openmausbot 0.1.72 → 0.1.73

package/dist-server/index.js

@@ -34774,6 +34774,10 @@ var CodexDriver = {
           stdio: ["pipe", "pipe", "pipe"]
         });
         let abandoned = false;
+        let codexThreadId = null;
+        let codexTurnId = null;
+        let startingNativeTurn = false;
+        const earlyNotifications = [];
         const state = {
           settled: false,
           lastText: "",
@@ -34807,6 +34811,18 @@ var CodexDriver = {
           rpcPending.set(id, {
             resolve: (v2) => {
               clearTimeout(timer);
+              if (method === "turn/start") {
+                if (typeof v2?.turn?.id !== "string" || !v2.turn.id) {
+                  reject(new Error("Codex did not return a native turn id"));
+                  return;
+                }
+                codexTurnId = v2.turn.id;
+                startingNativeTurn = false;
+                for (const notification of earlyNotifications.splice(0)) {
+                  if (state.settled) break;
+                  handleNotification(notification);
+                }
+              }
               resolve16(v2);
             },
             reject: (e) => {
@@ -34916,6 +34932,22 @@ var CodexDriver = {
         };
         const handleNotification = (msg) => {
           const p2 = msg.params ?? {};
+          const connectionError = msg.method === "error" && !("threadId" in p2) && !("turnId" in p2);
+          if (!connectionError) {
+            if (!codexThreadId || p2.threadId !== codexThreadId) return;
+            if (!codexTurnId) {
+              if (startingNativeTurn) {
+                if (earlyNotifications.length >= 1024) {
+                  void settle(false, "too_many_events_before_turn_start");
+                } else {
+                  earlyNotifications.push(msg);
+                }
+              }
+              return;
+            }
+            const eventTurnId = msg.method === "turn/started" || msg.method === "turn/completed" ? p2.turn?.id : p2.turnId;
+            if (eventTurnId !== codexTurnId) return;
+          }
           switch (msg.method) {
             // token-level chat text; the item/completed frame follows with the
             // whole message, so its delta is only a fallback when none streamed
@@ -35100,7 +35132,6 @@ var CodexDriver = {
           }
           autoAcceptPermissions = approvalMode === "full";
           const cursor = typeof turn.resumeCursor === "string" ? turn.resumeCursor : null;
-          let codexThreadId = null;
           let startedModel = null;
           if (cursor) {
             try {
@@ -35153,21 +35184,24 @@ var CodexDriver = {
             ...promptText ? [{ type: "text", text: promptText }] : [],
             ...(turn.images ?? []).map((image) => ({ type: "localImage", path: image.path }))
           ];
-          const startTurn2 = () => request2("turn/start", {
-            threadId: codexThreadId,
-            input: turnInput,
-            ...approvalParams.turn,
-            // Spread, not `effort: turn.effort ?? null`. Probed against
-            // codex-cli 0.146.0: null is indistinguishable from an absent key
-            // — both leave the thread's current effort alone, emitting no
-            // thread/settings/updated, and thread/resume reads the old value
-            // back. The app-server offers no way to clear a level either:
-            // "" is rejected outright and thread/start takes no effort at
-            // all. So a thread keeps the last level it was sent until it is
-            // sent another, and choosing Default lands on the bot's next new
-            // thread rather than the current one.
-            ...turn.effort ? { effort: turn.effort } : {}
-          });
+          const startTurn2 = () => {
+            startingNativeTurn = true;
+            return request2("turn/start", {
+              threadId: codexThreadId,
+              input: turnInput,
+              ...approvalParams.turn,
+              // Spread, not `effort: turn.effort ?? null`. Probed against
+              // codex-cli 0.146.0: null is indistinguishable from an absent key
+              // — both leave the thread's current effort alone, emitting no
+              // thread/settings/updated, and thread/resume reads the old value
+              // back. The app-server offers no way to clear a level either:
+              // "" is rejected outright and thread/start takes no effort at
+              // all. So a thread keeps the last level it was sent until it is
+              // sent another, and choosing Default lands on the bot's next new
+              // thread rather than the current one.
