@kevin5251984/guild 0.2.19 → 0.2.20

package/src/llm.ts

@@ -39,6 +39,16 @@ import {
   fromOpenAiUsage,
   withDuration,
 } from "./usage.ts";
+import {
+  attachReasoning,
+  clampEffort,
+  reasoningFor,
+  reasoningPayload,
+  refreshReasoningCatalog,
+  sanitizeEffort,
+} from "./reasoning-catalog.ts";
+
+export { refreshReasoningCatalog };
 import {
   OPENCODE_FREE_DEFAULT_MODEL,
   OPENCODE_FREE_PROVIDER_ID,
@@ -291,6 +301,7 @@ export function publicModels(dataDir: string, env: NodeJS.ProcessEnv = process.e
       return {
         id,
         ...provider,
+        models: attachReasoning(id, provider.models),
         apiKey: stored === "literal" ? "" : key,
         apiKeyPreview: key ? maskApiKey(key) : "",
         stored,
@@ -307,19 +318,19 @@ export function publicModels(dataDir: string, env: NodeJS.ProcessEnv = process.e
       ready:
         isKeylessProvider(p.id) ||
         Boolean(resolveApiKey(p.apiKey, env) || p.stored === "literal"),
-      models: p.models,
+      models: attachReasoning(p.id, p.models),
     })),
     ...subscriptions.map((s) => ({
       id: s.pickerId,
       name: s.name,
       kind: "oauth" as const,
       ready: s.ready,
-      models: s.models ?? [],
+      models: attachReasoning(s.pickerId, s.models ?? []),
     })),
   ];
   return {
     default: file.default ?? null,
-    reasoning: file.reasoning ?? "medium",
+    reasoning: file.reasoning ?? "",
     fast: Boolean(file.fast),
     aux: file.aux ?? {},
     auxRoles: AUX_ROLES,
@@ -483,9 +494,14 @@ export async function llmComplete(input: {
     spawnDepth: 0,
     allowWrite: true,
   };
+  const file = readModelsFile(input.dataDir);
+  const effort = clampEffort(
+    file.fast ? "low" : file.reasoning,
+    reasoningFor(target.providerId, target.model),
+    Boolean(file.fast),
+  );
   if (OAUTH_PICKER_IDS.has(target.providerId)) {
     try {
-      const file = readModelsFile(input.dataDir);
       return await completeOAuth({
         dataDir: input.dataDir,
         pickerId: target.providerId,
@@ -493,7 +509,7 @@ export async function llmComplete(input: {
         system: input.system,
         messages: input.messages,
         temperature: input.temperature ?? 0.4,
-        reasoning: file.fast ? "low" : file.reasoning,
+        reasoning: effort,
         tools: useTools,
         skills: input.skills,
         toolCtx,
@@ -521,6 +537,7 @@ export async function llmComplete(input: {
       input.temperature ?? 0.4,
       useTools,
       toolCtx,
+      effort,
     );
     if (!done) return null;
     return {
@@ -554,11 +571,12 @@ async function dispatchComplete(
   temperature: number,
   tools: boolean,
   ctx: ToolContext,
+  effort?: string,
 ): Promise<DispatchResult | null> {
   if (isKeylessProvider(target.providerId) && usesZenResponses(target.model)) {
     return tools
-      ? completeZenResponsesTools(target, system, messages, ctx)
-      : wrapText(await completeZenResponses(target, system, messages));
+      ? completeZenResponsesTools(target, system, messages, ctx, effort)
+      : wrapText(await completeZenResponses(target, system, messages, effort));
   }
   if (target.api === "openai-responses") {
     const text = await completeCodex(target, system, messages);
@@ -570,8 +588,8 @@ async function dispatchComplete(
       : wrapText(await completeAnthropic(target, system, messages));
   }
   return tools
-    ? completeOpenAiTools(target, system, messages, temperature, ctx)
-    : wrapText(await completeOpenAi(target, system, messages, temperature));
+    ? completeOpenAiTools(target, system, messages, temperature, ctx, effort)
+    : wrapText(await completeOpenAi(target, system, messages, temperature, effort));
 }
 
