zames_pro 2.94.5 → 2.94.7

package/CHANGELOG.md

@@ -7,6 +7,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [2.94.7] - 2026-11-03
+
+### Fixed
+
+- **deepseek:** recognise the server-busy notice from the SSE body
+
+## [2.94.6] - 2026-11-03
+
+### Fixed
+
+- **ui:** clear the generating badge as soon as an answer comes back
+
 ## [2.94.5] - 2026-11-03
 
 ### Changed

package/dist/agent-loop.js

@@ -49,6 +49,7 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
     // Coarse lifecycle of the send (generating/paused/settled) for the status
     // line. Same best-effort wiring as the other hooks.
     browser.onSendState = safe(onSendState);
+    const safeSendState = safe(onSendState);
     // Service notices (rate limit, server busy, resend) must reach the OPERATOR.
     // While a LineEditor is active a raw console.error is overwritten by the
     // editor's repaint, so the operator only saw the spinner. `onNotice` prints
@@ -425,6 +426,12 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
         }
         await reportChat();
         transcript?.log('assistant_raw', { response: rawResponse });
+        // A returned answer is a FINISHED generation: clear the "generating" badge
+        // NOW, not when the next ask() happens to start. The operator used to see
+        // "▶ генерация" for minutes after the answer had already been printed
+        // (the tool-call parsing, a long Bash run) - it looked like the agent was
+        // still generating with nothing in flight.
+        safeSendState('settled');
         // The user aborted generation (Esc/Ctrl+C).
         if (/^\s*\(прервано пользователем\)\s*$/.test(rawResponse)) {
             transcript?.log('user_aborted');

package/dist/browser.js

@@ -117,6 +117,10 @@ export class DeepSeekBrowser {
     // threw ds.send_no_new_answer while the chat plainly showed the rate limit.
     // Reset before a send.
     _netRateLimited;
+    // True when the LAST completion/continue body carried the server-busy
+    // notice ("Server busy, please try again later"). It arrives WITHOUT a
+    // DOM toast, so the SSE body is the authoritative source. Reset per send.
+    _netServerBusy;
     _netChatId;
     // The latest CONTEXT size (in tokens) DeepSeek reported in the current
     // chat. It is the `accumulated_token_usage` counter taken from the SSE
@@ -228,6 +232,7 @@ export class DeepSeekBrowser {
         this._netCaptureAt = 0;
         this._netNoAnswer = false;
         this._netRateLimited = false;
+        this._netServerBusy = false;
         this._netChatId = null;
         this._lastTokenUsage = null;
         this._toggles = { deepThinking, webSearch };
@@ -537,6 +542,13 @@ export class DeepSeekBrowser {
                         this._netRateLimited = true;
                         this._netCapture = '';
                     }
+                    // The SERVER-BUSY notice can arrive ONLY in the body: the toast is
+                    // short-lived, and without this the loop waited out the deadline and
+                    // reported "no answer" for a turn the server had refused.
+                    if (isServerBusyText(body)) {
+                        this._netServerBusy = true;
+                        this._netCapture = '';
+                    }
                     this._netCaptureAt = Date.now();
                 }
             }
@@ -2198,6 +2210,7 @@ export class DeepSeekBrowser {
         this._netCaptureAt = 0;
         this._netNoAnswer = false;
         this._netRateLimited = false;
+        this._netServerBusy = false;
         this._netBody = '';
         // Align the DeepSeek chat toggles (deep thinking / web search) with the
         // configured state BEFORE typing. Doing it here (after the send-pause) it
@@ -2324,6 +2337,12 @@ export class DeepSeekBrowser {
             if (this._netRateLimited && this._netCaptureAt >= this._lastSentAt) {
                 throw new RateLimitError(this._netBody.slice(0, 300) || 'rate limit (from network)');
             }
+            // Body-only server-busy notice: retry instead of waiting out the
+            // whole deadline (a live chat showed it while the agent reported
+            // "no answer").
+            if (this._netServerBusy && this._netCaptureAt >= this._lastSentAt) {
+                throw new ServerBusyError(this._netBody.slice(-300));
+            }
             // The turn was truncated by the server: the SSE body carries
             // finish_reason=generation_err / quasi_status=INCOMPLETE and the UI
             // shows "Continue". Retry instead of waiting out the whole timeout.
@@ -2579,6 +2598,12 @@ export class DeepSeekBrowser {
             if (this._netRateLimited && this._netCaptureAt >= this._lastSentAt) {
                 throw new RateLimitError(this._netBody.slice(0, 300) || 'rate limit (from network)');
             }
+            // Body-only server-busy notice: retry instead of waiting out the
+            // whole deadline (a live chat showed it while the agent reported
+            // "no answer").
+            if (this._netServerBusy && this._netCaptureAt >= this._lastSentAt) {
+                throw new ServerBusyError(this._netBody.slice(-300));
+            }
             // The exhausted context can likewise be visible only in the captured
             // body rather than in a toast, so ask() must see it here as well.
             if (this._netCaptureAt >= this._lastSentAt &&

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zames_pro",
-  "version": "2.94.5",
+  "version": "2.94.7",
   "description": "Terminal coding agent that drives a real browser to chat with DeepSeek, Qwen or GLM — no API key required. Reads and edits files, runs commands, commits to git, like Claude Code / Codex CLI.",
   "type": "module",
   "bin": {