@slope-dev/slope 1.25.1 → 1.25.2

package/dist/cli/loop/slope-executor.js

@@ -0,0 +1,794 @@
+/**
+ * SlopeExecutor — custom agentic tool loop using the Anthropic Messages API.
+ *
+ * Implements ExecutorAdapter with:
+ * - 6 tools: read_file, write_file, edit_file, bash, glob, grep
+ * - Token/cost tracking from response.usage
+ * - Full transcript recording
+ * - Stuck detection (repeated identical tool calls)
+ * - Timeout enforcement
+ */
+import { execSync, execFileSync } from 'node:child_process';
+import { existsSync, readFileSync, writeFileSync, mkdirSync } from 'node:fs';
+import { join, resolve, dirname } from 'node:path';
+import { loadConfig } from '../../core/config.js';
+import { loadScorecards } from '../../core/loader.js';
+import { computeHandicapCard } from '../../core/handicap.js';
+import { extractHazardIndex, filterCommonIssues } from '../../core/briefing.js';
+import { extractKeywords } from './planner.js';
+// ── Constants ───────────────────────────────────────
+const MAX_TURNS_DEFAULT = 50;
+const MAX_REPEATED_CALLS = 3;
+const MAX_GUARD_RETRIES = 2;
+const MAX_OVERLOAD_RETRIES = 3;
+const MAX_CONSECUTIVE_BASH = 5;
+/** Turn budget per club — smaller tickets get fewer turns to prevent flailing */
+const CLUB_TURN_LIMITS = {
+    putter: 20,
+    wedge: 30,
+    short_iron: 40,
+    long_iron: 50,
+    driver: 50,
+};
+const DEFAULT_MAX_TOKENS = 8192;
+const TOOL_BASH_TIMEOUT = 60_000;
+const TOOL_OUTPUT_CAP = 50_000;
+const FILE_READ_CAP = 100_000;
+const TRANSCRIPT_CAP = 2000;
+const TRUNCATE_KEEP_RECENT = 20; // keep last 10 turns fully intact
+// Approximate cost per million tokens — not intended to be precise
+const COST_TABLE = {
+    'claude-haiku-4-5': { in: 0.80, out: 4.00 },
+    'claude-sonnet-4-5': { in: 3.00, out: 15.00 },
+    'claude-sonnet-4-6': { in: 3.00, out: 15.00 },
+    'claude-opus-4-6': { in: 15.00, out: 75.00 },
+};
+const DEFAULT_COST = { in: 1.00, out: 5.00 };
+// Destructive command blocklist — [pattern, human-readable reason]
+const BLOCKED_COMMANDS = [
+    [/\brm\s+-\w*r\w*\s+\//, 'rm with recursive flag targeting absolute path'],
+    [/\bgit\s+push\b/, 'push is handled by the loop, not the executor'],
+    [/\bmkfs\b/, 'filesystem format commands are not allowed'],
+    [/\bdd\b.*\bof=\//, 'dd write to absolute path is not allowed'],
+    [/\b(shutdown|reboot|halt|poweroff)\b/, 'system power commands are not allowed'],
+    [/\bcurl\b.*\|\s*(ba)?sh/, 'piping curl to shell is not allowed'],
+    [/\bpnpm\s+slope\b|\bslope\s+/, 'use the slope tool instead of running slope via bash'],
+];
+// ── Tool definitions (Anthropic API format) ─────────
+const TOOLS = [
+    {
+        name: 'read_file',
+        description: 'Read the full contents of a file. Always read a file before editing it.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                path: { type: 'string', description: 'Relative path from the repo root' },
+            },
+            required: ['path'],
+        },
+    },
+    {
+        name: 'write_file',
+        description: 'Create a new file or completely overwrite an existing one. Use edit_file for targeted changes.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                path: { type: 'string', description: 'Relative path from the repo root' },
+                content: { type: 'string', description: 'Complete file content' },
+            },
+            required: ['path', 'content'],
+        },
+    },
+    {
+        name: 'edit_file',
+        description: 'Replace an exact string in a file. The old_string must match exactly (including whitespace/indentation). Only the first occurrence is replaced.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                path: { type: 'string', description: 'Relative path from the repo root' },
+                old_string: { type: 'string', description: 'Exact text to find (must be unique enough to match once)' },
+                new_string: { type: 'string', description: 'Replacement text' },
+            },
+            required: ['path', 'old_string', 'new_string'],
+        },
+    },
+    {
+        name: 'bash',
