@integrity-labs/agt-cli 0.27.150-test.15 → 0.27.150

package/dist/lib/manager-worker.js

@@ -15,10 +15,11 @@ import {
   provisionAutoKanbanProgressHook,
   provisionIsolationHook,
   provisionOrientHook,
+  provisionSessionStateHook,
   provisionStopHook,
   requireHost,
   safeWriteJsonAtomic
-} from "../chunk-24FTY53Z.js";
+} from "../chunk-CUHP2SVW.js";
 import {
   getProjectDir as getProjectDir2,
   getReadyTasks,
@@ -37,10 +38,13 @@ import {
   getSessionState,
   injectMessage,
   injectMessageWithStatus,
+  isAgentIdle,
   isAgentPromptReady,
   isSessionHealthy,
+  isStaleForToday,
   paneLogPath,
   parseEnvIntegrations,
+  peekCurrentSession,
   prepareForRespawn,
   probeMcpEnvSubstitution,
   readPaneLogTail,
@@ -49,12 +53,14 @@ import {
   rotateSessionForWedge,
   sanitizeMcpJson,
   sendToAgent,
+  sessionTranscriptDir,
   startPersistentSession,
   stopAllSessionsAndWait,
   stopPersistentSession,
   takeWatchdogGiveUpCount,
-  takeZombieDetection
-} from "../chunk-7GKJZBTB.js";
+  takeZombieDetection,
+  transcriptActivityAgeSeconds
+} from "../chunk-JLS7NQFE.js";
 import {
   KANBAN_CHECK_COMMAND,
   MAX_AVATAR_ENV_URL_BYTES,
@@ -83,13 +89,7 @@ import {
   resolveDmTarget,
   worseConnectivityOutcome,
   wrapScheduledTaskPrompt
-} from "../chunk-WOOYOAPG.js";
-import {
-  isAgentIdle,
-  isStaleForToday,
-  peekCurrentSession,
-  sessionTranscriptDir
-} from "../chunk-354FAVQR.js";
+} from "../chunk-A75AOK6E.js";
 import {
   parsePsRows,
   reapOrphanChannelMcps
@@ -1515,6 +1515,7 @@ function isUrgentUpgrade(opts) {
 }
 var RESTART_IDLE_THRESHOLD_SECONDS = 120;
 var RESTART_INBOUND_QUIET_SECONDS = 300;
+var RESTART_TRANSCRIPT_STALE_SECONDS = 60;
 var GATEABLE_RESTART_REASONS = /* @__PURE__ */ new Set([
   "model-change",
   "channel-set-change",
@@ -1533,9 +1534,11 @@ function decideRestartGate(opts) {
   }
   const paneThreshold = opts.idleThresholdSeconds ?? RESTART_IDLE_THRESHOLD_SECONDS;
   const inboundThreshold = opts.inboundQuietSeconds ?? RESTART_INBOUND_QUIET_SECONDS;
-  const paneBusy = opts.paneLogAgeSeconds !== null && opts.paneLogAgeSeconds < paneThreshold;
+  const transcriptThreshold = opts.transcriptStaleSeconds ?? RESTART_TRANSCRIPT_STALE_SECONDS;
+  const transcriptAge = opts.transcriptAgeSeconds ?? null;
+  const progressBusy = transcriptAge !== null ? transcriptAge < transcriptThreshold : opts.paneLogAgeSeconds !== null && opts.paneLogAgeSeconds < paneThreshold;
   const inboundBusy = opts.inboundAgeSeconds !== null && opts.inboundAgeSeconds < inboundThreshold;
-  if (paneBusy || inboundBusy) return "defer-idle";
+  if (progressBusy || inboundBusy) return "defer-idle";
   return "proceed";
 }
 
@@ -3377,13 +3380,18 @@ function clearAgentState(agentId, codeName) {
 // src/lib/wedge-detection.ts
 var DEFAULTS = {
   inboundWaitSeconds: 120,
-  // ENG-6238: the hard cap is now an ABSOLUTE BACKSTOP, not the primary
-  // discriminator — a still-producing session (e.g. a runaway loop) is only
-  // reaped here. Raised 300→1200 because the soft path now reliably catches
-  // the frozen-turn wedge via the transcript signal, so the hard cap no longer
-  // needs to fire early (which is exactly what false-killed kylie's long
-  // legitimate turns, ENG-6238).
-  inboundHardWaitSeconds: 1200,
+  // ENG-6264: DISABLED by default (0). A session that's actively producing
+  // tokens is never force-respawned — a working agent must not be killed just
+  // because a message has been queued behind its turn, no matter how long.
+  // ENG-6238 made this an absolute backstop (1200s) to still catch a
+  // producing-but-never-draining runaway loop, but that re-introduced the exact
+  // failure we set out to kill: cutting off real work on a long turn. Runaway
+  // token burn is owned by the cost guardrail (ENG-5556); a producing-but-silent
+  // loop still trips the synthetic-probe alarm. So the backstop is now opt-in:
+  // set AGT_WEDGE_INBOUND_HARD_WAIT_SECONDS to a positive value to re-enable it
+  // (floored at inboundWaitSeconds). 0 = the frozen/hung wedge (transcript
+  // static) is still caught by the soft path; only the *producing* path is spared.
