pi-requesty-discover 2.1.1 → 3.0.1

package/README.adoc

@@ -8,17 +8,19 @@ endif::[]
 
 = pi-requesty-discover
 
-image:https://img.shields.io/npm/v/pi-requesty-discover.svg[npm version, link=https://www.npmjs.com/package/pi-requesty-discover]
+image:https://img.shields.io/npm/v/pi-requesty-discover.svg[npm version,link=https://www.npmjs.com/package/pi-requesty-discover]
 
 This extension helps you discover currently available models in requesty.ai for your Pi Coding Agent.
 
 This is a fork of https://github.com/requestyai/pi-requesty[pi-requesty] with additional features:
 
-* Use `REQUESTY_API_KEY` from the environment
+* Use the `apiKey` from the model registry allowing you to use
+https://pi.dev/docs/latest/models#value-resolution[pi's full value resolution] capabilities
 * Health-check available models and write a log about them
 * Brief information about what was found, including added/removed models, health summary
 * Confirm before writing `models.json`, with a dry-run mode that only previews
 * A footer status line showing your monthly Requesty spend against your monthly limit
+* Optional Settings file in *json5* format
 
 == When and why use it
 
@@ -59,7 +61,7 @@ pi -e ./pi-requesty-discover
 == Configuration
 
 The extension reads the `requesty-export` provider from `~/.pi/agent/models.json`.
-You can override the provider id with the `REQUESTY_PROVIDER_ID` variable.
+You can override the provider id via the settings.
 
 Example:
 
@@ -70,46 +72,47 @@ Example:
     "requesty-export": {
       "name": "Requesty",
       "baseUrl": "https://router.requesty.ai/v1",
-      "apiKey": "$REQUESTY_API_KEY",
       "api": "openai-completions",
+      "apiKey": "$REQUESTY_API_KEY",
       "models": []
     }
   }
 }
 ----
 
-IMPORTANT: The `apiKey` from `models.json` is not used by this extension.
-It may contain Pi-specific environment-variable substitution that the extension cannot reliably resolve.
-Set `REQUESTY_API_KEY` instead. This variable is required.
-You can configure the real `apiKey` as you like, but it would be recommended to use the same variable as for the model discovery.
+IMPORTANT: Health checks are designed to use only a few tokens, but they still make real Requesty API calls and **will** consume credits.
 
-IMPORTANT: Health checks are designed to use only a few tokens, but they still make real Requesty API calls and may consume credits.
+On startup, the extension fetches `<baseUrl>/models` using the configured `apiKey` as the bearer token and registers discovered and healthy models with Pi.
 
-On startup, the extension fetches `<baseUrl>/models` using `REQUESTY_API_KEY` as the bearer token and registers discovered and healthy models with Pi.
+=== Discovery settings
 
-=== Environment variables
+The discovery workflow reads optional settings from `requesty-discovery-settings.json5` in the Pi config directory (`PI_CODING_AGENT_DIR`).
+The file uses JSON5, so it supports comments. It is optional; missing settings default to empty.
 
-[cols="1,1,1,1,2", options="header"]
-|===
-| Name | Required | Type | Default | Meaning
+Example:
 
-| `REQUESTY_API_KEY`
-| yes
-| `string`
-|
-| Requesty API key used for model discovery and health checks.
+[source,json5]
+----
+{
+  // override the default provider id
+  providerId: "requesty-export",
+  // configure the health check mode
+  healthCheckMode: "full", // or "basic", "off"
+  // models excluded from discovery; useful for models that keep failing health checks
+  bannedModels: [
+    // hangs on tool calls
+    'requesty/unstable-model',
+  ],
+}
+----
 
-| `REQUESTY_PROVIDER_ID`
-| no
-| `string`
-| `requesty-export`
-| The id of the provider.
+Banned models are excluded before health checks run (so they cost no credits) and are removed from `models.json` on the next discovery write.
 
-| `REQUESTY_HEALTH_CHECK_MODE`
-| no
-| `off`, `basic`, `full`
-| `full`
-| The health-check mode.
+=== Environment variables
+
+[cols="1,1,1,1,2",options="header"]
+|===
+| Name | Required | Type | Default | Meaning
 
 | `PI_CODING_AGENT_DIR`
 | no
@@ -132,12 +135,11 @@ The command:
 - discovers Requesty models (and optional health checks)
 - shows a summary notification (added/removed IDs, health summary)
 - asks for confirmation before writing `~/.pi/agent/models.json`
+- refreshes Pi's model registry so the new models are immediately available in `/model`
 - reports whether the file was updated
 
 If there are no model ID changes, confirmation still offers a metadata refresh write, which might have changed as well.
 
