agent-sh 0.13.3 → 0.13.4

package/dist/agent/agent-loop.js

@@ -87,7 +87,7 @@ export class AgentLoop {
     // doing X." Addresses Q3 in QUESTIONS.md.
     lastErrorByTool = new Map(); // tool → error summary
     lastErrorByFile = new Map(); // file path → error summary
-    static THINKING_LEVELS = ["off", "low", "medium", "high"];
+    static THINKING_LEVELS = ["off", "low", "medium", "high", "xhigh"];
     bus;
     llmClient;
     handlers;
@@ -168,8 +168,13 @@ export class AgentLoop {
             ];
             if (prev) {
                 const newIdx = this.modes.findIndex((m) => m.model === prev.model && m.provider === prev.provider);
-                if (newIdx !== -1)
+                if (newIdx !== -1) {
                     this.currentModeIndex = newIdx;
+                    const next = this.modes[newIdx];
+                    if (next.providerConfig && next.providerConfig !== prev.providerConfig) {
+                        this.llmClient.reconfigure({ ...next.providerConfig, model: next.model });
+                    }
+                }
             }
             if (activePreserved && prev) {
                 this.bus.emit("ui:info", {
@@ -507,7 +512,8 @@ export class AgentLoop {
             return mode.buildReasoningParams(this.thinkingLevel);
         if (this.thinkingLevel === "off")
             return {};
-        return { reasoning_effort: this.thinkingLevel };
+        const effort = this.thinkingLevel === "xhigh" ? "high" : this.thinkingLevel;
+        return { reasoning_effort: effort };
     }
     get currentMode() {
         return this.modes[this.currentModeIndex];
@@ -1076,12 +1082,15 @@ export class AgentLoop {
                 streamedCalls: streamedToolCalls,
             });
             fullResponseText += text;
-            // Record the assistant message via protocol
-            this.toolProtocol.recordAssistant(this.conversation, text, toolCalls, extras);
-            this.bus.emit("conversation:message-appended", {
-                role: "assistant",
-                content: text,
-            });
+            if (text || toolCalls.length > 0) {
+                this.toolProtocol.recordAssistant(this.conversation, text, toolCalls, extras);
+                this.bus.emit("conversation:message-appended", {
+                    role: "assistant",
+                    content: text,
+                });
+            }
+            if (signal.aborted)
+                break;
             // No tool calls → agent is done
             if (toolCalls.length === 0) {
                 this.conversation.eagerNucleateAgent(fullResponseText);
@@ -1502,83 +1511,90 @@ export class AgentLoop {
         };
         this.bus.emit("llm:request", requestParams);
         const stream = await this.llmClient.stream({ ...requestParams, signal });
-        for await (const chunk of stream) {
-            if (signal.aborted)
-                break;
-            this.bus.emit("llm:chunk", { chunk });
-            // Token usage (may arrive in a chunk with empty choices)
-            if (chunk.usage) {
-                const u = chunk.usage;
-                const promptTokens = u.prompt_tokens ?? 0;
-                this.bus.emit("agent:usage", {
-                    prompt_tokens: promptTokens,
-                    completion_tokens: u.completion_tokens ?? 0,
-                    total_tokens: u.total_tokens ?? 0,
-                });
-                // Feed accurate token count back to conversation state
-                if (promptTokens > 0) {
-                    this.conversation.updateApiTokenCount(promptTokens);
-                }
-            }
-            const choice = chunk.choices[0];
-            if (!choice)
-                continue;
-            const delta = choice.delta;
-            // Text content
-            if (delta?.content) {
-                text += delta.content;
-                // Filter tool tags from display output (inline mode)
-                const displayText = streamFilter
-                    ? streamFilter.feed(delta.content)
-                    : delta.content;
-                if (displayText) {
-                    this.bus.emitTransform("agent:response-chunk", {
-                        blocks: [{ type: "text", text: displayText }],
+        try {
+            for await (const chunk of stream) {
+                if (signal.aborted)
+                    break;
+                this.bus.emit("llm:chunk", { chunk });
+                // Token usage (may arrive in a chunk with empty choices)
+                if (chunk.usage) {
+                    const u = chunk.usage;
+                    const promptTokens = u.prompt_tokens ?? 0;
+                    this.bus.emit("agent:usage", {
+                        prompt_tokens: promptTokens,
+                        completion_tokens: u.completion_tokens ?? 0,
+                        total_tokens: u.total_tokens ?? 0,
                     });
+                    // Feed accurate token count back to conversation state
+                    if (promptTokens > 0) {
