self-reference

Files (1)

• SKILL.md

SKILL.md (first lines)

# Self-Reference

## Overview

This skill is a **self-knowledge interface** for bibo. When invoked, it tells you everything you need to know about yourself without having to parse AGENTS.md or SELF.md manually.

**Purpose**: Instead of reading through AGENTS.md to discover how to modify your extensions, skills, settings, or dashboard, this skill provides direct answers about your own architecture.

**When to use**:
- You need to modify an extension, skill, setting, or dashboard
- You need to understand your own architecture before making changes
- You need to know how to add a new tool, command, extension, or skill
- You want a quick reference for your own structure without reading full docs

## Parameters

- **query** (optional): Specific question about yourself. If omitted, provide a full self-knowledge index.
  - Examples: "How do I add a new extension?", "What are all my extensions?", "How does the dashboard work?", "What are my current model profiles?"
- **depth** (optional, default: `overview`): Level of detail for the response.
  - `overview`: High-level summary of the requested topic
  - `detailed`: Full details with file paths, code snippets, and examples
  - `complete`: Everything — full architecture, all extensions, all skills, all settings

## Steps

### 1. Parse the query

Determine what the user is asking about:
- **Extensions**: How to add/modify, list of all extensions, specific extension details
- **Skills**: How to add/modify, list of all skills, specific skill details
- **Settings**: Model profiles, compaction, retry, environment variables
- **Dashboard**: Architecture, API endpoints, frontend structure
- **Testing**: How to run tests, test coverage
- **System prompt**: How the system prompt is assembled, skill injection, knowledge injection
- **Quality mechanisms**: Quality monitor, output parser, browser extract retention, evidence preservation
- **General**: Full self-knowledge index

### 2. Retrieve self-knowledge

You already have this information in your context. Use the following structure:

#### Project Metadata
- Name: `bibo`, Version: `0.1.24`
- Description: A pi-based coding agent optimized for small local language models
- Dependency: `@mariozechner/pi-coding-agent` ^0.68.1
- Provider: `jackbox` (llama.cpp at `http://127.0.0.1:6969/v1`)
- Model: `bibo-qwen3.6` (128K context, 4096 max tokens, reasoning enabled)

#### Directory Structure
```
bibo/
├── AGENTS.md           # Project instructions (loaded as context)
├── SYSTEM.md           # System prompt override
├── SELF.md             # Internal architecture reference (read before modifying yourself)
├── package.json        # Dependencies + scripts
├── provider.json       # Provider config (local, not tracked)
├── provider.json.example
├── models.json         # Model definitions
├── tsconfig.json       # TypeScript config
├── vitest.config.ts    # Test config
├── .pi/
│   ├── settings.json   # Model profiles, compaction, retry
│   └── extensions/     # 21 TypeScript extension modules
│       ├── shared-state.ts  # Cross-extension state
│       ├── benchmark-profiles/
│       ├── tool-gating/
│       ├── permission-gate/
│       ├── checkpoint/
│       ├── skill-inject/
│       ├── knowledge-inject/
│       ├── quality-monitor/
│       ├── output-parser/
│       ├── thinking-budget/
│       ├── evidence/
│       ├── evidence-compact/
│       ├── browser-extract-retention/
│       ├── browser/
│       ├── hello/
│       ├── timer/
│       ├── shell-session/
│       ├── write-guard/
│       ├── context-mode/     # Multi-agent adapter (largest)
│       ├── llama-cpp-provider/
│       ├── extra-tools/
│       ├── breathing-border/
│       └── turn-cap/
├── skills/
│   ├── tools/          # 15 tool guidance files
│   ├── knowledge/      # 13 algorithm cheat sheets
│   └── protocols/      # 3 workflow protocols
├── dashboard/
│   ├── src/            # Node.js server + REST API
│   └── src-frontend/   # React 18 + Vite + TypeScript
├── bin/                # CLI scripts
├── sessions/           # Session data
└── node_modules/       # Dependencies
```

#### Extension Stack (21 extensions, in load order)

| # | Extension | Purpose | Key Events |
|---|-----------|---------|------------|
| 1 | **hello** | Startup notification | `session_start` |
| 2 | **timer** | Session/task duration tracking | `session_start`, `session_shutdown` |
| 3 | **benchmark-profiles** | Resolves model profiles, sets temperature=0.3 | `before_agent_start`, `before_provider_request` |
| 4 | **tool-gating** | Blocks tools not in whitelist | `before_agent_start`, `tool_call` |
| 5 | **permission-gate** | Whitelists safe bash commands | `tool_call` |
| 6 | **checkpoint** | Backs up files before Write/Edit | `session_start`, `tool_call` |
| 7 | **skill-inject** | Appends tool skill guidance (300-token budget) | Per-turn injection |
| 8 | **knowledge-inject** | Appends algorithm cheat sheets (200-token budget) | Per-turn injection |
| 9 | **quality-monitor** | Detects 4 failure modes at `turn_end` | `turn_end` |
| 10 | **output-parser** | Detects text-embedded tool calls | Per-turn |
| 11 | **thinking-budget** | Aborts when thinking tokens exceed budget | Per-turn |
| 12 | **evidence** | Per-session in-memory evidence store | Throughout session |
| 13 | **evidence-compact** | Preserves evidence after compaction | `compaction` |
| 14 | **browser-extract-retention** | Prunes old BrowserExtract results (keeps 2 newest) | Per-turn |
| 15 | **browser** | Per-session Playwright browser instance | `session_start`, `session_shutdown` |
| 16 | **shell-session** | Safe shell session management | Throughout session |
| 17 | **write-guard** | Write protection | `tool_call` |

From exoad/bibo-agent, Apache-2.0.

More from exoad/bibo-agent