@provos/ironcurtain 0.5.1 → 0.7.0

package/dist/pipeline/compile.js

@@ -1,321 +1,30 @@
 /**
  * CLI entry point for `ironcurtain compile-policy`.
  *
- * Orchestrates the policy compilation pipeline (stages 1-3):
- *   1. Compile constitution into declarative rules via LLM
- *   2. Generate test scenarios via LLM
- *   3. Verify compiled policy against real engine + LLM judge
+ * Thin wrapper over PipelineRunner that handles CLI-specific concerns:
+ * loading config, printing the progress banner, and exit codes.
  *
- * Requires tool-annotations.json to exist on disk. Run
- * `npm run annotate-tools` first to generate it.
+ * The core compile-verify-repair loop lives in PipelineRunner.
  */
 import { resolve } from 'node:path';
 import { fileURLToPath } from 'node:url';
-import { Client } from '@modelcontextprotocol/sdk/client/index.js';
-import { StdioClientTransport } from '@modelcontextprotocol/sdk/client/stdio.js';
 import chalk from 'chalk';
-import { extractServerDomainAllowlists } from '../config/index.js';
-import { PolicyEngine } from '../trusted-process/policy-engine.js';
-import { resolveRealPath } from '../types/argument-roles.js';
-import { buildCompilerSystemPrompt, compileConstitution, ConstitutionCompilerSession, validateCompiledRules, } from './constitution-compiler.js';
-import { getHandwrittenScenarios } from './handwritten-scenarios.js';
-import { resolveAllLists } from './list-resolver.js';
-import { computeHash, createPipelineLlm, loadExistingArtifact, loadPipelineConfig, showCached, writeArtifact, withSpinner, } from './pipeline-shared.js';
-import { applyScenarioCorrections, buildJudgeSystemPrompt, extractScenarioCorrections, filterStructuralConflicts, PolicyVerifierSession, verifyPolicy, } from './policy-verifier.js';
-import { buildGeneratorSystemPrompt, ScenarioGeneratorSession } from './scenario-generator.js';
-import { VERSION } from '../version.js';
-// ---------------------------------------------------------------------------
-// Content-Hash Caching (compile-specific)
-// ---------------------------------------------------------------------------
-function computePolicyHash(systemPrompt, annotations) {
-    return computeHash(systemPrompt, JSON.stringify(annotations));
-}
-function computeScenariosHash(systemPrompt, handwrittenScenarios) {
-    return computeHash(systemPrompt, JSON.stringify(handwrittenScenarios));
-}
-function extractPermittedDirectories(rules) {
-    const dirs = new Set();
-    for (const rule of rules) {
-        if (rule.if.paths?.within) {
-            dirs.add(rule.if.paths.within);
-        }
-    }
-    return [...dirs].sort();
-}
-function collectProbeScenarios(result) {
-    return result.rounds.flatMap((round) => round.newScenarios);
-}
-/**
- * Filters scenarios against structural invariants and logs any discarded ones.
- * Returns the valid scenarios and the full list of discarded scenarios.
- */
-function filterAndLogStructuralConflicts(engine, scenarios, label = 'Discarded scenario (structural conflict)') {
-    const { valid, discarded } = filterStructuralConflicts(engine, scenarios);
-    for (const d of discarded) {
-        const prefix = d.scenario.source === 'handwritten'
-            ? chalk.yellow('Warning: handwritten scenario conflicts with structural invariant:')
-            : chalk.dim(`${label}:`);
-        console.error(`  ${prefix} "${d.scenario.description}" — ${d.rule} always returns ${d.actual}`);
-    }
-    return { valid, discarded };
-}
-// ---------------------------------------------------------------------------
-// Scenario Merge (multi-turn regeneration)
-// ---------------------------------------------------------------------------
-/**
- * Merges replacement scenarios from the regeneration session into the
- * scenario list. Removes corrected and discarded scenarios, then adds
- * unique replacements that don't duplicate any remaining scenario.
- */
-export function mergeReplacements(scenarios, replacements, corrections, discardedScenarios) {
-    const removedDescriptions = new Set([
-        ...corrections.map((c) => c.scenarioDescription),
-        ...discardedScenarios.filter((d) => d.scenario.source !== 'handwritten').map((d) => d.scenario.description),
-    ]);
-    const kept = scenarios.filter((s) => !removedDescriptions.has(s.description));
-    const keptDescriptions = new Set(kept.map((s) => s.description));
-    const uniqueReplacements = replacements.filter((r) => !keptDescriptions.has(r.description));
-    return [...kept, ...uniqueReplacements];
-}
-// ---------------------------------------------------------------------------
-// Path Resolution (post-LLM normalization)
-// ---------------------------------------------------------------------------
-/**
- * Resolves all `paths.within` values in compiled rules to their real
- * filesystem paths, following symlinks. This ensures that symlinked
- * directories (e.g., ~/Downloads -> /mnt/c/.../Downloads on WSL) are
- * resolved to their canonical form so that runtime path comparisons
- * match correctly.
