mirror of
https://github.com/larchanka/manbot.git
synced 2026-05-13 13:39:40 +00:00
Redifine skills logic + add command to easy skill adding
This commit is contained in:
larchanka
committed by
Mikhail Larchanka
Mikhail Larchanka
parent
e948ba5ed1
commit
159e79fc67
@@ -15,8 +15,7 @@ AI-Agent/
|
||||
│ ├── INSTRUCTIONS.md # Workflow for task execution
|
||||
│ └── TASKS/ # Task specs (P1-01, P2-01, …)
|
||||
├── skills/ # Dynamic Skills System
|
||||
│ ├── CONFIG.md # Skill manifest and descriptions
|
||||
│ └── [skill-name]/ # SKILL.md for each skill
|
||||
│ └── [skill-name]/ # SKILL.md for each skill (first line has description)
|
||||
├── _docs/
|
||||
│ ├── ARCHITECTURE.md # Architectural patterns
|
||||
│ ├── CAPABILITY GRAPH.md # DAG format and node types
|
||||
|
||||
10
README.md
10
README.md
@@ -39,9 +39,9 @@ A multi-process AI platform with type-safe IPC and capability-graph execution. U
|
||||
Install and run Lemonade, then pull the models you need:
|
||||
|
||||
```bash
|
||||
ollama pull qwen2.5:0.5b
|
||||
ollama pull qwen2.5:1.5b
|
||||
ollama pull qwen3-vl
|
||||
lemonade pull qwen2.5:0.5b
|
||||
lemonade pull qwen2.5:1.5b
|
||||
lemonade pull qwen3-vl
|
||||
```
|
||||
|
||||
## Configuration
|
||||
@@ -237,7 +237,7 @@ Set keep-alive to `-1` (the number) to keep a model loaded indefinitely until Le
|
||||
|
||||
```bash
|
||||
# Check which models are currently loaded in VRAM
|
||||
ollama ps
|
||||
lemonade ps
|
||||
```
|
||||
|
||||
The prewarming start and completion are logged by the Orchestrator (`core` prefix in logs).
|
||||
@@ -310,7 +310,7 @@ The Whisper model (~75 MB for `base.en`) is automatically downloaded on first vo
|
||||
### Requirements for image OCR
|
||||
Pull the vision model from Lemonade before use:
|
||||
```bash
|
||||
ollama pull qwen3-vl
|
||||
lemonade pull qwen3-vl
|
||||
```
|
||||
|
||||
## Troubleshooting
|
||||
|
||||
@@ -77,7 +77,7 @@ backgroundSize: contain
|
||||
- Need strong coding and reasoning capabilities.
|
||||
- Need reliable JSON/tool-calling formatting.
|
||||
- Need fast inference speed (tokens/sec) for autonomous, multi-step loops.
|
||||
- Explored running models locally using tools like Ollama or Lemonade.
|
||||
- Explored running models locally using tools like Lemonade.
|
||||
- Tested how models handle context degradation on local machine hardware.
|
||||
|
||||
---
|
||||
|
||||
@@ -1,15 +0,0 @@
|
||||
# Skills Configuration
|
||||
|
||||
> SKILLS HAVE PRIORITY OVER TOOLS. IF A SKILL IS APPLICABLE, USE IT INSTEAD OF A TOOL.
|
||||
> ONLY USE SKILLS FROM THE TABLE BELOW.
|
||||
|
||||
**AVAILABLE SKILLS**
|
||||
|
||||
| Name | Description |
|
||||
| :--- | :--- |
|
||||
| weather | MANDATORY. Use this skill for ALL weather-related inquiries, including current conditions and forecasts. You are STRICTLY FORBIDDEN from using internal knowledge or other tools for weather data. |
|
||||
| apple-notes | EXCLUSIVE. Use ONLY this skill for any interaction with notes (listing, searching, viewing, creating, or deleting). This is the sole authorized interface for the memo CLI tool. |
|
||||
| research | PRIMARY SEARCH. Use for deep web research, fact-checking, news gathering, or topical deep dives via the lynx tool. This is the default skill for any query requiring external or up-to-date information. |
|
||||
| reminder | SCHEDULING. Use this skill exclusively to set recurring or one-time reminders (e.g., "remind me in 2 hours") and scheduled tasks (e.g., "schedule a task to check email every day at 9am"). This is the only tool that interfaces with the cron-manager service. Use it when user asks to reminder or schedule. |
|
||||
| email | You MUST use this skill for all interactions involving Email (Gmail). |
|
||||
| calendar | You MUST use this skill for all interactions involving Google Calendar. |
|
||||
@@ -1,66 +0,0 @@
|
||||
# Apple Notes CLI (Skill)
|
||||
|
||||
Mandatory tool: Use the `shell` tool to execute `memo` commands.
