Skip to content

release: 0.6.0 behavior-suite results and plan-only checkpoint fix #7

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We鈥檒l occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
sofus-common merged 2 commits into from
Oct 10, 2026
Merged

release: 0.6.0 behavior-suite results and plan-only checkpoint fix #7

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
580b6c7
fix: name the checkpoint-1 approval in plan-only Controlled drafts
sofus-common Oct 10, 2026
5caac69
release: date 0.6.0 changelog entry with behavior-suite results
sofus-common Oct 10, 2026
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
8 changes: 6 additions & 2 deletions docs/changelogs/CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,13 +1,17 @@
# Changelog

## 0.6.0 - Unreleased
## 0.6.0 - 2026-10-10

- Verification: the full 12-case behavior suite ran with independent transcript review on Claude Code 2.1.296 with `claude-sonnet-5-5` and Codex 0.161.0 with `gpt-6.1-sol` (reasoning effort low). Results were 11/12 and 12/12. All filesystem and executable acceptance checks passed. `claude-sonnet-5` was not re-run.
- Claude Code `payment` failed: the plan-only Controlled draft did not name the human approval needed after checkpoint 1. In a paired run on `claude-sonnet-5-5` the gap appeared in 1 of 3 runs on 0.5.1 and in 3 or 4 of 4 runs on 0.6.0 before the fix below, and in 1 or 2 of 4 runs after it. Small samples: a direction, not a rate. Codex named the approval in 2 of 2 runs after the fix.
- Run limits: both hosts ran with permission prompts and sandboxes off in disposable fixture folders, because the Windows Codex sandbox blocked or hung every command. Codex ran from an isolated `CODEX_HOME` with the plugin installed from GitHub; user-level skills in `~/.agents/skills` stayed discoverable but no scored session invoked them. The user's global CLAUDE.md stayed loaded on Claude Code.
- Observed: build and verify reports listed met and unmet counts (for example "4 met, 0 unmet"), and no changed file had placeholders. Met/unmet marking was uneven in Fast and read-only cases, and in two cases the final report did not re-measure figures it quoted from fixtures or workers.
- Fixed: the router's plan-only clause now lists the Controlled approval after checkpoint 1 among the approvals still needed, because `ai-native-plan`, which states that rule, was skipped in the failing run.
- Added completion rules informed by unlazy (MIT License). Wording only; no scripts or hooks bundled. See ADR-005.
- Router invariant: finish every in-scope part of the request, with no placeholders, stubs, `TODO`s, elided code, or silently deferred remainder. Minimal means the least code per part, not fewer parts. A part that cannot be finished is reported as unmet with its reason.
- Router output: mark each requested part met or unmet, and re-measure every reported number or label it `not measured`.
- Build: each checkpoint is implemented completely, and self-review also hunts defects, placeholders, and unhandled cases.
- Verify: step 8 re-reads the original request and maps every requested part to a supported claim or a reported unmet item; the evidence report ends with met and unmet counts.
- Not yet verified: the 12-case behavior suite has not run on this release.

## 0.5.1 - 2026-10-08

Expand Down
2 changes: 1 addition & 1 deletion skills/ai-native-sdlc/SKILL.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,7 @@ When a leaf skill opened this file, use only its lane and invariants. Do not rou
5. If `.sdlc/changes/*/state.md` identifies this work, resume its `next_action`; do not restart the lifecycle. For model selection, model qualification claims, or cost comparisons, read [model-selection.md](references/model-selection.md) before answering, including read-only and Fast work.
6. For Fast work, stay inline: state the short plan, make the smallest authorized edit, inspect the diff, run focused proof, and report fresh evidence. Use `ai-native-debug` first only when the cause is unknown. Do not create lifecycle files.
7. For Standard or Controlled work, invoke only the next matching skill through the host's skill mechanism; never do a leaf skill's work from this router:
- unclear outcome, new product behavior, no `spec.md` yet, or Controlled work without an approved `intent.md` -> `ai-native-shape`; a plan-only request still gets draft `spec.md` and `plan.md`, with nothing marked accepted and the approvals still needed listed
- unclear outcome, new product behavior, no `spec.md` yet, or Controlled work without an approved `intent.md` -> `ai-native-shape`; a plan-only request still gets draft `spec.md` and `plan.md`, with nothing marked accepted and the approvals still needed listed, including the Controlled approval after checkpoint 1
- accepted outcome with `spec.md` needing implementation design -> `ai-native-plan`
- accepted plan -> `ai-native-build`
- defect or failed check with unknown cause -> `ai-native-debug`
Expand Down
Loading