@j0hanz/code-review-analyst-mcp 1.4.4 → 1.5.1

package/dist/lib/model-config.d.ts

@@ -1,44 +1,39 @@
 /** Fast, cost-effective model for summarization and light analysis. */
-export declare const FLASH_MODEL = "gemini-2.5-flash";
+export declare const FLASH_MODEL = "gemini-3-flash-preview";
 /** High-capability model for deep reasoning, quality inspection, and reliable code generation. */
-export declare const PRO_MODEL = "gemini-2.5-pro";
-/**
- * Thinking budget (tokens) for Flash triage tools (impact, summary, API-breaking).
- * Explicitly disabled (0) — these are classification/extraction tasks that do not
- * benefit from a reasoning chain. Avoids default dynamic-thinking overhead.
- * Flash 2.5 range: 0–24_576.
- */
-export declare const FLASH_TRIAGE_THINKING_BUDGET: 0;
-/** Thinking budget (tokens) for Flash analysis tasks (test plans, complexity). */
-export declare const FLASH_THINKING_BUDGET: 16384;
-/** Thinking budget (tokens) for Pro model deep-analysis tasks (quality, patches). */
-export declare const PRO_THINKING_BUDGET: 24576;
-/** Output cap for Flash triage tools (impact, summary). */
-export declare const FLASH_TRIAGE_MAX_OUTPUT_TOKENS: 4096;
-/** Output cap for API breaking-change detection (migration guidance needs room). */
-export declare const FLASH_API_BREAKING_MAX_OUTPUT_TOKENS: 4096;
-/** Output cap for test-plan generation (includes pseudocode snippets). */
-export declare const FLASH_TEST_PLAN_MAX_OUTPUT_TOKENS: 8192;
-/** Output cap for Pro deep review findings. */
-export declare const PRO_REVIEW_MAX_OUTPUT_TOKENS: 12288;
-/** Output cap for Pro search/replace remediation blocks. */
-export declare const PRO_PATCH_MAX_OUTPUT_TOKENS: 8192;
-/** Output cap for Flash complexity analysis reports. */
-export declare const FLASH_COMPLEXITY_MAX_OUTPUT_TOKENS: 2048;
-/** Extended timeout for Pro model calls (ms). Pro thinks longer than Flash. */
+export declare const PRO_MODEL = "gemini-3-pro-preview";
+/** Default language hint. */
+export declare const DEFAULT_LANGUAGE = "detect";
+/** Default test-framework hint. */
+export declare const DEFAULT_FRAMEWORK = "detect";
+/** Extended timeout for Pro model calls (ms). */
 export declare const DEFAULT_TIMEOUT_PRO_MS = 120000;
-/** Temperature for triage/classification tools (deterministic structured extraction). */
-export declare const TRIAGE_TEMPERATURE: 0.1;
-/** Temperature for analytical tools (consistent algorithmic reasoning). */
-export declare const ANALYSIS_TEMPERATURE: 0.1;
-/** Temperature for code patch generation (maximum precision for search blocks). */
-export declare const PATCH_TEMPERATURE: 0;
-/** Temperature for creative synthesis tools (test plan generation). */
-export declare const CREATIVE_TEMPERATURE: 0.2;
 export declare const MODEL_TIMEOUT_MS: {
     readonly defaultPro: 120000;
 };
-/** Default language hint when not specified by the user. Tells the model to auto-detect. */
-export declare const DEFAULT_LANGUAGE = "detect";
-/** Default test-framework hint when not specified by the user. Tells the model to auto-detect. */
-export declare const DEFAULT_FRAMEWORK = "detect";
+/** Thinking level for Flash triage. */
+export declare const FLASH_TRIAGE_THINKING_LEVEL: "minimal";
+/** Thinking level for Flash analysis. */
+export declare const FLASH_THINKING_LEVEL: "medium";
+/** Thinking level for Pro deep analysis. */
+export declare const PRO_THINKING_LEVEL: "high";
+/** Output cap for Flash API breaking-change detection. */
+export declare const FLASH_API_BREAKING_MAX_OUTPUT_TOKENS: 4096;
+/** Output cap for Flash complexity analysis. */
+export declare const FLASH_COMPLEXITY_MAX_OUTPUT_TOKENS: 2048;
+/** Output cap for Flash test-plan generation. */
+export declare const FLASH_TEST_PLAN_MAX_OUTPUT_TOKENS: 8192;
+/** Output cap for Flash triage tools. */
+export declare const FLASH_TRIAGE_MAX_OUTPUT_TOKENS: 4096;
+/** Output cap for Pro patch generation. */
+export declare const PRO_PATCH_MAX_OUTPUT_TOKENS: 8192;
+/** Output cap for Pro deep review findings. */
+export declare const PRO_REVIEW_MAX_OUTPUT_TOKENS: 12288;
+/** Temperature for analytical tools. */
+export declare const ANALYSIS_TEMPERATURE: 1;
+/** Temperature for creative synthesis (test plans). */
+export declare const CREATIVE_TEMPERATURE: 1;
+/** Temperature for code patch generation. */
+export declare const PATCH_TEMPERATURE: 1;
+/** Temperature for triage/classification tools. */
+export declare const TRIAGE_TEMPERATURE: 1;

