qlogicagent 2.20.8 → 2.20.10

package/dist/types/agent/memory-recall-injection.d.ts

@@ -20,6 +20,8 @@ export interface RecallMemoriesDeps {
     turnId: string;
     /** Last user message (already truncated by the caller). */
     query: string;
+    /** Turn lifetime. Aborted recall must not be reported or injected as consumed. */
+    signal?: AbortSignal;
 }
 /**
  * Recall long-term memories for this turn and build the system-message content

package/dist/types/agent/tool-loop/loop-helpers.d.ts

@@ -1,7 +1,7 @@
 import type { ChatMessage, ToolCallExecutionEvidence } from "../types.js";
 export { isEmptyToolResult } from "./tool-result-classification.js";
 export declare const OUTPUT_REQUEST_CONTEXT_PATTERN: RegExp;
-export declare function findLastToolError(messages: unknown[], _toolName: string): string | undefined;
+export declare function findLastToolError(messages: unknown[], toolName: string, toolArguments?: string): string | undefined;
 export declare function looksLikeBuildRequest(messages?: readonly ChatMessage[]): boolean;
 export declare function isCapabilityDenial(text: string): boolean;
 export declare function resolveToolLoopBudget(requested?: number, messages?: readonly ChatMessage[]): number;

package/dist/types/cli/agent-runtime-session-state.d.ts

@@ -13,7 +13,10 @@ export interface SessionRuntimeState {
     hooks: HookRegistry;
     permission: PermissionState;
 }
+export declare const MCP_TOOL_SNAPSHOT_TTL_MS = 30000;
 export declare class AgentRuntimeSessionState {
+    private readonly now;
+    private readonly mcpToolSnapshotTtlMs;
     private readonly runtimeBySession;
     private readonly permissionLeaseCountBySession;
     private readonly retiredPermissionUnregisterBySession;
@@ -23,6 +26,7 @@ export declare class AgentRuntimeSessionState {
     memoryProvider: MemoryHandlerProvider | null;
     memoryDreamProvider: MemoryDreamRuntimeProvider | null;
     memoryUserId: string;
+    constructor(now?: () => number, mcpToolSnapshotTtlMs?: number);
     getPluginSkills(): ReturnType<PluginLoader["getPluginSkills"]>;
     applyBootstrap(sessionId: string, result: AgentRuntimeBootstrapResult): void;
     withPermissionLease<T>(sessionId: string, action: () => Promise<T>): Promise<T>;

package/dist/types/contracts/tool-execution-evidence.d.ts

@@ -1,8 +1,10 @@
 import type { FinalAnswerEvidence, ToolCallExecutionEvidence } from "../protocol/wire/notification-payloads.js";
+export declare function isFreshEvidenceToolCall(call: Pick<ToolCallExecutionEvidence, "identity">): boolean;
+export declare function hasSuccessfulFreshToolEvidence(toolCalls: readonly ToolCallExecutionEvidence[]): boolean;
 export declare function classifyToolFailureKind(result: {
     blocked?: boolean;
     blockReason?: string;
     error?: string;
     details?: Record<string, unknown>;
 }): ToolCallExecutionEvidence["failureKind"];
-export declare function buildFinalAnswerEvidence(freshEvidenceRequired: boolean, toolCalls: readonly ToolCallExecutionEvidence[]): FinalAnswerEvidence;
+export declare function buildFinalAnswerEvidence(freshEvidenceRequired: boolean, toolCalls: readonly ToolCallExecutionEvidence[], requiredSuccessfulLogicalToolNames?: readonly string[]): FinalAnswerEvidence;

package/dist/types/runtime/execution/tool-result-storage.d.ts

@@ -61,6 +61,7 @@ type SkillReadinessSet = {
 };
 type PendingSkillReadinessSet = SkillReadinessSet & {
     readPaths: Set<string>;
+    readCallIds: Set<string>;
     nextSets: SkillReadinessSet[];
 };
 type SkillReadinessActivation = {
@@ -73,6 +74,15 @@ export type PersistedToolResult = {
     preview: string;
     hasMore: boolean;
 };
+export type CompleteToolResultEvidence = {
+    schemaVersion: 1;
+    complete: true;
+    storage: "session-tool-result";
+    artifactPath: string;
+    contentSha256: string;
+    contentBytes: number;
+    contentChars: number;
+};
 export declare function createContentReplacementState(): ContentReplacementState;
 export declare function toolResultContextPolicy(details: Record<string, unknown> | undefined): ToolResultContextPolicy;
 export declare function toolResultContextBudget(policy: ToolResultContextPolicy): ToolResultContextBudget;
@@ -93,6 +103,7 @@ export declare function generatePreview(content: string, maxBytes: number): {
     hasMore: boolean;
 };
 export declare function persistToolResult(content: string, toolCallId: string, sessionId: string, projectRoot: string | undefined): Promise<PersistedToolResult | null>;
+export declare function persistCompleteToolResultEvidence(content: string, toolCallId: string, sessionId: string, projectRoot: string | undefined): Promise<CompleteToolResultEvidence | undefined>;
 /**
  * Build a reference message for large tool results (CC buildLargeToolResultMessage parity).
  */

