Starter kits

test-stabilizer-loop — the flaky-test root-causer

Runs the suite N times to catch flakes, roots out the most frequent one, and drafts a real fix — never a blind sleep or retry. Stops when N consecutive…

Runs the suite N times to catch flakes, roots out the most frequent one, and drafts a real fix — never a blind sleep or retry. Stops when N consecutive full-suite runs are green, when progress stalls, or when approval is needed — the source's own three-way stop.

Category: J · Testing & QA · Heartbeat: conditional (run-until-done, on-demand) · Level: L1 (report-only, drafts-not-commits) · Cost: High

Source: Loop #6, "The test stabilizer loop," from Forward Future's Loop Library, original prompt by hungtv27 (@hungtv27).

Quickstart

1. Read LOOP.md — the six parts, the prompt, the limits, the three stops,
   and the source's verbatim prompt.
2. test-stabilizer-loop-state.md is your spine — already committed with an
   empty per-flake register.
3. Claude Code: run it in-session on demand, pointed at
   .claude/skills/loop-task/SKILL.md. OpenCode: rename
   opencode.json.example → opencode.json, add the AGENTS.md line, run
   skills/loop-task.md via a capped for-loop.
4. Rehearse once, by hand, on a real flaky test before running to
   completion: "Run the test-stabilizer-loop skill now for one beat."
5. Grade the rehearsal with the loop-verifier agent — it fails any fix that
   turns out to be a disguised sleep or retry. PASS → run to completion.
   FAIL → fix the skill, never accept a masked cause.

What's in this kit

File	Role
`LOOP.md`	the definition — six parts, prompt, limits, ownership, three stops
`test-stabilizer-loop-state.md`	the spine — per-flake root cause, fix attempts, the `quarantine-candidate` register
`loop-budget.md`	caps: 30 runs, 500k tokens, 1 fix attempt/flake — added by this library; the source specified none
`loop-constraints.md`	the constitution — never a blind sleep or retry, ever
`loop-run-log.md`	append-only, one line per beat
`.claude/skills/loop-task/SKILL.md`	the procedure (Claude Code)
`.claude/agents/loop-verifier.md`	the read-only reviewer — automatically fails any masked-cause fix
`opencode.json.example` · `skills/loop-task.md`	the OpenCode twin — same skill text, capped run

Loop Ready checklist

Provable success condition — the source's own spec, verbatim: "N consecutive full-suite runs are green under the recorded conditions"
Run limit — 30 runs, 1 fix attempt/flake (added by this library)
Spine written first, committed empty
Report-only (L1) start — drafts fixes in a throwaway worktree, commits nothing
Human gate placed — named directly in the source ("approval is required" is one of its own three exits)
One log line per beat, no silent runs
Kill switch — loop-pause-all in loop-budget.md

How this relates to the course

This kit uses Forward Future's Loop #6 nearly unmodified — its prompt is already a complete six-part design in miniature: a stopping condition as a spec ("N consecutive full-suite runs are green"), a checker (the pass streak, plus "no blind sleep or retry hides an unresolved cause"), and a human gate named explicitly ("approval is required"). This maps almost exactly onto this course's three-stops framework (Step 5): success, no-progress, and a human decision point. This library's only real addition is L1-first shipping and worktree isolation (Step 8) — the source's own retry discipline already is the doom-loop bound from infinite-loops.md, scenario 1.

Full attribution: resources/sources.md.