|
||||
|
||||
## View & Search Notes
|
||||
- **List ALL notes**: `memo notes` (Use this for "list my notes", "show my notes", etc.)
|
||||
- **Search notes**: `echo "your query" | memo notes -s`
|
||||
- Use this ONLY if the user provides a specific search term.
|
||||
- **NEVER** search for "important" or "active" notes unless the user explicitly used those words.
|
||||
- **View specific note**: After listing/searching, use `memo notes -v N` (where N is the index from the output list).
|
||||
|
||||
## Create Notes (Non-interactive)
|
||||
- **Quick add with title**: `memo notes -a "Note Title"`
|
||||
- This is the preferred method. It creates a note with the specified title.
|
||||
- **Add to specific folder**: `memo notes -f "Folder Name" -a "Note Title"`
|
||||
|
||||
## Folders
|
||||
- **List all folders**: `memo notes -fl`
|
||||
|
||||
## Critical Rules
|
||||
1. **Prefer Listing**: When asked to "list" or "show" notes, ALWAYS start with `memo notes`.
|
||||
2. **No Default Filters**: Do not attempt to filter or search unless the user specified a term.
|
||||
3. **Pipes for Search**: Always use `echo "query" | memo notes -s` because the shell tool is non-interactive.
|
||||
4. **Sequential Operations**: To view a note's content, you first need its index from the list.
|
||||
5. **No Attachments**: This tool only supports plain text.
|
||||
6. **macOS Only**: Ensure you are on a macOS environment.
|
||||
|
||||
## Tool Call Examples (JSON)
|
||||
|
||||
When using this skill, format your tool calls as follows:
|
||||
|
||||
### List All Notes
|
||||
```json
|
||||
{
|
||||
"name": "shell",
|
||||
"arguments": {
|
||||
"command": "memo notes"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Search for "Project Phoenix"
|
||||
```json
|
||||
{
|
||||
"name": "shell",
|
||||
"arguments": {
|
||||
"command": "echo \"Project Phoenix\" | memo notes -s"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Create a Quick Note
|
||||
```json
|
||||
{
|
||||
"name": "shell",
|
||||
"arguments": {
|
||||
"command": "memo notes -a \"Meeting Notes\""
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
## Example Commands
|
||||
- `memo notes` (List everything - Default action)
|
||||
- `echo "project X" | memo notes -s` (Search for project X)
|
||||
- `memo notes -f "Inbox" -a "Buy milk"` (Create note)
|
||||
- `memo notes -fl` (List folders)
|
||||
@@ -1,3 +1,5 @@
|
||||
Description: You MUST use this skill for all interactions involving Google Calendar.
|
||||
|
||||
# Calendar Skill
|
||||
|
||||
Manage Google Calendar events using the `gog` CLI.
|
||||
|
||||
@@ -1,3 +1,5 @@
|
||||
Description: You MUST use this skill for all interactions involving Email (Gmail).
|
||||
|
||||
# Email Skill
|
||||
|
||||
Manage Gmail communications using the `gog` CLI.
|
||||
|
||||
@@ -1,3 +1,5 @@
|
||||
Description: SCHEDULING. Use this skill exclusively to set recurring or one-time reminders (e.g., "remind me in 2 hours") and scheduled tasks (e.g., "schedule a task to check email every day at 9am"). This is the only tool that interfaces with the cron-manager service. Use it when user asks to reminder or schedule.
|
||||
|
||||
# Reminder Skill
|
||||
|
||||
Set up one-time or recurring reminders for the user.
|
||||
|
||||
@@ -1,3 +1,5 @@
|
||||
Description: PRIMARY SEARCH. Use for deep web research, fact-checking, news gathering, or topical deep dives via the lynx tool. This is the default skill for any query requiring external or up-to-date information.
|
||||
|
||||
# Research Skill
|
||||
|
||||
Deep web research using the text-based browser `lynx` and DuckDuckGo HTML interface.
|
||||
|
||||
@@ -1,3 +1,5 @@
|
||||
Description: MANDATORY. Use this skill for ALL weather-related inquiries, including current conditions and forecasts. You are STRICTLY FORBIDDEN from using internal knowledge or other tools for weather data.
|
||||
|
||||
# Weather Skill
|
||||
|
||||
Get current weather conditions and forecasts.