package/dist/types/runtime/prompt/capability-routing-policy.d.ts

@@ -1,5 +1,4 @@
-import type { FreshWorkspaceEvidencePolicy } from "./fresh-workspace-evidence.js";
-export type CapabilityRoutingOwner = "structured-planner" | "direct-response" | "mcp-capability" | "typed-builtin" | "pinned-skill" | "host-skill-discovery";
+export type CapabilityRoutingOwner = "structured-planner" | "direct-response" | "mcp-capability" | "pinned-skill" | "host-skill-discovery";
 /**
  * Resolves exclusive capability ownership before any semantic Skill classification.
  * A lower-specificity route must never steal a turn already owned by a typed policy.
@@ -8,6 +7,5 @@ export declare function resolveCapabilityRoutingOwner(input: {
     structuredPlannerTurn: boolean;
     directResponse: boolean;
     mcpCapabilitySelection: boolean;
-    freshEvidencePolicy: FreshWorkspaceEvidencePolicy;
     pinnedSkillCount: number;
 }): CapabilityRoutingOwner;

package/dist/types/runtime/prompt/task-domain.d.ts

@@ -1,43 +1,20 @@
 /**
- * Task Domain Detection — selects which system prompt guidance is injected.
- *
- * qlogicagent handles coding, office, AND creative tasks; a coding-only prompt
- * hurts non-coding performance, so the per-domain CONTENT blocks are kept.
- *
- * Resolution is STATELESS and recomputed every turn (CC-aligned — no hidden
- * session/disk state, no stickiness latch, no auto-persisted guesses):
- *   1. Host override: `config.taskDomain` from thread.turn params (explicit).
- *   2. Auto-detect: keyword match on the current user message; "general" if ambiguous.
- *
- * The universal BEHAVIORAL rules (answer-directly, tools-not-default, conciseness)
- * live in universalGuidance() and apply in every domain, so misclassifying a turn
- * as "general" degrades only the domain-tailored content, never the core behavior.
- */
-/**
- * Task domain determines which system prompt sections are injected.
- *
- * - "coding" — software engineering: bug fixes, refactoring, code gen, etc.
- * - "office" — data analysis, reports, documents, spreadsheets, emails
- * - "creative" — writing, copywriting, marketing, storytelling, brainstorming
- * - "general" — undetectable or mixed domain, uses universal guidance
+ * Task-domain guidance is a Host-owned typed choice, not a language classifier.
+ * User text is intentionally not inspected for keywords or regular expressions.
  */
 export type TaskDomain = "coding" | "office" | "creative" | "general";
 /**
- * Detect task domain from user message text. Keyword matching — fast, no LLM call.
+ * Kept as a compatibility-neutral pure function for callers that only have text.
+ * Without a typed Host decision the safe domain is always general.
  */
-export declare function detectTaskDomain(userText: string): TaskDomain;
+export declare function detectTaskDomain(_userText: string): TaskDomain;
 export interface DomainResolutionContext {
-    /** Host-provided override from config.taskDomain (explicit, per-request). */
+    /** Host-provided typed override from thread.turn params. */
     hostOverride?: TaskDomain;
-    /** Current turn's user message text. */
+    /** Current user text is accepted for API stability but never classified. */
     userText: string;
 }
-/**
- * Resolve task domain, stateless and per-turn: explicit host override, else fresh
- * keyword auto-detect on this turn's message. No project file, no session stickiness,
- * no disk persistence.
- */
 export declare function resolveTaskDomain(ctx: DomainResolutionContext): {
     domain: TaskDomain;
-    source: "host-override" | "auto-detect";
+    source: "host-override" | "default";
 };

package/dist/types/skills/mcp/mcp-manager.d.ts

@@ -61,6 +61,7 @@ export interface McpManagerConfig {
     toolCatalog?: ToolRegistryLike;
 }
 export declare function resolveMcpCallTimeoutMs(value: number | undefined): number;
+export declare function withMcpDeadline<T>(operation: Promise<T>, timeoutMs: number, label: string): Promise<T>;
 /**
  * Run `attempt` up to `options.attempts` times, backing off `baseDelayMs * n`
  * between tries. Returns the first success; rethrows the last error if every

