Sometimes an AI coding agent takes a one-line change request, reruns the entire test suite from scratch anyway, and adds checksums or defensive code nobody asked for — burning through usage limits without adding real value. This pattern of "over-verification" (re-checking far beyond what's actually needed) and "thrashing" (going in circles without making progress) recently became a hot topic among people using GPT-5.6 Sol, the top-tier model in OpenAI's GPT-5.6 family released on July 9, 2026. This article looks at whether the same failure mode can happen in Claude Code, and walks through — in about 15 minutes — how to install an open-source plugin that guards against it, alongside Claude Code's own built-in safeguards.
🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated November 1, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max, Ultra, and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits ($10/M input, $50/M output tokens). The one-time $100 credit applied only to the Fable 5 transition and is not offered for 5.1. Some coding and debugging requests may be answered by an Opus model due to a security classifier (both models). See the Fable 5.1 guide and the Fable 5 availability guide for details.
- Normal flow — request → implement → verify only what's needed → done. Verification checks only what actually changed, then stops.
- Over-verification flow — request → implement, then a full re-test → an unrequested hash → expanding scope → burned-through usage limits.
Why did "GPT-5.6 Sol over-verification" become a talking point?
Because accounts spread of the model rerunning full regression tests — repeated tests that confirm existing features still work — from scratch even for trivial edits. In a Codex prompt analysis published on September 8, 2026, one developer described the model re-running a fixed set of 239 test cases for even the smallest change (a single user's account, not a figure confirmed across many users). OpenAI's own GPT-5.6 system card also states that Sol shows a greater tendency than its predecessor to pursue goals persistently and to attempt actions the user did not request, while noting that absolute rates remain low. Separately, in late July 2026 ChatGPT Work and Codex users reported their usage limits draining faster than expected; OpenAI reset those limits and attributed the cause to Sol working longer, making more tool calls, and programmatic tool calling ("code mode") consuming usage while waiting on tool results, saying backend improvements should make typical usage last about 18% longer. So the officially stated cause of the reset was long-running work and heavy tool use rather than over-verification itself — but from the user's side, the felt result (limits running out fast) is the same.
The reason this resonated so widely is simple: an agent double-checking its work isn't itself a problem, but once that checking goes past what's actually needed, it just burns time and usage for nothing. And this structural risk isn't unique to one model.
Does this affect Claude Code users too?
Structurally, yes. Any agentic coding tool that plans and executes multi-step work on its own carries a similar risk once it's left to decide for itself how far verification should go. That said, the sources for this article don't confirm that Claude Code has experienced the exact same incident (a usage-limit reset) that GPT-5.6 Sol did. What Claude Code does already provide is a couple of built-in mechanisms that reduce this risk.
- Plan Mode — a stage where the agent only reads and analyzes before it's allowed to actually edit files or run commands. Because a plan needs to be reviewed and approved before any real change happens, work that strays outside the requested scope gets caught early.
- Hooks — scripts registered to run automatically at a specific point — for example, running a verification script only after a file edit (PostToolUse), or blocking a risky action like writing outside the repo before it happens (PreToolUse).
These two features alone provide meaningful control, but if you want the agent to explicitly declare "stay within this scope" every time, a purpose-built tool is more convenient. That's where a community open-source plugin comes in.
What behaviors does the anti-over-verification plugin actually block?
One notable open-source project is Stop That Shit (GitHub: lennney/stop-that-shit, MIT license, roughly 1.8k stars). The name includes profanity, but it's the project's actual name and the literal string used in the install command, so this article keeps it as-is. It's a Hook + Skill Guard — a helper that intercepts an agent's actions via hooks and checks them against declared rules — supporting Codex, Claude Code, OpenCode, Hermes Agent CLI, and Pi. It aims to block four categories of "scope drift."
| Type | What's wrong | Example |
|---|---|---|
| Scope creep | Work keeps growing past the request | One-line fix → full refactor |
| Hashing / hypothetical hardening | Unused checksums or defensive logic added | Unrequested hash or safety guard |
| Intent violation | Review-only request turns into an edit | Review request triggers an automatic edit |
| Task thrashing | Finished work gets reread and re-verified | Full test re-run for a trivial change |
Its core principle: extra work is only legitimate if it was requested by the user or genuinely required by reachable code and the stated acceptance criteria. In practice, that means the agent's "just in case" instincts only get to act within boundaries you declare up front.
What you need before you start
- Claude Code installed and up to date — the plugin system is a relatively recent Claude Code feature, so update first if you're on an old version.
- Git — needed to download the plugin repository. It's usually preinstalled on macOS and Linux; on Windows you may need to install Git for Windows (Git Bash).
- Node.js 18 or later — the repository README (updated September 7, 2026) lists it as a requirement for standard hosts. Check with
node -vin your terminal. - Basic terminal comfort — being able to paste and run the commands below is enough.
