@quantiya/codevibe-core 2.0.47 → 2.0.49

package/dist/orchestration-shell/cli.d.ts

@@ -5,7 +5,7 @@ import { type ModeSelectionDeps } from './mode-selection';
 import { type RunOrchestrationShellArgs } from './index';
 import { type ShellEventEmit, type ShellEventEmitResult } from './emit-shell-event';
 import { type ShadowEnv } from '../local-executor/workspace-shadow';
-import { BackendPlannerClient, PlannerCacheLayer, PlannerHealthMachine, type LocalGemmaAdvisoryRunner, type LocalGemmaPlannerRunner } from '../planner';
+import { BackendPlannerClient, PlannerCacheLayer, PlannerHealthMachine, HostedGemmaAdvisoryRunner, type AdvisoryRunner, type LocalGemmaAdvisoryRunner, type LocalGemmaPlannerRunner } from '../planner';
 import type { PlannerAdapter, PlannerInput, PlannerDecision, PlannerProbeResult } from '../planner/adapter';
 import { keychainManager } from '../keychain';
 export { keychainManager };
@@ -270,6 +270,7 @@ export interface ProductionPlannerStackResult {
     adapter: BackendPlannerClient;
     cache: PlannerCacheLayer;
     health: PlannerHealthMachine;
+    hostedAdvisoryRunner: HostedGemmaAdvisoryRunner;
 }
 export declare function buildProductionPlannerStack(appsyncClient: AppSyncClient, emitter: (args: ShellEventEmit) => Promise<unknown>): ProductionPlannerStackResult;
 export interface ResolvePlannerAdapterArgs {
@@ -284,7 +285,7 @@ export interface ResolvedPlannerResult {
     adapter: PlannerAdapter;
     runtimeKind: NonNullable<RunOrchestrationShellArgs['plannerRuntimeKind']>;
     runtimeLabel: string;
-    advisoryRunner: LocalGemmaAdvisoryRunner | null;
+    advisoryRunner: AdvisoryRunner | null;
     classifyRunner: LocalGemmaPlannerRunner | null;
     cache: PlannerCacheLayer | null;
     health: PlannerHealthMachine | null;
@@ -295,14 +296,15 @@ export interface ResolvedPlannerResult {
  *
  * Precedence & Invariants:
  * 1. Forced hosted mode (`CODEVIBE_FORCE_HOSTED_PLANNER=1` or `CODEVIBE_PLANNER_PROVIDER=hosted`):
- *    - Allowed for PRO and MAX tiers. Returns the production hosted Bedrock planner stack.
+ *    - Allowed for PRO and MAX tiers. Returns the production hosted Bedrock planner stack and hosted advisory runner.
  *    - Disallowed for FREE tier (preserves $0 marginal cost invariant). Falls back to local adapter.
  * 2. Local-first preference:
- *    - When local Gemma is enabled and operational (`!localPlanner.reason`), returns the local adapter
- *      (`local_gemma_qat`).
+ *    - When local Gemma is enabled and operational (`!localPlanner.reason`), returns the composite adapter
+ *      and composite advisory runner for PRO/MAX (local-first with seamless hosted Bedrock fallback).
+ *      Returns local adapter/runner for FREE tier.
  * 3. Unavailable fallback:
  *    - When local Gemma is unavailable (`localPlanner.reason` is truthy):
- *      - PRO / MAX: Falls back to the hosted Bedrock planner stack (`hosted`).
+ *      - PRO / MAX: Falls back to the hosted Bedrock planner stack (`hosted`) and hosted advisory runner.
  *      - FREE: Retains the local unavailable adapter (`local_unavailable`), strictly preventing
  *        cloud inference costs.
  */

