Problem
Currently, the Gemma model's smaller context window limits its efficiency compared to Voyage models. This restricted context makes LLM recall inefficient, particularly for coding tasks.
Furthermore, the system currently lacks a "walking mode" for coding agents. When working on codebases, simple summary recall is insufficient; an agent needs the ability to navigate file-to-file to build a complete, project-scoped understanding of the architecture.
Proposed solution
-
Model Optionality: Introduce a configuration to optionally use Voyage-4-Nano (which supports a 32,000-token context window) instead of Gemma. This will significantly improve context retention and retrieval efficiency.
-
"Brain Mode" / Walking Mode for Coding: Implement a project-scoped context mechanism that allows the coding agent to walk file-to-file. Rather than relying purely on fragmented summaries, this mode would allow the agent to crawl directories, read full file contents in sequence, and maintain a persistent "mental map" of the specific project it is working on.
Alternatives considered
Continuing to rely on Gemma by implementing more aggressive chunking and summarization, though this often strips away the precise syntactical details required for accurate code generation and debugging.
Integrating other large-context models, though Voyage-4-Nano represents an ideal balance of retrieval quality, local development feasibility, and context size.
Additional context
Coding tasks inherently require deep, multi-file context. Enabling a project-scoped brain paired with Voyage's larger context window will drastically reduce hallucinations and improve the agent's ability to refactor or debug cross-file dependencies.
Problem
Currently, the Gemma model's smaller context window limits its efficiency compared to Voyage models. This restricted context makes LLM recall inefficient, particularly for coding tasks.
Furthermore, the system currently lacks a "walking mode" for coding agents. When working on codebases, simple summary recall is insufficient; an agent needs the ability to navigate file-to-file to build a complete, project-scoped understanding of the architecture.
Proposed solution
Model Optionality: Introduce a configuration to optionally use Voyage-4-Nano (which supports a 32,000-token context window) instead of Gemma. This will significantly improve context retention and retrieval efficiency.
"Brain Mode" / Walking Mode for Coding: Implement a project-scoped context mechanism that allows the coding agent to walk file-to-file. Rather than relying purely on fragmented summaries, this mode would allow the agent to crawl directories, read full file contents in sequence, and maintain a persistent "mental map" of the specific project it is working on.
Alternatives considered
Continuing to rely on Gemma by implementing more aggressive chunking and summarization, though this often strips away the precise syntactical details required for accurate code generation and debugging.
Integrating other large-context models, though Voyage-4-Nano represents an ideal balance of retrieval quality, local development feasibility, and context size.
Additional context
Coding tasks inherently require deep, multi-file context. Enabling a project-scoped brain paired with Voyage's larger context window will drastically reduce hallucinations and improve the agent's ability to refactor or debug cross-file dependencies.