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
/**
|
||||
* Integration test for conversation archiving flow (P6-06).
|
||||
* Verifies: task history retrieval by conversation_id, summary insertion into RAG, SQLite persistence.
|
||||
* Summarization step is mocked (no Ollama).
|
||||
* Summarization step is mocked (no Lemonade).
|
||||
*/
|
||||
|
||||
import { randomUUID } from "node:crypto";
|
||||
|
||||
@@ -114,7 +114,10 @@ const HELP_TEXT = `Commands:
|
||||
- Ask me to remind you: "Remind me in 5 minutes to check the oven"
|
||||
- Recurring reminders: "Remind me every day at 9am to take vitamins"
|
||||
- List reminders: /reminders
|
||||
- Cancel a reminder: /cancel_reminder <id>`;
|
||||
- Cancel a reminder: /cancel_reminder <id>
|
||||
|
||||
🛠 Skills:
|
||||
- Add a new skill: /add_skill <URL_TO_SKILL_MD> (downloads to an underscore-prefixed folder)`;
|
||||
|
||||
/** chatId -> current conversation ID for session grouping */
|
||||
const conversationIdByChat = new Map<number, string>();
|
||||
@@ -402,25 +405,73 @@ function main(): void {
|
||||
return;
|
||||
}
|
||||
|
||||
// /task [goal] — if goal is provided, create task; else show usage
|
||||
if (text.startsWith("/task")) {
|
||||
const goal = text.slice(5).trim();
|
||||
if (!goal) {
|
||||
sendToUser(chatId, "Usage: /task <your goal>. Example: /task Summarize the benefits of TypeScript.");
|
||||
// /add_skill [URL] — download a new skill
|
||||
if (text.startsWith("/add_skill")) {
|
||||
const url = text.slice(10).trim();
|
||||
if (!url) {
|
||||
sendToUser(chatId, "Usage: /add_skill {URL_TO_SKILL_MD}");
|
||||
return;
|
||||
}
|
||||
const payload: TelegramTaskCreatePayload = {
|
||||
chatId,
|
||||
userId: from.id,
|
||||
conversationId: getOrCreateConversationId(chatId),
|
||||
messageId: msg.message_id ?? 0,
|
||||
goal,
|
||||
...(from.username !== undefined && from.username !== "" && { username: from.username }),
|
||||
};
|
||||
base.send(createEnvelope<TelegramTaskCreatePayload>("task.create", "core", payload));
|
||||
|
||||
if (!url.toLowerCase().endsWith("skill.md")) {
|
||||
sendToUser(chatId, "❌ URL must end with SKILL.md");
|
||||
return;
|
||||
}
|
||||
|
||||
(async () => {
|
||||
try {
|
||||
const urlObj = new URL(url);
|
||||
const pathParts = urlObj.pathname.split("/").filter(p => p !== "");
|
||||
let folderName = "";
|
||||
|
||||
// if path is /foo/bar/SKILL.md, folderName is bar
|
||||
if (pathParts.length >= 2) {
|
||||
folderName = pathParts[pathParts.length - 2] ?? "";
|
||||
}
|
||||
|
||||
if (!folderName || folderName === "." || folderName === "..") {
|
||||
folderName = randomUUID();
|
||||
}
|
||||
|
||||
if (!folderName.startsWith("_")) {
|
||||
folderName = "_" + folderName;
|
||||
}
|
||||
|
||||
const cfg = getConfig();
|
||||
const skillsDir = resolve(process.cwd(), cfg.skills.skillsDir);
|
||||
const newSkillDir = resolve(skillsDir, folderName);
|
||||
|
||||
// Create folder
|
||||
await new Promise<void>((res, rej) =>
|
||||
mkdir(newSkillDir, { recursive: true }, (err) => (err ? rej(err) : res())),
|
||||
);
|
||||
|
||||
// Download file
|
||||
const resp = await fetch(url);
|
||||
if (!resp.ok) {
|
||||
throw new Error(`Failed to fetch SKILL.md: HTTP ${resp.status}`);
|
||||
}
|
||||
if (!resp.body) {
|
||||
throw new Error("Empty response body when fetching SKILL.md");
|
||||
}
|
||||
|
||||
const targetPath = resolve(newSkillDir, "SKILL.md");
|
||||
await pipeline(
|
||||
resp.body as unknown as NodeJS.ReadableStream,
|
||||
createWriteStream(targetPath)
|
||||
);
|
||||
|
||||
await sendToUser(chatId, `✅ Skill added to folder: <code>${folderName}</code>\n\nNotes:\n- The skill is currently <b>disabled</b> (starts with <code>_</code>).\n- Rename the folder to remove the underscore to enable it.\n- Use <code>/help</code> to see available commands.`);
|
||||
} catch (err) {
|
||||
console.error("[telegram-adapter] /add_skill error:", err);
|
||||
await sendToUser(chatId, `❌ Error adding skill: ${err instanceof Error ? err.message : String(err)}`);
|
||||
}
|
||||
})();
|
||||
return;
|
||||
}
|
||||
|
||||
// /task [goal] — if goal is provided, create task; else show usage
|
||||
|
||||
// Plain text: map to user goal and create task (same as /task <text>)
|
||||
const taskPayload: TelegramTaskCreatePayload = {
|
||||
chatId,
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
/**
|
||||
* Planner Agent: converts user intent into a structured execution DAG.
