dsh-local-ai 0.1.3 → 0.1.5

package/CHANGELOG.md

@@ -5,6 +5,24 @@ All notable changes to this project are documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.1.5] - 2026-08-23
+
+### Added
+
+- Endpoint liveness (M3): `scripts/check-endpoints.mjs` probes the configured Ollama HTTP endpoint (`/api/version`; 2xx = alive, any transport error or non-2xx = fail) with `OLLAMA_HOST` / `CHECK_ENDPOINTS` / `TIMEOUT_MS` overrides, and `.github/workflows/check-endpoints.yml` runs it monthly and on demand against a throwaway local Ollama; `test/check-endpoints.spec.ts` covers the endpoint-resolution, verdict, error-classification, and timeout helpers plus a plain-Node syntax check.
+
+## [0.1.4] - 2026-08-22
+
+### Added
+
+- **Vision support (issue #5)** — models whose `/api/show` capabilities include `vision` now declare `inputModalities: ['text', 'image']` and carry base64 image payloads on user messages (resolved through the optional `attachments` service); the `vision` config knob (default `true`) keeps the route text-only on opt-out, and text-only models still reject image content loudly (`UNSUPPORTED_CONTENT`).
+
+
+### Changed
+
+- Upgraded every `@deepseek-ai/dsh-*` dev dependency from `0.1.0-rc.8` to `0.1.1-rc.2` for DeepSeek Harness `0.1.1-rc.2` compatibility. Peer ranges stay `>=0.1.0-rc.8 <0.2.0`: no adapter, routing, tool, or command code uses an rc2-only API.
+- Repinned `minimumReleaseAgeExclude` to the whole `@deepseek-ai/*` scope and synchronized the `0.1.1-rc.2` baseline across the five-language READMEs, AGENTS.md, THIRD_PARTY_NOTICES.md, the CI workflow name, and the compat workflow.
+
 ## [0.1.3] - 2026-08-21
 
 ### Changed

package/README.es.md

@@ -24,7 +24,7 @@
 
 | Superficie | Estado |
 |---|---|
-| Harness | DeepSeek Harness `0.1.0-rc.8` |
+| Harness | DeepSeek Harness `0.1.1-rc.2` |
 | Node | `^22.19.0 \|\| >=24.0.0` |
 | Backend | [Ollama](https://ollama.com) (API HTTP local + sonda CLI) |
 | Modelo | Ruta solo texto (`inputModalities: ['text']`); se admiten llamadas y resultados de herramientas |
@@ -102,6 +102,7 @@ Todos los ajustes son campos `Config` de Schemastery (modificables desde cordis.
 | `defaultContextWindow` | `8192` | Capacidad de contexto cuando un modelo no tiene valor exacto |
 | `maxTokens` | `4096` | Límite de salida por solicitud cuando un modelo no tiene valor exacto |
 | `temperature` | *(none)* | Temperatura de muestreo por defecto (0..2); omitir deja el valor del proveedor |
+| `vision` | `true` | Declara y serializa el soporte de imágenes cuando el modelo informa vision; `false` mantiene la ruta solo texto |
 | `models` | `[]` | Mapeos de nombre visible → modelo Ollama |
 | `models[].name` | *(required)* | Nombre de modelo visible en el harness (`GenerateOptions.model`) |
 | `models[].model` | `= name` | Id del modelo Ollama |
@@ -143,22 +144,23 @@ Todos los ajustes son campos `Config` de Schemastery (modificables desde cordis.
 
 ## Known limitations
 
-- **Solo rc.8** — desarrollado y probado contra `@deepseek-ai/dsh@0.1.0-rc.8`; se espera que baselines más nuevos funcionen, pero los verifica el workflow mensual de compat.
-- **Ruta solo texto** — el contenido de imagen se rechaza (`UNSUPPORTED_CONTENT`); los modelos locales multimodales aún no están conectados.
+- **Solo rc.2** — desarrollado y probado contra `@deepseek-ai/dsh@0.1.1-rc.2`; se espera que baselines más nuevos funcionen, pero los verifica el workflow mensual de compat.
+- **Vision cuando el modelo la informa** — los modelos cuyas capacidades de `/api/show` incluyen `vision` declaran `inputModalities: ["text","image"]` y llevan cargas de imagen base64 en los mensajes de usuario (exclusión con `vision: false`); los modelos solo texto siguen rechazando el contenido de imagen (`UNSUPPORTED_CONTENT`).
 - **Respaldo a mitad de flujo** — una vez que una ruta local empezó a producir contenido, un fallo posterior se reenvía (no se retira); solo un fallo antes del primer token respalda a la nube.
 
 ## Development
 
 ```sh
 pnpm install        # node ^22.19 || >=24
-pnpm run typecheck  # tsc: src + tests contra los tipos publicados 0.1.0-rc.8
-pnpm run typecheck:ci  # tsc estricto contra los tipos publicados rc.8 (skipLibCheck off)
+pnpm run typecheck  # tsc: src + tests contra los tipos publicados 0.1.1-rc.2
+pnpm run typecheck:ci  # tsc estricto contra los tipos publicados rc.2 (skipLibCheck off)
 pnpm test           # vitest: costuras reales Context/LlmRuntime/ToolRuntime/CommandRuntime/subprocess
 pnpm run test:coverage  # puerta de cobertura (90/80/90/90)
 pnpm run build      # bundle tsdown + declaraciones tsc (lib/)
 pnpm run verify:self-contained  # las especificaciones de dependencias resuelven desde el registry
 pnpm run verify:artifacts       # cara ESM construida + bundle patch presentes
 node scripts/check-readme-sync.mjs  # puerta de sincronización de README en cinco idiomas
+node scripts/check-endpoints.mjs  # sonda de actividad M3 (Ollama /api/version)
 pnpm pack           # el tarball publicado
 ```
 

