feat(gate): route prompt-selection through mcpd's server Llm (credential tiering)

Cloud/server keys belong at the k8s/mcpd level; mcplocal handles only the user's
personal tokens. The gate's prompt-selection used mcplocal's LOCAL provider
registry (heavy=anthropic = the user's OAuth token, which is API-gated and 404s),
so selection always degraded.

Now LlmPromptSelector tries sources in priority order: (1) the project's server
Llm via mcpd's inference proxy (POST /api/v1/llms/:name/infer) — cloud/server
keys stay at k8s — then (2) the local personal-token provider as fallback. First
source that returns valid selection JSON wins. Extracted pickCompletionText
(content ?? reasoning_content) + extractSelection helpers.

Wiring: GatePluginConfig.llmProvider (from the project), threaded via
createDefaultPlugin; handleBeginSession builds serverInfer from ctx.postToMcpd.
Verified live: POST .../llms/vllm-current/infer returns valid selection JSON
(qwen3-thinking, 4000-token budget). + selector tests (server preferred, local
fallback on error, server-only). mcplocal suite green (755).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/mcplocal/src/gate/llm-selector.ts

@@ -1,12 +1,16 @@
 /**
  * LLM-based prompt selection for the gating flow.
  *
- * Sends tags + prompt index to the heavy LLM, which returns
- * a ranked list of relevant prompt names.
+ * Sends tags + prompt index to a heavy LLM, which returns a ranked list of
+ * relevant prompt names. Credential tiering (user rule): prefer mcpd's server
+ * `Llm` (cloud/server keys live at the k8s level) via the inference proxy, and
+ * fall back to the local (personal-token) provider registry only when the
+ * server path is unavailable. Cloud keys never live in mcplocal config.
  */
 
 import type { ProviderRegistry } from '../providers/registry.js';
 import type { SystemPromptFetcher } from '../proxymodel/types.js';
+import type { ChatMessage, CompletionOptions } from '../providers/types.js';
 
 export interface PromptIndexForLlm {
   name: string;
@@ -20,6 +24,23 @@ export interface LlmSelectionResult {
   reasoning: string;
 }
 
+/**
+ * Route a selection inference through mcpd's server `Llm` (typically a
+ * `POST /api/v1/llms/:name/infer`). Returns the completion text. Must throw on
+ * transport/HTTP failure so the selector can fall back to the local provider.
+ */
+export type ServerInfer = (
+  messages: ChatMessage[],
+  opts: { maxTokens: number; temperature: number; signal?: AbortSignal },
+) => Promise<string>;
+
+export interface SelectPromptsOptions {
+  getSystemPromptFn?: SystemPromptFetcher;
+  signal?: AbortSignal;
+  /** Preferred inference path: mcpd server Llm (cloud/server keys at k8s). */
+  serverInfer?: ServerInfer;
+}
+
 /**
  * Token budget for the selection call. A reasoning model spends most of its
  * budget inside reasoning_content before emitting the {selectedNames} JSON, so
@@ -31,17 +52,49 @@ const SELECT_MAX_TOKENS = (() => {
   return Number.isFinite(v) && v > 0 ? v : 4000;
 })();
 
+/**
+ * Pull the assistant text from an OpenAI-shaped completion, falling back to
+ * reasoning_content — thinking models emit their answer there with content null.
+ */
+export function pickCompletionText(resp: unknown): string {
+  const m = (resp as {
+    choices?: Array<{
+      message?: {
+        content?: string | null;
+        reasoning_content?: string | null;
+        provider_specific_fields?: { reasoning_content?: string | null };
+      };
+    }>;
+  }).choices?.[0]?.message;
+  const primary = m?.content ?? '';
+  if (primary !== '') return primary;
+  return m?.reasoning_content ?? m?.provider_specific_fields?.reasoning_content ?? '';
+}
+
+/** Extract the `{ "selectedNames": [...], "reasoning": "..." }` object, or null. */
+export function extractSelection(text: string): LlmSelectionResult | null {
+  const jsonMatch = text.match(/\{[\s\S]*"selectedNames"[\s\S]*\}/);
+  if (!jsonMatch) return null;
+  try {
+    const parsed = JSON.parse(jsonMatch[0]) as { selectedNames?: string[]; reasoning?: string };
+    return { selectedNames: parsed.selectedNames ?? [], reasoning: parsed.reasoning ?? '' };
+  } catch {
+    return null;
+  }
+}
+
 export class LlmPromptSelector {
   constructor(
-    private readonly providerRegistry: ProviderRegistry,
+    private readonly providerRegistry: ProviderRegistry | null,
   ) {}
 