- *
- * Falls back to path.resolve() if the path does not exist on disk.
- */
-export function resolveRulePaths(rules) {
-    return rules.map((rule) => {
-        if (!rule.if.paths?.within)
-            return rule;
-        const resolved = resolveRealPath(rule.if.paths.within);
-        if (resolved === rule.if.paths.within)
-            return rule;
-        return {
-            ...rule,
-            if: {
-                ...rule.if,
-                paths: { ...rule.if.paths, within: resolved },
-            },
-        };
-    });
-}
-async function compilePolicyRules(constitutionText, annotations, protectedPaths, inputHash, existingPolicy, llm, stepLabel = '[1/3]', system) {
-    const stepText = `${stepLabel} Compiling constitution`;
-    // Check cache: skip LLM call if inputs haven't changed.
-    // Still resolve paths in case symlink targets changed since last run.
-    if (existingPolicy && existingPolicy.inputHash === inputHash) {
-        showCached(stepText);
-        return {
-            rules: resolveRulePaths(existingPolicy.rules),
-            listDefinitions: existingPolicy.listDefinitions ?? [],
-            inputHash,
-        };
-    }
-    const session = new ConstitutionCompilerSession({
-        system: system ?? buildCompilerSystemPrompt(constitutionText, annotations, { protectedPaths }),
-        model: llm,
-        annotations,
-    });
-    const { result: compilationOutput } = await withSpinner(stepText, async (spinner) => {
-        const output = await session.compile((msg) => {
-            spinner.text = `${stepText} — ${msg}`;
-        });
-        const rules = resolveRulePaths(output.rules);
-        validateRulesOrThrow(rules, output.listDefinitions);
-        return { rules, listDefinitions: output.listDefinitions };
-    }, (output, elapsed) => `${stepText}: ${output.rules.length} rules compiled (${elapsed.toFixed(1)}s)`);
-    return {
-        rules: compilationOutput.rules,
-        listDefinitions: compilationOutput.listDefinitions,
-        inputHash,
-        session,
-    };
-}
-class RuleValidationError extends Error {
-    errors;
-    constructor(errors) {
-        super('Compiled rule validation failed');
-        this.errors = errors;
-        this.name = 'RuleValidationError';
-    }
-}
-function validateRulesOrThrow(rules, listDefinitions = []) {
-    const ruleValidation = validateCompiledRules(rules, listDefinitions);
-    if (ruleValidation.warnings.length > 0) {
-        for (const w of ruleValidation.warnings) {
-            console.error(`  ${chalk.yellow('Warning:')} ${w}`);
-        }
-    }
-    if (!ruleValidation.valid) {
-        throw new RuleValidationError(ruleValidation.errors);
-    }
-}
-async function compilePolicyRulesWithRepair(constitutionText, annotations, protectedPaths, baseInputHash, repairContext, llm, onProgress, system, session) {
-    let output;
-    if (session) {
-        output = await session.recompile(repairContext, onProgress);
-    }
-    else {
-        output = await compileConstitution(constitutionText, annotations, { protectedPaths }, llm, repairContext, onProgress, system);
-    }
-    const compiledRules = resolveRulePaths(output.rules);
-    validateRulesOrThrow(compiledRules, output.listDefinitions);
-    return {
-        rules: compiledRules,
-        listDefinitions: output.listDefinitions,
-        inputHash: `${baseInputHash}-repair`,
-        session,
-    };
-}
-// ---------------------------------------------------------------------------
-// Artifact Construction (pure data transformation)
-// ---------------------------------------------------------------------------
-function buildPolicyArtifact(constitutionHash, compilationResult) {
-    const artifact = {
-        generatedAt: new Date().toISOString(),
-        constitutionHash,
-        inputHash: compilationResult.inputHash,
-        rules: compilationResult.rules,
-    };
-    if (compilationResult.listDefinitions.length > 0) {
-        artifact.listDefinitions = compilationResult.listDefinitions;
-    }
-    return artifact;
-}