+                        this.conversation.updateApiTokenCount(promptTokens);
+                    }
                 }
-            }
-            const d = delta;
-            for (const name of ["reasoning", "reasoning_content"]) {
-                if (typeof d?.[name] === "string" && d[name].length > 0) {
-                    reasoning += d[name];
-                    reasoningField ??= name;
-                    this.bus.emit("agent:thinking-chunk", { text: d[name] });
-                }
-            }
-            if (Array.isArray(d?.reasoning_details)) {
-                for (const x of d.reasoning_details) {
-                    const idx = typeof x?.index === "number" ? x.index : reasoningDetailsByIndex.size;
-                    const prev = reasoningDetailsByIndex.get(idx);
-                    if (!prev) {
-                        reasoningDetailsByIndex.set(idx, { ...x });
+                const choice = chunk.choices[0];
+                if (!choice)
+                    continue;
+                const delta = choice.delta;
+                // Text content
+                if (delta?.content) {
+                    text += delta.content;
+                    // Filter tool tags from display output (inline mode)
+                    const displayText = streamFilter
+                        ? streamFilter.feed(delta.content)
+                        : delta.content;
+                    if (displayText) {
+                        this.bus.emitTransform("agent:response-chunk", {
+                            blocks: [{ type: "text", text: displayText }],
+                        });
                     }
-                    else {
-                        if (typeof x.text === "string")
-                            prev.text = (prev.text ?? "") + x.text;
-                        for (const [k, v] of Object.entries(x))
-                            if (k !== "text" && prev[k] === undefined)
-                                prev[k] = v;
+                }
+                const d = delta;
+                for (const name of ["reasoning", "reasoning_content"]) {
+                    if (typeof d?.[name] === "string" && d[name].length > 0) {
+                        reasoning += d[name];
+                        reasoningField ??= name;
+                        this.bus.emit("agent:thinking-chunk", { text: d[name] });
                     }
                 }
-            }
-            // Tool calls (streamed incrementally)
-            if (delta?.tool_calls) {
-                for (const tc of delta.tool_calls) {
-                    const idx = tc.index;
-                    if (!pendingToolCalls[idx]) {
-                        pendingToolCalls[idx] = {
-                            id: tc.id,
-                            name: tc.function.name,
-                            argumentsJson: "",
-                        };
+                if (Array.isArray(d?.reasoning_details)) {
+                    for (const x of d.reasoning_details) {
+                        const idx = typeof x?.index === "number" ? x.index : reasoningDetailsByIndex.size;
+                        const prev = reasoningDetailsByIndex.get(idx);
+                        if (!prev) {
+                            reasoningDetailsByIndex.set(idx, { ...x });
+                        }
+                        else {
+                            if (typeof x.text === "string")
+                                prev.text = (prev.text ?? "") + x.text;
+                            for (const [k, v] of Object.entries(x))
+                                if (k !== "text" && prev[k] === undefined)
+                                    prev[k] = v;
+                        }
                     }
-                    if (tc.function?.arguments) {
-                        pendingToolCalls[idx].argumentsJson +=
-                            tc.function.arguments;
+                }
+                // Tool calls (streamed incrementally)
+                if (delta?.tool_calls) {
+                    for (const tc of delta.tool_calls) {
+                        const idx = tc.index;
+                        if (!pendingToolCalls[idx]) {
+                            pendingToolCalls[idx] = {
+                                id: tc.id,
+                                name: tc.function.name,
+                                argumentsJson: "",
+                            };
+                        }
+                        if (tc.function?.arguments) {
+                            pendingToolCalls[idx].argumentsJson +=
+                                tc.function.arguments;
+                        }
                     }
                 }
             }
         }
+        catch (e) {
+            // On abort, fall through with whatever was accumulated so far.
+            if (!signal.aborted)
+                throw e;
+        }
         // Flush any buffered content from the stream filter
         if (streamFilter) {
             const remaining = streamFilter.flush();

