Known Limitations

Apex AI Brain is a powerful tool, but it has some known constraints and out-of-scope features. Understanding these limitations will help you use the system effectively.

Model-Specific Limitations

Claude

• Context Window: 200k tokens max (sufficient for most Salesforce projects)
• Best For: Complex analysis, long-form code generation, detailed planning
• Limitations: May be slower on very large code reviews

Gemini

• Speed: Faster response times, good for rapid iteration
• Cost: Lower token cost than Claude
• Limitations: May hallucinate on Salesforce-specific patterns more frequently than Claude

GPT-4 (Experimental)

• Status: Not recommended for production use
• Reason: Lacks Salesforce-specific guardrails compared to Claude and Gemini

Salesforce Feature Limitations

The following Salesforce features have limited or no support:

Feature	Status	Reason
Flows (Low-Code)	Partial	Can design but not generate working Flow XML
Metadata API	Read-only	Cannot deploy metadata directly (requires manual validation)
Streaming API	Not supported	Real-time event processing out of scope
Platform Events	Partial	Can design but limited deployment automation
Einstein AI	Not integrated	Requires separate Einstein API keys
FinanceCloud-specific	Limited	Not all industry cloud features documented
Gack Errors	Limited	Salesforce platform errors require human interpretation

Codebase Limitations

Project Size

• Max Codebase Size: 500+ files (scanner may timeout on massive orgs)
• Recommendation: Use /init --shallow for large projects or split into modules

Token Limits

• Max tokens per command: 200,000 (set by Claude's context window)
• Recommendation: Break large requests into smaller tasks using /plan

Agent Coordination

• Max concurrent agents: 5 (to avoid rate limiting)
• Recommendation: Use /spawn --sequential for large batches

Operational Limitations

No Real-Time Feedback

• Agents operate in "request-response" mode
• Cannot monitor ongoing deployments or tests in real-time
• Recommendation: Use --wait flag for explicit polling

No Persistent Session Memory Across Restarts

• Memory is stored locally per session
• Restarting the CLI clears in-memory context
• Workaround: Use /memory snapshot before shutdown

Internet Required

• All AI model APIs require internet connectivity
• Offline mode not supported
• Recommendation: Handle connection failures gracefully in scripts

Accuracy & Reliability

Code Generation Success Rates

Task	Success Rate	Notes
Apex CRUD Operations	95%+	Basic patterns highly reliable
LWC Component Structure	90%+	Good for standard components
SOQL Optimization	85%+	May miss edge cases
Trigger Logic	80%+	Complex business logic needs review
Test Coverage	75%+	Generated tests need augmentation
Flow Design	70%+	Recommend manual refinement

Why These Rates?

1. Salesforce API Complexity - Governor limits and edge cases are hard to predict
2. Org-Specific Variations - Custom objects, fields, and workflows vary significantly
3. Business Logic - AI cannot know your specific business requirements
4. Integration Patterns - External system integrations require domain knowledge

What You Must Do Manually

Apex AI Brain cannot automate these tasks:

1. Security Review - All generated code should go through security review
2. Performance Testing - Load testing and stress testing require manual validation
3. Data Migration - Moving data between environments requires careful planning
4. User Acceptance Testing - Business stakeholders must validate changes
5. Governance & Compliance - Industry-specific compliance (HIPAA, PCI, etc.) requires expert review
6. Deployment Coordination - Senior developers should manage production deployments

Rate Limiting

API Rate Limits

• Claude: 50 requests/min (standard tier)
• Gemini: 60 requests/min (standard tier)
• OpenAI GPT-4: 20 requests/min (default)

Workarounds

# Stagger requests manually
sfai /plan --agents 2  # Use fewer agents simultaneously

# Use caching
sfai /context cache --enable

# Batch operations
sfai /spawn --sequential --agents 5

Browser & UI Limitations

• No Web Dashboard - CLI-first tool only
• No Visual Flow Builder - Flows must be designed via YAML or JSON
• Limited Terminal Output - Large responses may be truncated (use --output file.md)
• No IDE Integration - Works with terminal only (VS Code extension planned)

Deployment Limitations

• No Rollback Automation - Must manually rollback failed deployments
• No Zero-Downtime Deploys - Blue-green deployments not automated
• Limited Approval Workflows - Cannot enforce multi-level approvals
• No Change Set Generation - Uses direct API deployments only

Support & Workarounds

For each limitation, we provide workarounds:

Large Projects? Use --shallow or split into modules Need Persistence? Use /memory snapshot and restore manually Offline Work? Generate code first, then work offline Complex Logic? Use /plan to break into smaller tasks Need Guarantees? Use --validation strict for safety checks

Future Improvements

We're actively working on:

• Visual Flow builder integration
• Better metadata deployment support
• Persistent cross-session memory
• VS Code extension
• Real-time deployment monitoring
• Industry cloud support (FinanceCloud, HealthCloud, etc.)

Report Limitations

If you find a limitation not listed here, please:

1. Check existing GitHub Issues
2. Create a new issue with: Limitation + Use Case + Workaround
3. Email the maintainer or check SUPPORT.md

---

Last Updated: March 19, 2026 Version: 1.0.0