-async function generateTestScenarios(constitutionText, annotations, allowedDirectory, permittedDirectories, inputHash, existingScenarios, llm, stepLabel = '[2/3]', system) {
-    const handwrittenScenarios = getHandwrittenScenarios(allowedDirectory);
-    const stepText = `${stepLabel} Generating test scenarios`;
-    // Check cache: skip LLM call if inputs haven't changed
-    if (existingScenarios && existingScenarios.inputHash === inputHash) {
-        showCached(stepText);
-        return { scenarios: existingScenarios.scenarios, inputHash };
-    }
-    // Create a multi-turn session so the repair loop can feed corrections back
-    const session = new ScenarioGeneratorSession({
-        system: system ?? buildGeneratorSystemPrompt(constitutionText, annotations, allowedDirectory, permittedDirectories),
-        model: llm,
-        annotations,
-        handwrittenScenarios,
-    });
-    const { result: scenarios } = await withSpinner(stepText, async (spinner) => session.generate((msg) => {
-        spinner.text = `${stepText} — ${msg}`;
-    }), (scenarios, elapsed) => {
-        const generatedCount = scenarios.length - handwrittenScenarios.length;
-        return `${stepText}: ${scenarios.length} scenarios (${handwrittenScenarios.length} handwritten + ${generatedCount} generated) (${elapsed.toFixed(1)}s)`;
-    });
-    return { scenarios, inputHash, session };
-}
-// ---------------------------------------------------------------------------
-// Policy Verification (LLM step)
-// ---------------------------------------------------------------------------
-async function verifyCompiledPolicy(constitutionText, compiledPolicyFile, toolAnnotationsFile, protectedPaths, scenarios, llm, allowedDirectory, maxRounds = 3, verbose = true, onProgress, serverDomainAllowlists, dynamicLists, system, session) {
-    const result = await verifyPolicy(constitutionText, compiledPolicyFile, toolAnnotationsFile, protectedPaths, scenarios, llm, maxRounds, allowedDirectory, onProgress, serverDomainAllowlists, dynamicLists, system, session);
-    if (!result.pass) {
-        if (verbose) {
-            console.error('');
-            console.error(chalk.red('Verification FAILED:'));
-            console.error(result.summary);
-            console.error('');
-            for (const f of result.failedScenarios) {
-                console.error(`  ${chalk.red('FAIL:')} ${f.scenario.description}`);
-                console.error(`    Expected: ${f.scenario.expectedDecision}, Got: ${f.actualDecision} (rule: ${f.matchingRule})`);
-            }
-        }
-        else {
-            console.error(`  ${result.failedScenarios.length} scenario(s) failed.`);
-        }
-    }
-    return result;
-}
-// ---------------------------------------------------------------------------
-// Artifact Output
-// ---------------------------------------------------------------------------
-function writeScenariosArtifact(generatedDir, constitutionHash, scenarioResult) {
-    const scenariosFile = {
-        generatedAt: new Date().toISOString(),
-        constitutionHash,
-        inputHash: scenarioResult.inputHash,
-        scenarios: scenarioResult.scenarios,
-    };
-    writeArtifact(generatedDir, 'test-scenarios.json', scenariosFile);
-}
-function writePolicyArtifact(generatedDir, compiledPolicyFile) {
-    writeArtifact(generatedDir, 'compiled-policy.json', compiledPolicyFile);
-}
-// ---------------------------------------------------------------------------
-// MCP Server Connection for Data-Backed Lists
-// ---------------------------------------------------------------------------
-/**
- * Connects to MCP servers needed for data-backed list resolution.
- * Only connects to servers hinted by list definitions with requiresMcp: true.
- * Returns a map keyed by server name for the resolver.
- */
-export async function connectMcpServersForLists(definitions, mcpServers) {
-    const mcpDefs = definitions.filter((d) => d.requiresMcp);
-    const hasUnhintedLists = mcpDefs.some((d) => !d.mcpServerHint);
-    // Connect all configured servers if any list lacks a hint,
-    // otherwise connect only the hinted servers.
