@integrity-labs/agt-cli 0.28.779 → 0.28.780

package/dist/{claude-pair-runtime-3J7KK6BP.js → claude-pair-runtime-CK34MZ6M.js}

@@ -100,7 +100,7 @@ async function spawnPairSession(session) {
     return { ok: true };
   } catch {
   }
-  const { resolveClaudeBinary } = await import("./persistent-session-GHIBAWHI.js");
+  const { resolveClaudeBinary } = await import("./persistent-session-ETZBTCXU.js");
   const claudeBin = resolveClaudeBinary();
   const pairEnv = {
     ...process.env,
@@ -373,4 +373,4 @@ export {
   startClaudePair,
   submitClaudePairCode
 };
-//# sourceMappingURL=claude-pair-runtime-3J7KK6BP.js.map
+//# sourceMappingURL=claude-pair-runtime-CK34MZ6M.js.map

package/dist/lib/manager-worker.js

@@ -58,7 +58,7 @@ import {
   safeWriteJsonAtomic,
   setConfigHash,
   tripClass
-} from "../chunk-3IDAWL35.js";
+} from "../chunk-VYFQRDSP.js";
 import {
   getProjectDir as getProjectDir2,
   getReadyTasks,
@@ -120,7 +120,7 @@ import {
   takeZombieDetection,
   toOpencodeModel,
   writeEgressAllowlist
-} from "../chunk-NTWABPKO.js";
+} from "../chunk-HHFDK7WB.js";
 import {
   ACCOUNT_ENFORCEMENT_MARKER_FILENAME,
   AnchorSessionClient,
@@ -143,6 +143,8 @@ import {
   SUPPRESS_SENTINEL,
   StreamEncoder,
   UNKNOWN_RATE_LIMIT,
+  agentHasRules,
+  agentRulesRestartBasis,
   appendDmFooter,
   attributeTranscriptUsageByRun,
   buildFailureCategoryPromptLines,
@@ -209,7 +211,7 @@ import {
   subagentActivityAgeSeconds,
   sumTranscriptUsageInWindow,
   transcriptActivityAgeSeconds
-} from "../chunk-LLBJLKGL.js";
+} from "../chunk-REPC7LF2.js";
 import {
   reapOrphanChannelMcps
 } from "../chunk-XWVM4KPK.js";
@@ -6045,14 +6047,29 @@ var BASELINE_CONCERNS = [
   // reads it once at boot, so a change has to be delivered by a respawn, and
   // the baseline has to survive a manager restart or every deploy would look
   // like a first poll and respawn the fleet (ENG-7771).
-  "telegramPolicy"
+  "telegramPolicy",
+  // ENG-9861: the agent's standing rules (ENG-9789). Same shape and same reason
+  // again — CLAUDE.md is read once at session start, so a rule added mid-session
+  // is on disk and inert until an unrelated restart, and the baseline has to
+  // survive a manager restart or every deploy looks like a first poll.
+  //
+  // Deliberately added WITHOUT bumping the format version. The version is
+  // coupled to the hash SEMANTICS of existing concerns (invariant 2) and none
+  // of those changed; a new key is additive, and an older sidecar lacking it
+  // loads as an empty map, degrading to the first-poll fail-safe for THIS
+  // concern only. Bumping instead would discard every agent's senderPolicy and
+  // telegramPolicy baseline and take the whole fleet through a fail-closed
+  // restart round to add a key — the precedent set when telegramPolicy itself
+  // was added.
+  "agentRules"
 ];
 function createDeliveryBaselineMaps() {
   return {
     senderPolicy: /* @__PURE__ */ new Map(),
     slackBehaviour: /* @__PURE__ */ new Map(),
     msteamsBehaviour: /* @__PURE__ */ new Map(),
-    telegramPolicy: /* @__PURE__ */ new Map()
+    telegramPolicy: /* @__PURE__ */ new Map(),
+    agentRules: /* @__PURE__ */ new Map()
   };
 }
 function getSenderPolicyBaselineFile(configDir) {
@@ -6631,6 +6648,13 @@ var agentState = {
   knownSenderPolicyHashes: /* @__PURE__ */ new Map(),
   /** ENG-9802: per-agent telegram chat-policy hash (allowlist / private-chats-only / peer kill switch). */
   knownTelegramPolicyHashes: /* @__PURE__ */ new Map(),
+  /**
+   * ENG-9861: per-agent hash of the RENDERED standing-rules section, so a rule
+   * added mid-session is delivered to the running agent rather than sitting on
+   * disk until an unrelated restart. Mirrors the sender/telegram policy maps
+   * above and is persisted through the same sidecar (`agentRules` concern).