+              ...turn.effort ? { effort: turn.effort } : {}
+            });
+          };
           try {
             await startTurn2();
           } catch (error51) {
@@ -51101,6 +51135,13 @@ var MANAGED_VPS_CONTAINER_NAME = /^openmausbot-vps-[a-z0-9]{1,12}-[a-f0-9]{12}$/
 var IMAGE_ID = /^sha256:[a-f0-9]{64}$/i;
 var PIDS_LIMIT2 = 512;
 var SCREENSHOT_PATH = "/tmp/openmausbot-vps-preview.png";
+var SCREENSHOT_TRANSFER = `/opt/venv/bin/python -I -c 'import base64, io, sys
+from PIL import Image
+with Image.open(sys.argv[1]) as image:
+    image.thumbnail((1280, 1280))
+    output = io.BytesIO()
+    image.convert("RGB").save(output, format="JPEG", quality=70)
+sys.stdout.write(base64.b64encode(output.getvalue()).decode("ascii"))' "$1" 2>/dev/null || { base64 < "$1" | tr -d "\\n"; }`;
 var INTERNAL_VIEWER_PORT2 = 6901;
 var VIEWER_VERSION = "1";
 var lifecycleLocks = /* @__PURE__ */ new Map();
@@ -51114,6 +51155,9 @@ function vpsLifecycleBusy() {
 var LOCK_ACQUIRE_TIMEOUT_MS = 5e3;
 var STATUS_CACHE_TTL_MS = 1e4;
 var statusCache = /* @__PURE__ */ new Map();
+var SCREENSHOT_BUDGET_MS = 45e3;
+var SCREENSHOT_CLEANUP_BUDGET_MS = 1e4;
+var pendingScreenshots = /* @__PURE__ */ new Map();
 var viewerConnections = /* @__PURE__ */ new Map();
 var desktopTunnels = /* @__PURE__ */ new Map();
 function containerNamePart(botId) {
@@ -51448,6 +51492,8 @@ async function vpsComputerStatus(cfg2, botId, runner = defaultRunner) {
   const key3 = vpsLockKey(cfg2, botId);
   const cacheable = runner === defaultRunner && key3 !== null;
   if (cacheable) {
+    await pendingScreenshots.get(key3)?.catch(() => {
+    });
     const cached3 = statusCache.get(key3);
     if (cached3 && cached3.expiresAt > Date.now()) return cached3.status;
   }
@@ -51892,21 +51938,49 @@ async function vpsComputerScreenshot(cfg2, botId, runner = defaultRunner) {
   const alias = vpsSshAlias(cfg2);
   if (!alias) throw Object.assign(new Error("VPS is not configured"), { status: 409 });
   const key3 = `${alias}:${vpsContainerName(botId)}`;
+  const pending = pendingScreenshots.get(key3);
+  if (pending) return pending;
   const cacheable = runner === defaultRunner;
-  return withVpsLifecycleLock(key3, async () => {
+  const deadline = Date.now() + SCREENSHOT_BUDGET_MS;
+  const workDeadline = deadline - SCREENSHOT_CLEANUP_BUDGET_MS;
+  let budgetExpired = false;
+  const timeoutError = () => Object.assign(new Error("The VPS screen preview timed out. Retry the preview when the connection recovers."), { status: 504 });
+  const boundedRunner = async (args, options = {}) => {
+    const remaining = workDeadline - Date.now() - COMMAND_TIMEOUT_KILL_GRACE_MS;
+    if (remaining <= 0) {
+      budgetExpired = true;
+      throw timeoutError();
+    }
+    try {
+      const result = await runner(args, { ...options, timeoutMs: Math.min(options.timeoutMs ?? 12e4, remaining) });
+      if (Date.now() >= workDeadline) {
+        budgetExpired = true;
+        throw timeoutError();
+      }
+      return result;
+    } catch (error51) {
+      if (Date.now() >= workDeadline - COMMAND_TIMEOUT_KILL_GRACE_MS) budgetExpired = true;
+      throw budgetExpired ? timeoutError() : error51;
+    }
+  };
+  const capture = withVpsLifecycleLock(key3, async () => {
     const cached3 = cacheable ? statusCache.get(key3) : void 0;
-    const status = cached3 && cached3.expiresAt > Date.now() ? cached3.status : await computeVpsComputerStatus(cfg2, botId, runner);
+    const status = cached3 && cached3.expiresAt > Date.now() ? cached3.status : await computeVpsComputerStatus(cfg2, botId, boundedRunner);
+    if (budgetExpired) {
+      if (cacheable) statusCache.delete(key3);
+      throw timeoutError();
+    }
     if (!status.ready) {