package/README.hi.md

@@ -24,7 +24,7 @@
 
 | सतह | स्थिति |
 |---|---|
-| Harness | DeepSeek Harness `0.1.0-rc.8` |
+| Harness | DeepSeek Harness `0.1.1-rc.2` |
 | Node | `^22.19.0 \|\| >=24.0.0` |
 | Backend | [Ollama](https://ollama.com) (स्थानीय HTTP API + CLI जाँच) |
 | Model | केवल-पाठ रूट (`inputModalities: ['text']`); टूल कॉल व परिणाम समर्थित हैं |
@@ -102,6 +102,7 @@ dsh --profile web --dump-config | grep -A2 'id: dsh-local-ai'
 | `defaultContextWindow` | `8192` | जब मॉडल का कोई सटीक मान न हो तो संदर्भ क्षमता |
 | `maxTokens` | `4096` | जब मॉडल का कोई सटीक मान न हो तो प्रति-अनुरोध आउटपुट सीमा |
 | `temperature` | *(none)* | डिफ़ॉल्ट सैंपलिंग तापमान (0..2); छोड़ने पर प्रदाता डिफ़ॉल्ट रहता है |
+| `vision` | `true` | मॉडल द्वारा vision रिपोर्ट करने पर छवि समर्थन घोषित व सीरियलाइज़ करता है; `false` रूट को केवल-टेक्स्ट रखता है |
 | `models` | `[]` | Harness-दृश्य नाम → Ollama मॉडल मैपिंग |
 | `models[].name` | *(required)* | Harness-दृश्य मॉडल नाम (`GenerateOptions.model`) |
 | `models[].model` | `= name` | Ollama मॉडल id |
@@ -143,22 +144,23 @@ dsh --profile web --dump-config | grep -A2 'id: dsh-local-ai'
 
 ## Known limitations
 
-- **केवल rc.8** — `@deepseek-ai/dsh@0.1.0-rc.8` के विरुद्ध विकसित व परीक्षित; नए harness बेसलाइन काम करने की अपेक्षा है, पर मासिक compat workflow उन्हें सत्यापित करता है।
-- **केवल-पाठ रूट** — छवि सामग्री अस्वीकृत होती है (`UNSUPPORTED_CONTENT`); मल्टीमॉडल स्थानीय मॉडल अभी जुड़े नहीं हैं।
+- **केवल rc.2** — `@deepseek-ai/dsh@0.1.1-rc.2` के विरुद्ध विकसित व परीक्षित; नए harness बेसलाइन काम करने की अपेक्षा है, पर मासिक compat workflow उन्हें सत्यापित करता है।
+- **मॉडल द्वारा vision रिपोर्ट करने पर विज़न** — जिन मॉडलों की `/api/show` capabilities में `vision` है, वे `inputModalities: ["text","image"]` घोषित करते हैं और उपयोगकर्ता संदेशों पर base64 छवि पेलोड ले जाते हैं (`vision: false` से बंद करें); केवल-टेक्स्ट मॉडल अब भी छवि सामग्री अस्वीकार करते हैं (`UNSUPPORTED_CONTENT`)।
 - **मध्य-स्ट्रीम वापसी** — एक बार स्थानीय रूट सामग्री बनाना शुरू कर दे, तो बाद की विफलता आगे भेजी जाती है (वापस नहीं ली जाती); केवल पहले टोकन से पहले की विफलता क्लाउड पर लौटती है।
 
 ## Development
 
 ```sh
 pnpm install        # node ^22.19 || >=24
-pnpm run typecheck  # tsc: src + परीक्षण, प्रकाशित 0.1.0-rc.8 प्रकारों के विरुद्ध
-pnpm run typecheck:ci  # सख्त tsc, प्रकाशित rc.8 प्रकारों के विरुद्ध (skipLibCheck बंद)
+pnpm run typecheck  # tsc: src + परीक्षण, प्रकाशित 0.1.1-rc.2 प्रकारों के विरुद्ध
+pnpm run typecheck:ci  # सख्त tsc, प्रकाशित rc.2 प्रकारों के विरुद्ध (skipLibCheck बंद)
 pnpm test           # vitest: वास्तविक Context/LlmRuntime/ToolRuntime/CommandRuntime/subprocess सीम
 pnpm run test:coverage  # कवरेज द्वार (90/80/90/90)
 pnpm run build      # tsdown बंडल + tsc घोषणाएँ (lib/)
 pnpm run verify:self-contained  # निर्भरता विनिर्देश registry से हल होते हैं
 pnpm run verify:artifacts       # निर्मित ESM फलक + बंडल पैच मौजूद
 node scripts/check-readme-sync.mjs  # पाँच-भाषा README समन्वय द्वार
+node scripts/check-endpoints.mjs  # M3 एंडपॉइंट-लाइवनेस प्रोब (Ollama /api/version)
 pnpm pack           # प्रकाशित tarball
 ```
 

package/README.md

@@ -1,6 +1,7 @@
 <div align="center">
 
 # 🤖 dsh-local-ai
+[![Gitee](https://img.shields.io/badge/Gitee-mirror-c71d23?logo=gitee)](https://gitee.com/perrylink/dsh-local-ai)
 