   async selectPrompts(
     tags: string[],
     promptIndex: PromptIndexForLlm[],
-    getSystemPromptFn?: SystemPromptFetcher,
-    signal?: AbortSignal,
+    opts: SelectPromptsOptions = {},
   ): Promise<LlmSelectionResult> {
+    const { getSystemPromptFn, signal, serverInfer } = opts;
+
     const DEFAULT_SYSTEM_PROMPT = `You are a context selection assistant. Given a developer's task keywords and a list of available project prompts, select which prompts are relevant to their work. Return a JSON object with "selectedNames" (array of prompt names) and "reasoning" (brief explanation). Priority 10 prompts must always be included.`;
     const systemPrompt = getSystemPromptFn
       ? await getSystemPromptFn('llm-gate-context-selector', DEFAULT_SYSTEM_PROMPT)
@@ -54,45 +107,62 @@ ${promptIndex.map((p) => `- ${p.name} (priority: ${p.priority}): ${p.summary ??
 
 Select the relevant prompts. Return JSON: { "selectedNames": [...], "reasoning": "..." }`;
 
-    const provider = this.providerRegistry.getProvider('heavy');
-    if (!provider) {
-      throw new Error('No heavy LLM provider available');
+    const messages: ChatMessage[] = [
+      { role: 'system', content: systemPrompt },
+      { role: 'user', content: userPrompt },
+    ];
+
+    // Sources in priority order: mcpd server Llm first (cloud/server keys at
+    // k8s), then the local personal-token provider as fallback.
+    const sources: Array<{ label: string; run: () => Promise<string> }> = [];
+    if (serverInfer) {
+      sources.push({
+        label: 'mcpd server Llm',
+        run: () => serverInfer(messages, { maxTokens: SELECT_MAX_TOKENS, temperature: 0, ...(signal ? { signal } : {}) }),
+      });
+    }
+    if (this.providerRegistry) {
+      sources.push({
+        label: 'local provider',
+        run: async () => {
+          const provider = this.providerRegistry!.getProvider('heavy');
+          if (!provider) throw new Error('No heavy LLM provider available');
+          const completionOptions: CompletionOptions = {
+            messages,
+            temperature: 0,
+            maxTokens: SELECT_MAX_TOKENS,
+            ...(signal ? { signal } : {}),
+          };
+          return (await provider.complete(completionOptions)).content;
+        },
+      });
+    }
+    if (sources.length === 0) {
+      throw new Error('No LLM provider available for prompt selection');
     }
 
-    // Deliberately NOT setting `model` here. Prompt-ranking only needs a
-    // working LLM; forcing the project's vLLM model onto the (anthropic) heavy
-    // provider made every selection fail silently. Use the provider's own model.
-    const completionOptions: import('../providers/types.js').CompletionOptions = {
-      messages: [
-        { role: 'system', content: systemPrompt },
-        { role: 'user', content: userPrompt },
-      ],
-      temperature: 0,
-      maxTokens: SELECT_MAX_TOKENS,
-      ...(signal ? { signal } : {}),
-    };
-
-    const result = await provider.complete(completionOptions);
-
-    const response = result.content;
-
-    // Parse JSON from response (may be wrapped in markdown code blocks)
-    const jsonMatch = response.match(/\{[\s\S]*"selectedNames"[\s\S]*\}/);
-    if (!jsonMatch) {
-      throw new Error('LLM response did not contain valid selection JSON');
-    }
-
-    const parsed = JSON.parse(jsonMatch[0]) as { selectedNames?: string[]; reasoning?: string };
-    const selectedNames = parsed.selectedNames ?? [];
-    const reasoning = parsed.reasoning ?? '';
-
-    // Always include priority 10 prompts
-    for (const p of promptIndex) {
-      if (p.priority === 10 && !selectedNames.includes(p.name)) {
-        selectedNames.push(p.name);
+    let lastErr: Error | null = null;
+    for (const src of sources) {
+      let text: string;
+      try {
+        text = await src.run();
+      } catch (err) {
+        lastErr = err instanceof Error ? err : new Error(String(err));
+        continue; // transport/HTTP failure → try the next source
       }
+      const sel = extractSelection(text);
+      if (!sel) {
+        lastErr = new Error(`${src.label}: LLM response did not contain valid selection JSON`);
+        continue;
+      }
+      // Always include priority 10 prompts.
+      for (const p of promptIndex) {
+        if (p.priority === 10 && !sel.selectedNames.includes(p.name)) {
+          sel.selectedNames.push(p.name);
+        }
+      }
+      return sel;
     }
-
-    return { selectedNames, reasoning };
+    throw lastErr ?? new Error('LLM prompt-selection failed');
   }
 }

