import { feature } from 'bun:bundle' import type { ContentBlockParam, TextBlockParam, } from '@anthropic-ai/sdk/resources' import { randomUUID } from 'crypto' import { setPromptId } from 'src/bootstrap/state.js' import { builtInCommandNames, type Command, type CommandBase, findCommand, getCommand, getCommandName, hasCommand, type PromptCommand, } from 'src/commands.js' import { NO_CONTENT_MESSAGE } from 'src/constants/messages.js' import type { SetToolJSXFn, ToolUseContext } from 'src/Tool.js' import type { AssistantMessage, AttachmentMessage, Message, NormalizedUserMessage, ProgressMessage, UserMessage, } from 'src/types/message.js' import { addInvokedSkill, getSessionId } from '../../bootstrap/state.js' import { COMMAND_MESSAGE_TAG, COMMAND_NAME_TAG } from '../../constants/xml.js' import type { CanUseToolFn } from '../../hooks/useCanUseTool.js' import { type AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, type AnalyticsMetadata_I_VERIFIED_THIS_IS_PII_TAGGED, logEvent, } from '../../services/analytics/index.js' import { getDumpPromptsPath } from '../../services/api/dumpPrompts.js' import { buildPostCompactMessages } from '../../services/compact/compact.js' import { resetMicrocompactState } from '../../services/compact/microCompact.js' import type { Progress as AgentProgress } from '@claude-code-best/builtin-tools/tools/AgentTool/AgentTool.js' import { runAgent } from '@claude-code-best/builtin-tools/tools/AgentTool/runAgent.js' import { renderToolUseProgressMessage } from '@claude-code-best/builtin-tools/tools/AgentTool/UI.js' import type { CommandResultDisplay } from '../../types/command.js' import { createAbortController } from '../abortController.js' import { getAgentContext } from '../agentContext.js' import { createAttachmentMessage, getAttachmentMessages, } from '../attachments.js' import { logForDebugging } from '../debug.js' import { isEnvTruthy } from '../envUtils.js' import { AbortError, MalformedCommandError } from '../errors.js' import { getDisplayPath } from '../file.js' import { extractResultText, prepareForkedCommandContext, } from '../forkedAgent.js' import { getFsImplementation } from '../fsOperations.js' import { isFullscreenEnvEnabled } from '../fullscreen.js' import { toArray } from '../generators.js' import { registerSkillHooks } from '../hooks/registerSkillHooks.js' import { logError } from '../log.js' import { enqueuePendingNotification } from '../messageQueueManager.js' import { createCommandInputMessage, createSyntheticUserCaveatMessage, createSystemMessage, createUserInterruptionMessage, createUserMessage, formatCommandInputTags, isCompactBoundaryMessage, isSystemLocalCommandMessage, normalizeMessages, prepareUserContent, } from '../messages.js' import type { ModelAlias } from '../model/aliases.js' import { parseToolListFromCLI } from '../permissions/permissionSetup.js' import { hasPermissionsToUseTool } from '../permissions/permissions.js' import { isOfficialMarketplaceName, parsePluginIdentifier, } from '../plugins/pluginIdentifier.js' import { isRestrictedToPluginOnly, isSourceAdminTrusted, } from '../settings/pluginOnlyPolicy.js' import { parseSlashCommand } from '../slashCommandParsing.js' import { sleep } from '../sleep.js' import { recordSkillUsage } from '../suggestions/skillUsageTracking.js' import { logOTelEvent, redactIfDisabled } from '../telemetry/events.js' import { buildPluginCommandTelemetryFields } from '../telemetry/pluginTelemetry.js' import { getAssistantMessageContentLength } from '../tokens.js' import { createAgentId } from '../uuid.js' import { getWorkload } from '../workloadContext.js' import type { ProcessUserInputBaseResult, ProcessUserInputContext, } from './processUserInput.js' type SlashCommandResult = ProcessUserInputBaseResult & { command: Command } // Poll interval and deadline for MCP settle before launching a background // forked subagent. MCP servers typically connect within 1-3s of startup; // 10s headroom covers slow SSE handshakes. const MCP_SETTLE_POLL_MS = 200 const MCP_SETTLE_TIMEOUT_MS = 10_000 /** * Executes a slash command with context: fork in a sub-agent. */ async function executeForkedSlashCommand( command: CommandBase & PromptCommand, args: string, context: ProcessUserInputContext, precedingInputBlocks: ContentBlockParam[], setToolJSX: SetToolJSXFn, canUseTool: CanUseToolFn, ): Promise { const agentId = createAgentId() const pluginMarketplace = command.pluginInfo ? parsePluginIdentifier(command.pluginInfo.repository).marketplace : undefined logEvent('tengu_slash_command_forked', { command_name: command.name as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, invocation_trigger: 'user-slash' as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, ...