 **Local-model (Ollama) integration for DeepSeek Harness.**
 
@@ -24,7 +25,7 @@
 
 | Surface | Status |
 |---|---|
-| Harness | DeepSeek Harness `0.1.0-rc.8` |
+| Harness | DeepSeek Harness `0.1.1-rc.2` |
 | Node | `^22.19.0 \|\| >=24.0.0` |
 | Backend | [Ollama](https://ollama.com) (local HTTP API + CLI probe) |
 | Model | Text-only route (`inputModalities: ['text']`); tool calls and tool results are supported |
@@ -102,6 +103,7 @@ All tunables are Schemastery `Config` fields (changeable from cordis.yml). An id
 | `defaultContextWindow` | `8192` | Context capacity used when a model has no exact value |
 | `maxTokens` | `4096` | Per-request output cap used when a model has no exact value |
 | `temperature` | *(none)* | Default sampling temperature (0..2); omitted leaves the provider default |
+| `vision` | `true` | Declare and serialize image support when the model reports vision; `false` keeps the route text-only |
 | `models` | `[]` | Harness-visible → Ollama model mappings |
 | `models[].name` | *(required)* | Harness-visible model name (`GenerateOptions.model`) |
 | `models[].model` | `= name` | Ollama model id |
@@ -143,22 +145,23 @@ All tunables are Schemastery `Config` fields (changeable from cordis.yml). An id
 
 ## Known limitations
 
-- **rc.8 only** — developed and tested against `@deepseek-ai/dsh@0.1.0-rc.8`; newer harness baselines are expected to work but are verified by the monthly compat workflow.
-- **Text-only route** — image content is rejected (`UNSUPPORTED_CONTENT`); multimodal local models are not wired up yet.
+- **rc.2 only** — developed and tested against `@deepseek-ai/dsh@0.1.1-rc.2`; newer harness baselines are expected to work but are verified by the monthly compat workflow.
+- **Vision when the model reports it** — models whose `/api/show` capabilities include `vision` declare `inputModalities: ["text","image"]` and carry base64 image payloads on user messages (opt out with `vision: false`); text-only models still reject image content (`UNSUPPORTED_CONTENT`).
 - **Mid-stream fallback** — once a local route has started producing content, a later failure is forwarded (not retracted); only a failure before the first token falls back to the cloud.
 
 ## Development
 
 ```sh
 pnpm install        # node ^22.19 || >=24
-pnpm run typecheck  # tsc: src + tests against the published 0.1.0-rc.8 types
-pnpm run typecheck:ci  # strict tsc against published rc.8 types (skipLibCheck off)
+pnpm run typecheck  # tsc: src + tests against the published 0.1.1-rc.2 types
+pnpm run typecheck:ci  # strict tsc against published rc.2 types (skipLibCheck off)
 pnpm test           # vitest: real Context/LlmRuntime/ToolRuntime/CommandRuntime/subprocess seams
 pnpm run test:coverage  # coverage gate (90/80/90/90)
 pnpm run build      # tsdown bundle + tsc declarations (lib/)
 pnpm run verify:self-contained  # dependency specs resolve from the registry
 pnpm run verify:artifacts       # built ESM face + bundle patch present
 node scripts/check-readme-sync.mjs  # five-language README sync gate
+node scripts/check-endpoints.mjs  # M3 endpoint-liveness probe (Ollama /api/version)
 pnpm pack           # the published tarball
 ```
 

package/README.pt.md

@@ -24,7 +24,7 @@
 
 | Superfície | Status |
 |---|---|
-| Harness | DeepSeek Harness `0.1.0-rc.8` |
+| Harness | DeepSeek Harness `0.1.1-rc.2` |
 | Node | `^22.19.0 \|\| >=24.0.0` |
 | Backend | [Ollama](https://ollama.com) (API HTTP local + sonda CLI) |
 | Modelo | Rota somente texto (`inputModalities: ['text']`); chamadas e resultados de ferramentas são suportados |
@@ -102,6 +102,7 @@ Todos os ajustes são campos `Config` de Schemastery (modificáveis pelo cordis.
 | `defaultContextWindow` | `8192` | Capacidade de contexto quando um modelo não tem valor exato |
 | `maxTokens` | `4096` | Limite de saída por solicitação quando um modelo não tem valor exato |
 | `temperature` | *(none)* | Temperatura de amostragem padrão (0..2); omitir mantém o padrão do provedor |
+| `vision` | `true` | Declara e serializa o suporte a imagens quando o modelo informa vision; `false` mantém a rota somente texto |
 | `models` | `[]` | Mapeamentos nome visível → modelo Ollama |
 | `models[].name` | *(required)* | Nome de modelo visível no harness (`GenerateOptions.model`) |
 | `models[].model` | `= name` | Id do modelo Ollama |
@@ -143,22 +144,23 @@ Todos os ajustes são campos `Config` de Schemastery (modificáveis pelo cordis.
 
 ## Known limitations
 
-- **Somente rc.8** — desenvolvido e testado contra `@deepseek-ai/dsh@0.1.0-rc.8`; espera-se que baselines mais novos funcionem, mas são verificados pelo workflow mensal de compat.
-- **Rota somente texto** — conteúdo de imagem é rejeitado (`UNSUPPORTED_CONTENT`); modelos locais multimodais ainda não estão conectados.
+- **Somente rc.2** — desenvolvido e testado contra `@deepseek-ai/dsh@0.1.1-rc.2`; espera-se que baselines mais novos funcionem, mas são verificados pelo workflow mensal de compat.
+- **Vision quando o modelo a informa** — modelos cujas capacidades de `/api/show` incluem `vision` declaram `inputModalities: ["text","image"]` e carregam payloads de imagem base64 nas mensagens do usuário (desative com `vision: false`); modelos somente texto continuam rejeitando conteúdo de imagem (`UNSUPPORTED_CONTENT`).
 - **Fallback no meio do fluxo** — uma vez que uma rota local começou a produzir conteúdo, uma falha posterior é reencaminhada (não retirada); apenas uma falha antes do primeiro token faz fallback para a nuvem.
 
