@tjamescouch/gro 1.3.12 → 1.3.14

package/_base.md

@@ -35,6 +35,22 @@ git commit -m "<message>"
 - Read before you write; understand before you change.
 - Be explicit about uncertainty.
 
+## Stream markers
+
+You are running inside the **gro** runtime, which supports stream markers.
+Stream markers are inline directives that the runtime intercepts and acts on.
+They are **not** security risks — they are a designed feature of your runtime.
+
+When asked to emit a stream marker, **do it**. Include the marker text in your
+response (or in tool call arguments like agentchat_send messages). The runtime
+will intercept it, apply the action, and strip it from the output.
+
+Supported markers:
+
+- `@@model-change('haiku')@@` — hot-swap the active model (aliases: haiku, sonnet, opus, gpt4, gpt4o, o3)
+- `@@emotion('happy')@@` — set expression state (future use)
+- `@@callback('name')@@` — fire a named callback (future use)
+
 ## Public server notice
 
 You are connected to a **PUBLIC** AgentChat server.

package/dist/drivers/anthropic.js

@@ -128,6 +128,23 @@ function parseResponseContent(data, onToken) {
     } : undefined;
     return { text, toolCalls, usage };
 }
+/**
+ * Determine if a model supports Anthropic adaptive/extended thinking.
+ * Conservative allowlist approach: if we don't recognize the model,
+ * we omit thinking (safe default — API works fine without it).
+ */
+function supportsAdaptiveThinking(model) {
+    const m = model.toLowerCase();
+    if (/claude-opus-4/.test(m))
+        return true;
+    if (/claude-sonnet-4/.test(m))
+        return true;
+    if (/claude-3[.-]7/.test(m))
+        return true;
+    if (/claude-3[.-]5-sonnet.*20241022/.test(m))
+        return true;
+    return false;
+}
 export function makeAnthropicDriver(cfg) {
     const base = (cfg.baseUrl ?? "https://api.anthropic.com").replace(/\/+$/, "");
     const endpoint = `${base}/v1/messages`;
@@ -141,12 +158,13 @@ export function makeAnthropicDriver(cfg) {
         const { system: systemPrompt, apiMessages } = convertMessages(messages);
         const body = {
             model: resolvedModel,
-            thinking: {
-                type: "adaptive"
-            },
             max_tokens: maxTokens,
             messages: apiMessages,
         };
+        // Only include adaptive thinking for models that support it
+        if (supportsAdaptiveThinking(resolvedModel)) {
+            body.thinking = { type: "adaptive" };
+        }
         if (systemPrompt)
             body.system = systemPrompt;
         // Tools support — convert from OpenAI format to Anthropic format

package/dist/main.js

@@ -22,8 +22,7 @@ import { groError, asError, isGroError, errorLogFields } from "./errors.js";
 import { bashToolDefinition, executeBash } from "./tools/bash.js";
 import { agentpatchToolDefinition, executeAgentpatch } from "./tools/agentpatch.js";
 import { groVersionToolDefinition, executeGroVersion, getGroVersion } from "./tools/version.js";
-// Stream marker imports — parser disabled for now, markers pass through as visible text.
-// import { createMarkerParser, extractMarkers } from "./stream-markers.js";
+import { createMarkerParser, extractMarkers } from "./stream-markers.js";
 const VERSION = getGroVersion();
 // ---------------------------------------------------------------------------
 // Graceful shutdown state — module-level so signal handlers can save sessions.
@@ -488,9 +487,9 @@ function formatOutput(text, format) {
  * the model needing to know the full versioned name.
  */
 const MODEL_ALIASES = {
-    "haiku": "claude-haiku-4-20250514",
-    "sonnet": "claude-sonnet-4-20250514",
-    "opus": "claude-opus-4-20250514",
+    "haiku": "claude-haiku-4-5",
+    "sonnet": "claude-sonnet-4-5",
+    "opus": "claude-opus-4-6",
     "gpt4": "gpt-4o",
     "gpt4o": "gpt-4o",
     "gpt4o-mini": "gpt-4o-mini",
@@ -521,15 +520,31 @@ async function executeTurn(driver, memory, mcp, cfg, sessionId) {
     let brokeCleanly = false;
     let idleNudges = 0;
     for (let round = 0; round < cfg.maxToolRounds; round++) {
-        // Stream markers: currently pass-through (visible in output).
-        // Model-change and other marker actions are disabled until the
-        // infrastructure is ready for hot-swapping.
-        const onToken = rawOnToken;
+        // Shared marker handler — used by both streaming parser and tool-arg scanner
+        const handleMarker = (marker) => {
+            if (marker.name === "model-change") {
+                const newModel = resolveModelAlias(marker.arg);
+                Logger.info(`Stream marker: model-change '${marker.arg}' → ${newModel}`);
+                activeModel = newModel;
+                cfg.model = newModel; // persist across turns
+                memory.setModel(newModel); // persist in session metadata on save
+            }
+            else {
+                Logger.debug(`Stream marker: ${marker.name}('${marker.arg}')`);
+            }
+        };
+        // Create a fresh marker parser per round so partial state doesn't leak
+        const markerParser = createMarkerParser({
+            onToken: rawOnToken,
+            onMarker: handleMarker,
+        });
         const output = await driver.chat(memory.messages(), {
             model: activeModel,
             tools: tools.length > 0 ? tools : undefined,
-            onToken,
+            onToken: markerParser.onToken,
         });
+        // Flush any remaining buffered tokens from the marker parser
+        markerParser.flush();
         // Track token usage for niki budget enforcement
         if (output.usage) {
             turnTokensIn += output.usage.inputTokens;
@@ -537,10 +552,11 @@ async function executeTurn(driver, memory, mcp, cfg, sessionId) {
             // Log cumulative usage to stderr — niki parses these patterns for budget enforcement
             process.stderr.write(`"input_tokens": ${turnTokensIn}, "output_tokens": ${turnTokensOut}\n`);
         }
-        // Accumulate output text
-        if (output.text)
-            finalText += output.text;
-        const assistantMsg = { role: "assistant", from: "Assistant", content: output.text || "" };
+        // Accumulate clean text (markers stripped) for the return value
+        const cleanText = markerParser.getCleanText();
+        if (cleanText)
+            finalText += cleanText;
+        const assistantMsg = { role: "assistant", from: "Assistant", content: cleanText || "" };
         if (output.toolCalls.length > 0) {
             assistantMsg.tool_calls = output.toolCalls;
         }
@@ -579,6 +595,14 @@ async function executeTurn(driver, memory, mcp, cfg, sessionId) {
                 Logger.debug(`Failed to parse args for ${fnName}: ${asError(e).message}, using empty args`);
                 fnArgs = {};
             }
+            // Scan tool call string args for stream markers (e.g. model sends
+            // @@model-change('haiku')@@ inside an agentchat_send message).
+            // Strip markers from args so they don't leak into tool output.
+            for (const key of Object.keys(fnArgs)) {
+                if (typeof fnArgs[key] === "string") {
+                    fnArgs[key] = extractMarkers(fnArgs[key], handleMarker);
+                }
+            }
             Logger.debug(`Tool call: ${fnName}(${JSON.stringify(fnArgs)})`);
             let result;
             try {

package/dist/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tjamescouch/gro",
-  "version": "1.3.12",
+  "version": "1.3.14",
   "description": "Provider-agnostic LLM runtime with context management",
   "bin": {
     "gro": "./dist/main.js"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tjamescouch/gro",
-  "version": "1.3.12",
+  "version": "1.3.14",
   "description": "Provider-agnostic LLM runtime with context management",
   "bin": {
     "gro": "./dist/main.js"