mr-onion

TestForge, Reconsidered

A reply to docs/REIMAGINING.md — Emil, 16 September 2026

Your brief made me look harder at this than I had in weeks. Here’s what’s changed since you wrote it, my answers to your four questions, and the shape of what I actually want to build.

First — a correction: the numbers here are already out of date

Phase 16 finished locally back in August, but never got pushed — so everything you looked at on GitHub was the pre-restructuring codebase. It’s on origin/main now. Concretely:

So “12,000 lines of plumbing around 13 good lines” isn’t quite the shape of the problem anymore — that part’s done. I take that as a good sign, not a rebuttal: we landed on almost the same diagnosis independently.

Your four questions

Q1. Is TestForge parked, or paused?

Paused, deliberately — not abandoned. The plan was always to get test-assist out the door as a polished product first, before opening this next question.

Q2. Does deleting the app feel like losing the work?

Yes, honestly — especially now that I can see how much of that “12,000 lines” is real, recently-finished product. But splitting out the prompt and a problem statement as a starting artifact makes sense to me now, and I expect more of this to surface as I keep building test-assist, in the direction your brief is pointing.

Q3. Which agent do you target first?

Claude Code. Making it portable to Copilot or Cursor is a real question, but a later one — I’m not designing for three targets before I’ve proven it works for one.

Q4. Is honest-coverage one skill, or the whole product?

Not enough on its own. Testing isn’t one discipline, and treating a single skill as the whole product undersells it — though honest-coverage specifically might stand alone for narrower uses. I want to be careful not to overstate that.

Where this leaves me: what I’d like to do

Not a rewrite, and not a deletion — additive. A skills/ directory that grows alongside TestForge, not instead of it. Three moves, in order:

  1. Package what’s already isolated. The prompt plus a problem statement from the README, built and tested as a real Claude Code skill — write-playwright-test, the floor skill from your brief, not the differentiator.
  2. Write the roadmap down, not just the code. A skills/ROADMAP.md recording your other four ideas — honest-coverage, dont-stub-what-you-prove, find-the-hidden-platform, plan-to-specs — as candidates, explicitly not commitments, so this doesn’t quietly collapse into one flagship skill by accident.
  3. Check before rebuilding. A short spike on whether Playwright MCP already covers what introspect/crawl do — the one genuine loss you flagged — before anyone spends time reconstructing roughly 1,200 lines of DOM-exploration logic that might not need reconstructing at all.

Ticketed as TFP-034, TFP-035, and TFP-036 in docs/decisions/phase-17-skills-extraction-backlog.md. Happy to walk through it in detail.