zames_pro 2.29.3 → 2.29.5

package/dist/browser.js

@@ -340,6 +340,13 @@ export class DeepSeekBrowser {
     _lastHistoryError;
     _netSniff;
     _netSniffLimit;
+    // The RAW SSE body of the LAST answer response. Unlike `_netCapture`, this
+    // is kept even when the body carries no RESPONSE text (a truncated turn:
+    // generation_err / INCOMPLETE, or FINISHED-with-only-reasoning). The
+    // incomplete/no-answer detectors MUST run on the raw body — guarding them on
+    // `_netCapture` (which is '' in exactly those cases) made them dead code and
+    // the agent hung until the timeout.
+    _netBody;
     _netHookInstalled;
     // Fired right when a message is actually typed/sent (AFTER the send-pause
     // and attachments). Used to start the "agent is working" spinner only when
@@ -393,6 +400,7 @@ export class DeepSeekBrowser {
         this._lastHistoryError = '';
         this._netSniff = [];
         this._netSniffLimit = 5;
+        this._netBody = '';
         this._netHookInstalled = false;
         this.onSendStart = null;
         this.onSendPause = null;
@@ -554,17 +562,31 @@ export class DeepSeekBrowser {
                     this._netChatId = cid[0];
                 this._netCaptureAt = Date.now();
             }
-            else if (/chat\/(completion|continue)/i.test(url) &&
-                isFinishedWithoutAnswer(body)) {
-                // A FINISHED turn with NO RESPONSE fragment (only reasoning): the UI
-                // shows a Continue button. Record it with a fresh timestamp so the
-                // ask() loops do not wait out the timeout — they click Continue.
-                this._netNoAnswer = true;
-                this._netCapture = '';
+            // Store the RAW body for the CURRENT answer (completion/continue) so the
+            // truncated/no-answer detectors can inspect it even when there is no
+            // RESPONSE text (extractAnswer() returns '' for those bodies).
+            if (/chat\/(completion|continue)/i.test(url)) {
+                this._netBody = body;
                 const cid = url.match(/[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}/);
                 if (cid)
                     this._netChatId = cid[0];
-                this._netCaptureAt = Date.now();
+                if (!extracted) {
+                    if (isFinishedWithoutAnswer(body)) {
+                        // A FINISHED turn with NO RESPONSE fragment (only reasoning): the
+                        // UI shows Stopped + Continue. Record it with a fresh timestamp
+                        // so the ask() loops do not wait out the timeout — they click
+                        // Continue / resend.
+                        this._netNoAnswer = true;
+                        this._netCapture = '';
+                    }
+                    // A truncated turn (generation_err / INCOMPLETE) also has no
+                    // RESPONSE; make the body visible to the incomplete check by
+                    // refreshing the timestamp.
+                    if (isGenerationIncompleteText(body)) {
+                        this._netCapture = '';
+                    }
+                    this._netCaptureAt = Date.now();
+                }
             }
         }
         catch { }
@@ -1652,6 +1674,7 @@ export class DeepSeekBrowser {
         this._netCapture = '';
         this._netCaptureAt = 0;
         this._netNoAnswer = 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
         // does not flip toggles for a generation that is not going to happen.
@@ -1711,10 +1734,13 @@ export class DeepSeekBrowser {
             // 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.
-            if (this._netCapture &&
-                this._netCaptureAt >= this._lastSentAt &&
-                isGenerationIncompleteText(this._netCapture)) {
-                throw new GenerationIncompleteError(this._netCapture.slice(-300));
+            // NOTE: the check runs on the RAW `_netBody`, not on `_netCapture`: a
+            // truncated body has NO RESPONSE fragment, so extractAnswer() returned
+            // '' and `_netCapture` was empty — guarding on it made this branch dead
+            // code and the agent hung on "Stopped" until the timeout.
+            if (this._netCaptureAt >= this._lastSentAt &&
+                isGenerationIncompleteText(this._netBody)) {
+                throw new GenerationIncompleteError(this._netBody.slice(-300));
             }
             // A FINISHED turn with NO answer (only reasoning): the UI shows a
             // Continue button. Click it (or, if autoContinue is off, keep waiting)
@@ -1805,6 +1831,14 @@ export class DeepSeekBrowser {
             // previous tool call.
             const cur = await this._readLastAnswerTextCleanDom().catch(() => '');
             const fresh = !!this._netCapture && this._netCaptureAt >= this._lastSentAt;
+            // The turn was truncated / produced no answer (generation_err,
+            // INCOMPLETE, or FINISHED-with-only-reasoning): the message WAS sent,
+            // the server just cut the turn. Resend (ask() handles it) instead of
+            // treating this as "send failed" and hammering Enter.
+            if (this._netCaptureAt >= this._lastSentAt &&
+                (isGenerationIncompleteText(this._netBody) || this._netNoAnswer)) {
+                throw new GenerationIncompleteError(this._netBody.slice(-300));
+            }
             if (cur &&
                 cur.trim() &&
                 normText(cur) !== normText(beforeText) &&
@@ -1875,10 +1909,11 @@ export class DeepSeekBrowser {
             // keeps the partial answer, so without this check the loop waits out the
             // whole timeout and throws ds.send_no_new_answer — the operator sees the
             // agent "stop" with a Continue button in the chat. Retry the send.
-            if (this._netCapture &&
-                this._netCaptureAt >= this._lastSentAt &&
-                isGenerationIncompleteText(this._netCapture)) {
-                throw new GenerationIncompleteError(this._netCapture.slice(-300));
+            // Runs on the RAW `_netBody` (a truncated body has no RESPONSE, so
+            // `_netCapture` is empty and the old guard was dead code).
+            if (this._netCaptureAt >= this._lastSentAt &&
+                isGenerationIncompleteText(this._netBody)) {
+                throw new GenerationIncompleteError(this._netBody.slice(-300));
             }
             // A FINISHED turn with NO answer (only reasoning): the UI shows a
             // Continue button. Click it and keep waiting — otherwise the loop spins

