dsh-model-vision

dsh-model-vision

Lets DSH recognise which models can actually see images.

Models that accept images get the "Image" input type checked automatically, and a · 可看图 marker after their name — no clicking through every model by hand, and it survives the next "Fetch available models".

No extra CLI to install, no second API key, no need to switch gateways. It only puts back a capability your gateway already advertises but DSH dropped.

You may be hitting this

• You paste an image and DSH answers "this model does not support image input", or the image is silently ignored
• Your gateway's docs say a model takes images, but DSH shows no "Image" option for it
• You checked "Image" by hand once, and the next "Fetch available models" cleared it

Why: when DSH reads a model catalog it keeps only the name, context window and output limit — it drops the modality. Every fetched model therefore falls back to text-only, and an image request is rejected before it leaves the client.

This plugin puts that field back.

Good fit: you use a third-party gateway / relay / self-hosted server (any OpenAI-compatible endpoint) and want image input to work. Not a fit: the model genuinely cannot see images. This plugin will not teach a text-only model to see.

What it does not do (so you can install it without worry):

• It does not make a text-only model see images
• It does not fill in API keys, and does not touch model names (except that trailing marker), context windows or pricing
• It does not guess a capability when the gateway says nothing
• It contacts no third-party host, uploads no conversation content, and has no telemetry

Install

Restart DSH afterwards.

dsh plugin --profile web add dsh-model-vision

This also registers the plugin in dsh.profile.bundles; no config editing needed. Replace web with your profile name if you use a different one.

You can also install straight from GitHub if you want to track the repo: dsh plugin --profile web add github:charovo0059/dsh-model-vision

Once it is listed you can also use Settings → Plugins → Plugin Market → search dsh-model-vision → Install.

From a local clone (when you are changing the code)

git clone https://github.com/charovo0059/dsh-model-vision.git
dsh plugin --profile web add link:<absolute path to the clone>

Self-check after installing

If you are unsure which step is missing:

node tools/check.mjs --profile web

See what it would change before installing

node tools/preview.mjs --profile web

Read-only: reads your config, asks the gateway once, prints the lines it would change — writes nothing. Safe to run while DSH is open.

Usage

It works in the background. You do not have to click anything.

It reconciles at three moments:

To check right now:

Settings → Models → scroll to the bottom of the model list → click Detect and fill in vision capability

The result appears in place (one sentence plus an expandable detail list).

How it decides

1. Read what the gateway says first — if it declares a modality, that wins. Cheap, and costs nothing.
2. Only probe the models the gateway says nothing about — send a solid-colour image and ask what colour it is, over two different colours:
   • Names a colour it was actually shown → it can see images
   • Names a different colour (shown red, answers "purple") → it cannot see images, recorded as text-only
   • Names no colour at all (answers with a paragraph) → inconclusive, nothing is written

Silence is not the same as text-only. If the gateway is silent and the probe is inconclusive, the plugin writes nothing — better nothing than wrong.

Questions you probably have:

Supported gateways

Anything DSH can use as a provider should work — it does not assume paths or field names.

For the exact field names it understands, see docs/how-it-works.md.

How to tell it is working

Most direct: Settings → Models → open a model you know accepts images → "Image" is already checked under Input types.

Read the sentence it leaves behind: after every run it writes a record. The message line is the human-readable one:

{
  "message": "扫了 58 个模型（网关报了 191 个），补了 34 个的「输入类型：图片」"
}

There are only a few possible messages:

The record lives at ~/.dsh/model-vision.<profile>.status.json (for the web profile that is model-vision.web.status.json). For the detailed fields (wrote / routes[].changes / credential) see docs/troubleshooting.md.

Will it break my config

• It writes one field only — the input type — and stops there
• Nothing to change → not a byte is written
• It compares line by line before writing, so comments, field order and !!js expressions survive
• If something ever goes wrong: back up ~/.dsh/profiles/web/cordis.patch.yml first, restore it to roll back
• To watch without writing: set dryRun: true; it still records, but touches no file

What it touches:

No third-party host is contacted, no conversation content is uploaded, no session history is read, and there is no telemetry.

Troubleshooting

Uninstall

dsh plugin --profile web remove dsh-model-vision

Restart DSH and it is gone — dsh plugin remove also drops it from dsh.profile.bundles, so no manual config editing is needed.

The "Image" checkboxes that were filled in stay — they are now ordinary entries in your own config and the plugin does not roll them back. Uncheck them in the settings page if you want them gone.

Configuration (usually unnecessary)

It works out of the box. To tune it, add an override in ~/.dsh/profiles/web/cordis.patch.yml. The name line must be exactly dsh-model-vision, otherwise the entry is skipped silently:

- id: model-vision
  name: dsh-model-vision
  config:
    mode: fill               # fill = only fill gaps (default); sync = also correct a stale declaration
    routes: []               # [] = all providers; or list specific route names
    dryRun: false            # true = look only, never write
    intervalMs: 300000       # reconcile every 5 minutes
    watchMs: 5000            # check the composed config every 5 seconds
    badgeSuffix: ' · 可看图'  # marker after the name; '' = no marker
    probe: false             # true = probe in the background too (default probes only on the button)
    probeRounds: 2           # colours asked per model
    maxProbes: 40            # probe cap per run
    timeoutMs: 15000         # model-list timeout
    probeTimeoutMs: 30000    # per-probe timeout
    statusFile: ''           # '' = default location; false = write nothing

Contributing

• Found a bug or want a feature → open an issue
• Sending a PR → read CONTRIBUTING.md first, then run npm test
• New gateway or field spelling it should recognise but does not → open an issue with the relevant fragment of your gateway's /models response; that is the most useful report

For implementation details (why DSH drops the field, why the marker lives in the name, the three triggers, the two write channels, the one-click endpoint, running the tests) → docs/how-it-works.md

License

MIT