feat: /compact, thinking display, transient retry, context windows (v1.3.0)

New: /compact command to manually compress history with token count display
New: Ink UI shows thinking content preview (last line) instead of just spinner
New: Transient error retry with exponential backoff (429, 500, timeout, etc.)
New: First-run tips for new users, GLM promo auto-expiry fallback
Fix: System prompt documents all 11 tools with constraints
Fix: 16 missing model shortcuts synced to terminal picker
Fix: Token estimation adds 16-token tool_use framing overhead
Fix: Router detects code blocks, uses byte-length token estimation
Fix: Context window registry expanded (20+ models + pattern fallback)
Fix: Glob only recurses on **, not on / patterns
Fix: WebSearch parser fallback + DDG internal link filtering
Fix: Tool descriptions document key limits in schemas

Author: 1bcMax

CHANGELOG.md

@@ -1,5 +1,30 @@
 # Changelog
 
+## 1.3.0 (2026-04-04)
+
+### New Features
+
+- **`/compact` command**: Manually compress conversation history to save tokens. Shows before/after token count.
+- **Thinking content display**: Ink UI now shows last line of model's thinking process (was only showing spinner).
+- **Transient error retry**: Network timeouts, 429 rate limits, and server errors now auto-retry with exponential backoff (up to 3 attempts) instead of terminating the session.
+- **First-run tips**: New users see helpful tip about `/model`, `/compact`, and `/help` on first launch.
+- **GLM promo auto-expiry**: Default model automatically switches from GLM-5 to Gemini Flash after promo ends.
+
+### Bug Fixes
+
+- **System prompt completeness**: All 11 tools now documented with constraints (was missing 5 tools)
+- **Model shortcuts synced**: 16 missing shortcuts added to terminal picker (was out of sync with proxy)
+- **Token estimation**: tool_use overhead now includes 16-token framing cost
+- **Router code block detection**: Triple backtick code blocks now boost complexity score
+- **Router token estimation**: Uses byte length for better accuracy with CJK/Unicode
+- **Context window registry**: Added 20+ missing models (xAI, GLM, Minimax, etc.) + pattern-based fallback
+- **Glob recursion**: Only `**` triggers full recursion (was over-recursing on `/` patterns)
+- **WebSearch parser**: Fallback regex when DuckDuckGo updates HTML; skip internal DDG links
+- **Tool descriptions**: Read, Grep, Glob schemas now document key limits and defaults
+- **Terminal UI commands**: `/model`, `/cost`, `/help`, `/compact` now work in piped/non-TTY mode
+- **Escape to abort**: Press Esc during generation to cancel current turn
+- **Per-turn cost display**: Session cost shown after every response
+
 ## 1.2.0 (2026-04-04)
 
