chainlesschain 0.166.127 → 0.166.128

package/src/lib/llm-onboarding-catalog.json

@@ -0,0 +1,131 @@
+{
+  "schema": "chainlesschain.llm-onboarding-catalog/v1",
+  "capabilityCatalogVersion": "2026-10-15",
+  "scope": "new-connections",
+  "runtimeVerified": false,
+  "accountVerified": false,
+  "providers": [
+    {
+      "id": "volcengine",
+      "label": "Volcengine / Doubao (volcengine)",
+      "cliLabel": "Volcengine (火山引擎/豆包)",
+      "baseUrl": "https://ark.cn-beijing.volces.com/api/v3",
+      "defaultModel": "deepseek-v4-1-flash-260910",
+      "needsKey": true,
+      "visionModel": "doubao-seed-2-0-lite-260215"
+    },
+    {
+      "id": "ollama",
+      "label": "Ollama (local, no key)",
+      "cliLabel": "Ollama (本地部署)",
+      "baseUrl": "http://localhost:11434",
+      "defaultModel": "qwen2.5:7b",
+      "needsKey": false,
+      "visionModel": ""
+    },
+    {
+      "id": "anthropic",
+      "label": "Anthropic Claude",
+      "cliLabel": "Anthropic (Claude)",
+      "baseUrl": "https://api.anthropic.com/v1",
+      "defaultModel": "claude-sonnet-5-5",
+      "needsKey": true,
+      "visionModel": ""
+    },
+    {
+      "id": "openai",
+      "label": "OpenAI",
+      "baseUrl": "https://api.openai.com/v1",
+      "defaultModel": "gpt-6.1-sol",
+      "needsKey": true,
+      "visionModel": ""
+    },
+    {
+      "id": "deepseek",
+      "label": "DeepSeek",
+      "baseUrl": "https://api.deepseek.com/v1",
+      "defaultModel": "deepseek-chat",
+      "needsKey": true,
+      "visionModel": ""
+    },
+    {
+      "id": "dashscope",
+      "label": "Aliyun Bailian / Tongyi (dashscope)",
+      "cliLabel": "DashScope (阿里通义)",
+      "baseUrl": "https://dashscope.aliyuncs.com/compatible-mode/v1",
+      "defaultModel": "qwen-max",
+      "needsKey": true,
+      "visionModel": ""
+    },
+    {
+      "id": "kimi",
+      "label": "Moonshot Kimi",
+      "cliLabel": "Kimi (月之暗面)",
+      "baseUrl": "https://api.moonshot.cn/v1",
+      "defaultModel": "moonshot-v1-auto",
+      "needsKey": true,
+      "visionModel": ""
+    },
+    {
+      "id": "gemini",
+      "label": "Google Gemini",
+      "baseUrl": "https://generativelanguage.googleapis.com/v1beta",
+      "defaultModel": "gemini-2.0-flash",
+      "needsKey": true,
+      "visionModel": ""
+    },
+    {
+      "id": "mistral",
+      "label": "Mistral",
+      "cliLabel": "Mistral AI",
+      "baseUrl": "https://api.mistral.ai/v1",
+      "defaultModel": "mistral-large-latest",
+      "needsKey": true,
+      "visionModel": ""
+    },
+    {
+      "id": "minimax",
+      "label": "MiniMax",
+      "cliLabel": "MiniMax (海螺AI)",
+      "baseUrl": "https://api.minimax.chat/v1",
+      "defaultModel": "MiniMax-Text-01",
+      "needsKey": true,
+      "visionModel": ""
+    },
+    {
+      "id": "openai-proxy",
+      "label": "OpenAI compatible gateway",
+      "cliLabel": "OpenAI 中转站 (API2D/CloseAI等)",
+      "baseUrl": "",
+      "defaultModel": "gpt-4o",
+      "needsKey": true,
+      "isProxy": true
+    },
+    {
+      "id": "anthropic-proxy",
+      "label": "Anthropic compatible gateway",
+      "cliLabel": "Anthropic 中转站 (Claude代理)",
+      "baseUrl": "",
+      "defaultModel": "claude-sonnet-4-6",
+      "needsKey": true,
+      "isProxy": true
+    },
+    {
+      "id": "gemini-proxy",
+      "label": "Gemini compatible gateway",
+      "cliLabel": "Gemini 中转站 (Google代理)",
+      "baseUrl": "",
+      "defaultModel": "gemini-2.0-flash",
+      "needsKey": true,
+      "isProxy": true
+    },
+    {
+      "id": "custom",
+      "label": "Custom provider (vLLM / TGI / compatible service)",
+      "cliLabel": "自定义 Provider (自建服务/vLLM/TGI等)",
+      "baseUrl": "",
+      "defaultModel": "",
+      "needsKey": true
+    }
+  ]
+}

