dsh-secret-paste

A DeepSeek Harness plugin that detects secrets pasted into the composer, stores them in the official credentials seam, and sends a [secret:REF] placeholder to the model instead of the plaintext. The real value never reaches the message, the session history, or any transcript.

Why

You paste an API key or a token into the chat. Without protection it goes straight to the model and into the session log — which is exactly where a secret should never land. This plugin intercepts the paste, stores the value in the official DSH credential seam (~/.dsh/.credentials.yaml, mode 0600), and replaces it in the draft with [secret:REF]. The model only ever sees the placeholder, and can read the real value on demand through a dedicated tool.

How it works

The plaintext value lives in exactly two places: (1) this tab's memory, and (2) ~/.dsh/.credentials.yaml (mode 0600). Every other surface — the sent message, session history, transcripts — sees only [secret:REF].

Installation

Prerequisites

• A DeepSeek Harness Web profile (the examples use web).
• The dsh CLI and pnpm on your PATH (dsh plugin is a thin pnpm forwarder).

Install from the npm registry (recommended)

dsh plugin --profile web add dsh-secret-paste

What this command does:

1. Runs pnpm add dsh-secret-paste inside the profile directory.
2. Installs the single runtime dependency (@sanity-labs/secret-scan).
3. Because the package declares dsh.bundle.patch, it is auto-registered in dsh.profile.bundles as a profile layer.

No build step runs on install — the browser bundle is prebuilt and shipped.

Restart

The server half (the secret_resolve tool) loads at boot, so restart the dsh web process after installing, then force-refresh the browser tab (Cmd+Shift+R / Ctrl+Shift+R) to load the client bundle:

# restart the dsh web process you manage (Ctrl-C + relaunch, or your process manager)

Verify

• Settings → Plugins lists dsh-secret-paste.
• Paste a GitHub token (ghp_...) into the composer: it is immediately replaced by [secret:PASTE_N] and a chip appears.

Installing from an agent or in automation

The exact same steps work when an AI agent or a CI job installs the plugin:

1. dsh plugin --profile web add dsh-secret-paste
2. Restart dsh web (required for the secret_resolve tool to register).
3. Force-refresh the browser tab.
4. Verify: paste ghp_... and expect [secret:PASTE_N], or confirm the served page's window.__DSH_BOOT__ contains a dsh-secret-paste entry.

Usage

Auto-detect and hide

Paste text that contains a high-confidence secret (ghp_..., sk-proj-..., JWT, …). It is stored and replaced with [secret:PASTE_N] immediately, and a chip shows:

🔒 Hidden · · PASTE_N · Undo

• Hover the chip to reveal the value in a tooltip — no layout shift.
• Undo restores the plaintext while you are still drafting.
• After you send, the chip stays visible but the undo button disappears; once the answer arrives, the chip is removed.

Medium confidence

confidence === 'medium' hits (e.g. a Bearer <token>) stay in the draft as-is and a "Suspected secret" chip asks you to confirm (Hide) or ignore (Ignore).

Manual marking

Formats the detector does not recognize (ark-..., some sk-...) are never guessed. Select the text and use the "Mark selection as secret" action, then "Hide & store".

Nested placeholders

A selection that already contains [secret:REF] can be wrapped again. The secret_resolve tool resolves such chains recursively down to plaintext (cycles or missing inner refs return found: false).

Model side: secret_resolve

When the model needs the real value, it calls the secret_resolve(ref) tool:

• Returns { found, value, source }.
• Resolves nested placeholders recursively to plaintext.
• The value is SENSITIVE: the tool description instructs the model never to echo, repeat, or write it into a reply, a file, a command, or a tool argument.

Security model

• Values never leak: only [secret:REF] appears in the message, history, or transcripts. Values exist only in tab memory and the 0600 credential file.
• Detection is conservative: high auto-hides, medium waits for a confirm, and unknown formats are never guessed (manual marking is the fallback).
• Collision-safe refs: if a ref is already configured, the next free ref is used instead of overwriting; the same value is stored once per session.
• No reveal endpoint: the credential seam reads value-free with no enumeration; after a refresh the in-memory value is gone and chips do not persist.
• Use-time visibility (v1): secret_resolve brings the value into model context only when the model asks for it.

Development

node scripts/build.mjs   # rebuild lib/client.js (no external bundler)
npm test                 # node --test tests/*.test.mjs

Structure

dsh-secret-paste/
├── package.json           # dsh.bundle.patch + dsh.client declarations
├── cordis.patch.yml       # mounts the server row
├── lib/
│   ├── index.js           # server: secret_resolve tool (recursive resolution)
│   └── client.js          # prebuilt browser bundle
├── src/
│   ├── resolve.js         # nested-placeholder resolver (shared with server)
│   ├── scan.js            # detection helpers (shared with tests)
│   └── client/index.js    # paste interception, chips, credentials.set
├── vendor/secret-scan.cjs # vendored @sanity-labs/secret-scan@1.1.0 (MIT)
├── scripts/build.mjs      # assembles lib/client.js
└── tests/                 # node:test unit tests

License

MIT. vendor/secret-scan.cjs is the compiled dist/index.cjs of @sanity-labs/secret-scan v1.1.0 (MIT), whose rules derive from gitleaks (MIT) and TruffleHog detectors; its license is preserved at vendor/secret-scan.LICENSE.