dsh-local-ai 0.2.10 → 0.2.11

package/CHANGELOG.md

@@ -6,6 +6,19 @@ 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).
 
 
+## [0.2.11] - 2026-09-18
+
+### Fixed
+
+- **A local route is never silently moved to the cloud by the image-offload circuit.** When the local stream failed with `IMAGE_OFFLOAD_REQUIRED` before producing content, the routing fallback treated it like any other failure and re-ran the request on the cloud — skipping the official offload circuit and leaking a local-only request to a remote provider. That code is now rethrown unchanged (both the finish-reason and the thrown-error paths), so the caller can offload the cited images and retry the same local route. Every other pre-content failure still falls back to the cloud.
+
+### Changed
+
+- Migrate to the renamed attachment contract: `@deepseek-ai/dsh-attachment` replaced `ImageRequestPolicy` (`maxPixels`/`maxBytes`) with a per-route `ImageRequestTarget` (`width`/`height`/`maxBytes`). The adapter now requests a 2048×2048 target with a 10 MiB byte cap, which preserves the previous 4 Mi-pixel / 10 MiB projection exactly (a target at or above the source keeps the source dimensions).
+- `/api/show` capability probes are cached for `visionCacheTtlMs` (default 30 s, `0` disables caching) instead of being issued once per request; `ollama_pull` and `ollama_remove` invalidate the cached entry for the affected model, and failures are cached for the window too so a down server is not probed per request.
+- Align the dev/test dependency graph to `0.1.6-alpha.2`: `dsh-agent`, `dsh-user-approval`, `dsh-invariants` and `dsh-scope` were still resolving to `0.1.1-rc.2` as transitive peers, which broke the strict published-types pass (`CallId` no longer exists on `dsh-llm`). Declaring them explicitly removes the mixed-generation graph.
+- Declare `dsh.manifestVersion: 1` and the canonical three-clause `engines.dsh`; refresh the `AGENTS.md`/CI baseline wording to the pinned line.
+
 ## [0.2.10] - 2026-09-12
 
 ### Changed

package/README-es.md

@@ -8,6 +8,7 @@
 *Descubre, descarga, elimina e inspecciona modelos locales, enruta solicitudes hacia ellos por tipo de tarea o palabra clave con respaldo automático a la nube, y obtén un resumen de estado de un vistazo con `/ollama`.*
 
 [![License](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](LICENSE)
+[![Gitee](https://img.shields.io/badge/Gitee-mirror-c71d23?logo=gitee)](https://gitee.com/perrylink/dsh-local-ai)
 [![DSH plugin](https://img.shields.io/badge/dsh--plugin-✅-green)](https://github.com/topics/dsh-plugin)
 [![dsh-doctor](https://raw.githubusercontent.com/PerryLink/dsh-plugin-doctor/main/badges/PerryLink__dsh-local-ai.svg)](https://github.com/PerryLink/dsh-plugin-doctor#verified-徽章)
 [![Node](https://img.shields.io/badge/node-%5E22.19%20%7C%7C%20%3E%3D24-brightgreen.svg)](#)
@@ -26,7 +27,7 @@
 
 | Superficie | Estado |
 |---|---|
-| Harness | DeepSeek Harness `dsh-v0.1.5-rc.2` (adaptado el 2026-09-09): el sobre de sesión conserva su campo ignorable solo para compatibilidad de lectura de logs almacenados - Session.append aún no puede estamparlo, por lo que el comportamiento de la puerta no cambia. Verificado el 2026-09-11 contra el checkout master dsh-v0.1.5-rc.2 (cadena de gates completa + smoke de instalación del profile). |
+| Harness | DeepSeek Harness `dsh-v0.1.6-alpha.2` (verificado el 2026-09-18: doble typecheck + 133 pruebas + puertas self-contained/artifacts). El rango de peers admite todas las líneas soportadas: `>=0.1.2-rc.1 <0.2.0 \|\| >=0.1.5-alpha.1 <0.2.0 \|\| >=0.1.6-0 <0.2.0`; los pines dev/test son `0.1.6-alpha.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 |
@@ -105,6 +106,7 @@ Todos los ajustes son campos `Config` de Schemastery (modificables desde cordis.
 | `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 |
+| `visionCacheTtlMs` | `30000` | Milisegundos que se conserva en caché la sonda `/api/show` (`0` desactiva la caché; pull/remove invalida ese modelo) |
 | `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 |
@@ -149,7 +151,7 @@ Todos los ajustes son campos `Config` de Schemastery (modificables desde cordis.
 - **Sin re-enrutamiento por defecto** — la lista `route` está vacía salvo que optes por lo contrario; una solicitud llega a un modelo local solo por una regla explícita o una selección explícita del proveedor `ollama`.
 - **Desinfección antes de mostrar** — las direcciones de endpoint y las rutas locales se desinfectan antes de llegar a la salida de herramientas, al comando `/ollama` o a los mensajes de error.
 - **Cero modelos empaquetados** — descargas y almacenamiento son responsabilidad de Ollama; nada se envía en el paquete.
-- **Fallo alto, fallo contenido** — la configuración inválida hace fallar el montaje; una ruta local que falla antes de producir contenido respalda a la nube (`next()`), de modo que un Ollama caído nunca bloquea una conversación.
+- **Fallo alto, fallo contenido** — la configuración inválida hace fallar el montaje; una ruta local que falla antes de producir contenido respalda a la nube (`next()`), de modo que un Ollama caído nunca bloquea una conversación. **Única excepción:** un fallo con `IMAGE_OFFLOAD_REQUIRED` se relanza en vez de reintentarse en la nube: ese código es el circuito oficial de descarga de imágenes pidiendo a esta misma ruta local que suelte imágenes retenidas, y el respaldo saltaría el circuito enviando en silencio una solicitud solo-local a un proveedor remoto.
 - **Visible para el modelo ⟺ registrado** — el enrutamiento solo cambia qué proveedor sirve una solicitud (el mensaje del asistente se registra con su procedencia `ollama`); no se inventa entrada visible nueva.
 
