@anthropic-ai/claude-agent-sdk 0.3.268 → 0.3.269

package/sdk.d.ts

@@ -2681,26 +2681,26 @@ export declare interface Query extends AsyncGenerator<SDKMessage, void> {
      */
     setMaxThinkingTokens(maxThinkingTokens: number | null, thinkingDisplay?: 'summarized' | 'omitted' | null): Promise<void>;
     /**
-     * Merge the provided settings into the flag settings layer, dynamically
-     * updating the active configuration. Equivalent to the inline `settings`
-     * option of `query()`, but applies mid-session. Flag settings sit above
-     * user/project/local settings and below managed policy settings in the
-     * precedence order.
+     * Merge settings into the flag settings layer. This is the inline `settings`
+     * option of `query()`, applied mid-session. Flag settings sit above
+     * user/project/local and below managed policy settings in precedence order.
      *
      * Successive calls shallow-merge top-level keys — a second call with
      * `{permissions: {...}}` replaces the entire `permissions` object from a
      * prior call. Pass `null` for a key to clear it from the flag layer and
      * fall back to lower-precedence sources (`undefined` is dropped by JSON
-     * serialization and has no effect).
-     *
+     * serialization and has no effect). Four keys instead reset session state
+     * and restore neither a `query()` option nor a settings-file value.
+     * `effortLevel` goes to the model's default effort, `model` to Claude Code's
+     * default model (not `ANTHROPIC_MODEL` or `settings.model`), `agent` to no
+     * main-thread agent, and `ultracode` to off with the current effort kept.
      * Only available in streaming input mode.
      *
-     * @param settings - A partial settings object to merge into the flag settings.
-     * Each top-level key also accepts `null` to clear it from the flag layer.
-     * `effortLevel` additionally accepts `'max'`, which is session-scoped: it
-     * applies for the rest of the session on models that support it and is
-     * never persisted to settings files (the persisted
-     * {@link Settings.effortLevel} excludes it for that reason).
+     * @param settings - A partial settings object to merge into the flag
+     * settings. `effortLevel` also accepts `'max'` (never written to settings
+     * files, so the persisted {@link Settings.effortLevel} excludes it): it is
+     * session-only, runs as `'high'` on a model without `'max'` support, and
+     * runs no higher than the organization's effort limit for the model.
      */
     applyFlagSettings(settings: {
         [K in keyof Settings]?: K extends 'effortLevel' ? EffortLevel | null : Settings[K] | null;
@@ -2806,6 +2806,7 @@ export declare interface Query extends AsyncGenerator<SDKMessage, void> {
     usage_EXPERIMENTAL_MAY_CHANGE_DO_NOT_RELY_ON_THIS_API_YET(opts?: {
         skipBehaviors?: boolean;
     }): Promise<SDKControlGetUsageResponse>;
+
     /**
      * Read a file from the session's filesystem for the remote sidebar
      * viewer. Path is resolved against cwd and gated by the same
@@ -3348,7 +3349,7 @@ export declare type SDKAssistantMessage = {
     session_id: string;
     request_id?: string;
     /**
-     * Client uuid of the user message this turn is answering (submitMessage options.uuid), stamped on a reply frame each time that send changes — the turn's FIRST reply frame, and then, for a turn started by a synthetic (meta) prompt, the first reply frame after each queued user message folded in mid-turn takes the echo over. In complete-message mode the frame is the first assistant message; with --include-partial-messages the stamp normally rides the first non-ping stream event instead (see SDKPartialAssistantMessage), and a turn that produces no stream events still stamps its first assistant message — so a consumer can bind the reply to the send it answers without waiting for the result; the server keeps the first stamp it sees per uuid. A turn started by a typed prompt keeps that uuid for its whole turn, so it stamps its first reply frame only. A meta turn's own uuid is stamped only when the host vouches it is the client event's own (on a hosted session, the uuid the session server persisted: delivered content such as a Slack owner ping, a Slack-bot observation or a client-injected synthetic turn), never for a prompt the CLI minted itself — except that the boot-time rescue turn re-running a turn a worker restart interrupted mid-way stamps the interrupted turn's own last user prompt (with resume_reason), the send that re-run answers; either way a user message folded into a meta turn takes the echo over from it (the rescue turn absorbing messages sent while the session was down; a bot-observation turn absorbing a human's post), and the first reply frame after that fold carries the folded message's uuid — the first reply that message got. Wrapper-level sibling — never inside `message.content` — so it is not replayed to the model. Absent on every other frame of the turn, on subagent frames (parent_tool_use_id set), on turns that neither had a client uuid nor folded a user message in, and from older producers.
+     * Client uuid of the user message this turn is answering (submitMessage options.uuid), stamped on an assistant message each time that send changes — the turn's FIRST top-level assistant message (which may carry only a thinking block, or be a synthetic API-error message), and then, for a turn started by a synthetic (meta) prompt, the first assistant message after each queued user message folded in mid-turn (the fold takes the echo over); with --include-partial-messages the turn's first non-ping stream event is stamped too, independently (see SDKPartialAssistantMessage), so the same uuid may appear on both — either binds the reply to the send it answers without waiting for the result; the server keeps the first stamp it sees per uuid. A turn started by a typed prompt keeps that uuid for its whole turn, so it stamps once per frame kind. A meta turn's own uuid is stamped only when the host vouches it is the client event's own (on a hosted session, the uuid the session server persisted: delivered content such as a Slack owner ping, a Slack-bot observation or a client-injected synthetic turn), never for a prompt the CLI minted itself — except that the boot-time rescue turn re-running a turn a worker restart interrupted mid-way stamps the interrupted turn's own last user prompt (with resume_reason), the send that re-run answers; either way a user message folded into a meta turn takes the echo over from it (the rescue turn absorbing messages sent while the session was down; a bot-observation turn absorbing a human's post), and the first reply frame of each kind after that fold carries the folded message's uuid — the first reply that message got. Wrapper-level sibling — never inside `message.content` — so it is not replayed to the model. Absent on every other frame of the turn, on subagent frames (parent_tool_use_id set), on turns that neither had a client uuid nor folded a user message in, and from older producers.
      */
     user_message_uuid?: string;
     /**
@@ -3758,6 +3759,151 @@ export declare type SDKControlGetContextUsageResponse = {
     } | null;
 };
 
