2026.8.13
·
37 commits
to main
since this release
What's Changed
- feat(init): scaffold docs/scope.md for factory-ready packages by @marccampbell in #102
- fix(cli): highest-semver short names + version-alias run by @marccampbell in #101
- feat(cli): adversary train from PR review history by @marccampbell in #103
- feat(cli): direct GitHub API, train without gh, PR URL run, review posting by @marccampbell in #104
- feat(train): bank human gold into agent/voice.md via apply issues by @marccampbell in #105
- feat(init): scaffold agent/voice.md for train voice banking by @marccampbell in #106
- feat(cli): reference voice.md example bank when generating PR comments by @marccampbell in #107
- feat(cli): compose adversaries via adversary.yaml uses by @marccampbell in #108
- fix(train): improve discovery and prevent issue floods by @marccampbell in #110
- chore(deps): bump github.com/santhosh-tekuri/jsonschema/v6 from 6.0.2 to 6.0.3 in the go-dependencies group by @dependabot[bot] in #111
- feat(train): always expand local uses; official jury separate by @marccampbell in #109
- Create issues directly from all-adversary train runs by @marccampbell in #113
- Generate human-written train issue briefs by @marccampbell in #114
- Ground train issue briefs in source evidence by @marccampbell in #115
- Refine source-specific train issue briefs by @marccampbell in #116
- Finalize every train issue brief with an LLM by @marccampbell in #117
- Budget train issue briefs for reasoning models by @marccampbell in #118
- Route training comments by package scope by @marccampbell in #119
- Enforce the training PR limit at result admission by @marccampbell in #120
- Require actionable corroborated training evidence by @marccampbell in #121
- Filter legacy train conversation noise by @marccampbell in #122
- Do not grade failed training reviews by @marccampbell in #123
- Reject automated training praise summaries by @marccampbell in #124
- Let the model judge singleton training evidence by @marccampbell in #125
- Deduplicate training issues by capability by @marccampbell in #126
- Cycle adversary training targets fairly by @marccampbell in #127
- Ignore review template guidance during training by @marccampbell in #128
- Require semantic overlap in training grades by @marccampbell in #129
- Replay cached training evidence on rate limits by @marccampbell in #130
- Prefer latest reviewer concerns in training by @marccampbell in #131
- Make local training execution reliable by @marccampbell in #132
- Preserve admitted training issue intent by @marccampbell in #133
- Exclude post-merge comments from training gold by @marccampbell in #134
- Skip contributor resolution replies in training by @marccampbell in #135
- Match recursive adversary file globs by @marccampbell in #136
- Filter inline children of automated reviews by @marccampbell in #137
- Retain independent training review rounds by @marccampbell in #138
- Cover pull author self-review filtering by @marccampbell in #139
- Keep dismissed train results terminal by @marccampbell in #140
- Bound train catalog discovery windows by @marccampbell in #141
- Let scope model train beyond runtime globs by @marccampbell in #144
- Report unique train result rows by @marccampbell in #146
- Preserve review thread context during training by @marccampbell in #147
- Deduplicate canonical adversary checkouts during training by @marccampbell in #149
- Reject explicitly nonlocal training gold by @marccampbell in #150
- Recognize imperative current review requests by @marccampbell in #151
- Route target-scoped training across all adversaries by @marccampbell in #152
- Match train stories to exact review comments by @marccampbell in #153
- Filter automated review training gold by @marccampbell in #154
- Keep explicit training targets isolated by @marccampbell in #155
Full Changelog: 2026.8.3...2026.8.13
Security
- Pinned scan:
govulncheckv1.6.0 found zero reachable or imported-package vulnerabilities. GO-2026-5970 is present in required modulegolang.org/x/text@v0.14.0, but no imported or called package path reaches it; disposition: not exploitable by this release.