Release v0.39.0

Half the cycle did not believe in agents. Verbaly has told agents to run it since 0.29.0, but only four of the twelve commands ever became tools, and the two it left out are the two an agent meets first: diagnosing a project and wrapping the text already in it. The same look found something worse next door. verbaly translate is the one command that spends money and minutes, and it was a single thread that threw the whole run away on one bad response: three answered batches, sixty translated messages, an exception, and nothing written to disk. No API is removed and every existing tool answers the same text it did before.

Highlights

• verbaly translate no longer loses work it already paid for. A batch the provider does not answer is retried, and if it still fails the run keeps going and writes everything else. Re-running asks only for what is left, instead of paying for all of it again.
• Translating is about four times faster. Batches now go out in parallel instead of one at a time, and a long run tells you what has landed while it runs, rather than staying silent for minutes.
• You can tell the translator what never to translate. Put your product names and required wordings in translate.glossary, and your tone in translate.instructions, instead of fixing the same word by hand after every run.
• Your coding agent gets the two tools it was missing: one to diagnose a project that is not working, and one to find hardcoded text in an existing codebase and wrap it. Onboarding an app to Verbaly no longer needs a terminal.
• Agents read numbers now, not sentences. Every tool answers with structured data as well as text, so a coding agent stops guessing at wording that can change between releases.
• Approving a machine translation is still something only a person can do, and that has not changed.

---

For the full details of this release, see the repository changelog.