/** * improver.ts * Reads diagnosis + prompt files, invokes AI, parses PromptEdit[] from response. * Validates: old_text verbatim match, char count limits. */ import fs from 'fs'; import path from 'path'; import type { PromptEdit, PromptProposal } from './types.js'; import { invokeLLM, type AgentType } from './invoke-llm.js'; const CHAR_LIMIT = 11_000; const IMPROVE_PROMPT_PATH = new URL('../src/prompts/improve.md', import.meta.url).pathname .replace(/^\/([A-Za-z]:)/, '$1'); // Fix Windows path // ── Helpers ─────────────────────────────────────────────────────────────────── function interpolate(template: string, vars: Record): string { let result = template; for (const [key, value] of Object.entries(vars)) { result = result.replaceAll(`{{${key}}}`, value); } return result; } function generateProposalId(): string { const now = new Date(); const ts = now.toISOString().replace(/[-:T.Z]/g, '').slice(0, 14); return `prop-${ts}`; } function readPromptFiles(plan2codeRoot: string): Record { const srcDir = path.join(plan2codeRoot, 'src'); const promptFiles = [ 'plan2code-1-plan.md', 'plan2code-1b-revise-plan.md', 'plan2code-2-document.md', 'plan2code-3-implement.md', 'plan2code-4-finalize.md', 'plan2code-init.md', 'plan2code-init-update.md', 'plan2code-quick-task.md', ]; const contents: Record = {}; for (const file of promptFiles) { try { contents[file] = fs.readFileSync(path.join(srcDir, file), 'utf8'); } catch { contents[file] = ''; } } return contents; } // ── Edit validation ─────────────────────────────────────────────────────────── export interface ValidationResult { valid: boolean; errors: string[]; warnings: string[]; } export function validateEdit( edit: PromptEdit, promptContents: Record, ): ValidationResult { const errors: string[] = []; const warnings: string[] = []; // Reject path traversal attempts if (edit.file.includes('..') || path.isAbsolute(edit.file)) { errors.push(`Rejected: "${edit.file}" contains path traversal or absolute path.`); return { valid: false, errors, warnings }; } // Check target file exists const fileContent = promptContents[edit.file]; if (fileContent === undefined) { errors.push(`Target file "${edit.file}" not found. Valid files: ${Object.keys(promptContents).join(', ')}`); return { valid: false, errors, warnings }; } // Check old_text exists verbatim in the file if (!fileContent.includes(edit.old_text)) { errors.push(`old_text not found verbatim in "${edit.file}". The AI may have hallucinated text.`); } else { // Warn if old_text appears more than once (ambiguous match) const occurrences = fileContent.split(edit.old_text).length - 1; if (occurrences > 1) { warnings.push(`old_text appears ${occurrences} times in "${edit.file}". Only the first occurrence will be replaced.`); } } // Check char count after edit const afterContent = fileContent.replace(edit.old_text, edit.new_text); if (afterContent.length > CHAR_LIMIT) { errors.push(`Edit would cause "${edit.file}" to exceed ${CHAR_LIMIT} char limit (would be ${afterContent.length} chars).`); } // Verify reported char counts match reality const actualBefore = fileContent.length; const actualAfter = afterContent.length; if (Math.abs(edit.char_count_before - actualBefore) > 10) { warnings.push(`Reported char_count_before (${edit.char_count_before}) differs from actual (${actualBefore}).`); } if (Math.abs(edit.char_count_after - actualAfter) > 10) { warnings.push(`Reported char_count_after (${edit.char_count_after}) differs from actual (${actualAfter}).`); } return { valid: errors.length === 0, errors, warnings }; } // ── AI response parsing ─────────────────────────────────────────────────────── export function parseProposalFromResponse(response: string): PromptEdit[] | null { // Look for JSON code block containing PromptEdit[] const jsonBlockMatch = response.match(/```(?:json)?\s*(\[[\s\S]*?