(command.pluginInfo && { _PROTO_plugin_name: command.pluginInfo.pluginManifest .name as AnalyticsMetadata_I_VERIFIED_THIS_IS_PII_TAGGED, ...(pluginMarketplace && { _PROTO_marketplace_name: pluginMarketplace as AnalyticsMetadata_I_VERIFIED_THIS_IS_PII_TAGGED, }), ...buildPluginCommandTelemetryFields(command.pluginInfo), }), }) const { skillContent, modifiedGetAppState, baseAgent, promptMessages } = await prepareForkedCommandContext(command, args, context) // Merge skill's effort into the agent definition so runAgent applies it const agentDefinition = command.effort !== undefined ? { ...baseAgent, effort: command.effort } : baseAgent logForDebugging( `Executing forked slash command /${command.name} with agent ${agentDefinition.agentType}`, ) // Assistant mode: fire-and-forget. Launch subagent in background, return // immediately, re-enqueue the result as an isMeta prompt when done. // Without this, N scheduled tasks on startup = N serial (subagent + main // agent turn) cycles blocking user input. With this, N subagents run in // parallel and results trickle into the queue as they finish. // // Gated on kairosEnabled (not CLAUDE_CODE_BRIEF) because the closed loop // depends on assistant-mode invariants: scheduled_tasks.json exists, // the main agent knows to pipe results through SendUserMessage, and // isMeta prompts are hidden. Outside assistant mode, context:fork commands // are user-invoked skills (/commit etc.) that should run synchronously // with the progress UI. if (feature('KAIROS') && (await context.getAppState()).kairosEnabled) { // Standalone abortController — background subagents survive main-thread // ESC (same policy as AgentTool's async path). They're cron-driven; if // killed mid-run they just re-fire on the next schedule. const bgAbortController = createAbortController() const commandName = getCommandName(command) // Workload: handlePromptSubmit wraps the entire turn in runWithWorkload // (AsyncLocalStorage). ALS context is captured when this `void` fires // and survives every await inside — isolated from the parent's // continuation. The detached closure's runAgent calls see the cron tag // automatically. We still capture the value here ONLY for the // re-enqueued result prompt below: that second turn runs in a fresh // handlePromptSubmit → fresh runWithWorkload boundary (which always // establishes a new context, even for `undefined`) → so it needs its // own QueuedCommand.workload tag to preserve attribution. const spawnTimeWorkload = getWorkload() // Re-enter the queue as a hidden prompt. isMeta: hides from queue // preview + placeholder + transcript. skipSlashCommands: prevents // re-parsing if the result text happens to start with '/'. When // drained, this triggers a main-agent turn that sees the result and // decides whether to SendUserMessage. Propagate workload so that // second turn is also tagged. const enqueueResult = (value: string): void => enqueuePendingNotification({ value, mode: 'prompt', priority: 'later', isMeta: true, skipSlashCommands: true, workload: spawnTimeWorkload, }) void (async () => { // Wait for MCP servers to settle. Scheduled tasks fire at startup and // all N drain within ~1ms (since we return immediately), capturing // context.options.tools before MCP connects. The sync path // accidentally avoided this — tasks serialized, so task N's drain // happened after task N-1's 30s run, by which time MCP was up. // Poll until no 'pending' clients remain, then refresh. const deadline = Date.now() + MCP_SETTLE_TIMEOUT_MS while (Date.now() < deadline) { const s = context.getAppState() if (!s.mcp.clients.some(c => c.type === 'pending')) break await sleep(MCP_SETTLE_POLL_MS) } const freshTools = context.options.refreshTools?.