@cloudcli-ai/cloudcli 1.37.0 → 1.37.2

package/server/modules/providers/list/claude/claude-models.provider.ts

@@ -9,12 +9,12 @@ import type {
 } from '@/shared/types.js';
 import { buildDefaultProviderCurrentActiveModel } from '@/shared/utils.js';
 
-export const CLAUDE_FALLBACK_MODELS: ProviderModelsDefinition = {
+export const CLAUDE_PREDEFINED_MODELS: ProviderModelsDefinition = {
   OPTIONS: [
     {
       value: 'default',
       label: 'Default (recommended)',
-      description: 'Use the Claude Code default model (currently Sonnet 4.6)',
+      description: 'Use the recommended model for your Claude account and deployment.',
       effort: {
         default: 'high',
         values: [
@@ -25,10 +25,25 @@ export const CLAUDE_FALLBACK_MODELS: ProviderModelsDefinition = {
         ],
       },
     },
+    {
+      value: 'best',
+      label: 'Best available',
+      description: 'Use Fable 5 when available, otherwise the latest Opus model.',
+      effort: {
+        default: 'high',
+        values: [
+          { value: 'low' },
+          { value: 'medium' },
+          { value: 'high' },
+          { value: 'xhigh' },
+          { value: 'max' },
+        ],
+      },
+    },
     {
       value: 'fable',
-      label: 'Fable',
-      description: 'Fable 5 · Most capable for your hardest and longest-running tasks · Uses your limits ~2× faster than Opus',
+      label: 'Fable 5',
+      description: 'Most capable Claude model for the hardest, longest-running tasks.',
       effort: {
         default: 'high',
         values: [
@@ -41,15 +56,16 @@ export const CLAUDE_FALLBACK_MODELS: ProviderModelsDefinition = {
       },
     },
     {
-      value: "sonnet",
-      label: "Sonnet",
-      description: "Sonnet 4.6 · Best for everyday tasks · $3/$15 per Mtok",
+      value: 'sonnet',
+      label: 'Sonnet',
+      description: 'Latest Sonnet model for everyday coding tasks.',
       effort: {
         default: 'high',
         values: [
           { value: 'low' },
           { value: 'medium' },
           { value: 'high' },
+          { value: 'xhigh' },
           { value: 'max' },
         ],
       },
@@ -57,13 +73,14 @@ export const CLAUDE_FALLBACK_MODELS: ProviderModelsDefinition = {
     {
       value: 'sonnet[1m]',
       label: 'Sonnet (1M context)',
-      description: 'Sonnet 4.6 for long sessions · $3/$15 per Mtok',
+      description: 'Latest Sonnet model with a 1M context window.',
       effort: {
         default: 'high',
         values: [
           { value: 'low' },
           { value: 'medium' },
           { value: 'high' },
+          { value: 'xhigh' },
           { value: 'max' },
         ],
       },
@@ -71,7 +88,7 @@ export const CLAUDE_FALLBACK_MODELS: ProviderModelsDefinition = {
     {
       value: 'opus',
       label: 'Opus',
-      description: 'Opus 4.8 · Best for everyday, complex tasks · ~2× usage vs Sonnet',
+      description: 'Latest Opus model for complex reasoning and coding tasks.',
       effort: {
         default: 'high',
         values: [
@@ -85,8 +102,8 @@ export const CLAUDE_FALLBACK_MODELS: ProviderModelsDefinition = {
     },
     {
       value: 'opus[1m]',
-      label: 'Opus 4.8 (1M context)',
-      description: 'Opus 4.8 with 1M context · Most capable for complex work · $5/$25 per Mtok',
+      label: 'Opus (1M context)',
+      description: 'Latest Opus model with a 1M context window.',
       effort: {
         default: 'high',
         values: [
@@ -101,7 +118,22 @@ export const CLAUDE_FALLBACK_MODELS: ProviderModelsDefinition = {
     {
       value: 'haiku',
       label: 'Haiku',
-      description: 'Haiku 4.5 · Fastest for quick answers · $1/$5 per Mtok',
+      description: 'Fast and efficient Claude model for simple tasks.',
+    },
+    {
+      value: 'opusplan',
+      label: 'Opus Plan',
+      description: 'Use Opus while planning, then switch to Sonnet for execution.',
+      effort: {
+        default: 'high',
+        values: [
+          { value: 'low' },
+          { value: 'medium' },
+          { value: 'high' },
+          { value: 'xhigh' },
+          { value: 'max' },
+        ],
+      },
     },
   ],
   DEFAULT: 'default',
@@ -113,7 +145,7 @@ export const findClaudeModelOption = (model: string | undefined | null): Provide
     return null;
   }
 
