kimaki 0.4.24 → 0.4.26

package/src/system-message.ts

@@ -1,3 +1,7 @@
+// OpenCode system prompt generator.
+// Creates the system message injected into every OpenCode session,
+// including Discord-specific formatting rules, diff commands, and permissions info.
+
 export function getOpencodeSystemMessage({ sessionId }: { sessionId: string }) {
   return `
 The user is reading your messages from inside Discord, via kimaki.xyz
@@ -14,24 +18,6 @@ Only users with these Discord permissions can send messages to the bot:
 - Manage Server permission
 - "Kimaki" role (case-insensitive)
 
-## changing the model
-
-To change the model used by OpenCode, edit the project's \`opencode.json\` config file and set the \`model\` field:
-
-\`\`\`json
-{
-  "model": "anthropic/claude-sonnet-4-20250514"
-}
-\`\`\`
-
-Examples:
-- \`"anthropic/claude-sonnet-4-20250514"\` - Claude Sonnet 4
-- \`"anthropic/claude-opus-4-20250514"\` - Claude Opus 4
-- \`"openai/gpt-4o"\` - GPT-4o
-- \`"google/gemini-2.5-pro"\` - Gemini 2.5 Pro
-
-Format is \`provider/model-name\`. You can also set \`small_model\` for tasks like title generation.
-
 ## uploading files to discord
 
 To upload files to the Discord thread (images, screenshots, long files that would clutter the chat), run:
@@ -52,7 +38,9 @@ bunx critique web -- path/to/file1.ts path/to/file2.ts
 
 You can also show latest commit changes using:
 
-bunx critique web HEAD~1
+bunx critique web HEAD
+
+bunx critique web HEAD~1 to get the one before last
 
 Do this in case you committed the changes yourself (only if the user asks so, never commit otherwise).
 
@@ -66,27 +54,9 @@ the max heading level is 3, so do not use ####
 
 headings are discouraged anyway. instead try to use bold text for titles which renders more nicely in Discord
 
-## tables
-
-discord does NOT support markdown gfm tables.
-
-so instead of using full markdown tables ALWAYS show code snippets with space aligned cells:
-
-\`\`\`
-Item        Qty   Price
-----------  ---   -----
-Apples      10    $5
-Oranges     3     $2
-\`\`\`
-
-Using code blocks will make the content use monospaced font so that space will be aligned correctly
-
-IMPORTANT: add enough space characters to align the table! otherwise the content will not look good and will be difficult to understand for the user
-
-code blocks for tables and diagrams MUST have Max length of 85 characters. otherwise the content will wrap
 
 ## diagrams
 
-you can create diagrams wrapping them in code blocks too.
+you can create diagrams wrapping them in code blocks.
 `
 }

package/src/tools.ts

@@ -1,3 +1,7 @@
+// Voice assistant tool definitions for the GenAI worker.
+// Provides tools for managing OpenCode sessions (create, submit, abort),
+// listing chats, searching files, and reading session messages.
+
 import { tool } from 'ai'
 import { z } from 'zod'
 import { spawn, type ChildProcess } from 'node:child_process'

package/src/utils.ts

@@ -1,3 +1,7 @@
+// General utility functions for the bot.
+// Includes Discord OAuth URL generation, array deduplication,
+// abort error detection, and date/time formatting helpers.
+
 import { PermissionsBitField } from 'discord.js'
 
 type GenerateInstallUrlOptions = {

package/src/voice-handler.ts

@@ -1,3 +1,7 @@
+// Discord voice channel connection and audio stream handler.
+// Manages joining/leaving voice channels, captures user audio, resamples to 16kHz,
+// and routes audio to the GenAI worker for real-time voice assistant interactions.
+
 import {
   VoiceConnectionStatus,
   EndBehaviorType,
@@ -430,14 +434,16 @@ export async function processVoiceAttachment({
   projectDirectory,
   isNewThread = false,
   appId,
-  sessionMessages,
+  currentSessionContext,
+  lastSessionContext,
 }: {
   message: Message
   thread: ThreadChannel
   projectDirectory?: string
   isNewThread?: boolean
   appId?: string
-  sessionMessages?: string
+  currentSessionContext?: string
+  lastSessionContext?: string
 }): Promise<string | null> {
   const audioAttachment = Array.from(message.attachments.values()).find(
     (attachment) => attachment.contentType?.startsWith('audio/'),
@@ -487,13 +493,22 @@ export async function processVoiceAttachment({
     }
   }
 
-  const transcription = await transcribeAudio({
-    audio: audioBuffer,
-    prompt: transcriptionPrompt,
-    geminiApiKey,
-    directory: projectDirectory,
-    sessionMessages,
-  })
+  let transcription: string
+  try {
+    transcription = await transcribeAudio({
+      audio: audioBuffer,
+      prompt: transcriptionPrompt,
+      geminiApiKey,
+      directory: projectDirectory,
+      currentSessionContext,
+      lastSessionContext,
+    })
+  } catch (error) {
+    const errMsg = error instanceof Error ? error.message : String(error)
+    voiceLogger.error(`Transcription failed:`, error)
+    await sendThreadMessage(thread, `⚠️ Transcription failed: ${errMsg}`)
+    return null
+  }
 
   voiceLogger.log(
     `Transcription successful: "${transcription.slice(0, 50)}${transcription.length > 50 ? '...' : ''}"`,

package/src/voice.ts

@@ -1,3 +1,7 @@
+// Audio transcription service using Google Gemini.
+// Transcribes voice messages with code-aware context, using grep/glob tools
+// to verify technical terms, filenames, and function names in the codebase.
+
 import {
   GoogleGenAI,
   Type,
@@ -48,7 +52,8 @@ async function runGrep({
       .join('\n')
 
