中文 | English

dsh-vision-bridge

A DSH profile bundle for sessions running models that cannot accept image input (DeepSeek and similar). Instead of the native "the current model does not support images" rejection, the plugin runs the following pipeline:

1. materializes the image into the session workspace (.dsh-vision-bridge/);
2. rewrites the user message (via the sanctioned agent/pre-step channel) into a [图片桥接] instruction carrying the image paths;
3. registers a system-prompt rule (order 150) requiring the model to call the configured vision tool (default vision_glance, fuzzy-selectable such as glance / ground / detect, matched against the agent's actual toolset) and answer from its analysis — the tool-call card and its result appear in the conversation;
4. patches llm.streamWithRegistration — the dispatch funnel behind both llm.stream and the agent loop's prepared-call path — to strip image blocks (including images nested inside tool results), replacing them with a placeholder before a text-only adapter serializes the conversation.

Sessions whose agent preset matches excluded presets (default minimal, the two-tool preset) are NOT bridged: their images are replaced with the original "model does not support images" notice text. The preset a session runs is resolved the same way DSH itself resolves it: the newest agent-preset/selected event wins, with the creation-time header as fallback.

Why the native rejection happens (DSH itself)

• dsh-host-apiproxy's prompt RPC rejects image messages before they enter the session when llm.resolveModelInfo().inputModalities lacks image (that is the banner shown in the UI).
• The DeepSeek adapter throws UNSUPPORTED_CONTENT on any message containing image blocks at serialization time.

This plugin intervenes at three layers:

Layer	Mechanism
Admission	Patches llm.resolveModelInfo so bridged models report an image input modality
Current message	agent/pre-step materializes the image and injects the vision-tool instruction
History	Patches llm.streamWithRegistration (the shared dispatch funnel) to strip image blocks before adapter serialization

Settings (Settings → 视觉桥接)

Setting	Default	Meaning
Enable vision bridge	on	Off = restore native DSH blocking entirely
Vision tool (fuzzy input)	vision_glance	The tool the model must call; fuzzy-matched against the agent's real toolset
Custom prompt (optional)	empty	Extra instruction passed to the vision tool
Bridge models	deepseek-official*,deepseek*	provider[:model] patterns, * wildcards
Excluded session presets	minimal	Multi-select + extra wildcard patterns; matched sessions are NOT bridged (original blocking text)

Configuration is persisted through the DSH settings service (namespace vision-bridge); it survives restarts.

Platforms: WSL/Ubuntu and Windows

• Image bytes are written via ctx.shell with base64 on stdin (no huge command lines): bash (base64 -d) on Linux/WSL, PowerShell ([IO.File]::WriteAllBytes) on native Windows — the shell flavor is auto-detected once (Write-Output probe + resolveExecutable('pwsh') fallback).
• Shell commands carry the session-resolved sandbox policy (like the bash tool): danger-full-access sessions run without a sandbox backend; confined sessions run inside the workspace fence when a backend exists.
• The vision tools themselves are native DSH tools (vision-tools skill); their runtime already handles cross-platform execution.

Requirements

• The session must have the vision-tools skill loaded (or another vision plugin) so a vision tool exists in the agent's toolset; otherwise the bridge degrades to a self-heal instruction (load the skill first) or a clear notice.
• Standard/cordis presets have the skill tool; the minimal preset does not (it is excluded by default).

Installation

Unzip the bundle, then add it to a profile:

unzip dsh-vision-bridge.zip -d dsh-vision-bridge
dsh plugin --profile web add ./dsh-vision-bridge

(dsh plugin forwards to pnpm in the profile directory. Alternatively, copy the extracted folder into the profile's node_modules/@dsh-external/ and reconcile plugins.)

Known limitations

• The admission patch is process-wide, so excluded sessions cannot get the native error banner — the same reason is rendered as text instead.
• Multiple sessions share the plugin's global effects (llm patches, pre-step listener, system rule) while the plugin is mounted.
• Bridged models never receive image blocks (the adapter forbids them); the attachment stays available in the session.