-    const neededServers = hasUnhintedLists
-        ? new Set(Object.keys(mcpServers))
-        : new Set(mcpDefs
-            .filter((d) => d.mcpServerHint != null)
-            .map((d) => d.mcpServerHint));
-    const connections = new Map();
-    for (const serverName of neededServers) {
-        const serverConfig = mcpServers[serverName];
-        // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition -- serverName may not exist in mcpServers at runtime (from mcpServerHint)
-        if (!serverConfig) {
-            console.error(`  ${chalk.yellow('Warning:')} MCP server "${serverName}" not configured — skipping`);
-            continue;
-        }
-        const transport = new StdioClientTransport({
-            command: serverConfig.command,
-            args: serverConfig.args,
-            env: serverConfig.env ? { ...process.env, ...serverConfig.env } : undefined,
-            stderr: 'pipe',
-        });
-        // Drain piped stderr to prevent backpressure
-        if (transport.stderr) {
-            transport.stderr.on('data', () => { });
-        }
-        const client = new Client({ name: 'ironcurtain-list-resolver', version: VERSION });
-        await client.connect(transport);
-        const toolsResult = await client.listTools();
-        connections.set(serverName, { client, tools: toolsResult.tools });
-    }
-    return connections;
-}
-export async function disconnectMcpServers(connections) {
-    for (const conn of connections.values()) {
-        try {
-            await conn.client.close();
-        }
-        catch {
-            // Ignore cleanup errors
-        }
-    }
-}
+import { loadPipelineConfig, loadToolAnnotationsFile } from './pipeline-shared.js';
+import { PipelineRunner, createPipelineModels } from './pipeline-runner.js';
+// Re-export utilities that test files import from this file.
+export { resolveRulePaths, mergeReplacements } from './pipeline-shared.js';
 // ---------------------------------------------------------------------------
 // Main Pipeline
 // ---------------------------------------------------------------------------
 export async function main() {
     const config = loadPipelineConfig();
-    // Load tool annotations from disk (produced by `npm run annotate-tools`)
-    const toolAnnotationsFile = loadExistingArtifact(config.generatedDir, 'tool-annotations.json', config.packageGeneratedDir);
+    // Load tool annotations early to validate they exist before printing the banner
+    const toolAnnotationsFile = loadToolAnnotationsFile(config.generatedDir, config.packageGeneratedDir);
     if (!toolAnnotationsFile) {
         console.error(chalk.red.bold("Error: tool-annotations.json not found. Run 'npm run annotate-tools' first to generate tool annotations."));
         process.exit(1);
     }
-    // Flatten annotations from all servers
     const allAnnotations = Object.values(toolAnnotationsFile.servers).flatMap((server) => server.tools);
-    // Compute domain allowlists for policy verification
-    const serverDomainAllowlists = extractServerDomainAllowlists(config.mcpServers);
     console.error(chalk.bold('Policy Compilation Pipeline'));
     console.error(chalk.bold('==========================='));
     console.error(`Constitution: ${chalk.dim(config.constitutionPath)}`);
@@ -323,278 +32,30 @@ export async function main() {
     console.error(`Output:       ${chalk.dim(config.generatedDir + '/')}`);
     console.error(`Annotations:  ${chalk.dim(`${allAnnotations.length} tools from ${Object.keys(toolAnnotationsFile.servers).length} server(s)`)}`);
     console.error('');
-    const { model: llm, logContext, logPath, cacheStrategy } = await createPipelineLlm(config.generatedDir, 'unknown');
-    // Build raw system prompt and derive hash + cache-wrapped version
-    const handwrittenScenarios = getHandwrittenScenarios(config.allowedDirectory);
-    const compilerPrompt = buildCompilerSystemPrompt(config.constitutionText, allAnnotations, { protectedPaths: config.protectedPaths }, handwrittenScenarios);
-    const compilerHash = computePolicyHash(compilerPrompt, allAnnotations);
-    const compilerSystem = cacheStrategy.wrapSystemPrompt(compilerPrompt);
-    // Load existing artifacts for cache comparison
-    const existingPolicy = loadExistingArtifact(config.generatedDir, 'compiled-policy.json', config.packageGeneratedDir);
-    const existingScenarios = loadExistingArtifact(config.generatedDir, 'test-scenarios.json', config.packageGeneratedDir);
-    // Compile constitution into policy rules (LLM-cacheable)
-    // Step numbering depends on whether list definitions are emitted,
-    // so compilation always uses [1/N] and the total is determined after.
-    logContext.stepName = 'compile-constitution';
-    let compilationResult = await compilePolicyRules(config.constitutionText, allAnnotations, config.protectedPaths, compilerHash, existingPolicy, llm, '[1/3]', compilerSystem);
-    // Build and write policy artifact immediately so it's available for
-    // inspection even if verification fails, and cached for next run.