\])\s*```/); if (!jsonBlockMatch) { // Try bare JSON array const bareMatch = response.match(/(\[[\s\S]*"old_text"[\s\S]*\])/); if (!bareMatch) return null; try { return JSON.parse(bareMatch[1]) as PromptEdit[]; } catch { return null; } } try { return JSON.parse(jsonBlockMatch[1]) as PromptEdit[]; } catch { return null; } } // ── Main improver ───────────────────────────────────────────────────────────── export interface ImproverOptions { diagnosisPath: string; // Path to diagnosis markdown file plan2codeRoot: string; // Path to plan2code repo root proposalsDir: string; // Where to save proposal JSON runsDir: string; // For tracking which runs this is based on model?: string; agent?: AgentType; // Agent to use (default: claude-code) } export interface ImproverResult { proposalPath: string; proposal: PromptProposal; validationResults: Array<{ edit: PromptEdit; result: ValidationResult }>; validEditCount: number; invalidEditCount: number; } export async function generateImprovement(opts: ImproverOptions): Promise { const { diagnosisPath, plan2codeRoot, proposalsDir, runsDir, model = 'default', agent = 'claude-code' } = opts; // Load diagnosis let diagnosisContent: string; try { diagnosisContent = fs.readFileSync(diagnosisPath, 'utf8'); } catch { throw new Error(`Could not read diagnosis file at ${diagnosisPath}`); } // Read prompt files const promptContents = readPromptFiles(plan2codeRoot); const srcDir = path.join(plan2codeRoot, 'src'); // Build char counts for each file const charCounts = Object.entries(promptContents) .map(([file, content]) => `| ${file} | ${content.length} | ${CHAR_LIMIT} | ${CHAR_LIMIT - content.length} headroom |`) .join('\n'); const promptContentsStr = Object.entries(promptContents) .map(([file, content]) => `## ${file} (${content.length} chars)\n\n${content}`) .join('\n\n---\n\n'); // Load improve prompt template let improveTemplate: string; try { improveTemplate = fs.readFileSync(IMPROVE_PROMPT_PATH, 'utf8'); } catch { const altPath = path.join(process.cwd(), 'src', 'prompts', 'improve.md'); improveTemplate = fs.readFileSync(altPath, 'utf8'); } const fullPrompt = interpolate(improveTemplate, { diagnosisContent, promptContents: promptContentsStr, charCounts: `| File | Current Chars | Limit | Headroom |\n|------|--------------|-------|----------|\n${charCounts}`, }); // Invoke Claude console.log(`\nInvoking AI improvement proposal (agent: ${agent}, model: ${model === 'default' ? 'user default' : model})...`); console.log('This may take a minute...\n'); let aiResponse: string; try { aiResponse = await invokeLLM({ prompt: fullPrompt, model, agent, timeout: 300_000, }); } catch (err) { throw new Error(`AI invocation failed: ${err instanceof Error ? err.message : String(err)}`); } // Parse edits const rawEdits = parseProposalFromResponse(aiResponse); if (!rawEdits || rawEdits.length === 0) { throw new Error('Could not parse PromptEdit[] from AI response. The AI may not have produced a valid JSON block.'); } // Enforce max edits per cycle const MAX_EDITS = 5; if (rawEdits.length > MAX_EDITS) { console.warn(`\n⚠ AI generated ${rawEdits.length} edits (max is ${MAX_EDITS}). Truncating to first ${MAX_EDITS}.`); rawEdits.length = MAX_EDITS; } // Validate each edit const validationResults: ImproverResult['validationResults'] = []; const validEdits: PromptEdit[] = []; for (const edit of rawEdits) { const result = validateEdit(edit, promptContents); validationResults.push({ edit, result }); if (result.valid) { // Compute accurate char counts const fileContent = promptContents[edit.file] ?? ''; const afterContent = fileContent.replace(edit.old_text, edit.new_text); edit.char_count_before = fileContent.length; edit.char_count_after = afterContent.length; edit.char_count_delta = afterContent.length - fileContent.length; validEdits.push(edit); } else { console.warn(`\n⚠ Edit rejected for "${edit.file}":`); for (const err of result.errors) { console.warn(` - ${err}`); } } for (const warn of result.warnings) { console.warn(` Warning: ${warn}`); } } // Get run IDs that contributed to this analysis const runIds: string[] = []; try { const files = fs.readdirSync(runsDir) .filter(f => f.startsWith('run-') && f.endsWith('.json')); runIds.push(...files.map(f => f.replace('.json', ''))); } catch { /* no runs dir */ } // Build proposal const proposalId = generateProposalId(); const proposal: PromptProposal = { proposal_id: proposalId, created_at: new Date().toISOString(), based_on_runs: runIds, analyst_model: model, proposals: validEdits, status: 'pending', diagnosis_file: path.basename(diagnosisPath), }; // Save proposal JSON fs.mkdirSync(proposalsDir, { recursive: true }); const proposalPath = path.join(proposalsDir, `${proposalId}.json`); fs.writeFileSync(proposalPath, JSON.stringify(proposal, null, 2), 'utf8'); // Also save raw AI response alongside const rawPath = path.join(proposalsDir, `${proposalId}-raw.md`); fs.writeFileSync(rawPath, aiResponse, 'utf8'); return { proposalPath, proposal, validationResults, validEditCount: validEdits.length, invalidEditCount: rawEdits.length - validEdits.length, }; }