Various fixes for tool calling

server/src/llm/chat-tools.ts

@@ -12,7 +12,7 @@ import { searchSearxng } from "../search/searxng.js";
 import { buildOpenAIConversationMessage } from "./message-content.js";
 import type { ChatMessage } from "./types.js";
 
-const MAX_TOOL_ROUNDS = 4;
+const MAX_TOOL_ROUNDS = env.CHAT_MAX_TOOL_ROUNDS;
 const DEFAULT_WEB_RESULTS = 5;
 const MAX_WEB_RESULTS = 10;
 const DEFAULT_FETCH_MAX_CHARACTERS = 12_000;
@@ -25,6 +25,7 @@ const MAX_SHELL_COMMAND_CHARACTERS = 20_000;
 const DEFAULT_SHELL_MAX_OUTPUT_CHARACTERS = 24_000;
 const MAX_SHELL_MAX_OUTPUT_CHARACTERS = 80_000;
 const REMOTE_EXEC_MAX_BUFFER_BYTES = 1_000_000;
+const MAX_DANGLING_TOOL_INTENT_RETRIES = 1;
 
 const execFileAsync = promisify(execFile);
 
@@ -70,7 +71,7 @@ const CODEX_EXEC_TOOL = {
   function: {
     name: "codex_exec",
     description:
-      "Delegate a coding, terminal, or multi-step software task to a persistent remote Codex CLI workspace. Use for complex code changes, repository inspection, running programs/tests, debugging build failures, or other tasks that need a real shell. Return the remote Codex summary and relevant stdout/stderr.",
+      "Delegate a coding, terminal, or multi-step software task to a persistent remote Codex CLI workspace. Use for complex code changes, repository inspection, running programs/tests, debugging build failures, or other tasks that need a real shell. The task runs non-interactively; the remote Codex instance must make reasonable assumptions, complete the task, and return a final summary with relevant stdout/stderr.",
     parameters: {
       type: "object",
       properties: {
@@ -191,11 +192,12 @@ export const CHAT_TOOL_SYSTEM_PROMPT =
   "You can use tools to gather up-to-date web information when needed. " +
   "Use web_search for discovery and recent facts, and fetch_url to read the full content of a specific page. " +
   "Prefer tools when the user asks for current events, verification, sources, or details you do not already have. " +
+  "When you decide tool use is needed, call the tool immediately in the same response; do not say you are running a tool unless you actually call it. " +
   (env.CHAT_CODEX_TOOL_ENABLED
-    ? "Use codex_exec when a request needs substantial coding work, repository inspection, shell commands, tests, debugging, or another complex task suited to a persistent Codex workspace. Provide codex_exec a complete prompt with the goal, constraints, and expected report-back format. "
+    ? "Use codex_exec when a request needs substantial coding work, repository inspection, shell commands, tests, debugging, or another complex task suited to a persistent Codex workspace. Provide codex_exec a complete prompt with the goal, constraints, assumptions, and expected report-back format. Never ask codex_exec to wait for user input or run interactive commands. "
     : "") +
   (env.CHAT_SHELL_TOOL_ENABLED
-    ? "Use shell_exec for direct command-line work on the remote devbox, including quick Python programs, calculations, file inspection, running tests, and small scripts. "
+    ? "Use shell_exec for direct non-interactive command-line work on the remote devbox, including quick Python programs, calculations, file inspection, running tests, and small scripts. "
     : "") +
   "Do not fabricate tool outputs; reason only from provided tool results.";
 
@@ -535,7 +537,20 @@ function buildDevboxSshTarget() {
 
 function buildRemoteCodexCommand(prompt: string) {
   const workdir = env.CHAT_CODEX_REMOTE_WORKDIR.trim();
-  const codexCommand = `codex exec ${shellQuote(prompt)}`;
+  const wrappedPrompt = [
+    "You are running in a non-interactive batch environment.",
+    "",
+    "Rules:",
+    "- Do not ask questions or wait for user input.",
+    "- Do not use interactive commands, editors, pagers, or prompts.",
+    "- If details are ambiguous, make a reasonable assumption and continue.",
+    "- Complete the task in one run, including any requested file edits, commands, and verification.",
+    "- End with a concise final report that includes changed files, commands run, and outcomes.",
+    "",
+    "Task:",
+    prompt,
+  ].join("\n");
+  const codexCommand = `codex exec --skip-git-repo-check ${shellQuote(wrappedPrompt)} < /dev/null`;
   return `mkdir -p ${shellQuote(workdir)} && cd ${shellQuote(workdir)} && ${codexCommand}`;
 }
 
@@ -595,6 +610,7 @@ async function runCodexExecTool(input: unknown): Promise<ToolRunOutcome> {
 
   const run = async (keyPath?: string) => {
     const sshArgs = [
+      "-n",
       "-o",
       "BatchMode=yes",
       "-o",
@@ -662,6 +678,7 @@ async function runShellExecTool(input: unknown): Promise<ToolRunOutcome> {
 
   const run = async (keyPath?: string) => {
     const sshArgs = [
+      "-n",
       "-o",
       "BatchMode=yes",
       "-o",
@@ -756,6 +773,31 @@ function buildEventArgs(name: string, args: Record<string, unknown>) {
   return args;
 }
 