+/**
+ * Returns the hooks listing the CLI's read-only /hooks menu renders: settings-file, session, and plugin hooks grouped by event and matcher, with display-ready strings (control characters revealed) and the policy and safe-mode state the menu banners on. A snapshot at request time; hosts re-request when their surface opens.
+ */
+declare type SDKControlGetHooksListingRequest = {
+    subtype: 'get_hooks_listing';
+};
+
+/**
+ * The hooks listing the CLI's /hooks menu renders, with display-ready strings.
+ */
+declare type SDKControlGetHooksListingResponse = {
+    /**
+     * Events that have at least one listed hook, in the /hooks menu’s lifecycle order.
+     */
+    events: {
+        /**
+         * Hook event name.
+         */
+        name: string;
+        /**
+         * One-line summary, as the /hooks event list shows.
+         */
+        summary: string;
+        /**
+         * Whether hooks on this event can carry a matcher.
+         */
+        supportsMatcher: boolean;
+        /**
+         * Number of listed hooks.
+         */
+        hookCount: number;
+    }[];
+    /**
+     * One row per listed hook: events in lifecycle order, matchers in the menu’s priority order.
+     */
+    hooks: {
+        event: string;
+        /**
+         * Matcher with control characters revealed; '' when the entry has none.
+         */
+        matcher: string;
+        /**
+         * Raw source name (userSettings, sessionHook, pluginHook, …).
+         */
+        source: string;
+        /**
+         * User-facing source description.
+         */
+        sourceLabel: string;
+        pluginName?: string;
+        /**
+         * Hook type (command, prompt, agent, http, mcp_tool, …).
+         */
+        type: string;
+        /**
+         * List-row label: statusMessage when set, else the identity text; one line, control characters revealed.
+         */
+        displayText: string;
+        /**
+         * Identity text — the literal command/prompt/URL that runs, control characters revealed.
+         */
+        commandText: string;
+        /**
+         * Label for commandText (Command, Prompt, URL, …).
+         */
+        contentLabel: string;
+        /**
+         * If-condition, revealed, when set.
+         */
+        condition?: string;
+        /**
+         * Timeout in seconds.
+         */
+        timeout?: number;
+        statusMessage?: string;
+        runsOnce?: boolean;
+        runsInBackground?: boolean;
+        /**
+         * True when the session's mode or policy keeps this hook from running (the policy block and safeMode/bareMode say why); absent on rows that run.
+         */
+        disabled?: true;
+        /**
+         * The entry as stored (raw matcher, '' when none, and raw hook object; no display escaping) for a host's edit form and as the target it names to `claude edit-hook`. Only on rows from a settings file this session reads and may write. HTTP header values are blanked (headersRedacted); a replace that sends no headers keeps the stored ones.
+         */
+        editable?: {
+            matcher: string;
+            config: Record<string, unknown>;
+            headersRedacted?: true;
+        };
+    }[];
+    /**
+     * Every hook event in lifecycle order (name, the /hooks summary, whether its hooks take a matcher), for an add-hook form.
+     */
+    eventCatalog: {
+        name: string;
+        summary: string;
+        supportsMatcher: boolean;
+    }[];
+    policy: {
+        /**
+         * Managed disableAllHooks — nothing runs at all.
+         */
+        disabledByPolicy: boolean;
+        /**
+         * Managed allowManagedHooksOnly — non-managed hooks are blocked and managed hooks are intentionally not listed.
+         */
+        managedOnly: boolean;
+        /**
+         * Managed strictPluginOnlyCustomization locks the hooks surface.
+         */
+        pluginOnly: boolean;
+        /**
+         * Effective disableAllHooks, whatever source set it.
+         */
+        allDisabled: boolean;
+        /**
+         * Hooks configured in managed settings (they run even under a non-managed disableAllHooks).
+         */
+        policyHookCount: number;
+    };
+    /**
+     * Present only when the session runs under --safe-mode.
+     */
+    safeMode?: {
+        managedHooksStillApply: boolean;
+        /**
+         * How to leave safe mode, per its activation source.
+         */
+        exitHint: string;
+    };
+    /**
+     * Present only under --bare / CLAUDE_CODE_SIMPLE with the hooks surface gated off: settings-file, flag, policy, and plugin hooks never fire there; session hooks still run.
+     */
+    bareMode?: {
+        /**
+         * How to leave bare mode, per its activation source.
+         */
+        exitHint: string;
+    };
+    /**
+     * Settings files skipped by the merge — their hooks are neither listed nor running.
+     */
+    errors?: coreTypes.SDKSettingsParseError[];
+};
+
 /**
  * Requests the formatted session cost summary (the same text /usage prints in non-interactive mode). Used by the thin-client /usage dialog to show the remote container cost instead of the local $0.00.
  */
@@ -3861,7 +4007,7 @@ export declare type SDKControlGetUsageResponse = {
             resets_at: string | null;
         } | null;
         /**
-         * Per-model weekly windows from the server limits[] array, filtered by the overage-included-models allowlist. Additive — present only when the server emits them.
+         * Per-model weekly windows from the server limits[] array, filtered by the overage-included-models allowlist. Additive: absent when nothing is known about them (an answer served from cached data, or rows the allowlist hides); an empty array means the endpoint itself answered and listed no per-model weekly window at all for this account, before the allowlist was applied.
          */
         model_scoped?: {
             /**
@@ -4114,6 +4260,7 @@ export declare type SDKControlInitializeResponse = {
 