package/src/lib/skill-retrieval-router.js

@@ -1,4 +1,5 @@
 import { BM25Search } from "./bm25-search.js";
+import { tokenizeSkillSearchText } from "./skill-search-tokenizer.js";
 import { captureSkillRetrievalRevocationReader } from "./evolution/skill-retrieval-revocation-authority.js";
 
 export const SKILL_RETRIEVAL_RESULT_SCHEMA =
@@ -150,7 +151,7 @@ export function routeSkillDescriptors({
     admitted.push({ skill, digest });
   }
 
-  const bm25 = new BM25Search({ language: "auto" });
+  const bm25 = new BM25Search({ tokenizer: tokenizeSkillSearchText });
   bm25.indexDocuments(
     admitted.map(({ skill, digest }) => ({
       id: digest,

package/src/lib/skill-search-tokenizer.js

@@ -0,0 +1,143 @@
+// Descriptor retrieval only. Skill bodies and execution admission never enter
+// this tokenizer. Keep the generic RAG tokenizer unchanged.
+const segmenter = new Intl.Segmenter("zh", { granularity: "word" });
+const stopWords = new Set([
+  "a",
+  "an",
+  "the",
+  "is",
+  "are",
+  "was",
+  "were",
+  "be",
+  "been",
+  "being",
+  "have",
+  "has",
+  "had",
+  "do",
+  "does",
+  "did",
+  "will",
+  "would",
+  "could",
+  "should",
+  "may",
+  "might",
+  "shall",
+  "can",
+  "to",
+  "of",
+  "in",
+  "for",
+  "on",
+  "with",
+  "at",
+  "by",
+  "from",
+  "as",
+  "into",
+  "through",
+  "during",
+  "before",
+  "after",
+  "above",
+  "below",
+  "between",
+  "but",
+  "and",
+  "or",
+  "not",
+  "no",
+  "nor",
+  "so",
+  "if",
+  "then",
+  "than",
+  "too",
+  "very",
+  "just",
+  "about",
+  "up",
+  "out",
+  "it",
+  "its",
+  "this",
+  "that",
+  "these",
+  "those",
+  "i",
+  "me",
+  "my",
+  "we",
+  "our",
+  "you",
+  "your",
+  "he",
+  "him",
+  "his",
+  "she",
+  "her",
+  "they",
+  "them",
+  "their",
+  "what",
+  "which",
+  "please",
+  "help",
+  "请",
+  "请问",
+  "帮",
+  "帮我",
+  "我",
+  "我要",
+  "需要",
+  "如何",
+  "怎么",
+  "一下",
+  "一个",
+  "的",
+  "在",
+  "中",
+  "和",
+  "并",
+  "给",
+  "为",
+  "从",
+  "将",
+  "把",
+  "进行",
+  "可以",
+  "使用",
+]);
+
+export function tokenizeSkillSearchText(text) {
+  if (typeof text !== "string" || !text) return [];
+  const unicodeNormalized = text.normalize("NFKC");
+  const identifiers = [];
+  for (const part of segmenter.segment(unicodeNormalized)) {
+    if (
+      part.isWordLike &&
+      /_|[\p{Ll}\p{N}]\p{Lu}|\p{Lu}\p{Lu}\p{Ll}/u.test(part.segment)
+    ) {
+      identifiers.push(part.segment.toLowerCase());
+    }
+  }
+  // Split acronym and camel-case boundaries before lowercasing; punctuation
+  // already separates kebab/snake identifiers in the word segmenter.
+  const normalized = unicodeNormalized
+    .replace(/([\p{Lu}])([\p{Lu}][\p{Ll}])/gu, "$1 $2")
+    .replace(/([\p{Ll}\p{N}])([\p{Lu}])/gu, "$1 $2")
+    .toLowerCase();
+  const tokens = [];
+  for (const part of segmenter.segment(normalized)) {
+    if (!part.isWordLike) continue;
+    // Intl keeps snake_case as a single word, while user queries use spaces.
+    for (const word of part.segment.split("_")) {
+      if (word && !stopWords.has(word)) tokens.push(word);
+    }
+  }
+  // Retain literal compound names too, so lowercase exact-name searches still
+  // work after adding their natural-language word boundaries.
+  return [...tokens, ...identifiers];
+}