 ## Development
 
 ```sh
 pnpm install        # node ^22.19 || >=24
-pnpm run typecheck  # tsc: src + testes contra os tipos publicados 0.1.0-rc.8
-pnpm run typecheck:ci  # tsc estrito contra os tipos publicados rc.8 (skipLibCheck off)
+pnpm run typecheck  # tsc: src + testes contra os tipos publicados 0.1.1-rc.2
+pnpm run typecheck:ci  # tsc estrito contra os tipos publicados rc.2 (skipLibCheck off)
 pnpm test           # vitest: costuras reais Context/LlmRuntime/ToolRuntime/CommandRuntime/subprocess
 pnpm run test:coverage  # porta de cobertura (90/80/90/90)
 pnpm run build      # bundle tsdown + declarações tsc (lib/)
 pnpm run verify:self-contained  # especificações de dependências resolvem do registry
 pnpm run verify:artifacts       # face ESM construída + bundle patch presentes
 node scripts/check-readme-sync.mjs  # porta de sincronização de README em cinco idiomas
+node scripts/check-endpoints.mjs  # sonda de atividade M3 (Ollama /api/version)
 pnpm pack           # o tarball publicado
 ```
 

package/README.zh.md

@@ -24,7 +24,7 @@
 
 | 项目 | 状态 |
 |---|---|
-| Harness | DeepSeek Harness `0.1.0-rc.8` |
+| Harness | DeepSeek Harness `0.1.1-rc.2` |
 | Node | `^22.19.0 \|\| >=24.0.0` |
 | 后端 | [Ollama](https://ollama.com)（本地 HTTP API + CLI 探测） |
 | 模型 | 纯文本路由（`inputModalities: ['text']`）；支持工具调用与工具结果 |
@@ -102,6 +102,7 @@ dsh --profile web --dump-config | grep -A2 'id: dsh-local-ai'
 | `defaultContextWindow` | `8192` | 模型无精确值时的上下文容量 |
 | `maxTokens` | `4096` | 模型无精确值时的单请求输出上限 |
 | `temperature` | *(none)* | 默认采样温度（0..2）；省略则用提供方默认值 |
+| `vision` | `true` | 模型报告 vision 能力时声明并序列化图片支持；`false` 保持纯文本路由 |
 | `models` | `[]` | Harness 可见名 → Ollama 模型映射 |
 | `models[].name` | *(required)* | Harness 可见模型名（`GenerateOptions.model`） |
 | `models[].model` | `= name` | Ollama 模型 id |
@@ -143,22 +144,23 @@ dsh --profile web --dump-config | grep -A2 'id: dsh-local-ai'
 
 ## Known limitations
 
-- **仅 rc.8** —— 针对 `@deepseek-ai/dsh@0.1.0-rc.8` 开发与测试；更新版本的 harness 基线预期可用，但由每月 compat workflow 验证。
-- **纯文本路由** —— 图片内容会被拒绝（`UNSUPPORTED_CONTENT`）；多模态本地模型尚未接入。
+- **仅 rc.2** —— 针对 `@deepseek-ai/dsh@0.1.1-rc.2` 开发与测试；更新版本的 harness 基线预期可用，但由每月 compat workflow 验证。
+- **模型声明 vision 时启用视觉** — `/api/show` capabilities 含 `vision` 的模型声明 `inputModalities: ["text","image"]`，并在用户消息上携带 base64 图片载荷（可用 `vision: false` 退出）；纯文本模型仍拒绝图片内容（`UNSUPPORTED_CONTENT`）。
 - **中途失败不回退** —— 本地路由一旦开始产出内容，之后的失败会透传（无法撤回）；只有首个 token 前的失败才回退云端。
 
 ## Development
 
 ```sh
 pnpm install        # node ^22.19 || >=24
-pnpm run typecheck  # tsc：src + 测试，对发布版 0.1.0-rc.8 类型
-pnpm run typecheck:ci  # 严格 tsc，对发布版 rc.8 类型（关闭 skipLibCheck）
+pnpm run typecheck  # tsc：src + 测试，对发布版 0.1.1-rc.2 类型
+pnpm run typecheck:ci  # 严格 tsc，对发布版 rc.2 类型（关闭 skipLibCheck）
 pnpm test           # vitest：真实 Context/LlmRuntime/ToolRuntime/CommandRuntime/subprocess 机制
 pnpm run test:coverage  # 覆盖率门禁（90/80/90/90）
 pnpm run build      # tsdown 打包 + tsc 声明（lib/）
 pnpm run verify:self-contained  # 依赖声明均来自 registry
 pnpm run verify:artifacts       # 构建产物 ESM 面 + bundle patch 存在
 node scripts/check-readme-sync.mjs  # 五语 README 同步门禁
+node scripts/check-endpoints.mjs  # M3 端点存活探测（Ollama /api/version）
 pnpm pack           # 发布用 tarball
 ```
 