package/dist/orchestration-shell/cli.js

@@ -3380,6 +3380,21 @@ var queries = {
     }
   `
   ),
+  // Hosted Gemma advisory text generation (browse synthesis, brainstorm, etc.).
+  // Lambda decrypts input prompt via session key, dispatches to Bedrock Gemma,
+  // encrypts response under session key, returns.
+  generatePlannerAdvisory: (
+    /* GraphQL */
+    `
+    mutation GeneratePlannerAdvisory($input: GeneratePlannerAdvisoryInput!) {
+      generatePlannerAdvisory(input: $input) {
+        advisoryText
+        serverLatencyMs
+        providerUsed
+      }
+    }
+  `
+  ),
   // CP-8 min Stage 1 R1 closure (HIGH-1) — applyUserDecision SDK call.
   // Implementation deviation from #C8M-6 literal wording: codevibe-core
   // does NOT take a runtime dep on `@quantiya/quorum-core` (per the
@@ -5660,6 +5675,31 @@ var AppSyncClient = class _AppSyncClient {
   async pingPlanner(input) {
     return (await this.graphqlRequest(mutations.pingPlanner, { input })).data.pingPlanner;
   }
+  /**
+   * Hosted Gemma advisory text generation (browse synthesis, brainstorm, etc.).
+   *
+   * Implements `PlannerAppSyncTransport.generatePlannerAdvisory`.
+   * Input `prompt` is base64 AES-GCM ciphertext under the session key.
+   * Output `advisoryText` is base64 AES-GCM ciphertext under the session key.
+   */
+  async generatePlannerAdvisory(input) {
+    if (!input || typeof input != "object")
+      throw new Error("generatePlannerAdvisory: input object is required");
+    if (!input.sessionId || typeof input.sessionId != "string")
+      throw new Error("generatePlannerAdvisory: sessionId is required");
+    if (!input.prompt || typeof input.prompt != "string")
+      throw new Error("generatePlannerAdvisory: prompt ciphertext is required");
+    if (input.prompt.length > 2e5)
+      throw new Error(`generatePlannerAdvisory: prompt ciphertext exceeds 200,000 chars ceiling (${input.prompt.length})`);
+    if (input.maxTokens !== void 0 && (input.maxTokens < 1 || input.maxTokens > 1536))
+      throw new Error(`generatePlannerAdvisory: maxTokens must be between 1 and 1536 (got ${input.maxTokens})`);
+    if (input.temperature !== void 0 && (input.temperature < 0 || input.temperature > 1))
+      throw new Error(`generatePlannerAdvisory: temperature must be between 0 and 1 (got ${input.temperature})`);
+    let response = await this.graphqlRequest(mutations.generatePlannerAdvisory, { input });
+    if (!response?.data?.generatePlannerAdvisory)
+      throw new Error("generatePlannerAdvisory: empty response returned from AppSync");
+    return response.data.generatePlannerAdvisory;
+  }
   // MARK: - CP-8 min applyUserDecision (Stage 1 R1 HIGH-1 closure)
   /**
    * CP-8 min Stage 1 R1 HIGH-1 closure (Hendry-approved 2026-05-19).
@@ -63041,30 +63081,30 @@ ${summary}`);
           numPredict: BROWSE_RETRY_NUM_PREDICT
         });
         if (signal?.aborted) return;
-        let retrySummary = sanitizeForTerminal(parseLocalGemmaBrowseSummary(retryRaw)).trim();
+        let trustLine2 = runner?.providerKind === "hosted" ? "No code was changed." : "No hosted model was called and no code was changed.", retrySummary = sanitizeForTerminal(parseLocalGemmaBrowseSummary(retryRaw)).trim();
         if (retrySummary.length > 0) {
           if (signal?.aborted) return;
           advise(
             store,
             headerAdvised ? `${retrySummary}
 
-No hosted model was called and no code was changed.` : `${header}
+${trustLine2}` : `${header}
 
 ${retrySummary}
 
-No hosted model was called and no code was changed.`
+${trustLine2}`
           );
           return;
         }
       } catch (retryErr) {
         if (signal?.aborted) return;
-        logger.warn("[orchestration-shell] local browse advisory retry failed", {
+        logger.warn("[orchestration-shell] browse advisory retry failed", {
           error: retryErr.message,
           runtimeLabel: runner.runtimeLabel
         });
       }
       if (signal?.aborted) return;
-      let extract = buildBrowseExtract(safeText);
+      let isHosted = runner?.providerKind === "hosted", modelNoun = isHosted ? "The model" : "The local model", deadEndNoun = isHosted ? "the model" : "the local model", trustLine = isHosted ? "No code was changed." : "No hosted model was called and no code was changed.", extract = buildBrowseExtract(safeText);
       if (extract) {
         if (signal?.aborted) return;
         let extractPrefix = headerAdvised ? "" : `${header}
@@ -63072,18 +63112,18 @@ No hosted model was called and no code was changed.`
 `;
         advise(
           store,
-          `${extractPrefix}The local model couldn't summarize this page, so here is the extracted page text (an extract, not a summary):
+          `${extractPrefix}${modelNoun} couldn't summarize this page, so here is the extracted page text (an extract, not a summary):
 