-  return CLAUDE_FALLBACK_MODELS.OPTIONS.find((option) => option.value === normalizedModel) ?? null;
+  return CLAUDE_PREDEFINED_MODELS.OPTIONS.find((option) => option.value === normalizedModel) ?? null;
 };
 type ClaudeInitEvent = {
   sessionId?: string;
@@ -237,7 +269,7 @@ export class ClaudeProviderModels implements IProviderModels {
     // const supportedModels = await queryInstance.supportedModels();
     // queryInstance.close();
     // return buildClaudeModelsDefinition(supportedModels);
-    return CLAUDE_FALLBACK_MODELS;
+    return CLAUDE_PREDEFINED_MODELS;
   }
 
   async getCurrentActiveModel(sessionId?: string): Promise<ProviderCurrentActiveModel> {

package/server/modules/providers/list/claude/claude-runtime.provider.js

@@ -24,10 +24,11 @@ import {
   buildClaudeUserContent,
   normalizeImageDescriptors
 } from '@/shared/image-attachments.js';
-import { CLAUDE_FALLBACK_MODELS } from '@/modules/providers/list/claude/claude-models.provider.js';
+import { CLAUDE_PREDEFINED_MODELS } from '@/modules/providers/list/claude/claude-models.provider.js';
 import { resolveClaudeCodeExecutablePath } from '@/shared/claude-cli-path.js';
 import {
   createNotificationEvent,
+  notifyBackgroundWorkCompleted,
   notifyRunFailed,
   notifyRunStopped,
   notifyUserIfEnabled
@@ -40,12 +41,35 @@ const pendingToolApprovals = new Map();
 // terminal `complete` (aborted: true) to the client, so the run loop must not
 // emit a second one when its generator winds down.
 const abortedSessionIds = new Set();
+// Query instances interrupted because a newer run took over their session id
+// (see addSession). Their run loops must stay silent on wind-down: the map
+// entry, the abort flag, and all client-facing events belong to the new run.
+const supersededInstances = new WeakSet();
 
 const TOOL_APPROVAL_TIMEOUT_MS = parseInt(process.env.CLAUDE_TOOL_APPROVAL_TIMEOUT_MS, 10) || 55000;
 
+// How long background work is allowed to keep running after a turn ends. This drives
+// two halves of the same behaviour:
+//
+//  1. Passed to the spawned CLI as CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS, which is how
+//     long it waits for still-running background *agents* before killing them.
+//  2. A backstop on how long we hold the SDK's stdin open after a turn's `result`.
+//     The SDK closes stdin as soon as a turn ends, and the CLI reads that EOF as
+//     "print wind-down" — killing background *shells* after a short grace period,
+//     which the ceiling above does not cover. Holding stdin open also lets the CLI
+//     push follow-up turns (background-task completions, Monitor notifications,
+//     scheduled wake-ups).
+//
+// The hold normally ends long before this: a turn with nothing outstanding closes
+// stdin immediately, background work releases it as soon as it reports back, and a
+// new turn supersedes the previous hold. This ceiling only catches background work
+// that never reports at all, so an abandoned session cannot leak a CLI process
+// forever. The timer resets on every message, so it measures silence, not total time.
+const BG_WAIT_CEILING_MS = 30 * 60 * 1000;
+
 const TOOLS_REQUIRING_INTERACTION = new Set(['AskUserQuestion', 'ExitPlanMode']);
 
