Test generation agents

🎉 Major Updates

• Hybrid C-Rust building: as translation progresses with every symbol, the tool now attempts to build a hybrid C-Rust artifact, and execute all tests (if available) using it. If at least one test fails, the translation of the current symbol is considered a failure and re-attempted. This represents a significant upgrade compared to just using cargo build as a correctness signal.
• Library test generation with the Berkeley KISS framework: we dispatch an autonomous RelentlessAgent to analyze a C target, collect input/output pairs, and write assertion tests for it in the Rust C FFI. The output tests are lite-verified for correct file locations, and can be used for the hybrid C-Rust build testing.
• Translating symbol cycles: our tool can now identify and jointly translate symbol cycles with length >1. Post-translation, C FFI wrappers are still generated per-symbol.

🎈 Minor Updates

• Migrated Rust workspace initialization and version control to the fully Python-based ideas.init.workspace module.
• Initial version of an sqlite3-backed symbol cache for resumable translations.
• Made ideas.translate resumable by re-reading the translation cache.
• Added environment variables HYBRID_BUILD and STATIC_TRANSLATIONS for controlling translation strategy and disabling bleeding edge features.

Robust parsing and Dockerfile

🎉 Major Updates

• Significantly improved C AST and Cmake parsing to handle more corner cases
• Added a Dockerfile for reproducible, isolated execution of C-to-Rust translation, in anticipation of AI agents
• Used cargo features to enable conditional compilation of wrappers in translation crate
• Reworked ideas.init to be modular (with two sub-modules), in anticipation of AI agents

🎈 Minor Updates

• Significant clean-up and deprecation of several modules: ideas.cover, ideas.logging

2026.1.1

🎉 Major Updates

• Added Python ideas.cmake module containing Python implementation of C build system instrumentation.
• Re-introduced ideas.ast_rust module for Rust abstract syntax tree (AST) parsing in Python.
• Added AST-based feedback to ideas.wrapper regarding deviations from the prescribed Rust template.

🎈 Minor Updates

• Upgraded dspy to 3.1.2.
• Added automated functionality for extracting and renaming all static global variables in C code, to avoid possible name conflict UB in complex projects.
• Correct usage of dspy.Code field annotations by creating persistent classes.
• Upgraded secondary dependencies.
• Attempting to build consolidated C code from ideas.init and error out if unsuccessful.

2026.1

🎉 Major Updates

• Added ideas.learn.translate GEPA-based automated prompt optimization module. By default, this performs teacher-student learning in two steps:
  • First, a translation dataset is generated using a powerful model, e.g. gpt-5.1.
  • Then, a prompt is optimized for a weaker model, e.g. gpt-5-mini using only the valid (building) translations from the teacher and a powerful GEPA reflection model
  • The metric function includes feedback for building and test correctness.
  • Launch the whole process in a single command using your own OPENROUTER_API_KEY and with:

make -f LEARNING.mk learn/translate

• Added support for git-based version control of the translation repo, in step with iterative AI translation and wrapper generation attempts.

🎈 Minor Updates

• Upgraded clang toolset to clang-21.
• Upgraded to Python 3.14. Don't forget to update your local uv installation!
• Significant code clean-up.

2025.11.1

🎈 Minor Updates

• Resolve source priority paths in ideas.init.
• Changed default model temperature to 0.0.
• Removed slow AST C parsing functionality.
• Add all past translated symbols to the current symbol in ideas.translate_recurrent
• Enable unsafe Rust compilation errors in ideas.translate_recurrent
• Clean-up ideas.utils

🚧 Fixes

• Fix a bug related to enum constant not being captured as standalone symbols
• Fix ideas.wrapper prompt to correctly reference the final on-disk location

2025.11

🎉 Major Updates

• ideas.init module for robust C project discovery (including linked libraries) and idiomatic use of cargo workspace for the translated Rust project. C artifacts are now translated to self-contained Rust crates in the same workspace.
• ideas.translate_recurrent module for symbol-by-symbol C-to-Rust translation using LLMs. The module is powerful enough to work across multi-artifact C projects and topologically sort symbols, assuming no name collisions or circular references exist.
• ideas.wrapper module for LLM-based C FFI wrapper generation for Rust functions. The module uses dspy.ChainOfThought with custom feedback and bindgen signatures derived from the original C function to improve C ABI compatibility of the translated Rust functions.

🎈 Minor Updates

• Expanded Rust AST parsing functionality and added an ensure_no_mangle flag to ideas.cover to deterministically guarantee that translated functions get exported with the correct #[unsafe(no_mangle)] attribute.
• Resolve source priority paths in ideas.init.
• Changed default model temperature to 0.0.
• Removed slow AST C parsing functionality.

---