agentic-qe 3.14.3 → 3.14.5

package/dist/shared/llm/model-registry.js

@@ -104,8 +104,8 @@ export const MODEL_REGISTRY = {
             supportsCodeExecution: false,
         },
         cost: {
-            inputCostPerMillion: 15.0,
-            outputCostPerMillion: 75.0,
+            inputCostPerMillion: 5.0,
+            outputCostPerMillion: 25.0,
         },
     },
     'claude-haiku-3-5': {
@@ -163,6 +163,33 @@ export const MODEL_REGISTRY = {
     // ==========================================================================
     // ADR-093: Opus 4.7, Sonnet 4.6, Haiku 4.5 (Apr 2026 cluster)
     // ==========================================================================
+    'claude-opus-4-6': {
+        name: 'Claude Opus 4.6',
+        family: 'claude',
+        tier: 'flagship',
+        description: 'Previous-generation Opus, still served. Accepts sampling params; no xhigh effort.',
+        recommended: false,
+        capabilities: {
+            contextLength: 1_000_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: true,
+            supportsMCP: true,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsAdaptiveThinking: true,
+            supportsEffortXHigh: false,
+            tokenizerVersion: 'legacy',
+        },
+        cost: {
+            inputCostPerMillion: 5.0,
+            outputCostPerMillion: 25.0,
+        },
+    },
     'claude-opus-4-7': {
         name: 'Claude Opus 4.7',
         family: 'claude',
@@ -186,6 +213,7 @@ export const MODEL_REGISTRY = {
             supportsAdaptiveThinking: true,
             supportsEffortXHigh: true,
             tokenizerVersion: 'opus-4-7',
+            supportsSamplingParams: false,
         },
         cost: {
             inputCostPerMillion: 5.0,
@@ -196,13 +224,13 @@ export const MODEL_REGISTRY = {
         name: 'Claude Sonnet 4.6',
         family: 'claude',
         tier: 'standard',
-        description: 'Steady-state agentic-coding workhorse (2026). Standard 200k context, same pricing as Sonnet 4. ' +
-            'ADR-093 Tier 3 default for the AQE fleet.',
+        description: 'Previous-generation workhorse (ADR-093 tier 3 until the 2026-09 refresh). ' +
+            '1M context, still served; accepts sampling params.',
         releaseDate: '2026-02-01',
         recommended: true,
         capabilities: {
-            contextLength: 200000,
-            maxOutputTokens: 64000,
+            contextLength: 1_000_000,
+            maxOutputTokens: 128000,
             supportsTools: true,
             supportsStreaming: true,
             supportsVision: true,
@@ -230,7 +258,7 @@ export const MODEL_REGISTRY = {
         recommended: true,
         capabilities: {
             contextLength: 200000,
-            maxOutputTokens: 16384,
+            maxOutputTokens: 64000,
             supportsTools: true,
             supportsStreaming: true,
             supportsVision: true,
@@ -250,8 +278,259 @@ export const MODEL_REGISTRY = {
         },
     },
     // ==========================================================================
+    // Claude 5 generation (verified against GET /v1/models + OpenRouter pricing,
+    // 2026-09-27). All reject sampling params; effort goes in output_config.
+    // ==========================================================================
+    'claude-sonnet-5': {
+        name: 'Claude Sonnet 5',
+        family: 'claude',
+        tier: 'standard',
+        description: 'AQE workhorse (tiers 2-3). Adaptive thinking runs by default.',
+        recommended: true,
+        capabilities: {
+            contextLength: 1_000_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: true, // adaptive thinking
+            supportsMCP: true,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsAdaptiveThinking: true,
+            supportsEffortXHigh: true,
+            tokenizerVersion: 'opus-4-7',
+            supportsSamplingParams: false,
+        },
+        cost: {
+            inputCostPerMillion: 2.0,
+            outputCostPerMillion: 10.0,
+        },
+    },
+    'claude-opus-5': {
+        name: 'Claude Opus 5',
+        family: 'claude',
+        tier: 'flagship',
+        description: 'Previous Opus (tier 4 before Opus 5.5). Thinking is on by default when omitted.',
+        recommended: true,
+        capabilities: {
+            contextLength: 1_000_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: true, // adaptive thinking
+            supportsMCP: true,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsAdaptiveThinking: true,
+            supportsEffortXHigh: true,
+            tokenizerVersion: 'opus-4-7',
+            supportsSamplingParams: false,
+        },
+        cost: {
+            inputCostPerMillion: 5.0,
+            outputCostPerMillion: 25.0,
+        },
+    },
+    'claude-opus-4-8': {
+        name: 'Claude Opus 4.8',
+        family: 'claude',
+        tier: 'flagship',
+        description: 'Previous-generation Opus; same request surface as Opus 4.7.',
+        recommended: false,
+        capabilities: {