 ## Known limitations
@@ -185,7 +187,7 @@ pnpm pack           # el tarball publicado
 
 ## PerryLink DSH Plugin Family
 
-Este proyecto es uno de los [37 complementos de DeepSeek Harness](https://github.com/PerryLink) mantenidos por [PerryLink](https://github.com/PerryLink). Si este te ayuda, probablemente los demás también:
+Este proyecto es uno de los [40 complementos de DeepSeek Harness](https://github.com/PerryLink) mantenidos por [PerryLink](https://github.com/PerryLink). Si este te ayuda, probablemente los demás también:
 
 | Plugin | One-liner |
 |---|---|
@@ -214,6 +216,7 @@ Este proyecto es uno de los [37 complementos de DeepSeek Harness](https://github
 | **[dsh-permission-rules](https://github.com/PerryLink/dsh-permission-rules)** | Reglas de permisos declarativas allow/deny/ask estilo Claude Code con auditoría | |
 | **[dsh-personal-directive](https://github.com/PerryLink/dsh-personal-directive)** | Inyector de directivas personales con interruptor en la barra superior (edición framework) |
 | **[dsh-plugin-guide](https://github.com/PerryLink/dsh-plugin-guide)** | Base de conocimiento de desarrollo de plugins como habilidad de agente bajo demanda | |
+| **[dsh-plugin-doctor](https://github.com/PerryLink/dsh-plugin-doctor)** | Zero-dependency static + sandbox smoke detector for DSH plugins | |
 | **[dsh-reach](https://github.com/PerryLink/dsh-reach)** | Puente multicanal de aprobación/preguntas: WeChat/Telegram/Feishu, consola de sesión |
 | **[dsh-research-report](https://github.com/PerryLink/dsh-research-report)** | Motor de informes de investigación verificables con evidencia direccionada por contenido | |
 | **[dsh-score](https://github.com/PerryLink/dsh-score)** | Puntuación de calidad multidimensional para plugins de DeepSeek Harness. | |
@@ -224,7 +227,16 @@ Este proyecto es uno de los [37 complementos de DeepSeek Harness](https://github
 | **[dsh-test-drive](https://github.com/PerryLink/dsh-test-drive)** | Pruebas de instalación y humo aisladas para plugins de DeepSeek Harness. | |
 | **[dsh-ticktick](https://github.com/PerryLink/dsh-ticktick)** | Puente de tareas TickTick/Dida365: panel de cabecera de sesión + 11 herramientas |
 | **[dsh-translate](https://github.com/PerryLink/dsh-translate)** | Traducción de parámetros entre proveedores y reparación determinista de JSON para DeepSeek Harness. | |
-| **[dsh-wechat](https://github.com/PerryLink/dsh-wechat)** | Puente WeChat ↔ DSH (bot Tencent iLink): texto/imagen/archivo/voz, aprobaciones en el chat |
+| **[dsh-wechat](https://github.com/pan17/dsh-wechat)** | Puente WeChat ↔ DSH (bot Tencent iLink): texto/imagen/archivo/voz, aprobaciones en el chat |
+| **[dsh-autotier](https://github.com/PerryLink/dsh-autotier)** | Automatic strong/cheap model-tier routing with deterministic risk guards and a `/tier` command | |
+| **[dsh-catalog](https://github.com/PerryLink/dsh-catalog)** | DSH Desktop Market standard catalog source for the PerryLink family | |
+| **[dsh-cert-mcp](https://github.com/PerryLink/dsh-cert-mcp)** | Read-only MCP server exposing the certification registry: grades, snapshots and five-dimension evidence | |
+| **[dsh-kit](https://github.com/PerryLink/dsh-kit)** | One-command starter pack that installs the core family | |
+| **[dsh-plugin-certification](https://github.com/PerryLink/dsh-plugin-certification)** | Community certification registry with repro-checkable grades and badges | |
+| **[dsh-plugin-kit](https://github.com/PerryLink/dsh-plugin-kit)** | Shared zero-runtime-dependency toolkit for the PerryLink DSH plugins | |
+| **[dsh-plugin-portal](https://github.com/PerryLink/dsh-plugin-portal)** | Zero-dependency static portal rendering the whole plugin family as one page | |
+| **[dsh-plugin-upgrade-015](https://github.com/PerryLink/dsh-plugin-upgrade-015)** | Merged `0.1.3-alpha.1` → `0.1.5-rc.1` upgrade corridor card plus a zero-dependency seam scanner | |
+| **[dsh-team-rooms](https://github.com/PerryLink/dsh-team-rooms)** | Cross-session team rooms: shared message bus, task board and timeline | |
 
 ## License
 

package/README-hi.md

@@ -8,6 +8,7 @@
 *स्थानीय मॉडल खोजें, खींचें, हटाएँ और जाँचें; कार्य-प्रकार या कीवर्ड के आधार पर अनुरोधों को स्थानीय मॉडल पर रूट करें और विफलता पर स्वतः क्लाउड पर वापस लौटें; `/ollama` से एक-झटके में स्थिति सारांश पाएँ।*
 
 [![License](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](LICENSE)
+[![Gitee](https://img.shields.io/badge/Gitee-mirror-c71d23?logo=gitee)](https://gitee.com/perrylink/dsh-local-ai)
 [![DSH plugin](https://img.shields.io/badge/dsh--plugin-✅-green)](https://github.com/topics/dsh-plugin)
 [![dsh-doctor](https://raw.githubusercontent.com/PerryLink/dsh-plugin-doctor/main/badges/PerryLink__dsh-local-ai.svg)](https://github.com/PerryLink/dsh-plugin-doctor#verified-徽章)
 [![Node](https://img.shields.io/badge/node-%5E22.19%20%7C%7C%20%3E%3D24-brightgreen.svg)](#)
@@ -26,7 +27,7 @@
 
