Try the demo: what should a coding-agent run prove before it continues?

[Keesan12]

The interactive demo is live now: https://martinloop.com/demo

The question I want to sharpen is simple: before an AI coding agent gets another attempt, what proof should it have to show?

The shape MartinLoop is testing is:

• remaining budget
• verifier movement
• failure class
• changed files
• rollback evidence
• explicit halt reason
• a run receipt a human can inspect later

The point is not to slow agents down. It is to stop the loop from spending credits on the same failure while still letting useful attempts continue.

If you are already running Claude Code, Codex, OpenCode, Cursor, or similar loops, I would love blunt feedback on the demo:

• what signal is missing?
• what would make you trust the stop/continue decision?
• what would make this worth trying from npm on a real run?

Fast path:

• demo: https://martinloop.com/demo
• npm: https://www.npmjs.com/package/martin-loop
• repo: https://github.com/Keesan12/martin-loop

If this kind of guardrail belongs in your stack, a star helps keep the OSS version alive.

[Keesan12]

Quick follow-up for anyone who tries the demo: we are using this thread to tune the public proof path, not just collect polite feedback.

The most useful signal is one of these:

• star the repo if the "budget, brakes, receipts" framing solves a real problem for your agent workflow
• run npx martin-loop demo and tell us where the receipt is confusing or missing a field
• drop the failure mode you most want Martin Loop to catch before an autonomous coding run keeps spending

A star helps the project get in front of more people who are already fighting runaway coding-agent loops. A concrete failure-mode reply helps us make the tool sharper.

[Keesan12]

Appreciate the blunt feedback. The demo is live here: https://martinloop.com/demo. If the stop/continue contract and receipts are useful in your loop, I would be grateful for a star on the repo so more builders can find it.

[Keesan12]

Update: the repo conversion surface has been tightened and merged in PR #59.

What changed:

• GitHub now detects the license correctly as Apache-2.0
• the README first screen is clearer about budgets, verifier gates, rollback evidence, and JSONL receipts
• npm metadata now uses the terms builders actually search for: MCP, Codex, Claude Code, agent runtime, AI governance, LLMOps, budget enforcement
• a social-preview asset is ready for launch graphics

If you try the demo and the receipt still does not answer the question you care about, drop that missing field here. That is the fastest way to make MartinLoop useful.