Security Policy

Threat model & data handling

RegressionLedger is designed to be safe to run on a hook surface that can read your code and deny tool calls. Its guarantees:

Local & offline. It makes no network calls, requires no API keys, and sends nothing anywhere. Everything runs on your machine.
No raw secrets stored. The ledger (.regressionledger/ledger.json) holds normalized tokens (string and number literals are abstracted to STR/NUM) plus a short ~120-char preview of each change and a parsed error signature — never your environment, keys, or full files. The ledger is git-ignored by default.
Fails open. If the hook hits any internal error it exits cleanly (exit 0, empty stdout) and never blocks or corrupts the agent's tool call. A guardrail that breaks the thing it protects is worse than no guardrail.
Zero dependencies. No third-party runtime packages, so there is no transitive supply-chain surface.

Every claim above is verifiable in src/ — start with src/hooks.js, src/ledger.js, and src/cli.js (cmdHook is the fail-open boundary).

Please report security issues privately via GitHub:

Repo Security tab → Report a vulnerability (private advisory).

Do not open a public issue for security problems. We'll acknowledge the report, investigate, and coordinate a fix and disclosure.

This project is pre-1.0; security fixes target the latest release on main.