-After a successful write, run `/reload`.
-
 [source,text]
 ----
 /requesty-discover --dry-run
@@ -145,22 +147,31 @@ After a successful write, run `/reload`.
 
 Same discovery and health checks, but never prompts to write and leaves `models.json` unchanged.
 
+[source,bash]
+----
+pi -p "/requesty-discover"
+pi -p "/requesty-discover --dry-run"
+----
+
+You can also run the command from the CLI in a non-interactive mode.
+
 == Usage status
 
 The extension shows your current Requesty monthly spend as a persistent line in Pi's footer, e.g.:
 
 [source,text]
 ----
-Requesty Usage (<API Key Name>): $63.55/$150.00 (42%)
+<API Key Name>: $63.55/$150.00 (42%)
 ----
 
 When your API key has no monthly limit (`monthly_limit` is `0`), the limit is treated as unlimited:
 
 [source,text]
 ----
-Requesty Usage (<API Key Name>): $63.55 (unlimited)
+<API Key Name>: $63.55 (unlimited)
 ----
 
-When you select a model from a provider other than `REQUESTY_PROVIDER_ID`, the status is cleared — there is no Requesty usage to show for it. If `REQUESTY_API_KEY` is not configured, the status is left untouched.
+When you select a model from a provider other than `REQUESTY_PROVIDER_ID`, the status is cleared — there is no Requesty usage to show for it.
 
-NOTE: The footer update is best-effort and never blocks a turn. If it fails (network error, stale context after `/reload`), it fails silently without spamming notifications.
+NOTE: The footer update is best-effort and never blocks a turn.
+If it fails (network error, stale context after `/reload`), it fails silently without spamming notifications.

package/README.md

@@ -6,11 +6,13 @@ This extension helps you discover currently available models in requesty.ai for
 
 This is a fork of [pi-requesty](https://github.com/requestyai/pi-requesty) with additional features:
 
-* Use `REQUESTY_API_KEY` from the environment
+* Use the `apiKey` from the model registry allowing you to use
+[pi’s full value resolution](https://pi.dev/docs/latest/models#value-resolution) capabilities
 * Health-check available models and write a log about them
 * Brief information about what was found, including added/removed models, health summary
 * Confirm before writing `models.json`, with a dry-run mode that only previews
 * A footer status line showing your monthly Requesty spend against your monthly limit
+* Optional Settings file in **json5** format
 
 ## When and why use it
 
@@ -48,7 +50,7 @@ pi -e ./pi-requesty-discover
 ## Configuration
 
 The extension reads the `requesty-export` provider from `~/.pi/agent/models.json`.
-You can override the provider id with the `REQUESTY_PROVIDER_ID` variable.
+You can override the provider id via the settings.
 
 Example:
 