package/dist/lib/model-config.js

@@ -1,104 +1,71 @@
 /** Fast, cost-effective model for summarization and light analysis. */
-export const FLASH_MODEL = 'gemini-2.5-flash';
+export const FLASH_MODEL = 'gemini-3-flash-preview';
 /** High-capability model for deep reasoning, quality inspection, and reliable code generation. */
-export const PRO_MODEL = 'gemini-2.5-pro';
-const THINKING_BUDGET_TOKENS = {
-    /**
-     * Disabled (0): triage/classification tasks need no reasoning chain.
-     * Flash 2.5 range: 0–24_576; 0 explicitly disables thinking.
-     */
-    flashTriage: 0,
-    /**
-     * Raised from 8_192 → half of Flash max (24_576).
-     * Used for analysis tasks (test plans, complexity) that benefit from
-     * multi-step reasoning but not from unbounded thinking tokens.
-     */
-    flash: 16_384,
-    /**
-     * Raised from 16_384 → 75 % of Pro max (32_768).
-     * Gives deep-review and patch-generation tools genuine headroom for
-     * complex multi-file diffs without switching to cost-unpredictable dynamic.
-     */
-    pro: 24_576,
+export const PRO_MODEL = 'gemini-3-pro-preview';
+/** Default hint for auto-detection. */
+const DEFAULT_DETECT_HINT = 'detect';
+/** Default language hint. */
+export const DEFAULT_LANGUAGE = DEFAULT_DETECT_HINT;
+/** Default test-framework hint. */
+export const DEFAULT_FRAMEWORK = DEFAULT_DETECT_HINT;
+/** Extended timeout for Pro model calls (ms). */
+export const DEFAULT_TIMEOUT_PRO_MS = 120_000;
+export const MODEL_TIMEOUT_MS = {
+    defaultPro: DEFAULT_TIMEOUT_PRO_MS,
+};
+Object.freeze(MODEL_TIMEOUT_MS);
+// ---------------------------------------------------------------------------
+// Budgets (Thinking & Output)
+// ---------------------------------------------------------------------------
+const THINKING_LEVELS = {
+    /** Minimal thinking for triage/classification. */
+    flashTriage: 'minimal',
+    /** Medium thinking for analysis tasks. */
+    flash: 'medium',
+    /** High thinking for deep review and patches. */
+    pro: 'high',
 };
+// Thinking budget in tokens for Flash and Pro tools. Note that these are not hard limits, but rather guidelines to encourage concise responses and manage latency/cost.
 const OUTPUT_TOKEN_BUDGET = {
-    flashTriage: 4_096,
-    /**
-     * Raised from 4_096: 15 test cases × pseudoCode@2_000 chars ≈ 7_500 tokens;
-     * staying at 4_096 risked MAX_TOKENS truncation on moderate test plans.
-     */
-    flashTestPlan: 8_192,
     flashApiBreaking: 4_096,
     flashComplexity: 2_048,
-    /**
-     * Raised from 8_192: 25 findings × (title+explanation+recommendation) can
-     * exceed 8_192 tokens for rich, high-finding-count reviews.
-     */
-    proReview: 12_288,
-    /**
-     * Raised from 4_096: 10 search/replace blocks with multi-line code context
-     * can exceed the previous cap and cause MAX_TOKENS truncation.
-     */
+    flashTestPlan: 8_192,
+    flashTriage: 4_096,
     proPatch: 8_192,
+    proReview: 12_288,
 };
-/**
- * Per-task temperature presets for structured JSON generation.
- * These are intentionally low: the model is already heavily constrained by
- * the responseSchema, so lower temperatures improve schema-validation
- * pass-through rates and reduce hallucinated field values.
- */
-const TOOL_TEMPERATURE = {
-    /** Triage/classification tasks — deterministic structured extraction. */
-    triage: 0.1,
-    /** Analytical reasoning — consistent algorithmic analysis. */
-    analysis: 0.1,
-    /** Code patch generation — maximum precision for exact-match search blocks. */
-    patch: 0.0,
-    /** Test plan generation — allow modest diversity in test-case synthesis. */
-    creative: 0.2,
-};
-const DEFAULT_DETECT_HINT = 'detect';
-/**
- * Thinking budget (tokens) for Flash triage tools (impact, summary, API-breaking).
- * Explicitly disabled (0) — these are classification/extraction tasks that do not
- * benefit from a reasoning chain. Avoids default dynamic-thinking overhead.
- * Flash 2.5 range: 0–24_576.
- */