() ?? context.options.tools const agentMessages: Message[] = [] for await (const message of runAgent({ agentDefinition, promptMessages, toolUseContext: { ...context, getAppState: modifiedGetAppState, abortController: bgAbortController, }, canUseTool, isAsync: true, querySource: 'agent:custom', model: command.model as ModelAlias | undefined, availableTools: freshTools, override: { agentId }, })) { agentMessages.push(message) } const resultText = extractResultText(agentMessages, 'Command completed') logForDebugging( `Background forked command /${commandName} completed (agent ${agentId})`, ) enqueueResult( `\n${resultText}\n`, ) })().catch(err => { logError(err) enqueueResult( `\n${err instanceof Error ? err.message : String(err)}\n`, ) }) // Nothing to render, nothing to query — the background runner re-enters // the queue on its own schedule. return { messages: [], shouldQuery: false, command } } // Collect messages from the forked agent const agentMessages: Message[] = [] // Build progress messages for the agent progress UI const progressMessages: ProgressMessage[] = [] const parentToolUseID = `forked-command-${command.name}` let toolUseCounter = 0 // Helper to create a progress message from an agent message const createProgressMessage = ( message: AssistantMessage | NormalizedUserMessage, ): ProgressMessage => { toolUseCounter++ return { type: 'progress', data: { message, type: 'agent_progress', prompt: skillContent, agentId, }, parentToolUseID, toolUseID: `${parentToolUseID}-${toolUseCounter}`, timestamp: new Date().toISOString(), uuid: randomUUID(), } } // Helper to update progress display using agent progress UI const updateProgress = (): void => { setToolJSX({ jsx: renderToolUseProgressMessage(progressMessages, { tools: context.options.tools, verbose: false, }), shouldHidePromptInput: false, shouldContinueAnimation: true, showSpinner: true, }) } // Show initial "Initializing…" state updateProgress() // Run the sub-agent try { for await (const message of runAgent({ agentDefinition, promptMessages, toolUseContext: { ...context, getAppState: modifiedGetAppState, }, canUseTool, isAsync: false, querySource: 'agent:custom', model: command.model as ModelAlias | undefined, availableTools: context.options.tools, })) { agentMessages.push(message) const normalizedNew = normalizeMessages([message]) // Add progress message for assistant messages (which contain tool uses) if (message.type === 'assistant') { // Increment token count in spinner for assistant messages const contentLength = getAssistantMessageContentLength(message as AssistantMessage) if (contentLength > 0) { context.setResponseLength(len => len + contentLength) } const normalizedMsg = normalizedNew[0] if (normalizedMsg && normalizedMsg.type === 'assistant') { progressMessages.push(createProgressMessage(message as AssistantMessage)) updateProgress() } } // Add progress message for user messages (which contain tool results) if (message.type === 'user') { const normalizedMsg = normalizedNew[0] if (normalizedMsg && normalizedMsg.type === 'user') { progressMessages.push(createProgressMessage(normalizedMsg as AssistantMessage)) updateProgress() } } } } finally { // Clear the progress display setToolJSX(null) } let resultText = extractResultText(agentMessages, 'Command completed') logForDebugging( `Forked slash command /${command.name} completed with agent ${agentId}`, ) // Prepend debug log for ant users so it appears inside the command output if (process.env.USER_TYPE === 'ant') { resultText = `[ANT-ONLY] API calls: ${getDisplayPath(getDumpPromptsPath(agentId))}\n${resultText}` } // Return the result as a user message (simulates the agent's output) const messages: UserMessage[] = [ createUserMessage({ content: prepareUserContent({ inputString: `/${getCommandName(command)} ${args}`.trim(), precedingInputBlocks, }), }), createUserMessage({ content: `\n${resultText}\n`, }), ] return { messages, shouldQuery: false, command, resultText, } } /** * Determines if a string looks like a valid command name. * Valid command names only contain letters, numbers, colons, hyphens, and underscores. * * @param commandName - The potential command name to check * @returns true if it looks like a command name, false if it contains non-command characters */ export function looksLikeCommand(commandName: string): boolean { // Command names should only contain [a-zA-Z0-9:_-] // If it contains other characters, it's probably a file path or other input return !/[^a-zA-Z0-9:\-_]/.test(commandName) } export async function processSlashCommand( inputString: string, precedingInputBlocks: ContentBlockParam[], imageContentBlocks: ContentBlockParam[], attachmentMessages: AttachmentMessage[], context: ProcessUserInputContext, setToolJSX: SetToolJSXFn, uuid?: string, isAlreadyProcessing?: boolean, canUseTool?: CanUseToolFn, ): Promise { const parsed = parseSlashCommand(inputString) if (!parsed) { logEvent('tengu_input_slash_missing', {}) const errorMessage = 'Commands are in the form `/command [args]`' return { messages: [ createSyntheticUserCaveatMessage(), ...attachmentMessages, createUserMessage({ content: prepareUserContent({ inputString: errorMessage, precedingInputBlocks, }), }), ], shouldQuery: false, resultText: errorMessage, } } const { commandName, args: parsedArgs, isMcp } = parsed const sanitizedCommandName = isMcp ? 