free-coding-models 0.5.83 → 0.5.85

package/src/core/model-family.js

@@ -0,0 +1,176 @@
+/**
+ * @file model-family.js
+ * @description Model family detection + family-preserving failover picker for the Smart Router.
+ *
+ * @details
+ *   📖 When a routed request fails, the router historically fell back to the
+ *   next entry in the set's priority order, which can be a completely different
+ *   model family (user asked for DeepSeek-class output and got a Qwen model).
+ *   This module normalizes every model to a `family` id and provides a
+ *   two-stage failover picker:
+ *     Stage 1: retry the same family on a DIFFERENT provider (same brain,
+ *              different host, e.g. `nvidiaNim/deepseek-v4` → `together/deepseek-v3`).
+ *     Stage 2: the historical set-order fallback (any healthy untried model).
+ *
+ *   📖 Family detection is a keyword scan over a "haystack" built from the
+ *   model id, label, and provider key (lowercased). Mapping ORDER matters:
+ *   more specific brands must be checked before generic substrings they
+ *   contain. `gpt` precedes `o1`/`o3` so `gpt-oss-120b` is classified as GPT,
+ *   and `nemotron` precedes `llama` so `llama-3.1-nemotron-70b` is classified
+ *   as Nemotron rather than Llama.
+ *
+ *   📖 `o1`/`o3` are short numeric keywords that would false-positive inside
+ *   ordinary ids (e.g. `mistral-o1x`), so they only match on word boundaries:
+ *   preceded/followed by a non-alphanumeric character or string edge.
+ *
+ * @functions
+ *   → detectFamily(model) - Normalize a model (string or object) to a family id
+ *   → getModelFamilyHaystack(model) - Lowercased id+label+provider scan string
+ *   → pickNextCandidate(opts) - Two-stage failover picker (family first, then set order)
+ *
+ * @exports BRAND_MAPPINGS, detectFamily, getModelFamilyHaystack, pickNextCandidate
+ *
+ * @see ./router-daemon.js - pickNextCandidate is used by the routeRequest failover loop
+ * @see ../config.js - per-set `familyFailover: true|false` toggle (default true)
+ */
+
+/**
+ * 📖 BRAND_MAPPINGS - ordered keyword → family table. FIRST match wins, so
+ * ordering is part of the contract: more specific keywords must come before
+ * generic ones they contain. Keep new entries sorted by specificity, not
+ * alphabetically.
+ */
+export const BRAND_MAPPINGS = [
+  { keywords: ['claude'], familyId: 'claude', familyName: 'Claude' },
+  { keywords: ['deepseek'], familyId: 'deepseek', familyName: 'DeepSeek' },
+  { keywords: ['gemini'], familyId: 'gemini', familyName: 'Gemini' },
+  // 📖 'gpt' MUST precede the openai-o entry: gpt-oss-120b contains neither
+  // o1 nor o3 on a word boundary, but ordering keeps that guarantee explicit.
+  { keywords: ['gpt', 'gpt-oss'], familyId: 'gpt', familyName: 'GPT' },
+  { keywords: ['nemotron'], familyId: 'nemotron', familyName: 'Nemotron' },
+  { keywords: ['llama'], familyId: 'llama', familyName: 'Llama' },
+  { keywords: ['minimax'], familyId: 'minimax', familyName: 'MiniMax' },
+  { keywords: ['qwen'], familyId: 'qwen', familyName: 'Qwen' },
+  { keywords: ['kimi', 'moonshot'], familyId: 'kimi', familyName: 'Kimi' },
+  { keywords: ['glm', 'chatglm'], familyId: 'glm', familyName: 'GLM' },
+  { keywords: ['mistral', 'mixtral'], familyId: 'mistral', familyName: 'Mistral' },
+  // 📖 o1/o3 are word-boundary matched (see matchesKeyword) to avoid
+  // false positives inside unrelated model ids.
+  { keywords: ['o1', 'o3'], familyId: 'openai-o', familyName: 'OpenAI o' },
+]
+
+/**
+ * 📖 getModelFamilyHaystack - build the lowercase scan string for a model.
+ * Accepts a plain string (model id) or an object with any of id/model/label/
+ * family/provider/providerKey fields, plus a nested `catalog` object as found
+ * on router routing candidates.
+ *
+ * @param {string|object|null} model
+ * @returns {string} lowercase haystack, '' when nothing usable is present
+ */
+export function getModelFamilyHaystack(model) {
+  if (typeof model === 'string') return model.toLowerCase()
+  if (!model || typeof model !== 'object') return ''