-function resolveClaudeEffort(model, effort, modelsDefinition = CLAUDE_FALLBACK_MODELS) {
+function resolveClaudeEffort(model, effort, modelsDefinition = CLAUDE_PREDEFINED_MODELS) {
   const selectedModel = modelsDefinition?.OPTIONS?.find((option) => option.value === model) || null;
   const allowedEfforts = selectedModel?.effort?.values
     ?.map((value) => value.value) || [];
@@ -165,7 +189,7 @@ function mapCliOptionsToSDK(options = {}) {
 
   // Forward all host env vars (e.g. ANTHROPIC_BASE_URL) to the subprocess.
   // Since SDK 0.2.113, options.env replaces process.env instead of overlaying it.
-  sdkOptions.env = { ...process.env };
+  sdkOptions.env = { ...process.env, CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS: String(BG_WAIT_CEILING_MS) };
 
   // Resolve the executable eagerly on Windows because the SDK uses raw child_process.spawn,
   // which does not reliably follow npm's shell wrappers like cross-spawn does.
@@ -209,12 +233,12 @@ function mapCliOptionsToSDK(options = {}) {
 
   sdkOptions.disallowedTools = settings.disallowedTools || [];
 
-  sdkOptions.model = options.model || CLAUDE_FALLBACK_MODELS.DEFAULT;
+  sdkOptions.model = options.model || CLAUDE_PREDEFINED_MODELS.DEFAULT;
 
   const resolvedEffort = resolveClaudeEffort(
     sdkOptions.model,
     effort,
-    options.effortModels || CLAUDE_FALLBACK_MODELS,
+    options.effortModels || CLAUDE_PREDEFINED_MODELS,
   );
   if (resolvedEffort) {
     sdkOptions.effort = resolvedEffort;
@@ -240,13 +264,37 @@ function mapCliOptionsToSDK(options = {}) {
  * @param {string} sessionId - Session identifier
  * @param {Object} queryInstance - SDK query instance
  * @param {Object} writer - WebSocket writer for reconnect support
+ * @param {Function} releaseInput - Closes the held stdin stream so the CLI can exit
  */
-function addSession(sessionId, queryInstance, writer = null) {
+function addSession(sessionId, queryInstance, writer = null, releaseInput = null) {
+  const existing = activeSessions.get(sessionId);
+  // A different live instance under the same key means an earlier run was
+  // superseded without being stopped (e.g. an abort that raced run setup and
+  // found nothing to interrupt). Overwriting it here would strand its
+  // generator forever — this map entry is the only handle for interrupting
+  // it. Stop it directly rather than via abortClaudeSDKSession, whose
+  // session-keyed abortedSessionIds flag would be consumed by the new run
+  // and suppress its terminal `complete`.
+  const superseding = Boolean(
+    existing && existing.status === 'active' && existing.instance && existing.instance !== queryInstance
+  );
+  if (superseding) {
+    supersededInstances.add(existing.instance);
+    Promise.resolve()
+      .then(() => existing.instance.interrupt())
+      .catch((error) => {
+        console.error(`Error interrupting superseded run for session ${sessionId}:`, error?.message || error);
+      });
+    existing.releaseInput?.();
+  }
+  const carried = superseding ? null : existing;
   activeSessions.set(sessionId, {
     instance: queryInstance,
-    startTime: Date.now(),
+    startTime: carried?.startTime || Date.now(),
     status: 'active',
-    writer
+    writer,
+    // Re-registered mid-run once the provider session id lands; keep the closer.
+    releaseInput: releaseInput || carried?.releaseInput || null
   });
 }
 