 ${extract}
 
-${BROWSE_AGENT_OFFER} No hosted model was called and no code was changed.`
+${BROWSE_AGENT_OFFER} ${trustLine}`
         );
         return;
       }
       if (signal?.aborted) return;
       advise(
         store,
-        `Read ${dispFinal} but the local model couldn't summarize it. ${BROWSE_AGENT_OFFER} No hosted model was called and no code was changed.`
+        `Read ${dispFinal} but ${deadEndNoun} couldn't summarize it. ${BROWSE_AGENT_OFFER} ${trustLine}`
       );
     }
   }
@@ -63225,30 +63265,30 @@ ${summary}`);
         if (signal?.aborted) return;
         let retryRaw = await runner.generateAdvisory(retryPrompt, { responseFormat: "text", numCtx: 16384, think: !1 });
         if (signal?.aborted) return;
-        let retrySummary = sanitizeForTerminal(parseLocalGemmaBrowseSummary(retryRaw)).trim();
+        let trustLine2 = runner?.providerKind === "hosted" ? "No code was changed." : "No hosted model was called and no code was changed.", retrySummary = sanitizeForTerminal(parseLocalGemmaBrowseSummary(retryRaw)).trim();
         if (retrySummary.length > 0) {
           if (signal?.aborted) return;
           advise(
             store,
             headerAdvised ? `${retrySummary}
 
-No hosted model was called and no code was changed.` : `${header}
+${trustLine2}` : `${header}
 
 ${retrySummary}
 
-No hosted model was called and no code was changed.`
+${trustLine2}`
           );
           return;
         }
       } catch (retryErr) {
         if (signal?.aborted) return;
-        logger.warn("[orchestration-shell] local multi-browse retry failed", {
+        logger.warn("[orchestration-shell] multi-browse retry failed", {
           error: retryErr.message,
           runtimeLabel: runner.runtimeLabel
         });
       }
       if (signal?.aborted) return;
-      let extracts = sources.map((s) => {
+      let isHosted = runner?.providerKind === "hosted", modelNoun = isHosted ? "The model" : "The local model", deadEndNoun = isHosted ? "the model" : "the local model", trustLine = isHosted ? "No code was changed." : "No hosted model was called and no code was changed.", extracts = sources.map((s) => {
         let ext = buildBrowseExtract(s.content);
         return `[${s.id}] ${s.title}
 ${ext || "(No extract available)"}`;
@@ -63262,11 +63302,11 @@ ${ext || "(No extract available)"}`;
 `;
         advise(
           store,
-          `${extractPrefix}The local model couldn't summarize these pages, so here are the extracted highlights:
+          `${extractPrefix}${modelNoun} couldn't summarize these pages, so here are the extracted highlights:
 