 | सतह | स्थिति |
 |---|---|
-| Harness | DeepSeek Harness `dsh-v0.1.5-rc.2` (2026-09-09 को अनुकूलित): सत्र लिफ़ाफ़ा अपना ignorable फ़ील्ड केवल संग्रहीत-लॉग पठन संगतता के लिए रखता है - Session.append अभी भी इसे स्टैम्प नहीं कर सकता, इसलिए गेट व्यवहार अपरिवर्तित है। 2026-09-11 को dsh-v0.1.5-rc.2 master checkout के विरुद्ध सत्यापित (पूर्ण गेट श्रृंखला + profile इंस्टॉल smoke)। |
+| Harness | DeepSeek Harness `dsh-v0.1.6-alpha.2` (2026-09-18 को सत्यापित: दोहरा typecheck + 133 टेस्ट + self-contained/artifacts गेट)। peer रेंज सभी समर्थित लाइनों को स्वीकार करती है: `>=0.1.2-rc.1 <0.2.0 \|\| >=0.1.5-alpha.1 <0.2.0 \|\| >=0.1.6-0 <0.2.0`; dev/test पिन `0.1.6-alpha.2` हैं। |
 | 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 |
@@ -105,6 +106,7 @@ dsh --profile web --dump-config | grep -A2 'id: dsh-local-ai'
 | `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 |
+| `visionCacheTtlMs` | `30000` | `/api/show` क्षमता जाँच कितने मिलीसेकंड कैश रहे (`0` = कैश बंद; pull/remove उस मॉडल को अमान्य करता है) |
 | `models` | `[]` | Harness-दृश्य नाम → Ollama मॉडल मैपिंग |
 | `models[].name` | *(required)* | Harness-दृश्य मॉडल नाम (`GenerateOptions.model`) |
 | `models[].model` | `= name` | Ollama मॉडल id |
@@ -149,7 +151,7 @@ dsh --profile web --dump-config | grep -A2 'id: dsh-local-ai'
 - **डिफ़ॉल्ट रूप से कोई पुनः-रूटिंग नहीं** — `route` सूची तब तक खाली है जब तक आप सहमति न दें; अनुरोध स्थानीय मॉडल तक केवल किसी स्पष्ट नियम या `ollama` प्रदाता के स्पष्ट चयन से पहुँचता है।
 - **दिखाने से पहले सैनिटाइज़** — एंडपॉइंट पते व स्थानीय पथ टूल आउटपुट, `/ollama` कमांड या त्रुटि संदेशों तक पहुँचने से पहले सैनिटाइज़ होते हैं।
 - **शून्य बंडल मॉडल** — डाउनलोड व स्टोरेज Ollama की अपनी ज़िम्मेदारी है; पैकेज में कुछ नहीं भेजा जाता।
-- **विफलता ज़ोरदार, विफलता सीमित** — अमान्य कॉन्फ़िग माउंट को विफल करता है; सामग्री बनाने से पहले विफल स्थानीय रूट क्लाउड पर वापस लौटता है (`next()`), इसलिए बंद Ollama कभी बातचीत को अटकाता नहीं।
+- **विफलता ज़ोरदार, विफलता सीमित** — अमान्य कॉन्फ़िग माउंट को विफल करता है; सामग्री बनाने से पहले विफल स्थानीय रूट क्लाउड पर वापस लौटता है (`next()`), इसलिए बंद Ollama कभी बातचीत को अटकाता नहीं। **एकमात्र अपवाद:** `IMAGE_OFFLOAD_REQUIRED` वाली विफलता क्लाउड पर दोबारा नहीं भेजी जाती बल्कि पुनः थ्रो होती है — यह कोड आधिकारिक इमेज-ऑफ़लोड सर्किट है जो इसी स्थानीय रूट से रखी गई छवियाँ हटाने को कहता है, और फ़ॉलबैक सर्किट को छोड़कर केवल-स्थानीय अनुरोध चुपचाप दूरस्थ प्रदाता को भेज देता।
 - **मॉडल-दृश्य ⟺ लॉग** — रूटिंग केवल यह बदलती है कि कौन-सा प्रदाता अनुरोध सेवा करता है (assistant संदेश अपनी `ollama` उत्पत्ति के साथ लॉग होता है); कोई नया मॉडल-दृश्य इनपुट नहीं गढ़ा जाता।
 
 ## Known limitations
@@ -185,7 +187,7 @@ pnpm pack           # प्रकाशित tarball
 
 ## PerryLink DSH Plugin Family
 
-यह प्रोजेक्ट [PerryLink](https://github.com/PerryLink) द्वारा अनुरक्षित [37 DeepSeek Harness प्लगइनों](https://github.com/PerryLink) में से एक है। अगर यह आपकी मदद करता है, तो बाकी भी करेंगे:
+यह प्रोजेक्ट [PerryLink](https://github.com/PerryLink) द्वारा अनुरक्षित [40 DeepSeek Harness प्लगइनों](https://github.com/PerryLink) में से एक है। अगर यह आपकी मदद करता है, तो बाकी भी करेंगे:
 
 | Plugin | One-liner |
 |---|---|
@@ -214,6 +216,7 @@ pnpm pack           # प्रकाशित tarball
 | **[dsh-permission-rules](https://github.com/PerryLink/dsh-permission-rules)** | ऑडिट के साथ Claude Code-शैली घोषणात्मक allow/deny/ask अनुमति नियम | |
 | **[dsh-personal-directive](https://github.com/PerryLink/dsh-personal-directive)** | शीर्ष-बार टॉगल के साथ व्यक्तिगत निर्देश इंजेक्टर (फ्रेमवर्क संस्करण) |
 | **[dsh-plugin-guide](https://github.com/PerryLink/dsh-plugin-guide)** | माँग पर एजेंट कौशल के रूप में प्लगइन-विकास ज्ञान आधार | |
+| **[dsh-plugin-doctor](https://github.com/PerryLink/dsh-plugin-doctor)** | Zero-dependency static + sandbox smoke detector for DSH plugins | |
 | **[dsh-reach](https://github.com/PerryLink/dsh-reach)** | मल्टी-चैनल अनुमोदन/प्रश्न ब्रिज: WeChat/Telegram/Feishu, सत्र कंसोल |
 | **[dsh-research-report](https://github.com/PerryLink/dsh-research-report)** | सामग्री-पता साक्ष्य और सीलबंद संस्करणों वाला सत्यापन-योग्य अनुसंधान-रिपोर्ट इंजन | |
 | **[dsh-score](https://github.com/PerryLink/dsh-score)** | DeepSeek Harness प्लगिनों की बहु-आयामी गुणवत्ता स्कोरिंग। | |
@@ -224,7 +227,16 @@ pnpm pack           # प्रकाशित tarball
 | **[dsh-test-drive](https://github.com/PerryLink/dsh-test-drive)** | DeepSeek Harness प्लगिनों के लिए पृथक इंस्टॉल-एंड-स्मोक टेस्ट ड्राइव। | |
 | **[dsh-ticktick](https://github.com/PerryLink/dsh-ticktick)** | TickTick/Dida365 कार्य ब्रिज: सत्र-हेडर पैनल + 11 टूल |
 | **[dsh-translate](https://github.com/PerryLink/dsh-translate)** | DeepSeek Harness के लिए वेंडर पैरामीटर अनुवाद और नियतात्मक JSON मरम्मत। | |
-| **[dsh-wechat](https://github.com/PerryLink/dsh-wechat)** | WeChat ↔ DSH ब्रिज (Tencent iLink bot): टेक्स्ट/इमेज/फ़ाइल/आवाज़, चैट में अनुमोदन |
+| **[dsh-wechat](https://github.com/pan17/dsh-wechat)** | WeChat ↔ DSH ब्रिज (Tencent iLink bot): टेक्स्ट/इमेज/फ़ाइल/आवाज़, चैट में अनुमोदन |
+| **[dsh-autotier](https://github.com/PerryLink/dsh-autotier)** | Automatic strong/cheap model-tier routing with deterministic risk guards and a `/tier` command | |
+| **[dsh-catalog](https://github.com/PerryLink/dsh-catalog)** | DSH Desktop Market standard catalog source for the PerryLink family | |
+| **[dsh-cert-mcp](https://github.com/PerryLink/dsh-cert-mcp)** | Read-only MCP server exposing the certification registry: grades, snapshots and five-dimension evidence | |
+| **[dsh-kit](https://github.com/PerryLink/dsh-kit)** | One-command starter pack that installs the core family | |
+| **[dsh-plugin-certification](https://github.com/PerryLink/dsh-plugin-certification)** | Community certification registry with repro-checkable grades and badges | |
+| **[dsh-plugin-kit](https://github.com/PerryLink/dsh-plugin-kit)** | Shared zero-runtime-dependency toolkit for the PerryLink DSH plugins | |
+| **[dsh-plugin-portal](https://github.com/PerryLink/dsh-plugin-portal)** | Zero-dependency static portal rendering the whole plugin family as one page | |
+| **[dsh-plugin-upgrade-015](https://github.com/PerryLink/dsh-plugin-upgrade-015)** | Merged `0.1.3-alpha.1` → `0.1.5-rc.1` upgrade corridor card plus a zero-dependency seam scanner | |
+| **[dsh-team-rooms](https://github.com/PerryLink/dsh-team-rooms)** | Cross-session team rooms: shared message bus, task board and timeline | |
 