-    let compiledPolicyFile = buildPolicyArtifact(config.constitutionHash, compilationResult);
-    writePolicyArtifact(config.generatedDir, compiledPolicyFile);
-    // Resolve dynamic lists if the compiler emitted list definitions
-    const hasLists = compilationResult.listDefinitions.length > 0;
-    const totalSteps = hasLists ? 4 : 3;
-    let dynamicLists;
-    if (hasLists) {
-        logContext.stepName = 'resolve-lists';
-        const existingLists = loadExistingArtifact(config.generatedDir, 'dynamic-lists.json', config.packageGeneratedDir);
-        // Connect to MCP servers if any list definitions require it
-        const needsMcp = compilationResult.listDefinitions.some((d) => d.requiresMcp);
-        let mcpConnections;
-        if (needsMcp) {
-            mcpConnections = await connectMcpServersForLists(compilationResult.listDefinitions, config.mcpServers);
-        }
-        const listStepText = `[2/${totalSteps}] Resolving dynamic lists`;
-        try {
-            const { result: resolvedLists } = await withSpinner(listStepText, async (spinner) => resolveAllLists(compilationResult.listDefinitions, { model: llm, mcpConnections }, existingLists, (msg) => {
-                spinner.text = `${listStepText} — ${msg}`;
-            }), (result, elapsed) => {
-                const count = Object.keys(result.lists).length;
-                return `${listStepText}: ${count} list(s) resolved (${elapsed.toFixed(1)}s)`;
-            });
-            dynamicLists = resolvedLists;
-            writeArtifact(config.generatedDir, 'dynamic-lists.json', dynamicLists);
-        }
-        finally {
-            if (mcpConnections) {
-                await disconnectMcpServers(mcpConnections);
-            }
-        }
+    const models = await createPipelineModels(config.generatedDir);
+    const runner = new PipelineRunner(models);
+    try {
+        await runner.run({
+            constitutionInput: config.constitutionText,
+            constitutionKind: 'constitution',
+            outputDir: config.generatedDir,
+            toolAnnotationsDir: config.generatedDir,
+            toolAnnotationsFallbackDir: config.packageGeneratedDir,
+            allowedDirectory: config.allowedDirectory,
+            protectedPaths: config.protectedPaths,
+            mcpServers: config.mcpServers,
+            llmLogPath: models.logPath,
+            preloadedToolAnnotations: toolAnnotationsFile,
+        });
     }
-    // Extract permitted directories from compiled rules for scenario generation
-    const permittedDirectories = extractPermittedDirectories(compilationResult.rules);
-    // Generate test scenarios (LLM-cacheable)
-    const scenarioStepLabel = `[${hasLists ? 3 : 2}/${totalSteps}]`;
-    logContext.stepName = 'generate-scenarios';
-    const scenarioPrompt = buildGeneratorSystemPrompt(config.constitutionText, allAnnotations, config.allowedDirectory, permittedDirectories, dynamicLists);
-    const scenarioHash = computeScenariosHash(scenarioPrompt, getHandwrittenScenarios(config.allowedDirectory));
-    const scenarioSystem = cacheStrategy.wrapSystemPrompt(scenarioPrompt);
-    const scenarioResult = await generateTestScenarios(config.constitutionText, allAnnotations, config.allowedDirectory, permittedDirectories, scenarioHash, existingScenarios, llm, scenarioStepLabel, scenarioSystem);
-    // Write scenarios to disk immediately so they're available for
-    // inspection even if verification fails, and cached for next run.
-    writeScenariosArtifact(config.generatedDir, config.constitutionHash, scenarioResult);
-    // Filter out scenarios that conflict with structural invariants
-    const filterEngine = new PolicyEngine(compiledPolicyFile, toolAnnotationsFile, config.protectedPaths, config.allowedDirectory, undefined, dynamicLists);
-    const { valid: initialValid, discarded: discardedScenarios } = filterAndLogStructuralConflicts(filterEngine, scenarioResult.scenarios, 'Discarded scenario (structural conflict)');
-    let filteredScenarios = initialValid;
-    // Verify compiled policy against scenarios (full depth)
-    const verifyStepLabel = `[${totalSteps}/${totalSteps}]`;
