Skip to content

TOOL_MODEL_MATRIX.md

Latest commit

 

History

History
34 lines (26 loc) · 2.26 KB

TOOL_MODEL_MATRIX.md

File metadata and controls

34 lines (26 loc) · 2.26 KB
Raw

Tool & Model Mapping Matrix

This document defines the exact technical stack used for each agent role. To maintain "Model Diversity" and avoid systemic blind spots, different tools and models are used across the pipeline.

Agent Stack Mapping

Agent Role Tool/CLI Primary Model Fallback/Escalation Key Purpose
Coordinator Hermes Agent algolia/xlarge claude-sonnet-4-6 Orchestration, Triage, Merging
Coder claude CLI algolia/xlarge claude-sonnet-4-6 Feature Implementation
QA opencode CLI algolia/medium algolia/xlarge Spec Verification, Test Execution
Reviewer codex CLI gpt-5.6-sol claude-sonnet-4-6 Security & Architecture Audit
Tech Docs claude CLI claude-sonnet-4-6 algolia/xlarge Structured Markdown, Architecture Docs
Marketing Site claude CLI claude-sonnet-4-6 algolia/xlarge HTML/CSS, Visual Design, Copy

Technical Implementation Details

1. The Proxy Layer (LiteLLM)

All CLI tools are routed through the LiteLLM proxy at 10.1.2.13:4000. This allows the system to:

  • Mimic APIs: The claude CLI believes it is talking to Anthropic, but LiteLLM routes it to a local algolia model.
  • Transparent Fallback: If a primary model is offline, LiteLLM can route to a fallback without the agent crashing.

2. Execution Patterns

  • The Coder (claude): Launched with --dangerously-skip-permissions and ANTHROPIC_BASE_URL pointing to the proxy.
  • The QA (opencode): Launched with -m litellm/<model> to specify the verification model.
  • The Reviewer (codex): Launched with --skip-git-repo-check for deep analysis of specific diffs.

Why Diversify Tools?

We use a "Multi-Tool Strategy" to prevent a single-point-of-failure in reasoning:

  • Coder (Claude CLI) is optimized for generating working code.
  • QA (OpenCode) is optimized for proving that code fails or passes.
  • Reviewer (Codex) is optimized for auditing code for non-functional flaws (security, leaks, complexity).

By using different tools, we ensure a "checks and balances" system where the agent that wrote the code is not the one verifying it, and the tool verifying it isn't the same one that reviewed it.