package/THIRD_PARTY_NOTICES.md

@@ -13,7 +13,7 @@ published tarball; these are install-time dependencies:
 | [typescript](https://github.com/microsoft/TypeScript) | `^5.9.0` | Apache-2.0 | Build-time declaration emission (`lib/types/`) |
 | [@deepseek-ai/cordis](https://www.npmjs.com/package/@deepseek-ai/cordis) | `^4.0.1` (peer) | See package | The plugin runtime |
 | [@deepseek-ai/schemastery](https://www.npmjs.com/package/@deepseek-ai/schemastery) | `^3.18.0` (peer) | See package | Configuration schema |
-| `@deepseek-ai/dsh-*` peers | `0.1.0-rc.8` (peer) | See packages | Official harness seams (`dsh-llm`, `dsh-tools`, `dsh-subprocess`, `dsh-commands`, `dsh-timeout`) |
+| `@deepseek-ai/dsh-*` peers | `0.1.1-rc.2` (peer) | See packages | Official harness seams (`dsh-llm`, `dsh-tools`, `dsh-subprocess`, `dsh-commands`, `dsh-timeout`) |
 
 At runtime the plugin only talks to the Ollama endpoint you configure (over
 its HTTP API and, for the process-liveness probe, its CLI); it performs no

package/cordis.patch.yml

@@ -21,6 +21,8 @@
         maxTokens: 4096
         # Default sampling temperature; omit to keep the provider default.
         # temperature: 0.7
+# Declare image support when the model reports vision; false keeps text-only.
+# vision: true
         # Harness-visible -> Ollama model mappings. Each entry maps a name you
         # select in the harness to an Ollama model id (defaults to the name).
         models: []

