fix(cli): prefer a cheap same-vendor small model for auxiliary tasks

.changeset/byok-cheap-small-model.md

@@ -0,0 +1,5 @@
+---
+"@kilocode/cli": patch
+---
+
+Prefer the session provider's cheapest chat-capable model for auxiliary tasks (session titles, prompt enhance, commit messages, branch names) instead of falling back to Kilo's auto small model. BYOK users no longer silently draw Kilo credits when their configured provider has no recognized small-model family.

packages/opencode/src/kilocode/provider/provider.ts

@@ -14,7 +14,7 @@ import { optionalOmitUndefined } from "@opencode-ai/core/schema"
 import { ProviderError } from "@/provider/error"
 import { Effect, Schema } from "effect"
 import type { LanguageModelV3 } from "@ai-sdk/provider"
-import { mapValues, omit, pickBy } from "remeda"
+import { mapValues, omit, pickBy, sortBy } from "remeda"
 import { reasoningSummary } from "./reasoning-summary"
 import type { Provider } from "@/provider/provider"
 import type { Auth } from "@/auth"
@@ -299,6 +299,87 @@ export function kiloSmallModelPriority(providerID: string): string[] | undefined
   return undefined
 }
 
+/**
+ * Picks the cheapest chat-capable model from the provider's own catalog as the
+ * small model for auxiliary tasks (session titles, prompt enhance, commit
+ * messages, branch names). Keeps those calls on the provider the user
+ * configured instead of falling through to Kilo's auto small model, which
+ * draws Kilo credits in BYOK setups. Chat-capable means text in and text out;
+ * tool calling is not required (e.g. Perplexity sonar and morph chat models
+ * advertise tool_call false). Models with no text output (image, audio, video)
+ * are skipped, and so are embedding and rerank models: they also take text in
+ * and return text scores, but requests to them cannot produce titles or commit
+ * messages. Missing cost data sorts as free; ties break by release date and
+ * id, matching the catalog ordering in getSmallModel.
+ *
+ * Non-generative exclusion is a case-insensitive id/family/api-id heuristic:
+ * besides the obvious embed/rerank substrings it covers common embedding
+ * families (bge, gte, e5, uae, instructor, colbert, voyage,
+ * sentence-similarity) whose ids carry no embed/rerank token (e.g.
+ * intfloat/e5-mistral-7b-instruct, bge-m3, Qwen/Qwen3-Reranker-0.6B), plus
+ * safety/classification models (guard, safeguard, moderation, shield,
+ * llama-guard, shieldgemma) whose text output is a label rather than
+ * generated prose. The api id is checked alongside the alias/family so an
+ * alias like `vectors` pointing at `nomic-embed-text` cannot bypass the
+ * exclusion.
+ */
+const NON_GENERATIVE_RE =
+  /embed|rerank|sentence[-_ ]similarity|text-embedding|bge|gte-?(?:large|base|small)?|e5-|-e5\b|\buae\b|instructor|colbert|voyage(?![a-z])|guard|safeguard|moderation|shield|llama[-_]?guard|shieldgemma|toxic|harm|pii|nsfw/i
+
+export function cheapestSmallModel(models: Provider.Model[]) {
+  return sortBy(
+    models.filter((model) => {
+      if (!model.capabilities.input.text || !model.capabilities.output.text) return false
+      return !NON_GENERATIVE_RE.test(`${model.family} ${model.id} ${model.api?.id ?? ""}`)
+    }),
+    [(model) => model.cost.input + model.cost.output, "asc"],
+    [(model) => model.release_date, "desc"],
+    [(model) => model.id, "desc"],
+  ).at(0)
+}
+
+/**
+ * Provider ids that always front a local model server (LM Studio, Ollama,
+ * Atomic Chat, Anaconda Desktop). Their catalogs list every known model, but
+ * the server only serves installed ones — regardless of which hostname the
+ * user points them at (localhost, a LAN ip, host.docker.internal, ...).
+ */
+const LOCAL_SERVER_IDS = new Set(["lmstudio", "ollama", "atomic-chat", "anaconda-desktop"])
+
+/**
+ * True when the provider routes to a local server (LM Studio, Ollama, ...).
+ * Local catalogs list every known model, but the server only serves installed
+ * ones — the cheapest catalog entry can name a model the user never installed
+ * (e.g. openai/gpt-oss-20b while the session uses Qwen). Callers fall back to
+ * the session's own model when getSmallModel returns undefined, so the cheap
+ * pick is skipped for local providers. A hostname check alone is not enough:
+ * a local server remains local when reached over the LAN
+ * (192.168.x.x, host.docker.internal, ollama.local), so known local-server
+ * ids always take the safe fallback. Only an unrelated remote endpoint keeps
+ * the cheap pick.
+ */
+export function isLoopbackProvider(provider: {
+  id?: unknown
+  options?: Record<string, unknown>
+  models: Record<string, Provider.Model>
+}) {
+  if (typeof provider.id === "string" && LOCAL_SERVER_IDS.has(provider.id)) return true
+  const raw = provider.options?.baseURL ?? provider.options?.baseUrl
+  if (typeof raw === "string" && raw !== "") return isLoopbackURL(raw)
+  const urls = Object.values(provider.models).map((model) => model.api?.url).filter(Boolean)
+  if (urls.length === 0) return false
+  return urls.every(isLoopbackURL)
+}
+
+function isLoopbackURL(raw: string) {
+  try {
+    const host = new URL(raw).hostname.toLowerCase()
+    return host === "localhost" || host === "127.0.0.1" || host === "::1" || host === "[::1]"
+  } catch {
+    return false
+  }
+}
+
 /**
  * True when the user has kilo credentials: a KILO_API_KEY env var, a stored
  * auth entry, or an apiKey in the kilo provider config. Mirrors the hasKey

packages/opencode/src/provider/provider.ts

@@ -45,6 +45,8 @@ import {
   publicKiloProvider,
   kiloSmallModelPriority,
   hasKiloCredentials,
+  cheapestSmallModel,
+  isLoopbackProvider,
   buildTimeoutSignal,
   requestTimeout,
   wrapFirstByte,
@@ -2141,6 +2143,14 @@ const layer = Layer.effect(
         if (candidates[0]) return candidates[0]
       }
 
+      // kilocode_change start - skip local-server providers: their catalog lists
+      // every known model but the server only serves installed ones, so the
+      // cheapest entry can name a model the user never installed. Returning
+      // undefined keeps the session-model fallback in the callers.
+      const cheap = isLoopbackProvider(provider) ? undefined : cheapestSmallModel(Object.values(provider.models))
+      if (cheap) return cheap
+      // kilocode_change end
+
       // kilocode_change start - fall back to kilo's auto small model only when the user actually has
       // kilo credentials. The kilo provider is always autoloaded (anonymous key), so checking it
       // unconditionally would route auxiliary tasks (session titles, commit messages, branch names)