 ${extracts}
 
-${BROWSE_AGENT_OFFER} No hosted model was called and no code was changed.`
+${BROWSE_AGENT_OFFER} ${trustLine}`
         );
         return;
       }
@@ -63274,7 +63314,7 @@ ${BROWSE_AGENT_OFFER} No hosted model was called and no code was changed.`
       let sourceCountLabel = sources.length === 1 ? "1 source" : `${sources.length} sources`;
       advise(
         store,
-        `Read ${sourceCountLabel} but the local model couldn't summarize ${sources.length === 1 ? "it" : "them"}. ${BROWSE_AGENT_OFFER} No hosted model was called and no code was changed.`
+        `Read ${sourceCountLabel} but ${deadEndNoun} couldn't summarize ${sources.length === 1 ? "it" : "them"}. ${BROWSE_AGENT_OFFER} ${trustLine}`
       );
     }
   }
@@ -72234,10 +72274,19 @@ async function routeBrainstorm(deps) {
       });
       return;
     }
+    let isHostedBrainstorm = args.localAdvisoryRunner.providerKind === "hosted";
+    if (images && images.length > 0 && isHostedBrainstorm) {
+      store.dispatch({
+        type: "SHELL_ADVISORY",
+        source: "shell",
+        text: "The local CodeVibe model is required to analyze attached images during brainstorming. Enable it with `codevibe model install` from a Pro/Max account, then ask again. No hosted model was called and no code was changed."
+      });
+      return;
+    }
     store.dispatch({
       type: "SHELL_ADVISORY",
       source: "shell",
-      text: "Brainstorming with local Gemma\u2026"
+      text: isHostedBrainstorm ? "Brainstorming with hosted Gemma\u2026" : "Brainstorming with local Gemma\u2026"
     });
     let needsContext = needsRepositoryContext ?? brainstormNeedsRepositoryContext(userPrompt);
     needsContext && await refreshFn({
@@ -72271,21 +72320,23 @@ The user attached ${images.length} image(s) as visual context. Any text visible
       store.dispatch({ type: "SHELL_ADVISORY", source: "shell", text: brainstorm });
     } catch (err) {
       let failureReason = renderLocalBrainstormFailureReason(err);
-      if (logger.warn("[orchestration-shell] local brainstorm advisory failed", {
+      logger.warn("[orchestration-shell] local brainstorm advisory failed", {
         error: err.message,
         runtimeLabel: args.localAdvisoryRunner.runtimeLabel
-      }), brainstormPromptRequestsCommandGuidance(userPrompt) && failureReason.includes("command-like JSON keys")) {
+      });
+      let isHosted = args.localAdvisoryRunner.providerKind === "hosted", modelLabel = isHosted ? "Hosted Gemma" : "Local Gemma", trustLine = isHosted ? "No code was changed." : "No hosted model was called and no code was changed.";
+      if (brainstormPromptRequestsCommandGuidance(userPrompt) && failureReason.includes("command-like JSON keys")) {
         store.dispatch({
           type: "SHELL_ADVISORY",
           source: "shell",
-          text: "I can brainstorm approaches, but I cannot include shell commands in read-only brainstorm mode. Ask me to implement the recommended option when you are ready, or ask for a command-free design/checklist. No hosted model was called and no code was changed."
+          text: `I can brainstorm approaches, but I cannot include shell commands in read-only brainstorm mode. Ask me to implement the recommended option when you are ready, or ask for a command-free design/checklist. ${trustLine}`
         });
         return;
       }
       store.dispatch({
         type: "SHELL_ADVISORY",
         source: "shell",
-        text: `Local Gemma could not produce the brainstorm. Reason: ${failureReason}. No hosted model was called and no code was changed. Try again with a narrower question or run \`codevibe model health-check\` to verify the local model.`
+        text: `${modelLabel} could not produce the brainstorm. Reason: ${failureReason}. ${trustLine} Try again with a narrower question or run \`codevibe model health-check\` to verify the model.`
       });
     }
   } finally {
@@ -72311,10 +72362,19 @@ async function routeAdvisory(deps) {
     });
     return;
   }
+  let isHostedAdvisory = args.localAdvisoryRunner.providerKind === "hosted";
+  if (hasImages && isHostedAdvisory) {
+    store.dispatch({
+      type: "SHELL_ADVISORY",
+      source: "shell",
+      text: "The local CodeVibe model is required to analyze an attached image. Enable it with `codevibe model install` from a Pro/Max account, then ask again. No hosted model was called and no code was changed."
+    });
+    return;
+  }
   store.dispatch({
     type: "SHELL_ADVISORY",
     source: "shell",
-    text: hasImages ? "Analyzing the attached image(s) with local Gemma\u2026" : "Answering with local Gemma\u2026"
+    text: hasImages ? "Analyzing the attached image(s) with local Gemma\u2026" : isHostedAdvisory ? "Answering with hosted Gemma\u2026" : "Answering with local Gemma\u2026"
   });
   let MAX_ADVISORY_CLARIFICATIONS = 4, MAX_ADVISORY_CLARIFICATION_QUESTION_CHARS = 400, MAX_ADVISORY_CLARIFICATION_ANSWER_CHARS = 1e3, MAX_ADVISORY_CANONICAL_CONTEXT_CHARS = 6e3, MAX_ADVISORY_USER_PROMPT_CHARS = 4e3, MAX_ADVISORY_PROMPT_CHARS = 42e3, MAX_ADVISORY_PAGE_CHARS = 24e3, MAX_ADVISORY_STATUS_CHARS = 1500, MAX_ADVISORY_PAGE_TOKENS = 6e3, MAX_ADVISORY_PROMPT_TOKENS = 13500, statusText = "";
   try {
@@ -72448,21 +72508,21 @@ async function routeAdvisory(deps) {
       });
       return;
     }
-    let fallback = deps.fallbackSummary?.trim();
+    let fallback = deps.fallbackSummary?.trim(), effectiveIsHosted = args.localAdvisoryRunner.providerKind === "hosted", modelNoun = effectiveIsHosted ? "The model" : "The local model", modelLabel = effectiveIsHosted ? "Hosted Gemma" : "Local Gemma", trustLine = effectiveIsHosted ? "No code was changed." : "No hosted model was called and no code was changed.", generationScope = effectiveIsHosted ? "hosted" : "local";
     if (fallback) {
       store.dispatch({
         type: "SHELL_ADVISORY",
         source: "shell",
         text: sanitizeForTerminal(`${fallback}
 
