dsh-tool-vision

GitHub: Scorp1o117/dsh-tool-vision · npm: dsh-tool-vision

Part of the DeepSeek Harness Enhancement Suite — Vision · Soul/Persona · Long-term Memory · Plugin Marketplace.

External vision model for DeepSeek Harness.

DSH 0.1.1 adds native image input for DeepSeek's vision catalog. This plugin remains useful when you want a separate OpenAI-compatible vision endpoint, pixel-level image tools, screenshots, or a text-model bridge. The harness derives every model request strictly from the session log (llm/stream requests must equal the durable derivation — the agent-loop invariant), so the bridge keeps its conversion inside that durable path:

1. inspect_image tool — sends an image (local file, or http(s) URL) to any OpenAI-compatible /chat/completions endpoint that supports image_url content parts, and returns the vision model's textual answer into the agent loop.
2. Image bridge (v0.2.1) — pasted images are turned into inspect_image hints before they enter the durable log, on the agent/pre-step waterfall (the one seam where the harness lets a plugin replace the messages of a proposed step). Images already logged by an older version are repaired lazily with a surface replace on the session's first pre-step. Only models listed in multimodalModels receive image blocks directly; a model's declared inputModalities are never consulted, because profiles routinely declare input: [text, image] on text-only models just to pass the harness's prompt-admission check.

• Zero dependencies beyond the dsh SDK — works with any compatible endpoint: OpenAI GPT-4o, Qwen-VL (DashScope), GLM-4V (Zhipu), Moonshot, Gemini compatible endpoints, local Ollama, etc.
• Registered on the global tools layer: every agent in the process can call inspect_image.
• Web UI settings section (v0.3.0): Settings → 视觉模型 edits the tool-vision namespace (API endpoint, write-only key, model, bridge options) in settings.yaml; changes hot-apply without a restart. The API key lives in settings.yaml, not the profile patch. Mount by package name (name: 'dsh-tool-vision') so the web client bundle is discovered.

Install

Mount in a profile patch ($DSH_HOME/profiles/<name>/cordis.patch.yml):

- insert:
    - id: tool-vision
      name: 'dsh-tool-vision'     # after: pnpm add dsh-tool-vision in the profile
      config:
        baseURL: 'https://api.openai.com/v1'
        apiKeyEnv: 'VISION_API_KEY'
        model: 'gpt-4o-mini'

Or load it from a local path without npm:

    - id: tool-vision
      name: './plugins/dsh-tool-vision/index.js'

Config

Field	Default	Meaning
enabled	true	Master switch (v0.8.0). Off unregisters everything this plugin contributes — inspect_image, the 14 vision_* tools, the image bridge, the preview route and the image-capability declaration. The settings section stays mounted so the switch can turn it back on. Hot-applies; no dsh restart.
baseURL	https://api.openai.com/v1	OpenAI-compatible API base URL.
apiKey	''	API key (takes precedence over env).
apiKeyEnv	VISION_API_KEY	Env var holding the key.
model	gpt-4o-mini	Vision model id.
maxTokens	1024	Max output tokens.
timeoutMs	60000	Per-request timeout.
maxImageBytes	10MB	Largest accepted local image.
description	default	Tool description shown to the model.
bridgeTextOnly	true	Bridge pasted images to text hints on models that cannot see images.
bridgeExportDir	temp	Export dir for bridged images (os.tmpdir()/dsh-vision-bridge).
multimodalModels	[]	Model list (comma-separated). Each entry is matched case-insensitively against the full id, its bare id after the last /, and provider/id, with * / ? globs (*vl*, deepseek/*). What the list means is set by the mode below.
multimodalListMode	whitelist	List mode (v0.9.0). whitelist: listed models receive image blocks directly (the historical behaviour). blacklist: listed models are forced through the bridge — the correction layer for a model that claims image support it does not have. off: the list is ignored. An unknown value falls back to whitelist.
autoDetectMultimodal	true	Auto-detect (v0.9.0). Decide from the current route's own declared inputModalities, then combine with the list (whitelist unions, blacklist subtracts). On by default: a text-only route is bridged, a multimodal one is treated like a whitelist member and gets images directly. The declaration is always read before this plugin's admission wrap, so bridgeAutoImage can never feed its own claim back in as evidence. Set false for the hand-maintained "list only" behaviour.
probeResults	{}	Measured verdicts (v0.9.0): "provider/model" → "yes"/"no", written by the vision_probe_model tool — do not edit by hand. A measurement outranks a declaration (a real request beats a claim) but not multimodalModels (explicit human intent has the last word).
bridgePreview	true	Inline preview for bridged images: thumbnail above the hint text in the user bubble (click to zoom).
bridgePreviewScanIntervalMs	2000	Fallback scan interval for the preview scanner (ms); 0 disables the fallback.
bridgePreviewHideHint	true	Hide the bridged hint text once the preview image has loaded (kept on failure — safe degradation).
bridgeAutoImage	true	While the bridge is on, report image input capability for every model to the host admission gate, so pasted images are accepted on text-only models without hand-editing provider configs.
sendSessionHeader	true	Send a stable session-id header on vision requests. OpenCode Go and similar gateways require x-opencode-session (one stable id per conversation); requests without it may error from 2026-09-06.
sessionHeaderName	x-opencode-session	Header name carrying the session id.
sessionId	''	Fixed session id for calls without a dsh session context; empty = auto (current dsh session id, else a stable per-process random id).

Image bridge setup

1. (Optional, usually not needed) If bridgeAutoImage is disabled, declare image input on the models you paste images onto, so the harness admits image messages (pi-ai style):

   llm-pi-ai:
     providers:
       your-provider:
         models:
           - id: deepseek-v4-flash
             input: [text, image]
   

2. List genuinely multimodal models in the plugin config so they receive image blocks untouched (see the next section for the list modes):

   - id: tool-vision
     name: 'dsh-tool-vision'
     config:
       multimodalListMode: whitelist    # default: listed models get images directly
       multimodalModels: ['mimo-v2.5', 'grok-4.5']
   

Then pasting an image while on a text-only model stores a hint like [User sent an image, exported to: <path>. Inspect it with the inspect_image tool...] in the transcript (the pasted image no longer renders as pixels in that message), and the agent inspects it through the configured vision endpoint.

Why not llm/stream? The harness freezes every request and the agent-loop invariant fails any request whose messages diverge from the session-log derivation (log-reconstruction desync), and this cordis waterfall's next() cannot replace request arguments. The agent/pre-step waterfall is the supported seam: its decision messages become the durable log, so the invariant stays satisfied.

Key resolution order: config.apiKey → process.env[apiKeyEnv] → process.env.OPENAI_API_KEY.

Bridge image preview (v0.4.0)

On text-only models, pasted images become [User sent an image...] hint text in the transcript. With bridgePreview enabled (default), the browser half renders those hints as inline thumbnails in the display layer only:

• Thumbnail + lightbox: click to zoom full-screen; click anywhere or press Esc to close;
• Immediate + fallback: new messages are handled by a MutationObserver; history is back-filled by a periodic scan (interval via bridgePreviewScanIntervalMs);
• Hide the hint (P2): with bridgePreviewHideHint on, the hint text is hidden once the image has loaded, leaving just the image; on load failure the text stays (safe degradation — never "no image AND no text");
• Precise identification: bridged hints carry an invisible prefix marker (\u200b[bridge]), so ordinary user text that happens to contain "exported to:" is never misidentified;
• Display-layer red line: persisted messages, the transcript, the model-facing text and the inspect_image chain are untouched.

Preview images are served by the same-origin loopback route /plugins/dsh-tool-vision/image: read-only access to the bridge export directory, localhost-only Host, image extensions only, ≤ 20MB per file, path-traversal protected.

Tool: inspect_image

Arg	Required	Meaning
path	✅	Image path (absolute, or relative to the current workspace) or http(s) URL.
question	–	Optional specific question about the image.
detail	–	auto / low / high resolution hint.

Example endpoints (baseURL):

• OpenAI: https://api.openai.com/v1 — gpt-4o, gpt-4o-mini
• Alibaba DashScope (Qwen-VL): https://dashscope.aliyuncs.com/compatible-mode/v1 — qwen-vl-plus, qwen-vl-max
• Zhipu (GLM-4V): https://open.bigmodel.cn/api/paas/v4 — glm-4v-flash (free tier), glm-4v-plus
• Moonshot (Kimi): https://api.moonshot.cn/v1 — moonshot-v1-8k-vision-preview
• Ollama local: http://localhost:11434/v1 — llama3.2-vision (no key)

Note for users

• This plugin is a standard profile bundle (dsh.bundle.patch): dsh plugin --profile web add dsh-tool-vision installs and mounts it in one step — no manual cordis.patch.yml edits needed.
• Settings changes hot-apply (no restart needed).
• Version 0.6.3 and newer require DSH 0.1.0-rc.7 or newer and are tested against 0.1.0-rc.7, 0.1.0-rc.8, and 0.1.1-rc.1.
• DSH 0.1.0-rc.6 users must pin dsh-tool-vision@0.6.1, the last release carrying the legacy settings-allowlist compatibility patch.

Pixel-level vision tools (v0.6.0, ported from dsh-vision-router)

14 vision_* tools driven by the same configured endpoint as inspect_image (baseURL/apiKey/model) — no provider chain, no local models, no extra settings:

Tool	Purpose
vision_describe	Image Q&A / multi-image comparison (optional structured JSON)
vision_ground	Locate a target and return its ORIGINAL-pixel bounding box
vision_detect	Enumerate elements (buttons, inputs, icons…) with numbered boxes
vision_crop	Crop a pixel region to a PNG artifact
vision_pixel_diff	Per-pixel comparison: ratio, worst regions, heatmap, report
vision_colors	Dominant-color quantization for palette matching
vision_ocr	Verbatim text transcription (letters only — not scene analysis)
vision_long_screenshot_ocr	Chunked long-screenshot transcription into Markdown
vision_trace	Potrace vectorization into colored SVG (worker-thread, safe)
vision_extract_foreground	Solid-background removal → transparent PNG
vision_html_screenshot	Headless render of a local .html (network blocked)
vision_screenshot	Desktop capture (privacy-gated: enable desktopScreenshot in settings; Win: PowerShell / macOS: screencapture / Linux: import/scrot)
vision_present	Publish a generated image to the user via the host attachment store
vision_materialize	Copy an attachment/local image into the workspace as a real path

Quality & safety details:

• Content-hash cache keyed by endpoint+model+image+question (no stale answers across model switches, failures are never cached).
• Uniform 4MP downscale before every model call; oversized inputs are rejected with a clear error (stat pre-check, 20MB cap on both file and attachment paths).
• Rate-limit / 5xx auto-retry with Retry-After-aware backoff; endpoint content-safety rejections are surfaced as VISION_CONTENT_FILTERED instead of a generic backend error.
• Long-OCR bounds: 120s total budget, 40-chunk cap, cancellation checks, stop-on-first-backend-failure.
• Path containment for relative inputs; artifacts land in <workspace>/.dsh-tool-vision/.

Requires sharp / potrace / puppeteer-core (declared as optional dependencies: a failed platform install never blocks the plugin; missing ones degrade lazily with an install hint and never break other tools).

vision_screenshot is privacy-sensitive and therefore not registered by default — set desktopScreenshot: true in the tool-vision settings to enable desktop capture.

v0.9.0: list modes, and auto-detection

The bridge answers one question: can the current model see images directly? v0.9.0 splits it into two independent inputs.

base = autoDetectMultimodal ? (route declares image) : {}
off        → direct = base              the list takes no part
whitelist  → direct = base ∪ list       the list only adds
blacklist  → direct = base \ list       the list only subtracts

A list hit always wins: the list is explicit user intent, so it outranks the model's own declaration — which is what makes it a usable correction layer.