@@ -363,38 +411,93 @@ function extractTokenBudget(sdkMessage) {
   };
 }
 
+// Tool calls that leave work running past the end of a turn. Bash only counts
+// when it is explicitly backgrounded; the rest defer or watch work by nature.
+const DEFERRED_WORK_TOOLS = new Set(['Monitor', 'ScheduleWakeup', 'CronCreate', 'TaskCreate']);
+
 /**
- * Builds the SDK `prompt` payload for one turn.
+ * Detects tool calls that keep working after the turn's `result` arrives.
  *
- * Plain text turns pass the string through unchanged. Turns with image
- * attachments use the SDK's streaming-input mode: a single SDKUserMessage
- * whose content carries the prompt text plus one base64 `image` block per
- * attachment (read from the global `~/.cloudcli/assets` folder).
+ * Only turns that start background work need their CLI process held open; every
+ * other turn can let it exit immediately, as it did before the hold existed.
+ *
+ * @param {Object} sdkMessage - SDK stream message
+ * @returns {boolean} True when the message launches work that outlives the turn
+ */
+function startsBackgroundWork(sdkMessage) {
+  const content = sdkMessage?.message?.content;
+  if (!Array.isArray(content)) {
+    return false;
+  }
+
+  return content.some((block) => {
+    if (block?.type !== 'tool_use') {
+      return false;
+    }
+    if (block.name === 'Bash') {
+      return block.input?.run_in_background === true;
+    }
+    return DEFERRED_WORK_TOOLS.has(block.name);
+  });
+}
+
+/**
+ * Builds the SDK user messages for one turn.
+ *
+ * Always returns SDKUserMessage records rather than a bare string: a string
+ * prompt makes the SDK flag the query as single-turn and close stdin the moment
+ * the turn's `result` arrives, which kills the CLI's background tasks. Plain
+ * text turns carry string content; turns with image attachments carry the
+ * prompt text plus one base64 `image` block per attachment (read from the
+ * global `~/.cloudcli/assets` folder).
  *
  * @param {string} command - User prompt
  * @param {Array} images - Image descriptors ({ path, name?, mimeType? })
  * @param {Array} files - Non-image attachment descriptors
  * @param {string} cwd - Project working directory attachment paths resolve against
- * @returns {Promise<string|AsyncIterable>} SDK prompt payload
+ * @returns {Promise<Array<Object>>} SDKUserMessage records for the turn
  */
-async function buildPromptPayload(command, images, files, cwd) {
+async function buildPromptMessages(command, images, files, cwd) {
   const promptWithFiles = appendFilesInputTag(command, files);
-  if (normalizeImageDescriptors(images).length === 0) {
-    return promptWithFiles;
-  }
+  const content = normalizeImageDescriptors(images).length === 0
+    ? promptWithFiles
+    : await buildClaudeUserContent(promptWithFiles, images, cwd);
+
+  return [{
+    type: 'user',
+    message: {
+      role: 'user',
+      content
+    },
+    parent_tool_use_id: null,
+    timestamp: new Date().toISOString()
+  }];
+}
 
-  const content = await buildClaudeUserContent(promptWithFiles, images, cwd);
-  return (async function* () {
-    yield {
-      type: 'user',
-      message: {
-        role: 'user',
-        content
-      },
-      parent_tool_use_id: null,
-      timestamp: new Date().toISOString()
-    };
+/**
+ * Wraps prompt messages in an async iterable that yields them and then parks.
+ *
+ * The SDK closes the CLI's stdin as soon as its input iterable is exhausted (and
+ * immediately on `result` for string prompts). The CLI reads that EOF as the end
+ * of the run and kills anything still going in the background, so the iterable
+ * has to stay pending until we actually want the process gone.
+ *
+ * @param {Array<Object>} messages - SDKUserMessage records to send
+ * @returns {{ stream: AsyncIterable, release: () => void }} Stream plus its closer
+ */
+function createHeldPromptStream(messages) {
+  let release;
+  const held = new Promise((resolve) => { release = resolve; });
+
+  const stream = (async function* () {
+    for (const message of messages) {
+      yield message;
+    }
+    // Keeps stdin open — the CLI stays alive until release() is called.
+    await held;
   })();
+
+  return { stream, release };
 }
 
 /**
@@ -483,9 +586,47 @@ async function queryClaudeSDK(command, options = {}, ws, context) {
     });
   };
 