 ### Bug Fixes (36 fixes across 5 rounds)

dist/agent/loop.js

@@ -214,6 +214,22 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
             break; // User wants to exit
         if (input === '')
             continue; // Empty input → re-prompt
+        // Handle /compact command — force compaction without sending to model
+        if (input === '/compact') {
+            const beforeTokens = estimateHistoryTokens(history);
+            const { history: compacted, compacted: didCompact } = await autoCompactIfNeeded(history, config.model, client, config.debug);
+            if (didCompact) {
+                history.length = 0;
+                history.push(...compacted);
+            }
+            const afterTokens = estimateHistoryTokens(history);
+            onEvent({ kind: 'text_delta', text: didCompact
+                    ? `Compacted: ~${beforeTokens.toLocaleString()} → ~${afterTokens.toLocaleString()} tokens\n`
+                    : `History is ${beforeTokens.toLocaleString()} tokens — no compaction needed.\n`
+            });
+            onEvent({ kind: 'turn_done', reason: 'completed' });
+            continue;
+        }
         history.push({ role: 'user', content: input });
         const abort = new AbortController();
         onAbortReady?.(() => abort.abort());
@@ -287,18 +303,33 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
             }
             catch (err) {
                 const errMsg = err.message || '';
+                const errLower = errMsg.toLowerCase();
                 // ── Prompt too long recovery ──
-                if (errMsg.toLowerCase().includes('prompt is too long') && recoveryAttempts < 3) {
+                if (errLower.includes('prompt is too long') && recoveryAttempts < 3) {
                     recoveryAttempts++;
                     if (config.debug) {
                         console.error(`[runcode] Prompt too long — forcing compact (attempt ${recoveryAttempts})`);
                     }
-                    // Force compaction by reducing history
                     const { history: compactedAgain } = await autoCompactIfNeeded(history, config.model, client, config.debug);
                     history.length = 0;
                     history.push(...compactedAgain);
                     continue; // Retry
                 }
+                // ── Transient error recovery (network, rate limit, server errors) ──
+                const isTransient = errLower.includes('429') || errLower.includes('rate')
+                    || errLower.includes('500') || errLower.includes('502') || errLower.includes('503')
+                    || errLower.includes('timeout') || errLower.includes('econnrefused')
+                    || errLower.includes('econnreset') || errLower.includes('fetch failed');
+                if (isTransient && recoveryAttempts < 3) {
+                    recoveryAttempts++;
+                    const backoffMs = Math.pow(2, recoveryAttempts) * 1000;
+                    if (config.debug) {
+                        console.error(`[runcode] Transient error — retrying in ${backoffMs / 1000}s (attempt ${recoveryAttempts}): ${errMsg.slice(0, 100)}`);
+                    }
+                    onEvent({ kind: 'text_delta', text: `\n*Retrying (${recoveryAttempts}/3)...*\n` });
+                    await new Promise(r => setTimeout(r, backoffMs));
+                    continue;
+                }
                 onEvent({ kind: 'turn_done', reason: 'error', error: errMsg });
                 break;
             }

dist/agent/tokens.js

@@ -60,23 +60,56 @@ export function estimateHistoryTokens(history) {
  * Context window sizes for known models.
  */
 const MODEL_CONTEXT_WINDOWS = {
+    // Anthropic
     'anthropic/claude-opus-4.6': 200_000,
     'anthropic/claude-sonnet-4.6': 200_000,
     'anthropic/claude-sonnet-4': 200_000,
     'anthropic/claude-haiku-4.5': 200_000,
+    'anthropic/claude-haiku-4.5-20251001': 200_000,
+    // OpenAI
     'openai/gpt-5.4': 128_000,
     'openai/gpt-5.4-pro': 128_000,
+    'openai/gpt-5.3': 128_000,
+    'openai/gpt-5.3-codex': 128_000,
+    'openai/gpt-5.2': 128_000,
     'openai/gpt-5-mini': 128_000,
+    'openai/gpt-5-nano': 128_000,
+    'openai/gpt-4.1': 1_000_000,
+    'openai/o3': 200_000,
+    'openai/o4-mini': 200_000,
+    // Google
     'google/gemini-2.5-pro': 1_000_000,
     'google/gemini-2.5-flash': 1_000_000,
+    'google/gemini-2.5-flash-lite': 1_000_000,
+    'google/gemini-3.1-pro': 1_000_000,
+    // DeepSeek
     'deepseek/deepseek-chat': 64_000,
     'deepseek/deepseek-reasoner': 64_000,
+    // xAI
+    'xai/grok-3': 131_072,
+    'xai/grok-4-0709': 131_072,
+    'xai/grok-4-1-fast-reasoning': 131_072,
+    // Others
+    'zai/glm-5': 128_000,
+    'moonshot/kimi-k2.5': 128_000,
+    'minimax/minimax-m2.7': 128_000,
 };
 /**
  * Get the context window size for a model, with a conservative default.
  */
 export function getContextWindow(model) {
-    return MODEL_CONTEXT_WINDOWS[model] ?? 128_000;
+    if (MODEL_CONTEXT_WINDOWS[model])
+        return MODEL_CONTEXT_WINDOWS[model];
+    // Pattern-based inference for unknown models
+    if (model.includes('gemini'))
+        return 1_000_000;
+    if (model.includes('claude'))
+        return 200_000;
+    if (model.includes('gpt-4.1'))
+        return 1_000_000;
+    if (model.includes('nemotron') || model.includes('qwen'))
+        return 128_000;
+    return 128_000;
 }
 /**
  * Reserved tokens for the compaction summary output.

