popeye-cli 1.10.0 → 2.1.0

package/skills/POPEYE_FULL_AUTONOMY_PIPELINE.md

@@ -1,6 +1,6 @@
 # Popeye Full Autonomy Pipeline Spec
-Version: 1.0
-Goal: 1 idea → 1 prompt → Popeye → production-ready system (PASS Production Gate)
+Version: 1.1 (Autonomy Hardening Gap Fixes)
+Goal: 1 idea -> 1 prompt -> Popeye -> production-ready system (PASS Production Gate)
 
 This spec defines the end-to-end autonomous workflow Popeye must execute, including:
 - Phase sequencing (no skipping)
@@ -90,6 +90,27 @@ Artifacts are never overwritten. New versions create new files.
 - Each phase may only transition to the next if all required artifacts exist and the gate passes.
 - Any failure routes to RECOVERY_LOOP, except missing inputs which route back to the phase that needs them.
 
+### v1.1 Gate Behaviors (Autonomy Hardening)
+
+Between phase execution and transition, the orchestrator now applies three additional checks:
+
+1. **Constitution Verification**: Before every `evaluateGate()` call, verifies that `skills/POPEYE_CONSTITUTION.md` has not been modified since pipeline start (SHA-256 hash comparison). If the constitution is invalid, the gate is blocked.
+
+2. **Gate Result Merge**: After gate evaluation, preserves `score`/`consensusScore` values stored by consensus phase handlers while updating `pass`/`blockers` from the gate engine. Prevents overwriting of consensus scores.
+
+3. **Change Request (CR) Routing**: After REVIEW and AUDIT phases pass their gates, checks `pipeline.pendingChangeRequests` for any `proposed` CRs. If found, deterministically routes to the appropriate consensus phase (based on change type) rather than continuing normal progression. CR lifecycle: `proposed` -> `approved` (when routed) or `rejected`.
+
+### v1.1 Pipeline State Additions
+
+```
+pendingChangeRequests?: Array<{
+  cr_id: string;
+  change_type: 'scope' | 'architecture' | 'dependency' | 'config' | 'requirement';
+  target_phase: PipelinePhase;
+  status: 'proposed' | 'approved' | 'rejected';
+}>;
+```
+
 ---
 
 ## 3) Phase-by-Phase Specification
@@ -321,15 +342,24 @@ Artifacts are never overwritten. New versions create new files.
   - evidence references
   - no shortcuts
   - integration wiring
+- Generate fresh repo snapshot and diff against baseline (role-plan-approval snapshot)
+- Detect implementation drift (config changes, significant line deltas)
 
 **Outputs**
-- Review decision doc → `/docs/consensus/review_<id>_<date>.md`
+- Review decision doc -> `/docs/consensus/review_<id>_<date>.md`
 - If rejected: structured blocking list
+- v1.1: Change Request artifacts for detected drift (stored as `change_request` artifacts)
+- v1.1: Pending CRs registered in `pipeline.pendingChangeRequests` for orchestrator routing
 
 **Gate conditions (PASS)**
 - Reviewer approves with evidence
 - No Constitution violations
 
+**v1.1 Post-Gate Behavior**
+- After REVIEW gate passes, orchestrator checks `pendingChangeRequests` for proposed CRs
+- If CRs exist, routes to the appropriate consensus phase (e.g., CONSENSUS_MASTER_PLAN for scope drift)
+- If no CRs, continues to AUDIT normally
+
 ---
 
 ### Phase 10 — AUDIT
@@ -339,7 +369,7 @@ Artifacts are never overwritten. New versions create new files.
 - Everything above + repo snapshot + logs
 
 **Actions (Auditor)**
-- Integration audit (FE↔BE, BE↔DB)
+- Integration audit (FE<->BE, BE<->DB)
 - Config/env audit
 - Tests/coverage audit
 - Migration audit
@@ -347,8 +377,10 @@ Artifacts are never overwritten. New versions create new files.
 - Deployment readiness audit
 
 **Outputs (required)**
-- Audit Report → `/docs/audit/audit_<id>_<date>.md`
-- Audit Report schema (optional json)
+- Audit Report -> `/docs/audit/audit_<id>_<date>.md`
+- Audit Report schema (JSON)
+- v1.1: Change Request artifacts for blocking architectural/security findings
+- v1.1: Pending CRs registered in `pipeline.pendingChangeRequests` for orchestrator routing
 
