recap-relay/startos/actions/setBackendRouting.ts

import { sdk } from '../sdk'
import { configFile } from '../file-models/config.json'

const { InputSpec, Value } = sdk

// Lets the operator tune which backend gets tried first per pipeline
// step (transcribe vs analyze) AND which Gemini SKU is used when
// Gemini is the backend. All four knobs live-reload — change them
// via this action and the next relay request honors the new values
// without a daemon restart.

const inputSpec = InputSpec.of({
  // ── Gemini model selection ──
  // Both fields are radio-select with curated options. The relay's
  // Gemini backend automatically falls back to lower-tier models in
  // this same list when the chosen one returns a 503 / capacity /
  // rate-limit error — see server/backends/gemini.js for the
  // fallback-chain logic.
  relay_gemini_transcription_model: Value.select({
    name: 'Gemini Transcription Model',
    description:
      "Primary Gemini SKU used when a transcription request is routed to Gemini. On 503/capacity/rate-limit failure, the relay falls back to lower-tier models in order (e.g. 3-flash → 2.5-flash → 2.0-flash).",
    default: 'gemini-3-flash-preview',
    values: {
      'gemini-3-flash-preview':
        'Gemini 3 Flash — latest, recommended (~$0.30/M in, $2.50/M out)',
      'gemini-2.5-flash':
        'Gemini 2.5 Flash — prior gen (same pricing as 3-flash)',
      'gemini-2.0-flash':
        'Gemini 2.0 Flash — older + cheapest (~$0.10/M in, $0.40/M out)',
    },
  }),
  relay_gemini_analysis_model: Value.select({
    name: 'Gemini Analysis Model',
    description:
      "Primary Gemini SKU used when an analysis request is routed to Gemini. On 503/capacity/rate-limit failure, the relay falls back to lower-tier models in order (e.g. 3.1-pro → 3-pro → 3-flash → 2.5-flash).",
    default: 'gemini-3.1-pro-preview',
    values: {
      'gemini-3.1-pro-preview':
        'Gemini 3.1 Pro — best quality on structured-JSON output ($5/M in, $25/M out)',
      'gemini-3-pro-preview':
        'Gemini 3 Pro — prior Pro gen (same pricing as 3.1)',
      'gemini-3-flash-preview':
        'Gemini 3 Flash — faster + ~20× cheaper than Pro; some loss of section-boundary precision on long transcripts',
      'gemini-2.5-flash':
        'Gemini 2.5 Flash — prior Flash gen',
    },
  }),

  // ── Backend routing preference per pipeline ──
  relay_transcribe_backend_preference: Value.select({
    name: 'Transcribe Backend Preference',
    description:
      'Routing strategy for transcription requests. The selected option controls the ORDER in which the relay tries each backend. The Gemini per-tier cap still applies regardless of this setting.',
    default: 'gemini_first',
    values: {
      gemini_first:
        'Gemini first → operator hardware (Parakeet) when cap exceeded',
      hardware_first: 'Operator hardware first → Gemini as fallback',
      gemini_only: 'Gemini only — fail when cap is exceeded',
      hardware_only:
        'Hardware only — fail when no Parakeet endpoint is configured',
    },
  }),
  relay_analyze_backend_preference: Value.select({
    name: 'Analyze Backend Preference',
    description:
      'Routing strategy for analysis requests. Same options as transcription but applies to the analyze step independently — you can route transcribe to hardware and analyze to Gemini, or vice versa.',
    default: 'gemini_first',
    values: {
      gemini_first:
        'Gemini first → operator hardware (Gemma) when cap exceeded',
      hardware_first: 'Operator hardware first → Gemini as fallback',
      gemini_only: 'Gemini only — fail when cap is exceeded',
      hardware_only:
        'Hardware only — fail when no Gemma endpoint is configured',
    },
  }),
})

export const setBackendRouting = sdk.Action.withInput(
  'set-backend-routing',

  async ({ effects }) => ({
    name: 'Set Backend Routing & Models',
    description:
      "Tune which Gemini SKUs the relay uses and the per-pipeline backend pecking order. Live-reloaded — changes take effect on the next request, no restart.",
    warning: null,
    allowedStatuses: 'any',
    group: 'AI Backends',
    visibility: 'enabled',
  }),

  inputSpec,

  async ({ effects }) => {
    const config = await configFile.read().once()
    // Coerce any previously-saved model name to a value in the new
    // select's options. Older 0.2.7-era saved configs could hold a
    // free-text value that's no longer in the dropdown — clamp to a
    // sensible default rather than presenting an invalid radio.
    const TX_OPTIONS = [
      'gemini-3-flash-preview',
      'gemini-2.5-flash',
      'gemini-2.0-flash',
    ] as const
    const AN_OPTIONS = [
      'gemini-3.1-pro-preview',
      'gemini-3-pro-preview',
      'gemini-3-flash-preview',
      'gemini-2.5-flash',
    ] as const
    const tx = config?.relay_gemini_transcription_model as
      | (typeof TX_OPTIONS)[number]
      | undefined
    const an = config?.relay_gemini_analysis_model as
      | (typeof AN_OPTIONS)[number]
      | undefined
    return {
      relay_gemini_transcription_model:
        tx && TX_OPTIONS.includes(tx) ? tx : 'gemini-3-flash-preview',
      relay_gemini_analysis_model:
        an && AN_OPTIONS.includes(an) ? an : 'gemini-3.1-pro-preview',
      relay_transcribe_backend_preference:
        (config?.relay_transcribe_backend_preference as
          | 'gemini_first'
          | 'hardware_first'
          | 'gemini_only'
          | 'hardware_only'
          | undefined) || 'gemini_first',
      relay_analyze_backend_preference:
        (config?.relay_analyze_backend_preference as
          | 'gemini_first'
          | 'hardware_first'
          | 'gemini_only'
          | 'hardware_only'
          | undefined) || 'gemini_first',
    }
  },

  async ({ effects, input }) => {
    await configFile.merge(effects, {
      relay_gemini_transcription_model: (
        input.relay_gemini_transcription_model || 'gemini-3-flash-preview'
      ).trim(),
      relay_gemini_analysis_model: (
        input.relay_gemini_analysis_model || 'gemini-3.1-pro-preview'
      ).trim(),
      relay_transcribe_backend_preference: input.relay_transcribe_backend_preference,
      relay_analyze_backend_preference: input.relay_analyze_backend_preference,
    })
    return null
  },
)