0.3.1 — Streaming Stability and Memory Persistence

0.3.1 is a patch release focused on streaming stability, memory persistence, proxy-aware tool scanning, and usage accounting for reasoning tokens.

Features

Reasoning Token Usage

thinkingTokens: Int? was added to both:

• ModelResponse
• UsageMetrics

OpenAI reasoning-token usage is parsed from completion_tokens_details.reasoning_tokens and wired through the engine. Azure OpenAI receives the same usage accounting path where the provider response includes the field.

Bug Fixes

Chat Memory Persistence

Conversation turns are now persisted correctly after:

• streaming completion
• structured output success

This closes a persistence gap where successful AI calls could update the caller-visible result without reliably recording the completed turn in configured chat memory.

Proxy-Aware Tool Scanning

Tool discovery now handles proxied Spring/MCP tool handlers more reliably. This matters when application infrastructure wraps tool beans and the scanner needs to inspect the user-facing tool contract rather than only the proxy class.

Refactoring

Provider streaming implementations were decomposed into smaller HTTP error handling and response parsing helpers.

Affected providers:

• OpenAI
• Anthropic
• Azure OpenAI

This keeps the streaming behavior easier to audit and reduces provider-specific parsing complexity.

Tests

Additional coverage was added around:

• proxy-aware tool scanning
• MCP tool handlers
• TokenAwareChatMemory
• streaming and memory persistence behavior

Notes

• TramAI 0.3.1 targets Java 21+.
• This release preserves the 0.3.0 public module surface.
• Consumer dependency coordinates should use dev.tramai:<module>:0.3.1 or the tramai-bom:0.3.1.