 function wrapText(text: string | null): DispatchResult | null {
@@ -584,6 +602,7 @@ async function completeOpenAiTools(
   messages: { role: "user" | "assistant"; content: string }[],
   temperature: number,
   ctx: ToolContext,
+  effort?: string,
 ): Promise<DispatchResult | null> {
   type ChatMsg = {
     role: string;
@@ -635,6 +654,7 @@ async function completeOpenAiTools(
               messages: msgs,
               tools: catalog,
               tool_choice: "auto",
+              ...reasoningPayload(target.providerId, target.baseUrl, effort),
             }),
             signal: roundSignal(ctx),
           });
@@ -901,6 +921,7 @@ async function completeZenResponses(
   target: LlmTarget,
   system: string,
   messages: { role: "user" | "assistant"; content: string }[],
+  effort?: string,
 ): Promise<string | null> {
   const response = await postZenResponses(
     target,
@@ -908,6 +929,7 @@ async function completeZenResponses(
       model: target.model,
       instructions: system,
       input: messages,
+      ...reasoningPayload(target.providerId, target.baseUrl, effort),
     },
     AbortSignal.timeout(STREAM_IDLE_TIMEOUT_MS),
   );
@@ -921,6 +943,7 @@ async function completeZenResponsesTools(
   system: string,
   messages: { role: "user" | "assistant"; content: string }[],
   ctx: ToolContext,
+  effort?: string,
 ): Promise<DispatchResult | null> {
   const input: ZenInput[] = messages.map((item) => ({
     role: item.role,
@@ -960,6 +983,7 @@ async function completeZenResponsesTools(
               input,
               tools,
               tool_choice: "auto",
+              ...reasoningPayload(target.providerId, target.baseUrl, effort),
             },
             roundSignal(ctx),
           );
@@ -1070,6 +1094,7 @@ async function completeOpenAi(
   system: string,
   messages: { role: "user" | "assistant"; content: string }[],
   temperature: number,
+  effort?: string,
 ): Promise<string | null> {
   const url = `${target.baseUrl}/chat/completions`;
   const response = await fetch(url, {
@@ -1079,6 +1104,7 @@ async function completeOpenAi(
       model: target.model,
       temperature,
       messages: [{ role: "system", content: system }, ...messages],
+      ...reasoningPayload(target.providerId, target.baseUrl, effort),
     }),
     signal: AbortSignal.timeout(STREAM_IDLE_TIMEOUT_MS),
   });
@@ -1278,6 +1304,7 @@ function sanitizeModels(file: ModelsFile): ModelsFile {
       .map((model) => ({
         id: model.id.trim(),
         name: model.name?.trim() || undefined,
+        ...(model.reasoning ? { reasoning: model.reasoning } : {}),
       }));
     if (models.length === 0) {
       throw new StoreError(400, `provider ${id} needs at least one model`);
@@ -1307,12 +1334,7 @@ function sanitizeModels(file: ModelsFile): ModelsFile {
   for (const [role, ref] of Object.entries(file.aux ?? {})) {
     aux[role as AuxRole] = validRef(ref as ModelRef | null);
   }
-  const reasoning =
-    file.reasoning === "minimal" ||
-    file.reasoning === "low" ||
-    file.reasoning === "high"
-      ? file.reasoning
-      : "medium";
+  const reasoning = sanitizeEffort(file.reasoning);
   const recent = (file.recent ?? [])
     .map((ref) => validRef(ref))
     .filter((ref): ref is ModelRef => Boolean(ref))

package/src/mcp.ts

@@ -1,4 +1,10 @@
-import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import {
+  chmodSync,
+  existsSync,
+  mkdirSync,
+  readFileSync,
+  writeFileSync,
+} from "node:fs";
 import { homedir } from "node:os";
 import { join } from "node:path";
 import { spawn, type ChildProcessWithoutNullStreams } from "node:child_process";
@@ -215,6 +221,8 @@ class McpSession {
 
 const sessions = new Map<string, McpSession>();
 
+const REDACTED_ENV_VALUE = "***";
+
 export function mcpPath(dataDir: string): string {
   return join(dataDir, "mcp.json");
 }
@@ -237,10 +245,29 @@ export function writeMcpFile(
   servers: Record<string, McpLaunch>,
 ): void {
   mkdirSync(dataDir, { recursive: true });
-  writeFileSync(
-    mcpPath(dataDir),
-    `${JSON.stringify({ mcpServers: servers }, null, 2)}\n`,
-  );
+  const file = mcpPath(dataDir);
+  writeFileSync(file, `${JSON.stringify({ mcpServers: servers }, null, 2)}\n`, {
+    mode: 0o600,
+  });
+  chmodSync(file, 0o600);
+}
+
+/**
+ * The wire shape for HTTP. `launch.env` holds API keys, so every value becomes
+ * a placeholder; the keys stay so the UI can still show that env is set.
+ * Spawning keeps using the full launch from `listGuildMcp` / `listActiveMcp`.
+ */
+export function publicMcpServer(server: McpServer): McpServer {
+  const env = server.launch?.env;
+  const launch: McpLaunch = { ...server.launch };
+  if (env && Object.keys(env).length) {
+    launch.env = Object.fromEntries(
+      Object.keys(env).map((key) => [key, REDACTED_ENV_VALUE]),
+    );
+  } else {
+    delete launch.env;
+  }
+  return { ...server, launch };
 }
 