+  const catalog = model.catalog && typeof model.catalog === 'object' ? model.catalog : {}
+  return [
+    model.id,
+    model.model,
+    model.label,
+    model.family,
+    model.name,
+    model.provider,
+    model.providerKey,
+    catalog.label,
+    catalog.model,
+  ]
+    .filter((value) => typeof value === 'string' && value.trim())
+    .join(' ')
+    .toLowerCase()
+}
+
+/**
+ * 📖 matchesKeyword - substring match, except for very short (<= 2 chars) or
+ * digit-leading keywords (o1, o3) which require word boundaries so they can
+ * never fire inside an unrelated identifier.
+ *
+ * @param {string} haystack lowercased scan string
+ * @param {string} keyword lowercase brand keyword (alphanumeric + dashes only)
+ * @returns {boolean}
+ */
+function matchesKeyword(haystack, keyword) {
+  if (keyword.length <= 2 || /^\d/.test(keyword)) {
+    const boundary = new RegExp(`(^|[^a-z0-9])${keyword}([^a-z0-9]|$)`)
+    return boundary.test(haystack)
+  }
+  return haystack.includes(keyword)
+}
+
+/**
+ * 📖 detectFamily - map a model to one of the BRAND_MAPPINGS family ids.
+ * Case-insensitive; unknown models return null so callers fall back to their
+ * default behaviour (for the router: plain set-order failover).
+ *
+ * @param {string|object|null} model model id, or object with id/label/provider fields
+ * @returns {string|null} family id (e.g. 'deepseek') or null
+ */
+export function detectFamily(model) {
+  const haystack = getModelFamilyHaystack(model)
+  if (!haystack) return null
+  for (const mapping of BRAND_MAPPINGS) {
+    if (mapping.keywords.some((keyword) => matchesKeyword(haystack, keyword))) {
+      return mapping.familyId
+    }
+  }
+  return null
+}
+
+/**
+ * 📖 pickNextCandidate - the two-stage failover policy, kept pure so it is
+ * unit-testable without the daemon.
+ *
+ * Stage 1 (family preserving): when `familyFailover` is enabled and the failed
+ * candidate's family is detected, pick the first eligible candidate of the SAME
+ * family hosted on a DIFFERENT provider. Candidates arrive in routing order
+ * (priority, circuit state, health - see getRoutingCandidates) so picking the
+ * first match respects the user's ranking inside the family.
+ *
+ * Stage 2 (set order): otherwise fall back to the historical behaviour - the
+ * first eligible candidate in routing order, whatever its family.
+ *
+ * @param {object} opts
+ * @param {Array<object>} opts.candidates routing candidates (ordered, already health-filtered)
+ * @param {object|null} opts.failedCandidate the candidate that just failed
+ * @param {Set<string>} opts.triedKeys candidate keys already attempted
+ * @param {Set<string>} opts.blockedProviders providers skipped for this request (auth failures)
+ * @param {boolean} [opts.familyFailover=true] per-set toggle (config.js normalization)
+ * @returns {{ candidate: object, reason: 'family_failover'|'set_order' }|null}
+ *   null when every candidate is exhausted
+ */
+export function pickNextCandidate({
+  candidates,
+  failedCandidate,
+  triedKeys,
+  blockedProviders,
+  familyFailover = true,
+}) {
+  if (!Array.isArray(candidates) || candidates.length === 0) return null
+  const tried = triedKeys instanceof Set ? triedKeys : new Set(triedKeys || [])
+  const blocked = blockedProviders instanceof Set ? blockedProviders : new Set(blockedProviders || [])
+  const eligible = candidates.filter((candidate) => {
+    if (!candidate || !candidate.key) return false
+    if (tried.has(candidate.key)) return false
+    if (blocked.has(candidate.provider)) return false
+    return true
+  })
+  if (familyFailover !== false && failedCandidate) {
+    const family = detectFamily(failedCandidate)
+    if (family) {
+      const sameFamily = eligible.find(
+        (candidate) => candidate.provider !== failedCandidate.provider && detectFamily(candidate) === family,
+      )
+      if (sameFamily) return { candidate: sameFamily, reason: 'family_failover' }
+    }
+  }
+  const next = eligible[0]
+  return next ? { candidate: next, reason: 'set_order' } : null
+}

