dsh-secret-scrub

In AI terminal and Agent workloads, API keys, bearer tokens, and private key blocks slip into persistent session logs or ride along in model requests all too easily—once leaked, the damage is irreversible. dsh-secret-scrub is a guard plugin built for exactly this gap: it intercepts text on its way to the log and the model, detects secret-shaped fragments against configurable rules, and irreversibly replaces them with a [REDACTED:category] placeholder before they are written or transmitted.

The plugin ships a three-tier rule set. Tier 0 (always active, non-disableable) covers a dozen-plus mainstream credential formats—AWS, GitHub, OpenAI, Anthropic, Stripe, and generic bearer/JWT tokens. Tier 1 adds environment-variable-style cloud provider keys at the balanced level. Tier 2, available only at aggressive, extends to email, phone numbers, national ID numbers, Luhn-checked credit cards, and a high-entropy fallback. Three waterfall listeners—user-message pre-step, tool-result post-execute, and code-dispatch log—guarantee that every write path is scrubbed. A pure-function scrubbing engine is also exported for use outside any Cordis runtime.

It is designed for teams operating AI terminals, headless Agent pipelines, or any Cordis-managed context where sensitive text must not appear in durable logs or outbound model calls. If you want placeholder-only logs, credential-free model requests, and a zero-dependency regex engine without building your own pipeline, dsh-secret-scrub is a drop-in, configure-and-go solution.