Self-Healing¶

Pisama's self-healing pipeline automatically generates, validates, and applies fixes when failure modes are detected in agent workflows. The system is designed with safety-first principles: every fix requires a checkpoint, every applied fix can be rolled back.

Pipeline Overview¶

Detection Result
       │
       ▼
┌──────────────────┐
│  1. Analyze      │  Identify failure root cause
│                  │  Determine fix category
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│  2. Generate Fix │  AI-powered fix suggestion
│                  │  Code-level or config-level
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│  3. Approval     │  Manual or automatic
│     Policy       │  Based on risk level
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│  4. Apply &      │  Execute with checkpoint
│     Validate     │  Run validation checks
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│  5. Rollback     │  If validation fails,
│     (if needed)  │  restore from checkpoint
└──────────────────┘

Fix Categories¶

Basic Fix Suggestions (All Plans)¶

Text-based suggestions that describe what to change:

• Loop breaking: "Consider adding a maximum iteration count or convergence check"
• Injection defense: "Add input sanitization before passing to LLM"
• Overflow prevention: "Implement conversation summarization when context exceeds 70%"

Code-Level Fixes (Startup+)¶

Specific code changes with copy-paste solutions:

# Example: Loop detection fix suggestion
# Add to your agent's step function:
if iteration_count > MAX_ITERATIONS:
    return {"status": "terminated", "reason": "max_iterations_exceeded"}

AI-Generated Runbooks (Growth+)¶

Operational documentation generated from detection patterns:

• Step-by-step remediation procedures
• Monitoring queries to verify the fix
• Prevention guidelines for the team

Self-Healing Playbooks (Growth+)¶

Pre-configured automated fix sequences:

AI-Generated Fixes (Enterprise)¶

Full automated fix generation using Claude:

• Analyzes the trace, detection, and codebase context
• Generates specific code patches
• Includes test cases for the fix
• Provides rollback instructions

Approval Policies¶

Fixes are categorized by risk level, and each level has a different approval requirement:

API Endpoints¶

List healing operations¶

curl "http://localhost:8000/api/v1/tenants/$TENANT_ID/healing/operations" \
  -H "Authorization: Bearer $TOKEN"

Get operation details¶

curl "http://localhost:8000/api/v1/tenants/$TENANT_ID/healing/operations/$OPERATION_ID" \
  -H "Authorization: Bearer $TOKEN"

Approve an operation¶

curl -X POST "http://localhost:8000/api/v1/tenants/$TENANT_ID/healing/operations/$OPERATION_ID/approve" \
  -H "Authorization: Bearer $TOKEN"

Execute a healing operation¶

curl -X POST "http://localhost:8000/api/v1/tenants/$TENANT_ID/healing/operations/$OPERATION_ID/execute" \
  -H "Authorization: Bearer $TOKEN"

Rollback a fix¶

curl -X POST "http://localhost:8000/api/v1/tenants/$TENANT_ID/healing/operations/$OPERATION_ID/rollback" \
  -H "Authorization: Bearer $TOKEN"

View healing history¶

curl "http://localhost:8000/api/v1/tenants/$TENANT_ID/healing/history" \
  -H "Authorization: Bearer $TOKEN"

Safety Guarantees¶

1. Checkpoint before apply: Every fix creates a state checkpoint before modifying anything
2. Rollback capability: Any applied fix can be rolled back to its checkpoint
3. Validation after apply: Fixes are validated immediately after application -- if validation fails, automatic rollback is triggered
4. Audit trail: Every healing operation is logged with timestamps, approvers, and outcomes
5. Canary deployment: Enterprise tier supports canary-style fix application -- apply to a subset, verify, then roll out

Availability by Plan¶