-    logContext.stepName = 'verify-policy';
-    const allAvailableTools = allAnnotations.map((a) => ({ serverName: a.serverName, toolName: a.toolName }));
-    const serverNamesList = [...new Set(allAnnotations.map((a) => a.serverName))];
-    const toolNamesList = [...new Set(allAnnotations.map((a) => a.toolName))];
-    let verifierSystem = cacheStrategy.wrapSystemPrompt(buildJudgeSystemPrompt(config.constitutionText, compiledPolicyFile, config.protectedPaths, allAvailableTools, dynamicLists, config.allowedDirectory));
-    let verifierSession = new PolicyVerifierSession({
-        system: verifierSystem,
-        model: llm,
-        serverNames: serverNamesList,
-        toolNames: toolNamesList,
-    });
-    const { result: verificationResultInitial } = await withSpinner(`${verifyStepLabel} Verifying policy`, async (spinner) => verifyCompiledPolicy(config.constitutionText, compiledPolicyFile, toolAnnotationsFile, config.protectedPaths, filteredScenarios, llm, config.allowedDirectory, 3, true, (msg) => {
-        spinner.text = `${verifyStepLabel} Verifying policy — ${msg}`;
-    }, serverDomainAllowlists, dynamicLists, verifierSystem, verifierSession), (r, elapsed) => r.pass
-        ? `${verifyStepLabel} Verified policy: ${r.rounds.length} round(s) (${elapsed.toFixed(1)}s)`
-        : `${verifyStepLabel} Verification completed with failures (${elapsed.toFixed(1)}s)`);
-    let verificationResult = verificationResultInitial;
-    // Collect probe scenarios from verifier across all attempts, filtering
-    // out any that conflict with structural invariants
-    const { valid: filteredInitialProbes } = filterAndLogStructuralConflicts(filterEngine, collectProbeScenarios(verificationResult), 'Discarded probe (structural conflict)');
-    const accumulatedProbes = filteredInitialProbes;
-    // Dual-channel compile-verify-repair loop (up to 2 repair attempts)
-    // The judge attributes each failure to 'rule', 'scenario', or 'both':
-    //   - scenario-blamed: patch the test expectation
-    //   - rule-blamed: recompile rules with failure feedback
-    //   - both: patch scenario AND recompile
-    const MAX_REPAIRS = 2;
-    let repairAttempts = 0;
-    let scenarioCorrectionsApplied = 0;
-    if (!verificationResult.pass) {
-        const baseInputHash = compilationResult.inputHash;
-        for (let attempt = 1; attempt <= MAX_REPAIRS; attempt++) {
+    catch (err) {
+        if (err instanceof Error && err.message.includes('Verification FAILED')) {
             console.error('');
-            // Gather judge analysis and attributions from the most recent verification
-            const lastRound = verificationResult.rounds[verificationResult.rounds.length - 1];
-            const judgeAnalysis = lastRound?.llmAnalysis ?? verificationResult.summary;
-            const attributedFailures = lastRound?.attributedFailures ?? [];
-            // Split failures into scenario corrections and rule-blamed failures
-            const allScenarios = [...scenarioResult.scenarios, ...accumulatedProbes];
-            const { corrections, handwrittenWarnings } = extractScenarioCorrections(attributedFailures, allScenarios);
-            for (const warning of handwrittenWarnings) {
-                console.error(`  ${chalk.yellow('Warning:')} ${warning}`);
-            }
-            // Apply scenario corrections (patch expectations on generated scenarios)
-            if (corrections.length > 0) {
-                scenarioResult.scenarios = applyScenarioCorrections(scenarioResult.scenarios, corrections);
-                const correctedProbes = applyScenarioCorrections(accumulatedProbes, corrections);
-                accumulatedProbes.splice(0, accumulatedProbes.length, ...correctedProbes);
-                scenarioCorrectionsApplied += corrections.length;
-                console.error(`  ${chalk.dim(`Corrected ${corrections.length} scenario expectation(s)`)}`);
-            }
-            // Feed corrections back to the scenario generator session so it can
-            // produce better replacement scenarios that avoid the same mistakes.
