Settings → Models: per-model reasoningEfforts missing from the editor; Fetch available models never consults the live endpoint for catalog routes

[Ali0600]

Title: Settings → Models: hand-added models get no Effort picker (reasoningEfforts not settable in the UI), and "Fetch available models" never consults the live endpoint for known providers

Environment

• DeepSeek Harness @deepseek-ai/dsh@0.1.1-rc.2 (installed via npx), web GUI
• macOS 14 (Apple Silicon), OpenRouter route with API key

Steps to reproduce

1. Find an OpenRouter model that is not in your build's shipped catalog — e.g. any newly listed or stealth/* model. At 0.1.1-rc.2 the bundled catalog (@earendil-works/pi-ai@0.82.1, dist/providers/data/openrouter.json) has 276 OpenRouter models while the live /api/v1/models lists 421+, including stealth/ox-alpha.
2. In Settings → Models → OpenRouter → Customized settings, add the model (id, name, context window, max tokens). It works — requests succeed.
3. Look at the composer's model picker: there is no Effort control for this model, even though the model reasons by default (OpenRouter metadata: reasoning.mandatory = true).

Observed behavior / root cause

• A hand-added entry has no installed-catalog twin, so resolveModelReasoning() in @deepseek-ai/dsh-llm-pi-ai resolves it as { reasoning: false } unless the entry declares reasoningEfforts.

• The Models page deliberately offers no effort field — its source comments explain that effort is a per-model capability and a per-provider control would be wrong. Agreed — but that argues for a per-model field, which is exactly what's missing.

• The only workaround is hand-editing ~/.dsh/settings.yaml:

  - id: stealth/ox-alpha
    name: Ox Alpha
    reasoningEfforts:
      low: low      # key = selectable level, value = wire spelling
      high: high    # dispatched as reasoning.effort on openrouter-format routes
      max: max

  After adding that block the Effort menu renders Low/High/Max correctly. The schema rules (allowed keys, omitted ⇒ unsupported, only off: may be empty) exist only in source docstrings, not user-facing docs.

• Separately, "Fetch available models" cannot help: discoverModels() returns catalogModels(provider) whenever the provider ships a catalog, so for known providers it re-lists the same bundled entries and never hits the network. Live listing only happens for custom providers where the user typed a base URL.

Requests

1. Expose reasoningEfforts as a per-model field in the Settings → Models editor (the models grid already edits id/name/contextWindow/maxTokens).
2. Give "Fetch available models" a live option for catalog routes — or fall back to {baseURL}/models when the user explicitly asks to fetch — so models newer than the package build can be discovered from the GUI. The live-fetch code path already exists and parses OpenAI-style { data: [...] } responses fine.
3. Consider surfacing a hint when a hand-added model is silently treated as non-reasoning ("no effort levels declared").

Happy to split into separate issues if you prefer. A community writeup of the full workaround lives here: https://github.com/Ali0600/dsh-openrouter-models.