package/dist/commands.js

@@ -197,60 +197,115 @@ export function resolveExtraDir(input, workdir) {
     }
     return { path: abs };
 }
-/** How many characters of the dialogue are printed by default. */
+/** How many messages of the dialogue are printed by default. */
 export const RESTORED_HISTORY_LIMIT = 20;
 /**
- * Is this message worth showing to the operator as part of the dialogue?
+ * DeepSeek stores the assistant turn VERBATIM: either a raw tool-call JSON
+ * ({"tool":"Read","args":{...}}, frequently dirty/truncated) or a `respond`
+ * call that carries the real message for the operator
+ * ({"tool":"respond","args":{"message":"..."}}). The operator only wants the
+ * `respond` message (and plain-text answers) — a tool-call is protocol noise.
  *
- * The chat history contains a lot of protocol noise that is meaningless in
- * the terminal: the system-prompt (a huge REQUEST), tool results
- * ("Tool result for ..."), the agent's raw tool-calls (JSON / DSML), the
- * "You stopped after a tool result..." nudges, and the agent's <system> notes.
- * We keep only real user turns and real assistant answers.
+ * Returns the tool name and, for `respond`, the unwrapped message. Tolerant on
+ * purpose: the stored JSON is often dirty (a stray `}}`, a missing "args"
+ * wrapper), so a bare JSON.parse is not enough.
  */
-export function isDisplayableMessage(m) {
+export function parseAssistantToolCall(text) {
+    const t = String(text || '').trim();
+    if (!t)
+        return null;
+    // Broken heads like `<|tool": ...` or `tool": "Bash"` are tolerated.
+    const m = t.match(/tool"?\s*:\s*"([^"\s]+)"/);
+    if (!m)
+        return null;
+    const tool = m[1];
+    if (tool !== 'respond')
+        return { tool };
+    const greedy = t.match(/"message"\s*:\s*"([\s\S]*)"\s*\}*\s*$/);
+    let raw = greedy ? greedy[1] : '';
+    if (!greedy) {
+        const idx = t.indexOf('"message"');
+        if (idx >= 0)
+            raw = t.slice(idx + 9).replace(/^\s*:\s*"/, '');
+    }
+    return { tool, message: unescapeJsonString(raw) };
+}
+/** Decode the JSON escape sequences that appear inside a captured string. */
+export function unescapeJsonString(s) {
+    return String(s || '')
+        .replace(/\\u([0-9a-fA-F]{4})/g, (_x, h) => String.fromCharCode(parseInt(h, 16)))
+        .replace(/\\n/g, NL)
+        .replace(/\\r/g, String.fromCharCode(13))
+        .replace(/\\t/g, String.fromCharCode(9))
+        .replace(/\\"/g, '"')
+        .replace(/\\\\/g, '\\');
+}
+/**
+ * Normalize one restored message for display:
+ *  - USER: drop the protocol noise (system-prompt, tool results, nudges) — it
+ *    is not the operator's words.
+ *  - ASSISTANT: unwrap a `respond` call to its message; drop any other
+ *    tool-call (protocol noise); keep plain-text answers.
+ * Returns null when the message must not be shown.
+ */
+export function normalizeRestoredMessage(m) {
     const text = String((m && m.text) || '').trim();
     if (!text)
-        return false;
+        return null;
     if (text.length > 100_000)
-        return false; // the system-prompt / a giant blob
+        return null; // the system-prompt / a giant blob
     if (m.role === 'user') {
-        // Tool output, the system-prompt and the corrective nudges are not the
-        // operator's words.
         if (/^Tool result for /.test(text))
-            return false;
+            return null;
         if (/^You are a coding agent running in a terminal/.test(text))
-            return false;
+            return null;
         if (/^You stopped after a tool result/.test(text))
-            return false;
+            return null;
         if (/^\[system\]/.test(text))
-            return false;
+            return null;
         if (/^The user ran \//.test(text))