|
||||
* Listens for plan.create, uses Ollama Adapter + Model Router, validates DAG, emits response.
|
||||
* Listens for plan.create, uses Lemonade Adapter + Model Router, validates DAG, emits response.
|
||||
*/
|
||||
|
||||
import { randomUUID } from "node:crypto";
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
/**
|
||||
* Generator Service: handles node.execute for generate_text and summarize.
|
||||
* Used by Executor when dispatching to "model-router"; calls Ollama via ModelRouter.
|
||||
* Used by Executor when dispatching to "model-router"; calls Lemonade via ModelRouter.
|
||||
* P6-04: summarize type uses summarizer prompt for memory extraction.
|
||||
*/
|
||||
|
||||
|
||||
@@ -226,7 +226,7 @@ export class LemonadeAdapter {
|
||||
|
||||
/**
|
||||
* Warmup model (loads it into memory).
|
||||
* Lemonade might not have a specific warmup endpoint identical to Ollama,
|
||||
* Lemonade might not have a specific warmup endpoint,
|
||||
* but sending a small message often works.
|
||||
*/
|
||||
async warmup(model: string): Promise<void> {
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
/**
|
||||
* Model Router: maps abstract complexity levels to Ollama model names.
|
||||
* Per _docs/TECH.md: small -> llama3:8b, medium -> mistral, large -> mixtral.
|
||||
* Model Router: maps abstract complexity levels to Lemonade model names.
|
||||
* Per _docs/TECH.md: small -> qwen2.5:0.5b, medium -> qwen2.5:1.5b, large -> qwen2.5:7b.
|
||||
*/
|
||||
|
||||
import { getConfig } from "../shared/config.js";
|
||||
@@ -22,7 +22,7 @@ export class ModelRouter {
|
||||
}
|
||||
|
||||
/**
|
||||
* Return the Ollama model name for the given complexity level.
|
||||
* Return the Lemonade model name for the given complexity level.
|
||||
*/
|
||||
getModel(complexity: ComplexityLevel): string {
|
||||
return this.config[complexity];
|
||||
|
||||
@@ -1,416 +0,0 @@
|
||||
/**
|
||||
* Ollama adapter: bridge to local Ollama instance for generate, chat, and streaming.
|
||||
* Uses fetch; supports timeout and retry for network errors.
|
||||
*/
|
||||
|
||||
import { readFile } from "node:fs/promises";
|
||||
import { getConfig } from "../shared/config.js";
|
||||
|
||||
|
||||
export interface GenerateOptions {
|
||||
timeoutMs?: number;
|
||||
keep_alive?: string | number;
|
||||
options?: Record<string, unknown>;
|
||||
}
|
||||
|
||||
export interface ChatOptions {
|
||||
timeoutMs?: number;
|
||||
keep_alive?: string | number;
|
||||
tools?: any[];
|
||||
options?: Record<string, unknown>;
|
||||
}
|
||||
|
||||
export interface GenerateResult {
|
||||
text: string;
|
||||
prompt_eval_count?: number;
|
||||
eval_count?: number;
|
||||
done: boolean;
|
||||
}
|
||||
|
||||
export interface ChatMessage {
|
||||
role: "system" | "user" | "assistant" | "tool";
|
||||
content: string;
|
||||
tool_calls?: ToolCall[];
|
||||
tool_call_id?: string;
|
||||
}
|
||||
|
||||
export interface ToolCall {
|
||||
id: string;
|
||||
type: string;
|
||||
function: {
|
||||
name: string;
|
||||
arguments: Record<string, any>;
|
||||
};
|
||||
}
|
||||
|
||||
export interface ChatResult {
|
||||
message: {
|
||||
role: string;
|
||||
content: string;
|
||||
tool_calls?: ToolCall[];
|
||||
};
|
||||
prompt_eval_count?: number;
|
||||
eval_count?: number;
|
||||
done: boolean;
|
||||
}
|
||||
|
||||
export interface StreamChunk {
|
||||
message?: { content: string; tool_calls?: ToolCall[] };
|
||||
done?: boolean;
|
||||
prompt_eval_count?: number;
|
||||
eval_count?: number;
|
||||
}
|
||||
|
||||
export interface EmbedResult {
|
||||
embedding: number[];
|
||||
prompt_eval_count?: number;
|
||||
}
|
||||
|
||||
export interface OllamaAdapterOptions {
|
||||
baseUrl?: string;
|
||||
timeoutMs?: number;
|
||||
retries?: number;
|
||||
}
|
||||
|
||||
export class OllamaAdapter {
|
||||
private readonly baseUrl: string;
|
||||
private readonly timeoutMs: number;
|
||||
private readonly retries: number;
|
||||
private readonly numCtx: number;
|
||||
|
||||
constructor(options: OllamaAdapterOptions = {}) {
|
||||
const c = getConfig().lemonade;
|
||||
this.baseUrl = options.baseUrl ?? c.baseUrl;
|
||||
this.timeoutMs = options.timeoutMs ?? c.timeoutMs;
|
||||
this.retries = options.retries ?? c.retries;
|
||||
this.numCtx = c.numCtx;
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate completion for a single prompt. Returns full response (stream: false).
