You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Documentation of the GitHub-sourced plugin install in a clean HOME, including enabled-plugin output, the installed skill path, and the first-push CI result.
Review tiers: a task is light only when every path in its Files list and actual diff is docs, tests or fixtures; any other path, and any instruction file (SKILL.md, AGENTS.md, CLAUDE.md, or files under skills/, commands/, agents/, prompts/, hooks/, .claude/, .github/), makes it full. Light tier skips only the Challenge; review authority, report compliance and lessons run on every tier. A plan may raise a tier, never lower it.
Glossary rows for independent reviewer, owner, correction attempt and review tier.
Changed
Breaking: fix rounds (up to five, the last two on a fresh implementer) are replaced by at most two correction attempts per task lineage, persisted across restarts and model or producer changes. Exhausting them escalates to the owner instead of adjudicating findings.
An independent reviewer, never chosen by the implementer, owns exact-revision acceptance. Rulings cannot override it, and acceptance grants no release, spending or disclosure authority.
The controller records each task's base commit in the ledger before the first dispatch and reviews <base>..<candidate>, so multi-commit tasks and resumed runs are reviewed whole.
The controller binds each review to git rev-parse HEAD, requires it to match the implementer's reported commit before review and before check-off, and ledgers the accepted SHA.
Implementers and reviewers write full reports to files and return at most 10 lines. The controller works from summaries and paths, batches ledger, lessons and checkbox edits into one message per task, and emits a resume line every 5 accepted tasks. Measured 2026-11-04: the controller session was 25 to 55% of each task's tokens.
Report-compliance gaps (missing Risk, Sources, Mocks or Verification) are Minor unless the reviewer shows the defect they hide. Minors never start a correction attempt.
Launch materials now record blog approval, the site branch and URL, and publication on 2026-11-03; the Show HN draft links to the published write-up.
The todo-api example notes that it ran under 1.0.0.