'mcp' : !builtInCommandNames().has(commandName) ? 'custom' : commandName // Check if it's a real command before processing if (!hasCommand(commandName, context.options.commands)) { // Check if this looks like a command name vs a file path or other input // Also check if it's an actual file path that exists let isFilePath = false try { await getFsImplementation().stat(`/${commandName}`) isFilePath = true } catch { // Not a file path — treat as command name } if (looksLikeCommand(commandName) && !isFilePath) { logEvent('tengu_input_slash_invalid', { input: commandName as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, }) const unknownMessage = `Unknown skill: ${commandName}` return { messages: [ createSyntheticUserCaveatMessage(), ...attachmentMessages, createUserMessage({ content: prepareUserContent({ inputString: unknownMessage, precedingInputBlocks, }), }), // gh-32591: preserve args so the user can copy/resubmit without // retyping. System warning is UI-only (filtered before API). ...(parsedArgs ? [ createSystemMessage( `Args from unknown skill: ${parsedArgs}`, 'warning', ), ] : []), ], shouldQuery: false, resultText: unknownMessage, } } const promptId = randomUUID() setPromptId(promptId) logEvent('tengu_input_prompt', {}) // Log user prompt event for OTLP void logOTelEvent('user_prompt', { prompt_length: String(inputString.length), prompt: redactIfDisabled(inputString), 'prompt.id': promptId, }) return { messages: [ createUserMessage({ content: prepareUserContent({ inputString, precedingInputBlocks }), uuid: uuid, }), ...attachmentMessages, ], shouldQuery: true, } } // Track slash command usage for feature discovery const { messages: newMessages, shouldQuery: messageShouldQuery, allowedTools, model, effort, command: returnedCommand, resultText, nextInput, submitNextInput, } = await getMessagesForSlashCommand( commandName, parsedArgs, setToolJSX, context, precedingInputBlocks, imageContentBlocks, isAlreadyProcessing, canUseTool, uuid, ) // Local slash commands that skip messages if (newMessages.length === 0) { const eventData: Record = { input: sanitizedCommandName as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, } // Add plugin metadata if this is a plugin command if (returnedCommand.type === 'prompt' && returnedCommand.pluginInfo) { const { pluginManifest, repository } = returnedCommand.pluginInfo const { marketplace } = parsePluginIdentifier(repository) const isOfficial = isOfficialMarketplaceName(marketplace) // _PROTO_* routes to PII-tagged plugin_name/marketplace_name BQ columns // (unredacted, all users); plugin_name/plugin_repository stay in // additional_metadata as redacted variants for general-access dashboards. eventData._PROTO_plugin_name = pluginManifest.name as AnalyticsMetadata_I_VERIFIED_THIS_IS_PII_TAGGED if (marketplace) { eventData._PROTO_marketplace_name = marketplace as AnalyticsMetadata_I_VERIFIED_THIS_IS_PII_TAGGED } eventData.plugin_repository = ( isOfficial ? repository : 'third-party' ) as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS eventData.plugin_name = ( isOfficial ? pluginManifest.name : 'third-party' ) as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS if (isOfficial && pluginManifest.version) { eventData.plugin_version = pluginManifest.version as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS } Object.assign( eventData, buildPluginCommandTelemetryFields(returnedCommand.pluginInfo), ) } logEvent('tengu_input_command', { ...eventData, invocation_trigger: 'user-slash' as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, ...(process.env.USER_TYPE === 'ant' && { skill_name: commandName as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, ...(returnedCommand.type === 'prompt' && { skill_source: returnedCommand.source as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, }), ...(returnedCommand.loadedFrom && { skill_loaded_from: returnedCommand.loadedFrom as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, }), ...