 export function listGuildMcp(dataDir: string): McpServer[] {

package/src/mention.ts

@@ -8,6 +8,15 @@ const LIST_MARK = /^[ \t]*(?:\d+[.)、]\s*|[-*•]\s+)/;
 /** Assignee right after the marker, or after 通過後/最後/再叫… — not the first @ anywhere on the line. */
 const LIST_ASSIGN =
   /^(?:`?@([A-Za-z0-9_-]+)`?|(?:通過後|最後|再叫|交給|交棒(?:給)?|指派|(?:call|ask)\s+)\s*`?@([A-Za-z0-9_-]+)`?)/i;
+/** Numbered/bullet item that names a later wave — do not start them this turn. */
+const LIST_LATER =
+  /^(?:通過後|最後|然後|之後|完成後)\s*`?@([A-Za-z0-9_-]+)`?/i;
+/** Prose: 四席完成後由 @infra / 才交 @infra. 再叫 is a handoff, not a defer. */
+const DEFER_AT =
+  /(?:完成後|通過後|最後(?:才|由)?|才需要|才叫|才交)[^@\n]{0,24}@([A-Za-z0-9_-]+)/gi;
+/** Bare handle after the same cues: 才需要call infra. */
+const DEFER_BARE =
+  /(?:完成後|才需要|才叫)\s*(?:call|ask|由)\s*`?@?([A-Za-z0-9_-]+)`?/gi;
 
 /** Drop fenced / inline code so `@pm` in a snippet does not dispatch. */
 export function stripMentionNoise(text: string): string {
@@ -40,6 +49,8 @@ export type MentionBlock = {
   kind: "lead" | "list";
 };
 
+export type MentionBot = { id: string; handle: string };
+
 function knownHandlesIn(chunk: string, known: Set<string>): string[] {
   const names: string[] = [];
   for (const row of chunk.matchAll(HANDLE)) {
@@ -97,17 +108,75 @@ export function lineStartMentions(text: string, handles: string[]): MentionBlock
   }));
 }
 
+/** Fenced samples only. Keep inline `code` so 「再叫 `@infra`」 still parses. */
+function stripFences(text: string): string {
+  return String(text ?? "").replace(/```[\s\S]*?```/g, " ");
+}
+
+/** Seats the same message says to start after someone else finishes. */
+export function deferredHandles(text: string, handles: string[]): string[] {
+  const raw = String(text ?? "");
+  const known = new Set(handles.map((handle) => handle.toLowerCase()));
+  const out: string[] = [];
+  const add = (name: string) => {
+    const key = name.toLowerCase();
+    if (!known.has(key) || out.includes(key)) return;
+    out.push(key);
+  };
+  const scan = stripFences(raw);
+  for (const row of scan.matchAll(new RegExp(DEFER_AT.source, "gi"))) {
+    if (row[1]) add(row[1]);
+  }
+  for (const row of scan.matchAll(new RegExp(DEFER_BARE.source, "gi"))) {
+    if (row[1]) add(row[1]);
+  }
+  let fence = false;
+  for (const part of raw.split(/(\r?\n)/)) {
+    if (part === "\n" || part === "\r\n") continue;
+    const trimmed = part.trim();
+    if (trimmed.startsWith("```")) {
+      fence = !fence;
+      continue;
+    }
+    if (fence || !LIST_MARK.test(part)) continue;
+    const hit = part.replace(LIST_MARK, "").match(LIST_LATER);
+    if (hit?.[1]) add(hit[1]);
+  }
+  return out;
+}
+
+export function withoutDeferredIds(
+  ids: string[],
+  text: string,
+  bots: MentionBot[],
+): string[] {
+  if (!ids.length) return ids;
+  const deferred = new Set(
+    deferredHandles(
+      text,
+      bots.map((bot) => bot.handle),
+    ),
+  );
+  if (!deferred.size) return ids;
+  return ids.filter((id) => {
+    const bot = bots.find((row) => row.id === id);
+    return !bot || !deferred.has(bot.handle.toLowerCase());
+  });
+}
+
 /**
  * Fallback when the client did not pick an assignee.
  * Every line-start @handle group, otherwise the first @handle in prose.
+ * Seats named as a later wave (完成後 / 最後 / 才叫) stay quiet this turn.
  */
 export function summonedHandles(text: string, handles: string[]): string[] {
   const known = new Set(handles.map((handle) => handle.toLowerCase()));
+  const deferred = new Set(deferredHandles(text, handles));
   const take = (names: string[]): string[] => {
     const out: string[] = [];
     for (const name of names) {
       const key = name.toLowerCase();
-      if (!known.has(key) || out.includes(key)) continue;
+      if (!known.has(key) || deferred.has(key) || out.includes(key)) continue;
       out.push(key);
     }
     return out;
@@ -117,7 +186,8 @@ export function summonedHandles(text: string, handles: string[]): string[] {
     const out: string[] = [];
     for (const block of blocks) {
       for (const handle of block.handles) {
-        if (!out.includes(handle)) out.push(handle);
+        if (deferred.has(handle) || out.includes(handle)) continue;
+        out.push(handle);
       }
     }
     return out;
@@ -128,11 +198,6 @@ export function summonedHandles(text: string, handles: string[]): string[] {
   return [];
 }
 