+   */
+  knownAgentRulesHashes: /* @__PURE__ */ new Map(),
   /**
    * ENG-6024: hash of the Teams behaviour subset (thread_auto_follow,
    * channel_response_mode, adaptive_cards_enabled/_ask_user_enabled) from the
@@ -6719,6 +6743,7 @@ function clearAgentState(agentId, codeName) {
   agentState.knownManagedMcpStructure.delete(agentId);
   agentState.knownSenderPolicyHashes.delete(agentId);
   agentState.knownTelegramPolicyHashes.delete(agentId);
+  agentState.knownAgentRulesHashes.delete(agentId);
   agentState.knownMsTeamsBehaviourHashes.delete(agentId);
   agentState.knownSlackBehaviourHashes.delete(agentId);
   agentState.knownGlobalSkillIds.delete(agentId);
@@ -12721,6 +12746,11 @@ function scheduleSessionRestart(codeName, delayMs, reason, breakerReason = "hot-
           baselineAgentId,
           agentState.knownTelegramPolicyHashes.get(baselineAgentId)
         );
+        persistDeliveryBaselineEntry(
+          "agentRules",
+          baselineAgentId,
+          agentState.knownAgentRulesHashes.get(baselineAgentId)
+        );
       }
     }
     runningMcpHashes.delete(codeName);
@@ -12870,7 +12900,7 @@ var pendingDayRolloverReset = /* @__PURE__ */ new Set();
 var dayRolloverInboundHold = /* @__PURE__ */ new Map();
 var INBOUND_HOLD_EPISODE_GAP_MS = 12e4;
 async function channelInboundActivityAgeSecondsFor(codeName) {
-  const { newestPendingInboundActivityMtimeMs } = await import("../responsiveness-probe-TT3LAALK.js");
+  const { newestPendingInboundActivityMtimeMs } = await import("../responsiveness-probe-EWUJOLVA.js");
   const newest = newestPendingInboundActivityMtimeMs(dirname10(paneLogPath(codeName)));
   if (newest === null) return null;
   return Math.max(0, Math.floor((Date.now() - newest) / 1e3));
@@ -13530,7 +13560,7 @@ var agentRestartTimezoneInputs = /* @__PURE__ */ new Map();
 var lastVersionCheckAt = 0;
 var VERSION_CHECK_INTERVAL_MS = 5 * 60 * 1e3;
 var lastResponsivenessProbeAt = 0;
-var agtCliVersion = true ? "0.28.779" : "dev";
+var agtCliVersion = true ? "0.28.780" : "dev";
 function resolveBrewPath(execFileSync2) {
   try {
     const out = execFileSync2("which", ["brew"], { timeout: 5e3 }).toString().trim();
@@ -14788,6 +14818,9 @@ function loadSenderPolicyBaseline2() {
   for (const [agentId, hash] of persistedDeliveryBaseline.telegramPolicy) {
     agentState.knownTelegramPolicyHashes.set(agentId, hash);
   }
+  for (const [agentId, hash] of persistedDeliveryBaseline.agentRules) {
+    agentState.knownAgentRulesHashes.set(agentId, hash);
+  }
 }
 var deliveryBaselineDirty = false;
 function saveDeliveryBaseline() {
@@ -15074,7 +15107,7 @@ function flushRestartedAgentDiagnostics(hostId, codeNames) {
   if (codeNames.length === 0) return;
   void (async () => {
     try {
-      const { collectDiagnostics } = await import("../persistent-session-GHIBAWHI.js");
+      const { collectDiagnostics } = await import("../persistent-session-ETZBTCXU.js");
       await api.post("/host/heartbeat", {
         host_id: hostId,
         agent_diagnostics: collectDiagnostics(codeNames, quarantineEntriesFor, claudeMdSizeFor, spawnOutcomeForDiagnostics, pidPressureFor, forwardedToolsFor)
@@ -15215,7 +15248,7 @@ async function pollCycleInner() {