(returnedCommand.kind && { skill_kind: returnedCommand.kind as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, }), }), }) return { messages: [], shouldQuery: false, model, nextInput, submitNextInput, } } // For invalid commands, preserve both the user message and error if ( newMessages.length === 2 && newMessages[1]!.type === 'user' && typeof newMessages[1]!.message.content === 'string' && newMessages[1]!.message.content.startsWith('Unknown command:') ) { // Don't log as invalid if it looks like a common file path const looksLikeFilePath = inputString.startsWith('/var') || inputString.startsWith('/tmp') || inputString.startsWith('/private') if (!looksLikeFilePath) { logEvent('tengu_input_slash_invalid', { input: commandName as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, }) } return { messages: [createSyntheticUserCaveatMessage(), ...newMessages], shouldQuery: messageShouldQuery, allowedTools, model, } } // A valid command const eventData: Record = { input: sanitizedCommandName as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, } // Add plugin metadata if this is a plugin command if (returnedCommand.type === 'prompt' && returnedCommand.pluginInfo) { const { pluginManifest, repository } = returnedCommand.pluginInfo const { marketplace } = parsePluginIdentifier(repository) const isOfficial = isOfficialMarketplaceName(marketplace) eventData._PROTO_plugin_name = pluginManifest.name as AnalyticsMetadata_I_VERIFIED_THIS_IS_PII_TAGGED if (marketplace) { eventData._PROTO_marketplace_name = marketplace as AnalyticsMetadata_I_VERIFIED_THIS_IS_PII_TAGGED } eventData.plugin_repository = ( isOfficial ? repository : 'third-party' ) as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS eventData.plugin_name = ( isOfficial ? pluginManifest.name : 'third-party' ) as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS if (isOfficial && pluginManifest.version) { eventData.plugin_version = pluginManifest.version as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS } Object.assign( eventData, buildPluginCommandTelemetryFields(returnedCommand.pluginInfo), ) } logEvent('tengu_input_command', { ...eventData, invocation_trigger: 'user-slash' as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, ...(process.env.USER_TYPE === 'ant' && { skill_name: commandName as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, ...(returnedCommand.type === 'prompt' && { skill_source: returnedCommand.source as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, }), ...(returnedCommand.loadedFrom && { skill_loaded_from: returnedCommand.loadedFrom as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, }), ...(returnedCommand.kind && { skill_kind: returnedCommand.kind as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS, }), }), }) // Check if this is a compact result which handle their own synthetic caveat message ordering const isCompactResult = newMessages.length > 0 && newMessages[0] && isCompactBoundaryMessage(newMessages[0]) return { messages: messageShouldQuery || newMessages.every(isSystemLocalCommandMessage) || isCompactResult ? newMessages : [createSyntheticUserCaveatMessage(), ...newMessages], shouldQuery: messageShouldQuery, allowedTools, model, effort, resultText, nextInput, submitNextInput, } } async function getMessagesForSlashCommand( commandName: string, args: string, setToolJSX: SetToolJSXFn, context: ProcessUserInputContext, precedingInputBlocks: ContentBlockParam[], imageContentBlocks: ContentBlockParam[], _isAlreadyProcessing?: boolean, canUseTool?: CanUseToolFn, uuid?: string, ): Promise { const command = getCommand(commandName, context.options.commands) // Track skill usage for ranking (only for prompt commands that are user-invocable) if (command.type === 'prompt' && command.userInvocable !== false) { recordSkillUsage(commandName) } // Check if the command is user-invocable // Skills with userInvocable === false can only be invoked by the model via SkillTool if (command.userInvocable === false) { return { messages: [ createUserMessage({ content: prepareUserContent({ inputString: `/${commandName}`, precedingInputBlocks, }), }), createUserMessage({ content: `This skill can only be invoked by Claude, not directly by users. Ask Claude to use the "${commandName}" skill for you.