@integrity-labs/agt-cli 0.28.763 → 0.28.765

package/dist/{chunk-HVOHM3PB.js → chunk-EWH4J63B.js}

@@ -7,13 +7,14 @@ import {
   buildRemoteMcpEntry,
   expandTemplateVars,
   getProjectDir,
+  hardBoundedExecFile,
   isolationMode,
   log,
   mergeEnvIntegrationsContent,
   parseEnvFileEntries,
   parseEnvIntegrations,
   shellQuote
-} from "./chunk-EU6RSU5V.js";
+} from "./chunk-LVBINNTQ.js";
 import {
   BIND_FAILURE_QUARANTINE_THRESHOLD,
   INTEGRATIONS_SECTION_END,
@@ -63,7 +64,7 @@ import {
   sessionTranscriptDir,
   worseConnectivityOutcome,
   wrapScheduledTaskPrompt
-} from "./chunk-L4KDOGHB.js";
+} from "./chunk-23KTLVYX.js";
 import {
   parsePsRows
 } from "./chunk-XWVM4KPK.js";
@@ -2551,9 +2552,24 @@ function provisionStopHook(codeName) {
     "# obligation (consume is the explicit handled-without-reply ack, the same",
     "# session-scoped drain the MCP applies), so either counts as answered. session_id",
     "# is unambiguous (one session per marker), so there is no cross-thread class here.",
+    "#",
+    "# ENG-9420: this predicate was the one of four that ENG-8142 missed. It came",
+    "# from ENG-7808 and had been error-blind ever since, so a direct_chat.reply",
+    '# whose tool_result came back is_error:true still satisfied "the agent',
+    '# replied" - the recovery outbox never written, the owed-reply re-prompt',
+    "# suppressed. Direct chat is where scheduled-task results and approval",
+    "# round-trips land, so that is a user asking a question and getting nothing.",
+    "#",
+    "# A FAILED consume is treated the same as a failed reply, deliberately.",
+    '# consume means "handled without replying" and drains the marker; if the',
+    "# drain failed, the marker was not drained and the conversation is not",
+    "# settled. Counting it would suppress the re-prompt for a message that is",
+    "# still genuinely pending. The failure direction of this choice is an extra",
+    "# re-prompt, which the user sees and can ignore; the other direction is a",
+    "# silently dropped message, which nobody sees at all.",
     "replied_this_conv_directchat() {",
     '  local session_id="$1" hit',
-    `  hit=$(printf '%s' "$LAST_ASSISTANT" | jq -r --arg sid "$session_id" '(.message.content // .content // []) | (if type=="array" then . else [] end) | (if any(.[]; (type=="object") and (.type=="tool_use") and ((.name|tostring)|test("(^|__)direct_chat[._](reply|consume)$")) and (((.input.session_id // "")|tostring)==$sid)) then "yes" else "no" end)' 2>/dev/null || echo no)`,
+    `  hit=$(printf '%s' "$LAST_ASSISTANT" | jq -r --arg sid "$session_id" --arg errids "$ERRORED_TOOL_IDS" '(($errids // "") | split("\\n") | map(select(length > 0))) as $bad | (.message.content // .content // []) | (if type=="array" then . else [] end) | (if any(.[]; (type=="object") and (.type=="tool_use") and ((.name|tostring)|test("(^|__)direct_chat[._](reply|consume)$")) and ((.id // "") as $tid | ($bad | index($tid)) == null) and (((.input.session_id // "")|tostring)==$sid)) then "yes" else "no" end)' 2>/dev/null || echo no)`,
     '  [ "$hit" = "yes" ]',
     "}",
     "recover_telegram_for() {",
@@ -6436,7 +6452,7 @@ function exchangeFailureKind(err) {
 }
 
 // src/lib/api-client.ts
-var agtCliVersion = true ? "0.28.763" : "dev";
+var agtCliVersion = true ? "0.28.765" : "dev";
 var lastConfigHash = null;
 function setConfigHash(hash) {
   lastConfigHash = hash && hash.length > 0 ? hash : null;
@@ -7043,6 +7059,12 @@ var HostFlagStore = class {
 