       if (cacheable) statusCache.delete(key3);
       throw Object.assign(new Error(status.problem ?? "The VPS computer is not ready"), { status: 409 });
     }
-    if (cacheable) statusCache.set(key3, { status, expiresAt: Date.now() + STATUS_CACHE_TTL_MS });
     const containerRef = status.container_id ?? status.container_name;
     if (!CONTAINER_ID.test(containerRef) && !CONTAINER_NAME.test(containerRef)) {
       throw Object.assign(new Error("the VPS container reference is malformed"), { status: 409 });
     }
+    let frame;
     try {
-      await runner(
+      await boundedRunner(
         vpsDockerArgs(
           alias,
           cuaExecArgs(
@@ -51916,30 +51990,41 @@ async function vpsComputerScreenshot(cfg2, botId, runner = defaultRunner) {
         ),
         { timeoutMs: 3e4 }
       );
-      const encoded = (await runner(vpsDockerArgs(alias, [
+      const encoded = (await boundedRunner(vpsDockerArgs(alias, [
         "exec",
         "-u",
         "cua",
         "-e",
         "HOME=/home/cua",
         containerRef,
-        "base64",
-        "-w0",
+        "sh",
+        "-c",
+        SCREENSHOT_TRANSFER,
+        "openmausbot-preview",
         SCREENSHOT_PATH
       ]), { timeoutMs: 3e4 })).stdout.trim();
       const checked = wholeScreenshot(Buffer.from(encoded, "base64"));
       if (!checked.ok) throw Object.assign(new Error("Cua Driver returned an incomplete VPS screenshot"), { status: 502 });
-      return { png: encoded, format: checked.mime === "image/jpeg" ? "jpeg" : "png" };
+      frame = { png: encoded, format: checked.mime === "image/jpeg" ? "jpeg" : "png" };
     } catch (error51) {
       if (cacheable) statusCache.delete(key3);
       throw error51;
     } finally {
-      await runner(vpsDockerArgs(alias, ["exec", "-u", "cua", containerRef, "rm", "-f", SCREENSHOT_PATH]), {
-        timeoutMs: 1e4
+      const cleanupMs = Math.min(5e3, deadline - Date.now() - COMMAND_TIMEOUT_KILL_GRACE_MS);
+      if (cleanupMs > 0) await runner(vpsDockerArgs(alias, ["exec", "-u", "cua", containerRef, "rm", "-f", SCREENSHOT_PATH]), {
+        timeoutMs: cleanupMs
       }).catch(() => {
       });
     }
+    if (cacheable) statusCache.set(key3, { status, expiresAt: Date.now() + STATUS_CACHE_TTL_MS });
+    return frame;
   });
+  pendingScreenshots.set(key3, capture);
+  try {
+    return await capture;
+  } finally {
+    if (pendingScreenshots.get(key3) === capture) pendingScreenshots.delete(key3);
+  }
 }
 
 // server/routines.ts
@@ -69491,6 +69576,7 @@ var localVmProvisionBusy = false;
 var localVmModeChangeBusy = false;
 var activeVpsThreads = /* @__PURE__ */ new Map();
 var boxLifecycleBusyBots = /* @__PURE__ */ new Set();
+var vpsPreviewRequests = /* @__PURE__ */ new Map();
 var orphanBoxLifecycleBusyIds = /* @__PURE__ */ new Set();
 var boxInventoryRequestsBusyIds = /* @__PURE__ */ new Set();
 var computerProviderConfigTransitions = /* @__PURE__ */ new Set();
@@ -69522,7 +69608,7 @@ function providerOperationConflict(provider) {
   if (managedBoxOwners().some((owner2) => owner2.inUse)) {
     return `stop active bot work and computer control before changing ${provider === "box" ? "the Box account" : "the VPS connection"}`;
   }
-  if (boxLifecycleBusyBots.size > 0) {
+  if (boxLifecycleBusyBots.size > 0 || vpsPreviewRequests.size > 0) {
     return "wait for cloud computer actions to finish before changing provider settings";
   }
   if (provider === "box") {
@@ -69568,10 +69654,13 @@ function claimManagedBoxMutation(instance) {
   }
   return claimBotComputerLifecycle(ownerBotId);
 }
-function claimBotComputerLifecycle(botId) {
+function claimBotComputerLifecycle(botId, allowPreview = false) {
   if (boxLifecycleBusyBots.has(botId)) {
