agentic-tool-calls   node: .json   core: onnx

Our vision is to transform Fetch into a highly responsive, zero-latency, cross-platform stationary software robot agent. Rather than memorizing complex CLI flags, users can naturally command their system using direct slash-commands, semantic text, or local voice triggers.

Below is our phased engineering plan:

• Unified Schema Registry: Consolidate ai-context.md and standard prompts into a single, machine-readable tools.json. Each tool will register its standard OpenAPI function schema (for the LLM) alongside a private execution template (for local terminal execution).
• Dual-Engine Triage Router: Implement a hybrid execution path in ai-agent.py:
  • Deterministic Jaccard Matching: Instantly intercept slash-commands and exact keyword shortcuts at <0.1ms latency.
  • Cactus Needle (26M SAN): Route natural, complex phrasing on-device at ~5ms latency using a quantized, local 14MB ONNX runtime.
• Fail-Safe Conversational Fallback: Establish a clean delegation pattern. When local routing returns no matched tools, seamlessly forward the context to your conversational baseline (gemini-3.1-flash-lite or Qwen-35B).

• Converse-to-Command Datasets: Synthesize thousands of diverse, natural-language training examples (utilizing frontier models as teachers) matching the specific administration tools of Fetch.
• Needle Gating Fine-Tuning: Fine-tune Cactus Needle's 26M attention weights locally (via CLI/Playground). Train the model to naturally isolate trigger patterns (e.g., "Fetch, scan me drive" vs "Hey Fetch, can you...") and reliably populate JSON argument structures.
• Robustness Evaluation: Build local evaluation validation datasets to verify routing accuracy, eliminating false positive command executions.

• Cross-Platform Script Abstraction: Abstract system administration bash scripts into OS-aware execution pipelines, preparing Fetch to run natively on Linux, macOS, and eventually Windows.
• Standalone Binary Compilations: Package the entire Python environment, dependencies, and the quantized 14MB routing model into a single, zero-dependency native binary executable.

• Offline Speech-to-Text (STT): Integrate a lightweight local speech transcriber.
• Ultra-Fast Text-to-Speech (TTS): Hook up highly efficient local speech synthesis (like koko) to allow the agent to talk back with negligible audio generation latency.
• Voice Trigger Hook: Establish an ambient system-level hotkey or wake-word listener to summon Fetch instantly without terminal typing.

• Abstract Syntax Tree (AST) Verification: Run local background compilers in Python to verify code structures generated by smaller models (like Qwen-2B) before outputting to the console.
• Auto-Repair Turn: If a syntax or indentation error is detected, automatically run a silent background patch turn to correct the code block dynamically.

All configurations and custom shortcuts are managed in ai-context.md.

• Direct (No Session): Sub-millisecond Jaccard matching (jaccard_search) instantly routes custom keywords to your local terminal.
• Single-Turn Agent (ai <query>): Returns a single response directly to your shell prompt without loading an active conversation.
• Multi-Turn Chat (ai alone): Starts a persistent terminal session with multi-turn context tracking.
• Workspace Agents (ai init <path>): Indexes your directory into a lightweight codebase graph and boots up a codebase-aware chat.

Click to expand ecosystem diagram

╭──────────────────────────────────────────────╮
│  >_ Fetch Robotics                           │
│                                              │
│  model:     Qwen3.6-35B-A3B.gguf             │
│  directory: ...-ai/projects/session-test     │
│  skill:     init codef                       │
│  database:  active (3 facts, 109 turns)      │
╰──────────────────────────────────────────────╯
[sys] Startup context: 210 tokens | Ctrl+C to exit.

Agent: Workspace loaded. Awaiting instructions.
❯

Evolving with your workspace, learning your habits, and standardizing your identity.

• Weaviate Engram's active reconciliation concepts with Noema's local Markdown file system.

Building semantic codebase maps and queryable relational graphs.

• Graphify's codebase mapping and codebase-memory-mcp's relational queries, supercharged with local semantic vector search via sqlite-vec.

Inspecting package updates, monitoring system health, and optimizing performance.

• log-checker and system-health live diagnostics with aur-audit, security-audit, update-inspector zero-trust auditing, system-optimizer resource adjustments, ai-status routing, and ai-commit hooks.

• Up / Down Arrow Keys: Cycle through available ranked selections.
• Enter: Execute the highlighted command (or initialize a workspace if the selection is a directory path).
• Esc / Right Arrow / Ctrl+C: Cancel/Skip the active menu, memory-recall, or tool authorization prompt cleanly.

~ ❯ weather
[01/02] ❯ [weather full] curl -s wttr.in | cat
:: ↵ run  Esc:

Manage your active cloud endpoints, inspect live API rankings, and toggle keys.

• Run model select directly from your terminal to launch the interactive Cloud Connection TUI.

Executed directly from your terminal prompt.

Typed directly inside an active chat session.

Typed inside an active chat session to adjust settings.

Add your shortcuts, commands, and workspaces to ai-context.md.

# --- Weather & Live Networking ---
[TOOL] curl -s wttr.in --cat ---> weather full, wttr, weather
[TOOL] curl -s "wttr.in/?format=3" --cat ---> weather simple, wttr, weather

# --- Fetch Agent Blueprint (CheatSheet) ---
~/.config/fetch/tools/blueprint --leaf ---> cheatsheet, blueprint, bp, cs

# 1. Optional: Install terminal rendering utilities
# (mdcat enables beautiful terminal markdown formatting)
yay -S mdcat

# 2. Install required system dependencies (Reduces latency)
# Debian/Ubuntu: sudo apt install python3-requests
# macOS / Other: pip install requests
sudo pacman -S python-requests

# 2.5 Optional: Install local vector-database extensions
# (Enables high-performance semantic search over your codebase)
# Debian/Ubuntu: pip install sqlite-vec --break-system-packages
# macOS / Other: pip install sqlite-vec
yay -S python-sqlite-vec

# 3. Clone the repository locally
git clone https://github.com/j5onrf/fetch.git ~/.config/fetch

# 4. Add the environment hook into Bash & reload your profile
echo '[ -f "$HOME/.config/fetch/ai-hook.sh" ] && source "$HOME/.config/fetch/ai-hook.sh"' >> ~/.bashrc
source ~/.bashrc

# 5. Create your private configuration file (No global exports needed!)
# Fill in only what you use; the rest defaults safely.
# The agent reads this dynamically on every run with zero terminal restarts.
nano ~/.config/fetch/.env

# ~/.config/fetch/.env
# use "ai status" and "model select"

# Claude API
CLAUDE_API_KEY="your-claude-api-key-here"
CLAUDE_MODEL="claude-fable-5"

# OpenAI API
OPENAI_API_KEY="your-openai-api-key-here"
OPENAI_MODEL="gpt-5.6"

# x.AI Grok API
XAI_API_KEY="xai-your-grok-api-key-here"
XAI_MODEL="grok-4.5"

# Google Gemini API
GEMINI_API_KEY="AIzaSyYourFullGeminiApiKeyHere"
GEMINI_MODEL="gemini-3.1-flash-lite"

# OpenRouter API
OPENROUTER_API_KEY="sk-or-v1-YourFullOpenRouterKeyHere"
OPENROUTER_MODEL="openrouter/free"

# Context Limits
AI_MAX_TOKENS=8192

• Origin: Based on the foundational Local-Ai Agent framework.
• This application incorporates the Cactus Needle routing model developed by Cactus Compute, which is licensed under the MIT License: Copyright (c) 2026 Cactus Compute (Standard MIT License)