0.9.7 — Definitions analysis in batches, with partial results

Released 2026-10-21.

The Definitions steps (R-D for requirements and missions, D-D for stories) no longer fail outright on a long text. Before this release a mission statement with about 300 undefined terms sent them all to the AI model in one call; the answer was cut off at the response limit and the whole step was lost.

What's new

Batching. Undefined terms are now drafted in batches of up to 25, one batch after another. The check that your defined terms are used consistently is a separate call. While a run is under way the status line on the Job Queue shows where it has got to, for example R-D: batch 2 of 12.

Partial results. If the AI model's answer to a batch cannot be read, REQQA sends that batch once more, on its own. If it still fails, every other batch's results are kept and the analysis finishes as Partial, with a MEDIUM drafting_failed finding naming the batch and its terms. A failed consistency check is reported as consistency_not_judged. The analysis summary says that some terms were not drafted or judged, and the readiness verdict is never READY for a partial result. After three batches in a row fail, REQQA stops sending further batches and reports the rest as not drafted.

Bug fixes

• Long mission statements no longer lose the whole Definitions step when the AI model's answer is too long to complete.

Behind the scenes

• Two new application settings, DEFINITIONS_BATCH_SIZE (1 to 100, default 25) and DEFINITIONS_FAILURE_LIMIT (at least 1, default 3). A missing or invalid value falls back to the default and logs a warning.
• Mission analysis jobs may now run for up to 90 minutes (previously 30), because a long mission's batches run one after another.
• 108 new automated test scenarios across 6 user stories.

Upgrade notes

No database change. Restart the web service and the background workers after upgrading, because the new settings are read when they start. A re-run of an analysis that was cut off earlier will now complete.