dsh-secret-paste

You paste an API key or token into the chat composer and it flows straight into the model's context, the session log, and every transcript—exactly where a secret should never land. dsh-secret-paste intercepts that moment: a detection engine identifies the credential, writes the real value into a local credentials file with 0600 permissions, and replaces it in your draft with a [secret:PASTE_N] placeholder. From that point on, messages, session history, and transcripts contain only the placeholder; the plaintext lives in exactly two places—this tab's memory and the credential file.

Detection is deliberately conservative. Drawing on 1,100+ rules derived from gitleaks and TruffleHog, the plugin splits hits into two tiers: high-confidence secrets (ghp_, sk-proj-, JWT, and similar) are auto-hidden immediately with a dismissible chip; medium-confidence matches (e.g. a Bearer prefix) surface a confirm-or-ignore prompt so you stay in control. Formats the detector cannot classify are never guessed—a manual "Mark selection as secret" action is provided as the fallback. On the model side, a dedicated secret_resolve tool returns the real value only when the model explicitly asks for it, and the tool's description instructs the model never to echo, repeat, or write the value into a reply, file, command, or tool argument, adding a second layer of leak prevention.

If you regularly handle tokens, API keys, JWTs, or other sensitive credentials inside DeepSeek Harness Web and want them kept out of plaintext conversations, session logs, and model context, this plugin gives you a local, zero-extra-server-route protection layer with a conservative, human-in-the-loop detection model.