+            contextLength: 1_000_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: true, // adaptive thinking
+            supportsMCP: true,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsAdaptiveThinking: true,
+            supportsEffortXHigh: true,
+            tokenizerVersion: 'opus-4-7',
+            supportsSamplingParams: false,
+        },
+        cost: {
+            inputCostPerMillion: 5.0,
+            outputCostPerMillion: 25.0,
+        },
+    },
+    'claude-opus-5-5': {
+        name: 'Claude Opus 5.5',
+        family: 'claude',
+        tier: 'flagship',
+        description: 'AQE tier 4 flagship. Thinking cannot be disabled; default effort medium; forced tool_choice rejected.',
+        releaseDate: '2026-09-22',
+        recommended: true,
+        capabilities: {
+            contextLength: 1_000_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: true, // adaptive thinking
+            supportsMCP: true,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsAdaptiveThinking: true,
+            supportsEffortXHigh: true,
+            tokenizerVersion: 'opus-4-7',
+            supportsSamplingParams: false,
+        },
+        cost: {
+            inputCostPerMillion: 4.0,
+            outputCostPerMillion: 20.0,
+        },
+    },
+    'claude-fable-5': {
+        name: 'Claude Fable 5',
+        family: 'claude',
+        tier: 'flagship',
+        description: 'Opt-in most-capable tier (superseded by Fable 5.1).',
+        recommended: false,
+        capabilities: {
+            contextLength: 1_000_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: true, // adaptive thinking
+            supportsMCP: true,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsAdaptiveThinking: true,
+            supportsEffortXHigh: true,
+            tokenizerVersion: 'opus-4-7',
+            supportsSamplingParams: false,
+        },
+        cost: {
+            inputCostPerMillion: 10.0,
+            outputCostPerMillion: 50.0,
+        },
+    },
+    'claude-fable-5-1': {
+        name: 'Claude Fable 5.1',
+        family: 'claude',
+        tier: 'flagship',
+        description: 'Opt-in most-capable model; 2x Opus pricing, thinking always on.',
+        recommended: false,
+        capabilities: {
+            contextLength: 1_000_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: true, // adaptive thinking
+            supportsMCP: true,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsAdaptiveThinking: true,
+            supportsEffortXHigh: true,
+            tokenizerVersion: 'opus-4-7',
+            supportsSamplingParams: false,
+        },
+        cost: {
+            inputCostPerMillion: 10.0,
+            outputCostPerMillion: 50.0,
+        },
+    },
+    // ==========================================================================
     // OpenAI Models
     // ==========================================================================
+    'gpt-6-astra': {
+        name: 'GPT-6 Astra',
+        family: 'gpt',
+        tier: 'flagship',
+        description: 'OpenAI flagship (verified in /v1/models 2026-09-27)',
+        recommended: false,
+        capabilities: {
+            contextLength: 1_050_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: false,
+            supportsMCP: false,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsSamplingParams: false,
+            usesMaxCompletionTokens: true,
+        },
+        cost: {
+            inputCostPerMillion: 10.0,
+            outputCostPerMillion: 50.0,
+        },
+    },
+    'gpt-6-sol': {
+        name: 'GPT-6 Sol',
+        family: 'gpt',
+        tier: 'standard',
+        description: 'OpenAI default for capable tasks',
+        recommended: true,
+        capabilities: {
+            contextLength: 1_050_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: false,
+            supportsMCP: false,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsSamplingParams: false,
+            usesMaxCompletionTokens: true,
+        },
+        cost: {
+            inputCostPerMillion: 2.0,
+            outputCostPerMillion: 10.0,
+        },
+    },
+    'gpt-6-luna': {
+        name: 'GPT-6 Luna',
+        family: 'gpt',
+        tier: 'economy',
+        description: 'OpenAI low-cost tier',
+        recommended: true,
+        capabilities: {
+            contextLength: 1_050_000,
+            maxOutputTokens: 128000,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: false,
+            supportsMCP: false,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+            supportsSamplingParams: false,
+            usesMaxCompletionTokens: true,
+        },
+        cost: {
+            inputCostPerMillion: 0.1,
+            outputCostPerMillion: 0.5,
+        },
+    },
     'gpt-4o': {
         name: 'GPT-4o',
         family: 'gpt',
@@ -408,6 +687,78 @@ export const MODEL_REGISTRY = {
     // ==========================================================================
     // Google Gemini Models
     // ==========================================================================