 ## License
 

package/README-pt.md

@@ -8,6 +8,7 @@
 *Descubra, baixe, remova e inspecione modelos locais, roteie solicitações para eles por tipo de tarefa ou palavra-chave com fallback automático para a nuvem e obtenha um resumo de status de relance com `/ollama`.*
 
 [![License](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](LICENSE)
+[![Gitee](https://img.shields.io/badge/Gitee-mirror-c71d23?logo=gitee)](https://gitee.com/perrylink/dsh-local-ai)
 [![DSH plugin](https://img.shields.io/badge/dsh--plugin-✅-green)](https://github.com/topics/dsh-plugin)
 [![dsh-doctor](https://raw.githubusercontent.com/PerryLink/dsh-plugin-doctor/main/badges/PerryLink__dsh-local-ai.svg)](https://github.com/PerryLink/dsh-plugin-doctor#verified-徽章)
 [![Node](https://img.shields.io/badge/node-%5E22.19%20%7C%7C%20%3E%3D24-brightgreen.svg)](#)
@@ -26,7 +27,7 @@
 
 | Superfície | Status |
 |---|---|
-| Harness | DeepSeek Harness `dsh-v0.1.5-rc.2` (adaptado em 2026-09-09): o envelope de sessão mantém seu campo ignorable apenas para compatibilidade de leitura de logs armazenados - o Session.append ainda não consegue estampá-lo, então o comportamento da porta não muda. Verificado em 2026-09-11 contra o checkout master dsh-v0.1.5-rc.2 (cadeia de gates completa + smoke de instalação do profile). |
+| Harness | DeepSeek Harness `dsh-v0.1.6-alpha.2` (verificado em 2026-09-18: typecheck duplo + 133 testes + portas self-contained/artifacts). O intervalo de peers admite todas as linhas suportadas: `>=0.1.2-rc.1 <0.2.0 \|\| >=0.1.5-alpha.1 <0.2.0 \|\| >=0.1.6-0 <0.2.0`; os pins dev/test são `0.1.6-alpha.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 |
@@ -105,6 +106,7 @@ Todos os ajustes são campos `Config` de Schemastery (modificáveis pelo cordis.
 | `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 |
+| `visionCacheTtlMs` | `30000` | Milissegundos em que a sondagem `/api/show` fica em cache (`0` desativa; pull/remove invalida esse modelo) |
 | `models` | `[]` | Mapeamentos nome visível → modelo Ollama |
 | `models[].name` | *(required)* | Nome de modelo visível no harness (`GenerateOptions.model`) |
 | `models[].model` | `= name` | Id do modelo Ollama |
@@ -149,7 +151,7 @@ Todos os ajustes são campos `Config` de Schemastery (modificáveis pelo cordis.
 - **Sem re-roteamento por padrão** — a lista `route` fica vazia salvo se você optar; uma solicitação chega a um modelo local apenas por uma regra explícita ou seleção explícita do provedor `ollama`.
 - **Sanitização antes de exibir** — endereços de endpoint e caminhos locais são sanitizados antes de chegar à saída de ferramentas, ao comando `/ollama` ou a mensagens de erro.
 - **Zero modelos empacotados** — downloads e armazenamento são responsabilidade do Ollama; nada é enviado no pacote.
-- **Falha alta, falha contida** — configuração inválida faz o montagem falhar; uma rota local que falha antes de produzir conteúdo faz fallback para a nuvem (`next()`), de modo que um Ollama caído nunca trava uma conversa.
+- **Falha alta, falha contida** — configuração inválida faz o montagem falhar; uma rota local que falha antes de produzir conteúdo faz fallback para a nuvem (`next()`), de modo que um Ollama caído nunca trava uma conversa. **Única exceção:** uma falha com `IMAGE_OFFLOAD_REQUIRED` é relançada em vez de repetida na nuvem — esse código é o circuito oficial de descarga de imagens pedindo a esta mesma rota local que libere imagens retidas, e o fallback pularia o circuito enviando em silêncio uma solicitação somente-local a um provedor remoto.
 - **Visível ao modelo ⟺ registrado** — o roteamento só muda qual provedor atende uma solicitação (a mensagem do assistente é registrada com sua proveniência `ollama`); nenhuma entrada visível nova é inventada.
 
 ## Known limitations
@@ -185,7 +187,7 @@ pnpm pack           # o tarball publicado
 
 ## PerryLink DSH Plugin Family
 
-Este projeto é um dos [37 plugins de DeepSeek Harness](https://github.com/PerryLink) mantidos por [PerryLink](https://github.com/PerryLink). Se este ajuda você, os outros provavelmente também:
+Este projeto é um dos [40 plugins de DeepSeek Harness](https://github.com/PerryLink) mantidos por [PerryLink](https://github.com/PerryLink). Se este ajuda você, os outros provavelmente também:
 