+  // Closes the held stdin stream so the CLI can wind down. Replaced once the
+  // stream exists; the finally block calls it no matter how the run ends.
+  let releasePromptStream = () => {};
+  let idleReleaseTimer = null;
+  // The client is told the turn is over as soon as `result` lands, even though
+  // the process lingers, so the UI never waits out the idle hold.
+  let turnCompleteSent = false;
+  // Set when a turn starts background work, cleared when the next `result`
+  // arrives — only turns with work still outstanding hold their process open.
+  let backgroundWorkPending = false;
+  // True while the process is being held open for background work, so a later
+  // `result` can be recognised as that work reporting back.
+  let heldForBackgroundWork = false;
+
+  // A new turn supersedes any earlier one still holding this session's process
+  // open, so held runs cannot stack up across a conversation.
+  if (sessionKey()) {
+    getSession(sessionKey())?.releaseInput?.();
+  }
+
+  // Arms (or re-arms) the idle countdown that eventually closes stdin.
+  const scheduleRelease = () => {
+    if (idleReleaseTimer) {
+      clearTimeout(idleReleaseTimer);
+      idleReleaseTimer = null;
+    }
+    idleReleaseTimer = setTimeout(() => {
+      idleReleaseTimer = null;
+      releasePromptStream();
+    }, BG_WAIT_CEILING_MS);
+    // Never let the hold keep the server process alive on its own.
+    idleReleaseTimer.unref?.();
+  };
+
+  // Hoisted above the try so the catch's cleanup can tell whether this run
+  // still owns the activeSessions entry (or was superseded by a newer run).
+  let queryInstance = null;
+
   try {
     const resolvedModel = await context.resolveResumeModel(sessionId, options.model);
-    let effortModels = CLAUDE_FALLBACK_MODELS;
+    let effortModels = CLAUDE_PREDEFINED_MODELS;
     try {
       effortModels = await context.getProviderModels();
     } catch (error) {
@@ -504,10 +645,10 @@ async function queryClaudeSDK(command, options = {}, ws, context) {
       sdkOptions.mcpServers = mcpServers;
     }
 
-    // Turns with image attachments switch to streaming input so the images
-    // ride along as real content blocks. Built per query attempt because an
-    // async generator cannot be replayed once consumed.
-    const createPrompt = () => buildPromptPayload(command, options.images, options.files, options.cwd);
+    // Every turn uses streaming input so stdin stays open past the turn's
+    // `result`. The message list is reusable, but each query attempt needs its
+    // own stream because an async generator cannot be replayed once consumed.
+    const promptMessages = await buildPromptMessages(command, options.images, options.files, options.cwd);
 
     sdkOptions.hooks = {
       Notification: [{
@@ -610,14 +751,11 @@ async function queryClaudeSDK(command, options = {}, ws, context) {
       return { behavior: 'deny', message: decision.message ?? 'User denied tool use' };
     };
 
-    // Query constructor reads this synchronously.
-    const prevStreamTimeout = process.env.CLAUDE_CODE_STREAM_CLOSE_TIMEOUT;
-    process.env.CLAUDE_CODE_STREAM_CLOSE_TIMEOUT = '300000';
-
-    let queryInstance;
+    let heldPrompt = createHeldPromptStream(promptMessages);
+    releasePromptStream = heldPrompt.release;
     try {
       queryInstance = query({
-        prompt: await createPrompt(),
+        prompt: heldPrompt.stream,
         options: sdkOptions
       });
     } catch (hookError) {
@@ -625,22 +763,19 @@ async function queryClaudeSDK(command, options = {}, ws, context) {
       // Keep notification behavior operational via runtime events even if hook registration fails.
       console.warn('Failed to initialize Claude query with hooks, retrying without hooks:', hookError?.message || hookError);
       delete sdkOptions.hooks;
+      // Discard the abandoned stream and build a fresh one for the retry.
+      heldPrompt.release();
+      heldPrompt = createHeldPromptStream(promptMessages);
+      releasePromptStream = heldPrompt.release;
       queryInstance = query({
-        prompt: await createPrompt(),
+        prompt: heldPrompt.stream,
         options: sdkOptions
       });
     }
 