package/package.json

@@ -1,6 +1,7 @@
 {
   "name": "qlogicagent",
-  "version": "2.20.8",
+  "version": "2.20.10",
+  "packageManager": "pnpm@10.23.0",
   "description": "XiaozhiClaw Agent CLI — subprocess architecture (JSON-RPC over stdio)",
   "type": "module",
   "main": "dist/index.js",
@@ -140,6 +141,10 @@
     "test:watch": "vitest",
     "benchmark:hermes-superiority": "tsx benchmarks/hermes-superiority/runner.ts",
     "benchmark:product-blackbox": "tsx benchmarks/product-blackbox/runner.ts",
+    "benchmark:product-blind-preflight": "tsx benchmarks/product-blackbox/runner.ts blind-preflight",
+    "benchmark:product-blind-calibration": "tsx benchmarks/product-blackbox/runner.ts prepare-blind-calibration",
+    "benchmark:product-blind-score": "tsx benchmarks/product-blackbox/runner.ts score-blind-calibration",
+    "benchmark:capability-state-fixtures": "tsx benchmarks/product-blackbox/run-capability-state-fixtures.ts",
     "lint": "oxlint .",
     "check:encoding-corruption": "node scripts/check-encoding-corruption.mjs",
     "check:mechanical-rules": "node scripts/check-mechanical-rules.mjs",
@@ -160,13 +165,13 @@
     "release": "node scripts/release.mjs"
   },
   "engines": {
-    "node": ">=22.0.0"
+    "node": ">=24.0.0 <25"
   },
   "dependencies": {
     "@agentclientprotocol/sdk": "^0.25.0",
     "@napi-rs/canvas": "0.1.100",
-    "@xiaozhiclaw/module-sdk": "0.2.4",
-    "@xiaozhiclaw/provider-core": "^0.1.30",
+    "@xiaozhiclaw/module-sdk": "0.2.5",
+    "@xiaozhiclaw/provider-core": "0.1.31",
     "better-sqlite3": "^12.10.0",
     "dotenv": "^17.3.1",
     "ioredis": "^5.11.1",

package/dist/types/runtime/prompt/fresh-workspace-evidence.d.ts

@@ -1,38 +0,0 @@
-import { type SystemPromptSection } from "./system-prompt-sections.js";
-export interface PromptMessageLike {
-    role?: string;
-    content?: unknown;
-}
-export interface ToolNameLike {
-    function?: {
-        name?: string;
-    };
-    name?: string;
-}
-export interface FreshWorkspaceEvidencePolicy {
-    requiresFreshTool: boolean;
-    reason?: "file-read" | "memory-mutation" | "media-generation" | "media-understanding" | "directory-list" | "content-search" | "shell" | "web";
-    mediaKind?: "image" | "video" | "tts" | "music";
-    allowedToolNames: string[];
-    /** When true, the typed route is an execution boundary rather than prompt guidance. */
-    enforceAllowedToolNames?: boolean;
-}
-export declare function getFreshWorkspaceEvidencePolicy(messages: readonly PromptMessageLike[] | string): FreshWorkspaceEvidencePolicy;
-export declare function isLightweightChatTurn(messages: readonly PromptMessageLike[] | string): boolean;
-export declare function isExplicitNoToolTurn(messages: readonly PromptMessageLike[] | string): boolean;
-/**
- * Apply a hard allowlist only for deterministic Host-owned operations such as
- * media generation/understanding. Other evidence policies keep the complete
- * turn surface so ordinary engineering and orchestration work is not narrowed
- * merely because its wording resembles an evidence request. Namespaced Host
- * tools are matched by their local MCP tool name.
- */
-export declare function selectFreshWorkspaceEvidenceTools<TTool extends ToolNameLike>(tools: readonly TTool[], policy: FreshWorkspaceEvidencePolicy): TTool[];
-/**
- * Prompt notice for an explicit "don't use tools" instruction. This INFORMS the
- * model and leaves the judgment to it — the engine no longer strips the tool list
- * (toolChoice="none"), which also silently disabled the reactive guards and
- * misfired on quoted/reported speech ("he said don't use tools").
- */
-export declare function createExplicitNoToolNoticeSection(): SystemPromptSection;
-export declare function createFreshWorkspaceEvidenceSection(policy: FreshWorkspaceEvidencePolicy, availableToolNames?: readonly string[]): SystemPromptSection | null;

package/dist/types/runtime/prompt/tool-necessity-policy.d.ts

@@ -1,12 +0,0 @@
-import type { FreshWorkspaceEvidencePolicy } from "./fresh-workspace-evidence.js";
-/**
- * Typed, conservative pre-tool routing for requests whose complete input is already
- * present in the latest user message. This is intentionally an operation taxonomy,
- * not a benchmark phrase list: every admitted class is closed over supplied text or
- * scalar values and therefore cannot gain correctness from filesystem, shell or MCP.
- */
-export interface ToolNecessityPolicy {
-    mode: "model-tools" | "direct-response";
-    reason?: "scalar-computation" | "text-transformation" | "structured-data-inspection" | "inline-classification" | "url-component-extraction";
-}
-export declare function resolveToolNecessityPolicy(text: string, evidencePolicy: FreshWorkspaceEvidencePolicy): ToolNecessityPolicy;