package/lib/index.js

@@ -23,6 +23,7 @@ const Config = z.object({
 	defaultContextWindow: z.number().default(8192),
 	maxTokens: z.number().default(4096),
 	temperature: z.number(),
+	vision: z.boolean().default(true),
 	models: z.array(z.object({
 		name: z.string().required(),
 		model: z.string(),
@@ -84,6 +85,7 @@ function resolveConfig(config = {}) {
 	assertPositiveInt("maxTokens", maxTokens);
 	const temperature = config.temperature;
 	if (temperature !== void 0) assertFiniteRange("temperature", temperature, 0, 2);
+	const vision = config.vision ?? true;
 	const seenNames = /* @__PURE__ */ new Set();
 	const models = (config.models ?? []).map((mapping, index) => {
 		if (typeof mapping.name !== "string" || mapping.name.trim().length === 0) throw new TypeError(`models[${index}].name must be a non-empty string`);
@@ -127,6 +129,7 @@ function resolveConfig(config = {}) {
 		defaultContextWindow,
 		maxTokens,
 		...temperature === void 0 ? {} : { temperature },
+		vision,
 		models,
 		route
 	};
@@ -242,6 +245,10 @@ function sanitizeText(value, maxChars = 4e3) {
 * `LlmError`. The fetch implementation is injectable for tests.
 * @module dsh-local-ai/ollama
 */
+/** True when the reported model capabilities include vision. */
+function hasVision(capabilities) {
+	return capabilities?.includes("vision") ?? false;
+}
 /** Build an absolute API URL from a normalized base URL. */
 function endpointUrl(baseURL, path) {
 	return `${baseURL}${path}`;
@@ -412,8 +419,10 @@ function contextLengthOf(show) {
 * vocabulary. User text is joined; assistant text becomes `content` and tool
 * calls become `tool_calls` (with arguments parsed from the raw JSON string to
 * the object Ollama expects); tool results become separate `tool` messages.
-* Core image blocks are rejected explicitly because this route is text-only;
-* unknown declaration-merged block types retain the documented extension
+* Top-level user-message image blocks map onto `images` (base64, no data-URI
+* prefix) when the request carries resolved payloads; images anywhere else —
+* or on a text-only route — still fail loud with `UNSUPPORTED_CONTENT`.
+* Unknown declaration-merged block types retain the documented extension
 * fallback (ignored for content, retained as text where text is expected).
 * @module dsh-local-ai/serialize
 */
@@ -421,9 +430,29 @@ function contextLengthOf(show) {
 function flattenText(blocks) {
 	return blocks.filter((block) => block.type === "text").map((block) => block.text).join("");
 }
-/** Reject core image content before any text-flattening path can silently erase it. */
-function assertTextOnly(blocks) {
-	if (contentHasImage(blocks)) throw new LlmError("The Ollama adapter does not support image content.", "UNSUPPORTED_CONTENT");
+/**
+* Reject image content the wire cannot carry. Top-level user-message images
+* map onto `images` when the request provides resolved payloads; every other
+* image (non-user roles, nested tool results, or a text-only route without
+* payloads) still fails loud instead of silently dropping the image.
+*/
+function assertImagesSupported(message, imagesByRef) {
+	for (const block of message.content) if (block.type === "tool-result" && contentHasImage(block.content)) throw new LlmError("The Ollama adapter does not support tool-result image content.", "UNSUPPORTED_CONTENT");
+	const hasTopLevelImage = message.content.some((block) => block.type === "image");
+	if (message.role !== "user" && hasTopLevelImage) throw new LlmError(`The Ollama adapter does not support image content on ${message.role} messages.`, "UNSUPPORTED_CONTENT");
+	if (hasTopLevelImage && imagesByRef === void 0) throw new LlmError("The Ollama adapter does not support image content for this model.", "UNSUPPORTED_CONTENT");
+}
+/** Collect the base64 payloads for one user message's top-level image blocks. */
+function imagesOf(message, imagesByRef) {
+	if (imagesByRef === void 0) return [];
+	const images = [];
+	for (const block of message.content) {
+		if (block.type !== "image") continue;
+		const data = imagesByRef.get(String(block.attachment.attachmentId));
+		if (data === void 0) throw new LlmError("The Ollama adapter could not resolve an image payload.", "UNSUPPORTED_CONTENT");
+		images.push(data);
+	}
+	return images;
 }
 /**
 * Parse a tool-call argument string into the object Ollama expects. The raw
@@ -468,7 +497,7 @@ function toolNameOf(callId, namesByCallId) {
 * @param messages - the harness conversation, in order.
 * @returns the wire messages; order preserved, each tool result expanded into its own entry.
 */
-function serializeMessages(messages) {
+function serializeMessages(messages, imagesByRef) {
 	const namesByCallId = /* @__PURE__ */ new Map();
 	for (const message of messages) {
 		if (message.role !== "assistant") continue;
@@ -476,7 +505,8 @@ function serializeMessages(messages) {
 	}
 	const wire = [];
 	for (const message of messages) {
-		assertTextOnly(message.content);
+		assertImagesSupported(message, imagesByRef);
+		const images = imagesOf(message, imagesByRef);
 		if (message.role === "system") {
 			wire.push({
 				role: "system",
@@ -490,9 +520,10 @@ function serializeMessages(messages) {
 		}
 		const toolResults = message.content.filter((block) => block.type === "tool-result");
 		const text = flattenText(message.content);
-		if (text.length > 0 || toolResults.length === 0) wire.push({
+		if (text.length > 0 || toolResults.length === 0 || images.length > 0) wire.push({
 			role: "user",
-			content: text
+			content: text,
+			...images.length > 0 ? { images } : {}
 		});
 		for (const result of toolResults) {
 			const name = toolNameOf(String(result.toolCallId), namesByCallId);
@@ -514,7 +545,7 @@ function serializeMessages(messages) {
 * @param resolved - the resolved plugin config.
 * @returns the `/api/chat` request body.
 */
-function serializeRequest(options, resolved) {
+function serializeRequest(options, resolved, imagesByRef) {
 	const mapping = resolved.models.find((entry) => entry.name === options.model);
 	const model = mapping?.model ?? options.model;
 	const messages = [];
@@ -522,7 +553,7 @@ function serializeRequest(options, resolved) {
 		role: "system",
 		content: options.system
 	});
-	messages.push(...serializeMessages(options.messages));
+	messages.push(...serializeMessages(options.messages, imagesByRef));
 	const temperature = options.temperature ?? mapping?.temperature ?? resolved.temperature;
 	const ollamaOptions = {};
 	if (temperature !== void 0) ollamaOptions.temperature = temperature;
@@ -855,25 +886,35 @@ function _usingCtx() {