package/src/core/models-dev-index.js

@@ -45,6 +45,7 @@ export const PROVIDER_ALIASES = {
   nvidia:     'nvidia',
   openrouter: 'openrouter',
   orcarouter: 'orcarouter',
+  'vercel-gateway': 'vercel',
   groq:       'groq',
   cerebras:   'cerebras',
   github:     'github-models',

package/src/core/opencode.js

@@ -432,6 +432,13 @@ export async function startOpenCode(model, fcmConfig) {
         options: { baseURL: 'https://api.orcarouter.ai/v1', apiKey: '{env:ORCAROUTER_API_KEY}' },
         models: {}
       }
+    } else if (providerKey === 'vercel-gateway') {
+      config.provider['vercel-gateway'] = {
+        npm: '@ai-sdk/openai-compatible',
+        name: 'Vercel AI Gateway',
+        options: { baseURL: 'https://ai-gateway.vercel.sh/v1', apiKey: '{env:VERCEL_AI_GATEWAY_API_KEY}' },
+        models: {}
+      }
     } else if (providerKey === 'huggingface') {
       config.provider.huggingface = {
         npm: '@ai-sdk/openai-compatible',
@@ -807,6 +814,13 @@ export async function startOpenCodeDesktop(model, fcmConfig) {
         options: { baseURL: 'https://api.orcarouter.ai/v1', apiKey: '{env:ORCAROUTER_API_KEY}' },
         models: {}
       }
+    } else if (providerKey === 'vercel-gateway') {
+      config.provider['vercel-gateway'] = {
+        npm: '@ai-sdk/openai-compatible',
+        name: 'Vercel AI Gateway',
+        options: { baseURL: 'https://ai-gateway.vercel.sh/v1', apiKey: '{env:VERCEL_AI_GATEWAY_API_KEY}' },
+        models: {}
+      }
     } else if (providerKey === 'huggingface') {
       config.provider.huggingface = {
         npm: '@ai-sdk/openai-compatible',

package/src/core/provider-key-tester.js

@@ -69,6 +69,7 @@ export const PROVIDER_AUTH_ENDPOINTS = {
   pollinations: null,
   requesty:     null,
   orcarouter:   { url: 'https://api.orcarouter.ai/v1/models',   method: 'GET' },
+  'vercel-gateway': { url: 'https://ai-gateway.vercel.sh/v1/models', method: 'GET' },
   together:     { url: 'https://api.together.xyz/v1/models',        method: 'GET' },
   perplexity:   { url: 'https://api.perplexity.ai/v1/models',       method: 'GET' },
   chutes:       { url: 'https://chutes.ai/v1/models',               method: 'GET' },

package/src/core/provider-metadata.js

@@ -66,6 +66,7 @@ export const ENV_VAR_NAMES = {
   siliconflow: 'SILICONFLOW_API_KEY',
   requesty:   'REQUESTY_API_KEY',
   orcarouter: 'ORCAROUTER_API_KEY',
+  'vercel-gateway': 'VERCEL_AI_GATEWAY_API_KEY',
 }
 
