Batched Bash

Batched bash

1. The project's own verbs beat hand-rolled pipelines. If the environment note or repo shows a Justfile, Makefile, Taskfile, or package.json scripts, use those (just build, make test, bun run check): they encode the flags and env the project actually uses. Hand-roll only what the project has not named.
2. One call per goal, not per command. Each tool round costs a full model round-trip; the commands themselves are nearly free. Write bash -lc "mkdir -p public && mv dashboard.html public/index.html && ls public". Never three separate calls for mkdir, mv, ls.
3. Label the stages so the output reads like a report and a failure names its stage: echo "=== install ===" && bun install 2>&1 | tail -2 && echo "=== typecheck ===" && bunx tsc --noEmit 2>&1 | head -20
4. Make pipelines self-verifying. Append the check to the action: grep -q "export default" index.ts && echo OK || echo MISSING.
5. Bound every noisy step inside the chain: | head -20, | tail -5, --stat, -n 5. The context you save is your own.
6. Batch independent probes with separators: ls src; echo ---; ls tests; echo ---; head -5 Cargo.toml.
7. Repo content lookups do not go through the shell. search_files and find_files already run ripgrep without the overhead; never shell out to rg, grep -r, find, or fd for repository files.