zames_pro 2.13.0 → 2.15.0

package/README.md

@@ -76,6 +76,45 @@ zames --version
 zames --help
 ```
 
+## Tools
+
+The agent has the same style of tools as Claude Code / Codex CLI:
+
+- **File tools** — `Read`, `Write`, `Edit`, `Bash`, `Glob`, `Grep`. `Read`
+  returns raw content by default; pass `numbered=true` to get `cat -n`-style
+  line numbers (for reference only — do not paste them into `Edit`).
+- **Extra tools** — `LS` (list a directory), `MultiEdit` (several edits to one
+  file applied atomically), `TodoWrite` (session task checklist), `ApplyPatch`
+  (multi-file patch in Codex's V4A format: `*** Begin Patch` … `*** End Patch`).
+- **Git** — `GitStatus`, `GitDiff`, `GitLog`, `GitAdd`, `GitCommit`, `GitPush`.
+- **Web** — `WebFetch`, `WebSearch`.
+- **Service** — `respond` (final answer to the operator, ends the task).
+
+All file tools stay inside the working directory (sandbox). `Write`/`Edit` and
+`MultiEdit`/`ApplyPatch` make a backup (undo) before touching a file.
+
+## Slash commands
+
+Type / in the prompt for hints (Tab completes). Besides the session and
+config commands (/new, /chats, /resume, /cd, /status, /config,
+/undo, /transcript, /mcp, /skills, /memory, /init, /reload,
+/debug-dom, /help, /exit) there are a few that mirror Claude Code /
+Codex CLI:
+
+- /diff [--staged] — show the working-tree git diff (--staged for the index).
+- /cost (alias /usage) — session stats: tasks, tool calls, duration
+ (DeepSeek web does not expose token counts).
+- /export [file] — write the session transcript to a Markdown file
+ (zames-export-<stamp>.md by default).
+- /doctor — diagnose node, git, config, browser, clipboard and MCP.
+- /permissions — show the confirmation settings (Write/Edit/Bash + the
+ alwaysConfirm regex list).
+- /add-dir <path> — validate an extra directory (the sandbox is fixed at
+ startup; relaunch with --dir to write there).
+- /review [focus] [--staged] — ask the agent to review uncommitted changes
+ and report findings (no code changes).
+
+
 ## Project context, skills and memory
 
 Like Codex / Claude Code, zames reads project instructions and reusable

package/dist/agent-loop.js

@@ -194,20 +194,20 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                 attempt: afterToolRetries,
                 error: e.message,
             });
-            safeWarning('browser.ask() не вернул ответ за ' +
+            safeWarning('browser.ask() did not return an answer within ' +
                 Math.round(askDeadlineMs / 1000) +
-                'с — повторяю запрос.');
+                's — retrying the request.');
             if (afterToolRetries < MAX_AFTER_TOOL_RETRIES) {
                 afterToolRetries++;
                 await new Promise((r) => setTimeout(r, 1500));
                 continue;
             }
             transcript?.log('ask_timeout_exhausted', {
-                message: 'ask() не вернул ответ и лимит повторов исчерпан',
+                message: 'ask() did not return an answer and the retry limit is exhausted',
             });
-            safeWarning('browser.ask() перестал отвечать; лимит повторов исчерпан, ' +
-                'останавливаюсь. Ответа модели нет — проверьте чат DeepSeek вручную.');
-            return 'ask() watchdog: ответ модели не получен';
+            safeWarning('browser.ask() stopped responding; retry limit exhausted, ' +
+                'stopping. No model answer — check the DeepSeek chat manually.');
+            return 'ask() watchdog: no model answer received';
         }
         await reportChat();
         transcript?.log('assistant_raw', { response: rawResponse });
@@ -262,10 +262,10 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                 response: String(rawResponse || '').slice(0, 200),
             });
             message =
