dsh-llm-capabilities

DSH plugin: auto-detect and configure model capabilities (reasoningEfforts + capacities) for llm-pi-ai. Successor to dsh-reasoning-efforts — one panel patches the fields the official llm-pi-ai UI leaves unconfigurable for self-hosted gateways.

English | 简体中文

Why

The official llm-pi-ai settings page for a custom provider only lets you edit contextWindow. What actually matters for routing stays hidden:

1. Thinking levels (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”.

Vision (input) used to be the second gap, but the official UI declares input modalities itself now — so since 0.2.0 this plugin leaves input strictly alone (existing values are preserved untouched on every save).

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.

What it does

• One settings page Settings → Model Capabilities
• Auto-detect on selection: picking a provider fetches GET {baseURL}/models server-side (credential resolved host-side, no key leak) and parses:
  • supported_features: ["reasoning"], supported_parameters, supports_reasoning, reasoning_effort … → reasoning
  • context_length, max_output_tokens … → capacities
• Catalog-aware: merges llm.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.
• Configured state echoes back: re-detection never erases saved reasoningEfforts — models already configured show their own levels (custom wire spellings included) whether or not the endpoint lists signals.
• New models: endpoint lists a model you haven't configured yet → it shows up with a not configured badge, saving will add it.
• Vision preserved: input is never written or removed; whatever the official UI declared stays as-is.
• Sticky save bar: a single “Save changes” button pinned to the bottom of the list while you scroll — no scrolling back to the top; frosted-glass surface so list content never bleeds through.
• Safe writes: settings.mutate with expectedRevision, deep-cloned models array, no other provider fields are touched.

Install

# 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.

Usage

1. Add your provider in Settings → Models as usual (route, baseURL, apiKeyEnv).
2. Open Settings → Model Capabilities and pick the provider — detection runs automatically.
3. Toggle Offer thinking levels per model: pick off/low/medium/high/xhigh/max; customize wire spellings via the pencil icon when needed.
4. Save changes → writes to llm-pi-ai (providers.<route>.models). The intensity slider updates immediately; input/vision stays however the official UI set it.

Go session header (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.

• Conversation keying (default): 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).
• Encrypted-reasoning self-heal: if the gateway still answers 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.
• Never touched: user-agent (stays deepseek-harness/... — Go requires honest identification) and authorization.
• Disable without uninstalling (profile cordis.patch.yml):

- insert:
    - id: model-capabilities
      name: dsh-llm-capabilities
      config:
        sessionHeaders:
          enabled: false

Configuration shape written

llm-pi-ai:
  providers:
    my-gateway:
      baseURL: https://gateway.example/v1
      api: openai-completions
      models:
        - id: gpt-4o
          input: [text, image]          # vision — written by the official UI, preserved by this plugin
          reasoningEfforts:              # thinking — written by this plugin
            off: null
            low: low
            medium: medium
            high: high
        - id: text-only-model
          input: [text]
          reasoningEfforts: false        # explicitly no reasoning

How detection works

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).

Development

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.

Relation to dsh-reasoning-efforts

	dsh-reasoning-efforts	dsh-llm-capabilities (this)
reasoning auto-detect	✅	✅ (ported, strict TS)
vision	❌	handled by the official UI; input preserved on save
host route	/thinking-levels/raw-models	/model-capabilities/raw-models + legacy compat

Migrate when ready — both can coexist during the transition.

License

MIT © hank reed (bamboostrip)