 // 📖 OPENCODE_MODEL_MAP: sparse table of model IDs that differ between sources.js and OpenCode's
@@ -339,6 +340,14 @@ export const PROVIDER_METADATA = {
     rateLimits: 'Free Hacker tier · zero token markup · 3 API keys',
     detailedLimits: 'Zero-markup AI gateway: token prices are passed through at provider rates.\nFree Hacker plan: 3 API keys, adaptive routing + automatic failover + guardrails included.\nOnly the explicit $-0 models are listed in this catalog; the orcarouter/fusion family is pay-as-you-go.',
   },
+  'vercel-gateway': {
+    label: 'Vercel AI Gateway',
+    color: chalk.rgb(140, 170, 255),
+    signupUrl: 'https://vercel.com/ai-gateway',
+    signupHint: 'AI Gateway → API Keys ($5 credits/30 days, no card)',
+    rateLimits: 'Free tier · $5 credits/30 days + $0 models',
+    detailedLimits: 'Official Vercel gateway at list prices, zero markup.\nEvery account gets $5 of credits every 30 days, no card required.\nThe monthly credit covers a subset of the catalog; the explicit $0 models stay free.\nCaution: purchasing credits once permanently moves the account to the paid tier and the monthly free credit stops.',
+  },
   'ollama-cloud': {
     label: 'Ollama Cloud',
     color: chalk.rgb(230, 230, 230),

package/src/core/quota-capabilities.js

@@ -65,6 +65,8 @@ export const PROVIDER_CAPABILITIES = {
   siliconflow: { telemetryType: 'unknown', supportsEndpoint: false, usageDisplay: 'ok', resetCadence: 'daily' },
   requesty: { telemetryType: 'unknown', supportsEndpoint: false, usageDisplay: 'ok', resetCadence: 'daily' },
   orcarouter: { telemetryType: 'unknown', supportsEndpoint: false, usageDisplay: 'ok', resetCadence: 'unknown' },
+  // 📖 Vercel AI Gateway exposes GET /v1/credits but no fetcher is wired yet; treat as monthly reset.
+  'vercel-gateway': { telemetryType: 'unknown', supportsEndpoint: false, usageDisplay: 'ok', resetCadence: 'monthly' },
 }
 
 /** Fallback for unrecognized providers */

package/src/core/router-daemon.js

@@ -54,6 +54,7 @@ import { sendUsageTelemetry } from './telemetry.js'
 import { TIER_ORDER } from './utils.js'
 import { atomicWriteJson, safeJsonParse, sleep, maskApiKey, isRouteableProvider } from './shared-helpers.js'
 import { normalizeRequestBody } from './schema-normalizer.js'
+import { pickNextCandidate } from './model-family.js'
 import {
   loadCache as loadProbeCache,
   flushCache as flushProbeCache,
@@ -1429,6 +1430,14 @@ class RouterRuntime {
   }
 
   addRequestLog(entry) {
+    // 📖 failover_reason (t8) - WHY this attempt was picked ('family_failover'
+    // or 'set_order'). The failover loop stores it on the active request right
+    // after choosing the next candidate; every log entry of that attempt
+    // (requestLog, /stats, SSE 'request' event) carries it from here on.
+    if (entry.failover === true && entry.request_id && entry.failover_reason === undefined) {
+      const active = this.activeRequests.get(entry.request_id)
+      if (active?.failoverReason) entry.failover_reason = active.failoverReason
+    }
     this.requestLog.unshift({ ...entry, at: nowIso() })
     while (this.requestLog.length > MAX_REQUEST_LOG) this.requestLog.pop()
     this.broadcast('request', entry)
@@ -2085,8 +2094,12 @@ class RouterRuntime {
       const tried = []
       const blockedProviders = new Set()
       let attemptIndex = 0
-      for (const candidate of candidates) {
-        if (attemptIndex >= maxAttempts) break
+      // 📖 attemptChain is a private copy of the routing order: on a family
+      // failover (t8) the same-family candidate is swapped in as the NEXT
+      // attempt, so the iteration order itself follows the two-stage policy.
+      const attemptChain = candidates.slice()
+      for (let index = 0; index < attemptChain.length && attemptIndex < maxAttempts; index += 1) {
+        const candidate = attemptChain[index]
         if (blockedProviders.has(candidate.provider)) continue
 
