Defines a ClientAdapter trait for pluggable client-specific behavior (e.g., OpencodeAdapter). Each adapter implements signature matching strategies (Fifo/Baseline), beta header injection, and thinking-budget fetching. The pattern enables new CLI tools to be supported by implementing a single trait without modifying core proxy logic.

Provides a CanonicalUsage structure that unifies usage metadata from Gemini, Anthropic, and OpenAI protocols. Handles cached/uncached/reasoning token aggregation and scaling (e.g., Claude token scaling). This component cleanly abstracts away provider-specific counting semantics and can be reused in any multi-provider proxy system.

Implements core context management: purification of tool call/result pairs, multi-language-aware token estimation (ASCII ~4 chars/token, CJK ~1.5 chars/token with 15% safety margin), and tool message trimming for aggressive context window bundling. A reusable algorithm for any LLM proxy needing efficient context usage.

Provides AES-256-GCM encryption for sensitive credentials (OAuth tokens, API keys) with keys derived from machine hardware IDs, supporting both legacy and v2 formats. Includes atomic file I/O for safe persistence. Well-isolated and could be extracted into a standalone credential vault helper.