-(Note: detailed local generation returned an empty answer; showing concise answer.)`)
+(Note: detailed ${generationScope} generation returned an empty answer; showing concise answer.)`)
       });
       return;
     }
     store.dispatch({
       type: "SHELL_ADVISORY",
       source: "shell",
-      text: hasImages ? "The local model returned no answer for the attached image(s). No code was changed." : "The local model returned an empty answer. Try rephrasing with a bit more detail. No code was changed."
+      text: hasImages ? `${modelNoun} returned no answer for the attached image(s). ${trustLine}` : `${modelNoun} returned an empty answer. Try rephrasing with a bit more detail. ${trustLine}`
     });
   } catch (err) {
     logger.warn("[orchestration-shell] local advisory failed", {
@@ -72470,7 +72530,7 @@ async function routeAdvisory(deps) {
       hasImages,
       runtimeLabel: args.localAdvisoryRunner.runtimeLabel
     });
-    let fallback = deps.fallbackSummary?.trim();
+    let fallback = deps.fallbackSummary?.trim(), effectiveIsHosted = args.localAdvisoryRunner.providerKind === "hosted", modelLabel = effectiveIsHosted ? "Hosted Gemma" : "Local Gemma", trustLine = effectiveIsHosted ? "No code was changed." : "No hosted model was called and no code was changed.", generationScope = effectiveIsHosted ? "hosted" : "local";
     if (fallback) {
       let reason = renderLocalBrainstormFailureReason(err);
       store.dispatch({
@@ -72478,14 +72538,14 @@ async function routeAdvisory(deps) {
         source: "shell",
         text: sanitizeForTerminal(`${fallback}
 
-(Note: detailed local generation was unavailable: ${reason}. Showing concise answer.)`)
+(Note: detailed ${generationScope} generation was unavailable: ${reason}. Showing concise answer.)`)
       });
       return;
     }
     store.dispatch({
       type: "SHELL_ADVISORY",
       source: "shell",
-      text: hasImages ? `Local Gemma could not analyze the attached image(s). Reason: ${renderLocalBrainstormFailureReason(err)}. No hosted model was called and no code was changed.` : `Local Gemma could not answer. Reason: ${renderLocalBrainstormFailureReason(err)}. No hosted model was called and no code was changed.`
+      text: hasImages ? `${modelLabel} could not analyze the attached image(s). Reason: ${renderLocalBrainstormFailureReason(err)}. ${trustLine}` : `${modelLabel} could not answer. Reason: ${renderLocalBrainstormFailureReason(err)}. ${trustLine}`
     });
   }
 }
@@ -75731,6 +75791,109 @@ Local planner error. Falling back to hosted Gemma planner on AWS.
   }
 };
 
+// src/planner/hosted-advisory.ts
+init_logger2();
+var HostedGemmaAdvisoryRunner = class {
+  constructor(deps) {
+    this.deps = deps;
+    this.runtimeLabel = "Hosted Gemma (AWS Bedrock)";
+    this.providerKind = "hosted";
+  }
+  async generateAdvisory(promptText, options) {
+    if (options?.images && options.images.length > 0)
+      throw new Error(
+        "HostedGemmaAdvisoryRunner: image attachments are not supported by hosted Bedrock Gemma. A local vision model is required."
+      );
+    let sessionId = this.deps.activeSessionIdHolder.value;
+    if (!sessionId)
+      throw new Error("Hosted Gemma advisory generation failed: no active session ID");
+    let sessionKey = await this.deps.sessionKeyResolver.getSessionKey(sessionId);
+    if (!sessionKey)
+      throw new Error(`Hosted Gemma advisory generation failed: session key unresolved for session ${sessionId}`);
+    let encryptedPrompt = this.deps.crypto.encryptString(promptText, sessionKey), maxTokens = options?.numPredict, temperature = options?.temperature, start = Date.now();
+    try {
+      let response = await this.deps.transport.generatePlannerAdvisory({
+        sessionId,
+        prompt: encryptedPrompt,
+        ...maxTokens !== void 0 ? { maxTokens } : {},
+        ...temperature !== void 0 ? { temperature } : {}
+      });
+      return this.deps.crypto.decryptString(response.advisoryText, sessionKey);
+    } catch (err) {
+      let elapsedMs = Date.now() - start;
+      throw logger.warn("[hosted-advisory] generatePlannerAdvisory failed", {
+        sessionId,
+        elapsedMs,
+        error: err?.message
+      }), err;
+    }
+  }
+  async warm() {
+    return !0;
+  }
+  async unload() {
+    return !0;
+  }
+};
+
+// src/planner/composite-advisory.ts
+init_logger2();
+var CompositeAdvisoryRunner = class {
+  constructor(opts) {
+    this.opts = opts;
+    this.localHealthy = !!opts.localRunner;
+  }
+  get runtimeLabel() {
+    return this.opts.tier === "FREE" || this.localHealthy ? this.opts.localRunner?.runtimeLabel ?? "Local Gemma" : this.opts.hostedRunner.runtimeLabel;
+  }
+  get providerKind() {
+    return this.opts.tier === "FREE" || this.localHealthy ? "local" : "hosted";
+  }
+  async generateAdvisory(promptText, options) {
+    if (this.opts.tier === "FREE") {
+      if (!this.opts.localRunner)
+        throw new Error(
+          "LocalModelUnavailable: local Gemma is not installed or enabled. Free tier operates in local-only mode."
+        );
+      return this.opts.localRunner.generateAdvisory(promptText, options);
+    }
+    if (options?.images && options.images.length > 0) {
+      if (!this.localHealthy || !this.opts.localRunner)
+        throw new Error(
+          "LocalModelUnavailable: local CodeVibe vision model is required to analyze attached images."
+        );
+      return this.opts.localRunner.generateAdvisory(promptText, options);
+    }
+    if (this.localHealthy && this.opts.localRunner)
+      try {
+        return await this.opts.localRunner.generateAdvisory(promptText, options);
+      } catch (err) {
+        logger.warn("[composite-advisory] local advisory failed; falling back to hosted Bedrock Gemma", {
+          error: err?.message
+        }), this.localHealthy = !1, this.opts.onFallbackToHosted?.(err);
+      }
+    return this.opts.hostedRunner.generateAdvisory(promptText, options);
+  }
+  async warm() {
+    if (this.opts.localRunner?.warm)
+      try {
+        return await this.opts.localRunner.warm();
+      } catch {
+        return !1;
+      }
+    return !0;
+  }
+  async unload() {
+    if (this.opts.localRunner?.unload)
+      try {
+        return await this.opts.localRunner.unload();
+      } catch {
+        return !1;
+      }
+    return !0;
+  }
+};
+
 // src/session/session-rekey.ts
 init_logger2();
 