     }
     try {
       const { detectHostSecurity } = await import("../host-security-6PDFG7F5.js");
-      const { collectDiagnostics } = await import("../persistent-session-GHIBAWHI.js");
+      const { collectDiagnostics } = await import("../persistent-session-ETZBTCXU.js");
       const diagCodeNames = [...agentState.persistentSessionAgents];
       const agentDiagnostics = diagCodeNames.length > 0 ? collectDiagnostics(diagCodeNames, quarantineEntriesFor, claudeMdSizeFor, spawnOutcomeForDiagnostics, pidPressureFor, forwardedToolsFor) : void 0;
       let tailscaleHostname;
@@ -15384,7 +15417,7 @@ async function pollCycleInner() {
         collectPanelessActivityProbes,
         getResponsivenessIntervalMs,
         occupancyQualificationClassifications
-      } = await import("../responsiveness-probe-TT3LAALK.js");
+      } = await import("../responsiveness-probe-EWUJOLVA.js");
       const probeIntervalMs = getResponsivenessIntervalMs();
       if (now - lastResponsivenessProbeAt > probeIntervalMs) {
         const probeCodeNames = [...agentState.persistentSessionAgents];
@@ -15494,7 +15527,7 @@ async function pollCycleInner() {
           collectResponsivenessProbes,
           livePendingInboundOldestAgeSeconds,
           parkPendingInbound
-        } = await import("../responsiveness-probe-TT3LAALK.js");
+        } = await import("../responsiveness-probe-EWUJOLVA.js");
         const { getProjectDir: wedgeProjectDir } = await import("../scheduler-engine-NDP36U7O.js");
         const wedgeNow = /* @__PURE__ */ new Date();
         const liveAgents = agentState.persistentSessionAgents;
@@ -15650,7 +15683,7 @@ async function pollCycleInner() {
     }
     try {
       const { scrapeMcpFailedBannerCount } = await import("../pane-mcp-banner-scraper-JA437JIB.js");
-      const { probeSessionAuth } = await import("../session-auth-dead-OHPGSHYL.js");
+      const { probeSessionAuth } = await import("../session-auth-dead-TDPTLVBO.js");
       const observations = [];
       const pendingCacheCommits = [];
       const modelApiErrorReportingOn = hostFlagStore().getBoolean("model-api-error-reporting");
@@ -17332,6 +17365,57 @@ async function processAgent(agent, agentStates, managedToolkits) {
       agentState.knownTelegramPolicyHashes.delete(agent.agent_id);
       removeDeliveryBaselineEntry("telegramPolicy", agent.agent_id);
     }
+    {
+      const agentRules = refreshData.agent_rules ?? [];
+      const agentRulesHash = createHash20("sha256").update(agentRulesRestartBasis(agentRules)).digest("hex");
+      const prevAgentRulesHash = agentState.knownAgentRulesHashes.get(agent.agent_id);
+      const rulesPresent = agentHasRules(agentRules);
+      const agentRulesDecision = decideSenderPolicyRestart({
+        previousHash: prevAgentRulesHash,
+        currentPolicyRestrictive: rulesPresent,
+        currentHash: agentRulesHash,
+        sessionMode: spSessionMode,
+        framework: spFramework,
+        sessionHealthy: spSessionHealthy,
+        // Any earlier restart this tick already respawns the session, which
+        // re-reads CLAUDE.md for free. A second would only coalesce and overwrite
+        // a friendlier notice.
