dsh-automode

🌐 简体中文: README.zh.md · English: README.md

Claude Code-style auto mode for DeepSeek Harness.

This is a guardrail plugin. It intercepts agent tool calls before execution and blocks actions that match deterministic deny rules, or the auto-mode classifier's block decision.

It is not a sandbox. The plugin runs in the DSH process, and a determined malicious plugin can do anything your user account can do. Use this to reduce unsafe autonomous tool use, not as an OS security boundary.

Install

dsh plugin add @log.li/dsh-automode

From a local checkout:

dsh plugin add ./path/to/dsh-automode

Restart dsh web after installing. The permission picker (bottom-left of the chat box) will show Auto mode alongside read-only / workspace-write / danger-full-access.

Commands

/auto           # switch this session to auto mode
/auto-status    # show diagnostics: preset, approval policy, breaker state

How it works

Tool call arrives
  │
  ├─ [pre-execute gate]  (all tools, first defense)
  │    ① Read-only tools → allow (unless deny matched)
  │    ② Deny rules (regex) → hard reject
  │    ③ Allow rules (prefix glob) → approve
  │    ④ In-tree file ops → approve (allowInsideWorkingDirectory)
  │    ⑤ Escalation intent → classifier pre-screen
  │    ⑥ Everything else → pass through
  │
  └─ [approval waterfall]
       ① Soft deny rules (prose) → reject
       ② Soft allow rules (prose) → approve
       ③ Read-only allowlist → approve
       ④ Verdict cache hit → reuse (no second LLM call)
       ⑤ Classifier (two-stage: one-token filter → structured review)
       ⑥ Failure → fail-closed

🖱️ Interactive version: docs/auto-mode-flow.html — pan/zoom, relationship tracing, dark mode. Diagram source: docs/auto-mode-flow.workflow.json.

The pre-execute gate intercepts ALL tool calls (including those inside the workspace sandbox that would never trigger the approval waterfall). The approval waterfall only runs for calls that actually need sandbox escalation. The pre-execute gate only applies to auto-mode sessions; in other presets (read-only / workspace-write / danger-full-access) it is a no-op so it never contradicts the sandbox the user chose.

Rules

The rule system has two layers:

Hard boundary (deterministic, never goes to classifier)

• deny — regex patterns that hard-reject. First match wins. Evaluated before everything else. Use for exfiltration, secrets, sensitive targets, dangerous commands.
• allow — prefix-glob patterns that approve without any LLM call. Evaluated after deny. Use for routine commands you trust completely.

Classifier guidance (prose, fed to the LLM)

• rules.deny — soft-deny descriptions. The classifier reads these as standing rejections. Can be overridden by direct user intent or a matching allow rule.
• rules.allow — soft-allow exceptions. The classifier reads these as standing approvals that override matching soft-deny rules.
• rules.environment — context facts (trusted repos, infrastructure, cloud buckets). The classifier uses these to judge whether an action is within the user's environment.

All rules.* arrays support $defaults: using ["$defaults", "my custom rule"] keeps the built-in rules while adding yours. Omitting $defaults replaces the entire built-in list for that section.

Configuration

Configuration goes in your profile's cordis.patch.yml. Everything has defaults; a bare {} config is valid.

- id: auto-mode
  name: dsh-automode
  config:
    # --- Hard boundary ---

    deny:
      - exfiltrat
      - 'curl\s+[^|]*\|\s*(?:ba)?sh'
      - authorized_keys
      # ... regex patterns

    allow:
      - 'trash *'
      - 'echo *'
      - 'git status'
      - 'ls*'
      # ... prefix globs

    readOnlyTools:
      - read
      - glob
      - grep
      - list
      - search

    allowPaths:
      - '~/Documents/'
      - '/tmp/'

    allowInsideWorkingDirectory: true

    # --- Classifier ---

    classifier:
      provider: ''             # empty = follow the session's active model (request header)
      model: ''                # empty = follow the session's active model (request header)
      maxTranscriptMessages: 40
      maxTokens: 2048
      temperature: 0
      reasoningLevel: low      # low / medium / high
      askFallback: false       # true = three-state (allow/ask/reject)

    rules:
      deny: ['$defaults']
      allow: ['$defaults']
      environment: ['$defaults']

    # --- Runtime ---

    failClosed: true           # reject on classifier failure
    preExecuteGate: true       # enable the pre-execute gate
    timeoutMs: 45000           # classifier call timeout
    classifyContextChars: 6000 # context budget for task alignment
    breakerConsecutive: 3      # consecutive DENY to trip breaker
    breakerTotal: 20           # total DENY to trip breaker

Key options

Permission preset icon

The auto-mode permission preset shows a bolt glyph in the permission picker. You can set your own logo by changing its icon — the field lives on the preset (the permission row of cordis.patch.yml, not auto-mode's own config) and is an SVG path drawn inside the shared shield outline:

