DSH Plugins Marketplace

DSH Plugins

Plugins

/

Security & Audit

/

dsh-jev-auto-approve

d

dsh-jev-auto-approve

Manifest valid

Reviews DeepSeek Harness approval requests with TypeSafe Jev and auto-approves low-risk, explicitly authorized actions; includes an English settings page for the key and token/cost usage. Each approva

UI (client)hasBundlePatch

dsh-jev-auto-approve

Jev-backed auto-approval for DeepSeek Harness (dsh).

The plugin asks TypeSafe Jev whether a pending approval request is safe to run, and answers it with allowed-once when the review passes. It does not change the session permission preset or the sandbox. Actions the plugin does not approve still reach the normal human prompt.

Experimental. Jev is an external model and can be wrong or manipulated by text in tool arguments. Auto-approval is a convenience, not a security boundary. Read SECURITY.md before using it with sensitive work.

Install

Install from npm:

dsh plugin --profile web add dsh-jev-auto-approve

Install directly from GitHub:

dsh plugin --profile web add github:Vultwo/dsh-jev-auto-approve

Or install the prebuilt release tarball:

dsh plugin --profile web add https://github.com/Vultwo/dsh-jev-auto-approve/releases/latest/download/dsh-jev-auto-approve.tgz

You can also open Plugins -> Add plugin in the Web UI and paste the GitHub URL. After installation, enable the bundle if DSH asks.

Requirements: DSH 0.2.2-rc.1 (tested) and a TypeSafe API key. The plugin uses Node 20.3+ APIs through DSH; it has no npm dependencies.

Setup

  1. Open Plugins -> Jev auto-approve -> Configure.
  2. Paste the TypeSafe key into API key and click Save key.
  3. Click Test connection. A successful test prints the Jev model and latency.
  4. Run a task. The Usage card shows reviews, auto-approved calls, calls that asked a human, input/output tokens, estimated cost, and the last review.

The key is stored in the DSH credential store as TYPESAFE_API_KEY, not in the profile patch, and the page never displays it again.

Fallbacks for headless or unattended use:

  1. Home file: printf '%s\n' 'YOUR_TYPESAFE_KEY' > ~/.dsh/jev-api-key && chmod 600 ~/.dsh/jev-api-key
  2. Environment: export TYPESAFE_API_KEY before launching DSH.
  3. Plugin config: add apiKey: YOUR_KEY to the jev-auto-approve row in cordis.patch.yml (plaintext; only if you accept that tradeoff).

Behavior

For each approval request the plugin sends Jev three calibrated yes/no questions:

QuestionMeaningDefault gate
safeThe action is reversible or clearly intended, and does not destroy data, exfiltrate secrets, force-push, or change security settings.>= 0.85
irreversibleIf the action goes wrong, it would irreversibly delete or overwrite pre-existing data, or make a hard-to-undo external change.<= 0.15
authorizedThe current user instruction or task context clearly authorizes this exact operation, target, and scope.>= 0.80
  • All three gates pass -> the plugin answers allowed-once.
  • Any gate fails, the key is missing, Jev errors, or the review times out -> the plugin delegates to the normal human approval prompt.
  • mode: shadow logs what it would do but always delegates to the human.
  • Read-only plugin_manager actions (list_plugins, list_bundles) take a deterministic fast path and are auto-approved without a Jev call. Set allowReadOnlyPluginManager: false to review them with Jev like every other action.

The plugin reviews approval requests only. Ordinary workspace operations that do not need approval are unaffected, and the plugin cannot override a never approval policy.

Configuration

The bundle is configured in cordis.patch.yml; the settings page edits the key and shows usage. All fields are optional.

