ECC's verification-loop skill stops reporting a failed build as a pass

The ECC plugin's verification-loop skill now reports a failed build as failed. Its build step piped the build into tail, so a broken build came back as a pass. If your agent ran it before a pull request, that pass may have been wrong.

verification-loop is ECC's check before a pull request. It runs the build, a type check, lint, tests and a secrets search, then writes a report that says ready or not ready. The build runs first.

In bash, a pipeline exits with the status of its last command. The build step sent npm run build into tail, so the pipeline reported tail's result, and a failed build exited zero, the code for success.

pipefail is the bash setting that fails a pipeline when any command in it fails. The skill switched it on only in the type check, the second phase, after the build had already run.

The fix turns pipefail on at the top of the build, lint and test steps, the three that pipe a check into head or tail. The same failed command now exits one, and the report says the build failed.

With pipefail on, head stops reading after thirty lines, and a clean linter still writing output can fail on a broken pipe. So the merged change swaps head for sed, which shows thirty lines but reads everything.

The fix covers the skills folder the Claude Code plugin installs, and the copy for Pi. The copy under dot agents, which Codex and Copilot CLI load, still has the old build line, and so do the dot kiro copy and six translations.

The fix is on main. ECC's latest tagged release, from October 1, came out before it. ECC's verification-loop skill stops reporting a failed build as a pass. It's pull request 3455 in the ECC repository.