     throw Object.assign(new Error("this bot's cloud computer is being changed \u2014 wait for it to finish"), { status: 409 });
   }
+  if (!allowPreview && vpsPreviewRequests.has(botId)) {
+    throw Object.assign(new Error("a screen preview is still refreshing \u2014 wait before changing this computer"), { status: 409 });
+  }
   boxLifecycleBusyBots.add(botId);
   return () => boxLifecycleBusyBots.delete(botId);
 }
@@ -73491,7 +73580,7 @@ var workspaceBackupRoutes = createWorkspaceBackupRoutes({
     return Boolean(current2?.scopes.includes("admin") && current2.kind === original.kind && (current2.kind !== "session" || original.kind === "session" && current2.session.id === original.session.id));
   },
   exclusive: (work, keepLocked) => workspaceMaintenance.run(work, {
-    idle: () => !providerFleetReloading && !providerAuthSessions.active && !browserEngineInstall && !routines?.isTicking && !calendarCalls?.isTicking && !localVmImageBusy && !localVmProvisionBusy && !localVmModeChangeBusy && !localVmLifecycleBusy.size && !boxLifecycleBusyBots.size && !orphanBoxLifecycleBusyIds.size && !computerProviderConfigTransitions.size && !checkpointRestoreLeases.size && store.bots.every((bot) => !botHasActiveTurn(bot.id) && !routines?.activeRunForBot(bot.id) && !computerControl.snapshot(bot.id).held) && store.groups.every((group) => !groupIsWorking(group)),
+    idle: () => !providerFleetReloading && !providerAuthSessions.active && !browserEngineInstall && !routines?.isTicking && !calendarCalls?.isTicking && !localVmImageBusy && !localVmProvisionBusy && !localVmModeChangeBusy && !localVmLifecycleBusy.size && !boxLifecycleBusyBots.size && !vpsPreviewRequests.size && !orphanBoxLifecycleBusyIds.size && !computerProviderConfigTransitions.size && !checkpointRestoreLeases.size && store.bots.every((bot) => !botHasActiveTurn(bot.id) && !routines?.activeRunForBot(bot.id) && !computerControl.snapshot(bot.id).held) && store.groups.every((group) => !groupIsWorking(group)),
     pause: () => {
       routines?.stop();
       calendarCalls?.stop();
@@ -78596,7 +78685,17 @@ data: ${JSON.stringify({ kind: "ping" })}
         return json3(res, 409, { error: "this bot's cloud computer is being changed \u2014 wait for it to finish" });
       }
       if (bot.cloudBackend === "vps") {
-        const releaseComputerLifecycle2 = claimBotComputerLifecycle(botId);
+        if (m2[2] === "screenshot") {
+          let preview = vpsPreviewRequests.get(botId);
+          if (!preview) {
+            preview = vpsComputerScreenshot(cfg, botId).finally(() => {
+              vpsPreviewRequests.delete(botId);
+            });
+            vpsPreviewRequests.set(botId, preview);
+          }
+          return json3(res, 200, await preview);
+        }
+        const releaseComputerLifecycle2 = claimBotComputerLifecycle(botId, m2[2] === "join");
         try {
           if (m2[2] === "exec") {
             return json3(res, 409, { error: "the VPS console is available to the bot through its scoped computer tools" });
@@ -78610,7 +78709,6 @@ data: ${JSON.stringify({ kind: "ping" })}
           if (m2[2] === "join") {
             return json3(res, 200, await vpsComputerJoin(cfg, botId));
           }
-          if (m2[2] === "screenshot") return json3(res, 200, await vpsComputerScreenshot(cfg, botId));
           const action = m2[2] === "provision" ? "provision" : m2[2] === "remove" ? "remove" : "stop";
           return json3(res, 200, await vpsComputerAction(action, cfg, botId));
         } finally {

package/dist-server/openmausbot.js

@@ -24762,6 +24762,10 @@ var init_codex = __esm({
               stdio: ["pipe", "pipe", "pipe"]
             });
             let abandoned = false;
+            let codexThreadId = null;