+  inboundHardWaitSeconds: 0,
   paneStaleSeconds: 120,
   transcriptStaleSeconds: 60,
   minCycles: 3
@@ -3402,10 +3410,12 @@ function resolveWedgeConfig(env = process.env) {
     DEFAULTS.inboundWaitSeconds,
     30
   );
-  const inboundHardWaitSeconds = Math.max(
-    inboundWaitSeconds,
-    parsePositiveInt(env.AGT_WEDGE_INBOUND_HARD_WAIT_SECONDS, DEFAULTS.inboundHardWaitSeconds, 30)
+  const inboundHardWaitRaw = parsePositiveInt(
+    env.AGT_WEDGE_INBOUND_HARD_WAIT_SECONDS,
+    DEFAULTS.inboundHardWaitSeconds,
+    0
   );
+  const inboundHardWaitSeconds = inboundHardWaitRaw <= 0 ? 0 : Math.max(inboundWaitSeconds, inboundHardWaitRaw);
   return {
     mode: parseMode(env.AGT_WEDGE_RESTART_MODE),
     inboundWaitSeconds,
@@ -3430,6 +3440,7 @@ function isWedgeCandidateCycle(signals, config2) {
   if (inboundAge === null) return false;
   if (inboundAge < config2.inboundWaitSeconds) return false;
   if (isSessionProducing(signals, config2)) {
+    if (config2.inboundHardWaitSeconds <= 0) return false;
     return inboundAge >= config2.inboundHardWaitSeconds;
   }
   return true;
@@ -4417,11 +4428,16 @@ function paneLogAgeSecondsFor(codeName) {
     return 0;
   }
 }
-function restartGateFor(codeName, breakerReason) {
-  if (!isGateableRestartReason(breakerReason)) return "bypass";
+function transcriptAgeSecondsFor(codeName) {
+  const sessionId = getSessionState(codeName)?.currentSessionId ?? null;
+  return transcriptActivityAgeSeconds(getProjectDir2(codeName), sessionId, /* @__PURE__ */ new Date());
+}
+function restartGateFor(codeName, reason) {
+  if (!isGateableRestartReason(reason)) return "bypass";
   return decideRestartGate({
     window: cachedMaintenanceWindow,
     paneLogAgeSeconds: paneLogAgeSecondsFor(codeName),
+    transcriptAgeSeconds: transcriptAgeSecondsFor(codeName),
     inboundAgeSeconds: inboundAgeSecondsFor(codeName),
     now: /* @__PURE__ */ new Date()
   });
@@ -4580,10 +4596,10 @@ async function runAgentConnectivityProbes(agent, integrations, projectDir) {
     );
   }
 }
-function stopPersistentSessionAndForgetMcpBaseline(codeName, breakerReason) {
-  const gate = restartGateFor(codeName, breakerReason);
+function stopPersistentSessionAndForgetMcpBaseline(codeName, breakerReason, gateReason = breakerReason) {
+  const gate = restartGateFor(codeName, gateReason);
   if (gate !== "bypass" && gate !== "proceed") {
-    log(`[maintenance-window] Deferring '${breakerReason}' restart for '${codeName}' (${gate})`);
+    log(`[maintenance-window] Deferring '${gateReason}' restart for '${codeName}' (${gate})`);
     return;
   }
   cancelPendingSessionRestart(codeName);
@@ -4760,7 +4776,7 @@ var cachedMaintenanceWindow = null;
 var lastVersionCheckAt = 0;
 var VERSION_CHECK_INTERVAL_MS = 5 * 60 * 1e3;
 var lastResponsivenessProbeAt = 0;
-var agtCliVersion = true ? "0.27.150-test.15" : "dev";
+var agtCliVersion = true ? "0.27.150" : "dev";
 function resolveBrewPath(execFileSync4) {
   try {
     const out = execFileSync4("which", ["brew"], { timeout: 5e3 }).toString().trim();
@@ -5958,7 +5974,7 @@ async function pollCycle() {
     }
     try {
       const { detectHostSecurity } = await import("../host-security-6PDFG7F5.js");
-      const { collectDiagnostics } = await import("../persistent-session-35PWSTLO.js");
+      const { collectDiagnostics } = await import("../persistent-session-ZLEK4KBF.js");
       const diagCodeNames = [...agentState.persistentSessionAgents];
       const agentDiagnostics = diagCodeNames.length > 0 ? collectDiagnostics(diagCodeNames) : void 0;
       let tailscaleHostname;
@@ -6045,12 +6061,12 @@ async function pollCycle() {
       const {
         collectResponsivenessProbes,
         getResponsivenessIntervalMs
-      } = await import("../responsiveness-probe-MA4M2QM4.js");
+      } = await import("../responsiveness-probe-3EUNCJDU.js");
       const probeIntervalMs = getResponsivenessIntervalMs();
       if (now - lastResponsivenessProbeAt > probeIntervalMs) {
         const probeCodeNames = [...agentState.persistentSessionAgents];
         if (probeCodeNames.length > 0) {
-          const { takeAcpxExecFailureCount, creditAcpxExecFailureCount } = await import("../persistent-session-35PWSTLO.js");
+          const { takeAcpxExecFailureCount, creditAcpxExecFailureCount } = await import("../persistent-session-ZLEK4KBF.js");
           const drainedGiveUps = /* @__PURE__ */ new Map();