`, }), ], shouldQuery: false, command, } } try { switch (command.type) { case 'local-jsx': { return new Promise(resolve => { let doneWasCalled = false const onDone = ( result?: string, options?: { display?: CommandResultDisplay shouldQuery?: boolean metaMessages?: string[] nextInput?: string submitNextInput?: boolean }, ) => { doneWasCalled = true // If display is 'skip', don't add any messages to the conversation if (options?.display === 'skip') { void resolve({ messages: [], shouldQuery: false, command, nextInput: options?.nextInput, submitNextInput: options?.submitNextInput, }) return } // Meta messages are model-visible but hidden from the user const metaMessages = (options?.metaMessages ?? []).map( (content: string) => createUserMessage({ content, isMeta: true }), ) // In fullscreen the command just showed as a centered modal // pane — the transient notification is enough feedback. The // "❯ /config" + "⎿ dismissed" transcript entries are // type:system subtype:local_command (user-visible but NOT sent // to the model), so skipping them doesn't affect model context. // Outside fullscreen keep them so scrollback shows what ran. // Only skip " dismissed" modal-close notifications — // commands that early-exit before showing a modal (/ultraplan // usage, /rename, /proactive) use display:system for actual // output that must reach the transcript. const skipTranscript = isFullscreenEnvEnabled() && typeof result === 'string' && result.endsWith(' dismissed') void resolve({ messages: options?.display === 'system' ? skipTranscript ? metaMessages : [ createCommandInputMessage( formatCommandInput(command, args), ), createCommandInputMessage( `${result}`, ), ...metaMessages, ] : [ createUserMessage({ content: prepareUserContent({ inputString: formatCommandInput(command, args), precedingInputBlocks, }), }), result ? createUserMessage({ content: `${result}`, }) : createUserMessage({ content: `${NO_CONTENT_MESSAGE}`, }), ...metaMessages, ], shouldQuery: options?.shouldQuery ?? false, command, nextInput: options?.nextInput, submitNextInput: options?.submitNextInput, }) } void command .load() .then(mod => mod.call(onDone, { ...context, canUseTool }, args)) .then(jsx => { if (jsx == null) return if (context.options.isNonInteractiveSession) { void resolve({ messages: [], shouldQuery: false, command, }) return } // Guard: if onDone fired during mod.call() (early-exit path // that calls onDone then returns JSX), skip setToolJSX. This // chain is fire-and-forget — the outer Promise resolves when // onDone is called, so executeUserInput may have already run // its setToolJSX({clearLocalJSX: true}) before we get here. // Setting isLocalJSXCommand after clear leaves it stuck true, // blocking useQueueProcessor and TextInput focus. if (doneWasCalled) return setToolJSX({ jsx, shouldHidePromptInput: true, showSpinner: false, isLocalJSXCommand: true, isImmediate: command.immediate === true, }) }) .catch(e => { // If load()/call() throws and onDone never fired, the outer // Promise hangs forever, leaving queryGuard stuck in // 'dispatching' and deadlocking the queue processor. logError(e) if (doneWasCalled) return doneWasCalled = true setToolJSX({ jsx: null, shouldHidePromptInput: false, clearLocalJSX: true, }) void resolve({ messages: [], shouldQuery: false, command }) }) }) } case 'local': { const displayArgs = command.isSensitive && args.trim() ? '***' : args const userMessage = createUserMessage({ content: prepareUserContent({ inputString: formatCommandInput(command, displayArgs), precedingInputBlocks, }), }) try { const syntheticCaveatMessage = createSyntheticUserCaveatMessage() const mod = await command.load() const result = await mod.call(args, context) if (result.type === 'skip') { return { messages: [], shouldQuery: false, command, } } // Use discriminated union to handle different result types if (result.type === 'compact') { // Append slash command messages to messagesToKeep so that // attachments and hookResults come after user messages const slashCommandMessages = [ syntheticCaveatMessage, userMessage, ...(result.displayText ? [ createUserMessage({ content: `${result.displayText}`, // --resume looks at latest timestamp message to determine which message to resume from // This is a perf optimization to avoid having to recaculcate the leaf node every time // Since we're creating a bunch of synthetic messages for compact, it's important to set // the timestamp of the last message to be slightly after the current time // This is mostly important for sdk / -p mode timestamp: new Date(Date.now() + 100).toISOString(), }), ] : []), ] const compactionResultWithSlashMessages = { ...result.compactionResult, messagesToKeep: [ ...