+            let codexTurnId = null;
+            let startingNativeTurn = false;
+            const earlyNotifications = [];
             const state = {
               settled: false,
               lastText: "",
@@ -24795,6 +24799,18 @@ var init_codex = __esm({
               rpcPending.set(id, {
                 resolve: (v) => {
                   clearTimeout(timer);
+                  if (method === "turn/start") {
+                    if (typeof v?.turn?.id !== "string" || !v.turn.id) {
+                      reject(new Error("Codex did not return a native turn id"));
+                      return;
+                    }
+                    codexTurnId = v.turn.id;
+                    startingNativeTurn = false;
+                    for (const notification of earlyNotifications.splice(0)) {
+                      if (state.settled) break;
+                      handleNotification(notification);
+                    }
+                  }
                   resolve13(v);
                 },
                 reject: (e) => {
@@ -24904,6 +24920,22 @@ var init_codex = __esm({
             };
             const handleNotification = (msg) => {
               const p = msg.params ?? {};
+              const connectionError = msg.method === "error" && !("threadId" in p) && !("turnId" in p);
+              if (!connectionError) {
+                if (!codexThreadId || p.threadId !== codexThreadId) return;
+                if (!codexTurnId) {
+                  if (startingNativeTurn) {
+                    if (earlyNotifications.length >= 1024) {
+                      void settle(false, "too_many_events_before_turn_start");
+                    } else {
+                      earlyNotifications.push(msg);
+                    }
+                  }
+                  return;
+                }
+                const eventTurnId = msg.method === "turn/started" || msg.method === "turn/completed" ? p.turn?.id : p.turnId;
+                if (eventTurnId !== codexTurnId) return;
+              }
               switch (msg.method) {
                 // token-level chat text; the item/completed frame follows with the
                 // whole message, so its delta is only a fallback when none streamed
@@ -25088,7 +25120,6 @@ var init_codex = __esm({
               }
               autoAcceptPermissions = approvalMode === "full";
               const cursor3 = typeof turn.resumeCursor === "string" ? turn.resumeCursor : null;
-              let codexThreadId = null;
               let startedModel = null;
               if (cursor3) {
                 try {
@@ -25141,21 +25172,24 @@ var init_codex = __esm({
                 ...promptText ? [{ type: "text", text: promptText }] : [],
                 ...(turn.images ?? []).map((image) => ({ type: "localImage", path: image.path }))
               ];
-              const startTurn = () => request("turn/start", {
-                threadId: codexThreadId,
-                input: turnInput,
-                ...approvalParams.turn,
-                // Spread, not `effort: turn.effort ?? null`. Probed against
-                // codex-cli 0.146.0: null is indistinguishable from an absent key
-                // — both leave the thread's current effort alone, emitting no
-                // thread/settings/updated, and thread/resume reads the old value
-                // back. The app-server offers no way to clear a level either:
-                // "" is rejected outright and thread/start takes no effort at
-                // all. So a thread keeps the last level it was sent until it is
-                // sent another, and choosing Default lands on the bot's next new
-                // thread rather than the current one.