+        channelSetRestartAlreadyScheduled: channelSetRestartScheduled || senderPolicyDecision.restart || msteamsBehaviourRestartScheduled || slackBehaviourRestartScheduled
+      });
+      if (agentRulesDecision.restart) {
+        const isRulesVerifyRestart = agentRulesDecision.firstPoll;
+        log(
+          isRulesVerifyRestart ? `[hot-reload] standing-rules delivery baseline unknown for '${agent.code_name}' with rules set (\u2192 ${agentRulesHash.slice(0, 8)}) \u2014 restarting session so the rendered rules are actually in force (fail closed, ENG-9861)` : `[hot-reload] standing rules changed for '${agent.code_name}' (${prevAgentRulesHash?.slice(0, 8) ?? "first"} \u2192 ${agentRulesHash.slice(0, 8)}) \u2014 restarting session`
+        );
+        const rulesNotice = isRulesVerifyRestart ? "Your manager was updated and needs to restart your session to make sure the standing rules you have been given are actually in force." : "Your standing rules have changed. Restarting session shortly so they take effect - until then you are still running on the previous set.";
+        const rulesDelivered = await injectMessage(
+          agent.code_name,
+          "system",
+          rulesNotice + BACK_ONLINE_GREETING_GUIDANCE,
+          { task_name: "standing-rules-update" },
+          log
+        ).catch(() => false);
+        const rulesDelay = rulesDelivered ? 8e3 : 3e3;
+        if (!rulesDelivered) {
+          log(
+            `[hot-reload] Inject notification unconfirmed for '${agent.code_name}' \u2014 proceeding with shorter delay`
+          );
+        }
+        scheduleSessionRestart(
+          agent.code_name,
+          rulesDelay,
+          isRulesVerifyRestart ? "standing-rules verify (unknown baseline)" : "standing-rules change",
+          "agent-rules-change"
+        );
+      }
+      agentState.knownAgentRulesHashes.set(agent.agent_id, agentRulesHash);
+      const hasLiveRulesConsumer = spSessionHealthy && spSessionMode === "persistent" && spFramework === "claude-code";
+      if (!hasLiveRulesConsumer) {
+        persistDeliveryBaselineEntry("agentRules", agent.agent_id, agentRulesHash);
+      } else if (agentRulesDecision.firstPoll && !rulesPresent) {
+        persistDeliveryBaselineEntry("agentRules", agent.agent_id, agentRulesHash);
+      }
+    }
   }
   const agentSessionMode = refreshData.agent.session_mode;
   if (agentSessionMode === "persistent" && (agentFrameworkCache.get(agent.code_name) ?? DEFAULT_FRAMEWORK) === "claude-code") {
@@ -19690,7 +19774,7 @@ async function handleRestartDoorbell(agentId, requestedAt, restartReason) {
     void api.post("/host/restart-ack", { host_id: hostId, agent_id: agentId, restart_requested_at: requestedAt }).catch((err) => log(`[restart-lane] ack failed for '${codeName}': ${err.message}`));
     void (async () => {
       try {
-        const { collectDiagnostics } = await import("../persistent-session-GHIBAWHI.js");
+        const { collectDiagnostics } = await import("../persistent-session-ETZBTCXU.js");
         await api.post("/host/heartbeat", {
           host_id: hostId,
           agent_diagnostics: collectDiagnostics([codeName], quarantineEntriesFor, claudeMdSizeFor, spawnOutcomeForDiagnostics, pidPressureFor, forwardedToolsFor)
@@ -19736,7 +19820,7 @@ async function respawnAgentAfterMcpStop(codeName, reason) {
     }
     try {
       const hostId = await getHostId();
-      const { collectDiagnostics } = await import("../persistent-session-GHIBAWHI.js");
+      const { collectDiagnostics } = await import("../persistent-session-ETZBTCXU.js");
       await api.post("/host/heartbeat", {
         host_id: hostId,
         agent_diagnostics: collectDiagnostics([codeName], quarantineEntriesFor, claudeMdSizeFor, spawnOutcomeForDiagnostics, pidPressureFor, forwardedToolsFor)
@@ -20372,7 +20456,7 @@ async function processClaudePairSessions(agents) {
     killPairSession,
     pairTmuxSession,
     finalizeClaudePairOnboarding
-  } = await import("../claude-pair-runtime-3J7KK6BP.js");
+  } = await import("../claude-pair-runtime-CK34MZ6M.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));