 // src/lib/stale-mcp-reaper.ts
 import { execFileSync } from "child_process";
+function livenessFromProbe(r, onNoAnswer) {
+  if (r.kind === "ok") return true;
+  if (r.kind === "failed" && r.status !== null) return false;
+  onNoAnswer?.(r.kind === "timed-out" ? "did not answer within its bound" : `could not run: ${r.message}`);
+  return true;
+}
 var rotationTimestamps = /* @__PURE__ */ new Map();
 var DEFAULT_ROTATION_GRACE_MS = 15e3;
 function rotationKey(codeName, serverKey) {
@@ -7256,19 +7278,30 @@ function findMcpChildrenForAgent(args) {
 function reapStaleMcpChildren(args) {
   const { log: log2, codeName, serverKeys, mcpJson, graceMs = 5e3, isolated = false } = args;
   if (serverKeys.length === 0) return [];
-  const runPs = args.runPs ?? (isolated ? () => execFileSync("docker", ["exec", `agt-${codeName}`, "ps", "-eo", "pid,ppid,args"], {
-    encoding: "utf-8",
-    timeout: 8e3
-  }) : () => execFileSync("ps", ["-eo", "pid,ppid,args"], { encoding: "utf-8", timeout: 5e3 }));
-  const killProcess = args.killProcess ?? (isolated ? (pid, signal) => {
-    try {
-      execFileSync(
+  const runPs = args.runPs ?? (isolated ? (
+    // ENG-9781: hard-bounded. See bounded-exec.ts — `execFileSync`'s own
+    // `timeout` is not a bound, and this runs per isolated agent per cycle
+    // on the manager-worker's critical path. Throws on a non-answer so the
+    // caller's existing catch skips the agent rather than acting on an
+    // empty process list.
+    () => {
+      const r = hardBoundedExecFile(
         "docker",
-        ["exec", `agt-${codeName}`, "kill", signal === "SIGKILL" ? "-KILL" : "-TERM", String(pid)],
-        { timeout: 8e3, stdio: "ignore" }
+        ["exec", `agt-${codeName}`, "ps", "-eo", "pid,ppid,args"],
+        { timeoutMs: 8e3, log: log2 }
+      );
+      if (r.kind === "ok") return r.stdout;
+      throw new Error(
+        r.kind === "timed-out" ? `docker exec ps did not answer within its bound (bounded=${r.bounded})` : `docker exec ps failed (status=${r.status}): ${r.message}`
       );
-    } catch {
     }
+  ) : () => execFileSync("ps", ["-eo", "pid,ppid,args"], { encoding: "utf-8", timeout: 5e3 }));
+  const killProcess = args.killProcess ?? (isolated ? (pid, signal) => {
+    hardBoundedExecFile(
+      "docker",
+      ["exec", `agt-${codeName}`, "kill", signal === "SIGKILL" ? "-KILL" : "-TERM", String(pid)],
+      { timeoutMs: 8e3, log: log2 }
+    );
   } : (pid, signal) => {
     try {
       process.kill(pid, signal);
@@ -7276,15 +7309,17 @@ function reapStaleMcpChildren(args) {
     }
   });
   const isAlive = args.isAlive ?? (isolated ? (pid) => {
-    try {
-      execFileSync("docker", ["exec", `agt-${codeName}`, "kill", "-0", String(pid)], {
-        timeout: 8e3,
-        stdio: "ignore"
-      });
-      return true;
-    } catch {
-      return false;
-    }
+    const r = hardBoundedExecFile(
+      "docker",
+      ["exec", `agt-${codeName}`, "kill", "-0", String(pid)],
+      { timeoutMs: 8e3, log: log2 }
+    );
+    return livenessFromProbe(
+      r,
+      (why) => log2(
+        `[stale-mcp-reaper] liveness probe for pid ${pid} in 'agt-${codeName}' ${why} \u2014 treating as ALIVE (no reap) rather than guessing`
+      )
+    );
   } : (pid) => {
     try {
       process.kill(pid, 0);
@@ -7453,10 +7488,33 @@ function reapMissingMcpSessions(args) {
     currentGenerationExternallyStarted = false
   } = args;
   const now = args.now ?? Date.now;
-  const runPs = args.runPs ?? (isolated ? () => execFileSync2("docker", ["exec", `agt-${codeName}`, "ps", "-eo", "pid,ppid,args"], {
-    encoding: "utf-8",
-    timeout: 8e3
-  }) : () => execFileSync2("ps", ["-eo", "pid,ppid,args"], { encoding: "utf-8", timeout: 5e3 }));
+  const runPs = args.runPs ?? (isolated ? (
+    // ENG-9781: HARD-bounded, and the distinction is the whole ticket. The
+    // previous `execFileSync(..., { timeout: 8_000 })` bounded nothing —
+    // Node fires SIGTERM at the deadline and then keeps waiting, so a
+    // docker CLI talking to a wedged daemon blocked the manager-worker's
+    // event loop and stopped scheduling for EVERY agent on the host
+    // (33m46s, twice, on 2026-09-01). This call runs once per isolated
+    // agent per poll cycle, so it is the single worst place on the manager
+    // for an unbounded subprocess.
+    //
+    // Throwing on a non-answer is deliberate rather than lazy: the caller
+    // already catches, logs and SKIPS the agent for the cycle, which is
+    // exactly right. A timeout must not become an empty process list —
+    // that reads as "every declared MCP server is missing" and restarts the
+    // session, turning a docker hiccup into a fleet-wide restart storm.
+    () => {
+      const r = hardBoundedExecFile(
+        "docker",
+        ["exec", `agt-${codeName}`, "ps", "-eo", "pid,ppid,args"],
+        { timeoutMs: 8e3, log: log2 }
+      );
+      if (r.kind === "ok") return r.stdout;
+      throw new Error(
+        r.kind === "timed-out" ? `docker exec ps did not answer within its bound (bounded=${r.bounded})` : `docker exec ps failed (status=${r.status}): ${r.message}`
+      );
+    }
+  ) : () => execFileSync2("ps", ["-eo", "pid,ppid,args"], { encoding: "utf-8", timeout: 5e3 }));
   if (!mcpJson?.mcpServers) {
     return { missing: [], restarted: false, reason: "no-mcp-json" };
   }
@@ -10593,4 +10651,4 @@ export {
   managerInstallSystemUnitCommand,
   managerUninstallSystemUnitCommand
 };
-//# sourceMappingURL=chunk-HVOHM3PB.js.map
+//# sourceMappingURL=chunk-EWH4J63B.js.map