-                'Ты остановился после результата инструмента. Продолжи работу: ' +
-                    'ответь РОВНО одним JSON-объектом вызова инструмента, без текста до и после, ' +
-                    'например: {"tool": "Bash", "args": {"command": "..."}}. ' +
-                    'Если задача действительно выполнена — вызови respond с итоговым сообщением.';
+                'You stopped after a tool result. Continue the work: ' +
+                    'reply with EXACTLY one JSON tool-call object, no text before or after, ' +
+                    'for example: {"tool": "Bash", "args": {"command": "..."}}. ' +
+                    'If the task is really done — call respond with the final message.';
             await new Promise((r) => setTimeout(r, 1500));
             continue;
         }
@@ -287,7 +287,7 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
         if (parsedCalls.some((p) => p && p._permissive)) {
             transcript?.log('permissive_parse', { response: rawResponse });
             if (debugLog) {
-                console.error('внимание: tool-call распознан нестрогим парсером');
+                console.error('warning: tool-call recognized by the permissive parser');
             }
         }
         if (!parsed) {
@@ -304,17 +304,17 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                     response: rawResponse,
                 });
                 if (debugLog) {
-                    console.error('внимание: ответ похож на tool-call, но не распознан (попытка ' +
+                    console.error('warning: the answer looks like a tool-call but was not recognized (attempt ' +
                         malformedRetries +
                         '/' +
                         MAX_MALFORMED_RETRIES +
                         ')');
                 }
                 message =
-                    'Твой предыдущий ответ не распознан как вызов инструмента. ' +
-                        'Ответь РОВНО одним JSON-объектом вызова инструмента, без текста до и после. ' +
-                        'НЕ используй XML/DSML-теги — только JSON. ' +
-                        'Например: {"tool": "Read", "args": {"path": "src/index.js"}}';
+                    'Your previous answer was not recognized as a tool call. ' +
+                        'Reply with EXACTLY one JSON tool-call object, no text before or after. ' +
+                        'Do NOT use XML/DSML tags — plain JSON only. ' +
+                        'For example: {"tool": "Read", "args": {"path": "src/index.js"}}';
                 continue;
             }
             const trimmed = (rawResponse || '').trim();
@@ -336,16 +336,16 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                     response: rawResponse,
                 });
                 if (debugLog) {
-                    console.error('внимание: пустой/служебный ответ, прошу продолжить (попытка ' +
+                    console.error('warning: empty/service answer, asking to continue (attempt ' +
                         stallRetries +
                         '/' +
                         MAX_STALL_RETRIES +
                         ')');
                 }
                 message =
-                    'Продолжи выполнение задачи. Если нужен инструмент — ответь РОВНО ' +
-                        'одним JSON-объектом вызова: {"tool": "...", "args": {...}}. ' +
-                        'Если задача выполнена — вызови инструмент respond с итоговым сообщением.';
+                    'Continue the task. If you need a tool — reply with EXACTLY ' +
+                        'one JSON tool-call object: {"tool": "...", "args": {...}}. ' +
+                        'If the task is done — call the respond tool with the final message.';
                 continue;
             }
             // The answer looks like "I'll call a tool now", but contains no call.