|
||||
*/
|
||||
async generate(
|
||||
prompt: string,
|
||||
model: string,
|
||||
opts: GenerateOptions = {},
|
||||
): Promise<GenerateResult> {
|
||||
const timeoutMs = opts.timeoutMs ?? this.timeoutMs;
|
||||
const url = `${this.baseUrl}/api/generate`;
|
||||
const body: Record<string, unknown> = {
|
||||
model,
|
||||
prompt,
|
||||
stream: false,
|
||||
options: {
|
||||
num_ctx: this.numCtx,
|
||||
...opts.options,
|
||||
},
|
||||
};
|
||||
if (opts.keep_alive !== undefined) body.keep_alive = opts.keep_alive;
|
||||
const res = await this.fetchWithRetry(url, body, timeoutMs);
|
||||
const data = (await res.json()) as {
|
||||
response?: string;
|
||||
done?: boolean;
|
||||
prompt_eval_count?: number;
|
||||
eval_count?: number;
|
||||
};
|
||||
const result: GenerateResult = {
|
||||
text: data.response ?? "",
|
||||
done: data.done ?? true,
|
||||
};
|
||||
if (data.prompt_eval_count !== undefined) result.prompt_eval_count = data.prompt_eval_count;
|
||||
if (data.eval_count !== undefined) result.eval_count = data.eval_count;
|
||||
return result;
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate with an image attachment.
|
||||
*/
|
||||
async generateWithImage(
|
||||
prompt: string,
|
||||
model: string,
|
||||
imagePath: string,
|
||||
opts: GenerateOptions = {},
|
||||
): Promise<GenerateResult> {
|
||||
const imageBytes = await readFile(imagePath);
|
||||
const base64Image = imageBytes.toString("base64");
|
||||
|
||||
const timeoutMs = opts.timeoutMs ?? this.timeoutMs;
|
||||
const url = `${this.baseUrl}/api/generate`;
|
||||
const body: Record<string, unknown> = {
|
||||
model,
|
||||
prompt,
|
||||
images: [base64Image],
|
||||
stream: false,
|
||||
options: {
|
||||
num_ctx: this.numCtx,
|
||||
...opts.options,
|
||||
},
|
||||
};
|
||||
if (opts.keep_alive !== undefined) body.keep_alive = opts.keep_alive;
|
||||
|
||||
const res = await this.fetchWithRetry(url, body, timeoutMs);
|
||||
const data = (await res.json()) as {
|
||||
response?: string;
|
||||
done?: boolean;
|
||||
prompt_eval_count?: number;
|
||||
eval_count?: number;
|
||||
};
|
||||
const result: GenerateResult = {
|
||||
text: data.response ?? "",
|
||||
done: data.done ?? true,
|
||||
};
|
||||
if (data.prompt_eval_count !== undefined) result.prompt_eval_count = data.prompt_eval_count;
|
||||
if (data.eval_count !== undefined) result.eval_count = data.eval_count;
|
||||
return result;
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Chat with messages. Returns full response (stream: false).
|
||||
*/
|
||||
async chat(
|
||||
messages: ChatMessage[],
|
||||
model: string,
|
||||
opts: ChatOptions = {},
|
||||
): Promise<ChatResult> {
|
||||
const timeoutMs = opts.timeoutMs ?? this.timeoutMs;
|
||||
const url = `${this.baseUrl}`;
|
||||
const body: Record<string, unknown> = {
|
||||
model,
|
||||
messages,
|
||||
stream: false,
|
||||
options: {
|
||||
num_ctx: this.numCtx,
|
||||
...opts.options,
|
||||
},
|
||||
};
|
||||
if (opts.keep_alive !== undefined) body.keep_alive = opts.keep_alive;
|
||||
if (opts.tools) body.tools = opts.tools;
|
||||
const res = await this.fetchWithRetry(url, body, timeoutMs);
|
||||
const data = (await res.json()) as {
|
||||
message?: { role: string; content: string; tool_calls?: ToolCall[] };
|
||||
done?: boolean;
|
||||
prompt_eval_count?: number;
|
||||
eval_count?: number;
|
||||
};
|
||||
const result: ChatResult = {
|
||||
message: data.message ?? { role: "assistant", content: "" },
|
||||
done: data.done ?? true,
|
||||
};
|
||||
if (data.prompt_eval_count !== undefined) result.prompt_eval_count = data.prompt_eval_count;
|
||||
if (data.eval_count !== undefined) result.eval_count = data.eval_count;
|
||||
return result;
|
||||
}
|
||||
|
||||
/**
|
||||
* Chat with an image attachment. Reads the image file at `imagePath`, encodes it
|
||||
* as base64, and injects it into the last user message via Ollama's `images` field.