@@ -58,8 +60,8 @@ Example:
     "requesty-export": {
       "name": "Requesty",
       "baseUrl": "https://router.requesty.ai/v1",
-      "apiKey": "$REQUESTY_API_KEY",
       "api": "openai-completions",
+      "apiKey": "$REQUESTY_API_KEY",
       "models": []
     }
   }
@@ -67,23 +69,37 @@ Example:
 ```
 
 **❗ IMPORTANT**\
-The `apiKey` from `models.json` is not used by this extension.
-It may contain Pi-specific environment-variable substitution that the extension cannot reliably resolve.
-Set `REQUESTY_API_KEY` instead. This variable is required.
-You can configure the real `apiKey` as you like, but it would be recommended to use the same variable as for the model discovery.
+Health checks are designed to use only a few tokens, but they still make real Requesty API calls and ***will*** consume credits.
 
-**❗ IMPORTANT**\
-Health checks are designed to use only a few tokens, but they still make real Requesty API calls and may consume credits.
+On startup, the extension fetches `<baseUrl>/models` using the configured `apiKey` as the bearer token and registers discovered and healthy models with Pi.
+
+### Discovery settings
+
+The discovery workflow reads optional settings from `requesty-discovery-settings.json5` in the Pi config directory (`PI_CODING_AGENT_DIR`).
+The file uses JSON5, so it supports comments. It is optional; missing settings default to empty.
+
+Example:
+
+```json5
+{
+  // override the default provider id
+  providerId: "requesty-export",
+  // configure the health check mode
+  healthCheckMode: "full", // or "basic", "off"
+  // models excluded from discovery; useful for models that keep failing health checks
+  bannedModels: [
+    // hangs on tool calls
+    'requesty/unstable-model',
+  ],
+}
+```
 
-On startup, the extension fetches `<baseUrl>/models` using `REQUESTY_API_KEY` as the bearer token and registers discovered and healthy models with Pi.
+Banned models are excluded before health checks run (so they cost no credits) and are removed from `models.json` on the next discovery write.
 
 ### Environment variables
 
 | Name | Required | Type | Default | Meaning |
 | --- | --- | --- | --- | --- |
-| `REQUESTY_API_KEY` | yes | `string` |  | Requesty API key used for model discovery and health checks. |
-| `REQUESTY_PROVIDER_ID` | no | `string` | `requesty-export` | The id of the provider. |
-| `REQUESTY_HEALTH_CHECK_MODE` | no | `off`, `basic`, `full` | `full` | The health-check mode. |
 | `PI_CODING_AGENT_DIR` | no | `string` | `~/.pi/agent` | pi configuration directory |
 
 ## Command
@@ -99,33 +115,40 @@ The command:
 * discovers Requesty models (and optional health checks)
 * shows a summary notification (added/removed IDs, health summary)
 * asks for confirmation before writing `~/.pi/agent/models.json`
+* refreshes Pi’s model registry so the new models are immediately available in `/model`
 * reports whether the file was updated
 
 If there are no model ID changes, confirmation still offers a metadata refresh write, which might have changed as well.
 
-After a successful write, run `/reload`.
-
 ```text
 /requesty-discover --dry-run
 ```
 
 Same discovery and health checks, but never prompts to write and leaves `models.json` unchanged.
 
+```bash
+pi -p "/requesty-discover"
+pi -p "/requesty-discover --dry-run"
+```
+
+You can also run the command from the CLI in a non-interactive mode.
+
 ## Usage status
 
 The extension shows your current Requesty monthly spend as a persistent line in Pi’s footer, e.g.:
 
 ```text
-Requesty Usage (<API Key Name>): $63.55/$150.00 (42%)
+<API Key Name>: $63.55/$150.00 (42%)
 ```
 
 When your API key has no monthly limit (`monthly_limit` is `0`), the limit is treated as unlimited:
 