-            return false;
-        return true;
+            return null;
+        return { role: 'user', text };
     }
-    // Assistant: a tool-call (JSON or DSML) is not an answer to show. The
-    // DSML markers use FULL-WIDTH vertical bars (｜｜DSML), so the regex must
-    // match them, not the ASCII pipe.
     if (/DSML/i.test(text))
-        return false;
+        return null;
     if (/^<system>/i.test(text))
-        return false;
-    if (/"tool"\s*:/.test(text.slice(0, 600))) {
-        return false;
+        return null;
+    // The model sometimes emits its own safety-classification block as a plain
+    // "answer"; it is not for the operator.
+    if (/<ds_safety>/i.test(text))
+        return null;
+    if (/^Safe$/i.test(text))
+        return null;
+    const call = parseAssistantToolCall(text);
+    if (call) {
+        if (call.tool !== 'respond')
+            return null;
+        const msg = (call.message || '').trim();
+        if (!msg)
+            return null;
+        return { role: 'assistant', text: msg };
     }
-    if (/^\s*[\[{]/.test(text) && /"args"\s*:/.test(text.slice(0, 600))) {
-        return false;
-    }
-    return true;
+    // A plain-text answer (written without the respond wrapper) — show as-is.
+    return { role: 'assistant', text };
+}
+/** Back-compat predicate: is the message worth showing at all? */
+export function isDisplayableMessage(m) {
+    return normalizeRestoredMessage(m) !== null;
 }
 /**
- * Keep only the last `limit` messages and drop empties/service noise. The
- * source (network history or DOM) may contain protocol messages; they are
- * filtered out here so the terminal output stays readable.
+ * Keep only the last `limit` messages, dropping protocol noise and UNWRAPPING
+ * the assistant `respond` calls to their operator-facing text.
  */
 export function trimRestoredMessages(messages, limit = RESTORED_HISTORY_LIMIT) {
-    const clean = (messages || []).filter((m) => m && typeof m.text === 'string' && isDisplayableMessage(m));
+    const clean = [];
+    for (const m of messages || []) {
+        if (!m || typeof m.text !== 'string')
+            continue;
+        const n = normalizeRestoredMessage(m);
+        if (n)
+            clean.push(n);
+    }
     if (limit > 0 && clean.length > limit)
         return clean.slice(clean.length - limit);
     return clean;

package/dist/index.js

@@ -828,6 +828,12 @@ async function printRestoredHistory(browser, ui, chatId = null) {
         all = await browser.readChatMessages().catch(() => []);
         domCount = all.length;
     }
+    // `all` is the raw history (mostly protocol noise). `displayable` is what
+    // survives normalization; `messages` is the last RESTORED_HISTORY_LIMIT of
+    // those. The "truncated" note must compare the DISPLAYABLE count with the
+    // limit (comparing `all.length` always fired, because tool-calls are
+    // dropped and the raw history is always longer).
+    const displayable = trimRestoredMessages(all, 0);
     const messages = trimRestoredMessages(all);
     // Always surface the diagnostic line when the history could not be turned
     // into anything printable — otherwise "the dialogue is empty" is a dead end
@@ -868,7 +874,7 @@ async function printRestoredHistory(browser, ui, chatId = null) {
                 NL);
         }
     }
-    if (all.length > messages.length) {
+    if (displayable.length > messages.length) {
         out(theme.dim(t('chats.history_truncated', { n: String(RESTORED_HISTORY_LIMIT) })));
     }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zames_pro",
-  "version": "2.29.3",
+  "version": "2.29.5",
   "description": "Terminal coding agent over chat.deepseek.com via Playwright",
   "type": "module",
   "bin": {