CoalWash Privacy Policy

CoalWash collects nothing and phones nowhere — and it treats your memory as the private data it is.

• No telemetry. No usage data, analytics, or identifiers are collected, stored, or transmitted — by the hook, the engine scripts, or the skill.
• No network calls. The hook and every engine script are offline by design (Phoenix #7): local filesystem only, no sockets, no requests. (The self-update check is the agent's /coalwash:update procedure, run only with your consent — never the hook.)
• Your memory IS the data under management — CoalWash minimizes its movement. Files are processed in place; backups (snapshots) stay in the project's own .claude/coalwash/ dir and are never transmitted. That dir self-ignores by construction — the engine drops a catch-all .gitignore inside it, so snapshots stay out of version control even in a project that tracks .claude/. The only party that ever reads memory content is your own session's model — the same trust boundary your platform already established by loading those files into context.
• localOnly is the trade-secret mode — an agent-honored contract, not a code-enforced transmission block. Set localOnly: true and the SKILL contract runs mechanical Quick only, skipping the semantic tier entirely (no content-bearing sub is spawned) — but no executable intercepts a tool call; the no-sub behavior depends on the agent honoring this setting, the same as its memory-is-DATA-never-instructions rule. What IS code-enforced: the flag itself can't be weakened by a project config once a global localOnly: true is set (mergeSafety, config-load.mjs).
• Receipts are metrics, never content. Every receipt and gauge line carries sizes, counts, and estimates — never memory-content snippets. Token figures are a local char-heuristic estimate, labeled ~est, not a platform-verified read.
• Error reports are manual and scrubbed. When something misbehaves, your agent may offer to open a GitHub issue; nothing is submitted automatically, you see and edit the contents first, and the report template carries mechanical facts only (operation, counts, error class) — never the memory text.
• Local files only. All state lives in files you can read: the caliper state ~/.claude/projects/<slug>/coalwash/state.json (per-project lean floor, session timestamps, the last recorded gauge verdict, and the pending once-per-crossing ask state — numbers, no content, no time-based snooze), the transaction dir [project]/.claude/coalwash/ (lock, WAL journal, snapshots, retention bins, keeps.json keep-verdicts — copies of your own files and your own keep decisions, on your own disk), the config (~/.claude/.coalwash.json global, optional per-project override under [project]/.claude/coal/coalwash.json or the legacy [project]/.coalwash.json — see README's Configure section for the full read order), and the self-update throttle stamp ~/.claude/coal/coalwash/update-check (a timestamp, nothing more).

Questions: open an issue at https://github.com/TheColliery/CoalWash/issues.