+        description: 'Run a shell command. Use for git operations, running tests (pnpm test), type checking (pnpm typecheck), and other CLI tools. Commands time out after 60s.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                command: { type: 'string', description: 'Shell command to execute' },
+            },
+            required: ['command'],
+        },
+    },
+    {
+        name: 'glob',
+        description: 'Find files matching a glob pattern. Returns relative paths, one per line.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                pattern: { type: 'string', description: 'Glob pattern (e.g., "src/**/*.ts", "*.json")' },
+            },
+            required: ['pattern'],
+        },
+    },
+    {
+        name: 'grep',
+        description: 'Search file contents with a regex pattern. Returns matching lines with file paths and line numbers.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                pattern: { type: 'string', description: 'Search pattern (regex supported)' },
+                path: { type: 'string', description: 'File or directory to search (default: current directory)' },
+                include: { type: 'string', description: 'File glob filter (e.g., "*.ts")' },
+            },
+            required: ['pattern'],
+        },
+    },
+    {
+        name: 'slope',
+        description: 'Run a SLOPE CLI command (read-only). Available: search, context, briefing, card, validate, map, plan, prep, status, next, flows, doctor. Example: slope({ command: "briefing --categories=testing" })',
+        input_schema: {
+            type: 'object',
+            properties: {
+                command: {
+                    type: 'string',
+                    description: 'SLOPE subcommand and flags (e.g., "search --query=handicap", "briefing", "map")',
+                },
+            },
+            required: ['command'],
+        },
+    },
+];
+/** Allowlisted read-only slope subcommands */
+const SLOPE_ALLOWLIST = new Set([
+    'search', 'context', 'briefing', 'card', 'validate',
+    'map', 'plan', 'prep', 'status', 'next', 'flows',
+    'doctor', 'version',
+]);
+const SLOPE_TOOL_OUTPUT_CAP = 4000;
+// ── Path security ───────────────────────────────────
+export function safePath(relPath, cwd) {
+    const abs = resolve(cwd, relPath);
+    if (!abs.startsWith(resolve(cwd))) {
+        throw new Error(`Path traversal blocked: ${relPath}`);
+    }
+    return abs;
+}
+// ── Model ID resolution ─────────────────────────────
+export function resolveModelId(model) {
+    return model
+        .replace(/^openrouter\/anthropic\//, '')
+        .replace(/^anthropic\//, '');
+}
+export function lookupCost(modelId) {
+    for (const [key, cost] of Object.entries(COST_TABLE)) {
+        if (modelId.includes(key))
+            return cost;
+    }
+    return DEFAULT_COST;
+}
+// ── Executor ────────────────────────────────────────
+export const slopeExecutor = {
+    id: 'slope',
+    async execute(ctx, config, cwd, log) {
+        const start = Date.now();
+        const transcript = [];
+        let totalIn = 0;
+        let totalOut = 0;
+        let client;
+        try {
+            const { default: AnthropicSDK } = await import('@anthropic-ai/sdk');
+            // Supports ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL from env
+            const baseURL = process.env.ANTHROPIC_BASE_URL;
+            client = new AnthropicSDK(baseURL ? { baseURL } : undefined);
+        }
+        catch (err) {
+            const msg = err instanceof Error ? err.message : String(err);
+            const hint = msg.includes('Cannot find') || msg.includes('MODULE_NOT_FOUND')
+                ? 'Install: pnpm add @anthropic-ai/sdk'
+                : 'Is ANTHROPIC_API_KEY set?';
+            log.error(`Anthropic client init failed (${hint}): ${msg}`);
+            return errorResult(transcript, start);
+        }
+        const modelId = resolveModelId(ctx.model);
+        const costRate = lookupCost(modelId);
+        const deadline = start + ctx.timeout * 1000;
+        const systemPrompt = buildSystemPrompt(ctx, config, cwd);
+        // Message history for the agentic loop
+        const messages = [
+            { role: 'user', content: ctx.prompt },
+        ];
+        const maxTurns = CLUB_TURN_LIMITS[ctx.ticket.club] ?? MAX_TURNS_DEFAULT;