+    'gemini-3.8-flash': {
+        name: 'Gemini 3.8 Flash',
+        family: 'gemini',
+        tier: 'economy',
+        description: 'Latest Gemini Flash (OpenRouter catalog 2026-09-27)',
+        recommended: true,
+        capabilities: {
+            contextLength: 1_048_576,
+            maxOutputTokens: 65536,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: false,
+            supportsMCP: false,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+        },
+        cost: {
+            inputCostPerMillion: 0.75,
+            outputCostPerMillion: 3.75,
+        },
+    },
+    'gemini-3.5-flash': {
+        name: 'Gemini 3.5 Flash',
+        family: 'gemini',
+        tier: 'standard',
+        description: 'Gemini 3.5 Flash',
+        recommended: false,
+        capabilities: {
+            contextLength: 1_048_576,
+            maxOutputTokens: 65536,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: false,
+            supportsMCP: false,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+        },
+        cost: {
+            inputCostPerMillion: 1.5,
+            outputCostPerMillion: 9.0,
+        },
+    },
+    'gemini-3.1-pro-preview': {
+        name: 'Gemini 3.1 Pro (preview)',
+        family: 'gemini',
+        tier: 'premium',
+        description: 'Gemini 3.1 Pro preview',
+        recommended: false,
+        capabilities: {
+            contextLength: 1_048_576,
+            maxOutputTokens: 65536,
+            supportsTools: true,
+            supportsStreaming: true,
+            supportsVision: true,
+            supportsJsonMode: true,
+            supportsSystemPrompt: true,
+            supportsExtendedThinking: false,
+            supportsMCP: false,
+            supportsEmbeddings: false,
+            supportsCodeExecution: false,
+        },
+        cost: {
+            inputCostPerMillion: 2.0,
+            outputCostPerMillion: 12.0,
+        },
+    },
     'gemini-pro': {
         name: 'Gemini Pro',
         family: 'gemini',
@@ -1199,23 +1550,43 @@ export function compareModels(modelA, modelB) {
 // ADR-093: Central Model Constants
 // ============================================================================
 /**
- * ADR-093 Tier 3 default Sonnet for the AQE fleet.
- * Standard 200k context (not 1M). Used as the workhorse model across routing,
- * advisor fallback, and provider defaults. Replaces hardcoded 'claude-sonnet-4-20250514'
- * in 25+ call sites.
+ * Default Sonnet for the AQE fleet (routing tiers 2-3, provider defaults,
+ * advisor fallback). Claude Sonnet 5 since the 2026-09 model refresh
+ * (was claude-sonnet-4-6 under ADR-093).
  */
-export const DEFAULT_SONNET_MODEL = 'claude-sonnet-4-6';
+export const DEFAULT_SONNET_MODEL = 'claude-sonnet-5';
 /**
- * ADR-093 Opus escalation target. Reserved for the ADR-092 MultiModelExecutor
- * when TinyDancer sets `triggerMultiModel=true` and the task warrants maximum
- * capability. Not the Tier 3 default — that remains Sonnet 4.6 per user
- * direction 2026-04-17.
+ * Opus escalation target (routing tier 4, ADR-092 MultiModelExecutor).
+ * Claude Opus 5.5 since the 2026-09 model refresh (was claude-opus-4-7):
+ * cheaper than Opus 5 ($4/$20 vs $5/$25). Its thinking cannot be disabled,
+ * forced tool_choice (`any`/`tool`) returns 400, and its default effort is
+ * `medium`. Fable 5.1 is registered but opt-in only.
  */
-export const DEFAULT_OPUS_MODEL = 'claude-opus-4-7';
+export const DEFAULT_OPUS_MODEL = 'claude-opus-5-5';
 /**
- * ADR-093 economy model. Replaces retired `claude-3-5-haiku-20241022`.
+ * Economy model (routing tier 1). Replaces retired `claude-3-5-haiku-20241022`.
  */
 export const DEFAULT_HAIKU_MODEL = 'claude-haiku-4-5';
+/**
+ * Canonical Claude model per AQE routing tier. Single source of truth for the
+ * tier maps in the router, task executor, and domain services — bump models
+ * here, not at each call site.
+ *
+ * Tier 0 is the Agent Booster (WASM) path and has no Claude model.
+ */
+export const CLAUDE_TIER_MODELS = {
+    1: 'claude-haiku-4-5-20251001',
+    2: DEFAULT_SONNET_MODEL,
+    3: DEFAULT_SONNET_MODEL, // extended-thinking tier, same model
+    4: DEFAULT_OPUS_MODEL,
+};
+/**
+ * Claude model for an AQE routing tier (1-4). Out-of-range tiers fall back to
+ * the tier 2/3 workhorse.
+ */
+export function getClaudeModelForTier(tier) {
+    return CLAUDE_TIER_MODELS[tier] ?? DEFAULT_SONNET_MODEL;
+}
 /**
  * ADR-093 registry of retiring/retired model IDs with their retirement dates.
  * Used by lints and ops tooling to prevent re-introduction of deprecated IDs.
@@ -1227,5 +1598,9 @@ export const RETIRING_MODELS = {
     'claude-opus-4-20250514': '2026-06-15',
     'claude-3-5-haiku-20241022': '2026-02-19',
     'claude-3-7-sonnet-20250219': '2026-02-19',
+    'claude-opus-4-1-20250805': '2026-08-05',
+    'claude-3-haiku-20240307': '2026-04-20',
+    'claude-3-5-sonnet-20241022': '2025-10-28',
+    'claude-3-5-sonnet-20240620': '2025-10-28',
 };
 //# sourceMappingURL=model-registry.js.map