@@ -81250,7 +81413,8 @@ ${err.message}
 function buildProductionPlannerStack(appsyncClient, emitter) {
   let transport = {
     classifyPlannerPrompt: (input) => appsyncClient.classifyPlannerPrompt(input),
-    pingPlanner: (input) => appsyncClient.pingPlanner(input)
+    pingPlanner: (input) => appsyncClient.pingPlanner(input),
+    generatePlannerAdvisory: (input) => appsyncClient.generatePlannerAdvisory(input)
   }, cache = new PlannerCacheLayer(), adaptedEmitter = async (e) => {
     await emitter({
       sessionId: e.sessionId,
@@ -81291,9 +81455,18 @@ function buildProductionPlannerStack(appsyncClient, emitter) {
     sessionKeyResolver,
     adaptedEmitter
   ), realSetActiveSession = adapter.setActiveSession.bind(adapter);
-  return adapter.setActiveSession = (sessionId) => {
+  adapter.setActiveSession = (sessionId) => {
     activeSessionIdHolder.value = sessionId, realSetActiveSession(sessionId);
-  }, { adapter, cache, health };
+  };
+  let hostedAdvisoryRunner = new HostedGemmaAdvisoryRunner({
+    transport: {
+      generatePlannerAdvisory: (input) => appsyncClient.generatePlannerAdvisory(input)
+    },
+    crypto: crypto33,
+    sessionKeyResolver,
+    activeSessionIdHolder
+  });
+  return { adapter, cache, health, hostedAdvisoryRunner };
 }
 function resolvePlannerAdapter(args) {
   let env = args.env ?? process.env, forceHosted = env.CODEVIBE_FORCE_HOSTED_PLANNER === "1" || env.CODEVIBE_PLANNER_PROVIDER === "hosted", tier = args.tier, isPaidTier = tier === "PRO" || tier === "MAX";
@@ -81304,7 +81477,7 @@ function resolvePlannerAdapter(args) {
         adapter: hostedStack.adapter,
         runtimeKind: "hosted",
         runtimeLabel: "Hosted Gemma (AWS Bedrock)",
-        advisoryRunner: args.localPlanner.advisoryRunner,
+        advisoryRunner: hostedStack.hostedAdvisoryRunner,
         classifyRunner: args.localPlanner.classifyRunner,
         cache: hostedStack.cache,
         health: hostedStack.health,
@@ -81336,6 +81509,10 @@ function resolvePlannerAdapter(args) {
         onNotice: (msg) => {
           process.stdout.write(msg);
         }
+      }), compositeAdvisoryRunner = new CompositeAdvisoryRunner({
+        localRunner: args.localPlanner.advisoryRunner,
+        hostedRunner: hostedStack.hostedAdvisoryRunner,
+        tier
       });
       return {
         adapter: compositeAdapter,
@@ -81345,7 +81522,7 @@ function resolvePlannerAdapter(args) {
         get runtimeLabel() {
           return compositeAdapter.runtimeLabel;
         },
-        advisoryRunner: args.localPlanner.advisoryRunner,
+        advisoryRunner: compositeAdvisoryRunner,
         classifyRunner: args.localPlanner.classifyRunner,
         cache: hostedStack.cache,
         health: hostedStack.health,
@@ -81378,7 +81555,7 @@ function resolvePlannerAdapter(args) {
       adapter: hostedStack.adapter,
       runtimeKind: "hosted",
       runtimeLabel: "Hosted Gemma (AWS Bedrock)",
-      advisoryRunner: null,
+      advisoryRunner: hostedStack.hostedAdvisoryRunner,
       classifyRunner: args.localPlanner.classifyRunner,
       cache: hostedStack.cache,
       health: hostedStack.health,