+        const recentSigs = [];
+        let outcome = 'completed';
+        let turn = 0;
+        let guardRetries = 0;
+        let overloadRetries = 0;
+        let innerGuardsPassed = false;
+        let consecutiveBash = 0;
+        // ── Agent loop ──
+        while (turn < maxTurns) {
+            const remaining = deadline - Date.now();
+            if (remaining <= 0) {
+                log.warn(`Timed out after ${ctx.timeout}s`);
+                outcome = 'timeout';
+                break;
+            }
+            turn++;
+            // Truncate old tool results to stay within context limits
+            truncateOldMessages(messages);
+            // AbortSignal for per-call timeout
+            const controller = new AbortController();
+            const callTimeout = setTimeout(() => controller.abort(), remaining);
+            let response;
+            try {
+                response = await client.messages.create({
+                    model: modelId,
+                    max_tokens: DEFAULT_MAX_TOKENS,
+                    system: systemPrompt,
+                    tools: TOOLS,
+                    messages,
+                }, { signal: controller.signal });
+            }
+            catch (err) {
+                const msg = err instanceof Error ? err.message : String(err);
+                if (controller.signal.aborted) {
+                    log.warn(`API call aborted (deadline reached)`);
+                    outcome = 'timeout';
+                    break;
+                }
+                log.error(`API error (turn ${turn}): ${msg}`);
+                if (msg.includes('overloaded') && overloadRetries < MAX_OVERLOAD_RETRIES) {
+                    overloadRetries++;
+                    await sleep(5000 * overloadRetries);
+                    continue;
+                }
+                outcome = 'error';
+                break;
+            }
+            finally {
+                clearTimeout(callTimeout);
+            }
+            // Accumulate usage
+            if (response.usage) {
+                totalIn += response.usage.input_tokens;
+                totalOut += response.usage.output_tokens;
+            }
+            // Model thinks it's done — run inner guards before accepting
+            if (response.stop_reason === 'end_turn') {
+                messages.push({ role: 'assistant', content: response.content });
+                // Skip inner guards if we've exhausted retries or are past deadline
+                if (guardRetries >= MAX_GUARD_RETRIES || Date.now() > deadline) {
+                    break;
+                }
+                const guardFailure = runInnerGuards(config, cwd, log);
+                if (!guardFailure) {
+                    log.info('Inner guards passed');
+                    innerGuardsPassed = true;
+                    break;
+                }
+                // Feed the error back so the model can self-correct
+                guardRetries++;
+                log.warn(`Inner guard failed (attempt ${guardRetries}/${MAX_GUARD_RETRIES}): ${guardFailure.guard}`);
+                messages.push({
+                    role: 'user',
+                    content: `Your changes have an issue that must be fixed before this ticket is complete.\n\n## ${guardFailure.guard} failed\n\`\`\`\n${guardFailure.output}\n\`\`\`\n\nPlease fix the issue and verify again.`,
+                });
+                continue;
+            }
+            if (response.stop_reason !== 'tool_use') {
+                log.warn(`Unexpected stop_reason: ${response.stop_reason}`);
+                break;
+            }
+            // Extract tool_use blocks
+            const toolBlocks = response.content.filter((b) => b.type === 'tool_use');
+            // Stuck detection — same tool calls N times in a row
+            const sig = toolBlocks.map(b => `${b.name}:${JSON.stringify(b.input)}`).join('|');
+            recentSigs.push(sig);
+            if (recentSigs.length > MAX_REPEATED_CALLS)
+                recentSigs.shift();
+            if (recentSigs.length === MAX_REPEATED_CALLS &&
+                recentSigs.every(s => s === sig)) {
+                log.warn('Stuck: identical tool calls repeated — stopping');
