Interactive REPL

The TRI CLI includes a full interactive REPL (Read-Eval-Print Loop) for continuous chat, code generation, and SWE agent operations without restarting the binary.

Starting the REPL

# Via zig build
zig build tri

# Or via the binary directly
./zig-out/bin/tri

When launched without arguments, TRI enters interactive mode:

═══════════════════════════════════════════════════════
  TRI CLI v3.0.0 — Trinity Unified CLI
  100% Local AI | Code | Chat | SWE | Swarm
  phi^2 + 1/phi^2 = 3 = TRINITY
═══════════════════════════════════════════════════════

  Mode: Chat | Language: Zig | Verbose: off
  Type a message or use /command

tri>

REPL Commands

All REPL commands are prefixed with /. Typing text without a prefix sends it as a message in the current mode.

Mode Switching

Command	Mode	Description
/chat	Chat	General-purpose AI conversation
/code	Code	Code generation from natural language prompts
/fix	BugFix	Detect and suggest fixes for bugs
/explain	Explain	Explain code structure or concepts
/test	Test	Generate comprehensive test suites
/doc	Document	Generate API documentation
/refactor	Refactor	Suggest refactoring improvements
/reason	Reason	Chain-of-thought reasoning mode

Example:

tri> /code
  Mode switched to: Code Generation

tri> Write a function that computes Fibonacci numbers
  [AI generates Fibonacci function in current language]

Language Switching

Command	Language	Description
/zig	Zig	Set output language to Zig (default)
/python	Python	Set output language to Python
/rust	Rust	Set output language to Rust
/js	JavaScript	Set output language to JavaScript
/javascript	JavaScript	Alias for /js

The language setting affects code generation, test generation, and documentation output.

Example:

tri> /python
  Language switched to: Python

tri> /code
tri> Implement a binary search tree
  [AI generates Python BST implementation]

Utility Commands

Command	Description
/stats	Display session statistics
/verbose	Toggle verbose output on/off
/help	Show REPL command reference
/quit	Exit the REPL
/exit	Exit the REPL (alias for /quit)
/q	Exit the REPL (alias for /quit)

Session Statistics

The /stats command displays accumulated metrics for the current session:

tri> /stats

Session Statistics
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  SWE Requests:     12
  Chat Queries:     34
  TVC Hits:         8
  TVC Misses:       26
  LLM Calls:        18
  Vision Calls:     2
  STT Calls:        1
  Context Messages:  47
  Summarizations:    3

Tracked metrics:

Metric	Description
SWE Requests	Total fix/explain/test/doc/refactor/reason invocations
Chat Queries	Total chat messages sent
TVC Hits	Responses served from TVC corpus cache
TVC Misses	Responses requiring fresh LLM generation
LLM Calls	Total API calls to language models
Vision Calls	Image analysis requests (via --image)
STT Calls	Speech-to-text invocations (via --voice)
Context Messages	Messages in sliding context window
Summarizations	Context window overflow summarizations

State Management

The REPL maintains a CLIState struct across the session:

Field	Default	Description
Mode	Chat	Current operating mode
Language	Zig	Current output language
Verbose	Off	Detailed output toggle
Stream	Off	Streaming output toggle
TVC Corpus	Loaded	10,000-entry ternary vector cache

TVC Corpus Persistence

The TVC (Ternary Vector Corpus) is loaded at startup and saved on exit:

• File: trinity_chat.tvc
• Capacity: 10,000 entries
• Vector dimension: 1,000 trits
• Threshold: $\phi^{-1} = 0.618$

When you chat, responses are cached in TVC. Subsequent similar queries return cached responses instantly.

Multi-Modal Support

The REPL supports multi-modal input through the chat command flags:

# In REPL mode, use chat with flags:
tri> chat --image photo.jpg "What's in this image?"
tri> chat --voice recording.wav "Transcribe this"
tri> chat --stream "Explain quantum computing"

Flag	Description
--stream	Enable streaming (typing effect) output
--image <path>	Attach image for vision analysis
--voice <path>	Attach audio for speech-to-text

Provider Configuration

The REPL auto-detects available API keys from environment variables:

Variable	Provider	Priority
GROQ_API_KEY	Groq	1 (fastest)
ANTHROPIC_API_KEY	Claude	2
OPENAI_API_KEY	OpenAI	3

If no API key is set, TRI falls back to the local GGUF model:

• Default path: models/tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf

Tips

• Quick mode switch + query: Type /code then your prompt on the next line
• Direct chat: Just type any text without / to send in current mode
• History: The REPL maintains a sliding context window of 20 messages
• Context overflow: When context exceeds the window, older messages are summarized automatically
• Exit: Use /quit, /exit, /q, or Ctrl+C

See Also

• Core Commands — Chat, code, and SWE agent details
• TVC Distributed Learning — TVC corpus architecture
• Sacred Constants — Mathematical constants reference