 **Gate conditions (PASS)**
 - No P0/P1 findings open
@@ -356,6 +388,13 @@ Artifacts are never overwritten. New versions create new files.
 - No hardcoded secrets
 - End-to-end wiring verified
 
+**v1.1 Post-Gate Behavior**
+- After AUDIT gate passes, orchestrator checks `pendingChangeRequests` for proposed CRs
+- Architectural findings (integration/schema blocking issues) create CRs targeting CONSENSUS_ARCHITECTURE
+- Security findings create CRs targeting CONSENSUS_MASTER_PLAN
+- If CRs exist, routes to the appropriate consensus phase before PRODUCTION_GATE
+- If no CRs, continues to PRODUCTION_GATE normally
+
 **If FAIL**
 - Enter RECOVERY_LOOP
 
@@ -389,7 +428,7 @@ Artifacts are never overwritten. New versions create new files.
 
 ---
 
-## 4) Recovery Loop (FAIL → Debugger → Plan → Fix → Retest)
+## 4) Recovery Loop (FAIL -> Debugger -> Plan -> Fix -> Retest)
 
 ### Phase 12 — RECOVERY_LOOP (conditional)
 **Purpose**: Self-heal deterministically using RCA, not guesswork.
@@ -406,6 +445,7 @@ Artifacts are never overwritten. New versions create new files.
    - ownership (which role must fix)
    - recommended corrective actions
    - prevention suggestion
+   - v1.1: `requires_phase_rewind_to` field specifying which phase to rewind to
 
 2) **Dispatcher generates Recovery Plan**
    - targeted tasks mapped to responsible role(s)
@@ -416,19 +456,23 @@ Artifacts are never overwritten. New versions create new files.
 
 4) **Implement fixes** (by responsible role)
 
-5) **Retest only the failed gate first**, then proceed forward:
-   - If failed at Audit → rerun Audit
-   - If failed at Production Gate → rerun Production checks
-   - If failed tests → rerun tests + QA validation
+5) **Post-recovery routing** (v1.1 — RCA-driven rewind):
+   - Orchestrator reads latest RCA JSON artifact from disk
+   - If RCA specifies `requires_phase_rewind_to`, pipeline rewinds to that phase
+   - Otherwise, retests the originally failed phase, then proceeds forward:
+     - If failed at Audit -> rerun Audit
+     - If failed at Production Gate -> rerun Production checks
+     - If failed tests -> rerun tests + QA validation
+   - If no failed phase is tracked, defaults to QA_VALIDATION
 
 **Outputs (required)**
-- RCA report → `/docs/incidents/rca_<id>_<date>.md`
+- RCA report -> `/docs/incidents/rca_<id>_<date>.md` (both markdown and JSON, JSON includes `requires_phase_rewind_to`)
 - Recovery plan packet + consensus packet
 - Updated artifacts if scope/contracts changed
 
 **Stop conditions**
 - Max recovery iterations (default: 5)
-- If exceeded → STUCK state with “Stuck Report”
+- If exceeded -> STUCK state with “Stuck Report”
 
 ---
 
@@ -479,6 +523,15 @@ If recovery iterations exceed max, Popeye must stop and output:
 - Every plan claim must be evidenced against repo snapshot
 - Production readiness is binary, not “looks good”
 
+### v1.1 Additions
+
+- Constitution integrity is verified at every gate (SHA-256 hash comparison). A modified constitution blocks all gate progression until resolved.
+- Consensus scores from phase handlers are never overwritten by gate engine evaluation (gate result merge rule).
+- REVIEW and AUDIT phases must register detected issues as Change Requests in `pipeline.pendingChangeRequests`.
+- CRs are routed deterministically by the orchestrator (not advisory). The routing is a deterministic transition, not a suggestion.
+- RCA packets must include a `requires_phase_rewind_to` field when the root cause originates in a phase earlier than the one that failed. The orchestrator reads this from the JSON artifact on disk.
+- CRs are processed one at a time (first `proposed` CR is routed and marked `approved` before the next is considered).
+
 ---
 
 End of Spec.

