zames_pro 2.15.3 → 2.15.5

package/dist/agent-loop.js

@@ -4,7 +4,7 @@ import { getGitContext, formatGitContext } from './gitTools.js';
 import { parseXmlToolCalls } from './xml-toolcall.js';
 import { translate } from './i18n.js';
 import { normText } from './browser.js';
-export async function runAgentLoop({ browser, tools, task, workdir, maxIterations = 200, freshChat = false, sendSystemPrompt = false, transcript = null, attachments = [], onThinking = () => { }, onAssistantThought = () => { }, onToolCall = () => { }, onToolResult = () => { }, onAssistantMessage = () => { }, onChatReady = () => { }, onWarning = () => { }, debugLog = false, locale = 'ru', }) {
+export async function runAgentLoop({ browser, tools, task, workdir, maxIterations = 0, freshChat = false, sendSystemPrompt = false, transcript = null, attachments = [], onThinking = () => { }, onAssistantThought = () => { }, onToolCall = () => { }, onToolResult = () => { }, onAssistantMessage = () => { }, onChatReady = () => { }, onWarning = () => { }, debugLog = false, locale = 'ru', }) {
     // UI callbacks must NEVER break the agent loop. A rendering error (a huge
     // tool result, a broken markdown frame, a closed terminal) used to throw
     // out of the loop right after a tool call — the session looked "stopped
@@ -83,8 +83,11 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
             commands: context?.commands.map((c) => c.name) ?? [],
         });
         safeThinking();
-        // system-prompt is an agent send: throttled (agent: true).
-        await browser.ask(systemPrompt, { timeout: 60_000, agent: true });
+        // system-prompt is the FIRST message of a fresh chat: the rate limit only
+        // applies to a rapid back-and-forth, so this send is not throttled
+        // (agent: false). Throttling it used to add a useless 15s pause at the
+        // start of every new session.
+        await browser.ask(systemPrompt, { timeout: 60_000, agent: false });
         await reportChat();
     }
     let message = task;
@@ -150,8 +153,14 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
     // not come back in time. This is separate from unparsedRetries because a
     // timeout is an infrastructure failure, not a model protocol violation.
     let afterToolRetries = 0;
+    // toolsRanInTask: how many tools actually ran in THIS task. The protocol
+    // guard uses it to tell "started, then slipped into chat mode" (a reasoning
+    // paragraph that is neither a tool call nor a real respond) from a genuine
+    // short answer. We key on the STRUCTURE (work already started), not words.
+    let toolsRanInTask = 0;
     const MAX_AFTER_TOOL_RETRIES = 6;
-    for (let i = 0; i < maxIterations; i++) {
+    const iterCap = maxIterations > 0 ? maxIterations : 100000;
+    for (let i = 0; i < iterCap; i++) {
         // NOTE: the spinner is NOT started here. browser.onSendStart fires it
         // right when the message is actually typed/sent (after the send-pause),
         // so no spinner runs during the pre-send phase.
@@ -413,6 +422,26 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                         '{\"tool\": \"respond\", \"args\": {\"message\": \"...\"}}';
                 continue;
             }
+            // STRUCTURAL guard against "the model slipped into chat mode":
+            // after work has already started (a tool really ran in this task), a
+            // plain-text answer that is neither a tool call nor a real respond must
+            // NOT be returned to the operator as the final result. Returning it is
+            // exactly the "agent stopped mid-task" symptom (e.g. it writes a reasoning
+            // paragraph "Now let me analyze..." instead of calling a tool). We do not
+            // match words here - the STRUCTURE (toolsRanInTask > 0) is the signal.
+            if (toolsRanInTask > 0) {
+                transcript?.log('protocol_violation_final', {
+                    response: rawResponse.slice(0, 500),
+                });
+                safeWarning(translate(locale)('msg.suspicious_stop'));
+                // The model stopped calling tools mid-task. Surface the last text as
+                // a READABLE report, but say explicitly that the task may be incomplete:
+                // never let a reasoning paragraph masquerade as a finished result.
+                const lastText = (rawResponse || '').trim();
+                return (lastText
+                    ? lastText + String.fromCharCode(10) + String.fromCharCode(10) + '(The model stopped calling tools before finishing. The task may be incomplete - check the DeepSeek chat.)'
+                    : 'The model stopped calling tools before finishing the task.');
+            }
             const suspiciousFinal = responseLooksLikeToolCall(rawResponse) ||
                 looksLikeUnfinishedWork((rawResponse || '').trim()) ||
                 !(rawResponse || '').trim();
@@ -514,6 +543,7 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
         // A tool just ran — the next answer is expected to be a fresh tool call.
         // Reset the watchdog so the next empty/repeated answer is nudged.
         justRanTool = true;
+        toolsRanInTask++;
         // The tool really executed on this turn. The NEXT iteration will treat
         // a repeat of this answer as stale only because of this flag (see NEXT,
         // not the current, justRanTool check).