-    // Restore immediately — Query constructor already captured the value
-    if (prevStreamTimeout !== undefined) {
-      process.env.CLAUDE_CODE_STREAM_CLOSE_TIMEOUT = prevStreamTimeout;
-    } else {
-      delete process.env.CLAUDE_CODE_STREAM_CLOSE_TIMEOUT;
-    }
-
     // Track the query instance for abort capability
     if (sessionKey()) {
-      addSession(sessionKey(), queryInstance, ws);
+      addSession(sessionKey(), queryInstance, ws, releasePromptStream);
     }
 
     // Process streaming messages
@@ -650,7 +785,7 @@ async function queryClaudeSDK(command, options = {}, ws, context) {
       if (message.session_id && !capturedSessionId) {
 
         capturedSessionId = message.session_id;
-        addSession(sessionKey(), queryInstance, ws);
+        addSession(sessionKey(), queryInstance, ws, releasePromptStream);
 
         // Set session ID on writer
         if (ws.setSessionId && typeof ws.setSessionId === 'function') {
@@ -685,36 +820,98 @@ async function queryClaudeSDK(command, options = {}, ws, context) {
       if (tokenBudgetData) {
         ws.send(createNormalizedMessage({ kind: 'status', text: 'token_budget', tokenBudget: tokenBudgetData, sessionId: capturedSessionId || sessionId || null, provider: 'claude' }));
       }
+
+      if (startsBackgroundWork(message)) {
+        backgroundWorkPending = true;
+      }
+
+      if (message.type === 'result') {
+        // The turn is done as far as the client is concerned.
+        const abortPending = sessionKey() ? abortedSessionIds.has(sessionKey()) : false;
+        if (!turnCompleteSent && !abortPending) {
+          turnCompleteSent = true;
+          ws.send(createCompleteMessage({ provider: 'claude', sessionId: capturedSessionId || sessionId || null, exitCode: 0 }));
+          notifyRunStopped({
+            userId: ws?.userId || null,
+            provider: 'claude',
+            sessionId: sessionId || capturedSessionId || null,
+            sessionName: sessionSummary,
+            stopReason: 'completed'
+          });
+        } else if (heldForBackgroundWork && !abortPending) {
+          // A result after the turn already reported complete means the work we
+          // held the process open for has finished and pushed a follow-up turn.
+          notifyBackgroundWorkCompleted({
+            userId: ws?.userId || null,
+            provider: 'claude',
+            sessionId: sessionId || capturedSessionId || null,
+            sessionName: sessionSummary
+          });
+        }
+        if (backgroundWorkPending) {
+          // Work started during this turn is still running. Hold the process
+          // open so it can finish and report back in a follow-up turn; the
+          // ceiling is only a backstop for work that never reports.
+          backgroundWorkPending = false;
+          heldForBackgroundWork = true;
+          scheduleRelease();
+        } else {
+          // Either nothing was backgrounded, or the background work just
+          // reported in — let the CLI exit now, as it always has.
+          heldForBackgroundWork = false;
+          releasePromptStream();
+        }
+      } else if (idleReleaseTimer) {
+        // Background activity after the turn — push the countdown back out.
+        scheduleRelease();
+      }
     }
 
-    // Clean up session on completion
-    if (sessionKey()) {
+    // Clean up session on completion — only while this run still owns the map
+    // entry. A superseding run may have replaced it, and deleting here would
+    // strand that run.
+    if (sessionKey() && getSession(sessionKey())?.instance === queryInstance) {
       removeSession(sessionKey());
     }
 
