← ClaudeAtlas

data-handlinglisted

Read before changing capture, indexing, redaction, learning, skill publication, engine execution, Actions bot setup, or any path that stores or sends user data.
theonly1me/shadowclone · ★ 26 · AI & Automation · score 82
Install: claude install-skill theonly1me/shadowclone
# Data handling Follow the code conventions in `CONTRIBUTING.md`. Preserve the boundaries documented in `docs/data-handling.md`. ## Consent and scope Every capture source has a named flag, off by default, and an entry in `docs/data-handling.md`. Widening locations or content counts as a source change. Transcript access does not enable Git metadata, and reading skills does not authorize editing them. Before consent, source discovery may return one ephemeral boolean for whether a configured root exists and is non-empty. Do not open entries or retain their names, paths, counts, timestamps, or provider identifiers. Preserve repository and remote-owner boundaries. Unknown origins stay isolated. Global guidance requires explicit global evidence or a user decision. Managed policy limits cannot be widened by user configuration. ## Learning input Events carry pointers. `resolveRedacted` is the learning-text boundary: eligible excerpts pass through `redactSecrets` when materialized. Do not add competing downstream redaction gates. Derive fingerprints, metadata, and model-facing text from the same bounded source snapshot. Exclude tool results, tool-returned file contents, thinking blocks, and data-access results from learning. Redaction reduces exposure without guaranteeing anonymity. Removing a redaction pattern needs a documented reason. New sources need integration tests proving the adapter-to-redaction path, including excluded fields and synthetic secrets. Demonstrate that