Skip to content

fix: review large changes in task units and find existing code by responsibility - #107

Merged
shinpr merged 2 commits into
mainfrom
fix/review-units
Nov 11, 2026
Merged

shinpr merged 2 commits into
mainfrom
fix/review-units

Conversation

@shinpr

@shinpr shinpr commented Nov 11, 2026

Copy link
Copy Markdown
Owner

Summary

  • Review Medium and Large changes in units of completed Work Plan tasks. A unit is a consuming task (a screen, route, command, or job) together with the tasks it depends on through Depends on; tasks that several units depend on form one shared unit. One code-reviewer judges each unit with the unit's tasks as task scope, so a large change no longer has one reviewer reading every file. security-reviewer still reviews the complete change once.
  • Reruns: the reviewer that reported a finding reruns it. After each correction, the orchestrator compares the diff with the scope of every other passed review (a code-reviewer's implementation files, applicable ACs, and contracts; security-reviewer's trust boundaries) and reruns only a reviewer whose scope the diff changes.
  • code-reviewer judges each applicable AC and material surface against unordered criteria instead of building a whole-change map, and no longer returns the acceptanceCriteria inventory, which no consumer read. Each unfulfilled AC still has one finding. The rerun boundary is derived from the correction diff.
  • Existing code is found by responsibility as well as by name: code with the same responsibility may share no name or reference with the change, including code added earlier in the same change. A search with no match shows only that the search found nothing. Code stays separate only when evidence shows different contracts or change reasons. The definition lives in coding-rules; codebase-analyzer carries it directly because it does not load coding-rules.
  • Task executors return completionCriteria (evidence per Completion Criteria item), which the orchestrator uses to mark completion, and findingsOutsideChange (problems observed but left outside the change, with no new search obligation). Unhandled items appear in the completion report.
  • Task files list a test file under Target Files only when the Work Plan fixes it, such as a generated skeleton; the executor selects other tests.
  • Code comments do not reference tasks, plans, or document sections.
  • recipe-front-adjust loads coding-rules, ai-development-guide, and implementation-approach before the parent session edits code, and testing before tests change.
  • Bump version to 1.8.8.

Validation

  • npm test: 23 Node tests and 6 Python tests passed.
  • Ran recipe-build with gpt-6.1-sol on a three-task Medium change (a shared calculation task and two consuming tasks):
    • Three unit code-reviewers and one security-reviewer ran. The shared-unit reviewer did not flag ACs delivered by the consuming tasks.
    • The shared-unit reviewer reported a rounding bug. After the fix, the affected code-reviewers rechecked only the correction diff, and security-reviewer was not rerun.
    • The executor reused an existing formatter found by responsibility. Executors returned completionCriteria and reported a seeded out-of-scope bug through findingsOutsideChange, which reached the completion report.

🤖 Generated with Claude Code

shinpr and others added 2 commits November 11, 2026 22:50
…code by responsibility

- Review Medium and Large changes in units of completed Work Plan tasks: a consuming task with the tasks it depends on, and one shared unit for tasks several units depend on. security-reviewer still reviews the complete change once.
- Rerun the reviewer that reported a finding, plus any passed reviewer whose scope (implementation files, applicable ACs, contracts, or trust boundaries) the correction diff changes.
- code-reviewer judges criteria instead of building a whole-change map and no longer returns the acceptanceCriteria inventory; its rerun boundary is derived from the correction diff.
- Find existing code by responsibility as well as by name, stop searching once the candidates that could change the decision have been read, and keep code separate only on evidence of different contracts or change reasons.
- Task executors report evidence per Completion Criteria item and problems observed outside the change; the orchestrator lists unhandled ones in the completion report.
- Task files list a test file only when the Work Plan fixes it.
- Code comments do not reference tasks, plans, or document sections.
- recipe-front-adjust loads the implementation skills before the parent session edits code.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@shinpr shinpr self-assigned this Nov 11, 2026
@shinpr
shinpr merged commit aa4e21d into main Nov 11, 2026
2 checks passed
@shinpr
shinpr deleted the fix/review-units branch November 11, 2026 13:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant