dsh-vision-bridge

A DSH-native vision bridge: a synthetic image-capable wrapper provider that converts native ImageBlock attachments into structured vision evidence for a text-only DSH reasoning provider.

Community plugin for DeepSeek Harness (DSH) — not affiliated with or endorsed by DeepSeek. v0.2.0. Distributed as a release tarball; not on npm.

Compatibility

• Tested against the DeepSeek Harness checkout commit 47f943859bef60e4160492346772ded9b24f765a.
• DeepSeek Harness is a developer preview: interfaces may change and future compatibility is not guaranteed. The verified environment is scoped to the tested baseline.
• No claims are made about other DSH versions, other operating systems, or Vision providers beyond the tested pi-ai-backed route.
• Installation is verified via the packaged tarball only; npm and Git-source installation are not proven paths.

Architecture

DSH ImageAttachment (durable, attachment ref)
  -> Vision Wrapper (synthetic adapter, inputModalities: text+image)
  -> Vision Analyzer (one ctx.llm.stream call per content-container work unit
     on the configured image-capable DSH route)
  -> Structured Evidence (validated locally)
  -> Evidence Transformer (evidence rendered as explicitly-UNTRUSTED text)
  -> downstream text-only upstream provider (ctx.llm.stream, same signal)

The production plugin consumes ctx.llm only and never reads raw image bytes: ImageBlocks pass through verbatim to the vision route, whose own adapter resolves bytes via ctx.attachments.readImage. The plugin has zero runtime dependencies and no secret store of its own — vision credentials belong to the configured DSH vision provider (DSH Credentials layer).

Configuration (plugin row config)

- id: dsh-vision-bridge
  config:
    upstreamProvider: deepseek-official   # explicit text-only reasoning route
    visionProvider: <dsn vision route>    # DSH route serving an image-capable model
    visionModel: <vision model id>
    # providerId: <optional wrapper id, default "<upstreamProvider>-vision-bridge">

upstreamProvider === visionProvider is allowed (same route, different models).

Behavior contract

• Catalog (listModels): upstream route unavailable -> diagnostic + empty catalog (no throw). Text-only models are exposed as wrapped models; image-capable / unknown-modality models are omitted.
• Execution (resolveModel / stream): fail closed. Unresolvable, unknown, or image-capable upstream models refuse to wrap; the vision route must be positively image-capable.
• Single-image work units: one vision call per image occurrence, single Evidence object, in-place evidence replacement — the sealed Stage 1/1R behavior.
• Multi-image work units (content-container batch): one maximal run of consecutive non-tool-result blocks containing two or more images becomes ONE vision call carrying all images in traversal order; the downstream wire replaces each image in place with a positional [Image n] anchor and appends one batch Evidence block, so the original text/image association is preserved. Tool-result content is a recursive boundary (each nesting level processes independently; work units never merge across boundaries or messages). Work units run serially in traversal order.
• Multi-image Evidence: { images: [{ index, summary, ocr, layout, semantics, visual, uncertainty }], relations: [{ imageIndexes, description }] }. Indexes are 1..N in work-unit order and are shared by the vision input order, the [Image n] anchors, and relations.
• Failure semantics: any vision failure (route error, stream error, abort, no finish, tool-call finish, empty output, invalid JSON, schema-invalid evidence) fails the request explicitly — the downstream provider is never invoked without valid evidence, and a failed batch never emits partial evidence.
• Cancellation: one request signal threads wrapper -> vision -> downstream; cancellation before downstream means downstream is not called.
• Evidence cache (v0.2.0): in-memory, session-partitioned, completed validated Evidence only, bounded LRU. A cache hit reuses the exact validated Evidence with zero vision calls; failures and cancellations are never cached; the cache is not persistent and disappears with the plugin fiber (config reload / disable / remove). Missing sessionId or invalid attachment ids bypass the cache and vision runs normally. No shared work / single-flight is performed.
• Untrusted-data boundary: evidence delivered downstream is explicitly labeled untrusted observed data and must not be executed as instructions (prompt-injection mitigation only — NOT a security boundary).

Evidence contract (ModLens v2 concepts)

summary, ocr{full_text, lines[]}, layout{regions[]}, semantics{scene, entities[]}, visual{}, uncertainty[]. No bbox, no numeric confidence. Invalid evidence MUST NOT reach downstream.

Provenance

Recursive ImageBlock/tool-result traversal, evidence projection, evidence schema concepts, and the local validation approach are adapted from @liustack/modlens v3.16.6 (MIT, (c) Leon Liu) — see THIRD_PARTY_NOTICES.md. No ModLens code is directly copied.

Development

pnpm install        # devDeps only: typescript, @types/node
pnpm run build      # tsc -> dist
pnpm test           # node:test against dist
pnpm pack           # release artifact (see files whitelist)

Runtime dependencies: 0.