package/dist/browser.js

@@ -144,7 +144,7 @@ export class DeepSeekBrowser {
     // a generation really begins, not during the pre-send phase (chat open,
     // throttle wait), which used to show a spinner with no work in flight.
     onSendStart;
-    constructor({ headless = false, debug = false, channel = 'chrome', answerTimeoutMs = 180000, askRetries = 3, stabilityChecks = 3, stabilityDelayMs = 1000, minSendIntervalMs = 15000, rateLimitWaitMs = 300000, maxRateLimitRetries = 6, maxServerBusyRetries = 5, serverBusyWaitMs = 3000, } = {}) {
+    constructor({ headless = false, debug = false, channel = 'chrome', answerTimeoutMs = 180000, askRetries = 3, stabilityChecks = 2, stabilityDelayMs = 400, minSendIntervalMs = 15000, rateLimitWaitMs = 300000, maxRateLimitRetries = 6, maxServerBusyRetries = 5, serverBusyWaitMs = 3000, } = {}) {
         this.headless = headless;
         this.debug = debug;
         this.channel = channel;
@@ -180,7 +180,10 @@ export class DeepSeekBrowser {
     async _launchOnce() {
         const options = {
             headless: this.headless,
-            slowMo: 30,
+            // 30ms slowMo per Playwright action added up over thousands of DOM
+            // actions per run. The waits we need are explicit; keep a small value
+            // for stability of clicks/typing.
+            slowMo: 10,
             args: ['--disable-blink-features=AutomationControlled'],
         };
         if (this.channel)
@@ -694,7 +697,7 @@ export class DeepSeekBrowser {
             await this._attachFiles(attachments);
         }
         await this._setInputText(input, prompt);
-        await this.page.waitForTimeout(200);
+        await this.page.waitForTimeout(50);
         let sent = false;
         for (const sel of SEND_SELECTORS) {
             const btn = this.page.locator(sel).last();
@@ -867,7 +870,11 @@ export class DeepSeekBrowser {
             const sameAsBefore = !!cur && !isNew && normText(cur) === normText(beforeText);
             if (isNew && cur === last) {
                 stable++;
-                if (stable >= 2) {
+                // `stabilityChecks` / `stabilityDelayMs` are the real knobs here.
+                // They used to be dead config (hardcoded 2 checks with an 800ms tick),
+                // so the FIN-loop always cost ~1.6s per answer. Defaults are now
+                // 2 checks x 400ms (~0.8s saved per turn) and the values are honored.
+                if (stable >= Math.max(1, this.stabilityChecks - 1)) {
                     if (isNew || !(await this._isGenerating())) {
                         this._askDebug('RETURN stable curLen=' + cur.length);
                         return cur;
@@ -880,7 +887,7 @@ export class DeepSeekBrowser {
             if (isNew)
                 last = cur;
             this._askDebug('FIN-loop isNew=' + isNew + ' sameAsBefore=' + sameAsBefore + ' stable=' + stable + ' curLen=' + cur.length + ' lastLen=' + last.length + ' netFresh=' + netFresh);
-            await this.page.waitForTimeout(800);
+            await this.page.waitForTimeout(Math.max(0, this.stabilityDelayMs));
         }
         if (last && (normText(last) !== normText(beforeText) || this._netCapture)) {
             this._askDebug('RETURN last len=' + last.length);