+                outcome = 'stuck';
+                break;
+            }
+            // Execute each tool
+            const toolResults = [];
+            for (const block of toolBlocks) {
+                const toolStart = Date.now();
+                const res = runTool(block.name, block.input, cwd, log);
+                const duration_ms = Date.now() - toolStart;
+                transcript.push({
+                    timestamp: new Date().toISOString(),
+                    tool: block.name,
+                    input: block.input,
+                    output: res.output.slice(0, TRANSCRIPT_CAP),
+                    duration_ms,
+                });
+                toolResults.push({
+                    type: 'tool_result',
+                    tool_use_id: block.id,
+                    content: res.output,
+                    ...(res.isError ? { is_error: true } : {}),
+                });
+            }
+            messages.push({ role: 'assistant', content: response.content });
+            messages.push({ role: 'user', content: toolResults });
+            log.info(`Turn ${turn}: ${toolBlocks.map(b => b.name).join(', ')}`);
+            // Bash-loop breaker: if the model runs bash N+ times in a row without
+            // reading or editing files, inject a nudge to break the spiral
+            const allBash = toolBlocks.every(b => b.name === 'bash');
+            consecutiveBash = allBash ? consecutiveBash + 1 : 0;
+            if (consecutiveBash >= MAX_CONSECUTIVE_BASH) {
+                log.warn(`Bash loop detected (${consecutiveBash} consecutive) — injecting nudge`);
+                messages.push({
+                    role: 'user',
+                    content: `You have run ${consecutiveBash} consecutive bash commands without reading or editing any files. This is usually a sign you are stuck. Stop and think about what is actually failing. Use read_file to examine the error, or use grep/glob to find the right file to edit. Do not run another bash command until you have read the relevant code.`,
+                });
+                consecutiveBash = 0;
+            }
+        }
+        if (turn >= maxTurns) {
+            log.warn(`Hit max turns (${maxTurns})`);
+            if (outcome === 'completed')
+                outcome = 'stuck';
+        }
+        // Commit any uncommitted changes
+        const filesChanged = gitCommit(ctx.ticketKey, ctx.ticket.title, cwd, log);
+        const duration_s = Math.round((Date.now() - start) / 1000);
+        const cost_usd = (totalIn * costRate.in + totalOut * costRate.out) / 1_000_000;
+        return {
+            outcome,
+            noop: false, // caller checks SHA diff
+            tokens_in: totalIn,
+            tokens_out: totalOut,
+            cost_usd: Math.round(cost_usd * 10000) / 10000,
+            duration_s,
+            transcript,
+            files_changed: filesChanged,
+            innerGuardsPassed: innerGuardsPassed && filesChanged.length === 0,
+        };
+    },
+};
+// ── System prompt ───────────────────────────────────
+export function buildSystemPrompt(ctx, config, cwd) {
+    let guide = '';
+    let guideWordCount = 0;
+    const guidePath = join(cwd, config.agentGuide);
+    if (existsSync(guidePath)) {
+        const raw = readFileSync(guidePath, 'utf8');
+        guideWordCount = raw.split(/\s+/).length;
+        if (guideWordCount <= config.agentGuideMaxWords) {
+            guide = `\n\n## Agent Guide\n${raw}`;
+        }
+    }
+    const modules = ctx.ticket.modules;
+    const scopeSection = modules.length > 0
+        ? `\n## Allowed Files (scope)\nOnly modify files in or related to these modules:\n${modules.map((m) => `- ${m}`).join('\n')}\nYou may create new test files for these modules. Do NOT edit files outside this scope.`
+        : '';
+    // SLOPE sprint context (Layer 1) — hazards, common issues, handicap
+    let slopeContext = '';
+    try {
+        slopeContext = buildSlopeContext(ctx.ticket, config, cwd, guideWordCount);
+        if (slopeContext)
+            slopeContext = '\n\n' + slopeContext;
+    }
+    catch { /* non-blocking */ }
+    return `You are an autonomous coding agent working on the SLOPE project.