-                ...turn.effort ? { effort: turn.effort } : {}
-              });
+              const startTurn = () => {
+                startingNativeTurn = true;
+                return request("turn/start", {
+                  threadId: codexThreadId,
+                  input: turnInput,
+                  ...approvalParams.turn,
+                  // Spread, not `effort: turn.effort ?? null`. Probed against
+                  // codex-cli 0.146.0: null is indistinguishable from an absent key
+                  // — both leave the thread's current effort alone, emitting no
+                  // thread/settings/updated, and thread/resume reads the old value
+                  // back. The app-server offers no way to clear a level either:
+                  // "" is rejected outright and thread/start takes no effort at
+                  // all. So a thread keeps the last level it was sent until it is
+                  // sent another, and choosing Default lands on the bot's next new
+                  // thread rather than the current one.
+                  ...turn.effort ? { effort: turn.effort } : {}
+                });
+              };
               try {
                 await startTurn();
               } catch (error51) {

package/dist-server/pair-cli.js

@@ -24762,6 +24762,10 @@ var init_codex = __esm({
               stdio: ["pipe", "pipe", "pipe"]
             });
             let abandoned = false;
+            let codexThreadId = null;
+            let codexTurnId = null;
+            let startingNativeTurn = false;
+            const earlyNotifications = [];
             const state = {
               settled: false,
               lastText: "",
@@ -24795,6 +24799,18 @@ var init_codex = __esm({
               rpcPending.set(id, {
                 resolve: (v) => {
                   clearTimeout(timer);
+                  if (method === "turn/start") {
+                    if (typeof v?.turn?.id !== "string" || !v.turn.id) {
+                      reject(new Error("Codex did not return a native turn id"));
+                      return;
+                    }
+                    codexTurnId = v.turn.id;
+                    startingNativeTurn = false;
+                    for (const notification of earlyNotifications.splice(0)) {
+                      if (state.settled) break;
+                      handleNotification(notification);
+                    }
+                  }
                   resolve13(v);
                 },
                 reject: (e) => {
@@ -24904,6 +24920,22 @@ var init_codex = __esm({
             };
             const handleNotification = (msg) => {
               const p = msg.params ?? {};
+              const connectionError = msg.method === "error" && !("threadId" in p) && !("turnId" in p);
+              if (!connectionError) {
+                if (!codexThreadId || p.threadId !== codexThreadId) return;
+                if (!codexTurnId) {
+                  if (startingNativeTurn) {
+                    if (earlyNotifications.length >= 1024) {
+                      void settle(false, "too_many_events_before_turn_start");
+                    } else {
+                      earlyNotifications.push(msg);
+                    }
+                  }
+                  return;
+                }
+                const eventTurnId = msg.method === "turn/started" || msg.method === "turn/completed" ? p.turn?.id : p.turnId;
+                if (eventTurnId !== codexTurnId) return;
+              }
               switch (msg.method) {
                 // token-level chat text; the item/completed frame follows with the
                 // whole message, so its delta is only a fallback when none streamed
@@ -25088,7 +25120,6 @@ var init_codex = __esm({
               }
               autoAcceptPermissions = approvalMode === "full";
               const cursor3 = typeof turn.resumeCursor === "string" ? turn.resumeCursor : null;
-              let codexThreadId = null;
               let startedModel = null;
               if (cursor3) {
                 try {
@@ -25141,21 +25172,24 @@ var init_codex = __esm({
                 ...promptText ? [{ type: "text", text: promptText }] : [],
                 ...(turn.images ?? []).map((image) => ({ type: "localImage", path: image.path }))
               ];
-              const startTurn = () => request("turn/start", {
-                threadId: codexThreadId,
-                input: turnInput,
-                ...approvalParams.turn,
-                // Spread, not `effort: turn.effort ?? null`. Probed against
-                // codex-cli 0.146.0: null is indistinguishable from an absent key
-                // — both leave the thread's current effort alone, emitting no
-                // thread/settings/updated, and thread/resume reads the old value
-                // back. The app-server offers no way to clear a level either:
-                // "" is rejected outright and thread/start takes no effort at
-                // all. So a thread keeps the last level it was sent until it is
-                // sent another, and choosing Default lands on the bot's next new
-                // thread rather than the current one.