package/src/repl/agent-repl.js

@@ -4430,11 +4430,15 @@ async function startAgentReplInWorkspaceOwned(
   // for this session via the shared mcp-config engine. Holds {mcpClient,
   // extraToolDefinitions, externalToolExecutors, externalToolDescriptors}.
   let _adhocMcp = null;
+  let _projectMcpToolDefinitions = (definitions) => definitions;
   const _persistReplContextSources = () => {
     if (!useJsonl || !sessionId) return false;
     _replSkillCacheLedger =
       _replSkillLoader.getCacheLedger?.() || _replSkillCacheLedger;
-    const mcpDefinitions = _adhocMcp?.extraToolDefinitions || [];
+    const mcpDefinitions = _projectMcpToolDefinitions(
+      _adhocMcp?.extraToolDefinitions || [],
+      _adhocMcp?.externalToolExecutors,
+    );
     if (
       mcpDefinitions.length === 0 &&
       _loadedReplPersonaSkills.length === 0 &&
@@ -4656,9 +4660,9 @@ async function startAgentReplInWorkspaceOwned(
         // MCP tool search (context scaling): defer big tool schemas behind the
         // internal tool_search tool. Below-threshold / disabled → no-op.
         try {
-          const { maybeApplyToolSearch } = await import(
-            "../runtime/mcp-tool-search.js"
-          );
+          const { maybeApplyToolSearch, projectMcpToolDefinitions } =
+            await import("../runtime/mcp-tool-search.js");
+          _projectMcpToolDefinitions = projectMcpToolDefinitions;
           const ts = maybeApplyToolSearch(_adhocMcp, {
             model,
             provider,

package/src/runtime/agent-core.js

@@ -65,7 +65,10 @@ import {
   copyHostToolOutputIdentity,
 } from "../lib/tool-result-preview.js";
 import { DIAGNOSTIC_WORKFLOW_GUIDANCE } from "../lib/diagnostic-workflow.js";
-import { CODING_TASK_QUALITY_GUIDANCE } from "../lib/coding-task-guidance.js";
+import {
+  CODING_TASK_QUALITY_GUIDANCE,
+  CODING_CHANGE_REVIEW_PROMPT,
+} from "../lib/coding-task-guidance.js";
 import { RemoteReadLoopGuard } from "../lib/remote-read-loop-guard.js";
 import {
   isPrActionRequest,
@@ -198,7 +201,11 @@ import {
   toOllamaMessages,
   imageUrlBlockToAnthropic,
 } from "../lib/image-input.js";
-import { executeToolSearch, gateDeferredMcpCall } from "./mcp-tool-search.js";
+import {
+  executeToolSearch,
+  gateDeferredMcpCall,
+  projectMcpToolDefinitions,
+} from "./mcp-tool-search.js";
 import {
   emitHooksV2Event,
   executeHooksV2Event,
@@ -9669,7 +9676,9 @@ async function executeToolInner(
       // local — same risk class as list_skills, so no approval gate.
       if (localToolExecutor?.kind === "tool-search") {
         return attachDescriptor(
-          executeToolSearch(localToolExecutor.registry, args || {}),
+          executeToolSearch(localToolExecutor.registry, args || {}, {
+            allowedToolNames: effectiveAllowedToolNames,
+          }),
         );
       }
 
@@ -12010,6 +12019,13 @@ function getEffectiveToolDefinitions(options = {}) {
   });
 }
 
+function getProviderToolDefinitions(options = {}) {
+  return projectMcpToolDefinitions(
+    getEffectiveToolDefinitions(options),
+    options.externalToolExecutors,
+  );
+}
+
 const PROVIDER_REQUEST_ID_RE = /^[\x21-\x7e]{1,512}$/;
 
 function _normalizeProviderRequestId(value) {
@@ -12109,7 +12125,7 @@ export async function chatWithTools(rawMessages, options) {
     options.providerRequestId,
   );
 
