Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion docs/changelogs/CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -1,7 +1,8 @@
# Changelog

## 0.5.0 - Unreleased
## 0.5.0 - 2026-10-03

- Verification at release: the full 12-case behavior suite ran on Claude Code (`claude-sonnet-5`) and Codex (`gpt-6-astra`) with independent transcript review, plus targeted re-runs after the final fixes. Known limits: `claude-sonnet-5-5` is not yet qualified for Controlled work; the separate test author is skipped in some Claude Code runs; the guard hook is Claude Code only; on Claude Code, `typo`, `payment`, `model-override`, `qualification`, and `cost-retry` missed parts of the rubric in the last full suite run; the visual gate and workspace marketplace import were not tested.
- Adopted a Helix-informed checkpoint convergence loop for Standard and Controlled work: `plan.md` holds an ordered checkpoint list approved at the plan gate, and each checkpoint passes behavior, visual (UI only), and self-review gates before the next. Plan checkpoints double as the Git commit points in `checkpoints.md`.
- Controlled work adds review checkpoint mode with two context-isolated reviewers, a separate test author when available, and a human stop after checkpoint 1; a recorded `checkpoint` approval covers a stated scope.
- Added a "Reference oracle" field to `spec.md`, the `checkpoint` approval gate, and human feedback in `state.md`. Verify step 5 is now a matched-state visual gate with severity, location, and INVALID recapture. Economy orders gates cheapest first.
Expand Down
Loading