Skip to content

README.md

Latest commit

 

History

History
97 lines (64 loc) · 3.28 KB

README.md

File metadata and controls

97 lines (64 loc) · 3.28 KB

Gradata — Claude Code Plugin

AI that learns your judgment. Every correction you make becomes a rule that sharpens future output.

Prerequisites

This plugin is a thin wrapper around the Gradata Python SDK. You must install the SDK first:

# Python 3.11+ required
pipx install gradata

Verify the install before enabling the plugin:

python -m gradata --version

If python is not on your PATH, or maps to Python < 3.11, fix that first — the plugin's hooks shell out to python -m gradata.hooks.* and will fail silently otherwise.

Install

/plugin marketplace add Gradata/gradata
/plugin install gradata

Claude Code will load four hooks on your next session:

Event Hook What it does
SessionStart gradata.hooks.inject_brain_rules Injects graduated rules into session context
UserPromptSubmit gradata.hooks.context_inject Injects brain context per user message
PostToolUse gradata.hooks.auto_correct Captures Edit/Write corrections as lessons
Stop gradata.hooks.session_close Emits SESSION_END + runs graduation sweep

Configure

Gradata stores its brain at $GRADATA_BRAIN_DIR if set, otherwise ./brain/ in the current working directory. To initialize a brain for the current project:

gradata init ./brain

To use a shared brain across projects:

export GRADATA_BRAIN_DIR=~/.gradata/default-brain
gradata init "$GRADATA_BRAIN_DIR"

Full SDK docs: https://gradata.ai

Uninstall

/plugin uninstall gradata

This removes the hooks from your Claude Code session. Your brain data (corrections, lessons, rules) stays on disk at $GRADATA_BRAIN_DIR. To fully remove:

pipx uninstall gradata
rm -rf ./brain         # or $GRADATA_BRAIN_DIR

Troubleshooting

Nothing seems to happen. The hooks fail silent by design so they never block Claude Code. Check:

python -m gradata.hooks.inject_brain_rules < /dev/null

If you see ModuleNotFoundError: No module named 'gradata', pipx install gradata did not succeed or python is pointing at a different interpreter than the one pipx used. Run pipx list to see where Gradata lives, then ensure that interpreter is first on your PATH.

Corrections are not being captured. Only Edit and Write tool calls trigger auto_correct. Plain chat edits (without a tool call) are not captured yet.

Rules are not being injected. Rules only inject once they've graduated past the PATTERN threshold (0.60 confidence). Check graduation state:

gradata rules --all

New corrections start at INSTINCT. It takes repeated reinforcement before a rule reaches injection-eligible confidence.

python not found on Windows. Install Python 3.11+ from python.org and check "Add to PATH" during install, or use py -m gradata.hooks.inject_brain_rules and alias python=py in your shell profile.

Where are my brain files? Run gradata doctor — it prints the resolved brain directory, lesson count, and rule graduation histogram.

Links