(result.compactionResult.messagesToKeep ?? []), ...slashCommandMessages, ], } // Reset microcompact state since full compact replaces all // messages — old tool IDs are no longer relevant. Budget state // (on toolUseContext) needs no reset: stale entries are inert // (UUIDs never repeat, so they're never looked up). resetMicrocompactState() return { messages: buildPostCompactMessages( compactionResultWithSlashMessages, ) as AssistantMessage[], shouldQuery: false, command, } } // Text result — use system message so it doesn't render as a user bubble return { messages: [ userMessage, createCommandInputMessage( `${result.value}`, ), ], shouldQuery: false, command, resultText: result.value, } } catch (e) { logError(e) return { messages: [ userMessage, createCommandInputMessage( `${String(e)}`, ), ], shouldQuery: false, command, } } } case 'prompt': { try { // Check if command should run as forked sub-agent if (command.context === 'fork') { return await executeForkedSlashCommand( command, args, context, precedingInputBlocks, setToolJSX, canUseTool ?? hasPermissionsToUseTool, ) } return await getMessagesForPromptSlashCommand( command, args, context, precedingInputBlocks, imageContentBlocks, uuid, ) } catch (e) { // Handle abort errors specially to show proper "Interrupted" message if (e instanceof AbortError) { return { messages: [ createUserMessage({ content: prepareUserContent({ inputString: formatCommandInput(command, args), precedingInputBlocks, }), }), createUserInterruptionMessage({ toolUse: false }), ], shouldQuery: false, command, } } return { messages: [ createUserMessage({ content: prepareUserContent({ inputString: formatCommandInput(command, args), precedingInputBlocks, }), }), createUserMessage({ content: `${String(e)}`, }), ], shouldQuery: false, command, } } } } } catch (e) { if (e instanceof MalformedCommandError) { return { messages: [ createUserMessage({ content: prepareUserContent({ inputString: e.message, precedingInputBlocks, }), }), ], shouldQuery: false, command, } } throw e } } function formatCommandInput(command: CommandBase, args: string): string { return formatCommandInputTags(getCommandName(command), args) } /** * Formats the metadata for a skill loading message. * Used by the Skill tool and for subagent skill preloading. */ export function formatSkillLoadingMetadata( skillName: string, _progressMessage: string = 'loading', ): string { // Use skill name only - UserCommandMessage renders as "Skill(name)" return [ `<${COMMAND_MESSAGE_TAG}>${skillName}`, `<${COMMAND_NAME_TAG}>${skillName}`, `true`, ].join('\n') } /** * Formats the metadata for a slash command loading message. */ function formatSlashCommandLoadingMetadata( commandName: string, args?: string, ): string { return [ `<${COMMAND_MESSAGE_TAG}>${commandName}`, `<${COMMAND_NAME_TAG}>/${commandName}`, args ? `${args}` : null, ] .filter(Boolean) .join('\n') } /** * Formats the loading metadata for a command (skill or slash command). * User-invocable skills use slash command format (/name), while model-only * skills use the skill format ("The X skill is running"). */ function formatCommandLoadingMetadata( command: CommandBase & PromptCommand, args?: string, ): string { // Use command.name (the qualified name including plugin prefix, e.g. // "product-management:feature-spec") instead of userFacingName() which may // strip the plugin prefix via displayName fallback. // User-invocable skills should show as /command-name like regular slash commands if (command.userInvocable !== false) { return formatSlashCommandLoadingMetadata(command.name, args) } // Model-only skills (userInvocable: false) show as "The X skill is running" if ( command.loadedFrom === 'skills' || command.loadedFrom === 'plugin' || command.loadedFrom === 'mcp' ) { return formatSkillLoadingMetadata(command.name, command.progressMessage) } return formatSlashCommandLoadingMetadata(command.name, args) } export async function processPromptSlashCommand( commandName: string, args: string, commands: Command[], context: ToolUseContext, imageContentBlocks: ContentBlockParam[] = [], ): Promise { const command = findCommand(commandName, commands) if (!command) { throw new MalformedCommandError(`Unknown command: ${commandName}`) } if (command.type !== 'prompt') { throw new Error( `Unexpected ${command.type} command. Expected 'prompt' command. Use /${commandName} directly in the main conversation.