Skip to content

Review large changes in units and give agents the frontend and test criteria they judge by - #200

Merged
shinpr merged 3 commits into
mainfrom
fix/review-proof-ownership
Nov 6, 2026
Merged

shinpr merged 3 commits into
mainfrom
fix/review-proof-ownership

Conversation

@shinpr

@shinpr shinpr commented Nov 6, 2026

Copy link
Copy Markdown
Owner

Problems and changes

One reviewer could not read a Large change in full

A single code-reviewer call covered every task of a Large change, so it read only a small part of the changed tests and missed duplicated proof and unused code.

  • Large changes are now reviewed in units of Work Plan tasks. A unit holds a consuming task with the tasks it depends on, and tasks shared by several units form one shared unit. At most 3 reviewer calls run at once.
  • Medium changes keep one code-reviewer. security-reviewer still reviews the complete change once.
  • The build and implement recipes pass each unit's files and tasks to its reviewer.

code-reviewer spent its effort on inventories instead of reading code

code-reviewer had to list every acceptance criterion, including fulfilled ones, and map the whole change in a fixed order before judging it. No consumer used the fulfilled entries.

  • The acceptance-criteria inventory is removed from the output and from the review recipes' reports. Unfulfilled criteria are still reported as findings.
  • The full-change map and the fixed order are replaced by the criteria the review judges.
  • The completion check requires every material surface in implementationFiles to be judged before the verdict.
  • Wording that assumed a workflow context is removed, so the agent behaves the same when called directly.

Frontend criteria did not reach shared agents

ai-development-guide and frontend-ai-guide repeated the same general criteria. Shared agents such as code-reviewer, investigator, verifier, and solver loaded only ai-development-guide, so they never received the React/TypeScript criteria.

  • frontend-ai-guide is removed. Its React/TypeScript-specific criteria move to ai-development-guide/references/frontend.md, and the frontend agents preload ai-development-guide.
  • The ai-development-guide and testing-principles descriptions now state which code and tests they cover.

Agents skipped references that they were told to read conditionally

When an agent was told to read a reference only under a condition, such as "for frontend code", it skipped the reference even when the condition held.

  • code-reviewer, integration-test-reviewer, investigator, verifier, and solver now read their references before starting, without conditions. The reason is stated with each instruction: those files hold criteria beyond the preloaded skills.
  • Executors, quality fixers, and acceptance-test-generator read the frontend and E2E references at the step that uses them.

Task files decided which tests to write

Task files listed test files in Target Files, and executors wrote tests file by file instead of choosing them by behavior. Executors also left test quality to downstream checks that only run tests.

  • Target Files list a test file only when the Work Plan fixes it, such as a generated skeleton. Executors select other tests under testing-principles.
  • The executor statement that downstream quality assurance re-checks test quality is removed, along with duplicated escalation rules.

Breaking change

The frontend-ai-guide skill no longer exists. Use ai-development-guide and its references/frontend.md instead.

🤖 Generated with Claude Code

shinpr and others added 2 commits November 6, 2026 12:59
…riteria they judge by

- Split post-implementation code review of Large changes into units of Work Plan tasks, run at most 3 at a time; Medium changes keep one reviewer
- Remove the acceptance-criteria inventory, full-change map, and fixed review order from code-reviewer, and require every material surface in implementationFiles to be judged before the verdict
- Merge frontend-ai-guide into ai-development-guide: React/TypeScript-specific criteria move to references/frontend.md and frontend agents preload ai-development-guide
- Have reviewers, investigators, and designers read the frontend and E2E references before their work without conditions, and have executors and fixers read them at the step that uses them
- List test files in task Target Files only when the Work Plan fixes them, and stop deferring test quality to downstream checks in task executors

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 6, 2026
- Remove the task executors' rule that synced [ ], [🔄], and [x] across the task file, Work Plan, and Design Doc
- Add a final gate item: for task-file execution, each Completion Criteria item in the task file is satisfied with evidence in the response

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@shinpr
shinpr merged commit d23f861 into main Nov 6, 2026
1 check passed
@shinpr
shinpr deleted the fix/review-proof-ownership branch November 6, 2026 04:18
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