|
||||
* Intended for vision/OCR models such as `glm-ocr:q8_0`.
|
||||
*
|
||||
* @param messages Conversation messages (same format as `chat()`).
|
||||
* @param model Ollama model name supporting vision (must accept `images` field).
|
||||
* @param imagePath Absolute local path to the image file (jpeg, png, webp, etc.).
|
||||
* @param opts Optional chat options (timeout, keep_alive, etc.).
|
||||
*/
|
||||
async chatWithImage(
|
||||
messages: ChatMessage[],
|
||||
model: string,
|
||||
imagePath: string,
|
||||
opts: ChatOptions = {},
|
||||
): Promise<ChatResult> {
|
||||
// Read and encode the image
|
||||
let imageBytes: Buffer;
|
||||
try {
|
||||
imageBytes = await readFile(imagePath);
|
||||
} catch (err) {
|
||||
throw new Error(
|
||||
`OllamaAdapter.chatWithImage: cannot read image at "${imagePath}": ${err instanceof Error ? err.message : String(err)
|
||||
}`,
|
||||
);
|
||||
}
|
||||
const base64Image = imageBytes.toString("base64");
|
||||
|
||||
// Clone messages; inject images into the last user message
|
||||
const messagesWithImage: Array<Record<string, unknown>> = messages.map(
|
||||
(msg, idx) => {
|
||||
const clone: Record<string, unknown> = { ...msg };
|
||||
if (idx === messages.length - 1 && msg.role === "user") {
|
||||
clone.images = [base64Image];
|
||||
}
|
||||
return clone;
|
||||
},
|
||||
);
|
||||
|
||||
// If no user message was found at the end, append one with just the image
|
||||
const lastMsg = messages[messages.length - 1];
|
||||
if (!lastMsg || lastMsg.role !== "user") {
|
||||
messagesWithImage.push({ role: "user", content: "", images: [base64Image] });
|
||||
}
|
||||
|
||||
const timeoutMs = opts.timeoutMs ?? this.timeoutMs;
|
||||
const url = `${this.baseUrl}`;
|
||||
const body: Record<string, unknown> = {
|
||||
model,
|
||||
messages: messagesWithImage,
|
||||
stream: false,
|
||||
options: {
|
||||
num_ctx: this.numCtx,
|
||||
...opts.options,
|
||||
},
|
||||
};
|
||||
if (opts.keep_alive !== undefined) body.keep_alive = opts.keep_alive;
|
||||
|
||||
const res = await this.fetchWithRetry(url, body, timeoutMs);
|
||||
const data = (await res.json()) as {
|
||||
message?: { role: string; content: string; tool_calls?: ToolCall[] };
|
||||
done?: boolean;
|
||||
prompt_eval_count?: number;
|
||||
eval_count?: number;
|
||||
};
|
||||
const result: ChatResult = {
|
||||
message: data.message ?? { role: "assistant", content: "" },
|
||||
done: data.done ?? true,
|
||||
};
|
||||
if (data.prompt_eval_count !== undefined) result.prompt_eval_count = data.prompt_eval_count;
|
||||
if (data.eval_count !== undefined) result.eval_count = data.eval_count;
|
||||
return result;
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate embedding for text. Uses POST /api/embed.
|
||||
*/
|
||||
|
||||
async embed(input: string, model: string, opts: { timeoutMs?: number } = {}): Promise<EmbedResult> {
|
||||
const timeoutMs = opts.timeoutMs ?? this.timeoutMs;
|
||||
const url = `${this.baseUrl}/api/embed`;
|
||||
const body = { model, input };
|
||||
const res = await this.fetchWithRetry(url, body, timeoutMs);
|
||||
const data = (await res.json()) as {
|
||||
embeddings?: number[][];
|
||||
prompt_eval_count?: number;
|
||||
};
|
||||
const embedding = Array.isArray(data.embeddings) && data.embeddings[0] ? data.embeddings[0] : [];
|
||||
const result: EmbedResult = { embedding };
|
||||
if (data.prompt_eval_count !== undefined) result.prompt_eval_count = data.prompt_eval_count;
|
||||
return result;
|
||||
}
|
||||
|
||||
/**
|
||||
* Warm up a model by sending a minimal prompt, ensuring it is loaded into memory.