package/src/adapters/gemini.ts

@@ -16,7 +16,7 @@ export type GeminiModel = string;
  * Default Gemini configuration
  */
 export const DEFAULT_GEMINI_CONFIG = {
-  model: 'gemini-2.0-flash' as GeminiModel,
+  model: 'gemini-2.5-flash' as GeminiModel,
   temperature: 0.3,
   maxTokens: 4096,
 };
@@ -94,7 +94,7 @@ export async function requestArbitration(
   scores: number[]
 ): Promise<ArbitrationResult> {
   const client = await createClient();
-  const generativeModel = client.getGenerativeModel({ model: 'gemini-2.0-flash' });
+  const generativeModel = client.getGenerativeModel({ model: 'gemini-2.5-flash' });
 
   const prompt = buildArbitrationPrompt(plan, reviewerFeedback, claudeFeedback, iterations, scores);
 
@@ -370,7 +370,7 @@ function parseList(text: string): string[] {
 export async function validateApiKey(): Promise<boolean> {
   try {
     const client = await createClient();
-    const model = client.getGenerativeModel({ model: 'gemini-2.0-flash' });
+    const model = client.getGenerativeModel({ model: 'gemini-2.5-flash' });
     await model.generateContent('Say "OK" if you can hear me.');
     return true;
   } catch {

package/src/adapters/openai.ts

@@ -322,7 +322,7 @@ Be specific and actionable. The specification should be detailed enough that a d
   }
 
   const completion = await client.chat.completions.create({
-    model: 'gpt-4o',
+    model: 'gpt-4.1',
     messages: [{ role: 'user', content: prompt }],
     temperature: 0.7,
     max_tokens: 4096,
@@ -360,7 +360,7 @@ Provide:
 Format your score as: CONSENSUS: [X]%`;
 
   const completion = await client.chat.completions.create({
-    model: 'gpt-4o',
+    model: 'gpt-4.1',
     messages: [{ role: 'user', content: prompt }],
     temperature: 0.3,
     max_tokens: 2048,

package/src/auth/gemini.ts

@@ -30,7 +30,7 @@ export interface GeminiAuthStatus {
 export async function validateGeminiToken(apiKey: string): Promise<boolean> {
   try {
     const client = new GoogleGenerativeAI(apiKey);
-    const model = client.getGenerativeModel({ model: 'gemini-2.0-flash' });
+    const model = client.getGenerativeModel({ model: 'gemini-2.5-flash' });
     // Test the key by making a simple request
     await model.generateContent('Say "OK"');
     return true;

package/src/cli/commands/create.ts

@@ -10,6 +10,7 @@ import { ProjectSpecSchema, type OutputLanguage, type OpenAIModel } from '../../
 import { requireAuth } from '../../auth/index.js';
 import { runWorkflow } from '../../workflow/index.js';
 import { generateProject, projectDirExists, cleanupProject } from '../../generators/index.js';
+import { readPopeyeMdConfig } from '../../config/popeye-md.js';
 import {
   printHeader,
   printSection,
@@ -41,8 +42,8 @@ export function createCreateCommand(): Command {
     )
     .option(
       '-m, --model <model>',
-      'OpenAI model for consensus (gpt-4o, gpt-4o-mini, o1-preview)',
-      'gpt-4o'
+      'OpenAI model for consensus (gpt-4.1, gpt-4o, o3, o4-mini)',
+      'gpt-4.1'
     )
     .option(
       '-o, --output <dir>',
@@ -93,9 +94,8 @@ export function createCreateCommand(): Command {
         }
 
         const model = options.model as OpenAIModel;
-        const validModels = ['gpt-4o', 'gpt-4o-mini', 'gpt-4-turbo', 'o1-preview', 'o1-mini'];
-        if (!validModels.includes(model)) {
-          printError(`Invalid model: ${model}. Use one of: ${validModels.join(', ')}`);
+        if (!model || model.trim().length === 0) {
+          printError('Model name must not be empty.');
           process.exit(1);
         }
 
@@ -173,15 +173,21 @@ export function createCreateCommand(): Command {
           succeedSpinner(`Created ${scaffoldResult.filesCreated.length} files`);
         }
 
-        // Run the workflow
+        // Run the workflow — merge CLI flags with popeye.md config
         printSection('Workflow Execution');
 