 
 
+
 };
 
 /**
@@ -4149,6 +4296,23 @@ declare type SDKControlListModelsRequest = {
     subtype: 'list_models';
 };
 
+/**
+ * Requests the session's live permission rules and workspace directories — the same data /permissions lists in the terminal: rules from settings files plus session-only approvals, slash-command grants, and --allowedTools flag rules, each with its source.
+ */
+declare type SDKControlListPermissionRulesRequest = {
+    subtype: 'list_permission_rules';
+};
+
+/**
+ * Success payload of list_permission_rules.
+ */
+export declare type SDKControlListPermissionRulesResponse = {
+    /**
+     * The session's live permission rules state, as list_permission_rules reports it.
+     */
+    state: SDKControlPermissionRulesState;
+};
+
 /**
  * Invokes an MCP tool via the subprocess MCP client without a model turn. No permission check (control channel is trusted, same as other subtypes). SDK-type MCP servers (config.type === "sdk") are rejected — they are caller-provided, so the caller can invoke them directly without the subprocess round-trip. Result content passes through the same processing as model-turn MCP calls. Session expiry is not retried automatically; callers can mcp_reconnect and retry. UrlElicitationRequired (-32042) tries Elicitation hooks; if no hook resolves, the call errors with the URL in the message — open it out-of-band, then retry mcp_call. STAGED calls (input_files/output_files declared) additionally stage lane rows in/out around the call — see the input_files describe. Staged failures come back as a success-subtype response whose staging field carries a typed error_code; subtype:error is emitted only when the call could not be attempted at all (server not connected, kill switch, dispatch failure) and means nothing ran. A target server that is not yet connected is brought up on demand: dispatch runs the deferred plugin/MCP startup resolution (the work a first model turn would have done) and waits up to 30s — shortened by expires_at when that is sooner — for the server to connect before answering "MCP server not connected", so a dispatch that races plugin startup (e.g. after an idle-wake reattach) succeeds instead of failing until a turn runs. Standard RPC semantics: a redelivered request_id supersedes the in-flight run (it is aborted and its response suppressed — exactly one response per request_id); conversion is idempotent, so re-running is safe. Cancellable via control_cancel_request.
  */
@@ -4295,6 +4459,26 @@ declare type SDKControlPermissionRequest = {
 
 };
 
+/**
+ * The session's live permission rules state, as list_permission_rules reports it.
+ */
+export declare type SDKControlPermissionRulesState = {
+    rules: SDKPermissionRuleEntry[];
+    workspaceDirectories: SDKPermissionWorkspaceDirectory[];
+    /**
+     * The session's original working directory.
+     */
+    originalCwd: string;
+    /**
+     * True when enterprise managed settings pin allowManagedPermissionRulesOnly: the session applies policy rules only, and rules from other settings files appear with notInEffect set.
+     */
+    managedOnly: boolean;
+    /**
+     * Settings parse and validation errors, as get_settings reports them. When non-empty, the listed files were skipped — their rules are not in the session and not listed above.
+     */
+    errors?: coreTypes.SDKSettingsParseError[];
+};
+
 /**