A blacklist never degrades into "everything unlisted is direct". Its base set is the auto-detected one; with auto-detection off that base set is empty, so an unlisted model is still bridged. That is deliberate: the alternative lets one typo push images at a text-only endpoint.

Matching: mimo-v2.5, xiaomi/mimo-v2.5 and commandcode/xiaomi/mimo-v2.5 all address the same route; * / ? are globs; matching is case-insensitive. v0.8.1 compared ids literally, so this README's own mimo-v2.5 example silently did nothing on a route spelled xiaomi/mimo-v2.5 — fixed here, and the fix only ever adds models to the direct set (no entry that used to force a model direct stops doing so).

In the panel (Settings → Vision Model):

• the list field is followed by a clickable list of the models dsh actually has configured (llm.listProviders() + llm.listModels()), grouped by provider and labelled with whether the route declares image input. Tick to add, untick to remove; the text field above still takes globs by hand. Both edit the same draft, persisted by Save — the picker head flags it as unsaved until then.
• a current route readout shows provider / model, whether images go direct or through the bridge, and why (list hit / auto-detect / default).

Why not just a <datalist>: a native datalist stays invisible until the user focuses the field and types, which made v0.9.0's first cut look like a dead panel. The list is now always visible, with the datalist kept as a typing aid.

Unticking removes the entry that actually matched — if mimo-v2.5 in the list is what covers xiaomi/mimo-v2.5, unticking drops mimo-v2.5 rather than inventing a full id. Which entry hit is computed server-side with the same matcher the bridge uses (matchedEntries in the payload), so the panel can never display a state that disagrees with the decision.

Read path (the easiest thing to get wrong here): autoDetectMultimodal MUST read the value from before resolveModelInfo was wrapped, or the "image support" that bridgeAutoImage stamps onto every model becomes evidence for itself. unwrappedResolveModelInfo() enforces that, with a dedicated regression test.

Candidates come from the plugin's own loopback route: GET /plugins/dsh-tool-vision/models (loopback Host only, read-only, no-store). It returns provider/model ids and one declared-capability boolean — no keys and no endpoint addresses. It is registered on the plugin fiber rather than the master switch's child fiber, so the panel keeps working while the plugin is off.

⚠️ inputModalities is a declaration, not a guarantee — profiles commonly set input: [text, image] on text-only models just to pass the admission gate. Upstream dsh-llm-pi-ai makes the same call for undeclared models, and its source says why: the two wrong answers do not cost the same. Under-claiming refuses the image before it is attached and names the model; over-claiming admits one the provider rejects mid-turn, after the message is already durable.

That is why detection is on by default with three backstops: (1) the first time a route is promoted purely by its own declaration, the log says so and names the fix; (2) the panel always shows the current route, the decision and the reason; (3) listing that model under blacklist mode forces the bridge back on.

Tests

npm test             # server-side unit tests (no extra dependencies)
npm run test:render  # panel render test (needs devDependencies)

npm run test:render loads the real client bundle in jsdom, drives the real registration path (apply → slots.register → the component), feeds it from the real server route handler, and asserts on the real DOM and the real settings writes.

It is a separate command and deliberately not part of npm test: it needs react / react-dom / jsdom, and a DOM test that silently skips when a dependency is missing is a false comfort. Install with npm i -D react@18 react-dom@18 jsdom.

Its reason to exist is specific: v0.9.0's first cut rendered the model list only into a native <datalist> — every server-side unit test passed while the panel looked completely dead. Nothing below the DOM can catch that class of bug.