+        const popeyeConfig = await readPopeyeMdConfig(projectDir);
         const workflowResult = await runWorkflow(spec, {
           projectDir,
           consensusConfig: {
             threshold,
             maxIterations,
             openaiModel: model,
+            reviewer: popeyeConfig?.reviewer ?? 'openai',
+            arbitrator: popeyeConfig?.arbitrator,
+            enableArbitration: popeyeConfig?.enableArbitration ?? false,
+            geminiModel: popeyeConfig?.geminiModel,
+            grokModel: popeyeConfig?.grokModel,
           },
           onProgress: (phase, message) => {
             handleProgressUpdate(phase, message);

package/src/cli/commands/resume.ts

@@ -12,6 +12,7 @@ import {
   resetWorkflow,
   cancelWorkflow,
 } from '../../workflow/index.js';
+import { readPopeyeMdConfig } from '../../config/popeye-md.js';
 import type { WorkflowPhase } from '../../types/workflow.js';
 import {
   printHeader,
@@ -94,13 +95,20 @@ export function createResumeCommand(): Command {
           process.exit(1);
         }
 
-        // Resume workflow
+        // Resume workflow — merge CLI flags with popeye.md config
         printSection('Resuming Workflow');
 
+        const popeyeConfig = await readPopeyeMdConfig(projectDir);
         const result = await resumeWorkflow(projectDir, {
           consensusConfig: {
             threshold,
             maxIterations,
+            reviewer: popeyeConfig?.reviewer ?? 'openai',
+            arbitrator: popeyeConfig?.arbitrator,
+            enableArbitration: popeyeConfig?.enableArbitration ?? false,
+            openaiModel: popeyeConfig?.openaiModel,
+            geminiModel: popeyeConfig?.geminiModel,
+            grokModel: popeyeConfig?.grokModel,
           },
           maxRetries,
           onProgress: (phase, message) => {

package/src/cli/interactive.ts

@@ -46,7 +46,7 @@ import { upgradeProject } from '../upgrade/index.js';
 import { buildUpgradeContext } from '../upgrade/context.js';
 import { OutputLanguageSchema, KNOWN_OPENAI_MODELS } from '../types/project.js';
 import type { ProjectSpec, OutputLanguage, OpenAIModel } from '../types/project.js';
-import { GeminiModelSchema, KNOWN_GEMINI_MODELS } from '../types/consensus.js';
+import { GeminiModelSchema, KNOWN_GEMINI_MODELS, KNOWN_GROK_MODELS } from '../types/consensus.js';
 import { OpenAIModelSchema } from '../types/project.js';
 import type { AIProvider, GeminiModel, GrokModel } from '../types/consensus.js';
 import {
@@ -1205,6 +1205,35 @@ async function handleReviewSlashCommand(state: SessionState, args: string[] = []
   }
 
   try {
+    // Check if project is pipeline-managed — route through bridge
+    const { loadProject } = await import('../state/index.js');
+    const projectState = await loadProject(state.projectDir);
+    const { isPipelineManaged, runReviewBridge } = await import('../pipeline/bridges/review-bridge.js');
+
+    if (isPipelineManaged(projectState)) {
+      printInfo('Pipeline-managed project detected — routing through pipeline bridge');
+      console.log();
+
+      const bridgeResult = await runReviewBridge({
+        projectDir: state.projectDir,
+        depth: options.depth,
+        strict: options.strict,
+        onProgress: (_stage, msg) => printInfo(msg),
+      });
+
+      if (bridgeResult.success) {
+        printSuccess(`Audit score: ${bridgeResult.overallScore}% — ${bridgeResult.recommendation}`);
+        printInfo(`${bridgeResult.findingsCount} finding(s), ${bridgeResult.changeRequestCount} CR(s) created, ${bridgeResult.artifactsCreated} artifact(s) stored`);
+        if (bridgeResult.changeRequestCount > 0) {
+          printInfo('Change Requests filed — run /resume to let the pipeline process them');
+        }
+      } else {
+        printError(bridgeResult.error ?? 'Review bridge failed');
+      }
+      return;
+    }
+
+    // Non-pipeline project — use legacy audit-mode
     const { runReview } = await import('./commands/review.js');
 
     console.log();
@@ -1484,7 +1513,7 @@ function handleLanguage(args: string[], state: SessionState): void {
 const KNOWN_MODELS: Record<string, readonly string[]> = {
   openai: KNOWN_OPENAI_MODELS,
   gemini: KNOWN_GEMINI_MODELS,
-  grok: ['grok-3', 'grok-3-mini', 'grok-2'],
+  grok: KNOWN_GROK_MODELS,
 };
 