-export const FLASH_TRIAGE_THINKING_BUDGET = THINKING_BUDGET_TOKENS.flashTriage;
-/** Thinking budget (tokens) for Flash analysis tasks (test plans, complexity). */
-export const FLASH_THINKING_BUDGET = THINKING_BUDGET_TOKENS.flash;
-/** Thinking budget (tokens) for Pro model deep-analysis tasks (quality, patches). */
-export const PRO_THINKING_BUDGET = THINKING_BUDGET_TOKENS.pro;
-/** Output cap for Flash triage tools (impact, summary). */
-export const FLASH_TRIAGE_MAX_OUTPUT_TOKENS = OUTPUT_TOKEN_BUDGET.flashTriage;
-/** Output cap for API breaking-change detection (migration guidance needs room). */
+/** Thinking level for Flash triage. */
+export const FLASH_TRIAGE_THINKING_LEVEL = THINKING_LEVELS.flashTriage;
+/** Thinking level for Flash analysis. */
+export const FLASH_THINKING_LEVEL = THINKING_LEVELS.flash;
+/** Thinking level for Pro deep analysis. */
+export const PRO_THINKING_LEVEL = THINKING_LEVELS.pro;
+/** Output cap for Flash API breaking-change detection. */
 export const FLASH_API_BREAKING_MAX_OUTPUT_TOKENS = OUTPUT_TOKEN_BUDGET.flashApiBreaking;
-/** Output cap for test-plan generation (includes pseudocode snippets). */
+/** Output cap for Flash complexity analysis. */
+export const FLASH_COMPLEXITY_MAX_OUTPUT_TOKENS = OUTPUT_TOKEN_BUDGET.flashComplexity;
+/** Output cap for Flash test-plan generation. */
 export const FLASH_TEST_PLAN_MAX_OUTPUT_TOKENS = OUTPUT_TOKEN_BUDGET.flashTestPlan;
+/** Output cap for Flash triage tools. */
+export const FLASH_TRIAGE_MAX_OUTPUT_TOKENS = OUTPUT_TOKEN_BUDGET.flashTriage;
+/** Output cap for Pro patch generation. */
+export const PRO_PATCH_MAX_OUTPUT_TOKENS = OUTPUT_TOKEN_BUDGET.proPatch;
 /** Output cap for Pro deep review findings. */
 export const PRO_REVIEW_MAX_OUTPUT_TOKENS = OUTPUT_TOKEN_BUDGET.proReview;
-/** Output cap for Pro search/replace remediation blocks. */
-export const PRO_PATCH_MAX_OUTPUT_TOKENS = OUTPUT_TOKEN_BUDGET.proPatch;
-/** Output cap for Flash complexity analysis reports. */
-export const FLASH_COMPLEXITY_MAX_OUTPUT_TOKENS = OUTPUT_TOKEN_BUDGET.flashComplexity;
-/** Extended timeout for Pro model calls (ms). Pro thinks longer than Flash. */
-export const DEFAULT_TIMEOUT_PRO_MS = 120_000;
 // ---------------------------------------------------------------------------
-// Temperature presets — see TOOL_TEMPERATURE constant for rationale.
+// Temperatures
 // ---------------------------------------------------------------------------
-/** Temperature for triage/classification tools (deterministic structured extraction). */
-export const TRIAGE_TEMPERATURE = TOOL_TEMPERATURE.triage;
-/** Temperature for analytical tools (consistent algorithmic reasoning). */
+const TOOL_TEMPERATURE = {
+    analysis: 1.0, // Gemini 3 recommends 1.0 for all tasks
+    creative: 1.0, // Gemini 3 recommends 1.0 for all tasks
+    patch: 1.0, // Gemini 3 recommends 1.0 for all tasks
+    triage: 1.0, // Gemini 3 recommends 1.0 for all tasks
+};
+/** Temperature for analytical tools. */
 export const ANALYSIS_TEMPERATURE = TOOL_TEMPERATURE.analysis;
-/** Temperature for code patch generation (maximum precision for search blocks). */
-export const PATCH_TEMPERATURE = TOOL_TEMPERATURE.patch;
-/** Temperature for creative synthesis tools (test plan generation). */
+/** Temperature for creative synthesis (test plans). */
 export const CREATIVE_TEMPERATURE = TOOL_TEMPERATURE.creative;
-export const MODEL_TIMEOUT_MS = {
-    defaultPro: DEFAULT_TIMEOUT_PRO_MS,
-};
-Object.freeze(MODEL_TIMEOUT_MS);
-/** Default language hint when not specified by the user. Tells the model to auto-detect. */
-export const DEFAULT_LANGUAGE = DEFAULT_DETECT_HINT;
-/** Default test-framework hint when not specified by the user. Tells the model to auto-detect. */
-export const DEFAULT_FRAMEWORK = DEFAULT_DETECT_HINT;
+/** Temperature for code patch generation. */
+export const PATCH_TEMPERATURE = TOOL_TEMPERATURE.patch;
+/** Temperature for triage/classification tools. */
+export const TRIAGE_TEMPERATURE = TOOL_TEMPERATURE.triage;

