Repository navigation
ThumbGate: local pre-action checks for Claude Code, looking for one reproducible failure #1942
IgorGanapolsky
started this conversation in
Show and tell
Replies: 1 comment
|
A code review can show what changed. It can't prove an agent's safety hook actually ran. I'm looking for one Claude Code operator with a repeated tool-call failure. Send a sanitized command, your agent/version, and what should have happened. I'll check whether ThumbGate supports that boundary before suggesting an install. No private logs, credentials or customer data. Our latest packaging repair has passed focused tests and an independent package review, but it is not released. I'm not asking you to install an unreleased fix or claiming it stops every mistake. If we can reproduce your case, the next step is a scoped fit check, not a subscription pitch. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
I build ThumbGate, an open-source local check on proposed tool calls in Claude Code.
I'm looking for one operator with a repeated mistake to test against, rather than another general agent-safety discussion.
The boundary matters: a prompt rule isn't an interception point. A check only helps when the actual tool call passes through its hook. It can't cover work that bypasses that hook.
The core is free. Matching destructive actions warn by default; strict mode can deny them. This isn't a promise to catch every mistake, and I'm not claiming a tested integration with this GitHub Action.
If you have a concrete case, reply with a sanitized proposed tool call, your agent/version, and what should have happened. No credentials, private logs, customer data, or real sensitive paths. I'll first check whether ThumbGate supports that boundary before suggesting a setup.
Code and setup: https://github.com/IgorGanapolsky/ThumbGate
All reactions