 /**
@@ -2976,7 +3005,7 @@ export async function startInteractiveMode(): Promise<void> {
     projectDir: process.cwd(),
     language: config.project.default_language,
     openaiModel: config.apis.openai.model,
-    geminiModel: 'gemini-2.0-flash',
+    geminiModel: 'gemini-2.5-flash',
     grokModel: config.apis.grok.model,
     claudeAuth: false,
     openaiAuth: false,

package/src/config/defaults.ts

@@ -19,10 +19,15 @@ export const DEFAULT_CONFIG: Config = {
   },
   apis: {
     openai: {
-      model: 'gpt-4o',
+      model: 'gpt-4.1',
       temperature: 0.3,
       max_tokens: 4096,
-      available_models: ['gpt-4o', 'gpt-4o-mini', 'gpt-4-turbo', 'o1-preview', 'o1-mini'],
+      available_models: [
+        'gpt-4.1', 'gpt-4.1-mini', 'gpt-4.1-nano',
+        'o3', 'o3-mini', 'o4-mini',
+        'gpt-4o', 'gpt-4o-mini',
+        'gpt-4-turbo', 'o1-preview', 'o1-mini',
+      ],
     },
     claude: {
       model: 'claude-sonnet-4-20250514',

package/src/config/popeye-md.ts

@@ -0,0 +1,139 @@
+/**
+ * Shared popeye.md configuration reader.
+ * Reads project-local config from the YAML frontmatter in popeye.md.
+ * Used by both interactive mode and CLI commands.
+ */
+
+import { promises as fs } from 'node:fs';
+import path from 'node:path';
+import { OutputLanguageSchema, type OutputLanguage } from '../types/project.js';
+import type { AIProvider, GeminiModel, GrokModel } from '../types/consensus.js';
+import type { OpenAIModel } from '../types/project.js';
+
+// ─── Types ───────────────────────────────────────────────
+
+/** Project-local configuration stored in popeye.md */
+export interface PopeyeMdConfig {
+  language: OutputLanguage;
+  reviewer: AIProvider;
+  arbitrator: AIProvider;
+  enableArbitration: boolean;
+  created: string;
+  lastRun: string;
+  projectName?: string;
+  description?: string;
+  notes?: string;
+  openaiModel?: OpenAIModel;
+  geminiModel?: GeminiModel;
+  grokModel?: GrokModel;
+}
+
+// ─── Reader ──────────────────────────────────────────────
+
+/**
+ * Read popeye.md from a project directory.
+ * Parses YAML frontmatter for project configuration.
+ *
+ * @param projectDir - Absolute path to the project root
+ * @returns Parsed config or null if file doesn't exist or is invalid
+ */
+export async function readPopeyeMdConfig(projectDir: string): Promise<PopeyeMdConfig | null> {
+  const configPath = path.join(projectDir, 'popeye.md');
+
+  try {
+    const content = await fs.readFile(configPath, 'utf-8');
+
+    // Parse YAML frontmatter
+    const frontmatterMatch = content.match(/^---\n([\s\S]*?)\n---/);
+    if (!frontmatterMatch) {
+      return null;
+    }
+
+    const frontmatter = frontmatterMatch[1];
+    const config: Partial<PopeyeMdConfig> = {};
+
+    // Parse each line of YAML
+    for (const line of frontmatter.split('\n')) {
+      const match = line.match(/^(\w+):\s*(.+)$/);
+      if (match) {
+        const [, key, value] = match;
+        const cleanValue = value.trim();
+
+        switch (key) {
+          case 'language':
+            if (OutputLanguageSchema.safeParse(cleanValue).success) {
+              config.language = cleanValue as OutputLanguage;
+            }
+            break;
+          case 'reviewer':
+            if (['openai', 'gemini', 'grok'].includes(cleanValue)) {
+              config.reviewer = cleanValue as AIProvider;
+            }
+            break;
+          case 'arbitrator':
+            if (['openai', 'gemini', 'grok', 'off'].includes(cleanValue)) {
+              if (cleanValue === 'off') {
+                config.enableArbitration = false;
+              } else {
+                config.arbitrator = cleanValue as AIProvider;
+                config.enableArbitration = true;
+              }
+            }
+            break;
+          case 'created':
+            config.created = cleanValue;
+            break;
+          case 'lastRun':
+            config.lastRun = cleanValue;
+            break;
+          case 'projectName':
+            config.projectName = cleanValue;
+            break;
+          case 'openaiModel':
+            if (cleanValue.length > 0) {
+              config.openaiModel = cleanValue;
+            }
+            break;
+          case 'geminiModel':
+            if (cleanValue.length > 0) {
+              config.geminiModel = cleanValue;
+            }
+            break;
+          case 'grokModel':
+            if (cleanValue.length > 0) {
+              config.grokModel = cleanValue;
+            }
+            break;
+        }
+      }
+    }
+
+    // Extract notes section if present
+    const notesMatch = content.match(/## Notes\n([\s\S]*?)(?=\n## |$)/);
+    if (notesMatch) {
+      config.notes = notesMatch[1].trim();
+    }
+
+    // Return config only if we have the essential fields
+    if (config.language && config.reviewer) {
+      return {
+        language: config.language,
+        reviewer: config.reviewer,
+        arbitrator: config.arbitrator || 'gemini',
+        enableArbitration: config.enableArbitration ?? true,
+        created: config.created || new Date().toISOString(),
+        lastRun: config.lastRun || new Date().toISOString(),
+        projectName: config.projectName,
+        description: config.description,
+        notes: config.notes,
+        openaiModel: config.openaiModel,
+        geminiModel: config.geminiModel,
+        grokModel: config.grokModel,
+      };
+    }
+
+    return null;
+  } catch {
+    return null;
+  }
+}