package/dist/lib/tool-contracts.d.ts

@@ -13,7 +13,7 @@ export interface ToolContract {
     model: string;
     /** Set to 0 for synchronous (non-Gemini) tools. */
     timeoutMs: number;
-    thinkingBudget?: number;
+    thinkingLevel?: 'minimal' | 'low' | 'medium' | 'high';
     /** Set to 0 for synchronous (non-Gemini) tools. */
     maxOutputTokens: number;
     /**
@@ -23,6 +23,8 @@ export interface ToolContract {
      * Omit to use the global default (0.2).
      */
     temperature?: number;
+    /** Enables deterministic JSON guidance and schema key ordering. */
+    deterministicJson?: boolean;
     params: readonly ToolParameterContract[];
     outputShape: string;
     gotchas: readonly string[];
@@ -48,11 +50,12 @@ export declare const TOOL_CONTRACTS: readonly [{
 }, {
     readonly name: "analyze_pr_impact";
     readonly purpose: "Assess severity, categories, breaking changes, and rollback complexity.";
-    readonly model: "gemini-2.5-flash";
+    readonly model: "gemini-3-flash-preview";
     readonly timeoutMs: 90000;
-    readonly thinkingBudget: 0;
+    readonly thinkingLevel: "minimal";
     readonly maxOutputTokens: 4096;
-    readonly temperature: 0.1;
+    readonly temperature: 1;
+    readonly deterministicJson: true;
     readonly params: readonly [{
         readonly name: "repository";
         readonly type: "string";
@@ -72,11 +75,12 @@ export declare const TOOL_CONTRACTS: readonly [{
 }, {
     readonly name: "generate_review_summary";
     readonly purpose: "Produce PR summary, risk rating, and merge recommendation.";
-    readonly model: "gemini-2.5-flash";
+    readonly model: "gemini-3-flash-preview";
     readonly timeoutMs: 90000;
-    readonly thinkingBudget: 0;
+    readonly thinkingLevel: "minimal";
     readonly maxOutputTokens: 4096;
-    readonly temperature: 0.1;
+    readonly temperature: 1;
+    readonly deterministicJson: true;
     readonly params: readonly [{
         readonly name: "repository";
         readonly type: "string";
@@ -96,11 +100,12 @@ export declare const TOOL_CONTRACTS: readonly [{
 }, {
     readonly name: "inspect_code_quality";
     readonly purpose: "Deep code review with optional full-file context.";
-    readonly model: "gemini-2.5-pro";
+    readonly model: "gemini-3-pro-preview";
     readonly timeoutMs: 120000;
-    readonly thinkingBudget: 24576;
+    readonly thinkingLevel: "high";
     readonly maxOutputTokens: 12288;
-    readonly temperature: 0.1;
+    readonly temperature: 1;
+    readonly deterministicJson: true;
     readonly params: readonly [{
         readonly name: "repository";
         readonly type: "string";
@@ -139,11 +144,12 @@ export declare const TOOL_CONTRACTS: readonly [{
 }, {
     readonly name: "suggest_search_replace";
     readonly purpose: "Generate verbatim search/replace fix blocks for one finding.";
-    readonly model: "gemini-2.5-pro";
+    readonly model: "gemini-3-pro-preview";
     readonly timeoutMs: 120000;
-    readonly thinkingBudget: 24576;
+    readonly thinkingLevel: "high";
     readonly maxOutputTokens: 8192;
-    readonly temperature: 0;
+    readonly temperature: 1;
+    readonly deterministicJson: true;
     readonly params: readonly [{
         readonly name: "findingTitle";
         readonly type: "string";
@@ -164,11 +170,12 @@ export declare const TOOL_CONTRACTS: readonly [{
 }, {
     readonly name: "generate_test_plan";
     readonly purpose: "Generate prioritized test cases and coverage guidance.";
-    readonly model: "gemini-2.5-flash";
+    readonly model: "gemini-3-flash-preview";
     readonly timeoutMs: 90000;
-    readonly thinkingBudget: 16384;
+    readonly thinkingLevel: "medium";
     readonly maxOutputTokens: 8192;
-    readonly temperature: 0.2;
+    readonly temperature: 1;
+    readonly deterministicJson: true;
     readonly params: readonly [{
         readonly name: "repository";
         readonly type: "string";
@@ -200,11 +207,12 @@ export declare const TOOL_CONTRACTS: readonly [{
 }, {
     readonly name: "analyze_time_space_complexity";
     readonly purpose: "Analyze Big-O complexity and detect degradations in changed code.";
-    readonly model: "gemini-2.5-flash";
+    readonly model: "gemini-3-flash-preview";
     readonly timeoutMs: 90000;
-    readonly thinkingBudget: 16384;
+    readonly thinkingLevel: "medium";
     readonly maxOutputTokens: 2048;
-    readonly temperature: 0.1;
+    readonly temperature: 1;
+    readonly deterministicJson: true;
     readonly params: readonly [{
         readonly name: "language";
         readonly type: "string";
@@ -218,11 +226,12 @@ export declare const TOOL_CONTRACTS: readonly [{
 }, {
     readonly name: "detect_api_breaking_changes";
     readonly purpose: "Detect breaking API/interface changes in a diff.";
-    readonly model: "gemini-2.5-flash";
+    readonly model: "gemini-3-flash-preview";
     readonly timeoutMs: 90000;
-    readonly thinkingBudget: 0;
+    readonly thinkingLevel: "minimal";
     readonly maxOutputTokens: 4096;
-    readonly temperature: 0.1;
+    readonly temperature: 1;
+    readonly deterministicJson: true;
     readonly params: readonly [{
         readonly name: "language";
         readonly type: "string";