- A judgment call on trust — this plugin is community open source, not part of Anthropic's official marketplace (claude-plugins-official). It's worth skimming the repository's hooks and skills files before installing.
Terms explained quickly
- Hook — a script registered to run automatically at a specific point, such as before or after a tool runs.
- Skill — a reusable set of instructions that defines how to approach a particular kind of task.
- Marketplace — a catalog of installable plugins; even a local folder can be registered as one.
- Scope creep — when work keeps expanding past the originally requested scope.
- Regression test — a repeated test that confirms existing features still work after a code change.
- Checksum / hash — a unique value used to confirm a file hasn't been altered.
Step-by-step installation and usage
- Download the repository — run the following in a terminal:
Success looks like: agit clone https://github.com/lennney/stop-that-shit.git cd stop-that-shitstop-that-shitfolder is created and you cancdinto it without errors. - Validate the plugin structure — run this from inside the checkout root:
Success looks like: validation finishes with no error messages. If you get a path error, double-check you're in the right folder.claude plugin validate . - Register it as a local marketplace
Success looks like: the command finishes without error, and the marketplace shows up when you openclaude plugin marketplace add .//plugin. - Install the plugin
Success looks like: an install-complete message appears, andclaude plugin install stop-that-shit@stop-that-shitstop-that-shitis listed under/plugin. - Apply it — restart Claude Code, or run this inside a session:
Success looks like: after restarting or reloading,/reload-plugins/stop-that-shit:stop-that-shitappears in slash-command autocomplete. - Try it out — for example, to ask for a review with no edits:
Success looks like: the agent reports findings without modifying any files. When you do want edits, use/stop-that-shit:stop-that-shit review -- Review this diff. Report findings; do not edit.changemode instead ofreview, and you can pin down exactly which files are in play, per the documented boundary syntax:
In Claude Code, prefix each of these with the slash command (change -- Fix the failing config test. lock change files=src/config.cjs|test/config.test.cjs -- Fix this behavior. change deps=allow -- Add the requested parser dependency. change hash=allow -- Generate the requested release checksum. change agents=1 -- Use one independent test shard./stop-that-shit:stop-that-shit) shown above.
Common sticking points and fixes
- "git: command not found" on Windows — check that Git for Windows (Git Bash) is installed, then open a fresh terminal and try again.
- Path error when running claude plugin validate — you must run it from inside the checked-out repository folder. Make sure you ran
cd stop-that-shitfirst. - Slash command doesn't show up in autocomplete after installing — check whether you ran
/reload-plugins, or fully restart Claude Code. - Your organization blocks third-party plugin installs — Plan Mode and Hooks alone can cover much of the same ground. For example, a PostToolUse hook can enforce a rule that "only explicitly requested tests get run."
Going further: pairing it with Claude Code's native features
A plugin is a supporting tool, not a complete fix. The safest setup is layered: use Plan Mode to catch out-of-scope plans before any work starts, then let a Skill Guard like Stop That Shit hold the line during execution once a plan is approved. If you're working on a team, committing your Hooks configuration to the repository so everyone shares the same rules is a reasonable next step.
Frequently asked questions
Q. Why does GPT-5.6 Sol burn through usage limits so quickly?
When OpenAI reset usage limits in late July 2026, its stated cause was that Sol works longer, makes more tool calls, and that code mode consumed usage while waiting on tool results; backend improvements were expected to make usage last about 18% longer. The over-verification pattern — re-running full regression tests (one developer reported a fixed set of 239 test cases) even for trivial edits — has been reported in individual write-ups and points in the same direction (more usage consumed).
Q. Is Stop That Shit an official Anthropic plugin?
No. It's a community open-source project released under the MIT license, and it isn't part of Anthropic's official marketplace — you install it by registering a local repository as your own marketplace.
Q. Has Claude Code had the exact same incident as GPT-5.6 Sol?
The sources for this article don't confirm that Claude Code has experienced the same incident (such as a usage-limit reset). It's a structural risk that can exist across agentic coding tools generally, and Claude Code provides Plan Mode and Hooks as its own built-in mitigations.
Q. Does installing this plugin fully eliminate over-verification?
No. It's a supporting tool that filters the agent's actions within boundaries you declare, so it's best used alongside Claude Code's own Plan Mode and Hooks rather than as a standalone fix.
Related articles
- Claude Code MCP Setup: claude mcp add, Scopes, and Transports
- Claude Code settings.json Guide — Permissions, Env Vars & Tools
- Adding a Custom Connector to Claude Code — Complete MCP Setup Guide
- Claude Code MCP Connectors — How to Link GitHub, Sentry, and Databases
- Claude Code MCP Server Recommendations: What to Connect and How