 | Plugin | One-liner |
 |---|---|
@@ -214,6 +216,7 @@ Este projeto é um dos [37 plugins de DeepSeek Harness](https://github.com/Perry
 | **[dsh-permission-rules](https://github.com/PerryLink/dsh-permission-rules)** | Regras de permissão declarativas allow/deny/ask estilo Claude Code com auditoria | |
 | **[dsh-personal-directive](https://github.com/PerryLink/dsh-personal-directive)** | Injetor de diretivas pessoais com alternância na barra superior (edição framework) |
 | **[dsh-plugin-guide](https://github.com/PerryLink/dsh-plugin-guide)** | Base de conhecimento de desenvolvimento de plugins como habilidade de agente sob demanda | |
+| **[dsh-plugin-doctor](https://github.com/PerryLink/dsh-plugin-doctor)** | Zero-dependency static + sandbox smoke detector for DSH plugins | |
 | **[dsh-reach](https://github.com/PerryLink/dsh-reach)** | Ponte multicanal de aprovação/perguntas: WeChat/Telegram/Feishu, console de sessão |
 | **[dsh-research-report](https://github.com/PerryLink/dsh-research-report)** | Motor de relatórios de pesquisa verificáveis com evidência endereçada por conteúdo | |
 | **[dsh-score](https://github.com/PerryLink/dsh-score)** | Pontuação de qualidade multidimensional para plugins de DeepSeek Harness. | |
@@ -224,7 +227,16 @@ Este projeto é um dos [37 plugins de DeepSeek Harness](https://github.com/Perry
 | **[dsh-test-drive](https://github.com/PerryLink/dsh-test-drive)** | Test drives isolados de instalação e smoke para plugins de DeepSeek Harness. | |
 | **[dsh-ticktick](https://github.com/PerryLink/dsh-ticktick)** | Ponte de tarefas TickTick/Dida365: painel no cabeçalho da sessão + 11 ferramentas |
 | **[dsh-translate](https://github.com/PerryLink/dsh-translate)** | Tradução de parâmetros entre fornecedores e reparo determinístico de JSON para DeepSeek Harness. | |
-| **[dsh-wechat](https://github.com/PerryLink/dsh-wechat)** | Ponte WeChat ↔ DSH (bot Tencent iLink): texto/imagem/arquivo/voz, aprovações no chat |
+| **[dsh-wechat](https://github.com/pan17/dsh-wechat)** | Ponte WeChat ↔ DSH (bot Tencent iLink): texto/imagem/arquivo/voz, aprovações no chat |
+| **[dsh-autotier](https://github.com/PerryLink/dsh-autotier)** | Automatic strong/cheap model-tier routing with deterministic risk guards and a `/tier` command | |
+| **[dsh-catalog](https://github.com/PerryLink/dsh-catalog)** | DSH Desktop Market standard catalog source for the PerryLink family | |
+| **[dsh-cert-mcp](https://github.com/PerryLink/dsh-cert-mcp)** | Read-only MCP server exposing the certification registry: grades, snapshots and five-dimension evidence | |
+| **[dsh-kit](https://github.com/PerryLink/dsh-kit)** | One-command starter pack that installs the core family | |
+| **[dsh-plugin-certification](https://github.com/PerryLink/dsh-plugin-certification)** | Community certification registry with repro-checkable grades and badges | |
+| **[dsh-plugin-kit](https://github.com/PerryLink/dsh-plugin-kit)** | Shared zero-runtime-dependency toolkit for the PerryLink DSH plugins | |
+| **[dsh-plugin-portal](https://github.com/PerryLink/dsh-plugin-portal)** | Zero-dependency static portal rendering the whole plugin family as one page | |
+| **[dsh-plugin-upgrade-015](https://github.com/PerryLink/dsh-plugin-upgrade-015)** | Merged `0.1.3-alpha.1` → `0.1.5-rc.1` upgrade corridor card plus a zero-dependency seam scanner | |
+| **[dsh-team-rooms](https://github.com/PerryLink/dsh-team-rooms)** | Cross-session team rooms: shared message bus, task board and timeline | |
 
 ## License
 

package/README-zh.md

@@ -10,6 +10,7 @@
 > **官方仓库。** 本仓库是 dsh-local-ai 的唯一官方仓库，由 PerryLink 维护。其他账号下的同名仓库与本项目无关。
 
 [![License](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](LICENSE)
+[![Gitee](https://img.shields.io/badge/Gitee-mirror-c71d23?logo=gitee)](https://gitee.com/perrylink/dsh-local-ai)
 [![DSH plugin](https://img.shields.io/badge/dsh--plugin-✅-green)](https://github.com/topics/dsh-plugin)
 [![dsh-doctor](https://raw.githubusercontent.com/PerryLink/dsh-plugin-doctor/main/badges/PerryLink__dsh-local-ai.svg)](https://github.com/PerryLink/dsh-plugin-doctor#verified-徽章)
 [![Node](https://img.shields.io/badge/node-%5E22.19%20%7C%7C%20%3E%3D24-brightgreen.svg)](#)
@@ -28,7 +29,7 @@
 
 | 项目 | 状态 |
 |---|---|
-| Harness | DeepSeek Harness `dsh-v0.1.5-rc.2`（2026-09-09 已适配）：会话信封保留 ignorable 字段但仅用于存量日志读取兼容——Session.append 仍无法盖章，门控行为不变。已于 2026-09-11 对照 dsh-v0.1.5-rc.2 master checkout 核验（完整门控链 + profile 安装冒烟）。 |
+| Harness | DeepSeek Harness `dsh-v0.1.6-alpha.2`（2026-09-18 核验：双 typecheck 尺子 + 133 项测试 + self-contained/artifacts 门）。peer 范围接纳全部受支持线：`>=0.1.2-rc.1 <0.2.0 \|\| >=0.1.5-alpha.1 <0.2.0 \|\| >=0.1.6-0 <0.2.0`；dev/test 钉号 `0.1.6-alpha.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 |
@@ -107,6 +108,7 @@ dsh --profile web --dump-config | grep -A2 'id: dsh-local-ai'
 | `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 |
+| `visionCacheTtlMs` | `30000` | `/api/show` 能力探测的缓存毫秒数（`0` 关闭缓存；pull/remove 会使该模型失效） |
 | `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 |
@@ -151,7 +153,7 @@ dsh --profile web --dump-config | grep -A2 'id: dsh-local-ai'
 - **默认不重路由** —— `route` 列表默认为空；请求只有通过显式规则或显式选择 `ollama` 提供方才会到达本地模型。
 - **展示前脱敏** —— 端点地址与本地路径在进入工具输出、`/ollama` 命令或错误消息前都会被脱敏。
 - **零捆绑模型** —— 下载与存储是 Ollama 自己的事；包内不含任何模型。
-- **失败响亮、失败可控** —— 非法配置导致挂载失败；本地路由在产出内容前失败会回退云端（`next()`），因此 Ollama 宕机不会卡死对话。
+- **失败响亮、失败可控** —— 非法配置导致挂载失败；本地路由在产出内容前失败会回退云端（`next()`），因此 Ollama 宕机不会卡死对话。**唯一例外**：带 `IMAGE_OFFLOAD_REQUIRED` 的失败会被重新抛出而不回退云端——该码是官方图像卸载回路在要求这条本地路由卸下保留图片，回退会跳过回路并把本应只在本地的请求静默发往远端提供方。
 - **模型可见 ⟺ 已记录** —— 路由只改变由哪个提供方服务请求（assistant 消息会以 `ollama` 来源记录）；不凭空新增模型可见输入。
 