 * resolution (one `() => ResolvedConfig` thunk re-read per operation), so a
 * changed base URL, model mapping, or timeout reaches the next request without
 * re-registration, while an in-flight stream keeps the facts it started with.
-* The adapter is text-only (`inputModalities: ['text']`); tool calls and tool
-* results are translated by {@link serializeRequest}.
+* Input modalities follow the model's reported `/api/show` capabilities
+* (`vision` → `['text', 'image']`) unless the `vision` config knob opts out;
+* image payloads resolve through the optional `attachments` service.
 * @module dsh-local-ai/adapter
 */
 /** Idle-timeout abort code stamped onto a stalled stream's timeout reason. */
 const STREAM_IDLE_TIMEOUT_CODE = "LLM_STREAM_IDLE_TIMEOUT";
+/**
+* Deterministic request-image policy: aspect-preserving projection at ≤ 4 MP
+* and a 10 MiB encoded-byte cap per image (a protocol default, not a tunable —
+* the attachment service owns admission limits).
+*/
+const REQUEST_IMAGE_POLICY = {
+	maxPixels: 4194304,
+	maxBytes: 10485760
+};
 /** The single provider route this adapter owns. */
 const OLLAMA_PROVIDER = "ollama";
 /** Reverse a model mapping: Ollama model id → harness-visible name. */
 function harnessNameOf(resolved, ollamaName) {
 	return resolved.models.find((entry) => entry.model === ollamaName)?.name ?? ollamaName;
 }
-/** Advertise one configured or discovered local model as text-only. */
-function modelInfo(provider, id, name) {
+/** Advertise one configured or discovered local model with its input modalities. */
+function modelInfo(provider, id, name, vision) {
 	return {
 		provider,
 		id,
 		name,
-		inputModalities: ["text"]
+		inputModalities: vision ? ["text", "image"] : ["text"]
 	};
 }
 /**
@@ -890,27 +931,45 @@ var OllamaAdapter = class extends LlmAdapter {
 	fetchImpl() {
 		return this.options.fetchImpl ?? ((input, init) => globalThis.fetch(input, init));
 	}
+	/** Probe one model's `/api/show` capabilities; failures degrade to text-only. */
+	async visionOf(resolved, name, signal) {
+		if (!resolved.vision) return false;
+		try {
+			return hasVision((await showModel(resolved.baseURL, name, this.fetchImpl(), signal)).capabilities);
+		} catch {
+			return false;
+		}
+	}
 	providerInfo(provider) {
 		return {
 			id: provider,
 			name: "Ollama (local)"
 		};
 	}
-	listModels(provider) {
+	async listModels(provider) {
 		const resolved = this.options.config();
-		return listModels(resolved.baseURL, this.fetchImpl()).then((models) => models.map((model) => modelInfo(provider, harnessNameOf(resolved, model.name), harnessNameOf(resolved, model.name)))).catch(() => resolved.models.map((entry) => modelInfo(provider, entry.name, entry.name)));
+		const models = await listModels(resolved.baseURL, this.fetchImpl()).catch(() => []);
+		const list = models.length > 0 ? models : resolved.models.map((entry) => ({
+			name: entry.model,
+			model: entry.model,
+			size: 0,
+			digest: ""
+		}));
+		const probes = await Promise.all(list.map((model) => this.visionOf(resolved, model.name)));
+		return list.map((model, index) => modelInfo(provider, harnessNameOf(resolved, model.name), harnessNameOf(resolved, model.name), probes[index] === true));
 	}
-	resolveModel(provider, model, _signal) {
+	async resolveModel(provider, model, signal) {
 		const resolved = this.options.config();
 		const mapping = resolved.models.find((entry) => entry.name === model);
-		return Promise.resolve({
+		const ollamaName = mapping?.model ?? model;
+		return {
 			provider,
 			id: model,
 			name: model,
-			inputModalities: ["text"],
+			inputModalities: await this.visionOf(resolved, ollamaName, signal) ? ["text", "image"] : ["text"],
 			context: { contextWindow: mapping?.contextWindow ?? resolved.defaultContextWindow },
 			defaultMaxTokens: mapping?.maxTokens ?? resolved.maxTokens
-		});
+		};
 	}
 	async *stream(options) {
 		try {
@@ -948,7 +1007,7 @@ var OllamaAdapter = class extends LlmAdapter {
 		}
 	}
 	async *request(options, resolved, signal) {
-		const body = serializeRequest(options, resolved);
+		const body = serializeRequest(options, resolved, await this.prepareImages(options, resolved, signal));
 		let response;
 		try {
 			response = await postStream(resolved.baseURL, "/api/chat", body, this.fetchImpl(), signal);
@@ -960,6 +1019,26 @@ var OllamaAdapter = class extends LlmAdapter {
 		if (!response.body) throw new LlmError("Ollama API returned no response body", "EMPTY_RESPONSE");
 		yield* translate(readNdjsonLines(response.body));
 	}
+	/**
+	* Resolve request-image payloads for one image-bearing request. A text-only
+	* route (config opt-out), a missing attachment service, or an unresolvable
+	* ref fails loud here — images are never silently dropped.
+	* @returns a `attachmentId → base64` map, or `undefined` when the request has no images.
+	*/
+	async prepareImages(options, resolved, signal) {
+		if (!options.messages.some((message) => contentHasImage(message.content))) return void 0;
+		if (!resolved.vision) throw new LlmError("The Ollama adapter does not support image content for this model (vision disabled).", "UNSUPPORTED_CONTENT");
+		const attachments = this.options.resolveAttachments?.();
+		if (attachments === void 0) throw new LlmError("Image content requires the attachments service; mount a profile with @deepseek-ai/dsh-attachment.", "UNSUPPORTED_CONTENT");
+		const map = /* @__PURE__ */ new Map();
+		for (const message of options.messages) for (const block of message.content) {
+			if (block.type !== "image") continue;
+			const ref = block.attachment;
+			const requestImage = await attachments.readImageRequest(ref, REQUEST_IMAGE_POLICY, signal);
+			map.set(String(ref.attachmentId), Buffer.from(requestImage.data).toString("base64"));
+		}
+		return map;
+	}
 };
 //#endregion
 //#region src/route.ts
@@ -1156,7 +1235,7 @@ async function checkHealth(baseURL, fetchImpl, subprocess, requestTimeoutMs, gra
 * The `dsh-local-ai` plugin version. The release script bumps this string
 * alongside `package.json`; `test/version.spec.ts` trips when the two drift.
 */
-const VERSION = "0.1.3";
+const VERSION = "0.1.5";
 //#endregion
 //#region src/index.ts
 const name = "local-ai";
@@ -1232,7 +1311,8 @@ function apply(ctx, config = {}) {
 	const fetchImpl = (input, init) => globalThis.fetch(input, init);
 	const adapter = new OllamaAdapter({
 		config: () => resolved,
-		fetchImpl
+		fetchImpl,
+		resolveAttachments: () => ctx.get("attachments")
 	});
 	ctx.llm.registerAdapter([OLLAMA_PROVIDER], adapter);
 	ctx.on("llm/stream", (options, next) => {