-            const { session } = scenarioResult;
-            if (session && (corrections.length > 0 || discardedScenarios.length > 0 || accumulatedProbes.length > 0)) {
-                const feedback = {
-                    corrections,
-                    discardedScenarios,
-                    probeScenarios: accumulatedProbes,
-                };
-                logContext.stepName = `repair-regenerate-${attempt}`;
-                const regenText = `Repair ${attempt}/${MAX_REPAIRS}: Regenerating scenarios`;
-                const { result: replacements } = await withSpinner(regenText, async (spinner) => session.regenerate(feedback, (msg) => {
-                    spinner.text = `${regenText} — ${msg}`;
-                }), (r, elapsed) => `${regenText}: ${r.length} replacement(s) (${elapsed.toFixed(1)}s)`);
-                // Merge: remove corrected/discarded scenarios and add unique replacements
-                scenarioResult.scenarios = mergeReplacements(scenarioResult.scenarios, replacements, corrections, discardedScenarios);
-            }
-            // Re-filter after corrections/regeneration (a corrected expected decision
-            // might now match a structural invariant result, making it valid)
-            ({ valid: filteredScenarios } = filterAndLogStructuralConflicts(filterEngine, scenarioResult.scenarios));
-            // Determine which failures need rule recompilation:
-            // - no attribution or blamed on 'rule'/'both' (conservative default)
-            // - handwritten scenarios blamed on 'scenario' (reclassified as rule issues)
-            const allRuleBlamedFailures = verificationResult.failedScenarios.filter((f) => {
-                const attr = attributedFailures.find((a) => a.scenarioDescription === f.scenario.description);
-                if (!attr || attr.blame.kind === 'rule' || attr.blame.kind === 'both')
-                    return true;
-                return handwrittenWarnings.some((w) => w.includes(f.scenario.description));
-            });
-            if (allRuleBlamedFailures.length > 0) {
-                // Recompile with failure feedback (always calls LLM, no cache)
-                const repairContext = {
-                    failedScenarios: allRuleBlamedFailures,
-                    judgeAnalysis,
-                    attemptNumber: attempt,
-                    existingListDefinitions: compilationResult.listDefinitions.length > 0 ? compilationResult.listDefinitions : undefined,
-                    handwrittenScenarios,
-                };
-                logContext.stepName = `repair-compile-${attempt}`;
-                const repairCompileText = `Repair ${attempt}/${MAX_REPAIRS}: Recompiling`;
-                const { result: repairCompileResult } = await withSpinner(repairCompileText, async (spinner) => compilePolicyRulesWithRepair(config.constitutionText, allAnnotations, config.protectedPaths, baseInputHash, repairContext, llm, (msg) => {
-                    spinner.text = `${repairCompileText} — ${msg}`;
-                }, compilerSystem, compilationResult.session), (r, elapsed) => `Repair ${attempt}/${MAX_REPAIRS}: Recompiled ${r.rules.length} rules (${elapsed.toFixed(1)}s)`);
-                compilationResult = repairCompileResult;
-                // Re-resolve dynamic lists if repair introduced new list definitions
-                if (dynamicLists && compilationResult.listDefinitions.length > 0) {
-                    const currentLists = dynamicLists;
-                    const newListDefs = compilationResult.listDefinitions.filter((def) => !(def.name in currentLists.lists));
-                    if (newListDefs.length > 0) {
-                        const mcpRequired = newListDefs.filter((d) => d.requiresMcp);
-                        if (mcpRequired.length > 0) {
-                            console.error(`  ${chalk.yellow('Warning:')} Repair introduced ${mcpRequired.length} new MCP-requiring list(s) — skipping resolution`);
-                        }
-                        const knowledgeDefs = newListDefs.filter((d) => !d.requiresMcp);
-                        if (knowledgeDefs.length > 0) {
-                            console.error(`  ${chalk.dim(`Resolving ${knowledgeDefs.length} new list(s) from repair...`)}`);
-                            const resolved = await resolveAllLists(knowledgeDefs, { model: llm }, currentLists);
-                            dynamicLists = {
-                                ...resolved,
-                                lists: { ...currentLists.lists, ...resolved.lists },
-                            };
-                            writeArtifact(config.generatedDir, 'dynamic-lists.json', dynamicLists);
-                        }
-                    }
-                }
-                // Write updated policy artifact
-                compiledPolicyFile = buildPolicyArtifact(config.constitutionHash, compilationResult);
-                writePolicyArtifact(config.generatedDir, compiledPolicyFile);
-                // Rebuild verifier system prompt and session with updated rules
-                verifierSystem = cacheStrategy.wrapSystemPrompt(buildJudgeSystemPrompt(config.constitutionText, compiledPolicyFile, config.protectedPaths, allAvailableTools, dynamicLists, config.allowedDirectory));
-                verifierSession = new PolicyVerifierSession({
-                    system: verifierSystem,
-                    model: llm,
-                    serverNames: serverNamesList,
