AI Native Landscape

Tura

Tracked

An open-source execution layer for AI coding agents that groups environment inspection, edits, builds, and tests into macro commands run in fewer model turns, with published DeepSWE benchmark artifacts.

Author Tura AI Open Sourced 2026-07-06 Last Commit Unknown

Overview

Tura is an open-source agent runtime harness for coding tasks. In a ReAct session the model must re-enter after every tool result, repeatedly carrying the system prompt and a growing context; Tura turns the same task into a runtime-managed command graph so deterministic execution continues without another model round trip. Its published DeepSWE comparison shows higher verifier success with fewer turns and tokens than Codex CLI.

Key Features

  • One macro tool command_run executes a multi-step tree (inspect, patch, build, test, lint) in a single LLM turn where a tool-calling agent needs five
  • Backward reasoning: instead of reasoning from current state to goal, Tura guides the model to estimate the pre-goal state first and reason back, reconstructing failure states before writing code
  • Context managed as runtime state: task status, runtime prompts, and recursive task manuals keep active context scoped rather than piling skills and stale history into one session
  • CLI compaction that preserves exact execution state (code locations, patches, tests) in task_status.compact_context instead of a loose summary
  • Multiple entry points: interactive TUI (tura), direct CLI runner (tura exec), gateway-backed run with streaming/history (tura run), local HTTP/SSE gateway with web GUI, and a desktop GUI workspace

Use Cases

  • Long-horizon debugging and repo-wide rewrite tasks where turn counts and token budgets dominate cost
  • Multi-session concurrent coding work with HTML rich-text support in both TUI and GUI
  • Teams wanting reproducible benchmark evidence — archived prompts, per-round tool calls, token usage, patches, and verifier results per session

Technical Details

  • Rust codebase organized as crates (runtime, router, gateway, session_log, tools) with TUI and GUI apps; npm distribution via tura-ai with no postinstall lifecycle script — the wrapper resolves and launches the platform package directly
  • Published benchmark artifacts cover 20 DeepSWE v1.1 tasks × 3 runs per agent plus rewrite and design tasks: Tura Balanced reached an 80.0% verifier success rate versus 63.3% for Codex CLI (+16.7 pp) with 31.1% fewer tokens; Tura Direct used 77.5% fewer tokens at a comparable 65.0% success rate
  • DeepSWE's own mini-swe-agent results show only an 8% gap between High and Medium reasoning effort, indicating higher reasoning effort alone does not explain the advantage; manifests for all replicates are published
  • Compaction is a CLI operation with explicit compact_context events: benchmark sessions resumed execution an average of 2.6 rounds after compaction (Codex estimated 5.4 from token-usage drops, as it exposes no equivalent events)
  • Provider-agnostic with documented evidence gaps: broader Anthropic/Gemini/local-provider, UI-latency, and cross-OS measurements tracked in the public roadmap and known-issues docs
  • Sessions persisted in a session DB with a runtime/session equivalence gate in tests; first launch requires configuring an LLM provider — no credentials bundled