Show HN: OSS Agent I built topped the TerminalBench on Gemini-3-flash-preview

TL;DR Highlight

Dirac cuts API costs 64.8% and achieves 65.2% on TerminalBench-2 with efficient context management.

Who Should Read

Developers burdened by API costs when using AI coding agents like Claude Code, Cline, and Aider, or those looking to integrate agents into large-scale codebase refactoring projects.

Core Mechanics

Dirac’s core philosophy stems from the well-known phenomenon that model inference ability degrades as context length increases. Maintaining a tight context improves both accuracy and cost.
Optimizing Hash-Anchored Edits—a method of fixing code positions with hashes before modifying them—significantly reduces token waste during file editing. Unlike agents that read and write entire files, Dirac precisely targets only the necessary changes.
By analyzing the Abstract Syntax Tree (AST), the model only includes the code snippets it actually needs in the context. Instead of reading the entire large file, it selectively retrieves only the required functions or classes.
Dirac processes large read/write operations in parallel. Unlike other agents that process tasks sequentially, it executes multiple file edits in batches, increasing speed and efficiency.
The model can directly write and execute bash/python/perl scripts, then analyze the results. This dynamic information gathering contrasts with statically reading files.
Dirac employs an 'opportunistic context update' strategy. It proactively populates the context with information the model is likely to request next, preventing unnecessary additional API calls.
Using Gemini-3-flash-preview, Dirac scored 65.2% on the TerminalBench-2 leaderboard, achieving first place. This is 17 percentage points higher than Google’s official result, demonstrating that agent harness quality significantly impacts performance, even with the same model.
Dirac does not use Model Context Protocol (MCP). While forked from Cline, it has evolved its own architecture and is released as open-source under the Apache 2.0 license.

Evidence

"The most resonant comment in discussions is that the harness has a greater impact on performance than the model itself. One comment noted that changing the harness had a larger impact on benchmark scores than switching from Gemini to Sonnet, and many developers agreed. A user shared their experience refactoring a Rust codebase using Kimi 2.6 and Dirac, finding it more productive than OpenCode, which corrupted .rs files. Concerns were raised about telemetry, with a user discovering Dirac sending data to dirac.run/v1/event, potentially including sensitive API error content. The opt-out mechanism was criticized as untrustworthy. Some argued that context management is a temporary fix for current model limitations and may become obsolete with future generations, like RAG. The benchmark was limited to Gemini 3 Flash, raising concerns about overfitting and the need for validation on other model families (e.g., Minimax 2.7)."

How to Apply

"If you’re experiencing excessive API costs with Cline or Aider, replace them with Dirac and compare costs and results. The claimed 64.8% cost reduction can be verified in your workflow. Dirac is well-suited for large-scale Rust/TypeScript codebase refactoring tasks, where context limits are easily reached, thanks to its AST-based context selection. If you use a company LLM proxy, a user reported successful connection with a few API/config file modifications. If telemetry is a concern, verify and disable the opt-out setting before use, especially in sensitive codebases where API errors could expose information."

Terminology

Hash-Anchored EditsA method of fixing the position of a line in a file by assigning a hash value, then precisely modifying only that position. This reduces costs by transmitting only the changed portions as tokens instead of rewriting the entire file.

ASTAbbreviation for Abstract Syntax Tree, a tree-like representation of source code parsed into structural units such as functions, classes, and variables. This allows for the extraction of only the necessary code blocks instead of the entire file.

HarnessAn external framework for running AI agents, encompassing prompt construction, tool calling, and context management. Performance can vary significantly depending on the harness, even with the same model.

TerminalBenchA benchmark that measures how well AI agents perform actual, complex development tasks in a terminal environment. It evaluates performance based on real repository modification tasks, not just simple code generation.

TelemetryThe automatic collection of usage data (error logs, feature usage patterns, etc.) by software and transmission to the developer's server. Opt-out telemetry requires users to disable it manually and may include sensitive information.

Opportunistic Context UpdateA strategy of proactively populating the context with information the model is likely to request next. This reduces costs and latency by preventing additional API calls.