+This is a TypeScript monorepo (pnpm, vitest, strict TypeScript).
+
+## Working Directory
+${cwd}
+
+## Ticket: ${ctx.ticketKey}
+${ctx.ticket.title}
+${scopeSection}
+
+## Rules
+- ALWAYS read a file before editing it — understand existing patterns first
+- Make real, substantive changes — never add only comments or whitespace
+- Keep changes minimal and focused on this ticket only
+- Do NOT edit files outside the allowed scope above
+- After all changes, run: pnpm typecheck && pnpm test
+- If tests fail, read the error output carefully before attempting a fix
+- If stuck after multiple bash attempts, stop and re-read the relevant source files
+- Do NOT run git commit — the system auto-commits after verification
+
+## Tools
+- read_file: Read file contents (always do this first)
+- edit_file: Surgical string replacement (preferred for changes)
+- write_file: Create new files or full rewrites only
+- bash: Shell commands (tests, typecheck, git, etc.)
+- glob: Find files by pattern
+- grep: Search file contents
+- slope: Query SLOPE sprint data (briefing, search, card, map, etc.)${guide}${slopeContext}`;
+}
+// ── SLOPE Context Injection (Layer 1) ───────────────
+/**
+ * Build SLOPE sprint context for injection into the system prompt.
+ * Each section is independently wrapped in try/catch for graceful degradation.
+ */
+export function buildSlopeContext(ticket, config, cwd, guideWordCount = 0) {
+    const wordBudget = Math.min(2000, config.agentGuideMaxWords - guideWordCount - 500);
+    if (wordBudget <= 0)
+        return '';
+    const sections = [];
+    const keywords = extractKeywords(`${ticket.title} ${ticket.description} ${ticket.modules.join(' ')}`, 5);
+    let slopeConfig;
+    try {
+        slopeConfig = loadConfig(cwd);
+    }
+    catch {
+        return '';
+    }
+    let scorecards;
+    try {
+        scorecards = loadScorecards(slopeConfig, cwd);
+    }
+    catch {
+        scorecards = [];
+    }
+    // Section 1: Hazard briefing
+    try {
+        if (scorecards.length > 0) {
+            const hazards = extractHazardIndex(scorecards);
+            const recentHazards = hazards.shot_hazards
+                .filter(h => keywords.some(kw => h.description.toLowerCase().includes(kw)))
+                .slice(0, 5);
+            if (recentHazards.length > 0) {
+                sections.push('### Hazard Warnings');
+                for (const h of recentHazards) {
+                    sections.push(`- [S${h.sprint}] ${h.type}: ${h.description}`);
+                }
+            }
+        }
+    }
+    catch { /* skip section */ }
+    // Section 2: Common issues
+    try {
+        const issuesPath = join(cwd, slopeConfig.commonIssuesPath);
+        if (existsSync(issuesPath)) {
+            const issues = JSON.parse(readFileSync(issuesPath, 'utf8'));
+            const filtered = filterCommonIssues(issues, { keywords });
+            if (filtered.length > 0) {
+                sections.push('### Known Gotchas');
+                for (const p of filtered.slice(0, 5)) {
+                    sections.push(`- [${p.category}] ${p.title}`);
+                    sections.push(`  Prevention: ${p.prevention.slice(0, 120)}`);
+                }
+            }
+        }
+    }
+    catch { /* skip section */ }
+    // Section 3: Codebase map section
+    try {
+        const mapPath = join(cwd, 'CODEBASE.md');
+        if (existsSync(mapPath)) {
+            const mapContent = readFileSync(mapPath, 'utf8');
+            const relevantSection = extractMapSection(mapContent, ticket.modules);
+            if (relevantSection) {
+                sections.push('### Codebase Context');
+                sections.push(relevantSection);
+            }
+        }
+    }
+    catch { /* skip section */ }
+    // Section 4: Handicap snapshot
+    try {
+        if (scorecards.length > 0) {
+            const card = computeHandicapCard(scorecards);