  * Read a file from the session filesystem for the remote sidebar viewer. Path is resolved against cwd and gated by the same read-permission rules as the Read tool.
  */
@@ -4408,6 +4592,14 @@ export declare type SDKControlReloadSkillsResponse = {
 declare type SDKControlRenameSessionRequest = {
     subtype: 'rename_session';
     title: string;
+    /**
+     * Who chose the title: 'remote' (the default) for a rename made on claude.ai and relayed to this process, 'host' for one the user made in the hosting application (an IDE), which the CLI counts as a user rename.
+     */
+    source?: 'remote' | 'host';
+    /**
+     * The session the title is for. When given and this process has since moved to another session (/clear, an in-session resume), the request is refused instead of naming the new session.
+     */
+    session_id?: string;
 };
 
 /**
@@ -4424,7 +4616,7 @@ export declare type SDKControlRequest = {
 
 };
 
-declare type SDKControlRequestInner = SDKControlInterruptRequest | SDKControlPermissionRequest | SDKControlInitializeRequest | SDKControlSetPermissionModeRequest | SDKControlSetModelRequest | SDKControlSetMaxThinkingTokensRequest | SDKControlRenameSessionRequest | SDKControlSetColorRequest | SDKControlMcpStatusRequest | SDKControlGetContextUsageRequest | SDKControlGetSessionCostRequest | SDKControlListModelsRequest | SDKControlGetUsageRequest | SDKControlGetBinaryVersionRequest | SDKControlMcpCallRequest | SDKControlFileSuggestionsRequest | SDKHookCallbackRequest | SDKControlMcpMessageRequest | SDKControlRewindFilesRequest | SDKControlCancelAsyncMessageRequest | SDKControlReadFileRequest | SDKControlSeedReadStateRequest | SDKControlMcpSetServersRequest | SDKControlRegisterRepoRootRequest | SDKControlReloadPluginsRequest | SDKControlReloadSkillsRequest | SDKControlReloadOutputStylesRequest | SDKControlMcpReconnectRequest | SDKControlMcpToggleRequest | SDKControlStopTaskRequest | SDKControlBackgroundTasksRequest | SDKControlApplyFlagSettingsRequest | SDKControlGetSettingsRequest | SDKControlUpdateSettingsRequest | SDKControlElicitationRequest | SDKControlRequestUserDialogRequest;
+declare type SDKControlRequestInner = SDKControlInterruptRequest | SDKControlPermissionRequest | SDKControlInitializeRequest | SDKControlSetPermissionModeRequest | SDKControlSetModelRequest | SDKControlSetMaxThinkingTokensRequest | SDKControlRenameSessionRequest | SDKControlSetColorRequest | SDKControlMcpStatusRequest | SDKControlGetContextUsageRequest | SDKControlGetSessionCostRequest | SDKControlListModelsRequest | SDKControlGetUsageRequest | SDKControlGetBinaryVersionRequest | SDKControlMcpCallRequest | SDKControlFileSuggestionsRequest | SDKHookCallbackRequest | SDKControlMcpMessageRequest | SDKControlRewindFilesRequest | SDKControlCancelAsyncMessageRequest | SDKControlReadFileRequest | SDKControlSeedReadStateRequest | SDKControlMcpSetServersRequest | SDKControlRegisterRepoRootRequest | SDKControlReloadPluginsRequest | SDKControlReloadSkillsRequest | SDKControlReloadOutputStylesRequest | SDKControlMcpReconnectRequest | SDKControlMcpToggleRequest | SDKControlStopTaskRequest | SDKControlBackgroundTasksRequest | SDKControlApplyFlagSettingsRequest | SDKControlGetSettingsRequest | SDKControlGetHooksListingRequest | SDKControlUpdateSettingsRequest | SDKControlElicitationRequest | SDKControlRequestUserDialogRequest | SDKControlListPermissionRulesRequest;
 