package/dist/lib/tool-contracts.js

@@ -1,4 +1,4 @@
-import { ANALYSIS_TEMPERATURE, CREATIVE_TEMPERATURE, DEFAULT_TIMEOUT_PRO_MS, FLASH_API_BREAKING_MAX_OUTPUT_TOKENS, FLASH_COMPLEXITY_MAX_OUTPUT_TOKENS, FLASH_MODEL, FLASH_TEST_PLAN_MAX_OUTPUT_TOKENS, FLASH_THINKING_BUDGET, FLASH_TRIAGE_MAX_OUTPUT_TOKENS, FLASH_TRIAGE_THINKING_BUDGET, PATCH_TEMPERATURE, PRO_MODEL, PRO_PATCH_MAX_OUTPUT_TOKENS, PRO_REVIEW_MAX_OUTPUT_TOKENS, PRO_THINKING_BUDGET, TRIAGE_TEMPERATURE, } from './model-config.js';
+import { ANALYSIS_TEMPERATURE, CREATIVE_TEMPERATURE, DEFAULT_TIMEOUT_PRO_MS, FLASH_API_BREAKING_MAX_OUTPUT_TOKENS, FLASH_COMPLEXITY_MAX_OUTPUT_TOKENS, FLASH_MODEL, FLASH_TEST_PLAN_MAX_OUTPUT_TOKENS, FLASH_THINKING_LEVEL, FLASH_TRIAGE_MAX_OUTPUT_TOKENS, FLASH_TRIAGE_THINKING_LEVEL, PATCH_TEMPERATURE, PRO_MODEL, PRO_PATCH_MAX_OUTPUT_TOKENS, PRO_REVIEW_MAX_OUTPUT_TOKENS, PRO_THINKING_LEVEL, TRIAGE_TEMPERATURE, } from './model-config.js';
 const DEFAULT_TIMEOUT_FLASH_MS = 90_000;
 export const INSPECTION_FOCUS_AREAS = [
     'security',
@@ -40,9 +40,10 @@ export const TOOL_CONTRACTS = [
         purpose: 'Assess severity, categories, breaking changes, and rollback complexity.',
         model: FLASH_MODEL,
         timeoutMs: DEFAULT_TIMEOUT_FLASH_MS,
-        thinkingBudget: FLASH_TRIAGE_THINKING_BUDGET,
+        thinkingLevel: FLASH_TRIAGE_THINKING_LEVEL,
         maxOutputTokens: FLASH_TRIAGE_MAX_OUTPUT_TOKENS,
         temperature: TRIAGE_TEMPERATURE,
+        deterministicJson: true,
         params: [
             {
                 name: 'repository',
@@ -73,9 +74,10 @@ export const TOOL_CONTRACTS = [
         purpose: 'Produce PR summary, risk rating, and merge recommendation.',
         model: FLASH_MODEL,
         timeoutMs: DEFAULT_TIMEOUT_FLASH_MS,
-        thinkingBudget: FLASH_TRIAGE_THINKING_BUDGET,
+        thinkingLevel: FLASH_TRIAGE_THINKING_LEVEL,
         maxOutputTokens: FLASH_TRIAGE_MAX_OUTPUT_TOKENS,
         temperature: TRIAGE_TEMPERATURE,
+        deterministicJson: true,
         params: [
             {
                 name: 'repository',
@@ -106,9 +108,10 @@ export const TOOL_CONTRACTS = [
         purpose: 'Deep code review with optional full-file context.',
         model: PRO_MODEL,
         timeoutMs: DEFAULT_TIMEOUT_PRO_MS,
-        thinkingBudget: PRO_THINKING_BUDGET,
+        thinkingLevel: PRO_THINKING_LEVEL,
         maxOutputTokens: PRO_REVIEW_MAX_OUTPUT_TOKENS,
         temperature: ANALYSIS_TEMPERATURE,
+        deterministicJson: true,
         params: [
             {
                 name: 'repository',
@@ -163,9 +166,10 @@ export const TOOL_CONTRACTS = [
         purpose: 'Generate verbatim search/replace fix blocks for one finding.',
         model: PRO_MODEL,
         timeoutMs: DEFAULT_TIMEOUT_PRO_MS,
-        thinkingBudget: PRO_THINKING_BUDGET,
+        thinkingLevel: PRO_THINKING_LEVEL,
         maxOutputTokens: PRO_PATCH_MAX_OUTPUT_TOKENS,
         temperature: PATCH_TEMPERATURE,
+        deterministicJson: true,
         params: [
             {
                 name: 'findingTitle',
@@ -198,9 +202,10 @@ export const TOOL_CONTRACTS = [
         purpose: 'Generate prioritized test cases and coverage guidance.',
         model: FLASH_MODEL,
         timeoutMs: DEFAULT_TIMEOUT_FLASH_MS,
-        thinkingBudget: FLASH_THINKING_BUDGET,
+        thinkingLevel: FLASH_THINKING_LEVEL,
         maxOutputTokens: FLASH_TEST_PLAN_MAX_OUTPUT_TOKENS,
         temperature: CREATIVE_TEMPERATURE,
+        deterministicJson: true,
         params: [
             {
                 name: 'repository',
@@ -245,9 +250,10 @@ export const TOOL_CONTRACTS = [
         purpose: 'Analyze Big-O complexity and detect degradations in changed code.',
         model: FLASH_MODEL,
         timeoutMs: DEFAULT_TIMEOUT_FLASH_MS,
-        thinkingBudget: FLASH_THINKING_BUDGET,
+        thinkingLevel: FLASH_THINKING_LEVEL,
         maxOutputTokens: FLASH_COMPLEXITY_MAX_OUTPUT_TOKENS,
         temperature: ANALYSIS_TEMPERATURE,
+        deterministicJson: true,
         params: [
             {
                 name: 'language',
@@ -269,9 +275,10 @@ export const TOOL_CONTRACTS = [
         purpose: 'Detect breaking API/interface changes in a diff.',
         model: FLASH_MODEL,
         timeoutMs: DEFAULT_TIMEOUT_FLASH_MS,
-        thinkingBudget: FLASH_TRIAGE_THINKING_BUDGET,
+        thinkingLevel: FLASH_TRIAGE_THINKING_LEVEL,
         maxOutputTokens: FLASH_API_BREAKING_MAX_OUTPUT_TOKENS,
         temperature: TRIAGE_TEMPERATURE,
+        deterministicJson: true,
         params: [
             {
                 name: 'language',