package/dist/planner/client.d.ts

@@ -26,6 +26,16 @@ export interface PlannerAppSyncTransport {
         ok: boolean;
         ms: number;
     }>;
+    generatePlannerAdvisory?(input: {
+        sessionId: string;
+        prompt: string;
+        maxTokens?: number;
+        temperature?: number;
+    }): Promise<{
+        advisoryText: string;
+        serverLatencyMs: number;
+        providerUsed: string;
+    }>;
 }
 /**
  * Crypto bridge — `encryptString` + `encryptJson` + `decryptJson`

package/dist/planner/composite-advisory.d.ts

@@ -0,0 +1,25 @@
+import type { Tier } from '../orchestration-shell/types';
+import type { AdvisoryOptions, AdvisoryRunner } from './local-advisory';
+export interface CompositeAdvisoryRunnerOptions {
+    localRunner?: AdvisoryRunner | null;
+    hostedRunner: AdvisoryRunner;
+    tier: Tier;
+    onFallbackToHosted?: (error: Error) => void;
+}
+/**
+ * CompositeAdvisoryRunner coordinates local and hosted advisory execution:
+ * - Free Tier Invariant: strictly isolates Free tier from cloud inference ($0 marginal cost).
+ * - Local-First Preference: routes natural language turns / browse to local Gemma when healthy.
+ * - Dynamic Failover: if local Gemma fails (Ollama down, crash, OOM), seamlessly falls back
+ *   to hosted Bedrock Gemma for Pro/Max users without interrupting the session.
+ */
+export declare class CompositeAdvisoryRunner implements AdvisoryRunner {
+    private readonly opts;
+    private localHealthy;
+    constructor(opts: CompositeAdvisoryRunnerOptions);
+    get runtimeLabel(): string;
+    get providerKind(): 'local' | 'hosted';
+    generateAdvisory(promptText: string, options?: AdvisoryOptions): Promise<string>;
+    warm(): Promise<boolean>;
+    unload(): Promise<boolean>;
+}

package/dist/planner/hosted-advisory.d.ts

@@ -0,0 +1,37 @@
+import type { AdvisoryOptions, AdvisoryRunner } from './local-advisory';
+import type { PlannerCryptoBridge, SessionKeyResolver } from './client';
+export interface HostedAdvisoryTransport {
+    generatePlannerAdvisory(input: {
+        sessionId: string;
+        prompt: string;
+        maxTokens?: number;
+        temperature?: number;
+    }): Promise<{
+        advisoryText: string;
+        serverLatencyMs: number;
+        providerUsed: string;
+    }>;
+}
+export interface HostedGemmaAdvisoryRunnerDeps {
+    transport: HostedAdvisoryTransport;
+    crypto: PlannerCryptoBridge;
+    sessionKeyResolver: SessionKeyResolver;
+    activeSessionIdHolder: {
+        value: string | null;
+    };
+}
+/**
+ * HostedGemmaAdvisoryRunner executes advisory text generation (browse synthesis,
+ * brainstorm, etc.) against Google Gemma 3 12B on AWS Bedrock via AppSync planner proxy.
+ *
+ * Wire payload is encrypted end-to-end with AES-256-GCM under the session key.
+ */
+export declare class HostedGemmaAdvisoryRunner implements AdvisoryRunner {
+    private readonly deps;
+    readonly runtimeLabel = "Hosted Gemma (AWS Bedrock)";
+    readonly providerKind: "hosted";
+    constructor(deps: HostedGemmaAdvisoryRunnerDeps);
+    generateAdvisory(promptText: string, options?: AdvisoryOptions): Promise<string>;
+    warm(): Promise<boolean>;
+    unload(): Promise<boolean>;
+}