 ```text
-Requesty Usage (<API Key Name>): $63.55 (unlimited)
+<API Key Name>: $63.55 (unlimited)
 ```
 
-When you select a model from a provider other than `REQUESTY_PROVIDER_ID`, the status is cleared — there is no Requesty usage to show for it. If `REQUESTY_API_KEY` is not configured, the status is left untouched.
+When you select a model from a provider other than `REQUESTY_PROVIDER_ID`, the status is cleared — there is no Requesty usage to show for it.
 
 **📌 NOTE**\
-The footer update is best-effort and never blocks a turn. If it fails (network error, stale context after `/reload`), it fails silently without spamming notifications.
+The footer update is best-effort and never blocks a turn.
+If it fails (network error, stale context after `/reload`), it fails silently without spamming notifications.

package/docs/DECISIONS.adoc

@@ -4,49 +4,43 @@ Short notes for non-obvious decisions: context, decision, why not the obvious al
 
 == Usage status stays on `setStatus` (shared, truncated footer line)
 
-Pi's `FooterComponent` joins all extensions' `setStatus` entries on one line and
-ellipsis-truncates the joined line to terminal width. No per-key layout, no wrap.
+Pi's `FooterComponent` joins all extensions' `setStatus` entries on one line and ellipsis-truncates the joined line to terminal width.
+No per-key layout, no wrap.
 
-`ctx.ui.setFooter(factory)` is replacement-only — it swaps the whole footer (pwd,
-tokens, context %, model, statuses), and the factory gets `footerData` but not
-token/model state. No "extra footer" API exists.
+`ctx.ui.setFooter(factory)` is replacement-only — it swaps the whole footer (pwd, tokens, context %, model, statuses), and the factory gets `footerData` but not token/model state.
+No "extra footer" API exists.
 
-Decision: keep `setStatus`. Revisit when the line overflows visibly with other
-extensions' statuses, or if the usage string grows past a typical width.
+Decision: keep `setStatus`.
+Revisit when the line overflows visibly with other extensions' statuses, or if the usage string grows past a typical width.
 
 == Hand-rolled `Try<T>` instead of an effects library
 
-Context: env loading and discovery evaluation both need a caught-error result value
-(`{ ok: true, value } | { ok: false, error }`). Libraries like Effect ship an
-industrial `Either`/`Effect.try` for exactly this.
+Context: env loading and discovery evaluation both need a caught-error result value (`{ ok: true, value } | { ok: false, error }`).
+Libraries like Effect ship an industrial `Either`/`Effect.try` for exactly this.
 
 Why not: Effect is a runtime, a generator DSL, and a worldview — not a utility.
-Two call sites in a Pi extension don't justify the dependency weight or the
-learning curve for contributors. The hand-rolled version is ~10 lines
-(`Try<T>`, `runCatching`, `runCatchingAsync`, named after Kotlin's `Result`),
-lives in `src/index.ts`, and documents intent just as well.
+Two call sites in a Pi extension don't justify the dependency weight or the learning curve for contributors.
+The hand-rolled version is ~10 lines (`Try<T>`, `runCatching`, `runCatchingAsync`, named after Kotlin's `Result`), lives in `utils.ts`, and documents intent just as well.
 
-Decision: keep the hand-rolled `Try<T>`. No `monads.ts` module either — rule of
-three. Revisit if the extension ever needs typed errors, retries, or resource
-scoping; that's the day Effect earns its place.
+Decision: keep the hand-rolled `Try<T>`.
+Revisit if the extension ever needs typed errors, retries, or resource scoping; that's the day Effect earns its place.
 
 == Not a provider — a models.json-writing extension
 
-Context: Requesty already works as a Pi provider via `models.json`. The temptation
-is to make this extension *be* the provider: read the model list from the Requesty
-API, hold it in memory, surface usage stats, let Pi manage the API key.
+Context: Requesty already works as a Pi provider via `models.json`.
+The temptation is to make this extension *be* the provider: read the model list from the Requesty API, hold it in memory, surface usage stats, let Pi manage the API key.
 
 Why not:
 
-- Health-checking is the core value and has no home in a provider. A provider serves
-  models; it doesn't vet them. The health check plus the diff/confirm/write flow is a
-  discovery-and-selection step, not serving.
-- A provider keeps its model list in-memory and refreshes occasionally. Here the
-  list is a persistent on-disk snapshot (`models.json`) — constant until you re-run
-  `/requesty-discover`. That matches the intent: an explicit, reviewed, user-confirmed write,
-  after which Pi's normal provider machinery serves the models.
+- Health-checking is the core value and has no home in a provider.
+A provider serves models; it doesn't vet them.
+The health check plus the diff/confirm/write flow is a discovery-and-selection step, not serving.