package/dist/lib/tool-factory.d.ts

@@ -55,10 +55,10 @@ export interface StructuredToolTaskConfig<TInput extends object = Record<string,
     transformResult?: (input: TInput, result: TResult, ctx: ToolExecutionContext) => TFinal;
     /** Optional validation hook for input parameters. */
     validateInput?: (input: TInput, ctx: ToolExecutionContext) => Promise<ReturnType<typeof createErrorToolResponse> | undefined> | ReturnType<typeof createErrorToolResponse> | undefined;
-    /** Optional Gemini model to use (e.g. 'gemini-2.5-pro'). */
+    /** Optional Gemini model to use (e.g. 'gemini-3-pro-preview'). */
     model?: string;
-    /** Optional thinking budget in tokens. */
-    thinkingBudget?: number;
+    /** Optional thinking level. */
+    thinkingLevel?: 'minimal' | 'low' | 'medium' | 'high';
     /** Optional timeout in ms for the Gemini call. Defaults to 90,000 ms. Use DEFAULT_TIMEOUT_PRO_MS for Pro model calls. */
     timeoutMs?: number;
     /** Optional max output tokens for Gemini. */
@@ -72,6 +72,10 @@ export interface StructuredToolTaskConfig<TInput extends object = Record<string,
     temperature?: number;
     /** Optional opt-in to Gemini thought output. Defaults to false. */
     includeThoughts?: boolean;
+    /** Optional deterministic JSON mode for stricter key ordering and repair prompting. */
+    deterministicJson?: boolean;
+    /** Optional batch execution mode. Defaults to runtime setting. */
+    batchMode?: 'off' | 'inline';
     /** Optional formatter for human-readable text output. */
     formatOutput?: (result: TFinal) => string;
     /** Optional context text used in progress messages. */
@@ -81,6 +85,7 @@ export interface StructuredToolTaskConfig<TInput extends object = Record<string,
     /** Builds the system instruction and user prompt from parsed tool input. */
     buildPrompt: (input: TInput, ctx: ToolExecutionContext) => PromptParts;
 }
+export declare function summarizeSchemaValidationErrorForRetry(errorMessage: string): string;
 export declare function wrapToolHandler<TInput, TResult extends CallToolResult>(options: {
     toolName: string;
     progressContext?: (input: TInput) => string;

package/dist/lib/tool-factory.js

@@ -1,6 +1,7 @@
 import { z } from 'zod';
 import { DefaultOutputSchema } from '../schemas/outputs.js';
 import { getDiff } from './diff-store.js';
+import { createCachedEnvInt } from './env-config.js';
 import { getErrorMessage, RETRYABLE_UPSTREAM_ERROR_PATTERN } from './errors.js';
 import { stripJsonSchemaConstraints } from './gemini-schema.js';
 import { generateStructuredJson, getCurrentRequestId } from './gemini.js';
@@ -14,6 +15,9 @@ const TIMEOUT_ERROR_PATTERN = /timed out|timeout/i;
 const BUDGET_ERROR_PATTERN = /exceeds limit|max allowed size|input too large/i;
 const BUSY_ERROR_PATTERN = /too many concurrent/i;
 const MAX_SCHEMA_RETRIES = 1;
+const DEFAULT_SCHEMA_RETRY_ERROR_CHARS = 1_500;
+const schemaRetryErrorCharsConfig = createCachedEnvInt('MAX_SCHEMA_RETRY_ERROR_CHARS', DEFAULT_SCHEMA_RETRY_ERROR_CHARS);
+const DETERMINISTIC_JSON_RETRY_NOTE = 'Deterministic JSON mode: keep key names exactly as schema-defined and preserve stable field ordering.';
 function createGeminiResponseSchema(config) {
     const sourceSchema = config.geminiSchema ?? config.resultSchema;
     return stripJsonSchemaConstraints(z.toJSONSchema(sourceSchema));
@@ -21,6 +25,35 @@ function createGeminiResponseSchema(config) {
 function parseToolInput(input, fullInputSchema) {
     return fullInputSchema.parse(input);
 }
+function extractResponseKeyOrdering(responseSchema) {
+    const schemaType = responseSchema.type;
+    if (schemaType !== 'object') {
+        return undefined;
+    }
+    const { properties } = responseSchema;