package/dist/shared/llm/openai-params.d.ts

@@ -0,0 +1,35 @@
+/**
+ * Agentic QE v3 - OpenAI Chat Completions parameter policy
+ *
+ * GPT-5+/GPT-6 and the o-series reject request fields the GPT-4 generation
+ * accepted: `max_tokens` returns 400 ("Use 'max_completion_tokens' instead")
+ * and any `temperature` other than the default 1 returns 400 ("Only the
+ * default (1) value is supported"). Verified live against gpt-6-sol and
+ * gpt-6-luna on 2026-09-27; gpt-4o still accepts the legacy shape. Every
+ * OpenAI-format request builder applies this policy so a model bump cannot
+ * silently break calls.
+ */
+export interface OpenAIParamPolicy {
+    /** Whether `temperature` (and other sampling params) may be sent. */
+    sendSampling: boolean;
+    /** Which field carries the output-token limit. */
+    maxTokensField: 'max_tokens' | 'max_completion_tokens';
+}
+/**
+ * Parameter policy for a model ID (canonical, `openai/…`, or Azure form).
+ * Registry capability flags win; models missing from the registry fall back
+ * to a name heuristic — GPT-5/GPT-6/o-series get the reasoning shape,
+ * everything else keeps the legacy `max_tokens` + `temperature` shape.
+ */
+export declare function getOpenAIParamPolicy(model: string): OpenAIParamPolicy;
+export interface OpenAIParamInput {
+    maxTokens?: number;
+    temperature?: number;
+}
+/**
+ * Add the output-token limit and sampling fields to an OpenAI Chat
+ * Completions request body according to the model's policy. Mutates and
+ * returns `body`.
+ */
+export declare function applyOpenAIParams<T extends object>(body: T, model: string, input: OpenAIParamInput): T;
+//# sourceMappingURL=openai-params.d.ts.map

