ebman CLAUDE.md

AI working rules for ebman

This file is read by Claude Code (and similar agents) on session start. Follow it.

What this project is

ebman is a Rust + ratatui TUI for AWS Elastic Beanstalk, k9s-styled. Source under src/. Backlog in BACKLOG.md. Tests live alongside the code in #[cfg(test)] mod tests blocks.

Where to start

PLAN.md is the rolling window of active work — read it first, work it top-down, prune and refill it before finishing. BACKLOG.md is the reservoir behind it. An item lives in exactly one of the two.

PLAN.md also carries the item classes (which stages an item needs) and the verify-the-claim gate. This file stays what it has always been: how to build, what green means, and the stop conditions.

Mandatory loop for autonomous work

When the user asks for autonomous work (e.g. "run autonomously", "build all the above", "next", or any directive to ship multiple items without per-step approval), you must:

1. Build green before claiming done. All three, in this order:

   cargo fmt --all
   cargo clippy --all-targets -- -D warnings
   cargo test
   

   Not cargo build. It does not run clippy, and it does not compile test targets — which is where most of the recent work lives, so "no new warnings" from a bare build covers none of it. CI runs fmt and clippy and will reject what a build-only check passed.

   Two gates run in CI that you cannot easily reproduce and should think about before pushing, not after: candor (a render path that reaches the network / LLM / filesystem fails the build) and cargo-deny.

2. Self-review every meaningful change. After each substantive feature or pass, perform a code review against the changes you made and surface bugs, design issues, dead code, missed edge cases, and inconsistencies. The review goes in your message back to the user — not just internal thinking.

3. Act on review findings — don't just list them, but keep them separable. Fix them in the same run; put them in separate commits. "Same turn" was being read as "same commit", which mixed a feature, three drive-by fixes and a refactor into one diff and destroyed git bisect on precisely the class of subtle regression this codebase keeps hitting. CONTRIBUTING.md asks contributors for a change that does one thing; the same applies here.