chainlesschain 0.166.127 → 0.166.129

package/src/lib/goal-condition-engine.js

@@ -157,7 +157,7 @@ export function checkDeterministicCondition(condition, evidence = {}) {
       };
     }
     case "exit-zero": {
-      const met = Number(evidence.exitCode) === 0;
+      const met = evidence.exitCode === 0;
       return {
         met,
         reason: `\`${condition.command}\` exited ${evidence.exitCode}`,
@@ -205,7 +205,10 @@ export function runDeterministicCheck(condition, deps = {}) {
       encoding: "utf8",
       windowsHide: true,
     });
-    evidence.exitCode = res?.status ?? (res?.error ? 1 : null);
+    // A timeout/signal or spawn error has no successful process exit, even if
+    // an adapter also supplies a status. Never coerce null/false/"0" to zero.
+    evidence.exitCode =
+      res?.error || res?.signal ? null : (res?.status ?? null);
   } else if (condition.kind === "file-exists") {
     const existsSync = deps.existsSync;
     evidence.fileExists = (p) =>
@@ -381,6 +384,25 @@ export class GoalConditionEngine {
 
   /** Restore a prior run (same clock injection contract). */
   static fromSnapshot(snap, { now } = {}) {
+    const state = snap?.state;
+    if (
+      !snap?.condition ||
+      !["model", "exit-zero", "file-exists", "contains", "regex"].includes(
+        snap.condition.kind,
+      ) ||
+      !state ||
+      !Number.isSafeInteger(state.outerTurns) ||
+      state.outerTurns < 0 ||
+      !Number.isSafeInteger(state.tokens) ||
+      state.tokens < 0 ||
+      typeof state.costUsd !== "number" ||
+      !Number.isFinite(state.costUsd) ||
+      state.costUsd < 0 ||
+      typeof state.startedAtMs !== "number" ||
+      !Number.isFinite(state.startedAtMs) ||
+      typeof state.done !== "boolean"
+    )
+      throw new TypeError("Invalid goal checkpoint state");
     const engine = new GoalConditionEngine({
       condition: snap.condition,
       budget: snap.budget,

package/src/lib/llm-connection-probe.js

@@ -3,6 +3,12 @@ import { estimateCost } from "./llm-pricing.js";
 import { providerFetch } from "./formal-budget-provider-scope.js";
 import { isFormalBudgetError } from "./formal-budget-errors.js";
 import { normalizeOllamaBaseUrl } from "./ollama-base-url.js";
+import { resolveModelCapabilityProfile } from "./model-capabilities.js";
+import {
+  createOpenAIResponsesBody,
+  normalizeOpenAIResponsesResponse,
+  normalizeOpenAIResponsesUsage,
+} from "./openai-responses.js";
 
 // Reasoning models spend the same output allowance on thinking before the
 // final answer. A 16-token probe can therefore receive HTTP 200 with only
@@ -13,10 +19,16 @@ export const LLM_PROBE_SCHEMA = "chainlesschain.llm-connection-probe/v1";
 const count = (value) => Number.isSafeInteger(value) && value >= 0;
 
 /** Preserve missing/malformed usage as unknown, including cached prompt tokens. */
-export function normalizeLlmProbeUsage(provider, data) {
+export function normalizeLlmProbeUsage(provider, data, runtimeProtocol = null) {
   if (!data || typeof data !== "object" || Array.isArray(data)) return null;
   let usage;
-  if (provider === "ollama") {
+  if (provider === "openai" && runtimeProtocol === "openai-responses") {
+    // The existing Responses codec owns cached/reasoning token accounting.
+    // Do not invent a priced cache-write split that the probe has not verified.
+    const writes = data.usage?.input_tokens_details?.cache_write_tokens;
+    if (writes !== undefined && writes !== 0) return null;
+    usage = normalizeOpenAIResponsesUsage(data.usage);
+  } else if (provider === "ollama") {
     usage = {
       input_tokens: data.prompt_eval_count,
       output_tokens: data.eval_count,
@@ -145,6 +157,15 @@ export async function probeLlmConnection(
       ? normalizeOllamaBaseUrl(target.baseUrl)
       : target.baseUrl.replace(/\/+$/, "");
   const headers = { "Content-Type": "application/json" };
+  const profile =
+    target.provider === "openai"
+      ? resolveModelCapabilityProfile({
+          provider: target.provider,
+          model: target.model,
+          baseUrl: target.baseUrl,
+        })
+      : null;
+  const nativeResponses = profile?.runtimeProtocol === "openai-responses";
   let url, body;
   if (target.provider === "ollama") {
     url = `${base}/api/generate`;
@@ -172,6 +193,15 @@ export async function probeLlmConnection(
         contents: [{ parts: [{ text: "Say hi in one word." }] }],
         generationConfig: { maxOutputTokens: PROBE_OUTPUT_TOKENS },
       };
+    } else if (nativeResponses) {
+      url = `${base}/responses`;
+      headers.Authorization = `Bearer ${target.apiKey}`;
+      body = createOpenAIResponsesBody({
+        model: target.model,
+        messages: [{ role: "user", content: "Say hi in one word." }],
+        maxOutputTokens: PROBE_OUTPUT_TOKENS,
+        modelProfile: profile,
+      });
     } else {
       url = `${base}/chat/completions`;
       headers.Authorization = `Bearer ${target.apiKey}`;
@@ -204,7 +234,19 @@ export async function probeLlmConnection(
         }),
       );
     const data = await response.json();
-    usage = normalizeLlmProbeUsage(target.provider, data);
+    usage = normalizeLlmProbeUsage(
+      target.provider,
+      data,
+      profile?.runtimeProtocol,
+    );
+    if (nativeResponses && (data?.error || data?.status !== "completed")) {
+      throw new Error(
+        data?.status === "incomplete" &&
+        data?.incomplete_details?.reason === "max_output_tokens"
+          ? "The model responded but exhausted the test output limit before a final answer; this does not mean the connection settings are invalid"
+          : "The Responses endpoint did not complete the connection test",
+      );
+    }
     const text =
       target.provider === "ollama"
         ? data.response
@@ -218,7 +260,9 @@ export async function probeLlmConnection(
                 ?.filter((p) => p.thought !== true)
                 ?.map((p) => p.text || "")
                 .join("")
-            : data.choices?.[0]?.message?.content;
+            : nativeResponses
+              ? normalizeOpenAIResponsesResponse(data).message.content
+              : data.choices?.[0]?.message?.content;
     if (typeof text !== "string" || !text.trim()) {
       const reachedLimit =
         data.choices?.[0]?.finish_reason === "length" ||

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,