src/mcplocal/src/http/project-mcp-endpoint.ts

@@ -147,6 +147,10 @@ export function registerProjectMcpEndpoint(app: FastifyInstance, mcpdClient: Mcp
       providerRegistry: effectiveRegistry,
     };
     if (resolvedModel) pluginConfig.modelOverride = resolvedModel;
+    // Route gate prompt-selection through this project's server Llm (mcpd
+    // inference proxy) so cloud/server keys stay at the k8s level; the local
+    // personal-token provider is the fallback. See credential-tiering.
+    if (mcpdConfig.llmProvider) pluginConfig.llmProvider = mcpdConfig.llmProvider;
     const basePlugin = createDefaultPlugin(pluginConfig);
     // Optional favourite-index presentation: curated favourite/<tool> + full
     // all/<server>/<tool> + a "prefer favourite/" instruction. Composed AFTER

src/mcplocal/src/proxymodel/plugins/gate.ts

@@ -13,7 +13,7 @@ import type { ProxyModelPlugin, PluginSessionContext } from '../plugin.js';
 import { SessionGate } from '../../gate/session-gate.js';
 import { TagMatcher, extractKeywordsFromToolCall, tokenizeDescription } from '../../gate/tag-matcher.js';
 import type { TagMatchResult } from '../../gate/tag-matcher.js';
-import { LlmPromptSelector } from '../../gate/llm-selector.js';
+import { LlmPromptSelector, pickCompletionText, type ServerInfer } from '../../gate/llm-selector.js';
 import type { ProviderRegistry } from '../../providers/registry.js';
 import { withTimeout, TimeoutError } from '../../util/with-timeout.js';
 
@@ -26,6 +26,13 @@ export interface GatePluginConfig {
   providerRegistry?: ProviderRegistry | null;
   modelOverride?: string;
   byteBudget?: number;
+  /**
+   * Name of the project's server `Llm` (mcpd). When set (and not 'none'), gate
+   * prompt-selection routes through mcpd's inference proxy first — keeping
+   * cloud/server keys at the k8s level — and only falls back to the local
+   * personal-token provider registry. See credential-tiering principle.
+   */
+  llmProvider?: string;
 }
 
 const MAX_RESPONSE_CHARS = 24_000;
@@ -33,8 +40,12 @@ const MAX_RESPONSE_CHARS = 24_000;
 export function createGatePlugin(config: GatePluginConfig = {}): ProxyModelPlugin {
   const isGated = config.gated !== false;
   const tagMatcher = new TagMatcher(config.byteBudget);
-  const llmSelector = config.providerRegistry
-    ? new LlmPromptSelector(config.providerRegistry)
+  // Prefer routing selection through mcpd's server Llm (cloud/server keys at
+  // k8s); the local provider registry (personal tokens) is the fallback. A gate
+  // with a server Llm but no local providers still selects (server-only).
+  const serverLlm = config.llmProvider && config.llmProvider !== 'none' ? config.llmProvider : undefined;
+  const llmSelector = (config.providerRegistry || serverLlm)
+    ? new LlmPromptSelector(config.providerRegistry ?? null)
     : null;
 
   // Per-session state tracking (plugin-scoped, not global SessionGate)
@@ -49,7 +60,7 @@ export function createGatePlugin(config: GatePluginConfig = {}): ProxyModelPlugi
 
       // Register begin_session virtual tool
       ctx.registerTool(getBeginSessionTool(llmSelector), async (args, callCtx) => {
-        return handleBeginSession(args, callCtx, sessionGate, tagMatcher, llmSelector);
+        return handleBeginSession(args, callCtx, sessionGate, tagMatcher, llmSelector, serverLlm);
       });
 