+            const last5 = card.last_5;
+            sections.push('### Handicap Snapshot (last 5)');
+            sections.push(`- Handicap: +${last5.handicap.toFixed(1)}`);
+            sections.push(`- GIR: ${last5.gir_pct.toFixed(1)}%`);
+            sections.push(`- Avg Putts: ${last5.avg_putts.toFixed(1)}`);
+            sections.push(`- Penalties: ${last5.penalties_per_round.toFixed(1)}/round`);
+            const mp = last5.miss_pattern;
+            const totalMisses = mp.long + mp.short + mp.left + mp.right;
+            if (totalMisses > 0) {
+                const dirs = ['long', 'short', 'left', 'right']
+                    .filter(d => mp[d] > 0)
+                    .map(d => `${d}:${mp[d]}`);
+                sections.push(`- Miss pattern: ${dirs.join(' ')}`);
+            }
+        }
+    }
+    catch { /* skip section */ }
+    if (sections.length === 0)
+        return '';
+    let result = '## Sprint Context\n\n' + sections.join('\n');
+    const words = result.split(/\s+/);
+    if (words.length > wordBudget) {
+        result = words.slice(0, wordBudget).join(' ') + '\n...(truncated)';
+    }
+    return result;
+}
+/**
+ * Extract the relevant section from CODEBASE.md based on module paths.
+ */
+export function extractMapSection(mapContent, modules) {
+    if (modules.length === 0)
+        return null;
+    const lines = mapContent.split('\n');
+    const matchedLines = [];
+    let capturing = false;
+    let captureDepth = 0;
+    for (const line of lines) {
+        const headingMatch = line.match(/^(#{1,3})\s+(.+)/);
+        if (headingMatch) {
+            const depth = headingMatch[1].length;
+            const title = headingMatch[2].toLowerCase();
+            const matches = modules.some(mod => {
+                const parts = mod.split('/').filter(p => p.length > 2 && !p.includes('.'));
+                return parts.some(part => title.includes(part.toLowerCase()));
+            });
+            if (matches && !capturing) {
+                capturing = true;
+                captureDepth = depth;
+                matchedLines.push(line);
+            }
+            else if (capturing && depth <= captureDepth) {
+                capturing = false;
+                if (matches) {
+                    capturing = true;
+                    captureDepth = depth;
+                    matchedLines.push(line);
+                }
+            }
+            else if (capturing) {
+                matchedLines.push(line);
+            }
+        }
+        else if (capturing) {
+            matchedLines.push(line);
+        }
+    }
+    const result = matchedLines.join('\n').trim();
+    return result.length > 0 ? result : null;
+}
+export function runTool(name, input, cwd, log) {
+    try {
+        switch (name) {
+            case 'read_file': return toolReadFile(input, cwd);
+            case 'write_file': return toolWriteFile(input, cwd);
+            case 'edit_file': return toolEditFile(input, cwd);
+            case 'bash': return toolBash(input, cwd);
+            case 'glob': return toolGlob(input, cwd);
+            case 'grep': return toolGrep(input, cwd);
+            case 'slope': return toolSlope(input, cwd);
+            default: return { output: `Unknown tool: ${name}`, isError: true };
+        }
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        log.warn(`Tool ${name} error: ${msg}`);
+        return { output: msg, isError: true };
+    }
+}
+function toolReadFile(input, cwd) {
+    const abs = safePath(input.path, cwd);
+    if (!existsSync(abs))
+        return { output: `File not found: ${input.path}`, isError: true };
+    let content = readFileSync(abs, 'utf8');
+    if (content.length > FILE_READ_CAP) {
+        content = content.slice(0, FILE_READ_CAP) + '\n... (truncated at 100KB)';
+    }
+    return { output: content, isError: false };
+}
+function toolWriteFile(input, cwd) {