package/dist/planner/index.d.ts

@@ -4,5 +4,7 @@ export { PlannerCacheLayer } from './cache';
 export { PlannerHealthMachine } from './health-state';
 export { BackendPlannerClient, PlannerBudgetExceededError, PlannerTierGateRejectedError, type PlannerAppSyncTransport, type PlannerCryptoBridge, type SessionKeyResolver, type ShellEventEmit, type EmitShellEventFn, } from './client';
 export { LocalGemmaPlannerAdapter, parseLocalGemmaPlannerDecision, PlannerOutputUnparseableError, renderLocalGemmaPlannerPrompt, type LocalGemmaPlannerRunner, } from './local-gemma';
-export { parseLocalGemmaAdvisorySummary, renderLocalGemmaFamiliarizePrompt, renderLocalGemmaMultiBrowsePrompt, renderLocalGemmaBrowsePrompt, MAX_BROWSE_CONTENT_CHARS, MAX_MULTI_BROWSE_TOTAL_CONTENT_CHARS, MAX_MULTI_BROWSE_RENDERED_PROMPT_CHARS, type LocalGemmaAdvisoryRunner, } from './local-advisory';
+export { parseLocalGemmaAdvisorySummary, renderLocalGemmaFamiliarizePrompt, renderLocalGemmaMultiBrowsePrompt, renderLocalGemmaBrowsePrompt, MAX_BROWSE_CONTENT_CHARS, MAX_MULTI_BROWSE_TOTAL_CONTENT_CHARS, MAX_MULTI_BROWSE_RENDERED_PROMPT_CHARS, type AdvisoryRunner, type AdvisoryOptions, type AdvisoryResponseFormat, type LocalGemmaAdvisoryRunner, type LocalGemmaAdvisoryOptions, type LocalGemmaAdvisoryResponseFormat, } from './local-advisory';
 export { CompositePlannerAdapter, type CompositePlannerAdapterOptions, } from './composite-adapter';
+export { HostedGemmaAdvisoryRunner, type HostedGemmaAdvisoryRunnerDeps, type HostedAdvisoryTransport, } from './hosted-advisory';
+export { CompositeAdvisoryRunner, type CompositeAdvisoryRunnerOptions, } from './composite-advisory';

package/dist/planner/local-advisory.d.ts

@@ -1,6 +1,6 @@
 import type { StructuralSummary } from '../structural-summary/types';
-export interface LocalGemmaAdvisoryRunner {
-    generateAdvisory(promptText: string, options?: LocalGemmaAdvisoryOptions): Promise<string>;
+export interface AdvisoryRunner {
+    generateAdvisory(promptText: string, options?: AdvisoryOptions): Promise<string>;
     /**
      * Dogfood #618 — optional fire-and-forget model pre-warm (load + keep_alive).
      * Never throws; resolves false on failure. Called once at shell startup.
@@ -13,10 +13,11 @@ export interface LocalGemmaAdvisoryRunner {
      */
     unload?(): Promise<boolean>;
     runtimeLabel: string;
+    readonly providerKind?: 'local' | 'hosted';
 }
-export type LocalGemmaAdvisoryResponseFormat = 'json' | 'text';
-export interface LocalGemmaAdvisoryOptions {
-    responseFormat?: LocalGemmaAdvisoryResponseFormat;
+export type AdvisoryResponseFormat = 'json' | 'text';
+export interface AdvisoryOptions {
+    responseFormat?: AdvisoryResponseFormat;
     numPredict?: number;
     /**
      * RAW base64 images (NO `data:` prefix) for the MULTIMODAL local answerer
@@ -38,7 +39,15 @@ export interface LocalGemmaAdvisoryOptions {
      * Omitted unless set; browse callers pass false (R8-F2, R9-F4).
      */
     think?: boolean;
+    /**
+     * Temperature hint for advisory generation.
+     */
+    temperature?: number;
 }
+/** Backward-compatibility aliases for local runner consumers */
+export type LocalGemmaAdvisoryRunner = AdvisoryRunner;
+export type LocalGemmaAdvisoryResponseFormat = AdvisoryResponseFormat;
+export type LocalGemmaAdvisoryOptions = AdvisoryOptions;
 export declare const MAX_RENDERED_PROMPT_CHARS = 30000;
 export declare const MAX_RENDERED_PROMPT_TOKENS = 7500;
 export declare function redactAbsoluteLocalPaths(text: string): string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@quantiya/codevibe-core",
-  "version": "2.0.47",
+  "version": "2.0.49",
   "description": "Core library for CodeVibe plugins - shared keychain, crypto, AppSync, and auth functionality",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",