-/** Fenced samples only. Keep inline `code` so 「再叫 `@infra`」 still parses. */
-function stripFences(text: string): string {
-  return String(text ?? "").replace(/```[\s\S]*?```/g, " ");
-}
-
 /**
  * Bot replies: line-start specs, else first prose @handle, plus assignment verbs
  * even when the handle is wrapped in backticks (models quote @infra a lot).
@@ -143,18 +208,17 @@ const HANDOFF_ASK =
 
 export function handoffHandles(text: string, handles: string[]): string[] {
   const known = new Set(handles.map((handle) => handle.toLowerCase()));
+  const deferred = new Set(deferredHandles(text, handles));
   const out = summonedHandles(text, handles);
   const ask = new RegExp(HANDOFF_ASK.source, "gi");
   for (const row of stripFences(text).matchAll(ask)) {
     const key = row[1].toLowerCase();
-    if (!known.has(key) || out.includes(key)) continue;
+    if (!known.has(key) || deferred.has(key) || out.includes(key)) continue;
     out.push(key);
   }
   return out;
 }
 
-export type MentionBot = { id: string; handle: string };
-
 export function sanitizeMentionIds(
   raw: unknown,
   bots: MentionBot[],
@@ -188,7 +252,7 @@ export function parseMentionIds(
     if (!bot || out.includes(bot.id)) continue;
     out.push(bot.id);
   }
-  return out;
+  return withoutDeferredIds(out, text, bots);
 }
 
 /**
@@ -206,7 +270,11 @@ export function messageMentionIds(
         bots,
         message.author === "you" ? "user" : "bot",
       );
-  return ids.filter((id) => id !== message.author);
+  return withoutDeferredIds(
+    ids.filter((id) => id !== message.author),
+    message.body,
+    bots,
+  );
 }
 
 /**

package/src/oauth.ts

@@ -1242,7 +1242,7 @@ export async function completeOAuth(input: {
   system: string;
   messages: { role: "user" | "assistant"; content: string }[];
   temperature?: number;
-  reasoning?: "minimal" | "low" | "medium" | "high";
+  reasoning?: string;
   tools?: boolean;
   skills?: SkillRef[];
   toolCtx?: ToolContext;
@@ -1301,7 +1301,7 @@ export async function completeOAuth(input: {
   });
   const options: {
     temperature?: number;
-    reasoning?: "minimal" | "low" | "medium" | "high";
+    reasoning?: string;
     signal?: AbortSignal;
     transformHeaders?: (
       headers: Record<string, string | null>,

package/src/opencode-free.ts

@@ -1,4 +1,5 @@
 import type { ModelEntry, ProviderEntry } from "@guild/protocol";
+import { guildUserAgent } from "./version.ts";
 
 /** Built-in alias `free` → this provider, matching Hermes Agent (2026-08). */
 export const OPENCODE_FREE_PROVIDER_ID = "opencode-free";
@@ -94,7 +95,7 @@ export function llmRequestHeaders(target: {
       "content-type": "application/json",
       "http-referer": "https://github.com/Jakevin/guild",
       "x-title": "Guild",
-      "user-agent": "Guild/0.2.19",
+      "user-agent": guildUserAgent(),
       ...extra,
     };
     delete headers.authorization;
@@ -126,7 +127,7 @@ export async function fetchOpenCodeFreeModels(
     const res = await fetch(`${OPENCODE_FREE_BASE_URL}/models`, {
       headers: {
         accept: "application/json",
-        "user-agent": "Guild/0.2.19",
+        "user-agent": guildUserAgent(),
         "http-referer": "https://github.com/Jakevin/guild",
         "x-title": "Guild",
       },