|
||||
* The keep_alive parameter controls how long the model stays in memory after the call.
|
||||
*/
|
||||
async warmup(model: string, keepAlive: string | number): Promise<void> {
|
||||
const url = `${this.baseUrl}`;
|
||||
const body = {
|
||||
model,
|
||||
messages: [{ role: "user", content: "hello" }],
|
||||
stream: false,
|
||||
keep_alive: keepAlive,
|
||||
};
|
||||
try {
|
||||
await this.fetchWithRetry(url, body, this.timeoutMs);
|
||||
} catch (err) {
|
||||
throw new Error(
|
||||
`OllamaAdapter.warmup failed for model "${model}": ${err instanceof Error ? err.message : String(err)
|
||||
}`,
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Stream chat response. Returns async iterator of chunks (NDJSON).
|
||||
*/
|
||||
async *streamChat(
|
||||
messages: ChatMessage[],
|
||||
model: string,
|
||||
opts: ChatOptions = {},
|
||||
): AsyncGenerator<StreamChunk> {
|
||||
const timeoutMs = opts.timeoutMs ?? this.timeoutMs;
|
||||
const url = `${this.baseUrl}`;
|
||||
const body: Record<string, unknown> = { model, messages, stream: true };
|
||||
if (opts.keep_alive !== undefined) body.keep_alive = opts.keep_alive;
|
||||
if (opts.options !== undefined) body.options = opts.options;
|
||||
const res = await this.fetchWithRetry(url, body, timeoutMs);
|
||||
if (!res.body) return;
|
||||
const reader = res.body.getReader();
|
||||
const decoder = new TextDecoder();
|
||||
let buffer = "";
|
||||
try {
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) break;
|
||||
buffer += decoder.decode(value, { stream: true });
|
||||
const lines = buffer.split("\n");
|
||||
buffer = lines.pop() ?? "";
|
||||
for (const line of lines) {
|
||||
const trimmed = line.trim();
|
||||
if (!trimmed) continue;
|
||||
try {
|
||||
const chunk = JSON.parse(trimmed) as StreamChunk;
|
||||
yield chunk;
|
||||
} catch {
|
||||
// skip malformed line
|
||||
}
|
||||
}
|
||||
}
|
||||
if (buffer.trim()) {
|
||||
try {
|
||||
yield JSON.parse(buffer.trim()) as StreamChunk;
|
||||
} catch {
|
||||
// skip
|
||||
}
|
||||
}
|
||||
} finally {
|
||||
reader.releaseLock();
|
||||
}
|
||||
}
|
||||
|
||||
private async fetchWithRetry(
|
||||
url: string,
|
||||
body: unknown,
|
||||
timeoutMs: number,
|
||||
): Promise<Response> {
|
||||
let lastError: unknown;
|
||||
for (let attempt = 0; attempt <= this.retries; attempt++) {
|
||||
try {
|
||||
const controller = new AbortController();
|
||||
const timeout = setTimeout(() => controller.abort(), timeoutMs);
|
||||
const res = await fetch(url, {
|
||||
method: "POST",
|
||||
headers: { "Content-Type": "application/json" },
|
||||
body: JSON.stringify(body),
|
||||
signal: controller.signal,
|
||||
});
|
||||
clearTimeout(timeout);
|
||||
if (!res.ok) {
|
||||
const text = await res.text();
|
||||
throw new Error(`Ollama ${res.status}: ${text}`);
|
||||
}
|
||||
return res;
|
||||
} catch (err) {
|
||||
lastError = err;
|
||||
const isRetryable =
|
||||
err instanceof Error &&
|
||||
(err.name === "AbortError" ||
|
||||
err.message.includes("fetch") ||
|
||||
err.message.includes("ECONNREFUSED") ||
|
||||
err.message.includes("network") ||
|
||||
err.message.includes("reset") ||
|
||||
err.message.includes("hangup"));
|
||||
|
||||
if (attempt === this.retries || !isRetryable) {
|
||||
if (err instanceof Error && (err as any).cause) {
|
||||
const cause = (err as any).cause;
|
||||
const causeMsg = cause instanceof Error ? cause.message : String(cause);
|
||||
err.message += ` (Cause: ${causeMsg})`;
|
||||
}
|
||||
throw err;
|
||||
}
|
||||
// Wait before retry: 1s, 2s...