 ## Known limitations
@@ -187,7 +189,7 @@ pnpm pack           # 发布用 tarball
 
 ## PerryLink DSH Plugin Family
 
-这是 [PerryLink](https://github.com/PerryLink) 维护的 [37 个 DeepSeek Harness 插件](https://github.com/PerryLink) 之一。如果它能帮到你，其他的也会：
+这是 [PerryLink](https://github.com/PerryLink) 维护的 [40 个 DeepSeek Harness 插件](https://github.com/PerryLink) 之一。如果它能帮到你，其他的也会：
 
 | Plugin | One-liner |
 |---|---|
@@ -216,6 +218,7 @@ pnpm pack           # 发布用 tarball
 | **[dsh-permission-rules](https://github.com/PerryLink/dsh-permission-rules)** | Claude Code 风格声明式 allow/deny/ask 权限规则，带审计 | |
 | **[dsh-personal-directive](https://github.com/PerryLink/dsh-personal-directive)** | 个人指令注入器：顶栏开关（框架版） |
 | **[dsh-plugin-guide](https://github.com/PerryLink/dsh-plugin-guide)** | 作为按需代理技能的插件开发知识库 | |
+| **[dsh-plugin-doctor](https://github.com/PerryLink/dsh-plugin-doctor)** | Zero-dependency static + sandbox smoke detector for DSH plugins | |
 | **[dsh-reach](https://github.com/PerryLink/dsh-reach)** | 多渠道审批/提问桥接：微信/Telegram/飞书，会话控制台 |
 | **[dsh-research-report](https://github.com/PerryLink/dsh-research-report)** | 可验证研究报告引擎：内容寻址证据账本与封存版本 | |
 | **[dsh-score](https://github.com/PerryLink/dsh-score)** | DeepSeek Harness 插件的多维质量评分。 | |
@@ -226,7 +229,16 @@ pnpm pack           # 发布用 tarball
 | **[dsh-test-drive](https://github.com/PerryLink/dsh-test-drive)** | DeepSeek Harness 插件的隔离试装冒烟。 | |
 | **[dsh-ticktick](https://github.com/PerryLink/dsh-ticktick)** | TickTick/滴答清单任务桥接：会话头面板 + 11 个工具 |
 | **[dsh-translate](https://github.com/PerryLink/dsh-translate)** | DeepSeek Harness 的厂商参数翻译与确定性 JSON 修复。 | |
-| **[dsh-wechat](https://github.com/PerryLink/dsh-wechat)** | 微信 ↔ DSH 桥接（Tencent iLink 机器人）：文本/图片/文件/语音，聊天内审批卡片 |
+| **[dsh-wechat](https://github.com/pan17/dsh-wechat)** | 微信 ↔ DSH 桥接（Tencent iLink 机器人）：文本/图片/文件/语音，聊天内审批卡片 |
+| **[dsh-autotier](https://github.com/PerryLink/dsh-autotier)** | Automatic strong/cheap model-tier routing with deterministic risk guards and a `/tier` command | |
+| **[dsh-catalog](https://github.com/PerryLink/dsh-catalog)** | DSH Desktop Market standard catalog source for the PerryLink family | |
+| **[dsh-cert-mcp](https://github.com/PerryLink/dsh-cert-mcp)** | Read-only MCP server exposing the certification registry: grades, snapshots and five-dimension evidence | |
+| **[dsh-kit](https://github.com/PerryLink/dsh-kit)** | One-command starter pack that installs the core family | |
+| **[dsh-plugin-certification](https://github.com/PerryLink/dsh-plugin-certification)** | Community certification registry with repro-checkable grades and badges | |
+| **[dsh-plugin-kit](https://github.com/PerryLink/dsh-plugin-kit)** | Shared zero-runtime-dependency toolkit for the PerryLink DSH plugins | |
+| **[dsh-plugin-portal](https://github.com/PerryLink/dsh-plugin-portal)** | Zero-dependency static portal rendering the whole plugin family as one page | |
+| **[dsh-plugin-upgrade-015](https://github.com/PerryLink/dsh-plugin-upgrade-015)** | Merged `0.1.3-alpha.1` → `0.1.5-rc.1` upgrade corridor card plus a zero-dependency seam scanner | |
+| **[dsh-team-rooms](https://github.com/PerryLink/dsh-team-rooms)** | Cross-session team rooms: shared message bus, task board and timeline | |
 
 ## License
 

package/README.md

@@ -29,7 +29,7 @@
 
 | Surface | Status |
 |---|---|
-| Harness | DeepSeek Harness `dsh-v0.1.5-rc.2` (adapted 2026-09-09): the session envelope keeps its ignorable field for stored-log read compatibility only - Session.append still cannot stamp it, so audit-gate behavior is unchanged. Verified 2026-09-11 against the dsh-v0.1.5-rc.2 master checkout (full gate chain + profile install smoke). |
+| Harness | DeepSeek Harness `dsh-v0.1.6-alpha.2` (verified 2026-09-18: dual typecheck rulers + 133 tests + self-contained/artifacts gates). The peer range admits every supported line: `>=0.1.2-rc.1 <0.2.0 \|\| >=0.1.5-alpha.1 <0.2.0 \|\| >=0.1.6-0 <0.2.0`; dev/test pins are `0.1.6-alpha.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 |
@@ -110,6 +110,7 @@ All tunables are Schemastery `Config` fields (changeable from cordis.yml). An id
 | `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 |
+| `visionCacheTtlMs` | `30000` | Milliseconds a `/api/show` capability probe stays cached (`0` disables caching; a pull or remove invalidates that model) |
 | `models` | `[]` | Harness-visible → Ollama model mappings |
 | `models[].name` | *(required)* | Harness-visible model name (`GenerateOptions.model`) |
 | `models[].model` | `= name` | Ollama model id |
@@ -154,7 +155,7 @@ All tunables are Schemastery `Config` fields (changeable from cordis.yml). An id
 - **No re-routing by default** — the `route` list is empty unless you opt in; a request reaches a local model only through an explicit rule or an explicit `ollama` provider selection.
 - **Sanitize before display** — endpoint addresses and local paths are sanitized before they reach tool output, the `/ollama` command, or error messages.
 - **Zero bundled models** — downloads and storage are Ollama's own responsibility; nothing is shipped in the package.
-- **Failure loud, failure contained** — invalid config fails the mount; a local route that fails before producing content falls back to the cloud (`next()`), so a down Ollama never bricks a conversation.
+- **Failure loud, failure contained** — invalid config fails the mount; a local route that fails before producing content falls back to the cloud (`next()`), so a down Ollama never bricks a conversation. **One carve-out:** a failure carrying `IMAGE_OFFLOAD_REQUIRED` is rethrown instead of retried on the cloud — that code is the official offload circuit asking this same local route to drop retained images, and falling back would skip the circuit while silently sending a local-only request to a remote provider.
 - **Model-visible ⟺ logged** — routing only changes which provider serves a request (the assistant message is logged with its `ollama` provenance); no new model-visible input is invented.
 
