@integrity-labs/agt-cli 0.28.158 → 0.28.159

package/dist/{chunk-E6ABR63O.js → chunk-TNKKZHEQ.js}

@@ -3,7 +3,7 @@ import {
   formatMissingVar,
   isClaudeFastMode,
   probeMcpEnvSubstitution
-} from "./chunk-TLX6ORJ5.js";
+} from "./chunk-SAAK2IP4.js";
 import {
   reapOrphanChannelMcps
 } from "./chunk-XWVM4KPK.js";
@@ -1588,4 +1588,4 @@ export {
   stopAllSessionsAndWait,
   getProjectDir
 };
-//# sourceMappingURL=chunk-E6ABR63O.js.map
+//# sourceMappingURL=chunk-TNKKZHEQ.js.map

package/dist/{claude-pair-runtime-SXO5Q5R6.js → claude-pair-runtime-E2E2FJHS.js}

@@ -100,7 +100,7 @@ async function spawnPairSession(session) {
     return { ok: true };
   } catch {
   }
-  const { resolveClaudeBinary } = await import("./persistent-session-N7WFONZC.js");
+  const { resolveClaudeBinary } = await import("./persistent-session-K5CFAUY7.js");
   const claudeBin = resolveClaudeBinary();
   const pairEnv = {
     ...process.env,
@@ -373,4 +373,4 @@ export {
   startClaudePair,
   submitClaudePairCode
 };
-//# sourceMappingURL=claude-pair-runtime-SXO5Q5R6.js.map
+//# sourceMappingURL=claude-pair-runtime-E2E2FJHS.js.map

package/dist/lib/manager-worker.js

@@ -28,7 +28,7 @@ import {
   requireHost,
   safeWriteJsonAtomic,
   setConfigHash
-} from "../chunk-3OLFOA4F.js";
+} from "../chunk-PGW3CTYR.js";
 import {
   getProjectDir as getProjectDir2,
   getReadyTasks,
@@ -70,8 +70,9 @@ import {
   takeZombieDetection,
   transcriptActivityAgeSeconds,
   writeEgressAllowlist
-} from "../chunk-E6ABR63O.js";
+} from "../chunk-TNKKZHEQ.js";
 import {
+  CONVERSATION_FAILURE_CATEGORIES,
   FLAGS_SCHEMA_VERSION,
   FLAG_REGISTRY,
   KANBAN_CHECK_COMMAND,
@@ -80,6 +81,7 @@ import {
   StreamEncoder,
   appendDmFooter,
   attributeTranscriptUsageByRun,
+  buildFailureCategoryPromptLines,
   buildScheduledTaskContextBlocks,
   classifyActor,
   classifyOutput,
@@ -89,6 +91,7 @@ import {
   formatActorId,
   formatRunMarker,
   getFramework,
+  isConversationFailureCategory,
   isEmptyTotals,
   isOnboardingArea,
   isParseError,
@@ -107,7 +110,7 @@ import {
   resolveChannels,
   resolveDmTarget,
   sumTranscriptUsageInWindow
-} from "../chunk-TLX6ORJ5.js";
+} from "../chunk-SAAK2IP4.js";
 import {
   parsePsRows,
   reapOrphanChannelMcps
@@ -2498,9 +2501,10 @@ 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)
 - summary: ONE short sentence (max 140 chars) explaining the score. Do NOT quote sensitive user content.
+- 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:
-{"score": 0-100, "verdict": "success|partial|failure", "summary": "..."}`;
+{"score": 0-100, "verdict": "success|partial|failure", "summary": "...", "failure_category": "${CONVERSATION_FAILURE_CATEGORIES.join("|")} or null"}`;
 }
 function parseVerdict(raw) {
   const match = raw.match(/\{[\s\S]*\}/);
@@ -2515,7 +2519,8 @@ function parseVerdict(raw) {
   if (!Number.isFinite(score) || score < 0 || score > 100) return null;
   if (obj.verdict !== "success" && obj.verdict !== "partial" && obj.verdict !== "failure") return null;
   const summary = typeof obj.summary === "string" ? obj.summary.slice(0, 140) : "";
-  return { score: Math.round(score), verdict: obj.verdict, summary };
+  const failureCategory = obj.verdict === "failure" && isConversationFailureCategory(obj.failure_category) ? obj.failure_category : null;
+  return { score: Math.round(score), verdict: obj.verdict, summary, failureCategory };
 }
 async function maybeEvaluateConversations(args) {
   const { api: api2, backend, codeName, agentId, log: log2 } = args;
@@ -2585,6 +2590,7 @@ async function maybeEvaluateConversations(args) {
         score: verdict.score,
         verdict: verdict.verdict,
         summary: verdict.summary,
+        failure_category: verdict.failureCategory,
         model: backend.model
       });
       log2(
@@ -6871,7 +6877,7 @@ var agentRestartTimezoneInputs = /* @__PURE__ */ new Map();
 var lastVersionCheckAt = 0;
 var VERSION_CHECK_INTERVAL_MS = 5 * 60 * 1e3;
 var lastResponsivenessProbeAt = 0;
-var agtCliVersion = true ? "0.28.158" : "dev";
+var agtCliVersion = true ? "0.28.159" : "dev";
 function resolveBrewPath(execFileSync4) {
   try {
     const out = execFileSync4("which", ["brew"], { timeout: 5e3 }).toString().trim();
@@ -8004,7 +8010,7 @@ async function pollCycle() {
     }
     try {
       const { detectHostSecurity } = await import("../host-security-6PDFG7F5.js");
-      const { collectDiagnostics } = await import("../persistent-session-N7WFONZC.js");
+      const { collectDiagnostics } = await import("../persistent-session-K5CFAUY7.js");
       const diagCodeNames = [...agentState.persistentSessionAgents];
       const agentDiagnostics = diagCodeNames.length > 0 ? collectDiagnostics(diagCodeNames) : void 0;
       let tailscaleHostname;
@@ -8152,7 +8158,7 @@ async function pollCycle() {
       const {
         collectResponsivenessProbes,
         getResponsivenessIntervalMs
-      } = await import("../responsiveness-probe-UXJY4YRL.js");
+      } = await import("../responsiveness-probe-5BQI3RNU.js");
       const probeIntervalMs = getResponsivenessIntervalMs();
       if (now - lastResponsivenessProbeAt > probeIntervalMs) {
         const probeCodeNames = [...agentState.persistentSessionAgents];
@@ -8184,7 +8190,7 @@ async function pollCycle() {
           collectResponsivenessProbes,
           livePendingInboundOldestAgeSeconds,
           parkPendingInbound
-        } = await import("../responsiveness-probe-UXJY4YRL.js");
+        } = await import("../responsiveness-probe-5BQI3RNU.js");
         const { getProjectDir: wedgeProjectDir } = await import("../claude-scheduler-FATCLHDM.js");
         const wedgeNow = /* @__PURE__ */ new Date();
         const liveAgents = agentState.persistentSessionAgents;
@@ -11685,7 +11691,7 @@ async function processClaudePairSessions(agents) {
     killPairSession,
     pairTmuxSession,
     finalizeClaudePairOnboarding
-  } = await import("../claude-pair-runtime-SXO5Q5R6.js");
+  } = await import("../claude-pair-runtime-E2E2FJHS.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));