 /**
  * Progress for a long-running client-originated control_request (currently only side_question), correlated by request_id. status 'started' means the worker accepted the request and launched the work; 'api_retry' carries the same retry counters as SDKAPIRetryMessage and is present only for that status.
@@ -4906,7 +5098,7 @@ export declare type SDKPartialAssistantMessage = {
     session_id: string;
     ttft_ms?: number;
     /**
-     * Client uuid of the user message this turn is answering (submitMessage options.uuid), stamped on a non-ping stream event each time that send changes: the turn's FIRST non-ping stream event (normally the frame that triggers the turn's initial ack), and, for a turn started by a synthetic (meta) prompt, the first non-ping stream event after each queued user message folded in mid-turn takes the echo over (see SDKAssistantMessage.user_message_uuid for the rule) — so a consumer can bind the reply stream to the send it answers without waiting for the result. A turn started by a typed prompt stamps its first non-ping stream event only. Absent on every other stream event of the turn, on turns that neither had a client uuid nor folded a user message in, and from older producers.
+     * Client uuid of the user message this turn is answering (submitMessage options.uuid), stamped on a non-ping stream event each time that send changes: the turn's FIRST non-ping stream event (normally the frame that triggers the turn's initial ack), and, for a turn started by a synthetic (meta) prompt, the first non-ping stream event after each queued user message folded in mid-turn takes the echo over (see SDKAssistantMessage.user_message_uuid for the rule) — so a consumer can bind the reply stream to the send it answers without waiting for the result. A turn started by a typed prompt stamps only its first non-ping stream event; independently, its first complete assistant message is stamped as well (see SDKAssistantMessage.user_message_uuid), so the same uuid may appear on both. Absent on every other stream event of the turn, on turns that neither had a client uuid nor folded a user message in, and from older producers.
      */
     user_message_uuid?: string;
     /**
@@ -4926,7 +5118,7 @@ export declare type SDKPermissionDenial = {
 };
 
 /**
- * Emitted when a tool call is auto-denied without an interactive permission prompt (e.g. auto-mode classifier, dontAsk mode, headless-agent auto-deny, or a deny rule). With a permission prompt surface (stdio/SDK canUseTool), the 'ask' path surfaces via a can_use_tool control_request and this event covers the 'deny' short-circuit. Without one (bare -p / SDK query() with no canUseTool), 'ask' decisions are terminal, so this event also covers those implicit denials. Best-effort advisory: in rare races a denial can book without a frame or a frame can lack a booking twin — result.permission_denials is the authoritative record. Denials that resolve before canUseTool runs — PreToolUse hook denies, and deny-rule overrides of hook allow/ask decisions — are not covered here, and neither is the MCP --permission-prompt-tool surface (the prompt tool is the host there).
+ * Emitted when a tool call is auto-denied without an interactive permission prompt (e.g. auto-mode classifier, dontAsk mode, headless-agent auto-deny, or a deny rule). With a permission prompt surface (stdio/SDK canUseTool), the 'ask' path surfaces via a can_use_tool control_request and this event covers the 'deny' short-circuit. Without one (bare -p / SDK query() with no canUseTool), 'ask' decisions are terminal, so this event also covers those implicit denials. Best-effort advisory: in rare races a denial can book without a frame or a frame can lack a booking twin — result.permission_denials is the authoritative record. Denials that resolve before canUseTool runs — PreToolUse hook denies, deny-rule overrides of hook allow/ask decisions, and file-tool calls (Read, Edit, Write) refused by a path-scoped deny rule — are not covered here, and neither is the MCP --permission-prompt-tool surface (the prompt tool is the host there).
  */
 export declare type SDKPermissionDeniedMessage = {
     type: 'system';
@@ -4953,6 +5145,53 @@ export declare type SDKPermissionDeniedMessage = {
     session_id: string;
 };
 