+    // A superseded run winds down silently: the map entry, the abort flag,
+    // and all client-facing events belong to the run that replaced it.
+    const superseded = supersededInstances.has(queryInstance);
+
     // Send the terminal completion event — skipped for aborted runs, whose
-    // terminal `complete` (aborted: true) was already sent by abort-session.
-    const wasAborted = sessionKey() ? abortedSessionIds.delete(sessionKey()) : false;
-    if (!wasAborted) {
-      ws.send(createCompleteMessage({ provider: 'claude', sessionId: capturedSessionId || sessionId || null, exitCode: 0 }));
+    // terminal `complete` (aborted: true) was already sent by abort-session, and
+    // for runs that already reported completion when their `result` arrived.
+    const wasAborted = !superseded && sessionKey() ? abortedSessionIds.delete(sessionKey()) : false;
+    if (!turnCompleteSent && !superseded) {
+      turnCompleteSent = true;
+      if (!wasAborted) {
+        ws.send(createCompleteMessage({ provider: 'claude', sessionId: capturedSessionId || sessionId || null, exitCode: 0 }));
+      }
+      notifyRunStopped({
+        userId: ws?.userId || null,
+        provider: 'claude',
+        sessionId: sessionId || capturedSessionId || null,
+        sessionName: sessionSummary,
+        stopReason: wasAborted ? 'aborted' : 'completed'
+      });
     }
-    notifyRunStopped({
-      userId: ws?.userId || null,
-      provider: 'claude',
-      sessionId: sessionId || capturedSessionId || null,
-      sessionName: sessionSummary,
-      stopReason: wasAborted ? 'aborted' : 'completed'
-    });
     // Complete
 
   } catch (error) {
     console.error('SDK query error:', error);
 
-    // Clean up session on error
-    if (sessionKey()) {
+    // Clean up session on error — only while this run still owns the map entry
+    // (a superseding run may have replaced it).
+    if (sessionKey() && getSession(sessionKey())?.instance === queryInstance) {
       removeSession(sessionKey());
     }
 
+    if (supersededInstances.has(queryInstance)) {
+      // Interrupted because a newer run took over this session id; that run
+      // owns the abort flag and all further client-facing events.
+      return;
+    }
+
     const wasAborted = sessionKey() ? abortedSessionIds.delete(sessionKey()) : false;
     if (wasAborted) {
       // The abort already produced the terminal complete; a generator throw
@@ -728,9 +925,13 @@ async function queryClaudeSDK(command, options = {}, ws, context) {
       ? 'Claude Code is not installed. Please install it first: https://docs.anthropic.com/en/docs/claude-code'
       : error.message;
 
-    // Send error to WebSocket, then the terminal complete
+    // Send error to WebSocket, then the terminal complete. A run that already
+    // reported completion and then failed during its post-turn hold still
+    // surfaces the error, but must not emit a second terminal complete.
     ws.send(createNormalizedMessage({ kind: 'error', content: errorContent, sessionId: capturedSessionId || sessionId || null, provider: 'claude' }));
-    ws.send(createCompleteMessage({ provider: 'claude', sessionId: capturedSessionId || sessionId || null, exitCode: 1 }));
+    if (!turnCompleteSent) {
+      ws.send(createCompleteMessage({ provider: 'claude', sessionId: capturedSessionId || sessionId || null, exitCode: 1 }));
+    }
     notifyRunFailed({
       userId: ws?.userId || null,
       provider: 'claude',
@@ -738,6 +939,14 @@ async function queryClaudeSDK(command, options = {}, ws, context) {
       sessionName: sessionSummary,
       error
     });
+  } finally {
+    // Always close stdin — otherwise an aborted or failed run leaves the CLI
+    // process (and its MCP servers) alive until the server exits.
+    if (idleReleaseTimer) {
+      clearTimeout(idleReleaseTimer);
+      idleReleaseTimer = null;
+    }
+    releasePromptStream();
   }
 }
 
@@ -764,6 +973,10 @@ async function abortClaudeSDKSession(sessionId) {
     // Call interrupt() on the query instance
     await session.instance.interrupt();
 
+    // Release the held stdin stream; without this the CLI stays up for the rest
+    // of the post-turn hold even though the user cancelled.
+    session.releaseInput?.();
+
     // Update session status
     session.status = 'aborted';