Show your loop

[cobusgreyling]

Show your loop

This is the permanent show-and-tell thread for Loop Engineering.

Post anything you’re proud of, stuck on, or surprised by — real runs preferred over theory.

Template (copy/paste)

### What I ran
- **Pattern:** daily-triage / pr-babysitter / ci-sweeper / other:
- **Tool:** grok / claude / codex / cursor / opencode / other:
- **Command:** `npx @cobusgreyling/loop-init . --pattern … --tool …`

### Outcome
- **Loop Ready score:** (before → after if you have it)
- **What worked:**
- **What failed or surprised me:**
- **Would I run this again next week?** yes / no / with changes:

### Optional proof
- Badge: `npx @cobusgreyling/loop-audit . --badge`
- Link to PR, story, or adopter issue

Start here if you’re new

npx @cobusgreyling/loop-init . --pattern daily-triage --tool grok
npx @cobusgreyling/loop-audit . --suggest

Demo: Loop Ready score climbing

Week-one culture: report-only (L1) before auto-fix (L2). See LOOP.md.

Also useful

• Add your project to adopters
• Share a production story
• v1.6.0 release
• Interactive showcase

Reply below — first loops, failures, and score screenshots all count.

[cobusgreyling]

Maintainer note: this is the pinned-style hub for first runs and demos.

If you just tried loop-init / loop-audit, drop your score and pattern below — even if week one was report-only only.

Latest release: v1.6.0

[KongFangXun]

What I ran

• Pattern: other: commit-time audit gate — a pr-babysitter variant that watches every AI-authored commit, not just PRs
• Tool: other: own harness — sofagent (MIT), with DeepSeek Harness + GLM as execution backends
• Command: not loop-init — the loop is the product itself: sofagent-audit --install-hook wires pre-commit / commit-msg / post-commit hooks into the repo

Outcome

• Loop Ready score: n/a (our gate predates loop-audit; didn't fake a score)
• What worked: deterministic rules over vibes. 24 zero-token rules scan every git diff (secret leaks, out-of-scope edits, blind modifications, prompt-injection traces) and veto bad AI commits before they land. Every AI commit passes through it — including the commits that fixed the gate itself.
• What failed or surprised me: two things. ① Network: our push channel kept dying mid-loop (github.com 443 flapping), so we fell back to the Git Data API (blobs → trees → commits → refs PATCH) — the loop survived; our network was the bottleneck, not GitHub. ② Community: our first awesome-list submission got auto-rejected by checklist bots because our fork was behind upstream — the diff showed other people's newly-merged entries and the bots flagged those. Rebasing the fork to upstream HEAD before submitting fixed it.
• Would I run this again next week? yes — it runs on every commit, and it just vetoed one of my own doc-only changes for a regression during this week's release.

Optional proof

• Adopters entry merged: Add sofagent to adopters (PR audit gate, L2) #573 (thanks!)
• Real audit output from today's commit: ✅ sofagent 审计 · 17 项检查 · 24 条规则 · PASS
• Also live as a GitHub Marketplace action for teams who want the same gate on PRs

[burgerseater]

We have open-sourced LoopArena, a benchmark for comparing models as runtime Controllers of coding-agent loops.

The evaluated component is the Controller. Across Controller-model comparisons, the coding Worker and execution setup are held fixed, so the benchmark compares how different models guide the same Worker under common conditions.

LoopArena evaluates this role at three scopes: execution-validated next-step decisions, repeated control over task slices, and complete software tasks.

• GitHub: https://github.com/AMAP-ML/LoopArena
• Hugging Face paper: https://huggingface.co/papers/2608.28281
• ModelScope paper: https://www.modelscope.cn/papers/2608.28281

Disclosure: I am one of the authors/maintainers. We would be interested in hearing how others evaluate the model or policy responsible for runtime loop control.