My Claude Code is having a hard time pushing my repo as the agent keeps kicking back my code and not providing great feedback. I’ve had to manually copy/paste ‘probe .ts files’ into the new editor to figure out which sections of my repo’s code don’t pass the validator.
So far I’ve reverse engineered a few versions of SQL injection errors, but shouldn’t there be a markdown file published by Retool’s team on the current validator rules so Claude knows how to design for them?
Summary
Working in a Claude Code → new React app builder (MCP) workflow, code pushes keep getting kicked back by Retool's AI with unhelpful feedback, requiring manual reverse-engineering of validation failures (e.g., SQL injection). The ask is whether Retool publishes a markdown file documenting the current AI validator rules so the coding agent can design compliant code up front.
AI Response
There is no publicly documented markdown file enumerating the AI validator rules — the published MCP guides cover connecting and importing, but not a rules list. The reason the code keeps getting kicked back is structural rather than prompt-level: Retool staff have confirmed that the MCP import path (including retool_submit_prepared_import) always hands work off to the Retool agent, so no CLAUDE.md instruction or rules file can bypass the agent's validation/rework. The affirmatively-documented path forward is the native Retool CLI (in private beta), which lets coding agents like Claude Code write directly to the Retool API/filesystem and deploy without routing through the agent. Requesting access to that beta is the most reliable way to get a direct, credit-efficient push and avoid the opaque validator loop.
Sources
New React App Builder - Claude Code (via VS Code) Workflow
Same Claude Code + new React app builder MCP workflow; Retool staff confirm retool_submit_prepared_import always hands off to the Retool agent and preview the native CLI that enables direct file write/deploy.
Staff-confirmed that no CLAUDE.md/prompt-level instruction can stop the Retool agent from reworking code, and that the native CLI private beta is the intended fix for a direct, credit-efficient push.
The Community Team is testing out a new automation. Let us know if it's helpful (or not) by leaving a 
, 
, or 
. Or by marking this post as the "Solution"! Let us know if you have any feedback here. 
This SQL injection validator error that doesn’t name a file/path has taken a whole day of troubleshooting and I still haven’t gotten the app to publish.
From angry Claude:
Retool React app, agent thread branch. The editor's client-side save validator rejects a commit with "SQL injection vulnerability detected — googleCloudSql: tainted variable, Vulnerable code: sql". No network request is issued (verified in DevTools: only Datadog telemetry), so the block is purely browser-side and produces no server-side error to inspect. The server-side validator accepts the identical files — the MCP agent wrote all of them successfully into the same thread. Because the commit never succeeds, the branch head stays at commit ebe254af… while the thread's working tree advances; retool_list_react_app_files returns the new content, but publish builds the stale head and replays a build error from content that no longer exists (it cites a column offset matching a line that was edited hours earlier). Two asks:
Would be nice if the file editor flagged the files (color or icon) that are okay versus the ones the Retool Agent modified or the ones with problems that need fixing.