+    if (typeof properties !== 'object' || properties === null) {
+        return undefined;
+    }
+    return Object.keys(properties);
+}
+export function summarizeSchemaValidationErrorForRetry(errorMessage) {
+    const maxChars = Math.max(200, schemaRetryErrorCharsConfig.get());
+    const compact = errorMessage.replace(/\s+/g, ' ').trim();
+    if (compact.length <= maxChars) {
+        return compact;
+    }
+    return `${compact.slice(0, maxChars - 3)}...`;
+}
+function createSchemaRetryPrompt(prompt, errorMessage, deterministicJson) {
+    const summarizedError = summarizeSchemaValidationErrorForRetry(errorMessage);
+    const deterministicNote = deterministicJson
+        ? `\n${DETERMINISTIC_JSON_RETRY_NOTE}`
+        : '';
+    return {
+        summarizedError,
+        prompt: `${prompt}\n\nCRITICAL: The previous response failed schema validation. Error: ${summarizedError}${deterministicNote}`,
+    };
+}
 function createGenerationRequest(config, promptParts, responseSchema, onLog, signal) {
     const request = {
         systemInstruction: promptParts.systemInstruction,
@@ -31,8 +64,8 @@ function createGenerationRequest(config, promptParts, responseSchema, onLog, sig
     if (config.model !== undefined) {
         request.model = config.model;
     }
-    if (config.thinkingBudget !== undefined) {
-        request.thinkingBudget = config.thinkingBudget;
+    if (config.thinkingLevel !== undefined) {
+        request.thinkingLevel = config.thinkingLevel;
     }
     if (config.timeoutMs !== undefined) {
         request.timeoutMs = config.timeoutMs;
@@ -46,13 +79,23 @@ function createGenerationRequest(config, promptParts, responseSchema, onLog, sig
     if (config.includeThoughts !== undefined) {
         request.includeThoughts = config.includeThoughts;
     }
+    if (config.deterministicJson) {
+        const responseKeyOrdering = extractResponseKeyOrdering(responseSchema);
+        if (responseKeyOrdering !== undefined) {
+            request.responseKeyOrdering = responseKeyOrdering;
+        }
+    }
+    if (config.batchMode !== undefined) {
+        request.batchMode = config.batchMode;
+    }
     if (signal !== undefined) {
         request.signal = signal;
     }
     return request;
 }
+const VALIDATION_ERROR_PATTERN = /validation/i;
 function classifyErrorMeta(error, message) {
-    if (error instanceof z.ZodError || /validation/i.test(message)) {
+    if (error instanceof z.ZodError || VALIDATION_ERROR_PATTERN.test(message)) {
         return {
             kind: 'validation',
             retryable: false,
@@ -341,11 +384,16 @@ export function registerStructuredToolTask(server, config) {
                                 throw error;
                             }
                             const errorMessage = getErrorMessage(error);
+                            const schemaRetryPrompt = createSchemaRetryPrompt(prompt, errorMessage, config.deterministicJson === true);
                             await onLog('warning', {
                                 event: 'schema_validation_failed',
-                                details: { attempt, error: errorMessage },
+                                details: {
+                                    attempt,
+                                    error: schemaRetryPrompt.summarizedError,
+                                    originalChars: errorMessage.length,
+                                },
                             });
-                            retryPrompt = `${prompt}\n\nCRITICAL: The previous response failed schema validation. Error: ${errorMessage}`;
+                            retryPrompt = schemaRetryPrompt.prompt;
                         }
                     }
                     if (!parsed) {