+/**
+ * The CLI's plain-language reading of a rule (e.g. "Any Bash command starting with npm run"), split into parts so hosts can render the rule-derived fragment the way the terminal does (bold). prefix and suffix are fixed words; emphasis is rule content — apply display hygiene (invisible-character escaping) before rendering it.
+ */
+export declare type SDKPermissionRuleDescription = {
+    prefix: string;
+    emphasis?: string;
+    suffix?: string;
+};
+
+/**
+ * One permission rule with its provenance and where it lives.
+ */
+export declare type SDKPermissionRuleEntry = {
+    behavior: 'allow' | 'deny' | 'ask';
+    /**
+     * Where the rule comes from. Mirrors PermissionRuleSource (permissionRuleLookup.ts PERMISSION_RULE_SOURCES); the parity test in test/cli/headlessControl/listPermissionRules.test.ts keeps the two aligned.
+     */
+    source: 'userSettings' | 'projectSettings' | 'localSettings' | 'flagSettings' | 'policySettings' | 'cliArg' | 'command' | 'session' | 'toolsNarrowing' | 'mcpServerPolicy' | 'hostCredential';
+    /**
+     * The stored rule string VERBATIM, exactly as the session holds it. Two stored spellings that parse identically each get their own entry. Can carry invisible or control characters by design — escape at display.
+     */
+    rule: string;
+    /**
+     * Plain-language reading of the rule; absent where the terminal shows no subtitle either.
+     */
+    description?: SDKPermissionRuleDescription;
+    /**
+     * Where the rule lives: 'persistent' (userSettings/projectSettings/localSettings — saved in a settings file), 'session' (cliArg/session — in memory only, for the rest of this session), or 'readonly' (policySettings/flagSettings/command and every other source — the set the terminal's /permissions treats as read-only). Informational for hosts; this request never changes rules.
+     */
+    editability: 'persistent' | 'session' | 'readonly';
+    /**
+     * Present (true) when enterprise managed settings pin allowManagedPermissionRulesOnly and this rule, read from a non-policy settings file, is ignored by the session. Such rows are readonly.
+     */
+    notInEffect?: boolean;
+};
+
+/**
+ * One additional working directory in the permission scope.
+ */
+export declare type SDKPermissionWorkspaceDirectory = {
+    path: string;
+    /**
+     * Where the directory grant came from: a settings source (e.g. 'localSettings'), 'cliArg' (--add-dir), or 'session' (/add-dir, IDE workspace folders).
+     */
+    source: string;
+};
+
 /**
  * Configuration for loading a plugin.
  */
@@ -6437,6 +6676,10 @@ export declare interface Settings {
      * Default shell for input-box ! commands. Defaults to 'bash' on all platforms (no Windows auto-flip).
      */
     defaultShell?: 'bash' | 'powershell';
+    /**
+     * Whether the Bash tool shows a diff of the files a Bash command changed (PostToolUse Bash hooks get the changed-file list in tool_response). Set to false to turn that off. Default: on when the Bash tool handles file edits. Only user, flag or policy settings can turn it on outside auto and bypassPermissions modes.
+     */
+    bashEditDiffEnabled?: boolean;
     /**
      * How many characters of a successful Bash or PowerShell command's output Claude receives inline (default 30000; values clamp to 4000-128000). Output past this is saved to a file and Claude receives a short preview plus the path. When set, this also replaces BASH_MAX_OUTPUT_LENGTH, which on its own only sizes the read-back window.
      */