- id: permission
  config:
    presets:
      auto-mode:
        sandbox: workspace-write
        approval: ask
        name: Auto mode
        description: ...
        icon: '<your-svg-path-d>'   # default bolt: 'M9.15 3.4L5.85 8.55H7.95L7.05 12.6L10.45 7.25H8.25L9.15 3.4Z'

Whether it shows. The picker draws a preset's icon only when the DSH reads preset icons. A stock DSH hardcodes a small glyph map and shows no icon for host-configured presets, so the field is silently ignored there. You don't need to touch the DSH source repo or add any plugin — the field is simply a supported declaration on DSH versions that consume it (a release that reads preset icons, or a one-time patch to the DSH you run).

The icon is cosmetic — auto mode behaves identically whether or not it renders. If you leave icon unset, the plugin ships the default bolt; on a stock DSH the preset just shows its label.

Two-stage classifier

The classifier uses two stages to minimize LLM cost:

1. One-token filter (~1 token): asks the LLM for a single digit (0 = safe, 1 = needs review). Uses a generous token budget and robust digit parsing so reasoning models aren't starved, and honors classifier.reasoningLevel as the reasoning effort. Most routine actions return 0 and skip stage 2.
2. Structured review: only runs when stage 1 flags the action. Returns a full verdict with reason.

This means most tool calls cost ~1 token of classifier overhead. Only borderline actions incur the full classifier cost.

Circuit breaker

When the classifier denies 3 actions in a row or 20 total in a session, the breaker trips and auto mode pauses. The approval waterfall delegates to the human answerer. Any human decision (allowing OR rejecting the action) resumes auto mode and resets all counters — a human in the loop breaks the silent-denial loop the breaker exists to catch. If the human cancels the request or no answerer is available, the breaker stays tripped.

At the moment the breaker trips, the plugin injects a hint telling the model to request danger-full-access sandbox escalation directly on its next attempt (which surfaces the human approval window immediately), instead of the "try at current level → hit a denied error → then escalate" round-trip.

Classifier failures (timeout, parse error, empty response) are NOT counted toward the breaker.

Denial guidance & diagnostics

When an action is denied, the model is told to try a safer alternative. If no safer alternative exists, it is instructed to stop retrying and ask the user for explicit permission — a denied action will keep failing, and only explicit user approval lets a later attempt pass (the classifier weighs the user's recent explicit intent via <recent_user_intent>).

Classifier failures are logged to the DSH log (not the decisions log) with the resolved route (provider/model), the underlying error code/message, and the raw model output. This makes a recurring classifier returned no verdict diagnosable — the usual cause is the classifier inheriting the session's heavy reasoning model, whose chain-of-thought starves the classifier token budget (or a slow/free gateway timing out).

Verdict cache

Classifier verdicts are cached per session by tool + command signature. If the same action is requested again (e.g., from the approval waterfall after a pre-execute classify), the cached verdict is reused without a second LLM call. Cache entries expire after 5 minutes.

Logging

All decisions are logged to ~/.dsh/auto-mode/decisions.jsonl (JSONL format, append-only, survives restarts). Each entry includes:

• at — ISO timestamp
• event — decision / pre-execute-deny / pre-execute-allow / breaker / resume / boot
• outcome — allowed-once / rejected / cancelled
• tool — tool name
• tier — deny / allow / classify:monitor / classify:cache / classify:fail / ...
• detail — human-readable reason
• sessionId — session identifier

Use the review script to analyze the log and identify rule optimization opportunities:

node scripts/auto-mode-review.mjs

System prompt shadowing

When auto mode is active, the plugin shadows the approval-policy system prompt so the model sees "auto" instead of "ask". This tells the model that tool rejections come from the automated reviewer, not from a human. The model adjusts its retry strategy accordingly (try a smaller/safer action instead of asking the user).

Architecture

src/
  index.ts         Main entry: preset management, approval answerer, breaker reset, commands, system-prompt shadowing
  config.ts        Config schema + $defaults mechanism + built-in rule lists
  bands.ts         Deterministic band engine (deny regex + allow glob)
  pre-execute.ts   Pre-execute gate (first defense; real-path trust, classifier pre-screen, breaker-trip hint)
  classifier.ts    Two-stage classifier (+ robust parser, reasoning effort, diagnostics)
  rules.ts         Prose rule matching for the classifier
  prompt.ts        Classifier prompt construction (<recent_user_intent> + intent weighting)
  cache.ts         Verdict cache (shared across enforcement points)
  breaker.ts       Circuit breaker (3 consecutive / 20 total)
  log.ts           Shared appendDecision JSONL logger

License

MIT