         const activeReq = this.activeRequests.get(requestId)
@@ -2103,8 +2116,35 @@ class RouterRuntime {
         attemptIndex += 1
         if (result.authFailure) blockedProviders.add(candidate.provider)
         if (result.failoverToNext && attemptIndex < maxAttempts) {
-          const next = candidates.find((entry) => !tried.includes(entry.key) && !blockedProviders.has(entry.provider))
-          this.logger.warn(`Failover ${candidate.key}${next ? ` -> ${next.key}` : ''}`, { request_id: requestId, reason: result.reason })
+          // 📖 Two-stage failover (t8): prefer a healthy model of the SAME
+          // family on another provider (DeepSeek down on NIM -> DeepSeek on
+          // Together) so the user's output style doesn't change mid-request.
+          // Falls back to the historical set-order pick, and is disabled
+          // per-set via familyFailover: false.
+          const pick = pickNextCandidate({
+            candidates: attemptChain,
+            failedCandidate: candidate,
+            triedKeys: new Set(tried),
+            blockedProviders,
+            familyFailover: set.familyFailover !== false,
+          })
+          const next = pick?.candidate || null
+          // 📖 Reorder the remaining chain so `next` is genuinely the following
+          // attempt. With set-order picks this is already the case (or the
+          // skipped entries are blocked anyway), so behaviour is unchanged.
+          if (next && attemptChain[index + 1] !== next) {
+            const nextIndex = attemptChain.indexOf(next)
+            if (nextIndex > index) {
+              attemptChain.splice(nextIndex, 1)
+              attemptChain.splice(index + 1, 0, next)
+            }
+          }
+          const activeReqForReason = this.activeRequests.get(requestId)
+          if (next && activeReqForReason) activeReqForReason.failoverReason = pick.reason
+          this.logger.warn(
+            `Failover ${candidate.key}${next ? ` -> ${next.key}` : ''}${pick?.reason === 'family_failover' ? ' [family]' : ''}`,
+            { request_id: requestId, reason: result.reason },
+          )
           void sendUsageTelemetry(this.config, {}, {
             event: 'app_router_failover',
             mode: 'daemon',
@@ -2112,6 +2152,7 @@ class RouterRuntime {
               from_model: candidate.key,
               to_model: next?.key || null,
               reason: result.reason,
+              failover_reason: pick?.reason || null,
               attempt_number: attemptIndex,
             },
           })

package/src/core/router-dashboard.js

@@ -251,6 +251,9 @@ function normalizeRequestEntry(entry) {
     latency_ms: toFiniteNumber(item.latency_ms, null),
     tokens: toFiniteNumber(item.tokens, 0),
     failover: item.failover === true,
+    // 📖 failover_reason (t8) - 'family_failover' | 'set_order'; absent on
+    // older daemons, guarded so the dashboard renders fine either way.
+    failover_reason: safeString(item.failover_reason, null),
     stream: item.stream === true,
     error: safeString(item.error, null),
   }
@@ -1066,7 +1069,9 @@ export function renderRouterDashboard(state, deps = {}) {
       const statusColor = statusText.startsWith('2') ? themeColors.success : statusText === 'ERR' ? themeColors.error : themeColors.warning
       const latency = Number.isFinite(row.latency_ms) ? `${Math.round(row.latency_ms)}ms` : '—'
       const detail = [
-        row.failover ? 'failover' : '',
+        // 📖 t8 - family hops get their own tag so a same-family retry is
+        // visibly different from a plain set-order failover.
+        row.failover ? (row.failover_reason === 'family_failover' ? 'family' : 'failover') : '',
         row.stream ? 'stream' : '',
         row.error || '',
       ].filter(Boolean).join(', ') || '—'