@@ -361,18 +361,18 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                     response: rawResponse,
                 });
                 if (debugLog) {
-                    console.error('внимание: ответ похож на незавершённую работу, прошу продолжить (попытка ' +
+                    console.error('warning: the answer looks like unfinished work, asking to continue (attempt ' +
                         looksDoneRetries +
                         '/' +
                         MAX_LOOKSDONE_RETRIES +
                         ')');
                 }
                 message =
-                    'Похоже, ты собирался вызвать инструмент, но не вызвал. ' +
-                        'Если задача ещё не выполнена — ответь РОВНО одним JSON-объектом ' +
-                        'вызова инструмента, без текста до и после. ' +
-                        'Если задача действительно выполнена — вызови respond с итоговым ' +
-                        'сообщением оператору.';
+                    'It looks like you meant to call a tool but did not. ' +
+                        'If the task is not finished — reply with EXACTLY one JSON ' +
+                        'tool-call object, no text before or after. ' +
+                        'If the task is really done — call respond with the final ' +
+                        'message to the operator.';
                 continue;
             }
             // STRICT: only tool calls and respond reach the operator. Plain text is
@@ -387,12 +387,12 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                 });
                 message =
                     (justRanTool
-                        ? 'Ты остановился после вызова инструмента и написал обычный текст. '
-                        : 'Ты написал обычный текст без вызова инструмента. ') +
-                        'Задача ещё не завершена. Ответь РОВНО одним JSON-объектом вызова ' +
-                        'инструмента, без текста до и после, например: ' +
+                        ? 'You stopped after a tool call and wrote plain text. '
+                        : 'You wrote plain text without a tool call. ') +
+                        'The task is not finished. Reply with EXACTLY one JSON tool-call ' +
+                        'object, no text before or after, for example: ' +
                         '{\"tool\": \"Bash\", \"args\": {\"command\": \"...\"}}. ' +
-                        'Если задача действительно выполнена — вызови respond с итоговым сообщением.';
+                        'If the task is really done — call respond with the final message.';
                 continue;
             }
             // Budget exhausted. Ask for respond EXACTLY once more; if the model
@@ -404,8 +404,8 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                     response: rawResponse.slice(0, 500),
                 });
                 message =
-                    'Последний шаг: вызови инструмент respond с итоговым сообщением ' +
-                        'оператору. Не пиши обычный текст — только вызов respond, например: ' +
+                    'Last step: call the respond tool with the final message ' +
+                        'to the operator. Do not write plain text — only a respond call, for example: ' +
                         '{\"tool\": \"respond\", \"args\": {\"message\": \"...\"}}';
                 continue;
             }
@@ -450,16 +450,16 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                 stallRetries++;
                 transcript?.log('empty_respond', { attempt: stallRetries });
                 if (debugLog) {
-                    console.error('внимание: пустой respond, прошу продолжить (попытка ' +
+                    console.error('warning: empty respond, asking to continue (attempt ' +
                         stallRetries +
                         '/' +
                         MAX_STALL_RETRIES +
                         ')');
                 }
                 message =
-                    'Ты вызвал respond с пустым message. Если задача выполнена — ' +
-                        'вызови respond с итоговым сообщением оператору. Если нет — ' +
-                        'продолжи работу вызовом инструмента.';
+                    'You called respond with an empty message. If the task is done — ' +
+                        'call respond with the final message to the operator. If not — ' +
+                        'continue the work with a tool call.';
                 continue;
             }
             // An EMPTY respond must NEVER be a silent final: the operator would see
@@ -469,9 +469,9 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
             // the task normally.
             if (!isMeaningfulRespond(msg)) {
                 transcript?.log('empty_respond_exhausted', { response: rawResponse });
-                safeWarning('Модель вызвала respond без текста, и лимит повторов исчерпан. ' +
-                    'Проверьте чат DeepSeek вручную.');
-                return 'Модель не сформировала итоговое сообщение (пустой respond).';
+                safeWarning('The model called respond without text, and the retry limit is exhausted. ' +
+                    'Check the DeepSeek chat manually.');
+                return 'The model produced no final message (empty respond).';
             }
             safeAssistantMessage(msg);
             transcript?.log('assistant_final', { message: msg });
@@ -485,7 +485,7 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
         for (const call of callsToRun) {
             const tool = tools.find((t) => t.name === call.tool);
             if (!tool) {
-                const err = `Неизвестный инструмент: ${call.tool}`;
+                const err = `Unknown tool: ${call.tool}`;
                 safeToolResult(err);
                 transcript?.log('tool_error', { tool: call.tool, error: err });
                 results.push({ tool: call.tool, result: err });
@@ -498,7 +498,7 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                 result = await tool.fn(call.args);
             }
             catch (e) {
-                result = `Ошибка: ${e.message}`;
+                result = `Error: ${e.message}`;
             }
             safeToolResult(result);
             transcript?.log('tool_result', {
@@ -536,7 +536,7 @@ export async function runAgentLoop({ browser, tools, task, workdir, maxIteration
                 .join('\n\n');
         }
     }
-    return 'Достигнут лимит итераций.';
+    return 'Iteration limit reached.';
 }
 // The answer looks like a tool call, but parseToolCall() did not recognize it.
