dshbase

Blog · Guide

DeepSeek Harness FAQ

August 14, 2026 · dshbase

These are the questions people are actually asking in the DeepSeek Harness community — not the ones in the docs. Every answer below comes from real usage on dsh 0.1.0-rc.6, and from answering these exact questions in the official discussions.

1. Why is the permission system so "dumb"?

The most common complaint. dsh doesn't ask for permission when you use git/node — it uses permission presets: a fixed policy decided at startup, not per-tool prompting. If git is blocked, it's because the current preset put it in the "approval" or "deny" tier.

Fix: change the permission preset in Settings, or install dsh-tool-approval for true per-step manual approval. Both are in the plugin directory.

2. Why can't I paste images or documents?

Dragging an image and getting "switch model" means the current model doesn't declare vision modality (its input doesn't include image). dsh won't silently convert an image to text.

Fix: either add input: [text, image] for the model in settings.yaml, or use a vision plugin — dsh-vision-toolkit (image Q&A, OCR, chart parsing for text-only models) or dsh-drop-to-path (drag-to-path).

3. Token usage explodes on long sessions — how do I cut cost?

Three levers that actually work:

  • PTC mode — dsh's prompt-token-cache mode is built for DeepSeek's prefix caching and sharply cuts repeated-prefix cost.
  • Compact early — use /compact manually rather than waiting for auto-compaction.
  • Fewer plugins — every tool's description consumes tokens; install only what you need.

We measured the numbers in our cost & caching analysis.

4. Fresh install crashes with "--expose-internals is required"

A cordis HMR-loader issue on clean environments (macOS arm64 / NixOS / Linux). Workaround:

node --expose-internals $(which dsh) web

Or set NODE_OPTIONS=--expose-internals. This and every other install error we've hit are in the troubleshooting guide.

5. Is there a Docker image?

No official image yet. But dsh is plain Node.js — Node 18+ and npx @deepseek-ai/dsh web is all you need, and headless mode drops neatly into a container or CI. Community distributions like oh-dsh bundle TUI + desktop + web together. See the install guide.

6. What about long-term memory?

The memory ecosystem has exploded in week one — several approaches, all in the directory:

  • dsh-memory — cited memory over the lossless session log
  • dsh-mnemon — runtime + retrievable archive + supervised memory
  • sage-mem — Chinese-first full-text search memory
  • dsh-sgme — multi-agent shared memory via HTTP

7. Will there be a desktop app?

Not officially, but the TUI and oh-dsh distribution cover the "native-ish" experience. The web UI is the primary surface.

Have a question we didn't cover? Ask in the official community.

All articles →

🌐 中文