+    const abs = safePath(input.path, cwd);
+    mkdirSync(dirname(abs), { recursive: true });
+    writeFileSync(abs, input.content);
+    return { output: `Written: ${input.path}`, isError: false };
+}
+function toolEditFile(input, cwd) {
+    const abs = safePath(input.path, cwd);
+    if (!existsSync(abs))
+        return { output: `File not found: ${input.path}`, isError: true };
+    const content = readFileSync(abs, 'utf8');
+    const oldStr = input.old_string;
+    if (!content.includes(oldStr)) {
+        return {
+            output: `old_string not found in ${input.path}. Ensure exact match including whitespace and indentation.`,
+            isError: true,
+        };
+    }
+    writeFileSync(abs, content.replace(oldStr, input.new_string));
+    return { output: `Edited: ${input.path}`, isError: false };
+}
+function toolBash(input, cwd) {
+    const cmd = input.command;
+    // Block destructive commands — the model should not push (loop handles it),
+    // delete broad filesystem paths, or run system-level commands
+    const blocked = BLOCKED_COMMANDS.find(([re]) => re.test(cmd));
+    if (blocked) {
+        return { output: `Blocked: ${blocked[1]}`, isError: true };
+    }
+    try {
+        const output = execSync(cmd, {
+            cwd,
+            encoding: 'utf8',
+            timeout: TOOL_BASH_TIMEOUT,
+            maxBuffer: 1024 * 1024,
+            stdio: ['pipe', 'pipe', 'pipe'],
+        });
+        return {
+            output: (output.length > TOOL_OUTPUT_CAP
+                ? output.slice(0, TOOL_OUTPUT_CAP) + '\n... (truncated)'
+                : output) || '(no output)',
+            isError: false,
+        };
+    }
+    catch (err) {
+        const e = err;
+        const out = (e.stderr || '') + (e.stdout || '') || e.message || 'Command failed';
+        return {
+            output: typeof out === 'string' ? out.slice(0, TOOL_OUTPUT_CAP) : 'Command failed',
+            isError: true,
+        };
+    }
+}
+function toolGlob(input, cwd) {
+    const pattern = input.pattern;
+    try {
+        const output = execFileSync('git', ['ls-files', '--cached', '--others', '--exclude-standard', '--', pattern], { cwd, encoding: 'utf8', timeout: 10_000 });
+        const files = output.split('\n').filter(Boolean);
+        if (files.length === 0)
+            return { output: '(no matches)', isError: false };
+        return { output: files.slice(0, 200).join('\n'), isError: false };
+    }
+    catch {
+        return { output: '(no matches)', isError: false };
+    }
+}
+function toolGrep(input, cwd) {
+    const pattern = input.pattern;
+    const searchPath = input.path || '.';
+    const include = input.include;
+    try {
+        const args = ['-rn', '--color=never'];
+        if (include) {
+            args.push(`--include=${include}`);
+        }
+        else {
+            args.push('--include=*.ts', '--include=*.js', '--include=*.json', '--include=*.md', '--include=*.sh');
+        }
+        args.push('--', pattern, searchPath);
+        const output = execFileSync('grep', args, {
+            cwd,
+            encoding: 'utf8',
+            timeout: 15_000,
+            maxBuffer: 512 * 1024,
+        });
+        const lines = output.split('\n').filter(Boolean);
+        if (lines.length === 0)
+            return { output: '(no matches)', isError: false };
+        return { output: lines.slice(0, 100).join('\n'), isError: false };
+    }
+    catch {
+        // grep exit code 1 = no matches
+        return { output: '(no matches)', isError: false };
+    }
+}
+function toolSlope(input, cwd) {
+    const command = (input.command ?? '').trim();
+    const parts = command.split(/\s+/);
+    const subcommand = parts[0];
+    if (!subcommand || !SLOPE_ALLOWLIST.has(subcommand)) {
+        return {
+            output: `Command "${subcommand ?? ''}" not in allowlist. Use one of: ${[...SLOPE_ALLOWLIST].join(', ')}`,