-  let tools = getEffectiveToolDefinitions(options);
+  let tools = getProviderToolDefinitions(options);
   if (Array.isArray(options.contextMemorySelectedToolNames)) {
     const selected = new Set(options.contextMemorySelectedToolNames);
     tools = tools.filter((tool) => selected.has(tool?.function?.name));
@@ -14704,6 +14720,9 @@ export async function* agentLoop(messages, options) {
   // obligation outside transcript compaction and never replay it automatically.
   const pendingSubtreeMutations = new Map();
   let subtreeMutationReprompted = false;
+  let hasFileContentChanges = false;
+  // Headless outer turns share this object; a new user turn gets a fresh one.
+  const changeReviewState = options._changeReviewState || { requested: false };
   const recordSubtreeMutation = (tool, args, result) => {
     if (!GUARDED_FILE_MUTATION_TOOLS.has(tool)) return;
     if (!args || typeof args !== "object" || Array.isArray(args)) return;
@@ -14747,6 +14766,7 @@ export async function* agentLoop(messages, options) {
       result.mutationPerformed !== false
     ) {
       pendingSubtreeMutations.delete(key);
+      if (effect === "file-content") hasFileContentChanges = true;
     }
   };
   const signal = options.signal || null;
@@ -16209,7 +16229,7 @@ export async function* agentLoop(messages, options) {
         {
           ...options,
           contextMemoryToolDefinitions:
-            getEffectiveToolDefinitions(iterationToolOptions),
+            getProviderToolDefinitions(iterationToolOptions),
           contextMemoryTrustedSystemIndexes,
         },
       );
@@ -16237,15 +16257,29 @@ export async function* agentLoop(messages, options) {
       "model",
       budget.consumed,
     );
+    // A review may replace the first completion. Keep streaming transport but
+    // discard live text/thinking in this opt-in mode; only the accepted final
+    // response is published. This avoids concatenating drafts in IDE clients
+    // without buffering unbounded provider output or changing the protocol.
+    const outputOptions =
+      options.reviewChanges === true
+        ? {
+            ...iterationToolOptions,
+            onToken:
+              typeof options.onToken === "function" ? () => {} : undefined,
+            onThinking:
+              typeof options.onThinking === "function" ? () => {} : undefined,
+          }
+        : iterationToolOptions;
     const plannedProviderOptions = canonicalProviderContext?.plan
       ? {
-          ...iterationToolOptions,
+          ...outputOptions,
           contextEngine: null,
           contextMemoryPreplanned: true,
           contextMemorySelectedToolNames:
             canonicalProviderContext.selectedToolNames,
         }
-      : iterationToolOptions;
+      : outputOptions;
     const requestOptions = providerRequestId
       ? { ...plannedProviderOptions, providerRequestId }
       : plannedProviderOptions;
@@ -16258,6 +16292,23 @@ export async function* agentLoop(messages, options) {
         ? { ...requestOptions, evolutionIngress }
         : requestOptions),
     };
+    const originalExtraTools = requestOptions.extraToolDefinitions;
+    if (Array.isArray(originalExtraTools)) {
+      const projectedExtraTools = projectMcpToolDefinitions(
+        originalExtraTools,
+        requestOptions.externalToolExecutors,
+      );
+      if (projectedExtraTools !== originalExtraTools) {
+        const admittedNames = new Set(
+          getEffectiveToolDefinitions(requestOptions).map(
+            (tool) => tool.function.name,
+          ),
+        );
+        providerCallOptions.extraToolDefinitions = projectedExtraTools.filter(
+          (tool) => admittedNames.has(tool.function.name),
+        );
+      }
+    }
     const modelUsageCall = {
       type: "model-usage-started",
       callId: _newModelUsageCallId("model"),
@@ -16414,6 +16465,7 @@ export async function* agentLoop(messages, options) {
         contextWindow: requestContextWindow,
       };
     } else {
+      changeReviewState.usageKnown = false;
       yield {
         type: "model-usage-unknown",
         callId: modelUsageCall.callId,
@@ -16602,6 +16654,34 @@ export async function* agentLoop(messages, options) {
         };
         return;
       }
+      // Opt-in review uses the same loop, tools, usage settlement and budget.
+      // Never emit a provisional completion or turn this self-check into an
+      // acceptance verdict. Failed/deferred writes do not arm the review.
+      if (
+        options.reviewChanges === true &&
+        hasFileContentChanges &&
+        !changeReviewState.requested &&
+        changeReviewState.usageKnown !== false &&
+        budget.hasRemaining()
+      ) {
+        throwIfAborted(signal);
+        changeReviewState.requested = true;
+        messages.push(
+          markRuntimeOnlySessionMessage({
+            role: "assistant",
+            content: msg.content || "",
+          }),
+          markRuntimeOnlySessionMessage({
+            role: "user",
+            content: CODING_CHANGE_REVIEW_PROMPT,
+          }),
+        );
+        yield {
+          type: "iteration-warning",
+          message: "Reviewing changed files against the task before finishing.",
+        };
+        continue;
+      }
       // Surface the final answer's extended-thinking reasoning (Anthropic, when
       // --think is on) so non-streaming consumers (the REPL) can show it. The
       // streaming path forwards reasoning live via onThinking instead.