package/dist/agent/index.js

@@ -15,7 +15,9 @@ function persistedModelFor(providerName) {
     return getSettings().providers?.[providerName]?.defaultModel;
 }
 function defaultReasoningBuilder(level) {
-    return level === "off" ? {} : { reasoning_effort: level };
+    if (level === "off")
+        return {};
+    return { reasoning_effort: level === "xhigh" ? "high" : level };
 }
 function mergeCaps(settingsCaps, payloadCaps, modelIds) {
     if (!settingsCaps)
@@ -118,11 +120,12 @@ export default function agentBackend(ctx) {
     ctx.define("llm:get-client", () => llmClient);
     ctx.define("llm:invoke", (messages, opts) => {
         const effort = opts?.reasoningEffort;
+        const clampedEffort = effort === "xhigh" ? "high" : effort;
         return llmClient.complete({
             messages: messages,
             max_tokens: opts?.maxTokens,
             model: opts?.model,
-            ...(effort && effort !== "off" ? { reasoning_effort: effort } : {}),
+            ...(clampedEffort && clampedEffort !== "off" ? { reasoning_effort: clampedEffort } : {}),
         });
     });
     let modes = [];

package/dist/agent/providers/openai.d.ts

@@ -2,6 +2,8 @@
  * Cloud OpenAI (api.openai.com). reasoning_effort vocabulary diverges per
  * family: o-series has no off; gpt-5-codex floors at "low"; plain gpt-5
  * floors at "minimal"; gpt-5.1+ accepts "none" as documented full off.
+ * Top tier: only gpt-5.1-codex-max and gpt-5.[4-9]+ accept "xhigh"; others
+ * clamp to "high".
  */
 import type { AgentContext } from "../host-types.js";
 export default function activate(ctx: AgentContext): void;

package/dist/agent/providers/openai.js

@@ -18,9 +18,16 @@ function offEffortFor(model) {
         return "minimal";
     return null;
 }
+function supportsXhigh(model) {
+    if (model.startsWith("gpt-5.1-codex-max"))
+        return true;
+    return /^gpt-5\.[4-9]/.test(model);
+}
 function buildReasoningParams(level, model) {
-    if (level !== "off")
-        return { reasoning_effort: level };
+    if (level !== "off") {
+        const effort = level === "xhigh" && !(model && supportsXhigh(model)) ? "high" : level;
+        return { reasoning_effort: effort };
+    }
     const off = model ? offEffortFor(model) : null;
     return off ? { reasoning_effort: off } : {};
 }

package/examples/extensions/ash-acp-bridge/src/index.ts

@@ -446,10 +446,13 @@ function getModelsPayload(): Record<string, unknown> | undefined {
   if (!core) return undefined;
   const info = core.bus.emitPipe("config:get-models", { models: [], active: null });
   if (!info.models.length) return undefined;
+  const idFor = (m: { model: string; provider: string }) =>
+    m.provider ? `${m.model}@${m.provider}` : m.model;
+  const current = info.active ?? info.models[0]!;
   return {
-    currentModelId: info.active?.model ?? info.models[0]?.model,
+    currentModelId: idFor(current),
     availableModels: info.models.map((m) => ({
-      modelId: m.model,
+      modelId: idFor(m),
       name: m.provider ? `${m.provider}/${m.model}` : m.model,
       description: m.provider ? `Provider: ${m.provider}` : "",
     })),

package/examples/extensions/ashi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@guanyilun/ashi",
-  "version": "0.1.1",
+  "version": "0.1.3",
   "description": "Ash in an interactive TUI — agent-sh's built-in agent without the shell underneath",
   "type": "module",
   "main": "dist/cli.js",
@@ -48,7 +48,7 @@
   },
   "dependencies": {
     "@earendil-works/pi-tui": "^0.74.0",
-    "agent-sh": "^0.13.2",
+    "agent-sh": "^0.13.3",
     "chalk": "^5.5.0",
     "cli-highlight": "^2.1.11"
   },

package/examples/extensions/ollama.ts

@@ -1,44 +1,46 @@
 /**
- * Ollama provider extension — local daemon and Ollama Cloud.
+ * Ollama provider extension — local daemon or Ollama Cloud.
  *
- *   OLLAMA_API_KEY  → Ollama Cloud (https://ollama.com)
- *   OLLAMA_HOST     → local host override (default http://localhost:11434)
+ * Cloud auth (any of):
+ *   agent-sh auth login ollama-cloud   # preferred
+ *   OLLAMA_API_KEY=...                 # env fallback
+ *
+ * Local host:
+ *   OLLAMA_HOST (default http://localhost:11434)
  *
  * Catalog comes from /api/tags; per-model context length is fetched
  * from /api/show (model_info["${arch}.context_length"]). Chat goes
  * through the OpenAI-compatible /v1/chat/completions shim.
  *
- * Setup (cloud):
- *   export OLLAMA_API_KEY="your-key"
- *
- * Setup (local):
- *   ollama serve   # default http://localhost:11434
- *
  * Usage:
  *   agent-sh -e ./examples/extensions/ollama.ts
  *
  *   # Or add to settings.json:
  *   { "extensions": ["./examples/extensions/ollama.ts"] }
  */
-import type { ExtensionContext } from "agent-sh/types";
+import { resolveApiKey } from "agent-sh/auth";
+import type { AgentContext } from "agent-sh/types";
 
 const ECHO_REASONING_PATTERNS: RegExp[] = [/deepseek/i];
 
-export default function activate(ctx: ExtensionContext): void {
-  const apiKey = process.env.OLLAMA_API_KEY;
-  const host = apiKey
+export default function activate(ctx: AgentContext): void {
+  const cloudKey = resolveApiKey("ollama-cloud").key ?? process.env.OLLAMA_API_KEY;
+  const host = cloudKey
     ? "https://ollama.com"
     : (process.env.OLLAMA_HOST ?? "http://localhost:11434").replace(/\/$/, "");
-  const id = apiKey ? "ollama-cloud" : "ollama";
+  const id = cloudKey ? "ollama-cloud" : "ollama";
 