`, ) } return getMessagesForPromptSlashCommand( command, args, context, [], imageContentBlocks, ) } async function getMessagesForPromptSlashCommand( command: CommandBase & PromptCommand, args: string, context: ToolUseContext, precedingInputBlocks: ContentBlockParam[] = [], imageContentBlocks: ContentBlockParam[] = [], uuid?: string, ): Promise { // In coordinator mode (main thread only), skip loading the full skill content // and permissions. The coordinator only has Agent + TaskStop tools, so the // skill content and allowedTools are useless. Instead, send a brief summary // telling the coordinator how to delegate this skill to a worker. // // Workers run in-process and inherit CLAUDE_CODE_COORDINATOR_MODE from the // parent env, so we also check !context.agentId: agentId is only set for // subagents, letting workers fall through to getPromptForCommand and receive // the real skill content when they invoke the Skill tool. if ( feature('COORDINATOR_MODE') && isEnvTruthy(process.env.CLAUDE_CODE_COORDINATOR_MODE) && !context.agentId ) { const metadata = formatCommandLoadingMetadata(command, args) const parts: string[] = [ `Skill "/${command.name}" is available for workers.`, ] if (command.description) { parts.push(`Description: ${command.description}`) } if (command.whenToUse) { parts.push(`When to use: ${command.whenToUse}`) } const skillAllowedTools = command.allowedTools ?? [] if (skillAllowedTools.length > 0) { parts.push( `This skill grants workers additional tool permissions: ${skillAllowedTools.join(', ')}`, ) } parts.push( `\nInstruct a worker to use this skill by including "Use the /${command.name} skill" in your Agent prompt. The worker has access to the Skill tool and will receive the skill's content and permissions when it invokes it.`, ) const summaryContent: ContentBlockParam[] = [ { type: 'text', text: parts.join('\n') }, ] return { messages: [ createUserMessage({ content: metadata, uuid }), createUserMessage({ content: summaryContent, isMeta: true }), ], shouldQuery: true, model: command.model, effort: command.effort, command, } } const result = await command.getPromptForCommand(args, context) // Register skill hooks if defined. Under ["hooks"]-only (skills not locked), // user skills still load and reach this point — block hook REGISTRATION here // where source is known. Mirrors the agent frontmatter gate in runAgent.ts. const hooksAllowedForThisSkill = !isRestrictedToPluginOnly('hooks') || isSourceAdminTrusted(command.source) if (command.hooks && hooksAllowedForThisSkill) { const sessionId = getSessionId() registerSkillHooks( context.setAppState, sessionId, command.hooks, command.name, command.type === 'prompt' ? command.skillRoot : undefined, ) } // Record skill invocation for compaction preservation, scoped by agent context. // Skills are tagged with their agentId so only skills belonging to the current // agent are restored during compaction (preventing cross-agent leaks). const skillPath = command.source ? `${command.source}:${command.name}` : command.name const skillContent = result .filter((b): b is TextBlockParam => b.type === 'text') .map(b => b.text) .join('\n\n') addInvokedSkill( command.name, skillPath, skillContent, getAgentContext()?.agentId ?? null, ) const metadata = formatCommandLoadingMetadata(command, args) const additionalAllowedTools = parseToolListFromCLI( command.allowedTools ?? [], ) // Create content for the main message, including any pasted images const mainMessageContent: ContentBlockParam[] = imageContentBlocks.length > 0 || precedingInputBlocks.length > 0 ? [...imageContentBlocks, ...precedingInputBlocks, ...result] : result // Extract attachments from command arguments (@-mentions, MCP resources, // agent mentions in SKILL.md). skipSkillDiscovery prevents the SKILL.md // content itself from triggering discovery — it's meta-content, not user // intent, and a large SKILL.md (e.g. 110KB) would fire chunked AKI queries // adding seconds of latency to every skill invocation. const attachmentMessages = await toArray( getAttachmentMessages( result .filter((block): block is TextBlockParam => block.type === 'text') .map(block => block.text) .join(' '), context, null, [], // queuedCommands - handled by query.ts for mid-turn attachments context.messages, 'repl_main_thread', { skipSkillDiscovery: true }, ), ) const messages = [ createUserMessage({ content: metadata, uuid, }), createUserMessage({ content: mainMessageContent, isMeta: true, }), ...attachmentMessages, createAttachmentMessage({ type: 'command_permissions', allowedTools: additionalAllowedTools, model: command.model, }), ] return { messages, shouldQuery: true, allowedTools: additionalAllowedTools, model: command.model, effort: command.effort, command, } }