KeyDefaultMeaning
enabledtrueMaster switch.
modeenforceenforce auto-approves; shadow only logs.
baseURLhttps://api.typesafe.ai/v1/systemoneTypeSafe Jev endpoint.
modeljev-latestJev model id.
apiKeyemptyFallback only; the settings page stores the key in the DSH credential store.
apiKeyEnvTYPESAFE_API_KEYEnvironment variable and credential reference fallback.
apiKeyFile~/.dsh/jev-api-keyFile fallback (one line, the key).
timeoutMs4000Per-review wall-clock budget.
safeThreshold0.85Minimum safe probability.
irreversibleMax0.15Maximum irreversible probability.
authorizedThreshold0.80Minimum authorized probability.
includeUserMessagetrueInclude the most recent user message in the review state.
allowReadOnlyPluginManagertrueAuto-approve plugin_manager actions list_plugins and list_bundles without a Jev call.
logFile~/.dsh/jev-auto-approve.jsonlAppend-only JSONL decision log.

Commands

CommandEffect
/jev-autoShow enabled/mode, key source, thresholds, counters, token totals, last review, log path.
/jev-auto testMake one real Jev request to verify the key, endpoint, and model.
/jev-auto on / /jev-auto offToggle this process only.
/jev-auto shadow / /jev-auto enforceSwitch this process only.

What is sent to TypeSafe

The review state contains the tool name, the approval reason, the tool arguments (up to 6000 characters), the working directory, and, unless includeUserMessage is false, the most recent user message (up to 1200 characters). No environment variables, credentials, or full session history are sent. The key itself is sent only in the HTTP Authorization header.

Logs

Each decision appends one JSON line to logFile:

{"ts":"2026-11-11T15:00:00.000Z","outcome":"allowed","tool":"bash","callId":"call_...","sessionId":"...","mode":"enforce","scores":{"safe":0.96,"irreversible":0.05,"authorized":0.91},"latencyMs":420,"reviewReason":"all thresholds met","keySource":"credentials:TYPESAFE_API_KEY","usage":{"inputTokens":501,"outputTokens":0,"costUsd":0.00002},"cwd":"/path/to/workspace"}

The log never contains the API key or the full tool arguments.

Development

node --test test/*.test.mjs

The tests use a local mock Jev server; no key or network is needed. The CI workflow runs the same tests on Node 24.

License

MIT. See LICENSE.

Comments

Loading…

Similar plugins

dsh-auto-review-jev

DeepSeek Harness approval guardian: a reviewer model decides permission requests before they reach you, with a System One decision-API backend (TypeSafe Jev, SiliconFlow, self-hosted Laya), a configur

Security & AuditManifest valid

★ 0

dsh plugin --profile web add dsh-auto-review-jev

by kiiiiile

AI-reviewed auto-approval for DeepSeek Harness: safe tool asks granted automatically, risky ones escalated with the review attached

Manifest valid

★ 0

MIT

TypeScript

Sep 14, 2026

dsh plugin --profile web add dsh-approval-ai-review

by Elaina-real

Tiered auto-review for DeepSeek Harness (DSH): static-rule safety net + LLM reviewer + human fallback. Auto-approve safe actions, auto-deny irreversible ones, ask a human for the rest. PURE VIBE CODING - not audited, use at your own risk.

Security & AuditTools & CapabilitiesManifest valid

★ 0

↓ 22/wk

MIT

JavaScript

Aug 14, 2026

dsh plugin --profile web add dsh-tiered-approval

by flyhai

Interception and audit frame for DeepSeek Harness: one approval ladder, hash-chained audit, result redaction.

Security & AuditManifest valid

★ 0

BSD-3-Clause

JavaScript

Nov 17, 2026

dsh plugin --profile web add dsh-auto-approval-frame

by StyxNether

Auto-Approval: a middle permission tier for DeepSeek Harness between workspace-write and danger-full-access, auto-approving harmless commands and trusted-area targets

Security & AuditManifest valid

★ 6

↓ 59/wk

MIT

JavaScript

Sep 6, 2026

dsh plugin --profile web add dsh-auto-approval-plugin

by aa2246740

Codex-aligned Auto-review and Approve for me mode for DeepSeek Harness

Security & AuditTerminal & ClientsManifest valid

★ 0

MIT

TypeScript

Nov 3, 2026

dsh plugin --profile web add dsh-approve-for-me