cartethyia 2.5.3 → 2.5.6

package/src/providers/discovery/openai-model-discovery.ts

@@ -49,6 +49,11 @@ export async function fetchOpenAICompatibleModels(
       const parsedContext =
         positiveNumber(entry.context_length) ??
         positiveNumber((entry as Record<string, unknown>).input_token_limit) ??
+        // `context_window` is the spelling an OpenAI-shaped catalog uses for
+        // its own field. Unread, every model on such an upstream fell to the
+        // 200_000 default below and was advertised at the same wrong window no
+        // matter what the upstream reported — a 1M model shown as 200k.
+        positiveNumber((entry as Record<string, unknown>).context_window) ??
         positiveNumber(entry.contextWindow) ??
         positiveNumber(entry.max_input_tokens) ??
         modalityContextLimit(entry);

package/src/providers/discovery/openrouter-free-roster.ts

@@ -1,11 +1,12 @@
 /**
  * The free tier shared by the OpenRouter-backed providers.
  *
- * Cline, Kilo Code and OpenRouter serve one model catalog: the same 469 ids,
- * and the same 9 of them carrying a `:free` suffix. Verified against all three
- * live listings — every id below is present under `openrouter`, `kilo` and
- * `cline` — so this roster is a catalog fact, not a per-provider guess, and it
- * lives in one place rather than being restated three times.
+ * Cline, Kilo Code and OpenRouter serve one model catalog, and every id below
+ * was verified to exist under `openrouter`, `kilo` and `cline` — so this
+ * roster is a catalog fact, not a per-provider guess, and it lives in one
+ * place rather than being restated three times. Dropping a row here unlists
+ * it from all three providers' seeded catalog at the next boot; live
+ * discovery will still find it, which is the intended prune.
  *
  * The limits, modalities and capability flags are read from OpenRouter's own
  * `/api/v1/models`, which states `context_length`,
@@ -13,6 +14,10 @@
  * `supported_parameters`. That is the serving catalog's declaration, so these
  * are measurements rather than defaults.
  *
+ * The set is curated: a 2.6B toy (`liquid/lfm-2.5-2.6b:free`) and the
+ * upstream content-safety classifier (`nvidia/nemotron-3.5-content-safety:free`)
+ * are deliberately not seeded — a prefill classifier is not a chat model.
+ *
  * This is a floor, not the authority: each provider also discovers live, and a
  * row the listing carries that this roster does not is still found there. The
  * roster exists so a fresh install has a usable free tier before any fetch, and
@@ -32,15 +37,16 @@ interface FreeRow {
   readonly ctx: number;
   readonly out: number;
   readonly vision?: boolean;
-  /** `supported_parameters` omits `tools` on the safety classifier only. */
+  /** `false` only when `supported_parameters` omits `tools`. */
   readonly toolCall?: boolean;
 }
 
 /**
- * The 9 `:free` ids, with the limits the catalog states for each.
+ * The shared `:free` ids still worth seeding, with the limits the catalog
+ * states for each.
  *
- * Every row declares `reasoning`: `supported_parameters` lists it on all
- * nine. `toolCall` is `true` unless stated otherwise, because a row that
+ * Every row declares `reasoning`: `supported_parameters` lists it on every
+ * entry. `toolCall` is `true` unless stated otherwise, because a row that
  * advertises tools must keep them — a false `toolCall` makes the preflight
  * strip `tools` from the request, which is a wrong capability rather than a
  * conservative one.
@@ -48,7 +54,6 @@ interface FreeRow {
 const FREE_ROWS: readonly FreeRow[] = [
   { id: "cohere/north-mini-code:free", ctx: 256_000, out: 64_000 },
   { id: "dots-studio/dots-3-note-preview:free", ctx: 512_000, out: 460_800, vision: true },
-  { id: "liquid/lfm-2.5-2.6b:free", ctx: 65_536, out: 8_192 },
   { id: "minimax/minimax-m3.1-flash:free", ctx: 1_000_000, out: 524_288, vision: true },
   {
     id: "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free",
@@ -57,13 +62,6 @@ const FREE_ROWS: readonly FreeRow[] = [
     vision: true,
   },
   { id: "nvidia/nemotron-3-ultra-550b-a55b:free", ctx: 1_000_000, out: 65_536 },
-  {
-    id: "nvidia/nemotron-3.5-content-safety:free",
-    ctx: 128_000,
-    out: 8_192,
-    vision: true,
-    toolCall: false,
-  },
   { id: "nvidia/nemotron-3.5-lightning:free", ctx: 1_000_000, out: 65_536 },
   { id: "nvidia/nemotron-3.5-super-vl:free", ctx: 1_000_000, out: 131_072, vision: true },
 ];

package/src/providers/discovery/probe-phases.ts

@@ -315,10 +315,10 @@ export async function resolveProbeTarget(args: {
       endpointPath = endpointForFamily(wireFamily);
     }
   }
-  // Wire family and source surface were the same vocabulary until `gemini`
-  // landed, so this used to be a plain assignment. They are not the same thing:
-  // `gemini` names a wire (native generateContent) but has no canonical request
-  // surface of its own, while `completion` is a surface with no matching wire.
+  // Wire family and source surface are not the same vocabulary, so this cannot
+  // be a plain assignment: `gemini` names a wire (native generateContent) but
+  // has no canonical request surface of its own, while `completion` is a
+  // surface with no matching wire.
   // Probing a Gemini wire therefore has to name the surface whose codec builds
   // the request, rather than pretending the wire is a surface.
   const sourceSurface: SourceSurface = wireFamily === "gemini" ? "chat" : wireFamily;

package/src/providers/discovery/probe-wire.ts

@@ -6,15 +6,8 @@
  * answer is a layered merge — operator overrides, the discovery module's own
  * definition, the provider's static catalog, then a generic default — and it
  * lives here so `probing-service.ts` owns orchestration rather than merge
- * precedence.
- *
- * This module used to carry its own `DISCOVERY_CONFIG_BY_PROVIDER` table of
- * endpoint paths. It was a byte-identical second copy of
- * `PROVIDER_CAPABILITIES[id].endpointPathsByWireFamily` in the provider
- * registry, and it had already drifted: `ollamacloud` was listed here with four
- * paths while the registry map was the only one the catalog service validated
- * against. `discoveryPathsFor` now reads the registry, so a provider declares
- * its endpoint paths once.
+ * precedence. Endpoint paths come from the provider registry, so a provider
+ * declares its endpoint paths once rather than in a second table here.
  */
 import type { WireFamily } from "../../protocol/canonical-model";
 import type { CompatibilityProfile } from "../provider-metadata";
@@ -176,10 +169,10 @@ export function staticEndpointForWire(
  * Compiled-pattern memo.
  *
  * `matchModelWireFamily` runs per model on every discovery pass and on every
- * manual registration, and it used to compile each rule's `RegExp` inside that
- * loop — so the same handful of patterns were recompiled once per model per
- * pass. Keyed by the pattern source, which is exactly what `new RegExp(source)`
- * derives from.
+ * manual registration, so patterns are compiled once here instead of inside
+ * that loop — otherwise the same handful of patterns are recompiled once per
+ * model per pass. Keyed by the pattern source, which is exactly what
+ * `new RegExp(source)` derives from.
  *
  * A pattern that fails to compile is not memoized; the tolerance stays at read
  * time because an invalid pattern is rejected at write time, and a stored one