package/src/config/schema.ts

@@ -23,14 +23,17 @@ export const ConsensusSettingsSchema = z.object({
  * OpenAI API settings schema
  */
 export const OpenAISettingsSchema = z.object({
-  model: z
-    .enum(['gpt-4o', 'gpt-4o-mini', 'gpt-4-turbo', 'o1-preview', 'o1-mini'])
-    .default('gpt-4o'),
+  model: z.string().min(1).default('gpt-4.1'),
   temperature: z.number().min(0).max(2).default(0.3),
   max_tokens: z.number().min(100).max(32000).default(4096),
   available_models: z
     .array(z.string())
-    .default(['gpt-4o', 'gpt-4o-mini', 'gpt-4-turbo', 'o1-preview', 'o1-mini']),
+    .default([
+      'gpt-4.1', 'gpt-4.1-mini', 'gpt-4.1-nano',
+      'o3', 'o3-mini', 'o4-mini',
+      'gpt-4o', 'gpt-4o-mini',
+      'gpt-4-turbo', 'o1-preview', 'o1-mini',
+    ]),
 });
 
 /**
@@ -55,10 +58,15 @@ export const GrokSettingsSchema = z.object({
  */
 export const APISettingsSchema = z.object({
   openai: OpenAISettingsSchema.default({
-    model: 'gpt-4o',
+    model: 'gpt-4.1',
     temperature: 0.3,
     max_tokens: 4096,
-    available_models: ['gpt-4o', 'gpt-4o-mini', 'gpt-4-turbo', 'o1-preview', 'o1-mini'],
+    available_models: [
+      'gpt-4.1', 'gpt-4.1-mini', 'gpt-4.1-nano',
+      'o3', 'o3-mini', 'o4-mini',
+      'gpt-4o', 'gpt-4o-mini',
+      'gpt-4-turbo', 'o1-preview', 'o1-mini',
+    ],
   }),
   claude: ClaudeSettingsSchema.default({
     model: 'claude-sonnet-4-20250514',
@@ -139,10 +147,15 @@ export const ConfigSchema = z.object({
   }),
   apis: APISettingsSchema.default({
     openai: {
-      model: 'gpt-4o',
+      model: 'gpt-4.1',
       temperature: 0.3,
       max_tokens: 4096,
-      available_models: ['gpt-4o', 'gpt-4o-mini', 'gpt-4-turbo', 'o1-preview', 'o1-mini'],
+      available_models: [
+        'gpt-4.1', 'gpt-4.1-mini', 'gpt-4.1-nano',
+        'o3', 'o3-mini', 'o4-mini',
+        'gpt-4o', 'gpt-4o-mini',
+        'gpt-4-turbo', 'o1-preview', 'o1-mini',
+      ],
     },
     claude: {
       model: 'claude-sonnet-4-20250514',