 // Used as a safeguard against "the agent called a tool and stopped": in that

package/dist/commands.js

@@ -0,0 +1,201 @@
+import path from 'path';
+// Helpers for the extra slash commands (/diff, /cost, /export, /doctor,
+// /permissions, /review, /add-dir). Pure functions, unit-tested without a
+// live agent/browser. index.ts only renders their output.
+const NL = String.fromCharCode(10);
+const CR = String.fromCharCode(13);
+// ---------- /diff ----------
+export function formatDiff(diffText, opts = {}) {
+    const maxLines = opts.maxLines ?? 200;
+    const trimmed = String(diffText ?? '').split(CR + NL).join(NL).trimEnd();
+    if (!trimmed)
+        return '(no changes)';
+    const lines = trimmed.split(NL);
+    if (lines.length <= maxLines)
+        return trimmed;
+    const head = lines.slice(0, maxLines).join(NL);
+    const rest = lines.length - maxLines;
+    return (head +
+        NL +
+        '... [' + rest + ' more lines, use Bash git diff for the full output]');
+}
+export function diffGitArgs(staged = false) {
+    return staged ? 'git diff --staged' : 'git diff';
+}
+export function summarizeTranscript(entries) {
+    const stats = {
+        turns: 0,
+        toolCalls: 0,
+        toolCounts: {},
+        durationMs: 0,
+        startedAt: null,
+    };
+    for (const e of entries) {
+        if (!e || typeof e !== 'object')
+            continue;
+        if (e.type === 'user_task')
+            stats.turns++;
+        if (e.type === 'tool_call') {
+            stats.toolCalls++;
+            const tool = String(e.tool ?? '?');
+            stats.toolCounts[tool] = (stats.toolCounts[tool] || 0) + 1;
+        }
+        if (typeof e.elapsed === 'number' && e.elapsed > stats.durationMs) {
+            stats.durationMs = e.elapsed;
+        }
+        if (!stats.startedAt && typeof e.ts === 'string')
+            stats.startedAt = e.ts;
+    }
+    return stats;
+}
+export function parseTranscript(body) {
+    const out = [];
+    for (const line of String(body ?? '').split(NL)) {
+        const t = line.trim();
+        if (!t)
+            continue;
+        try {
+            const v = JSON.parse(t);
+            if (v && typeof v === 'object')
+                out.push(v);
+        }
+        catch {
+            // skip malformed lines
+        }
+    }
+    return out;
+}
+export function formatDuration(ms) {
+    const total = Math.max(0, Math.round(ms / 1000));
+    const h = Math.floor(total / 3600);
+    const m = Math.floor((total % 3600) / 60);
+    const s = total % 60;
+    const pad = (n) => String(n).padStart(2, '0');
+    if (h > 0)
+        return h + 'h ' + pad(m) + 'm ' + pad(s) + 's';
+    if (m > 0)
+        return m + 'm ' + pad(s) + 's';
+    return s + 's';
+}
+export function renderCost(stats, transcriptFile) {
+    const lines = [];
+    lines.push('Session stats (DeepSeek web does not expose token counts):');
+    lines.push(' tasks: ' + stats.turns);
+    lines.push(' tool calls: ' + stats.toolCalls);
+    const top = Object.entries(stats.toolCounts).sort((a, b) => b[1] - a[1]);
+    if (top.length) {
+        lines.push(' by tool:');
+        for (const [name, n] of top)
+            lines.push(' ' + name + ': ' + n);
+    }
+    lines.push(' duration: ' + formatDuration(stats.durationMs));
+    if (stats.startedAt)
+        lines.push(' started: ' + stats.startedAt);
+    lines.push(' transcript: ' + (transcriptFile || '(off)'));
+    return lines.join(NL);
+}
+// ---------- /export ----------
+export function formatExport(entries, meta = {}) {
+    const out = [];
+    out.push('# zames session export');
+    out.push('');
+    if (meta.workdir)
+        out.push('- workdir: ' + meta.workdir);
+    if (meta.chatId)
+        out.push('- chat: ' + meta.chatId);
