@integrity-labs/agt-cli 0.28.461 → 0.28.463

package/dist/lib/manager-worker.js

@@ -42,7 +42,7 @@ import {
   resolveEffectivePinRaw,
   safeWriteJsonAtomic,
   setConfigHash
-} from "../chunk-WCYGJANF.js";
+} from "../chunk-VCZNGC6J.js";
 import {
   getProjectDir as getProjectDir2,
   getReadyTasks,
@@ -9976,7 +9976,7 @@ var agentRestartTimezoneInputs = /* @__PURE__ */ new Map();
 var lastVersionCheckAt = 0;
 var VERSION_CHECK_INTERVAL_MS = 5 * 60 * 1e3;
 var lastResponsivenessProbeAt = 0;
-var agtCliVersion = true ? "0.28.461" : "dev";
+var agtCliVersion = true ? "0.28.463" : "dev";
 function resolveBrewPath(execFileSync2) {
   try {
     const out = execFileSync2("which", ["brew"], { timeout: 5e3 }).toString().trim();
@@ -13792,6 +13792,67 @@ async function processAgent(agent, agentStates) {
               );
             });
           },
+          // ENG-8225: a stale-mcp-reaper rotation's grace window expired with the
+          // child still absent — nothing made the tool call ENG-5344 was banking on
+          // (dwight's `anchor-browser`, dead from 16:00:05, never called again all
+          // evening). Arm ENG-7429's prompt bind-probe budget so the next tool call
+          // is MANUFACTURED, which is the only lever that exists: a stdio MCP child
+          // belongs to Claude Code on the session's stdio pipes, so a
+          // manager-spawned one would be invisible to findMcpChildrenForAgent's
+          // ppid-to-`claude` walk AND unreachable by any tool call.
+          //
+          // Returns whether the probe was actually armed. Gated on the same
+          // `session-tool-probe` flag as every other arming site (2014, 6780) — when
+          // it's off, arming is inert, so we report false and the reaper falls back
+          // to its pre-ENG-8225 restart rather than suppressing a restart with no
+          // repair behind it.
+          onRotationRespawnStalled: (codeName, serverKey, { attempt, maxAttempts }) => {
+            if (!hostFlagStore().getBoolean("session-tool-probe")) {
+              log(
+                `[mcp-presence-reaper] '${codeName}:${serverKey}' rotation respawn stalled but the 'session-tool-probe' flag is off \u2014 cannot arm the ENG-7429 prompt probe; falling back to the session restart (ENG-8225)`
+              );
+              return false;
+            }
+            promptBindProbeAttempts.set(codeName, PROMPT_BIND_PROBE_ATTEMPTS);
+            log(
+              `[mcp-presence-reaper] '${codeName}:${serverKey}' armed force-due prompt bind-probe budget (${PROMPT_BIND_PROBE_ATTEMPTS}) to force a tool call, attempt ${attempt}/${maxAttempts} (ENG-8225)`
+            );
+            return true;
+          },
+          // ENG-8225 (AC3): the prompt-probe budget is spent and the rotated child
+          // never came back. Escalate down the two paths this area already uses —
+          // mark the integration unhealthy (`/host/integration-health`, the ENG-5292
+          // onGiveUp path) and open a `reaper_action` alert (`/host/reaper-event`,
+          // the ENG-6184 onRestart path). No new alerting channel.
+          onRotationRespawnGaveUp: (codeName, serverKey, { attempts }) => {
+            const reason = `rotation respawn never happened: stale-mcp-reaper SIGTERM'd '${serverKey}' and the child never returned after ${attempts} prompt bind-probe(s) (ENG-8225)`;
+            const mapping = managedToolkitIdByAgentAndServerId.get(
+              managedToolkitMapKey(agent.agent_id, serverKey)
+            );
+            if (mapping) {
+              const { toolkitId, connectionKey } = mapping;
+              api.post("/host/integration-health", {
+                agent_id: agent.agent_id,
+                toolkit_id: toolkitId,
+                connection_key: connectionKey,
+                action: "mark_unhealthy",
+                reason
+              }).then(() => {
+                log(
+                  `[mcp-presence-reaper] marked integration unhealthy for '${codeName}' toolkit '${toolkitId}' after a never-respawned rotation (ENG-8225)`
+                );
+              }).catch((err) => {
+                log(
+                  `[mcp-presence-reaper] failed to mark integration unhealthy for '${codeName}' toolkit '${toolkitId}' after a never-respawned rotation (ENG-8225): ${err.message} \u2014 local escalation state still recorded`
+                );
+              });
+            } else {
+              log(
+                `[mcp-presence-reaper] never-respawned rotation on '${codeName}:${serverKey}' \u2014 not a managed toolkit (no toolkit_id mapping), skipping integration-health API call (ENG-8225)`
+              );
+            }
+            void reportReaperEvent(agent.agent_id, "mcp-presence-reaper", "kill", reason);
+          },
           // ENG-5286: record the restart event with the API so the
           // agent-heartbeat-monitor cron can aggregate per-agent counts
           // into the McpReaperRestartsHourly CloudWatch metric. Fire-and-
@@ -15121,7 +15182,12 @@ async function processDirectChatMessage(agent, msg) {
       await api.post("/host/direct-chat/reply", {
         agent_id: agent.agentId,
         session_id: msg.session_id,
-        content: "[Expired] This message timed out before it could be processed. Please resend it."
+        content: "[Expired] This message timed out before it could be processed. Please resend it.",
+        // ENG-8274: kind:'notice' — the manager is draining a stale backlog
+        // message. That is the system reporting a delivery outcome, not the agent
+        // answering, and it owes no reply. Without this it rendered as an agent
+        // bubble, so a timeout the agent never even saw read as the agent's words.
+        kind: "notice"
       });
     } catch {
     }