← statichum.studio

A no-progress 'circuit breaker' that auto-halts a coding agent stuck repeating failed actions — triggered by looping, not by a dollar/token cap

dev tool real project ••• trending

Developers running Claude Code / Codex autonomously report agents falling into doom loops: re-running tests against a small self-inflicted error (missing import, wrong param) and burning enormous token volume with zero progress. The specific pain is detecting and halting on repeated failed/no-diff actions, which current guardrails treat as a weak heuristic bolted onto a deterministic budget cap that fires a turn or more too late. The wedge is loop/no-progress detection as the PRIMARY halt trigger, independent of spend — and crucially one that catches paraphrased retries, not just exact-match repeats.

builder note

Distinct from a budget cap or cost dashboard — users say those fire too late. The defensible product leads with semantic detection (repeated edit-test-fail cycles, no-diff progress over N turns) and halts immediately. The hard part is catching paraphrased retry loops, because naive exact-hash matching misses them; that detection is the moat, not the dashboard.

landscape (3 existing solutions)

Existing guardrails anchor on dollar/token budget caps and treat loop detection as an afterthought, so they halt only after the waste happens; nothing leads with semantic 'agent is repeating failed actions / making no progress' detection as the primary, immediate trigger.

Burnstop Verified by fetch: 'Loop detection is a heuristic; the budget cap is the deterministic backstop.' Loop detection is explicitly secondary, fires roughly one turn behind, and 'can't stop mid-turn' — so a real loop is only caught after the spend, not on the no-progress signal itself.
Loopers (tryloopers.com) Primarily a token/dollar budget firewall; 'loop detection' is listed but under-specified, with no distinction between semantic no-progress detection and statistical thresholds.
Codex built-in Stop The vendor's own halt is reactive and slow (10-20s inside a tool-call loop) and there is no automatic detection that fires without a human watching.

sources (3)

github https://github.com/openai/codex/issues/30061 "stuck in a runaway inference loop... very high cpu usage and token burn" 2026-06-25
github https://github.com/openai/codex/issues/30209 "Stop button takes 10-20s when model is in a tool-call loop" 2026-06-26
hn https://news.ycombinator.com/item?id=48643569 "if you run the 'loops'... it will burn through enormous amounts of usage" 2026-06-25
ai-agentsclaude-codecodexloop-detectionguardrails