dsh-vision-tool

Image routing for text-only models in DeepSeek Harness (DSH).

Text-only models (e.g. deepseek-v4-flash) cannot see images. This bundle gives
them eyes in two coordinated steps:

1. vision-prompt — shadows POST /api/session.prompt. When the active
   session model does not support image input, pasted images are persisted as
   content-addressed attachments and rewritten in place into text prompts that
   carry the full attachment reference JSON. Any other request (no image, or a
   model that already supports images) is forwarded unchanged, with the
   official /api trust fence (DNS rebinding / cross-site defense) reimplemented.
2. vision-tool — registers a global analyze_image tool. The model calls
   it with the attachment reference (or a local file path); the tool routes
   the image to a vision model and returns its text answer.

Verified end-to-end: paste an image → attachment persisted → model calls
analyze_image → vision model describes the image.

Install

Requires the dsh CLI (see package and install a plugin).

# from npm (once published)
dsh plugin --profile <name> add dsh-vision-tool

# straight from a git host (pin a commit; no build step needed — pure ESM)
dsh plugin --profile <name> add github:<you>/dsh-vision-tool#<sha>

# or from a local tarball
pnpm pack
dsh plugin --profile <name> add ./dsh-vision-tool-0.1.0.tgz

The bundle's cordis.patch.yml inserts two rows: vision-tool and
vision-prompt. Restart the profile afterwards:

dsh --profile <name>

Verify the layer landed without booting:

dsh --profile <name> --dump-config   # look for the "# == dsh-vision-tool" layer

Configuration

Credential (required)

The tool resolves KIMI_CODE_API_KEY — from $DSH_HOME/.credentials.yaml or an
environment variable of the same name. Get the key from your
Kimi Code subscription page (sk-kimi- prefix).

# $DSH_HOME/.credentials.yaml
KIMI_CODE_API_KEY: sk-kimi-...

Switching vision models

The defaults target kimi-for-coding at https://api.kimi.com/coding/v1.
Override the row in your profile's cordis.patch.yml (a patch replaces the
whole config, so restate every key you keep):

- id: vision-tool
  name: dsh-vision-tool
  config:
    baseURL: https://api.kimi.com/coding/v1
    model: kimi-for-coding
    apiKeyEnv: KIMI_CODE_API_KEY
    maxImageBytes: 20971520
    timeoutMs: 120000

Note: kimi-for-coding only accepts temperature: 1 (anything else is
rejected with HTTP 400). The tool hard-codes temperature: 1 as its default
and is not configurable for this model. Other OpenAI-compatible vision
endpoints generally work as long as they accept temperature: 1.

Supported inputs

• attachment — full reference JSON injected by the paste-rewrite mechanism
  ({"attachmentId":"sha256:...","mediaType":...,"bytes":N,"width":N,"height":N}).
  Pass it verbatim; do not strip fields.
• path — local image file (absolute, or relative to the session cwd).
• Formats: png / jpg / jpeg / webp / gif. Local files up to
  maxImageBytes (default 20 MB). Attachments are bounded by the harness
  attachment store limits.

Security

• vision-prompt reimplements the official /api trust fence: loopback /
  trustedHosts host check, sec-fetch-site and Origin checks.
• Request bodies are capped at 160 MB (413 otherwise), matching the harness
  http-bridge default.
• Any failure degrades to passthrough — the original request is forwarded
  unchanged, never swallowed or mangled.
• The tool only reads the attachment you reference and your configured
  credential; it never stores prompt or image content beyond the attachment
  store the harness itself maintains.

Diagnostics

Both plugins append to $DSH_HOME/vision-trace.log:

handle: rewrite result = REWRITTEN
vision-tool: execute: resolve KIMI_CODE_API_KEY -> source=file len=72

License

MIT