package/dist/config.js

@@ -6,7 +6,7 @@ const ZAMES_HOME = path.join(os.homedir(), '.zames');
 const HOME_CONFIG = path.join(ZAMES_HOME, 'config.json');
 const PROJECT_CONFIG = path.join(process.cwd(), '.zamesrc.json');
 export const DEFAULTS = {
-    maxIterations: 200,
+    maxIterations: 0,
     headless: false,
     debug: false,
     browserChannel: null,
@@ -36,8 +36,8 @@ export const DEFAULTS = {
     browser: {
         answerTimeoutMs: 180000,
         askRetries: 3,
-        stabilityChecks: 3,
-        stabilityDelayMs: 1000,
+        stabilityChecks: 2,
+        stabilityDelayMs: 400,
         minSendIntervalMs: 15000,
         rateLimitWaitMs: 300000,
         maxRateLimitRetries: 6,
@@ -77,7 +77,7 @@ function deepMerge(target, source) {
 }
 export const CONFIG_SCHEMA = [
     { path: 'ui.locale', type: 'enum', values: ['ru', 'en'], labelKey: 'cfg.f.ui_locale', groupKey: 'cfg.group.ui' },
-    { path: 'maxIterations', type: 'number', min: 1, max: 100000, labelKey: 'cfg.f.maxIterations', groupKey: 'cfg.group.agent' },
+    { path: 'maxIterations', type: 'number', min: 0, max: 100000, labelKey: 'cfg.f.maxIterations', groupKey: 'cfg.group.agent' },
     { path: 'headless', type: 'boolean', labelKey: 'cfg.f.headless', groupKey: 'cfg.group.agent' },
     { path: 'debug', type: 'boolean', labelKey: 'cfg.f.debug', groupKey: 'cfg.group.agent' },
     { path: 'hotReload', type: 'boolean', labelKey: 'cfg.f.hotReload', groupKey: 'cfg.group.agent' },

package/dist/i18n.js

@@ -45,7 +45,7 @@ const CATALOG = {
     'help.opt.resume_last': { ru: 'вернуться в последний сохранённый чат', en: 'resume the last saved chat' },
     'help.opt.new_chat': { ru: 'начать новый чат (по умолчанию)', en: 'start a new chat (default)' },
     'help.opt.resend_prompt': { ru: 'дослать system-prompt в существующий чат', en: 'resend system-prompt into an existing chat' },
-    'help.opt.max_iter': { ru: 'лимит итераций (по умолчанию {n})', en: 'iteration limit (default {n})' },
+    'help.opt.max_iter': { ru: 'лимит итераций, 0 = без лимита ({n})', en: 'iteration limit, 0 = unlimited ({n})' },
     'help.opt.headless': { ru: 'браузер без UI', en: 'headless browser' },
     'help.opt.debug': { ru: 'подробный лог', en: 'verbose log' },
     'help.opt.calibrate': { ru: 'режим калибровки селекторов', en: 'selector calibration mode' },
@@ -144,7 +144,7 @@ const CATALOG = {
     'status.last_chat': { ru: 'Last chat: {v}', en: 'Last chat: {v}' },
     'status.sessions': { ru: 'Сессии: {v}', en: 'Sessions: {v}' },
     'status.dev': { ru: 'Dev mode (auto-reload): {v}', en: 'Dev mode (auto-reload): {v}' },
-    'status.max_iter': { ru: 'Лимит итераций: {v}', en: 'Iteration limit: {v}' },
+    'status.max_iter': { ru: 'Лимит итераций: {v} (0 = без лимита)', en: 'Iteration limit: {v} (0 = unlimited)' },
     'status.headless': { ru: 'Headless: {v}', en: 'Headless: {v}' },
     'status.debug': { ru: 'Debug: {v}', en: 'Debug: {v}' },
     'status.undo': { ru: 'Undo: {v}', en: 'Undo: {v}' },

package/dist/index.js

@@ -53,8 +53,8 @@ const t = (key, params) => translate(currentLocale)(key, params);
 const headless = hasFlag('--headless') || config.headless;
 const debug = hasFlag('--debug') || config.debug;
 const calibrate = hasFlag('--calibrate');
-const maxIter = Number(getArg('--max-iter', String(config.maxIterations))) ||
-    config.maxIterations;
+const maxIterArg = getArg('--max-iter', null);
+const maxIter = maxIterArg !== null ? Number(maxIterArg) : config.maxIterations;
 const positional = getPositional();
 const task = getArg('--task', positional.join(' ').trim() || null);
 const chatIdArg = getArg('--chat', null);

package/dist/net-capture.js

@@ -133,7 +133,12 @@ export function extractAnswer(body) {
 }
 // Saves the DeepSeek network response body to disk for post-mortem analysis.
 // The files live in ~/.zames/net-log — from them the real answer format is visible.
+// DEBUG ONLY: disabled unless ZAMES_NET_DEBUG=1. It writes a file per network
+// response (thousands of files / tens of MB) and is not needed for the agent
+// to work — the answer is taken from extractAnswer() in memory.
 export async function dumpNetBody(url, body) {
+    if (!process.env.ZAMES_NET_DEBUG)
+        return;
     try {
         const dir = path.join(os.homedir(), '.zames', 'net-log');
         await fs.mkdir(dir, { recursive: true });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zames_pro",
-  "version": "2.15.3",
+  "version": "2.15.5",
   "description": "Terminal coding agent over chat.deepseek.com via Playwright",
   "type": "module",
   "bin": {