 ## Known limitations
@@ -190,7 +191,7 @@ pnpm pack           # the published tarball
 
 ## PerryLink DSH Plugin Family
 
-This project is one of the [37 DeepSeek Harness plugins](https://github.com/PerryLink) maintained by [PerryLink](https://github.com/PerryLink). If this one helps you, the others likely will too:
+This project is one of the [40 DeepSeek Harness plugins](https://github.com/PerryLink) maintained by [PerryLink](https://github.com/PerryLink). If this one helps you, the others likely will too:
 
 | Plugin | One-liner |
 |---|---|
@@ -219,6 +220,7 @@ This project is one of the [37 DeepSeek Harness plugins](https://github.com/Perr
 | **[dsh-permission-rules](https://github.com/PerryLink/dsh-permission-rules)** | Claude Code-style declarative allow/deny/ask permission rules with audit | |
 | **[dsh-personal-directive](https://github.com/PerryLink/dsh-personal-directive)** | Personal directive injector with top-bar toggle (framework edition) |
 | **[dsh-plugin-guide](https://github.com/PerryLink/dsh-plugin-guide)** | Plugin-development knowledge base as an on-demand agent skill | |
+| **[dsh-plugin-doctor](https://github.com/PerryLink/dsh-plugin-doctor)** | Zero-dependency static + sandbox smoke detector for DSH plugins | |
 | **[dsh-reach](https://github.com/PerryLink/dsh-reach)** | Multi-channel approval/question bridge: WeChat/Telegram/Feishu, session console |
 | **[dsh-research-report](https://github.com/PerryLink/dsh-research-report)** | Verifiable research-report engine: content-addressed evidence ledger and sealed versions | |
 | **[dsh-score](https://github.com/PerryLink/dsh-score)** | Multi-dimensional quality scoring for DeepSeek Harness plugins. | |
@@ -229,7 +231,16 @@ This project is one of the [37 DeepSeek Harness plugins](https://github.com/Perr
 | **[dsh-test-drive](https://github.com/PerryLink/dsh-test-drive)** | Isolated install-and-smoke test drives for DeepSeek Harness plugins. | |
 | **[dsh-ticktick](https://github.com/PerryLink/dsh-ticktick)** | TickTick/Dida365 task bridge: session-header panel + 11 tools |
 | **[dsh-translate](https://github.com/PerryLink/dsh-translate)** | Vendor parameter translation and deterministic JSON repair for DeepSeek Harness. | |
-| **[dsh-wechat](https://github.com/PerryLink/dsh-wechat)** | WeChat ↔ DSH bridge (Tencent iLink bot): text/image/file/voice, approvals in chat |
+| **[dsh-wechat](https://github.com/pan17/dsh-wechat)** | WeChat ↔ DSH bridge (Tencent iLink bot): text/image/file/voice, approvals in chat |
+| **[dsh-autotier](https://github.com/PerryLink/dsh-autotier)** | Automatic strong/cheap model-tier routing with deterministic risk guards and a `/tier` command | |
+| **[dsh-catalog](https://github.com/PerryLink/dsh-catalog)** | DSH Desktop Market standard catalog source for the PerryLink family | |
+| **[dsh-cert-mcp](https://github.com/PerryLink/dsh-cert-mcp)** | Read-only MCP server exposing the certification registry: grades, snapshots and five-dimension evidence | |
+| **[dsh-kit](https://github.com/PerryLink/dsh-kit)** | One-command starter pack that installs the core family | |
+| **[dsh-plugin-certification](https://github.com/PerryLink/dsh-plugin-certification)** | Community certification registry with repro-checkable grades and badges | |
+| **[dsh-plugin-kit](https://github.com/PerryLink/dsh-plugin-kit)** | Shared zero-runtime-dependency toolkit for the PerryLink DSH plugins | |
+| **[dsh-plugin-portal](https://github.com/PerryLink/dsh-plugin-portal)** | Zero-dependency static portal rendering the whole plugin family as one page | |
+| **[dsh-plugin-upgrade-015](https://github.com/PerryLink/dsh-plugin-upgrade-015)** | Merged `0.1.3-alpha.1` → `0.1.5-rc.1` upgrade corridor card plus a zero-dependency seam scanner | |
+| **[dsh-team-rooms](https://github.com/PerryLink/dsh-team-rooms)** | Cross-session team rooms: shared message bus, task board and timeline | |
 
 ### Install from the DSH Desktop Market
 

package/cordis.patch.yml

@@ -23,6 +23,9 @@
         # temperature: 0.7
 # Declare image support when the model reports vision; false keeps text-only.
 # vision: true
+        # Milliseconds a /api/show capability probe stays cached (0 disables
+        # caching; pull/remove invalidate the entry for that model).
+        # visionCacheTtlMs: 30000
         # 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