+            isError: true,
+        };
+    }
+    try {
+        const output = execFileSync('pnpm', ['slope', ...parts], {
+            cwd,
+            encoding: 'utf8',
+            timeout: 30_000,
+        });
+        if (output.length > SLOPE_TOOL_OUTPUT_CAP) {
+            return { output: output.slice(0, SLOPE_TOOL_OUTPUT_CAP) + '\n...(output truncated)', isError: false };
+        }
+        return { output: output || '(no output)', isError: false };
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        return { output: `slope ${command} failed: ${msg.slice(0, 200)}`, isError: true };
+    }
+}
+// ── Git helpers ─────────────────────────────────────
+function gitCommit(ticketKey, title, cwd, log) {
+    try {
+        const status = execFileSync('git', ['status', '--porcelain'], {
+            cwd,
+            encoding: 'utf8',
+        }).trim();
+        if (!status)
+            return [];
+        const files = status.split('\n').map(l => l.slice(3).trim()).filter(Boolean);
+        execFileSync('git', ['add', '-A'], { cwd, stdio: 'pipe' });
+        execFileSync('git', ['commit', '-m', `${ticketKey}: ${title}`], {
+            cwd,
+            stdio: 'pipe',
+        });
+        log.info(`Committed: ${ticketKey}: ${title} (${files.length} files)`);
+        return files;
+    }
+    catch (err) {
+        log.warn(`Commit helper failed: ${err instanceof Error ? err.message : err}`);
+        return [];
+    }
+}
+/**
+ * Run typecheck + tests inside the executor loop, giving the model
+ * a chance to self-correct before the outer guards revert everything.
+ * Returns null on success, or the failure details.
+ */
+function runInnerGuards(config, cwd, log) {
+    // Guard 1: Typecheck
+    try {
+        execSync('pnpm typecheck', { cwd, stdio: 'pipe', timeout: 120_000 });
+        log.info('Inner guard: typecheck passed');
+    }
+    catch (err) {
+        const e = err;
+        const output = ((e.stderr || '') + (e.stdout || '')).slice(0, 3000) || 'typecheck failed';
+        log.warn(`Inner guard failed: typecheck`);
+        return { guard: 'typecheck', output };
+    }
+    // Guard 2: Tests
+    try {
+        execSync(config.loopTestCmd, { cwd, stdio: 'pipe', timeout: 300_000 });
+        log.info('Inner guard: tests passed');
+    }
+    catch (err) {
+        const e = err;
+        const output = ((e.stderr || '') + (e.stdout || '')).slice(0, 3000) || 'tests failed';
+        log.warn(`Inner guard failed: tests`);
+        return { guard: 'tests', output };
+    }
+    return null;
+}
+// ── Message truncation ──────────────────────────────
+/**
+ * Truncate old tool result contents to prevent context window overflow.
+ * Keeps the first message (task prompt) and recent turns fully intact.
+ * Older tool results are shrunk to 200 chars.
+ */
+function truncateOldMessages(messages) {
+    if (messages.length <= TRUNCATE_KEEP_RECENT + 2)
+        return;
+    for (let i = 2; i < messages.length - TRUNCATE_KEEP_RECENT; i++) {
+        const msg = messages[i];
+        if (msg.role === 'user' && Array.isArray(msg.content)) {
+            for (const block of msg.content) {
+                if (block.type === 'tool_result' &&
+                    typeof block.content === 'string' &&
+                    block.content.length > 200) {
+                    block.content = block.content.slice(0, 200) + '\n[truncated]';
+                }
+            }
+        }
+    }
+}
+// ── Utilities ───────────────────────────────────────
+function sleep(ms) {
+    return new Promise(r => setTimeout(r, ms));
+}
+function errorResult(transcript, start) {
+    return {
+        outcome: 'error',
+        noop: false,
+        tokens_in: 0,
+        tokens_out: 0,
+        cost_usd: 0,
+        duration_s: Math.round((Date.now() - start) / 1000),
+        transcript,
+        files_changed: [],
+    };
+}
+//# sourceMappingURL=slope-executor.js.map