@cloudpeers-jkl/model-router 0.2.1 → 0.2.2

package/README.md

@@ -68,7 +68,9 @@ plus the provider keys (`ANTHROPIC_API_KEY`, `GEMINI_SERVER_API_KEY`/`GEMINI_API
 
 ## Publishing
 
-Not yet published. Consumed in-repo via relative imports (see
-`server/lib/model-router.ts`). To publish for external consumers
-(carepeers seams, gemini-proxy in Deno): `npm run build && npm publish`
-from this directory — `dist/` is what ships.
+Published on npm since 2026-07-15 (`@cloudpeers-jkl/model-router`, current
+0.2.1); carepeers consumes the npm package. mcp itself consumes this source
+in-repo via relative imports (see `server/lib/model-router.ts`) — source and
+package must not drift, so publish after any change here:
+`npm publish --access public` from this directory (`prepublishOnly` builds
+`dist/`; requires JKL's npm browser 2FA).

package/dist/policy.d.ts

@@ -27,6 +27,13 @@ export declare function gateEnforced(): boolean;
  * Never triggers for aggregate_only/externalizable (spec scope: local_only).
  */
 export declare function evaluateGate(sovereigntyClass: SovereigntyClass, tier: TierDef): GateOutcome;
+/**
+ * §4 — the caller's `model` is honored when it belongs to the selected
+ * tier's provider family (a gemini caller may pick flash vs pro); otherwise
+ * the tier's policy model serves. Tier 0 always serves its own model — the
+ * self-hosted endpoint decides what it hosts, not the caller.
+ */
+export declare function modelMatchesProvider(model: string, provider: Provider): boolean;
 /**
  * Resolve the tier chain for a request: the static task-class chain,
  * optionally reordered by an advisory backend_hint (§4). The hint never adds

package/dist/policy.js

@@ -31,6 +31,19 @@ export function evaluateGate(sovereigntyClass, tier) {
         return 'allowed'; // Stage 2 governed egress
     return gateEnforced() ? 'blocked' : 'would_block'; // Stage 3 : Stage 1
 }
+/**
+ * §4 — the caller's `model` is honored when it belongs to the selected
+ * tier's provider family (a gemini caller may pick flash vs pro); otherwise
+ * the tier's policy model serves. Tier 0 always serves its own model — the
+ * self-hosted endpoint decides what it hosts, not the caller.
+ */
+export function modelMatchesProvider(model, provider) {
+    if (provider === 'gemini')
+        return /^(models\/)?gemini/i.test(model);
+    if (provider === 'anthropic')
+        return /^claude/i.test(model);
+    return false; // selfhosted: never caller-selected
+}
 /**
  * Resolve the tier chain for a request: the static task-class chain,
  * optionally reordered by an advisory backend_hint (§4). The hint never adds

package/dist/router.js

@@ -14,7 +14,7 @@
  * stable interface: the Stage-3 enforcement flip decision is made from them.
  */
 import { RouterBlockedError } from './errors.js';
-import { evaluateGate, resolveChain, TIERS } from './policy.js';
+import { evaluateGate, modelMatchesProvider, resolveChain, TIERS } from './policy.js';
 import { parseRouterRequest } from './schema.js';
 function firstText(content) {
     return content.find((b) => b.type === 'text')?.text ?? '';
@@ -34,6 +34,10 @@ function stripEnvelope(req, model) {
     const { cloudpeers: _cloudpeers, stream: _stream, ...rest } = req;
     return { ...rest, model };
 }
+/** §4 — caller model honored when it belongs to the tier's provider family. */
+function servedModel(req, tier) {
+    return req.model && modelMatchesProvider(req.model, tier.provider) ? req.model : tier.model;
+}
 /**
  * §7.1 #2/#4 + A1.2 — pre-call enforcement that applies to the request as a
  * whole. Fail closed: a named instrument with no wired verifier blocks.
@@ -160,7 +164,7 @@ async function postCall(candidate, stripped, response, latencyMs, req, ctx, deps
         .meter({
         userId: ctx.userId,
         serviceId: ctx.serviceId,
-        model: candidate.tier.model,
+        model: stripped.model ?? candidate.tier.model,
         operation: ctx.operation,
         promptTokens: usage.tokens_in,
         completionTokens: usage.tokens_out,
@@ -192,7 +196,7 @@ export async function routeMessages(input, ctx, deps) {
     let lastError = null;
     let bestEffort = null;
     for await (const candidate of eligibleTiers(req, ctx, deps, state)) {
-        const stripped = stripEnvelope(req, candidate.tier.model);
+        const stripped = stripEnvelope(req, servedModel(req, candidate.tier));
         const started = performance.now();
         try {
             const response = await candidate.adapter.invoke(stripped, {
@@ -205,7 +209,7 @@ export async function routeMessages(input, ctx, deps) {
             const result = {
                 response,
                 provider: candidate.tier.provider,
-                model: candidate.tier.model,
+                model: stripped.model ?? candidate.tier.model,
                 tier: candidate.tier.tier,
                 escalations,
                 degraded: false,
@@ -244,7 +248,7 @@ export async function routeMessagesStream(input, ctx, deps) {
     for await (const candidate of eligibleTiers(req, ctx, deps, state)) {
         if (!candidate.adapter.supports_streaming)
             continue;
-        const stripped = stripEnvelope(req, candidate.tier.model);
+        const stripped = stripEnvelope(req, servedModel(req, candidate.tier));
         const gate = state.requestGate === 'would_block' ? 'would_block' : candidate.gate;
         const started = performance.now();
         const events = candidate.adapter.invokeStream(stripped, {
@@ -277,7 +281,7 @@ export async function routeMessagesStream(input, ctx, deps) {
                 type: 'message',
                 role: 'assistant',
                 content: [],
-                model: candidate.tier.model,
+                model: stripped.model ?? candidate.tier.model,
                 stop_reason: stopReason,
                 usage: { input_tokens: inputTokens, output_tokens: outputTokens },
             };
@@ -285,7 +289,7 @@ export async function routeMessagesStream(input, ctx, deps) {
         };
         return {
             provider: candidate.tier.provider,
-            model: candidate.tier.model,
+            model: stripped.model ?? candidate.tier.model,
             tier: candidate.tier.tier,
             gate,
             stream: metered(),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cloudpeers-jkl/model-router",
-  "version": "0.2.1",
+  "version": "0.2.2",
   "description": "cloudpeers inference router — embedded library (no central data plane). Sovereignty privacy gate, static task-class routing, Anthropic Messages lingua franca, uniform metering.",
   "type": "module",
   "license": "MIT",
@@ -21,9 +21,9 @@
     "test": "vitest run",
     "prepublishOnly": "npm run build"
   },
-  "peerDependencies": {
-    "@google/generative-ai": ">=0.24.0",
-    "zod": ">=3.22.0"
+  "dependencies": {
+    "@google/generative-ai": "^0.24.1",
+    "zod": "^3.24.2"
   },
   "repository": {
     "type": "git",