     return output.slice(0, 2000)
-  } catch {
+  } catch (e) {
+    voiceLogger.error('grep search failed:', e)
     return 'grep search failed'
   }
 }
@@ -300,7 +305,8 @@ export async function transcribeAudio({
   temperature,
   geminiApiKey,
   directory,
-  sessionMessages,
+  currentSessionContext,
+  lastSessionContext,
 }: {
   audio: Buffer | Uint8Array | ArrayBuffer | string
   prompt?: string
@@ -308,7 +314,8 @@ export async function transcribeAudio({
   temperature?: number
   geminiApiKey?: string
   directory?: string
-  sessionMessages?: string
+  currentSessionContext?: string
+  lastSessionContext?: string
 }): Promise<string> {
   try {
     const apiKey = geminiApiKey || process.env.GEMINI_API_KEY
@@ -334,6 +341,22 @@ export async function transcribeAudio({
 
     const languageHint = language ? `The audio is in ${language}.\n\n` : ''
 
+    // build session context section
+    const sessionContextParts: string[] = []
+    if (lastSessionContext) {
+      sessionContextParts.push(`<last_session>
+${lastSessionContext}
+</last_session>`)
+    }
+    if (currentSessionContext) {
+      sessionContextParts.push(`<current_session>
+${currentSessionContext}
+</current_session>`)
+    }
+    const sessionContextSection = sessionContextParts.length > 0
+      ? `\nSession context (use to understand references to files, functions, tools used):\n${sessionContextParts.join('\n\n')}`
+      : ''
+
     const transcriptionPrompt = `${languageHint}Transcribe this audio for a coding agent (like Claude Code or OpenCode).
 
 CRITICAL REQUIREMENT: You MUST call the "transcriptionResult" tool to complete this task.
@@ -347,30 +370,30 @@ This is a software development environment. The speaker is giving instructions t
 - File paths, function names, CLI commands, package names, API endpoints
 
 RULES:
-1. You have LIMITED tool calls - use grep/glob sparingly, call them in parallel
-2. If audio is unclear, transcribe your best interpretation
-3. If audio seems silent/empty, call transcriptionResult with "[inaudible audio]"
-4. When warned about remaining steps, STOP searching and call transcriptionResult immediately
+1. If audio is unclear, transcribe your best interpretation, interpreting words event with strong accents are present, identifying the accent being used first so you can guess what the words meawn
+2. If audio seems silent/empty, call transcriptionResult with "[inaudible audio]"
+3. Use the session context below to understand technical terms, file names, function names mentioned
 
 Common corrections (apply without tool calls):
 - "reacked" → "React", "jason" → "JSON", "get hub" → "GitHub", "no JS" → "Node.js", "dacker" → "Docker"
 
-Project context for reference:
-<context>
+Project file structure:
+<file_tree>
 ${prompt}
-</context>
-${sessionMessages ? `\nRecent session messages:\n<session_messages>\n${sessionMessages}\n</session_messages>` : ''}
+</file_tree>
+${sessionContextSection}
 
 REMEMBER: Call "transcriptionResult" tool with your transcription. This is mandatory.
 
 Note: "critique" is a CLI tool for showing diffs in the browser.`
 
-    const hasDirectory = directory && directory.trim().length > 0
+    // const hasDirectory = directory && directory.trim().length > 0
     const tools = [
       {
         functionDeclarations: [
           transcriptionResultToolDeclaration,
-          ...(hasDirectory ? [grepToolDeclaration, globToolDeclaration] : []),
+          // grep/glob disabled - was causing transcription to hang
+          // ...(hasDirectory ? [grepToolDeclaration, globToolDeclaration] : []),
         ],
       },
     ]

package/src/worker-types.ts

@@ -1,3 +1,7 @@
+// Type definitions for worker thread message passing.
+// Defines the protocol between main thread and GenAI worker for
+// audio streaming, tool calls, and session lifecycle management.
+
 import type { Tool as AITool } from 'ai'
 
 // Messages sent from main thread to worker

package/src/xml.ts

@@ -1,3 +1,7 @@
+// XML/HTML tag content extractor.
+// Parses XML-like tags from strings (e.g., channel topics) to extract
+// Kimaki configuration like directory paths and app IDs.
+
 import { DomHandler, Parser, ElementType } from 'htmlparser2'
 import type { ChildNode, Element, Text } from 'domhandler'
 import { createLogger } from './logger.js'

package/README.md

@@ -1,48 +0,0 @@
-# Kimaki Discord Bot
-
-A Discord bot that integrates OpenCode coding sessions with Discord channels and voice.
-
-## Installation
-
-```bash
-npm install -g kimaki
-```
-
-## Setup
-
-Run the interactive setup:
-
-```bash
-kimaki
-```
-
-This will guide you through:
-1. Creating a Discord application at https://discord.com/developers/applications
-2. Getting your bot token
-3. Installing the bot to your Discord server
-4. Creating channels for your OpenCode projects
-
-## Commands
-
-### Start the bot
-
-```bash
-kimaki
-```
-
-## Discord Slash Commands
-
-Once the bot is running, you can use these commands in Discord:
-
-- `/session <prompt>` - Start a new OpenCode session
-- `/resume <session>` - Resume an existing session
-- `/add-project <project>` - Add a new project to Discord
-- `/accept` - Accept a permission request
-- `/accept-always` - Accept and auto-approve similar requests
-- `/reject` - Reject a permission request
-
-## Voice Support
-
-Join a voice channel that has an associated project directory, and the bot will join with Jarvis-like voice interaction powered by Gemini.
-
-Requires a Gemini API key (prompted during setup).