package/dist/lib/types.d.ts

@@ -1,14 +1,20 @@
 export type JsonObject = Record<string, unknown>;
 export type GeminiLogHandler = (level: string, data: unknown) => Promise<void>;
+export interface GeminiFunctionCallingContext {
+    readonly modelParts: readonly unknown[];
+}
 export interface GeminiRequestExecutionOptions {
     maxRetries?: number;
     timeoutMs?: number;
     temperature?: number;
     maxOutputTokens?: number;
-    thinkingBudget?: number;
+    thinkingLevel?: 'minimal' | 'low' | 'medium' | 'high';
     includeThoughts?: boolean;
     signal?: AbortSignal;
     onLog?: GeminiLogHandler;
+    responseKeyOrdering?: readonly string[];
+    functionCallingContext?: GeminiFunctionCallingContext;
+    batchMode?: 'off' | 'inline';
 }
 export interface GeminiStructuredRequestOptions extends GeminiRequestExecutionOptions {
     model?: string;

package/dist/prompts/index.js

@@ -43,9 +43,9 @@ function getToolGuide(tool) {
     if (!contract) {
         return `Use \`${tool}\` to analyze your code changes.`;
     }
-    const { thinkingBudget } = contract;
-    const modelLine = thinkingBudget !== undefined
-        ? `Model: ${contract.model} (thinking budget ${thinkingBudget}, output cap ${contract.maxOutputTokens}).`
+    const { thinkingLevel } = contract;
+    const modelLine = thinkingLevel !== undefined
+        ? `Model: ${contract.model} (thinking level ${thinkingLevel}, output cap ${contract.maxOutputTokens}).`
         : `Model: ${contract.model} (output cap ${contract.maxOutputTokens}).`;
     return `Tool: ${contract.name}\n${modelLine}\nOutput: ${contract.outputShape}\nUse: ${contract.purpose}`;
 }

package/dist/resources/instructions.js

@@ -23,9 +23,9 @@ function formatToolSection(contract) {
 ${parameterLines.join('\n')}
 - Output shape: \`${contract.outputShape}\``;
     }
-    const thinkingLine = contract.thinkingBudget === undefined
-        ? '- Thinking budget: disabled'
-        : `- Thinking budget: ${contract.thinkingBudget}`;
+    const thinkingLine = contract.thinkingLevel === undefined
+        ? '- Thinking level: disabled'
+        : `- Thinking level: ${contract.thinkingLevel}`;
     return `### \`${contract.name}\`
 - Purpose: ${contract.purpose}
 - Model: \`${contract.model}\`