-                    toolNames: toolNamesList,
-                });
-            }
-            else {
-                console.error(`  ${chalk.dim('No rule-blamed failures — skipping recompilation')}`);
-            }
-            // Verify with reduced depth, using base scenarios + accumulated probes
-            logContext.stepName = `repair-verify-${attempt}`;
-            const repairScenarios = [...filteredScenarios, ...accumulatedProbes];
-            const repairVerifyText = `Repair ${attempt}/${MAX_REPAIRS}: Verifying`;
-            const { result: repairVerifyResult } = await withSpinner(repairVerifyText, async (spinner) => verifyCompiledPolicy(config.constitutionText, compiledPolicyFile, toolAnnotationsFile, config.protectedPaths, repairScenarios, llm, config.allowedDirectory, 1, true, (msg) => {
-                spinner.text = `${repairVerifyText} — ${msg}`;
-            }, serverDomainAllowlists, dynamicLists, verifierSystem, verifierSession), (r, elapsed) => r.pass
-                ? `Repair ${attempt}/${MAX_REPAIRS}: Verified (${elapsed.toFixed(1)}s)`
-                : `Repair ${attempt}/${MAX_REPAIRS}: ${r.failedScenarios.length} failure(s) (${elapsed.toFixed(1)}s)`);
-            verificationResult = repairVerifyResult;
-            // Accumulate any new probe scenarios, filtering structural conflicts
-            const { valid: validRepairProbes } = filterAndLogStructuralConflicts(filterEngine, collectProbeScenarios(verificationResult), 'Discarded probe (structural conflict)');
-            accumulatedProbes.push(...validRepairProbes);
-            repairAttempts = attempt;
-            if (verificationResult.pass) {
-                // Run final full verification with all accumulated scenarios
-                logContext.stepName = 'final-verify';
-                const finalScenarios = [...filteredScenarios, ...accumulatedProbes];
-                // New session for final verification (fresh history for clean evaluation)
-                const finalSession = new PolicyVerifierSession({
-                    system: verifierSystem,
-                    model: llm,
-                    serverNames: serverNamesList,
-                    toolNames: toolNamesList,
-                });
-                const { result: finalVerifyResult } = await withSpinner('Final full verification', async (spinner) => verifyCompiledPolicy(config.constitutionText, compiledPolicyFile, toolAnnotationsFile, config.protectedPaths, finalScenarios, llm, config.allowedDirectory, 3, true, (msg) => {
-                    spinner.text = `Final full verification — ${msg}`;
-                }, serverDomainAllowlists, dynamicLists, verifierSystem, finalSession), (r, elapsed) => r.pass
-                    ? `Final full verification: passed (${elapsed.toFixed(1)}s)`
-                    : `Final full verification: ${r.failedScenarios.length} failure(s) (${elapsed.toFixed(1)}s)`);
-                verificationResult = finalVerifyResult;
-                // Accumulate any new probes from final verification, filtering structural conflicts
-                const { valid: validFinalProbes } = filterAndLogStructuralConflicts(filterEngine, collectProbeScenarios(verificationResult), 'Discarded probe (structural conflict)');
-                accumulatedProbes.push(...validFinalProbes);
-                break;
-            }
+            console.error(chalk.red.bold('Verification FAILED — artifacts written but policy may need review.'));
+            process.exit(1);
         }
+        throw err;
     }
-    // Re-write scenarios artifact if the repair loop modified them
-    if (repairAttempts > 0) {
-        writeScenariosArtifact(config.generatedDir, config.constitutionHash, scenarioResult);
-    }
-    // Deduplicate accumulated probes by description
-    const seenDescriptions = new Set(scenarioResult.scenarios.map((s) => s.description));
-    const uniqueProbes = accumulatedProbes.filter((s) => {
-        if (seenDescriptions.has(s.description))
-            return false;
-        seenDescriptions.add(s.description);
-        return true;
-    });
-    const totalScenariosTested = filteredScenarios.length + uniqueProbes.length;
-    console.error('');
-    console.error(`  Rules: ${compilationResult.rules.length}`);
-    console.error(`  Scenarios tested: ${totalScenariosTested}`);
-    if (discardedScenarios.length > 0) {
-        console.error(`  Scenarios discarded (structural conflicts): ${discardedScenarios.length}`);
-    }
-    if (uniqueProbes.length > 0) {
-        console.error(`  Probe scenarios accumulated: ${uniqueProbes.length}`);
-    }
-    if (repairAttempts > 0) {
-        console.error(`  Repair attempts: ${repairAttempts}`);
-    }
-    if (scenarioCorrectionsApplied > 0) {
-        console.error(`  Scenario corrections: ${scenarioCorrectionsApplied}`);
-    }
-    console.error(`  Artifacts written to: ${chalk.dim(config.generatedDir + '/')}`);
-    console.error(`  LLM interaction log: ${chalk.dim(logPath)}`);
-    if (!verificationResult.pass) {
-        console.error('');
-        console.error(chalk.red.bold('Verification FAILED — artifacts written but policy may need review.'));
-        process.exit(1);
-    }
-    console.error('');
-    console.error(chalk.green.bold('Policy compilation successful!'));
 }
 // Only run when executed directly (not when imported by cli.ts or tests)
 if (process.argv[1] && resolve(process.argv[1]) === fileURLToPath(import.meta.url)) {