#2546

test-driven-development / executing-plans: "batch the test and the code to save a turn" is missing from the rationalization table, and fix-pass ledger lines don't require RED output

Openachievebysteveopened 13d ago
  • I searched existing issues and this has not been proposed before

Related but distinct: #994 (closed — RED evidence in TDD subagent reports), #2146 (RED passing for the wrong reason), #2230.

What problem does this solve?

During an executing-plans Final Review fix pass, the agent wrote the new failing tests and the implementation in consecutive responses without running the tests in between. The context monitor had warned the window was nearly full, and the user had just added a new request mid-turn. The first run after both were written showed two defects that a RED run would have exposed before the implementation existed:

  • one test used an id that the implementation was about to make valid, so it could never have failed on the old code;
  • a concurrency test asserted only that output "starts with 'beat '", which beat skipped: ... also satisfies, so it could pass while losing beats.

The only protection in play was skill text. The "each fix RED→GREEN" line in executing-plans was loaded and skipped anyway.

Proposed solution

  1. test-driven-development, rationalization table, new row: "I'll batch the test and the code to save a turn" → "The RED run is how you learn the test can fail. A batched run only proves the pair agrees with itself."
  2. executing-plans, Final Review fix pass: each fix's ledger line quotes the RED output (the failing assertion text), not just the test name. A line without RED output is an unverified fix. This makes a skipped step visible in the artefact instead of relying on the rule being remembered under pressure.

What alternatives did you consider?

Stronger wording on the existing RED→GREEN line. Rejected: the line was already loaded and clear; the failure was skipping it under time/context pressure, which a louder sentence doesn't change. Requiring the RED output in the ledger turns the skip into a visible gap.

Is this appropriate for core Superpowers?

Yes — it's about the TDD and executing-plans workflow itself, independent of project domain.

Environment (required)

FieldValue
Superpowers version6.4.2
Harness (Claude Code, Cursor, etc.)Claude Code (Claude desktop app, Windows)
Harness versiondesktop app, October 2026
Your model + versionClaude Opus 5.5 (claude-opus-5-5)
All plugins installedsuperpowers, watch-video, playwright, ghost-writer, remotion

Context

Real session; observed 2026-10-15 while applying four review findings to a small web page project.

Activity

0

No activity yet.