-                ...turn.effort ? { effort: turn.effort } : {}
-              });
+              const startTurn = () => {
+                startingNativeTurn = true;
+                return request("turn/start", {
+                  threadId: codexThreadId,
+                  input: turnInput,
+                  ...approvalParams.turn,
+                  // Spread, not `effort: turn.effort ?? null`. Probed against
+                  // codex-cli 0.146.0: null is indistinguishable from an absent key
+                  // — both leave the thread's current effort alone, emitting no
+                  // thread/settings/updated, and thread/resume reads the old value
+                  // back. The app-server offers no way to clear a level either:
+                  // "" is rejected outright and thread/start takes no effort at
+                  // all. So a thread keeps the last level it was sent until it is
+                  // sent another, and choosing Default lands on the bot's next new
+                  // thread rather than the current one.
+                  ...turn.effort ? { effort: turn.effort } : {}
+                });
+              };
               try {
                 await startTurn();
               } catch (error51) {

package/dist-server/server/drivers/agents-proxy.js

@@ -531,7 +531,7 @@ const TOOLS = [
             additionalProperties: false,
             properties: {
                 action: { type: "string", enum: ["append", "replace", "remove", "supersede"] },
-                text: { type: "string", minLength: 1, pattern: "\\S", description: "Non-blank new text for append, replace, or supersede: the fact itself, without a date or bullet. Omit for remove; use remove to delete a passage." },
+                text: { type: "string", minLength: 1, description: "Non-blank new text for append, replace, or supersede: the fact itself, without a date or bullet. Omit for remove; use remove to delete a passage." },
                 old_text: { type: "string", minLength: 1, description: "Exact unique existing passage for replace, supersede, or remove. Omit for append." },
             },
             required: ["action"],
@@ -544,7 +544,7 @@ const TOOLS = [
             type: "object",
             additionalProperties: false,
             properties: {
-                text: { type: "string", minLength: 1, pattern: "\\S", description: "One line about what happened, in plain words." },
+                text: { type: "string", minLength: 1, description: "One line about what happened, in plain words." },
             },
             required: ["text"],
         },

package/dist-server/server/drivers/codex.js

@@ -515,6 +515,10 @@ export const CodexDriver = {
                     stdio: ["pipe", "pipe", "pipe"],
                 });
                 let abandoned = false;
+                let codexThreadId = null;
+                let codexTurnId = null;
+                let startingNativeTurn = false;
+                const earlyNotifications = [];
                 const state = {
                     settled: false,
                     lastText: "",
@@ -553,6 +557,21 @@ export const CodexDriver = {
                     rpcPending.set(id, {
                         resolve: (v) => {
                             clearTimeout(timer);
+                            if (method === "turn/start") {
+                                if (typeof v?.turn?.id !== "string" || !v.turn.id) {
+                                    reject(new Error("Codex did not return a native turn id"));
+                                    return;
+                                }
+                                // Bind synchronously: a single stdout chunk can contain the
+                                // response, streamed events, completion and a late request.