           const drainedAcpxFailures = /* @__PURE__ */ new Map();
           const probes = collectResponsivenessProbes(probeCodeNames).map((p) => {
@@ -6084,8 +6100,7 @@ async function pollCycle() {
           collectResponsivenessProbes,
           livePendingInboundOldestAgeSeconds,
           deadLetterPendingInbound
-        } = await import("../responsiveness-probe-MA4M2QM4.js");
-        const { transcriptActivityAgeSeconds } = await import("../daily-session-PNQX5URX.js");
+        } = await import("../responsiveness-probe-3EUNCJDU.js");
         const { getProjectDir: wedgeProjectDir } = await import("../claude-scheduler-FATCLHDM.js");
         const wedgeNow = /* @__PURE__ */ new Date();
         const liveAgents = agentState.persistentSessionAgents;
@@ -6149,6 +6164,21 @@ async function pollCycle() {
             log(
               `[wedge] forced fresh respawn ${detail} \u2192 new session ${newId} (transcript preserved${deadNote})`
             );
+            const wedgeAgentId = agentState.codeNameToAgentId.get(codeName);
+            if (wedgeAgentId) {
+              api.post("/host/wedge-respawn", {
+                agent_id: wedgeAgentId,
+                code_name: codeName,
+                dead_lettered_count: deadLettered,
+                pane_age_seconds: signals.paneActivityAgeSeconds,
+                inbound_age_seconds: signals.pendingInboundOldestAgeSeconds,
+                transcript_age_seconds: transcriptAge
+              }).catch((err) => {
+                log(
+                  `[wedge] failed to record respawn event for '${codeName}' (ENG-6265): ${err.message} \u2014 respawn proceeded; CloudWatch metric will under-count this event`
+                );
+              });
+            }
             const inProgressCardIds = (kanbanBoardCache.get(codeName) ?? []).filter((item) => item.status === "in_progress" && isHybridActionable(item)).map((item) => item.id);
             const cardStates = wedgeRestartsByCard.get(codeName) ?? /* @__PURE__ */ new Map();
             const { next, newlyPoisoned } = recordWedgeForCards(
@@ -7868,7 +7898,19 @@ async function processAgent(agent, agentStates) {
         // isolated and the agent keeps running degraded instead of being
         // paused wholesale. reaperRestartBreakerReason() encodes that
         // single-vs-multi decision; undefined means "restart, don't count".
-        stopSession: (codeName, ctx) => stopPersistentSessionAndForgetMcpBaseline(codeName, reaperRestartBreakerReason(ctx.activeKeys)),
+        //
+        // ENG-6264: the breaker-count reason (above, undefined for a single dead
+        // MCP) is decoupled from the GATE reason. Pre-6264 an undefined breaker
+        // reason also made the restart non-gateable → 'bypass' → the session was
+        // torn down mid-turn (the common single-MCP case interrupted busy
+        // agents). Always pass 'mcp-presence-reaper' as the gate reason so the
+        // restart defers-until-idle, while breakerReason still governs whether it
+        // counts against the breaker.
+        stopSession: (codeName, ctx) => stopPersistentSessionAndForgetMcpBaseline(
+          codeName,
+          reaperRestartBreakerReason(ctx.activeKeys),
+          "mcp-presence-reaper"
+        ),
         // ENG-5292: when the reaper gives up on a managed MCP (cap from
         // ENG-5279 + state-preservation from ENG-5285 both said "this
         // MCP keeps failing after 3 restart cycles"), mark the matching
@@ -9087,6 +9129,11 @@ ${truncateForLog(ctx.tail)}` : `; pane_tail_hash=sha256:${createHash3("sha256").
     } catch (err) {
       log(`[persistent-session] Failed to provision auto-progress hook for '${codeName}': ${err.message}`);
     }
+    try {
+      provisionSessionStateHook(codeName);
+    } catch (err) {
+      log(`[persistent-session] Failed to provision session-state hook for '${codeName}': ${err.message}`);
+    }
     const sessionRunResult = await startRun({
       agent_id: agent.agent_id,
       source_type: "system",
@@ -10615,7 +10662,7 @@ async function processClaudePairSessions(agents) {
     killPairSession,
     pairTmuxSession,
     finalizeClaudePairOnboarding
-  } = await import("../claude-pair-runtime-GIUCD7IG.js");
+  } = await import("../claude-pair-runtime-3ZIOY3Z5.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));