GITAUTO.md Restrictions

Rules in coding_standards.xml restrict what can go into GITAUTO.md: only reusable, repo-specific learnings. No task-specific details, no temporary workarounds, no project documentation. This keeps GITAUTO.md focused and useful across future sessions.

Why This Exists

Without restrictions, the model dumps entire debugging sessions into GITAUTO.md. It writes entries like "Fixed issue #123 by changing line 45 of auth.ts" or "Tried approach A, failed, then tried approach B." Over time, GITAUTO.md grows to thousands of lines of irrelevant noise that pollutes future sessions. The model reads all of it at the start of every run, wasting tokens on one-off debugging notes that will never be relevant again.

Why Models Over-Write to Memory Files

Models treat any writable file as a place to dump information. Without restrictions, the model records every debugging step, every failed approach, and every one-off fix. This is the same behavior as excessive logging - the model errs on the side of "more information is better" with no sense of what's reusable vs what's noise. Filtering requires judgment about future sessions that the model doesn't have. Models are trained to be helpful and thorough, so recording more information feels "helpful" because training rewards comprehensive responses. The concept of "this is noise, not signal" requires domain judgment that training does not provide.

How It Works

The coding_standards.xml file includes explicit rules about GITAUTO.md content: entries must be reusable patterns that apply to future sessions (e.g., "this repo uses Jest with ts-jest transformer"), not task-specific details (e.g., "fixed bug in PR #456"). Temporary workarounds and step-by-step debugging logs are prohibited. The model is instructed to think about whether each potential entry would help a future agent session before writing it.

Related Features

• GITAUTO.md - how GITAUTO.md is configured and used
• Coding Standards - the XML file that contains GITAUTO.md restriction rules