Unreleased¶
These notes cover changes merged after v0.9.0-alpha.
Added¶
- Per-job Kilo state isolation and optional
kilo run --puresupport, including preflight capability checks. - A 64-job
markdown-it-pycase study in which Loreley found a separately validated 6.75% throughput improvement. - A 64-job
python-pathspeccase study documenting a four-generation 25.14% diagnostic candidate and the allocation-selection failure that invalidated the preregistered outcome. - A fresh Zstandard V19 case study in which the registered winner improved sealed-holdout compression by 1.02% with neutral decompression, plus a Top-10 follow-up that confirmed a generation-4 candidate on a new corpus.
- A three-case-study evidence report that keeps prospective, post-hoc, and supplemental results distinct.
- Phase-specific Kilo planning and coding models, plus explicit provider, thinking, and reasoning controls for trajectory summaries.
- Evaluator-provided candidate identities and exact Git-tree identities for archive deduplication and safe contract-scoped result reuse.
Fixed¶
- Include descendant Kilo sessions and reasoning output in usage accounting, while keeping repeated coding invocations distinct.
- Preserve bounded token/request failure reason codes through agent retries.
- Fetch candidate commits explicitly when narrow clone refspecs omit their branches.
- Terminate the complete Kilo process group after a POSIX timeout so descendant processes cannot continue making API requests after their job fails.
- Read provider-reported Kilo tokens and cost from session-tree aggregates, avoiding descendant double counting and local price reconstruction.
- Keep Kilo workers non-interactive even with native provider configuration by disabling interactive and suggestion tools in the injected headless profile.
- Make campaign constraints override generic planning and coding validation advice, and allow noisy root calibration to retry after cooldown.
- Derive sampling from persistent per-island ordinals and cool down recent
(base, inspirations)recipes so scheduler restarts do not replay a campaign. - Reuse coding or planning summaries for commit messages instead of making a separate commit-summary model request.