Revision is the part that changes outcomes: a wrong assumption at thought three can be replaced in place instead of quietly carried through everything after it, and a branch lets two candidate fixes be developed before one is picked. Because the model has to be told to use it, the prompt templates are the reliable entry point — pick bug-analysis or refactoring-plan and the structured pass happens without you having to ask for it in words.
One tool that structures the model's own reasoning about code. Each call is a single thought carrying its number, an estimate of the total, and whether another is needed — and the estimate is allowed to move as the problem gets clearer.
- A numbered chain of thoughts, each one call, with total_thoughts adjustable mid-chain and next_thought_needed set to false only when the reasoning is actually finished
- Revision in place: is_revision plus revises_thought rewrites an earlier step rather than piling a correction on top of it
- Branching with branch_from_thought and branch_id, so two approaches can be carried side by side instead of one being abandoned
- needs_more_thoughts extends a chain that turned out to be longer than it looked
- A stop at 20 thought steps, so a loop ends rather than running away
- Five prompt templates for the moments this actually helps: architecture-decision, bug-analysis, code-review, feature-planning and refactoring-plan
- Every template takes working_directory, so the reasoning is anchored to the project you are in
- Argument values persist between sessions, so a template you use often does not have to be refilled
No account and no key — it reasons, it does not call anything out. It runs as npx -y @mettamatt/code-reasoning, or as the code-reasoning command once installed, with --debug for verbose logging and --config-dir to move the configuration directory. MIT.
One command — npx -y @mettamatt/code-reasoning