       // Register read_prompts virtual tool (available even when ungated)
@@ -250,6 +261,7 @@ async function handleBeginSession(
   sessionGate: SessionGate,
   tagMatcher: TagMatcher,
   llmSelector: LlmPromptSelector | null,
+  serverLlm?: string,
 ): Promise<unknown> {
   const rawTags = args['tags'] as string[] | undefined;
   const description = args['description'] as string | undefined;
@@ -287,8 +299,25 @@ async function handleBeginSession(
         chapters: p.chapters,
       }));
       const getSystemPromptFn = ctx.getSystemPrompt.bind(ctx);
+      // Prefer mcpd's server Llm (keeps cloud/server keys at k8s); the selector
+      // falls back to the local personal-token provider on failure.
+      const serverInfer: ServerInfer | undefined = serverLlm
+        ? async (messages, o) => {
+            const resp = await ctx.postToMcpd(`/api/v1/llms/${encodeURIComponent(serverLlm)}/infer`, {
+              messages,
+              temperature: o.temperature,
+              max_tokens: o.maxTokens,
+              stream: false,
+            });
+            return pickCompletionText(resp);
+          }
+        : undefined;
       const llmResult = await withTimeout(
-        (signal) => llmSelector.selectPrompts(tags, llmIndex, getSystemPromptFn, signal),
+        (signal) => llmSelector.selectPrompts(tags, llmIndex, {
+          getSystemPromptFn,
+          signal,
+          ...(serverInfer ? { serverInfer } : {}),
+        }),
         GATE_LLM_TIMEOUT_MS,
         'gate LLM prompt-selection',
       );

src/mcplocal/tests/llm-selector.test.ts

@@ -109,12 +109,49 @@ describe('LlmPromptSelector', () => {
     const selector = new LlmPromptSelector(registry);
     const ac = new AbortController();
 
-    await selector.selectPrompts(['test'], sampleIndex, undefined, ac.signal);
+    await selector.selectPrompts(['test'], sampleIndex, { signal: ac.signal });
 
     const call = (provider.complete as ReturnType<typeof vi.fn>).mock.calls[0]![0] as CompletionOptions;
     expect(call.signal).toBe(ac.signal);
   });
 
+  it('prefers the mcpd server Llm (serverInfer) over the local provider', async () => {
+    const provider = makeMockProvider('{ "selectedNames": ["mqtt-config"], "reasoning": "local" }');
+    const registry = makeRegistry(provider);
+    const selector = new LlmPromptSelector(registry);
+    const serverInfer = vi.fn().mockResolvedValue('{ "selectedNames": ["zigbee-pairing"], "reasoning": "server" }');
+
+    const result = await selector.selectPrompts(['x'], sampleIndex, { serverInfer });
+
+    expect(serverInfer).toHaveBeenCalledOnce();
+    expect(provider.complete).not.toHaveBeenCalled(); // server succeeded → no local call
+    expect(result.selectedNames).toContain('zigbee-pairing');
+    expect(result.reasoning).toBe('server');
+  });
+
+  it('falls back to the local provider when serverInfer throws', async () => {
+    const provider = makeMockProvider('{ "selectedNames": ["mqtt-config"], "reasoning": "local" }');
+    const registry = makeRegistry(provider);
+    const selector = new LlmPromptSelector(registry);
+    const serverInfer = vi.fn().mockRejectedValue(new Error('mcpd HTTP 503'));
+
+    const result = await selector.selectPrompts(['x'], sampleIndex, { serverInfer });
+
+    expect(serverInfer).toHaveBeenCalledOnce();
+    expect(provider.complete).toHaveBeenCalledOnce();
+    expect(result.selectedNames).toContain('mqtt-config');
+    expect(result.reasoning).toBe('local');
+  });
+
+  it('works server-only (null registry) when a serverInfer is supplied', async () => {
+    const selector = new LlmPromptSelector(null);
+    const serverInfer = vi.fn().mockResolvedValue('{ "selectedNames": ["mqtt-config"], "reasoning": "s" }');
+
+    const result = await selector.selectPrompts(['x'], sampleIndex, { serverInfer });
+
+    expect(result.selectedNames).toContain('mqtt-config');
+  });
+
   it('throws when no heavy provider is available', async () => {
     const registry = new ProviderRegistry(); // Empty registry
     const selector = new LlmPromptSelector(registry);