dist/commands/start.js

@@ -24,7 +24,9 @@ export async function startCommand(options) {
         model = configModel;
     }
     else {
-        model = 'zai/glm-5'; // Default: GLM-5 promo ($0.001/call)
+        // Default: GLM-5 promo if still active, otherwise Gemini Flash (cheap & reliable)
+        const promoExpiry = new Date('2026-04-15');
+        model = Date.now() < promoExpiry.getTime() ? 'zai/glm-5' : 'google/gemini-2.5-flash';
     }
     // Auto-create wallet if needed (no interruption — free models work without funding)
     let walletAddress = '';
@@ -52,6 +54,10 @@ export async function startCommand(options) {
     console.log(chalk.dim(`  Model:  ${model}`));
     console.log(chalk.dim(`  Wallet: ${walletAddress || 'not set'}`));
     console.log(chalk.dim(`  Dir:    ${workDir}`));
+    // First-run tip: show if no config file exists yet
+    if (!configModel && !options.model) {
+        console.log(chalk.dim(`\n  Tip: /model to switch models · /compact to save tokens · /help for all commands`));
+    }
     console.log('');
     // Fetch balance in background (don't block startup)
     const walletInfo = {

dist/tools/websearch.js

@@ -44,11 +44,16 @@ async function execute(input, _ctx) {
 }
 function parseDuckDuckGoResults(html, maxResults) {
     const results = [];
-    // Match result blocks: <a class="result__a" href="...">title</a>
+    // Primary parser: match result blocks by class names
     const linkRegex = /<a[^>]*class="result__a"[^>]*href="([^"]*)"[^>]*>([\s\S]*?)<\/a>/gi;
     const snippetRegex = /<a[^>]*class="result__snippet"[^>]*>([\s\S]*?)<\/a>/gi;
-    const links = [...html.matchAll(linkRegex)];
-    const snippets = [...html.matchAll(snippetRegex)];
+    let links = [...html.matchAll(linkRegex)];
+    let snippets = [...html.matchAll(snippetRegex)];
+    // Fallback parser if primary finds nothing (DDG may have updated HTML)
+    if (links.length === 0) {
+        const fallbackLink = /<a[^>]*class="[^"]*result[^"]*"[^>]*href="([^"]*)"[^>]*>([\s\S]*?)<\/a>/gi;
+        links = [...html.matchAll(fallbackLink)];
+    }
     for (let i = 0; i < Math.min(links.length, maxResults); i++) {
         const link = links[i];
         const snippet = snippets[i];
@@ -58,6 +63,9 @@ function parseDuckDuckGoResults(html, maxResults) {
         if (uddgMatch) {
             url = decodeURIComponent(uddgMatch[1]);
         }
+        // Skip internal DDG links
+        if (url.startsWith('/') || url.includes('duckduckgo.com'))
+            continue;
         results.push({
             title: stripTags(link[2] || '').trim(),
             url,

