dsh-llm-capabilities
DSH 插件：为 llm-pi-ai 自动检测并配置模型能力（reasoningEfforts + 输入模态）。dsh-reasoning-efforts 的后继者。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-llm-capabilities@0.1.4Llm Capabilities 以 dsh-llm-capabilities 发布，当前版本为 0.1.4。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
DSH plugin: auto-detect and configure model capabilities (reasoningEfforts + input modalities) for
llm-pi-ai. Successor todsh-reasoning-efforts— one panel patches the two fields the officialllm-pi-aiUI leaves unconfigurable for self-hosted gateways.
English | 简体中文
The official llm-pi-ai settings page for a custom provider only lets you edit contextWindow. Two capabilities that actually matter for routing stay hidden:
reasoningEfforts) — which off/minimal/low/medium/high/xhigh/max the model accepts, and the wire spelling (max: ultra) each level maps to. Without this the intensity slider says “this model provides no selectable thinking levels”.input: ["text","image"]) — whether the model accepts images. Without this a gateway model that does do vision is still ["text"] locally, so every image attachment is rejected pre-flight with UNSUPPORTED_CONTENT before it ever reaches the endpoint.Both are per-model fields on providers.<route>.models[] that pi-ai already understands. This plugin just makes them editable and auto-detectable.
If you came from dsh-reasoning-efforts: keep it, it still works. dsh-llm-capabilities serves both endpoints (/model-capabilities/raw-models and legacy /thinking-levels/raw-models) so old clients keep working. New installs should use this package.
Settings → Model CapabilitiesGET {baseURL}/models server-side (credential resolved host-side, no key leak), parses:
supported_features: ["reasoning"], supported_parameters, supports_reasoning, reasoning_effort … → reasoningmodalities / input_modalities / supported_modalities / supports_vision / capabilities: ["vision"] … → visioncontext_length, max_output_tokens … → capacitiesllm.models() (pi-ai's own knowledge) as a complementary source — the endpoint wins for yes/no, the catalog refines the level set and stands in when the endpoint says nothing.not configured badge, applying will add it.inherit (omit input, use catalog/default) / image ✓ (["text","image"]) / text-only (["text"])settings.mutate with expectedRevision, deep-cloned models array, no other provider fields are touched.# published
dsh plugin --profile web add dsh-llm-capabilities
# local dev
dsh plugin --profile web add link:D:\AllCode\dsh\dsh-llm-capabilities
Requires DHS >= 0.1.1-rc.2, node >= 22.13.
Settings → Models as usual (route, baseURL, apiKeyEnv).Settings → Model Capabilities, pick the provider.Offer thinking levels → pick off/low/medium/high/xhigh/max, edit wire spellings inlineVision: inherit / image ✓ / text-onlyllm-pi-ai (providers.<route>.models). Intensity slider and image admission update immediately.x-opencode-session)OpenCode Go requires third-party agents to send a stable x-opencode-session
header for prompt-cache affinity (docs).
The host half wraps globalThis.fetch and sets it on requests whose URL contains
opencode.ai/zen/go (covers ocg-c / ocg-r / ocg-a). All other requests
pass through untouched.
ses_ + 32 hex derived from the dsh
conversation id (sha256), so the same chat keeps the same OpenCode session
across restarts. Title calls use an isolated namespace; compaction rides the
conversation's namespace because it replays history (reasoning
encrypted_content is bound to the account that issued it).400 … reasoning 'encrypted_content' was not issued to this caller
(session→account affinity lost on the gateway side), the wrapper retries
once with the replayed encrypted reasoning items stripped instead of
failing the turn. Opt out with sessionHeaders.encryptedContentRetry: false.user-agent (stays deepseek-harness/... — Go requires honest
identification) and authorization.cordis.patch.yml):- insert:
- id: model-capabilities
name: dsh-llm-capabilities
config:
sessionHeaders:
enabled: false
llm-pi-ai:
providers:
my-gateway:
baseURL: https://gateway.example/v1
api: openai-completions
models:
- id: gpt-4o
input: [text, image] # vision
reasoningEfforts: # thinking
off: null
low: low
medium: medium
high: high
- id: text-only-model
input: [text]
reasoningEfforts: false # explicitly no reasoning
llm.discoverModels is intentionally narrowed host-side to id/name/contextWindow/maxTokens. This plugin adds a host route that returns the raw listing:
GET /model-capabilities/raw-models?route=<route>
It reads baseURL/apiKeyEnv from your own llm-pi-ai settings, resolves the credential via ctx.credentials, fetches {baseURL}/models, and returns {ok, data} without ever echoing the secret. Only routes you already configured are reachable (browser trust fence: rejects cross-site, checks Origin/Host).
pnpm install
pnpm run typecheck
pnpm run build
pnpm run dev # watch
The web bundle is a window.__ModuleLoader__.load({id, factory}) closure — not ESM — so keep client imports to platform modules (react, @deepseek-ai/dsh-client-*) only.
| dsh-reasoning-efforts | dsh-llm-capabilities (this) | |
|---|---|---|
| reasoning auto-detect | ✅ | ✅ (ported, strict TS) |
| vision auto-detect | ❌ | ✅ |
| input write | ❌ | ✅ |
| host route | /thinking-levels/raw-models | /model-capabilities/raw-models + legacy compat |
Migrate when ready — both can coexist during the transition.
MIT © hank reed (bamboostrip)