   // OpenAI SDK rejects an empty apiKey; the local daemon ignores the value.
-  const sdkKey = apiKey || "no-key";
+  const sdkKey = cloudKey || "no-key";
   const baseURL = `${host}/v1`;
   const headers: Record<string, string> = {};
-  if (apiKey) headers.Authorization = `Bearer ${apiKey}`;
+  if (cloudKey) headers.Authorization = `Bearer ${cloudKey}`;
 
   ctx.agent.providers.configure(id, {
-    reasoningParams: (level) => ({ reasoning_effort: level === "off" ? "none" : level }),
+    reasoningParams: (level) => {
+      if (level === "off") return { reasoning_effort: "none" };
+      return { reasoning_effort: level === "xhigh" ? "high" : level };
+    },
   });
 
   ctx.bus.emit("provider:register", { id, apiKey: sdkKey, baseURL, models: [] });

package/examples/extensions/zai-coding-plan.ts

@@ -19,7 +19,8 @@ const DEFAULT_MODELS = [
 
 function buildReasoningParams(level: string, _model?: string): Record<string, unknown> {
   if (level === "off") return { thinking: { type: "disabled" } };
-  return { thinking: { type: "enabled" }, reasoning_effort: level };
+  const effort = level === "xhigh" ? "high" : level;
+  return { thinking: { type: "enabled" }, reasoning_effort: effort };
 }
 
 export default function activate(ctx: AgentContext): void {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-sh",
-  "version": "0.13.3",
+  "version": "0.13.4",
   "description": "A shell-first terminal where AI is one keystroke away",
   "type": "module",
   "main": "dist/core/index.js",

package/examples/extensions/ollama-cloud.ts

@@ -1,78 +0,0 @@
-/**
- * Ollama Cloud — hosted Ollama models (https://ollama.com).
- *
- * Auth:  agent-sh auth login ollama-cloud
- * Usage: agent-sh -e ./examples/extensions/ollama-cloud.ts
- */
-import { resolveApiKey } from "agent-sh/auth";
-import type { AgentContext } from "agent-sh/types";
-
-const HOST = "https://ollama.com";
-const BASE_URL = `${HOST}/v1`;
-const ID = "ollama-cloud";
-
-function buildReasoningParams(level: string, _model?: string): Record<string, unknown> {
-  return { reasoning_effort: level === "off" ? "none" : level };
-}
-
-async function fetchModels(apiKey: string) {
-  const headers: Record<string, string> = { Authorization: `Bearer ${apiKey}` };
-  const tagsRes = await fetch(`${HOST}/api/tags`, { headers });
-  if (!tagsRes.ok) return [];
-  const tagsData = await tagsRes.json() as { models?: { name: string }[] };
-  const names = (tagsData.models ?? []).map((m) => m.name);
-  if (!names.length) return [];
-
-  const ctxs = await Promise.all(
-    names.map((name) =>
-      fetch(`${HOST}/api/show`, {
-        method: "POST",
-        headers: { ...headers, "Content-Type": "application/json" },
-        body: JSON.stringify({ name }),
-      })
-        .then((r) => r.ok ? r.json() as Promise<{ model_info?: Record<string, unknown> }> : null)
-        .then((d) => {
-          if (!d?.model_info) return undefined;
-          const info = d.model_info;
-          const arch = info["general.architecture"] as string | undefined;
-          if (arch) {
-            const ctx = info[`${arch}.context_length`];
-            if (typeof ctx === "number") return ctx;
-          }
-          for (const [k, v] of Object.entries(info)) {
-            if (k.endsWith(".context_length") && typeof v === "number") return v;
-          }
-          return undefined;
-        })
-        .catch(() => undefined),
-    ),
-  );
-
-  return names.map((name, i) => ({
-    id: name,
-    contextWindow: ctxs[i],
-    echoReasoning: /deepseek/i.test(name),
-  }));
-}
-
-export default function activate(ctx: AgentContext): void {
-  const { key } = resolveApiKey(ID);
-  const apiKey = key ?? process.env.OLLAMA_API_KEY;
-  if (!apiKey) return;
-
-  ctx.agent.providers.configure(ID, { reasoningParams: buildReasoningParams });
-
-  // Register placeholder while catalog loads
-  ctx.bus.emit("provider:register", { id: ID, apiKey, baseURL: BASE_URL, models: [] });
-
-  fetchModels(apiKey).then((models) => {
-    if (!models.length) return;
-    ctx.bus.emit("provider:register", {
-      id: ID,
-      apiKey,
-      baseURL: BASE_URL,
-      defaultModel: models[0]!.id,
-      models,
-    });
-  }).catch(() => { /* keep placeholder */ });
-}