+- A provider keeps its model list in-memory and refreshes occasionally.
+Here the list is a persistent on-disk snapshot (`models.json`) — constant until you re-run
+`/requesty-discover`.
+That matches the intent: an explicit, reviewed, user-confirmed write, after which Pi's normal provider machinery serves the models.
 
-The provider-like parts (API model fetch, usage footer, wanting Pi-managed auth) are
-real, but the health check and the snapshot model break the provider fit.
+The provider-like parts (API model fetch, usage footer, wanting Pi-managed auth) are real, but the health check and the snapshot model break the provider fit.
 
 Decision: stay a `/requesty-discover` command that writes `models.json`.
+

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-requesty-discover",
-  "version": "2.1.1",
+  "version": "3.0.1",
   "description": "Discover available Requesty models for the Pi Coding Agent",
   "repository": {
     "type": "git",
@@ -47,7 +47,8 @@
     ]
   },
   "devDependencies": {
-    "@earendil-works/pi-tui": "^0.80.3",
+    "@earendil-works/pi-tui": "^0.85.0",
+    "@earendil-works/pi-coding-agent": "^0.85.0",
     "@eslint/js": "^10.0.1",
     "@types/node": "^26.1.0",
     "@vitest/coverage-v8": "^4.1.9",
@@ -63,6 +64,7 @@
     "vitest": "^4.1.9"
   },
   "dependencies": {
+    "json5": "^2.2.3",
     "zod": "^4.4.3"
   }
 }

package/src/discovery.ts

@@ -0,0 +1,222 @@
+import type { ProviderModelConfig } from '@earendil-works/pi-coding-agent'
+import { type Env } from './env'
+import {
+  type GetApiKey,
+  type ModelsDiff,
+  type ModelsJson,
+  diffModels,
+  formatModelsDiffSummary,
+  getRequestyConfig,
+  updateModelsJson,
+} from './models-json'
+import { checkModels, formatHealthSummary, writeHealthCheckLog } from './health-check'
+import { discoverModels } from './requesty-api'
+import type { DiscoverySettings } from './settings'
+import { formatErrorMessage, runCatchingAsync } from './utils'
+
+const DRY_RUN_ARG = '--dry-run'
+
+interface AutocompleteItem {
+  value: string
+  label: string
+  description: string
+}
+
+export type NotificationLevel = 'info' | 'warning' | 'error'
+
+/** Everything the discovery workflow needs from the UI. One port, two adapters (tui/console) in index.ts. */
+export type DiscoveryUi = {
+  notify(message: string, level: NotificationLevel): void
+  confirm(title: string, message: string): Promise<boolean>
+  setStatus(message: string): void
+}
+
+export type RefreshModelsRegistry = () => Promise<void>
+
+export type DiscoveryEvaluation = {
+  dryRun: boolean
+  modelCount: number
+  failedCount: number
+  warningCount: number
+  passing: ProviderModelConfig[]
+  diff: ModelsDiff
+  healthCheckSummary: string
+  logNote: string
+  data: ModelsJson
+}
+
+export function formatDiscoveryFailure(error: unknown): string {
+  const detail = formatErrorMessage(error)
+  return `Discovery failed: ${detail}`
+}
+
+export async function evaluateDiscovery(
+  args: string,
+  settings: DiscoverySettings,
+  env: Env,
+  ui: DiscoveryUi,
+  getApiKey: GetApiKey,
+): Promise<DiscoveryEvaluation> {
+  ui.setStatus('Discovering Requesty models...')
+  const dryRun = args.split(' ').includes(DRY_RUN_ARG)
+
+  const { data, provider, existingModelIds } = await getRequestyConfig(getApiKey, settings, env)
+  const bannedModels = new Set(settings.bannedModels)
+  const allModels = await discoverModels(provider)
+  const foundBannedModels = allModels
+    .filter(model => bannedModels.has(model.id))
+    .map(model => model.id)
+    .toSorted()
+  const models = allModels.filter(model => !bannedModels.has(model.id))
+  const modelsMap = new Map(models.map(m => [m.id, m]))
+
+  let diff: ModelsDiff
+  let failedCount = 0
+  let warningCount = 0
+  let passing: ProviderModelConfig[] = []
+  let logNote = ''
+  let healthCheckSummary = ''
+
+  if (settings.healthCheckMode !== 'off') {
+    if (models.length > 0) {
+      ui.setStatus(`Checking models 0/${models.length}...`)
+      const healthResults = await checkModels(provider, models, settings.healthCheckMode === 'full', {
+        onProgress: ({ completed, total }) => {
+          ui.setStatus(`Checking models ${completed}/${total}...`)
+        },
+      })
+      const sortedResults = healthResults.toSorted((a, b) => a.modelId.localeCompare(b.modelId))
+      failedCount = sortedResults.filter(r => r.status === 'error').length
+      warningCount = sortedResults.filter(r => r.status === 'warning').length
+      passing = sortedResults.flatMap(r => {
+        const model = modelsMap.get(r.modelId)
+        if (!model) return []
+        if (r.status === 'ok') return [model]
+        // warnings are transient: assume that the error will go away and add them anyway
+        if (r.status === 'warning') return [model]
+        return []
+      })
+      healthCheckSummary = formatHealthSummary(sortedResults)
