@nanhara/hara 0.109.2 → 0.109.3

package/CHANGELOG.md

@@ -5,6 +5,19 @@ All notable changes to `@nanhara/hara`.
 > Versioning (pre-1.0, SemVer-style): the **minor** (middle) number bumps for a **new feature**; the
 > **patch** (last) number bumps for **optimizations/fixes of existing features**.
 
+## 0.109.3 — an empty model response no longer looks like a hang
+
+- **A turn that comes back with no text and no tool calls is no longer silently dropped.** The agent
+  loop treated an empty completion as "done" and returned with zero feedback — so after e.g. a "继续"
+  the box just sat there, idle, looking frozen for hours (CPU 0%, waiting at the prompt) when really
+  the turn had vanished. It now **retries once** with a nudge (an empty response is usually a transient
+  hiccup), and if it's still empty says so plainly (`✻ empty response — nothing to do. Rephrase…`)
+  instead of disappearing. Matches how Claude Code / codex refuse to end on a blank turn.
+- **Closed a matching spin:** a `tool_use` stop reason carrying an *empty* tool list used to push an
+  empty tool round and re-request in a loop; it's now bounded by the same single-retry guard. (The
+  120s stall-watchdog already covered a genuinely dead/silent socket — this covers the case where the
+  request *succeeds* but returns nothing.)
+
 ## 0.109.2 — long paste no longer freezes the input box
 
 - **The input box now draws a bottom-anchored viewport, not every wrapped row.** A long multi-line

package/dist/agent/loop.js

@@ -97,6 +97,7 @@ export async function runAgent(history, opts) {
     const permRules = loadPermissionRules(ctx.cwd); // command-level allow/ask/deny policy for the bash tool
     let activeProvider = provider; // may switch to a fallback model on a recoverable error (app-failover)
     let triedFallback = false;
+    let emptyRetried = false; // one-shot: a genuinely empty model turn gets a single nudge before we give up
     // Stuck/loop guard — only in headless chat (`hara gateway`), where a wrong approach can grind forever with
     // nobody to hit Esc (e.g. screenshots it can't read). Once per run, when the agent keeps repeating one
     // non-read tool or acting blind, we inject a reflection nudge so it steps back instead of spinning.
@@ -281,7 +282,37 @@ export async function runAgent(history, opts) {
             }
             return;
         }
-        if (r.stop !== "tool_use")
+        // Empty-turn guard. The model returned nothing actionable — no text AND no tool calls (a blank
+        // completion, or a "tool_use" stop with an empty tool list). Silently returning here leaves the
+        // user at a dead prompt with ZERO feedback: it reads as a 15-hour hang when really the turn just
+        // vanished. Retry ONCE with a nudge (usually a transient hiccup), then, if still empty, say so
+        // plainly and end — never loop forever, never disappear. (Claude Code / codex both guard this.)
+        if (!r.text.trim() && r.toolUses.length === 0) {
+            if (!emptyRetried) {
+                emptyRetried = true;
+                history.pop(); // drop the empty assistant turn before re-asking
+                history.push({ role: "user", content: "(Your previous response was empty. Continue the task now — take the next concrete step with a tool, or reply with text. Do not return an empty response.)" });
+                if (!opts.quiet) {
+                    const note = "✻ empty response — retrying once…";
+                    if (sink)
+                        sink.notice(note);
+                    else
+                        out(c.dim(`${note}\n`));
+                }
+                continue;
+            }
+            const note = "✻ the model returned an empty response — nothing to do. Rephrase your request, or press Enter to try again.";
+            if (!opts.quiet) {
+                if (sink)
+                    sink.notice(note);
+                else
+                    out(c.dim(`${note}\n`));
+            }
+            return;
+        }
+        // A "tool_use" stop with text but no tools (rare) has nothing to execute — end after showing the text
+        // rather than pushing an empty tool round and re-requesting in a loop.
+        if (r.stop !== "tool_use" || r.toolUses.length === 0)
             return;
         const plans = [];
         // Extra (per-run) tools win over the registry so a run-scoped tool can't be shadowed by a global one.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nanhara/hara",
-  "version": "0.109.2",
+  "version": "0.109.3",
   "description": "hara — a coding agent CLI that runs like an engineering org.",
   "bin": {
     "hara": "dist/index.js"