package/dist/shared/llm/openai-params.js

@@ -0,0 +1,54 @@
+/**
+ * Agentic QE v3 - OpenAI Chat Completions parameter policy
+ *
+ * GPT-5+/GPT-6 and the o-series reject request fields the GPT-4 generation
+ * accepted: `max_tokens` returns 400 ("Use 'max_completion_tokens' instead")
+ * and any `temperature` other than the default 1 returns 400 ("Only the
+ * default (1) value is supported"). Verified live against gpt-6-sol and
+ * gpt-6-luna on 2026-09-27; gpt-4o still accepts the legacy shape. Every
+ * OpenAI-format request builder applies this policy so a model bump cannot
+ * silently break calls.
+ */
+import { getModelCapabilities } from './model-registry.js';
+/**
+ * Model families that use the reasoning-model request shape, for IDs the
+ * registry does not know (optionally `openai/`-prefixed).
+ */
+const REASONING_MODEL_PATTERN = /^(openai\/)?(gpt-5|gpt-6|o\d)/i;
+const LEGACY_POLICY = { sendSampling: true, maxTokensField: 'max_tokens' };
+const REASONING_POLICY = { sendSampling: false, maxTokensField: 'max_completion_tokens' };
+/**
+ * Parameter policy for a model ID (canonical, `openai/…`, or Azure form).
+ * Registry capability flags win; models missing from the registry fall back
+ * to a name heuristic — GPT-5/GPT-6/o-series get the reasoning shape,
+ * everything else keeps the legacy `max_tokens` + `temperature` shape.
+ */
+export function getOpenAIParamPolicy(model) {
+    try {
+        const caps = getModelCapabilities(model);
+        return {
+            sendSampling: caps.supportsSamplingParams !== false,
+            maxTokensField: caps.usesMaxCompletionTokens === true ? 'max_completion_tokens' : 'max_tokens',
+        };
+    }
+    catch {
+        return REASONING_MODEL_PATTERN.test(model) ? { ...REASONING_POLICY } : { ...LEGACY_POLICY };
+    }
+}
+/**
+ * Add the output-token limit and sampling fields to an OpenAI Chat
+ * Completions request body according to the model's policy. Mutates and
+ * returns `body`.
+ */
+export function applyOpenAIParams(body, model, input) {
+    const policy = getOpenAIParamPolicy(model);
+    const target = body;
+    if (input.maxTokens !== undefined) {
+        target[policy.maxTokensField] = input.maxTokens;
+    }
+    if (policy.sendSampling && input.temperature !== undefined) {
+        target.temperature = input.temperature;
+    }
+    return body;
+}
+//# sourceMappingURL=openai-params.js.map

package/dist/shared/llm/provider-manager.js

@@ -27,6 +27,7 @@ import { GeminiProvider } from './providers/gemini.js';
 import { AzureOpenAIProvider } from './providers/azure-openai.js';
 import { BedrockProvider } from './providers/bedrock.js';
 import { toError } from '../error-utils.js';
+import { DEFAULT_SONNET_MODEL } from './model-registry.js';
 import { createRegisteredProvider } from './provider-registry.js';
 /**
  * Default provider manager configuration
@@ -682,17 +683,17 @@ export function createQEProviderManager() {
         loadBalancing: 'least-cost',
         providers: {
             claude: {
-                model: 'claude-sonnet-4-6',
+                model: DEFAULT_SONNET_MODEL,
                 maxTokens: 8192,
                 temperature: 0.3, // Lower for QE tasks
             },
             openai: {
-                model: 'gpt-4o',
+                model: 'gpt-6-sol',
                 maxTokens: 8192,
                 temperature: 0.3,
             },
             ollama: {
-                model: 'llama3.1',
+                model: 'qwen3-coder:30b',
                 maxTokens: 4096,
                 temperature: 0.3,
             },

package/dist/shared/llm/providers/azure-openai.d.ts

@@ -58,9 +58,20 @@ export declare class AzureOpenAIProvider implements LLMProvider {
     readonly name: string;
     private config;
     private requestId;
+    /** Model the caller configured explicitly (not the default), for parameter policy. */
+    private readonly explicitModel?;
     constructor(config: Partial<AzureOpenAIConfig> & {
         deploymentId: string;
     });
+    /**
+     * Model ID used to pick the request-parameter policy. The deployment decides
+     * which model actually runs, so the default `model` is not trusted here: an
+     * explicit model wins, otherwise the deployment name is looked up. A
+     * deployment name that hides the base model (e.g. 'prod-chat') falls back to
+     * the name heuristic in openai-params, i.e. the legacy max_tokens shape —
+     * GPT-5/6 deployments with opaque names must therefore set `model`.
+     */
+    private paramPolicyModel;
     /**
      * Check if Azure OpenAI is available and configured
      */