+      writeHealthCheckLog(
+        provider,
+        sortedResults,
+        diffModels(existingModelIds, passing),
+        { providerId: settings.providerId, bannedModels: foundBannedModels },
+        env,
+      )
+    }
+    diff = diffModels(existingModelIds, passing)
+    logNote = `Full health check log: ${env.health_check_log_path}\n`
+  } else {
+    passing = models
+    diff = diffModels(existingModelIds, passing)
+  }
+
+  return {
+    dryRun,
+    modelCount: models.length,
+    failedCount,
+    warningCount,
+    passing,
+    diff,
+    healthCheckSummary,
+    logNote,
+    data,
+  }
+}
+
+export async function finalizeDiscovery(
+  evaluation: DiscoveryEvaluation,
+  settings: DiscoverySettings,
+  env: Env,
+  ui: DiscoveryUi,
+  refresh?: RefreshModelsRegistry,
+): Promise<void> {
+  const level = notificationLevel(evaluation)
+  const summary = buildDiscoverySummary(evaluation)
+
+  // Always surface the discovery result first (toast styling), then decide.
+  if (evaluation.dryRun) {
+    ui.notify(
+      `${summary}
+Dry run: left models.json unchanged.`,
+      level,
+    )
+    return
+  }
+
+  if (evaluation.passing.length === 0) {
+    ui.notify(
+      `${summary}
+Left models.json unchanged.`,
+      level,
+    )
+    return
+  }
+
+  ui.notify(summary, level)
+  const { title, message } = buildConfirmPrompt(evaluation)
+  const shouldUpdate = await ui.confirm(title, message)
+  if (shouldUpdate) {
+    updateModelsJson(evaluation.data, evaluation.passing, settings, env)
+    if (!refresh) {
+      ui.notify('Updated models.json.', 'info')
+      return
+    }
+    const refreshResult = await runCatchingAsync(() => refresh())
+    if (refreshResult.ok) {
+      ui.notify('Updated models.json. New models are available in /model.', 'info')
+    } else {
+      ui.notify(
+        `Updated models.json, but the model registry could not be refreshed: ${formatErrorMessage(refreshResult.error)}. Run /reload or restart Pi to use the changes.`,
+        'warning',
+      )
+    }
+    return
+  }
+
+  ui.notify('Left models.json unchanged.', 'info')
+}
+
+function buildDiscoverySummary(evaluation: DiscoveryEvaluation): string {
+  return [
+    `Discovered ${evaluation.modelCount} Requesty model(s).`,
+    evaluation.healthCheckSummary.trimEnd(),
+    formatModelsDiffSummary(evaluation.diff),
+    evaluation.logNote.trimEnd(),
+  ]
+    .filter(part => part.length > 0)
+    .join('\n')
+}
+
+function buildConfirmPrompt(evaluation: DiscoveryEvaluation): { title: string; message: string } {
+  const hasIdChanges = evaluation.diff.added.length > 0 || evaluation.diff.removed.length > 0
+  if (hasIdChanges) {
+    return {
+      title: `Write ${evaluation.passing.length} model(s)?`,
+      message: 'Update models.json with the discovery result above.',
+    }
+  }
+  return {
+    title: 'Refresh models.json?',
+    message: 'No model ID changes. Rewrite the file to refresh metadata anyway?',
+  }
+}
+
+function notificationLevel(evaluation: DiscoveryEvaluation): NotificationLevel {
+  if (evaluation.failedCount > 0) return 'error'
+  if (evaluation.warningCount > 0) return 'warning'
+  return 'info'
+}
+
+export function getArgumentCompletions(prefix: string): AutocompleteItem[] {
+  const options = [
+    {
+      value: DRY_RUN_ARG,
+      label: DRY_RUN_ARG,
+      description: 'Preview discovery without offering to write models.json',
+    },
+  ]
+  if (!prefix) return options
+  return options.filter(o => o.value.toLowerCase().startsWith(prefix.toLowerCase()))
+}

package/src/env.ts

@@ -1,37 +1,17 @@
 import path from 'node:path'
-import process from 'node:process'
-import { prettifyError, z } from 'zod'
 import { getAgentDir } from '@earendil-works/pi-coding-agent'
 
-const HealthCheckModeSchema = z.enum(['off', 'basic', 'full']).default('full')
-
-export const DEFAULT_PROVIDER_ID = 'requesty-export'
-
 export type Env = {
   models_json_path: string
   health_check_log_path: string
-  provider_id: string
-  requesty_api_key: string
-  health_check_mode: z.infer<typeof HealthCheckModeSchema>
+  settings_path: string
 }
 
 export function getEnv(): Env {
-  const envVars = process.env
-  const result = HealthCheckModeSchema.safeParse(envVars.REQUESTY_HEALTH_CHECK_MODE)
-  if (!result.success) {
-    throw new Error(prettifyError(result.error))
-  }
-
-  const apiKey = envVars.REQUESTY_API_KEY
-  if (!apiKey) {
-    throw new Error(`apiKey must be set via REQUESTY_API_KEY env var`)
-  }
   const agentPath = getAgentDir()
   return {
     models_json_path: path.join(agentPath, 'models.json'),
     health_check_log_path: path.join(agentPath, 'requesty-health-check.log'),
-    provider_id: envVars.REQUESTY_PROVIDER_ID ?? DEFAULT_PROVIDER_ID,
-    requesty_api_key: apiKey,
-    health_check_mode: result.data,
+    settings_path: path.join(agentPath, 'requesty-discovery-settings.json5'),
   }
 }