@@ -1,6 +1,6 @@
 import { defineTool } from "@deepseek-ai/dsh-tools";
 import z from "@deepseek-ai/schemastery";
-import { EMPTY_RESPONSE_CODE, LlmAdapter, LlmError, attributionHeaders, contentHasImage } from "@deepseek-ai/dsh-llm";
+import { EMPTY_RESPONSE_CODE, IMAGE_OFFLOAD_REQUIRED_CODE, LlmAdapter, LlmError, attributionHeaders, contentHasImage } from "@deepseek-ai/dsh-llm";
 import { idleWatchdog, timeoutOf } from "@deepseek-ai/dsh-timeout";
 import { tmpdir } from "node:os";
 //#region src/config.ts
@@ -24,6 +24,7 @@ const Config = z.object({
 	maxTokens: z.number().default(4096),
 	temperature: z.number(),
 	vision: z.boolean().default(true),
+	visionCacheTtlMs: z.number().default(3e4),
 	models: z.array(z.object({
 		name: z.string().required(),
 		model: z.string(),
@@ -57,6 +58,10 @@ const Config = z.object({
 function assertPositiveInt(name, value) {
 	if (!Number.isSafeInteger(value) || value <= 0) throw new TypeError(`${name} must be a positive safe integer, got ${String(value)}`);
 }
+/** Throw unless `value` is a non-negative safe integer (`0` disables a cache). */
+function assertNonNegativeInt(name, value) {
+	if (!Number.isSafeInteger(value) || value < 0) throw new TypeError(`${name} must be a non-negative safe integer, got ${String(value)}`);
+}
 /** Throw unless `value` is a finite number in `[min, max]`. */
 function assertFiniteRange(name, value, min, max) {
 	if (typeof value !== "number" || !Number.isFinite(value) || value < min || value > max) throw new TypeError(`${name} must be a finite number in [${min}, ${max}], got ${String(value)}`);
@@ -101,6 +106,8 @@ function resolveConfig(config = {}) {
 	const temperature = config.temperature;
 	if (temperature !== void 0) assertFiniteRange("temperature", temperature, 0, 2);
 	const vision = config.vision ?? true;
+	const visionCacheTtlMs = config.visionCacheTtlMs ?? 3e4;
+	assertNonNegativeInt("visionCacheTtlMs", visionCacheTtlMs);
 	const models = resolveMappings(config.models, "models");
 	const backends = (config.backends ?? []).map((backend, index) => {
 		if (typeof backend.name !== "string" || backend.name.trim().length === 0) throw new TypeError(`backends[${index}].name must be a non-empty string`);
@@ -149,6 +156,7 @@ function resolveConfig(config = {}) {
 		maxTokens,
 		...temperature === void 0 ? {} : { temperature },
 		vision,
+		visionCacheTtlMs,
 		models,
 		backends,
 		route
@@ -956,12 +964,18 @@ function _usingCtx() {
 /** Idle-timeout abort code stamped onto a stalled stream's timeout reason. */
 const STREAM_IDLE_TIMEOUT_CODE$1 = "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).
+* Deterministic request-image target for this adapter's image projection. The
+* 0.1.6 attachment service replaced the per-request policy object
+* (`ImageRequestPolicy`: `maxPixels`/`maxBytes`) with a per-route target
+* (`ImageRequestTarget`: `width`/`height`/`maxBytes`), where a target at or
+* above the source keeps the source dimensions. 2048×2048 therefore preserves
+* the previous 4 Mi-pixel ceiling exactly, and `maxBytes` keeps the 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,
+const REQUEST_IMAGE_TARGET = {
+	width: 2048,
+	height: 2048,
 	maxBytes: 10485760
 };
 /** The single provider route this adapter owns. */
@@ -986,6 +1000,12 @@ function modelInfo(provider, id, name, vision) {
 */
 var OllamaAdapter = class extends LlmAdapter {
 	options;
+	/**
+	* `/api/show` capability cache keyed by `baseURL|model`, so a conversation
+	* that resolves the same model repeatedly probes once per TTL window instead
+	* of once per request. Pull/remove invalidate explicitly.
+	*/
+	visionCache = /* @__PURE__ */ new Map();
 	constructor(options) {
 		super();
 		this.options = options;
@@ -993,14 +1013,38 @@ var OllamaAdapter = class extends LlmAdapter {
 	fetchImpl() {
 		return this.options.fetchImpl ?? ((input, init) => globalThis.fetch(input, init));
 	}
+	/**
+	* Drop cached capability probes. Called after a model is pulled or removed,
+	* where the cached answer is known to be stale.
+	* @param name - the Ollama model id, or undefined to clear every entry.
+	*/
+	invalidateVision(name) {
+		if (name === void 0) {
+			this.visionCache.clear();
+			return;
+		}
+		for (const key of [...this.visionCache.keys()]) if (key.endsWith(`|${name}`)) this.visionCache.delete(key);
+	}
 	/** Probe one model's `/api/show` capabilities; failures degrade to text-only. */
 	async visionOf(resolved, name, signal) {
 		if (!resolved.vision) return false;
+		const key = `${resolved.baseURL}|${name}`;
+		const ttl = resolved.visionCacheTtlMs;
+		if (ttl > 0) {
+			const cached = this.visionCache.get(key);
+			if (cached !== void 0 && Date.now() - cached.at < ttl) return cached.vision;
+		}
+		let vision = false;
 		try {
-			return hasVision((await showModel(resolved.baseURL, name, this.fetchImpl(), signal)).capabilities);
+			vision = hasVision((await showModel(resolved.baseURL, name, this.fetchImpl(), signal)).capabilities);
 		} catch {
-			return false;
+			vision = false;
 		}
+		if (ttl > 0) this.visionCache.set(key, {
+			vision,
+			at: Date.now()
+		});
+		return vision;
 	}
 	providerInfo(provider) {
 		return {
@@ -1096,7 +1140,7 @@ var OllamaAdapter = class extends LlmAdapter {
 		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);
+			const requestImage = await attachments.readImageRequest(ref, REQUEST_IMAGE_TARGET, signal);
 			map.set(String(ref.attachmentId), Buffer.from(requestImage.data).toString("base64"));
 		}
 		return map;
@@ -1606,6 +1650,16 @@ var OpenAICompatibleAdapter = class extends LlmAdapter {
 //#endregion
 //#region src/route.ts
 /**
+* Whether a local-route failure is the official image-offload circuit asking
+* THIS route to drop retained images and retry: the request fits nowhere until
+* the named occurrences are offloaded.
+* @param failure - the finish reason's failure facts, when present.
+* @returns true for `IMAGE_OFFLOAD_REQUIRED`.
+*/
+function isImageOffloadRequired(failure) {
+	return failure?.code === IMAGE_OFFLOAD_REQUIRED_CODE;
+}
+/**
 * Whether a stream chunk carries visible model output for fallback timing.
 * The host moved this helper out of `@deepseek-ai/dsh-llm` on 0.1.2-alpha.2
 * (it now lives in the client ui-chat package), so the plugin pins the same
@@ -1680,6 +1734,11 @@ function decideRoute(options, resolved) {
 * local route keeps retry and failure normalization). If the local route
 * finishes with an error or aborts before any token delta, `next()` (the
 * cloud) is streamed instead; otherwise the local stream is forwarded.
+*
+* One failure is deliberately NOT retried on the cloud: `IMAGE_OFFLOAD_REQUIRED`
+* is the official image-offload circuit asking this same route to drop retained
+* images. Falling back there would skip the circuit and silently move a
+* local-only request to a remote provider, so the signal is rethrown instead.
 * @param streamLocal - produces the local stream for a re-routed request.
 * @param options - the original request.
 * @param decision - the local model to route to.
@@ -1702,6 +1761,7 @@ async function* routeLocal(streamLocal, options, decision, next) {
 			}
 			if (chunk.type === "finish") {
 				if (chunk.reason.kind === "error" || chunk.reason.kind === "aborted") {
+					if (isImageOffloadRequired(chunk.reason.failure)) throw new LlmError(chunk.reason.failure.message, chunk.reason.failure.code);
 					yield* next();
 					return;
 				}
@@ -1717,12 +1777,13 @@ async function* routeLocal(streamLocal, options, decision, next) {
 		}
 		for (const buffered of pending) yield buffered;
 	} catch (error) {
-		if (!producedContent) {
-			yield* next();
-			return;
+		if (producedContent) {
+			for (const buffered of pending) yield buffered;
+			throw error;
 		}
-		for (const buffered of pending) yield buffered;
-		throw error;
+		if (isImageOffloadRequired(error)) throw error;
+		yield* next();
+		return;
 	}
 }
 //#endregion
@@ -1818,7 +1879,7 @@ 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 = "0.2.10";
+const VERSION = "0.2.11";
 //#endregion
 //#region src/index.ts
 const name = "local-ai";
@@ -1983,9 +2044,11 @@ function apply(ctx, config = {}) {
 		},
 		async execute(args, exec) {
 			const name = args.name;
+			const result = await pullModel(resolved.baseURL, name, fetchImpl, exec.signal);
+			adapter.invalidateVision(name);
 			return {
 				name,
-				status: (await pullModel(resolved.baseURL, name, fetchImpl, exec.signal)).status
+				status: result.status
 			};
 		}
 	}));
@@ -2007,6 +2070,7 @@ function apply(ctx, config = {}) {
 		async execute(args, exec) {
 			const name = args.name;
 			await removeModel(resolved.baseURL, name, fetchImpl, exec.signal);
+			adapter.invalidateVision(name);
 			return {
 				name,
 				removed: true