Slash command parse and dispatch

Overview

This flow covers how a user-typed string starting with / is parsed into a command name and arguments, then dispatched to the correct handler fn:processSlashCommand. It is triggered from the prompt input submission path after the base user-input processor detects a slash prefix fn:processUserInputBase.

Steps

1. The user submits text from the prompt input component, which routes non-bash input into the base user-input processor fn:PromptInput. That processor normalizes images and content blocks before considering slash-command handling fn:processUserInputBase.
2. For bridge-origin input the base processor calls parseSlashCommand early to decide whether the command is safe to run remotely, short-circuiting unsafe commands with a local stdout message fn:processUserInputBase. Otherwise it hands the raw string to processSlashCommand fn:processUserInputBase.
3. processSlashCommand invokes parseSlashCommand, which trims the input, verifies the leading /, splits on spaces, extracts commandName, detects an (MCP) marker, and joins the rest as args into a ParsedSlashCommand fn:parseSlashCommand schema:ParsedSlashCommand. A null result makes processSlashCommand return a synthetic "Commands are in the form /command [args]" error fn:processSlashCommand.
4. processSlashCommand then checks hasCommand; if unknown it stats /commandName via the fs implementation to distinguish a file path from a typo, using looksLikeCommand as a character-class filter fn:processSlashCommand fn:getFsImplementation fn:looksLikeCommand. Unknown-but-command-shaped input logs tengu_input_slash_invalid and returns an "Unknown skill" message fn:processSlashCommand fn:logEvent.
5. For a recognized command processSlashCommand delegates to getMessagesForSlashCommand with the parsed name, args, and ProcessUserInputContext fn:processSlashCommand schema:ProcessUserInputContext. That function resolves the Command object via getCommand, which calls findCommand against context.options.commands and throws a ReferenceError listing available commands if missing fn:getMessagesForSlashCommand fn:getCommand.
6. If the resolved command is a user-invocable prompt skill, usage is recorded before dispatch via the debounced tracker that persists through saveGlobalConfig fn:getMessagesForSlashCommand fn:recordSkillUsage fn:saveGlobalConfig. Commands with userInvocable === false are rejected with a message telling the user to ask Claude to invoke the skill fn:getMessagesForSlashCommand.
7. getMessagesForSlashCommand then switches on command.type: local-jsx commands resolve via a Promise-wrapped JSX handler, while prompt commands with context === 'fork' path into executeForkedSlashCommand which prepares a forked agent context and runs the sub-agent fn:getMessagesForSlashCommand fn:executeForkedSlashCommand fn:prepareForkedCommandContext fn:runAgent.
8. Back in processSlashCommand, the returned SlashCommandResult is used to log tengu_input_command with plugin telemetry fields and to assemble the final ProcessUserInputBaseResult messages that feed the main query loop fn:processSlashCommand fn:buildPluginCommandTelemetryFields schema:SlashCommandResult schema:ProcessUserInputBaseResult.
9. The SkillTool path reuses the same parsing/dispatch machinery by importing processPromptSlashCommand and calling findCommand + recordSkillUsage itself when the model (not the user) invokes a skill fn:call fn:recordSkillUsage.

State touched

• state:sentSkillNames
• state:globalConfigCache

Decisions

(none in whitelist) fn:processSlashCommand