Loop Engineer v0.3.4 — make your agent loop prove it's done

[SollanSystems]

Long-running agent loops fail in a specific, repeatable way: they say they're
done. After compaction they forget what "done" meant, optimize to the visible
test, and report success with no independent evidence. We call that a false
completion, and this whole repo is built to catch it.

Loop Engineer is a proof-of-done layer that sits above your agent runtime:

• 7 typed terminal states — Succeeded requires evidence; FailedUnverifiable,
  FailedBlocked, FailedBudget, FailedSafety, FailedSpecGap, AbortedByHuman
  give a loop honest ways to stop that aren't a silent "completed".
• A contract validator (doctor) and inspector (inspect) that grade a loop's
  proof machinery on invocation evidence, not self-asserted flags.
• A held-out gate + anti-cheat scan built to catch a loop gaming its own verifier.

Try it in 5 seconds, no install, no API key:

git clone https://github.com/SollanSystems/loop-engineer.git
cd loop-engineer
python3 -m loop inspect examples/naive-loop        # score 0, weak — self-asserted "done"
python3 -m loop inspect examples/coverage-repair   # score 90, strong — gate-backed proof

The demo GIF in the README is filmed on exactly those two commands — nothing
staged, and the strong example's false_completion: false is backed by a
committed real gate verdict, not a hand-set flag.

Tiers 1–2 (the contract format + proof toolchain) are pure-stdlib Python over
files on disk — runtime-neutral. The 9 Claude Code skills are the reference
runner, not a requirement.

Feedback welcome — especially from anyone running long-horizon loops on other
runtimes (LangGraph, OpenHands, ruflo): what would it take to put this proof
layer above your loop?