Simplify tool approval chat flow

duck_core/api.py

@@ -508,13 +508,6 @@ def create_app() -> FastAPI:
                 )
                 messages = await runtime.context_builder.build_async_messages(task)
                 tool_observations = [tool_observation]
-                if approval.decision != "deny" and not has_password_request(tool_observations):
-                    tool_observations = await runtime._run_action_loop(
-                        task_id,
-                        messages,
-                        task.workspace,
-                        initial_observations=tool_observations,
-                    )
                 async for tool_event in emit_tool_events(task_id, continued_event.sequence):
                     yield tool_event
                 if has_password_request(tool_observations):
@@ -696,13 +689,6 @@ def create_app() -> FastAPI:
                 )
                 messages = await runtime.context_builder.build_async_messages(task)
                 tool_observations = [tool_observation]
-                if not has_password_request(tool_observations):
-                    tool_observations = await runtime._run_action_loop(
-                        task_id,
-                        messages,
-                        task.workspace,
-                        initial_observations=tool_observations,
-                    )
                 async for tool_event in emit_tool_events(task_id, continued_event.sequence):
                     yield tool_event
                 if has_password_request(tool_observations):

duck_core/web/static/app.js

@@ -175,35 +175,39 @@ function formatToolCommand(tool, args) {
 }
 
 function formatToolStart(tool, args) {
-  const lines = [formatToolCommand(tool, args)];
+  return formatToolCommand(tool, args);
-  const serializedArgs = JSON.stringify(args || {}, null, 2);
-  if (serializedArgs !== "{}") lines.push(serializedArgs);
-  return lines.join("\n");
 }
 
 function appendToolTerminal(article, eventPayload) {
   const paragraph = article?.querySelector("p");
+  const payload = eventPayload.payload || eventPayload;
+  const existing = findToolTerminal(article, payload.index);
+  if (existing) return existing;
   const terminal = createToolTerminal(eventPayload);
   paragraph?.before(terminal);
   document.querySelector("#messages").scrollTop = document.querySelector("#messages").scrollHeight;
+  return terminal;
+}
+
+function findToolTerminal(article, index) {
+  return article?.querySelector(`.tool-terminal[data-tool-index="${index || ""}"]`);
 }
 
 function updateToolTerminal(article, eventPayload) {
   const payload = eventPayload.payload || eventPayload;
-  const terminal = article?.querySelector(`.tool-terminal[data-tool-index="${payload.index || ""}"]`);
+  const terminal = findToolTerminal(article, payload.index);
   const body = terminal?.querySelector(".tool-terminal-body");
   const status = terminal?.querySelector(".tool-terminal-status");
   const result = payload.result || {};
   if (!body || !status) return;
   terminal.classList.toggle("is-error", !result.ok);
+  terminal.classList.remove("is-waiting");
   status.textContent = result.ok ? "ok" : "error";
 
-  const parts = [body.textContent.trim()];
+  const title = terminal.querySelector(".tool-terminal-title")?.textContent || body.textContent.trim();
+  const parts = [title];
   if (result.output) parts.push("\nstdout\n" + result.output.trimEnd());
   if (result.error) parts.push("\nstderr\n" + result.error.trimEnd());
-  if (result.metadata && Object.keys(result.metadata).length) {
-    parts.push("\nmetadata\n" + JSON.stringify(result.metadata, null, 2));
-  }
   body.textContent = parts.join("\n");
   document.querySelector("#messages").scrollTop = document.querySelector("#messages").scrollHeight;
 }
@@ -211,14 +215,13 @@ function updateToolTerminal(article, eventPayload) {
 function appendApprovalTerminal(article, eventPayload) {
   const payload = eventPayload.payload || eventPayload;
   const action = payload.action || {};
-  appendToolTerminal(article, {
+  const terminal = appendToolTerminal(article, {
     payload: {
       index: payload.index,
       tool: payload.tool,
       args: action.args || {},
     },
   });
-  const terminal = article?.querySelector(`.tool-terminal[data-tool-index="${payload.index || ""}"]`);
   const body = terminal?.querySelector(".tool-terminal-body");
   const status = terminal?.querySelector(".tool-terminal-status");
   const command = formatToolCommand(action.tool || payload.tool, action.args || {});
@@ -228,14 +231,15 @@ function appendApprovalTerminal(article, eventPayload) {
   if (status) status.textContent = "approval";
   if (body) {
     body.textContent = [
-      body.textContent,
-      "",
-      "approval required",
       command,
+      "",
+      "Требуется разрешение на выполнение.",
       payload.reason || "",
     ].filter(Boolean).join("\n");
   }
-  terminal?.append(createInlineApprovalActions(payload.approval_id, command));
+  if (!terminal?.querySelector(".tool-approval-actions")) {
+    terminal?.append(createInlineApprovalActions(payload.approval_id, command));
+  }
 }
 
 function appendPasswordPrompt(article, eventPayload) {
@@ -318,9 +322,13 @@ async function resolveInlineApproval(button) {
   const decision = action === "deny" ? "denied" : "allowed";
 
   terminal.classList.toggle("is-error", action === "deny");
-  terminal.classList.remove("is-waiting");
+  terminal.classList.toggle("is-waiting", action !== "deny");
-  if (status) status.textContent = decision;
+  if (status) status.textContent = action === "deny" ? decision : "running";
-  if (body) body.textContent = `${command}\n\n${decision}: ${humanApprovalDecision(action)}`;
+  if (body) {
+    body.textContent = action === "deny"
+      ? `${command}\n\n${decision}: ${humanApprovalDecision(action)}`
+      : `${command}\n\nРазрешено. Выполняю команду...`;
+  }
   actions?.remove();
   if (taskId) {
     await continueAfterInlineApproval(terminal.closest(".message"), taskId, approvalId);

prompts/roles/thinker.md

@@ -5,3 +5,7 @@ When asked who or what you are, answer as DuckLM. You may mention that DuckLM is
 powered by a local model, but do not answer as the base model identity.
 
 Respond clearly and briefly unless the task needs detail.
+Do not reveal hidden reasoning, chain-of-thought, drafts, or step-by-step
+internal analysis. Return only the user-facing answer. If tool observations are
+present, summarize the useful result without echoing raw tool JSON or metadata
+unless the user explicitly asks for raw output.

tests/smoke/test_api_stream_chat.py

@@ -110,9 +110,12 @@ def test_stream_chat_endpoint_executes_tool_before_streaming_answer(tmp_path, mo
 
 def test_continue_stream_executes_approved_tool_and_streams_answer(tmp_path, monkeypatch):
     monkeypatch.setenv("DUCK_DB_PATH", str(tmp_path / "duck.sqlite3"))
+    action_calls = 0
 
     async def fake_chat(self, role, messages, temperature=None, max_output_tokens=None, response_format=None):
+        nonlocal action_calls
         assert role == "action"
+        action_calls += 1
         if any("tool_observations" in message["content"] for message in messages):
             actions = []
         else:
@@ -173,6 +176,7 @@ def test_continue_stream_executes_approved_tool_and_streams_answer(tmp_path, mon
     assert "event: content_delta" in body
     assert "continued after approval" in body
     assert "event: done" in body
+    assert action_calls == 1
     conversation_id = re.search(r'"conversation_id"\s*:\s*"([^"]+)"', initial_body).group(1)
     conversation = client.get(f"/v1/conversations/{conversation_id}").json()
     assert conversation["messages"][-1]["content"] == "continued after approval"