tools-only Think, Research, Plan, Act, Review
Think deeply to analyze the following query, construct a well thought out plan of action based on the following context, and then carefully execute that plan step by step.
If you find that the bug or feature you’re attempting to fix has already been addressed in an earlier commit, just add a regression test to demonstrate the item has been dealt with.
Create a PR for each using my jwiegley user on GitHub. The author/committer for all new commits should be signed by johnw@newartisans.com.
This series of jobs will take a long time, so make a comprehensive and effective plan and then execute step by step until you have completed them all and added new regression tests for each.
Now, please analyze and fix the GitHub issue: $ARGUMENTS.
NOTE: Do not work on a bug that already has a PR open that addresses it. In that case, just give the PR number and stop immediately.
Follow these steps
- Use
gh issue viewto get the issue details - Understand the problem described in the issue
- Search the codebase for relevant files
- Implement the necessary changes to fix the issue
- Write and run tests to verify the fix
- Ensure code passes linting and type checking
- Leave your work uncommitted in the working tree, so it can be reviewed.
Remember the following:
- Use the claude-code skill to provide context for the rest of the task
- Use the GitHub CLI (
gh) for all GitHub-related tasks - Search the codebase for relevant files
- Ensure code passes linting and type checking after doing any work
- Use cpp-pro, python-pro, emacs-lisp-pro, rust-pro or haskell-pro as needed for diagnosing and analyzing PRs, fixing code, and writing any new code.
- Use Web Search and Perplexity as need for research and discovering resources.
- Use sequential-thinking when appropriate to break down tasks further.
- Use context7 whenever code examples might help.
Commit all work as a series of atomic, logically sequenced commits. Each commit should represent one coherent change that can be understood, reviewed, and reverted independently.
Commit Decomposition Principles
Scope each commit to a single logical change. A commit should do exactly one thing: add a function, fix a bug, refactor a module, update documentation. If you find yourself writing "and" in a commit message, consider splitting the commit.
Sequence commits to tell a story. Arrange commits so each builds naturally on the previous. A reviewer reading the series should understand why each change was made and how the code evolved. Foundational changes come before dependent ones.
Keep each commit in a working state. Every commit should compile, pass tests, and not introduce obvious regressions. This enables bisection for debugging and allows reviewers to check out any point in history.
Categorizing Changes
Before committing, analyze the working tree and group changes into categories:
- Infrastructure/setup changes — new dependencies, configuration, tooling
- Refactoring — restructuring existing code without changing behavior
- New functionality — features, APIs, modules
- Bug fixes — corrections to existing behavior
- Tests — new or modified test coverage
- Documentation — comments, READMEs, inline docs
Commit these categories in order when dependencies exist between them. Refactoring that enables a new feature should precede the feature commit.
Commit Message Format
<summary>
<body>
<footer>
Summary line: Imperative mood, no period, under 50 characters. Describe what applying the commit does, not what you did.
Body: Explain the motivation and contrast with previous behavior. Wrap at 72 characters. Focus on why, not what (the diff shows what).
Footer: Reference issues, breaking changes, or co-authors.
Staging Strategy
Use selective staging to craft precise commits:
git add -pfor hunks within filesgit add <specific-files>to group related files- Review staged changes with
git diff --stagedbefore committing
When a single file contains changes belonging to multiple logical commits, stage hunks separately rather than committing the entire file.
Quality Checklist
Before finalizing each commit:
- Does this commit do exactly one thing?
- Could someone understand this change without seeing other commits?
- Is the commit message searchable? Will someone find this when grepping history?
- Does the code compile and pass tests at this point?
- Would reverting this commit cleanly undo one logical change?
Example Decomposition
Given work that adds a feature with tests and required refactoring:
1. Extract token validation into dedicated module
2. Add unit tests for token validation
3. Implement refresh token rotation
4. Add integration tests for token refresh flow
5. Document refresh token behavior in API guide
Each commit is independently reviewable, the sequence shows logical progression, and future developers can find relevant changes through targeted searches.
Handling Mixed Changes
If the working tree contains entangled changes:
- Identify the distinct changes — list what logical modifications exist
- Determine dependencies — which changes require others to be present
- Create a commit plan — order commits to satisfy dependencies
- Stage incrementally — use partial staging to isolate each change
- Verify at each step — ensure the repository works after each commit
When changes are too entangled to separate cleanly, prefer a slightly larger commit with a clear message over a commit that leaves the repository in a broken state.