$ the-wire · showcase
Repo2RLEnv parsers learn cargo-nextest and Django, Tasksmith ships to PyPI
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
Repo2RLEnv widened the set of test runners its log parsers and verifier actually understand and put Tasksmith on PyPI as a standalone CLI, while downstream repo work in timm and speech-to-speech stayed incremental.
log_parsers: parse cargo-nextest status lines huggingface/Repo2RLEnv
cargo nextest run is now recognized as its own runner in both log_parsers and the standalone verifier, with status lines parsed into binary-id test::name -> status. Retries resolve to the last attempt (TRY n FAIL followed by FLAKY 2/3 ends as passed), in-progress SLOW [> ...] lines are ignored, and indented libtest output under a failure is no longer misread as a result.
log_parsers: read unittest and Django test-runner output huggingface/Repo2RLEnv
Verbose unittest and Django TextTestRunner output is now parsed in the generation parser and the standalone verifier, covering python -m unittest, manage.py test and runtests.py, with the verbosity each runner needs added automatically. Subtests group under their parent with the worst status winning, and unittest count footers no longer leak into the pytest fallback.
feat: publish Tasksmith as a PyPI CLI package huggingface/Repo2RLEnv
pip install tasksmith now installs Repo2RLEnv with Tasksmith, Daytona and Harbor support, and tasksmith run ... forwards to the existing repo2rlenv tasksmith run ... implementation, with Modal as an optional extra. The launcher carries its own 0.0.1 version and requires Repo2RLEnv >=0.9.3,<0.10; the root package version and release workflow are untouched.
cli: load .env from the working directory huggingface/Repo2RLEnv
_load_dotenv_if_present() now calls find_dotenv(usecwd=True), so a .env in the working directory is loaded under uv tool install, pipx, or any venv outside the project, where the old bare load_dotenv() searched upward from the installed cli.py and never found it. Existing environment variables still win and PYTHON_DOTENV_DISABLED still turns loading off.
Limit generated Realtime function call IDs to 32 characters huggingface/speech-to-speech
Generated Realtime function call IDs were 37 characters, which LiveKit Agents 1.8.5 truncates to 32, so tool output no longer matched the pending call and the agent went silent after executing the tool. The shared helper now emits call_ plus 27 UUID hex characters, keeping IDs within 32 across local, Chat Completions, and Responses backends.
Record Kokoro TTS first-audio latency on the turn tracker huggingface/speech-to-speech
Kokoro TTS streamed audio but never published first-audio timings, so a turn could report a populated e2e while tts_ttfa stayed n/a. First-provider and first-yielded audio are now recorded on the existing turn tracker for both the native and MLX paths, with the merge in #699 and a follow-up fixing the latency input types and regression test stubs.
