deepseek-autoreview

A「替我审核」(review-on-my-behalf) permission preset for DeepSeek Harness: a fourth permission mode beside read-only / workspace-write / danger-full-access. Escalation asks run through a three-tier funnel — a conservative whitelist (zero-token fast approval), a danger blocklist (zero-token interception), and an independent reviewer model for the gray zone. Safe operations auto-approve, unsafe/uncertain ones fall through to the human approval prompt, and obviously malicious asks can be denied outright. Fail-closed by construction, with a user-switchable reviewer model in General settings.

The four modes

Install

Requires a dsh profile (~/.dsh/profiles/<name>, new dsh.profile.bundles + cordis.patch.yml structure).

# npm (after publishing)
dsh plugin add deepseek-autoreview
# straight from GitHub
dsh plugin add github:Violet-Evergarden-Xiya/deepseek-AutoReview
# local checkout
dsh plugin add file:/path/to/deepseek-AutoReview

Then register the bundle in the profile's package.json (dsh plugin add only installs the dependency):

{
  "dependencies": { "deepseek-autoreview": "1.1.0" },
  "dsh": {
    "profile": {
      "bundles": ["@deepseek-ai/dsh-base", "@deepseek-ai/dsh-web-app", "deepseek-autoreview"]
    }
  }
}

Restart dsh (no hot reload for the host patch layer):

dsh web

The preset picker then shows the fourth entry「替我审核」.

Usage

Switch a session to「替我审核」:

• workspace edits run silently;
• escalations pass the funnel: structurally verifiable commands (npm/pnpm/pip install, git clone/pull/fetch, no shell metacharacters) fast-approve with zero tokens; catastrophic shapes and sensitive paths (rm -rf /, curl | sh, writes to /etc …) intercept with zero tokens; the gray zone is judged by the reviewer model;
• auto-approvals are rate-limited (5/min, 30/hour by default) — over-budget asks delegate to the human.

Switching the reviewer model in settings

After install, Settings → General gains a「替我审核模型」row, switchable at any time with immediate effect:

• Follow main conversation (default): the reviewer reuses the session's current model;
• Fixed model: fill Provider / Model (e.g. deepseek-official / deepseek-v4-flash); the host validates the route against the LLM catalog and falls back to the session route (flagged in the audit) when invalid;
• Priority: settings (fixed) > composition config > session route.

Configuration (profile cordis.patch.yml)

- id: permission-review
  name: deepseek-autoreview
  config:
    provider: deepseek-official   # optional deployment-level fallback route (paired with model)
    model: deepseek-v4-flash

How it works

• Preset table: the bundle patch overrides the permission row's preset table with review appended (workspace-write + ask).
• Three-tier funnel: whitelist (zero tokens) → blocklist (zero tokens) → reviewer model over the real tool arguments recovered from the session log by callId, the current turn's user request, and the asker's self-reported reason (least trusted).
• Answerer: a PREPENDED approval/request listener ahead of the human channel, gated on the review preset.
• Closed vocabulary: allow / refer / deny (deny rides the platform's rejected semantics).

Audit and records

• review/verdict (log-only): { toolName, callId?, decision, rationale?, source, model?, latencyMs?, evidenceSha256?, rateLimited?, routeFallback?, error? } — source names the deciding layer, evidenceSha256 fingerprints the operation.
• One concise in-conversation notice per decision.

Boundaries

• Delegated subagents keep their pinned never policy — the reviewer never sees child escalations.
• The whitelist only approves structurally verifiable commands (single verb, no ; | & < > $ = \ ( ) { }` anywhere), so composite or obfuscated commands always land in the blocklist or the model.
• The preset-table override replaces the whole presets key (the three shipped entries are restated + review); hand-merge review into custom tables.
• Network access is outside the review scope.
• Judgment quality depends on the chosen model; iterate on the rubric.

Build & test from source

pnpm install
pnpm typecheck
pnpm test      # 27 cases incl. adversarial samples (command chaining, path tricks), fake adapter, zero tokens
pnpm build     # tsc declarations → tsdown lib/ (host half + browser client bundle)

Publish:

pnpm publish   # runs prepublishOnly (typecheck + test) first