+                                codexTurnId = v.turn.id;
+                                startingNativeTurn = false;
+                                for (const notification of earlyNotifications.splice(0)) {
+                                    if (state.settled)
+                                        break;
+                                    handleNotification(notification);
+                                }
+                            }
                             resolve(v);
                         },
                         reject: (e) => {
@@ -710,6 +729,33 @@ export const CodexDriver = {
                 };
                 const handleNotification = (msg) => {
                     const p = msg.params ?? {};
+                    // An app-server also emits notifications for native helper threads.
+                    // Only this request's parent may write its transcript/usage or settle
+                    // its run. Requests still use the approval broker above, including
+                    // helper requests; ignoring child *notifications* must not grant tools.
+                    const connectionError = msg.method === "error" &&
+                        !("threadId" in p) && !("turnId" in p);
+                    if (!connectionError) {
+                        if (!codexThreadId || p.threadId !== codexThreadId)
+                            return;
+                        if (!codexTurnId) {
+                            // Some servers stream before acknowledging turn/start. Retain a
+                            // bounded prefix, then filter against the authoritative response.
+                            if (startingNativeTurn) {
+                                if (earlyNotifications.length >= 1024) {
+                                    void settle(false, "too_many_events_before_turn_start");
+                                }
+                                else {
+                                    earlyNotifications.push(msg);
+                                }
+                            }
+                            return;
+                        }
+                        const eventTurnId = msg.method === "turn/started" || msg.method === "turn/completed"
+                            ? p.turn?.id : p.turnId;
+                        if (eventTurnId !== codexTurnId)
+                            return;
+                    }
                     switch (msg.method) {
                         // token-level chat text; the item/completed frame follows with the
                         // whole message, so its delta is only a fallback when none streamed
@@ -959,7 +1005,6 @@ export const CodexDriver = {
                     // on start AND resume so Codex owns their lifetime through compaction.
                     // Removed bot rules are cleared without dropping native configured rules.
                     const cursor = typeof turn.resumeCursor === "string" ? turn.resumeCursor : null;
-                    let codexThreadId = null;
                     let startedModel = null;
                     if (cursor) {
                         try {
@@ -1020,21 +1065,24 @@ export const CodexDriver = {
                         ...(promptText ? [{ type: "text", text: promptText }] : []),
                         ...(turn.images ?? []).map((image) => ({ type: "localImage", path: image.path })),
                     ];
-                    const startTurn = () => request("turn/start", {
-                        threadId: codexThreadId,
-                        input: turnInput,
-                        ...approvalParams.turn,
-                        // Spread, not `effort: turn.effort ?? null`. Probed against
-                        // codex-cli 0.146.0: null is indistinguishable from an absent key
-                        // — both leave the thread's current effort alone, emitting no
-                        // thread/settings/updated, and thread/resume reads the old value
-                        // back. The app-server offers no way to clear a level either:
-                        // "" is rejected outright and thread/start takes no effort at
-                        // all. So a thread keeps the last level it was sent until it is
-                        // sent another, and choosing Default lands on the bot's next new
-                        // thread rather than the current one.
-                        ...(turn.effort ? { effort: turn.effort } : {}),
-                    });
+                    const startTurn = () => {
+                        startingNativeTurn = true;
+                        return request("turn/start", {
+                            threadId: codexThreadId,
+                            input: turnInput,
+                            ...approvalParams.turn,
+                            // Spread, not `effort: turn.effort ?? null`. Probed against
+                            // codex-cli 0.146.0: null is indistinguishable from an absent key
+                            // — both leave the thread's current effort alone, emitting no
+                            // thread/settings/updated, and thread/resume reads the old value
+                            // back. The app-server offers no way to clear a level either:
+                            // "" is rejected outright and thread/start takes no effort at
+                            // all. So a thread keeps the last level it was sent until it is
+                            // sent another, and choosing Default lands on the bot's next new
+                            // thread rather than the current one.
+                            ...(turn.effort ? { effort: turn.effort } : {}),
+                        });
+                    };
                     try {
                         await startTurn();
                     }