dist/ui/app.js

@@ -51,6 +51,7 @@ function RunCodeApp({ initialModel, workDir, walletAddress, walletBalance, chain
     const [showHelp, setShowHelp] = useState(false);
     const [showWallet, setShowWallet] = useState(false);
     const [balance, setBalance] = useState(walletBalance);
+    const [thinkingText, setThinkingText] = useState('');
     // Key handler for picker + esc + abort
     const isPickerOrEsc = mode === 'model-picker' || (mode === 'input' && ready && !input) || !ready;
     useInput((ch, key) => {
@@ -151,6 +152,9 @@ function RunCodeApp({ initialModel, workDir, walletAddress, walletBalance, chain
                     setStatusMsg('Conversation cleared');
                     setTimeout(() => setStatusMsg(''), 3000);
                     return;
+                case '/compact':
+                    onSubmit('/compact');
+                    return;
                 default:
                     setStatusMsg(`Unknown command: ${cmd}. Try /help`);
                     setTimeout(() => setStatusMsg(''), 3000);
@@ -161,6 +165,7 @@ function RunCodeApp({ initialModel, workDir, walletAddress, walletBalance, chain
         setInput('');
         setStreamText('');
         setThinking(false);
+        setThinkingText('');
         setTools(new Map());
         setReady(false);
         setWaiting(true);
@@ -184,6 +189,11 @@ function RunCodeApp({ initialModel, workDir, walletAddress, walletBalance, chain
                     case 'thinking_delta':
                         setWaiting(false);
                         setThinking(true);
+                        setThinkingText(prev => {
+                            // Keep last 500 chars of thinking for display
+                            const updated = prev + event.text;
+                            return updated.length > 500 ? updated.slice(-500) : updated;
+                        });
                         break;
                     case 'capability_start':
                         setWaiting(false);
@@ -239,9 +249,9 @@ function RunCodeApp({ initialModel, workDir, walletAddress, walletBalance, chain
                 }), _jsx(Text, { children: " " })] }));
     }
     // ── Normal Mode ──
-    return (_jsxs(Box, { flexDirection: "column", children: [statusMsg && (_jsx(Box, { marginLeft: 2, children: _jsx(Text, { color: "green", children: statusMsg }) })), showHelp && (_jsxs(Box, { flexDirection: "column", marginLeft: 2, marginTop: 1, marginBottom: 1, children: [_jsx(Text, { bold: true, children: "Commands" }), _jsx(Text, { children: " " }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/model" }), " [name]  Switch model (picker if no name)"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/wallet" }), "        Show wallet address & balance"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/cost" }), "          Session cost & savings"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/clear" }), "         Clear conversation display"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/help" }), "          This help"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/exit" }), "          Quit"] }), _jsx(Text, { children: " " }), _jsx(Text, { dimColor: true, children: "  Shortcuts: sonnet, opus, gpt, gemini, deepseek, flash, free, r1, o4, nano, mini, haiku" })] })), showWallet && (_jsxs(Box, { flexDirection: "column", marginLeft: 2, marginTop: 1, marginBottom: 1, children: [_jsx(Text, { bold: true, children: "Wallet" }), _jsx(Text, { children: " " }), _jsxs(Text, { children: ["  Chain:   ", _jsx(Text, { color: "magenta", children: chain })] }), _jsxs(Text, { children: ["  Address: ", _jsx(Text, { color: "cyan", children: walletAddress })] }), _jsxs(Text, { children: ["  Balance: ", _jsx(Text, { color: "green", children: balance })] })] })), Array.from(tools.values()).map((tool, i) => (_jsx(Box, { marginLeft: 1, children: tool.done ? (tool.error
+    return (_jsxs(Box, { flexDirection: "column", children: [statusMsg && (_jsx(Box, { marginLeft: 2, children: _jsx(Text, { color: "green", children: statusMsg }) })), showHelp && (_jsxs(Box, { flexDirection: "column", marginLeft: 2, marginTop: 1, marginBottom: 1, children: [_jsx(Text, { bold: true, children: "Commands" }), _jsx(Text, { children: " " }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/model" }), " [name]  Switch model (picker if no name)"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/wallet" }), "        Show wallet address & balance"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/cost" }), "          Session cost & savings"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/compact" }), "       Compress conversation history"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/clear" }), "         Clear conversation display"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/help" }), "          