package/lib/types/adapter.d.ts

@@ -4,12 +4,14 @@
  * resolution (one `() => ResolvedConfig` thunk re-read per operation), so a
  * changed base URL, model mapping, or timeout reaches the next request without
  * re-registration, while an in-flight stream keeps the facts it started with.
- * The adapter is text-only (`inputModalities: ['text']`); tool calls and tool
- * results are translated by {@link serializeRequest}.
+ * Input modalities follow the model's reported `/api/show` capabilities
+ * (`vision` → `['text', 'image']`) unless the `vision` config knob opts out;
+ * image payloads resolve through the optional `attachments` service.
  * @module dsh-local-ai/adapter
  */
 import { LlmAdapter } from '@deepseek-ai/dsh-llm';
 import type { GenerateOptions, LlmModelInfo, LlmProviderInfo, LlmResolvedModelInfo, StreamChunk } from '@deepseek-ai/dsh-llm';
+import type { AttachmentStore } from '@deepseek-ai/dsh-attachment';
 import type { FetchLike } from './ollama.js';
 import type { ResolvedConfig } from './config.js';
 /** Constructor options for {@link OllamaAdapter}. */
@@ -18,6 +20,8 @@ export interface OllamaAdapterOptions {
     config: () => ResolvedConfig;
     /** Fetch implementation, injectable for tests; defaults to `globalThis.fetch`. */
     fetchImpl?: FetchLike;
+    /** Optional attachment service resolving durable image refs to request bytes. */
+    resolveAttachments?: () => AttachmentStore | undefined;
 }
 /** The single provider route this adapter owns. */
 export declare const OLLAMA_PROVIDER = "ollama";
@@ -30,10 +34,19 @@ export declare class OllamaAdapter extends LlmAdapter {
     private readonly options;
     constructor(options: OllamaAdapterOptions);
     private fetchImpl;
+    /** Probe one model's `/api/show` capabilities; failures degrade to text-only. */
+    private visionOf;
     providerInfo(provider: string): LlmProviderInfo;
     listModels(provider: string): Promise<readonly LlmModelInfo[]>;
-    resolveModel(provider: string, model: string, _signal?: AbortSignal): Promise<LlmResolvedModelInfo>;
+    resolveModel(provider: string, model: string, signal?: AbortSignal): Promise<LlmResolvedModelInfo>;
     stream(options: GenerateOptions): AsyncIterable<StreamChunk>;
     private request;
+    /**
+     * Resolve request-image payloads for one image-bearing request. A text-only
+     * route (config opt-out), a missing attachment service, or an unresolvable
+     * ref fails loud here — images are never silently dropped.
+     * @returns a `attachmentId → base64` map, or `undefined` when the request has no images.
+     */
+    private prepareImages;
 }
 //# sourceMappingURL=adapter.d.ts.map

package/lib/types/adapter.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"adapter.d.ts","sourceRoot":"","sources":["../../src/adapter.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,EAAE,UAAU,EAAY,MAAM,sBAAsB,CAAA;AAC3D,OAAO,KAAK,EAAE,eAAe,EAAE,YAAY,EAAE,eAAe,EAAE,oBAAoB,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAA;AAI7H,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,aAAa,CAAA;AAG5C,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,aAAa,CAAA;AAKjD,qDAAqD;AACrD,MAAM,WAAW,oBAAoB;IACnC,2DAA2D;IAC3D,MAAM,EAAE,MAAM,cAAc,CAAA;IAC5B,kFAAkF;IAClF,SAAS,CAAC,EAAE,SAAS,CAAA;CACtB;AAED,mDAAmD;AACnD,eAAO,MAAM,eAAe,WAAW,CAAA;AAavC;;;;GAIG;AACH,qBAAa,aAAc,SAAQ,UAAU;IAC/B,OAAO,CAAC,QAAQ,CAAC,OAAO;gBAAP,OAAO,EAAE,oBAAoB;IAI1D,OAAO,CAAC,SAAS;IAIR,YAAY,CAAC,QAAQ,EAAE,MAAM,GAAG,eAAe;IAI/C,UAAU,CAAC,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,SAAS,YAAY,EAAE,CAAC;IAO9D,YAAY,CACnB,QAAQ,EAAE,MAAM,EAChB,KAAK,EAAE,MAAM,EACb,OAAO,CAAC,EAAE,WAAW,GACpB,OAAO,CAAC,oBAAoB,CAAC;IAaxB,MAAM,CAAC,OAAO,EAAE,eAAe,GAAG,aAAa,CAAC,WAAW,CAAC;YA8CpD,OAAO;CA0BxB"}
+{"version":3,"file":"adapter.d.ts","sourceRoot":"","sources":["../../src/adapter.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAmB,UAAU,EAAY,MAAM,sBAAsB,CAAA;AAC5E,OAAO,KAAK,EAAE,eAAe,EAAE,YAAY,EAAE,eAAe,EAAE,oBAAoB,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAA;AAC7H,OAAO,KAAK,EAAE,eAAe,EAAsB,MAAM,6BAA6B,CAAA;AAItF,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,aAAa,CAAA;AAG5C,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,aAAa,CAAA;AAYjD,qDAAqD;AACrD,MAAM,WAAW,oBAAoB;IACnC,2DAA2D;IAC3D,MAAM,EAAE,MAAM,cAAc,CAAA;IAC5B,kFAAkF;IAClF,SAAS,CAAC,EAAE,SAAS,CAAA;IACrB,iFAAiF;IACjF,kBAAkB,CAAC,EAAE,MAAM,eAAe,GAAG,SAAS,CAAA;CACvD;AAED,mDAAmD;AACnD,eAAO,MAAM,eAAe,WAAW,CAAA;AAavC;;;;GAIG;AACH,qBAAa,aAAc,SAAQ,UAAU;IAC/B,OAAO,CAAC,QAAQ,CAAC,OAAO;gBAAP,OAAO,EAAE,oBAAoB;IAI1D,OAAO,CAAC,SAAS;IAIjB,iFAAiF;YACnE,QAAQ;IAUb,YAAY,CAAC,QAAQ,EAAE,MAAM,GAAG,eAAe;IAIzC,UAAU,CAAC,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,SAAS,YAAY,EAAE,CAAC;IAa9D,YAAY,CACzB,QAAQ,EAAE,MAAM,EAChB,KAAK,EAAE,MAAM,EACb,MAAM,CAAC,EAAE,WAAW,GACnB,OAAO,CAAC,oBAAoB,CAAC;IAexB,MAAM,CAAC,OAAO,EAAE,eAAe,GAAG,aAAa,CAAC,WAAW,CAAC;YA8CpD,OAAO;IA4BvB;;;;;OAKG;YACW,aAAa;CAwB5B"}