+function looksLikeDanglingToolIntent(text: string) {
+  const normalized = text
+    .toLowerCase()
+    .replace(/[`*_>#-]/g, " ")
+    .replace(/\s+/g, " ")
+    .trim();
+  if (!normalized) return false;
+  if (normalized.length > 800) return false;
+  if (/\blet me know\b/.test(normalized) || /\bif you (want|would like)\b/.test(normalized)) return false;
+  return (
+    /\b(calling|running|executing|trying|checking|testing)\b.{0,80}\b(now|it|tool|command|shell_exec|codex_exec)\b/.test(normalized) ||
+    /\b(let me|i'?ll|i will)\b.{0,120}\b(run|execute|call|try|check|test)\b/.test(normalized) ||
+    /\b(stand by|hang on|one moment)\b/.test(normalized)
+  );
+}
+
+function appendDanglingToolIntentCorrection(conversation: any[], text: string) {
+  conversation.push({ role: "assistant", content: text });
+  conversation.push({
+    role: "system",
+    content:
+      "Internal correction: the previous assistant message claimed it would run a tool, but no tool call was made. If the task needs an available tool, call it now. Otherwise provide the final answer directly without saying you will run a tool.",
+  });
+}
+
 function mergeUsage(acc: Required<ToolAwareUsage>, usage: any) {
   if (!usage) return false;
   acc.inputTokens += usage.prompt_tokens ?? 0;
@@ -833,6 +875,7 @@ export async function runToolAwareOpenAIChat(params: ToolAwareCompletionParams):
   const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
   let sawUsage = false;
   let totalToolCalls = 0;
+  let danglingToolIntentRetries = 0;
 
   for (let round = 0; round < MAX_TOOL_ROUNDS; round += 1) {
     const completion = await params.client.chat.completions.create({
@@ -858,8 +901,14 @@ export async function runToolAwareOpenAIChat(params: ToolAwareCompletionParams):
 
     const toolCalls = Array.isArray(message.tool_calls) ? message.tool_calls : [];
     if (!toolCalls.length) {
+      const text = typeof message.content === "string" ? message.content : "";
+      if (danglingToolIntentRetries < MAX_DANGLING_TOOL_INTENT_RETRIES && looksLikeDanglingToolIntent(text)) {
+        danglingToolIntentRetries += 1;
+        appendDanglingToolIntentCorrection(conversation, text);
+        continue;
+      }
       return {
-        text: typeof message.content === "string" ? message.content : "",
+        text,
         usage: sawUsage ? usageAcc : undefined,
         raw: { responses: rawResponses, toolCallsUsed: totalToolCalls },
         toolEvents,
@@ -914,6 +963,7 @@ export async function* runToolAwareOpenAIChatStream(
   const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
   let sawUsage = false;
   let totalToolCalls = 0;
+  let danglingToolIntentRetries = 0;
 
   for (let round = 0; round < MAX_TOOL_ROUNDS; round += 1) {
     const stream = await params.client.chat.completions.create({
@@ -938,9 +988,6 @@ export async function* runToolAwareOpenAIChatStream(
       const deltaText = choice?.delta?.content ?? "";
       if (typeof deltaText === "string" && deltaText.length) {
         roundText += deltaText;
-        if (roundToolCalls.size === 0) {
-          yield { type: "delta", text: deltaText };
-        }
       }
 
       const deltaToolCalls = Array.isArray(choice?.delta?.tool_calls) ? choice.delta.tool_calls : [];
@@ -969,6 +1016,14 @@ export async function* runToolAwareOpenAIChatStream(
       }));
 
     if (!normalizedToolCalls.length) {
+      if (danglingToolIntentRetries < MAX_DANGLING_TOOL_INTENT_RETRIES && looksLikeDanglingToolIntent(roundText)) {
+        danglingToolIntentRetries += 1;
+        appendDanglingToolIntentCorrection(conversation, roundText);
+        continue;
+      }
+      if (roundText) {
+        yield { type: "delta", text: roundText };
+      }
       yield {
         type: "done",
         result: {
@@ -982,7 +1037,7 @@ export async function* runToolAwareOpenAIChatStream(
     }
 
     totalToolCalls += normalizedToolCalls.length;
-    conversation.push({
+    const assistantToolCallMessage: any = {
       role: "assistant",
       tool_calls: normalizedToolCalls.map((call) => ({
         id: call.id,
@@ -992,7 +1047,11 @@ export async function* runToolAwareOpenAIChatStream(
           arguments: call.arguments,
         },
       })),
-    });
+    };
+    if (roundText) {
+      assistantToolCallMessage.content = roundText;
+    }
+    conversation.push(assistantToolCallMessage);
 
     for (const call of normalizedToolCalls) {
       const { event, toolResult } = await executeToolCallAndBuildEvent(call, params);