|
||||
console.warn(`[ollama-adapter] Fetch failed: ${err.message}. Retrying in ${(attempt + 1)}s... (Attempt ${attempt + 1}/${this.retries})`);
|
||||
await new Promise((resolve) => setTimeout(resolve, (attempt + 1) * 1000));
|
||||
}
|
||||
}
|
||||
throw lastError;
|
||||
}
|
||||
}
|
||||
@@ -1,6 +1,6 @@
|
||||
/**
|
||||
* RAG Service: semantic memory via embeddings and similarity search.
|
||||
* Uses Ollama for embeddings; SQLite for persistent storage.
|
||||
* Uses Lemonade for embeddings; SQLite for persistent storage.
|
||||
* P6-01: SQLite persistence. P7-01: sqlite-vss for scalable KNN with fallback to dot-product.
|
||||
*/
|
||||
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
import { readFileSync, existsSync } from "node:fs";
|
||||
import { readFileSync, existsSync, readdirSync } from "node:fs";
|
||||
import { join, resolve } from "node:path";
|
||||
import { getConfig } from "../shared/config.js";
|
||||
import { TELEGRAM_HTML_FORMAT_INSTRUCTION } from "../agents/prompts/telegram-html.js";
|
||||
@@ -19,17 +19,48 @@ export class SkillManager {
|
||||
}
|
||||
|
||||
/**
|
||||
* List all available skills from CONFIG.md.
|
||||
* List all available skills by scanning the skills directory.
|
||||
* Each skill is a subdirectory containing a SKILL.md file.
|
||||
* The description is extracted from the first line of SKILL.md.
|
||||
*/
|
||||
public listSkills(): SkillInfo[] {
|
||||
const configPath = join(this.skillsDir, "CONFIG.md");
|
||||
if (!existsSync(configPath)) return [];
|
||||
if (!existsSync(this.skillsDir)) return [];
|
||||
|
||||
try {
|
||||
const content = readFileSync(configPath, "utf-8");
|
||||
return this.parseConfig(content);
|
||||
const entries = readdirSync(this.skillsDir, { withFileTypes: true });
|
||||
const skills: SkillInfo[] = [];
|
||||
|
||||
for (const entry of entries) {
|
||||
// Ignore hidden directories and those starting with underscore (disabled)
|
||||
if (entry.isDirectory() && !entry.name.startsWith(".") && !entry.name.startsWith("_")) {
|
||||
const skillMdPath = join(this.skillsDir, entry.name, "SKILL.md");
|
||||
if (existsSync(skillMdPath)) {
|
||||
try {
|
||||
const content = readFileSync(skillMdPath, "utf-8");
|
||||
const lines = content.split("\n").filter(l => l.trim() !== "");
|
||||
const firstLine = lines[0]?.trim();
|
||||
if (!firstLine) continue;
|
||||
|
||||
let description = firstLine;
|
||||
if (firstLine.toLowerCase().startsWith("description:")) {
|
||||
description = firstLine.substring("description:".length).trim();
|
||||
}
|
||||
|
||||
if (description) {
|
||||
skills.push({
|
||||
name: entry.name,
|
||||
description: description
|
||||
});
|
||||
}
|
||||
} catch (err) {
|
||||
console.error(`Failed to read SKILL.md for ${entry.name}:`, err);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
return skills;
|
||||
} catch (err) {
|
||||
console.error("Failed to load skills config:", err);
|
||||
console.error("Failed to list skills by scanning directory:", err);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
@@ -52,41 +83,5 @@ export class SkillManager {
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Simple markdown table/list parser for CONFIG.md.
|
||||
*/
|
||||
private parseConfig(content: string): SkillInfo[] {
|
||||
const lines = content.split("\n");
|
||||
const skills: SkillInfo[] = [];
|
||||
|
||||
for (const line of lines) {
|
||||
// Handle table rows: | name | description |
|
||||
if (line.includes("|")) {
|
||||
const parts = line.split("|").map(p => p.trim()).filter(p => p !== "");
|
||||
if (parts.length >= 2 && parts[0] && parts[1]) {
|
||||
const name = parts[0];
|
||||
const desc = parts[1];
|
||||
if (name.toLowerCase() !== "name" && !name.startsWith("---")) {
|
||||
skills.push({
|
||||
name,
|
||||
description: desc
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
// Handle list items: - name: description or names: description
|
||||
else if (line.trim().startsWith("-") || line.trim().startsWith("*")) {
|
||||
const clean = line.trim().substring(1).trim();
|
||||
const colonIndex = clean.indexOf(":");
|
||||
if (colonIndex !== -1) {
|
||||
const name = clean.substring(0, colonIndex).trim();
|
||||
const description = clean.substring(colonIndex + 1).trim();
|
||||
if (name && description) {
|
||||
skills.push({ name, description });
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
return skills;
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user