Capability probe: vision_probe_model (v0.9.0)

Every other signal here rests on what a model says about itself. This tool sends a real image to the route and reports what it does — the only ground truth in the plugin.

base = autoDetect ? route declares image : {}
a measured verdict (if any) overrides base    measurement > claim
a list hit (if any) overrides everything      human intent > measurement

Why one request is not enough (each of these was learned the hard way):

• Models guess. One may answer "blue" to an orange square, so a probe asks about two different colors and only passes a model that reads both.
• Reasoning models return empty content when the thinking budget eats max_tokens; the default is 2048 and the reader falls back to reasoning_content.
• A control group is mandatory. A text-only request runs first to prove the route works at all; without it a 401 or a timeout would be misread as "cannot see images" and push a perfectly good multimodal route back onto the bridge permanently.
• An accepting endpoint is not an image-reading model. Measured against a real gateway: meituan/LongCat-2.0:free returned HTTP 200 for the image request and answered "I can't see any image." Only an endpoint that actively rejects the image part is a conclusive negative.

So a probe ends in one of three verdicts: yes (control passed, both colors correct), no (the endpoint rejected the image part, or answered without reading it), or unknown (network, auth or protocol trouble — never turned into a capability claim).

How it gets a route's endpoint and credential (no new configuration): llm.listConfigurableProviders() names the provider's settings namespace and path → settings.get(ns) resolves baseURL/apiKeyEnv/api → credentials.resolve(apiKeyEnv) yields the secret (.value) — the same path dsh-llm-pi-ai uses for a real call. Read-only, and neither the endpoint nor the key ever appears in a probe result.

Linking: a verdict is written to probeResults and takes effect in the bridge decision immediately — yes sends images directly, no forces the bridge back on — regardless of the list mode. (That is also why it does not silently write into multimodalModels: a blacklist list means the opposite, so an automatic entry there would produce exactly the wrong result.) Each row in the panel's picker carries a badge:

• measured: reads images / measured: no image reading (blue / red), shown in preference to the declaration;
• a never-probed route falls back to the declares image badge;
• the current-route readout reports measured (a real image was read / NOT read).

A route on an unknown protocol (e.g. openai-responses) returns unknown with a reason, rather than guessing confidently with chat/completions.

v0.8.0: master switch, and the save-path fix

Master switch. enabled, plus a one-click button at the top of the section (Disable all / Re-enable). Registrations are effects on the cordis fiber that makes them, so the plugin now puts every tool, the image bridge, the preview route and the image-capability declaration in a child fiber: turning the switch off disposes it, and all 15 tools leave the model's tool list together. The settings section stays on the parent fiber, so the switch can turn the plugin back on. No dsh restart.

Save-path fix. The form used to submit its 18 fields as parallel scope.set()/unset() calls. Each write carries its own revision fence, a fence behind the Host document is refused with settings/conflict, and a refused write still resolves — the scope's contract is "settle after the write and any recovery read", not "throw on refusal". The section therefore reported "Saved" while the edits silently reverted, which reads as "settings cannot be saved at all".

Writes are now one atomic mutate(), so the whole batch shares one fence and one persistence decision, and the section is inspected after the write settles: "Saved" only when the change is really there, otherwise "Write did not take effect" plus a reload of the form. Hosts without mutate() fall back to sequential writes (each waits for its predecessor, keeping the revision chain intact) — never parallel.

Also removes three if (typeof scope.load === "function") scope.load() guards. The SettingsScope seam has never had load() — it is getSnapshot / subscribe / mutate / set / unset, and reads ride the shared describe mirror driven by the Host's settings/document-updated. Those guards were dead code that read like a refresh which never happened, and they made the missing write verification look intentional.

Limitations

• A bridged image enters the conversation as a text hint (a transcript, not pixels) — pixel-precise in-context reasoning is not available to text-only models; the vision model's description comes back through inspect_image.
• The bridge is a one-way door: an image pasted on a text-only model is rewritten into the durable log at agent/pre-step, so switching to a multimodal model later does not turn it back into an image block. (The other direction — multimodal to text-only — is repaired automatically by repairLoggedImages.)
• Images are base64-transferred; mind privacy and size limits.
• Independent of the dsh-llm routing/retry system; failures return clear errors to the agent.

License

MIT — bridge preview & integration: xing666173. Pixel vision tools ported from dsh-vision-router (© ysr666, MIT) with gratitude.