+    out.push('- exported: ' + new Date().toISOString());
+    out.push('');
+    for (const e of entries) {
+        const type = String(e.type ?? '?');
+        if (type === 'user_task') {
+            out.push('## Task');
+            out.push('');
+            out.push(String(e.task ?? ''));
+            out.push('');
+        }
+        else if (type === 'tool_call') {
+            out.push('- tool ' +
+                String(e.tool ?? '?') +
+                ' ' +
+                JSON.stringify(e.args ?? {}));
+        }
+        else if (type === 'tool_result') {
+            out.push(' -> ' +
+                String(e.tool ?? '?') +
+                ': ' +
+                String(e.result ?? '').slice(0, 500));
+        }
+        else if (type === 'assistant_final') {
+            out.push('## Answer');
+            out.push('');
+            out.push(String(e.message ?? ''));
+            out.push('');
+        }
+    }
+    return out.join(NL);
+}
+export function defaultExportPath(workdir, now = new Date()) {
+    const stamp = now
+        .toISOString()
+        .replace(/[:.]/g, '-')
+        .replace('T', '_')
+        .replace('Z', '');
+    return path.join(workdir, 'zames-export-' + stamp + '.md');
+}
+export function renderDoctor(d) {
+    const rows = [];
+    const row = (ok, label, value) => {
+        rows.push(' ' + (ok ? '[OK] ' : '[WARN]') + ' ' + label.padEnd(16) + ' ' + value);
+    };
+    row(true, 'node', d.nodeVersion);
+    row(true, 'platform', d.platform);
+    row(true, 'workdir', d.workdir);
+    row(d.gitOk, 'git', d.gitOk ? 'repo (' + (d.gitBranch || 'detached') + ')' : 'not a repository');
+    row(d.configOk, 'config', d.configOk ? 'loaded' : 'error: ' + (d.configError || 'unknown'));
+    row(true, 'browser', d.browserChannel ? d.browserChannel : 'bundled chromium');
+    row(!!d.clipboardTool, 'clipboard', d.clipboardTool || 'no tool found (image paste disabled)');
+    row(true, 'mcp', d.mcpServers + ' server(s), ' + d.mcpTools + ' tool(s)');
+    row(d.transcriptOk, 'transcript', d.transcriptOk ? 'on' : 'off');
+    return 'Doctor:' + NL + rows.join(NL);
+}
+export function renderPermissions(p) {
+    const onoff = (b) => (b ? 'ask' : 'allow');
+    const lines = [];
+    lines.push('Tool permissions (confirmation settings):');
+    lines.push(' Write: ' + onoff(p.write));
+    lines.push(' Edit: ' + onoff(p.edit));
+    lines.push(' Bash: ' + onoff(p.bash));
+    if (p.alwaysConfirm.length) {
+        lines.push(' Always confirm (regex):');
+        for (const re of p.alwaysConfirm)
+            lines.push(' ' + re);
+    }
+    lines.push('');
+    lines.push('Change with: /config set confirmation.write false (and .edit / .bash)');
+    return lines.join(NL);
+}
+// ---------- /add-dir ----------
+export function resolveExtraDir(input, workdir) {
+    const raw = String(input ?? '').trim();
+    if (!raw)
+        return { error: 'Usage: /add-dir <path>' };
+    const abs = path.resolve(workdir, raw);
+    if (abs === path.resolve(workdir)) {
+        return { error: 'This is already the working directory.' };
+    }
+    return { path: abs };
+}
+// ---------- /review ----------
+export function buildReviewPrompt(focus, hasStaged = false) {
+    const scope = hasStaged ? 'staged' : 'uncommitted';
+    const f = String(focus ?? '').trim();
+    let s = 'Review the ' +
+        scope +
+        ' changes in this repository (use GitDiff). ' +
+        'Focus on real bugs: correctness, race conditions, error handling, ' +
+        'security, and missing tests. Do NOT invent problems and do NOT change ' +
+        'code: only report findings with file:line and a short rationale.';
+    if (f)
+        s += NL + 'Extra focus: ' + f;
+    return s;
+}