Pattern Proposal: Provider Abstraction for Deterministic Agent Execution

Summary

LLM access is isolated behind a client interface so agent steps can run against real or mock providers.

Candidate Pattern

• ID: llm-provider-abstraction
• Classification: intelligence
• Maturity: established
• Type: implemented

Observed Signals

No specific signals listed.

Evidence

Supporting Evidence

• src/cas_runner/llm/litellm_client.py (LiteLLMClient) - Real provider adapter
• src/cas_runner/llm/mock_client.py (MockLLMClient) - Test double

Scope Assessment

• Evidence Strength: high
• Confidence: high
• Applies to: ideas-to-life-cas

Rationale

• Intent: Decouple agent execution from LLM backend to enable testing and portability.
• Context: Used by the CAS runner for all LLM-backed agent steps.
• Force: Need to test agent flows without external LLM costs
• Force: Support provider switching to avoid vendor lock-in
• Force: Keep agent code independent of SDK specifics

Traceability

Artifact source: proposed-patterns