivenessIntervalMs
-      } = await import("../responsiveness-probe-O4F7H7IB.js");
+      } = await import("../responsiveness-probe-6HC3PENG.js");
       const probeIntervalMs = getResponsivenessIntervalMs();
       if (now - lastResponsivenessProbeAt > probeIntervalMs) {
         const probeCodeNames = [...agentState.persistentSessionAgents];
@@ -7669,7 +7675,7 @@ async function pollCycle() {
           collectResponsivenessProbes,
           livePendingInboundOldestAgeSeconds,
           parkPendingInbound
-        } = await import("../responsiveness-probe-O4F7H7IB.js");
+        } = await import("../responsiveness-probe-6HC3PENG.js");
         const { getProjectDir: wedgeProjectDir } = await import("../claude-scheduler-FATCLHDM.js");
         const wedgeNow = /* @__PURE__ */ new Date();
         const liveAgents = agentState.persistentSessionAgents;
@@ -10644,7 +10650,7 @@ async function processClaudePairSessions(agents) {
     killPairSession,
     pairTmuxSession,
     finalizeClaudePairOnboarding
-  } = await import("../claude-pair-runtime-COEJFJJX.js");
+  } = await import("../claude-pair-runtime-VBTV4KZO.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));