This help"] }), _jsxs(Text, { children: ["  ", _jsx(Text, { color: "cyan", children: "/exit" }), "          Quit"] }), _jsx(Text, { children: " " }), _jsx(Text, { dimColor: true, children: "  Shortcuts: sonnet, opus, gpt, gemini, deepseek, flash, free, r1, o4, nano, mini, haiku" })] })), showWallet && (_jsxs(Box, { flexDirection: "column", marginLeft: 2, marginTop: 1, marginBottom: 1, children: [_jsx(Text, { bold: true, children: "Wallet" }), _jsx(Text, { children: " " }), _jsxs(Text, { children: ["  Chain:   ", _jsx(Text, { color: "magenta", children: chain })] }), _jsxs(Text, { children: ["  Address: ", _jsx(Text, { color: "cyan", children: walletAddress })] }), _jsxs(Text, { children: ["  Balance: ", _jsx(Text, { color: "green", children: balance })] })] })), Array.from(tools.values()).map((tool, i) => (_jsx(Box, { marginLeft: 1, children: tool.done ? (tool.error
                     ? _jsxs(Text, { color: "red", children: ["  \u2717 ", tool.name, " ", _jsxs(Text, { dimColor: true, children: [tool.elapsed, "ms"] })] })
-                    : _jsxs(Text, { color: "green", children: ["  \u2713 ", tool.name, " ", _jsxs(Text, { dimColor: true, children: [tool.elapsed, "ms \u2014 ", tool.preview.slice(0, 60), tool.preview.length > 60 ? '...' : ''] })] })) : (_jsxs(Text, { color: "cyan", children: ["  ", _jsx(Spinner, { type: "dots" }), " ", tool.name, "..."] })) }, i))), thinking && (_jsx(Box, { marginLeft: 1, children: _jsxs(Text, { color: "magenta", children: ["  ", _jsx(Spinner, { type: "dots" }), " thinking..."] }) })), waiting && !thinking && tools.size === 0 && (_jsx(Box, { marginLeft: 1, children: _jsxs(Text, { color: "yellow", children: ["  ", _jsx(Spinner, { type: "dots" }), " ", _jsx(Text, { dimColor: true, children: currentModel })] }) })), streamText && (_jsx(Box, { marginTop: 0, marginBottom: 0, children: _jsx(Text, { children: streamText }) })), ready && (turnTokens.input > 0 || turnTokens.output > 0) && streamText && (_jsx(Box, { marginLeft: 1, marginTop: 0, children: _jsxs(Text, { dimColor: true, children: [turnTokens.input.toLocaleString(), " in / ", turnTokens.output.toLocaleString(), " out", totalCost > 0 ? `  ·  $${totalCost.toFixed(4)} session` : ''] }) })), ready && (_jsx(InputBox, { input: input, setInput: setInput, onSubmit: handleSubmit, model: currentModel, balance: balance, focused: mode === 'input' }))] }));
+                    : _jsxs(Text, { color: "green", children: ["  \u2713 ", tool.name, " ", _jsxs(Text, { dimColor: true, children: [tool.elapsed, "ms \u2014 ", tool.preview.slice(0, 60), tool.preview.length > 60 ? '...' : ''] })] })) : (_jsxs(Text, { color: "cyan", children: ["  ", _jsx(Spinner, { type: "dots" }), " ", tool.name, "..."] })) }, i))), thinking && (_jsxs(Box, { flexDirection: "column", marginLeft: 1, children: [_jsxs(Text, { color: "magenta", children: ["  ", _jsx(Spinner, { type: "dots" }), " thinking..."] }), thinkingText && (_jsxs(Text, { dimColor: true, wrap: "truncate-end", children: ["  ", thinkingText.split('\n').pop()?.slice(0, 80)] }))] })), waiting && !thinking && tools.size === 0 && (_jsx(Box, { marginLeft: 1, children: _jsxs(Text, { color: "yellow", children: ["  ", _jsx(Spinner, { type: "dots" }), " ", _jsx(Text, { dimColor: true, children: currentModel })] }) })), streamText && (_jsx(Box, { marginTop: 0, marginBottom: 0, children: _jsx(Text, { children: streamText }) })), ready && (turnTokens.input > 0 || turnTokens.output > 0) && streamText && (_jsx(Box, { marginLeft: 1, marginTop: 0, children: _jsxs(Text, { dimColor: true, children: [turnTokens.input.toLocaleString(), " in / ", turnTokens.output.toLocaleString(), " out", totalCost > 0 ? `  ·  $${totalCost.toFixed(4)} session` : ''] }) })), ready && (_jsx(InputBox, { input: input, setInput: setInput, onSubmit: handleSubmit, model: currentModel, balance: balance, focused: mode === 'input' }))] }));
 }
 export function launchInkUI(opts) {
     let resolveInput = null;

dist/ui/terminal.js

@@ -260,6 +260,7 @@ export class TerminalUI {
                 console.error(chalk.bold('\n  Commands:'));
                 console.error('  /model [name]  — switch model (e.g. /model sonnet)');
                 console.error('  /cost          — session cost and tokens');
+                console.error('  /compact       — compress conversation history');
                 console.error('  /exit          — quit');
                 console.error('  /help          — this help\n');
                 console.error(chalk.dim('  Shortcuts: sonnet, opus, gpt, gemini, deepseek, flash, free, r1, o4\n'));

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/runcode",
-  "version": "1.2.1",
+  "version": "1.3.0",
   "description": "RunCode — AI coding agent powered by 41+ models. Pay per use with USDC.",
   "type": "module",
   "bin": {