🔱 TRIDENT CLI

Three Prongs. One Power. All Yours.

TRIDENT is an all-powerful agentic AI coding assistant that runs entirely in your terminal. Inspired by Claude Code, OpenAI Codex CLI, and Qwen Code CLI — built to surpass them all.

---

Architecture: The Three Prongs

🔱 TRIDENT
├── ⚡ FORGE     — Agentic coding engine (tool loop, file ops, shell execution)
├── 🔮 ORACLE    — Project context engine (TRIDENT.md, framework detection, tree scanning)
└── 🛡 WARDEN    — Safety & approval layer (risk classification, diffs, session logging)

---

Quickstart

1. Install

# Clone or copy this project, then:
npm install

# Make globally available
npm link

# Or run directly:
npx tsx src/index.ts

2. Set your API key

export ANTHROPIC_API_KEY=sk-ant-...

3. Run doctor check

trident doctor

4. Initialize your project

cd /your/project
trident init   # Generates TRIDENT.md — the AI memory file

5. Start building

trident                           # Interactive mode
trident "add input validation"    # One-shot task

---

Commands

Command	Description
trident	Interactive REPL mode
trident "task"	One-shot task execution
trident init	Generate TRIDENT.md for current project
trident config	Show full configuration
trident config <key> <value>	Set a config value
trident doctor	Check environment & API keys
trident review	Review last session action log
trident --help	Show all options

Flags

Flag	Description
-m, --model <model>	Override model (e.g. claude-opus-4-5)
--mode <mode>	Approval mode: yolo, review, lockdown
--max-turns <n>	Max agent loop iterations (default: 50)
--budget <usd>	Max spend in USD for this session

Interactive Commands

While in interactive mode:

• exit or quit — Exit TRIDENT
• init — Generate TRIDENT.md
• mode yolo|review|lockdown — Switch approval mode
• model <name> — Switch model

---

Approval Modes

Mode	Description
review (default)	Auto-approve reads; ask for writes, shell, destructive
yolo	Auto-approve everything — fastest, zero interruptions
lockdown	Ask for every single action — maximum safety

Risk Classification

Every tool call is classified before execution:

Risk	Color	Actions
READ	🟢 Green	read_file, list_dir, search, web_fetch
WRITE	🟡 Yellow	write_file, edit_file
EXECUTE	🟣 Magenta	run_command
DESTRUCTIVE	🔴 Red	delete_file

---

TRIDENT.md — Project Memory

TRIDENT reads TRIDENT.md from your project root and injects it into every agent prompt. This gives the AI persistent knowledge about your project.

# TRIDENT Project Context

## Project
- Name, languages, frameworks, package manager

## Commands
- Install, dev, test, build, lint

## Do Not Touch
- Files/dirs TRIDENT should never modify

## Context for AI
- Any conventions, rules, or extra context you want TRIDENT to know

Run trident init to auto-generate it. Edit it freely.

---

Agent Tools

TRIDENT's agent has access to 10 tools:

Tool	Description
read_file	Read file contents
write_file	Write/create a file
edit_file	Surgical string-replace edits
delete_file	Delete a file (DESTRUCTIVE)
list_dir	List directory (with recursive option)
run_command	Execute shell commands
search_codebase	Grep across project files
web_fetch	Fetch URLs (docs, APIs)
ask_user	Ask you a clarifying question
final_answer	Signal task completion

All tools have a 30-second timeout and full session logging.

---

Configuration

trident config                        # Show all config
trident config model claude-sonnet-4-5  # Switch model
trident config mode yolo              # Default to YOLO mode
trident config maxTurns 100           # More iterations

Config stored at ~/.trident-cli/config.json.

---

Session Logs

All actions are logged to ~/.trident/logs/<session-id>.jsonl.

Review the latest session:

trident review

Each log entry contains: timestamp, tool name, input, result, approved/denied, risk level.

---

Models

TRIDENT uses Anthropic's models by default:

Model	Best For
claude-opus-4-5	Complex agentic tasks (default)
claude-sonnet-4-5	Balanced speed/quality
claude-haiku-4-5-20251001	Fast, cheap tasks

---

Project Structure

trident/
├── src/
│   ├── index.ts              ← CLI entry point & interactive REPL
│   ├── config.ts             ← Config management (conf-backed)
│   ├── agent/
│   │   ├── loop.ts           ← Main agent loop (streaming + tool calling)
│   │   └── tools.ts          ← Tool definitions + executors
│   ├── providers/
│   │   └── anthropic.ts      ← Streaming Anthropic provider
│   ├── oracle/
│   │   └── index.ts          ← Project scanner, TRIDENT.md, system prompt
│   ├── ui/
│   │   ├── renderer.ts       ← Terminal UI (chalk-based)
│   │   └── diff.ts           ← Colored diff viewer
│   └── warden/
│       └── index.ts          ← Risk classifier, approval, session logger
├── package.json
├── tsconfig.json
└── README.md

---

Philosophy

1. Think before acting — The agent plans before calling tools
2. Minimal blast radius — Surgical edits preferred over full rewrites
3. Verify work — Tests/linters run after changes when available
4. Ask when uncertain — Clarification over wrong assumptions
5. Full transparency — Every action logged, every risk shown
6. Complete tasks — The agent doesn't stop until it's done

---

Built with ⚡ TypeScript, Anthropic SDK, Commander, Chalk, Inquirer, Execa