dsh-vision-tool

dsh-vision-tool

A vision tool for DeepSeek Harness agents: sends a local image to the OpenAI-compatible vision endpoint you configure and returns a text description — for when the current session model can't ingest images directly.

The plugin ships no built-in defaults: provider (baseURL), credentials, and models are all configured at install time. If configuration is incomplete the plugin loads normally but does not register the tool (log line shows what's missing) — it can never crash the host.

Features

• Agent-callable tool: vision image_path=/tmp/shot.png
• Any OpenAI-compatible /chat/completions endpoint (vendor-agnostic)
• Credential resolution chain: inline apiKey → apiKeyEnv (environment variable) → ~/.dsh/.credentials.yaml same-name key
• Optional cross-check: query several models and merge answers (cross_check=true) to guard against hallucinated descriptions
• maxTokens / maxImageBytes limits (optional)
• Requests go out via a curl subprocess — inherits host proxy env vars and avoids Cloudflare 403/1010 on default urllib/undici user agents

Install

Tarball

tar xzf dsh-vision-tool-install.tar.gz && cd dsh-vision-tool-install

# interactive: just run it and answer the prompts
bash install.sh --restart

# or parameterized (scriptable / repeatable deploys)
bash install.sh --restart \
  --vision-base-url https://example.com/v1 \
  --vision-api-key-env MY_VISION_KEY \
  --vision-model gpt-4o \
  --vision-models gpt-4o,qwen-vl-max,kimi-latest \
  --vision-max-tokens 2000

npm

dsh plugin --profile web add dsh-vision-tool
# then configure: edit the dsh-vision-tool config section in ~/.dsh/profiles/web/cordis.patch.yml

Install parameters (--vision-*)

No parameters + interactive terminal → guided prompts. No parameters + non-interactive → empty config written (tool not registered); re-run with parameters, or edit the config section in ~/.dsh/profiles/web/cordis.patch.yml.

Usage (inside the agent)

vision image_path=/tmp/shot.png
vision image_path=/tmp/shot.png question="What text is in this image?"
vision image_path=/tmp/shot.png model=gpt-4o
vision image_path=/tmp/shot.png cross_check=true   # requires --vision-models configured at install

Security

• No built-in keys — credentials come only from your install parameters / environment / credential file
• ~/.dsh/.credentials.yaml should be chmod 600

Troubleshooting

• Tool doesn't appear: log shows [dsh-vision-tool] missing ... — re-run install.sh with --vision-* parameters and restart
• Credential resolution fails: at least one of apiKey / apiKeyEnv (incl. credential-file same-name key)
• Endpoint incompatible: baseURL must be OpenAI-compatible (/chat/completions, Authorization: Bearer)
• Empty output from reasoning models: configure --vision-max-tokens 2000 or higher

License

MIT. Chinese documentation: README.zh.md

---

Part of dsh-plugins — a small monorepo of DSH plugins: dsh-lan-gateway, dsh-vision-tool, dsh-mobile-ui.