Release v0.1.9

AceAI v0.1.9

Features

• Tool tags for grouping tools
• Max tool usage per run
• Tool registry and dynamic tool descriptions

Release v0.1.8

AceAI v0.1.8

Highlights

• Agent instruction handling now deduplicates added instructions while keeping system_message in sync.

Tooling

• scripts/release.py adds branch-based version parsing with validation and keeps --version optional; make release conditionally forwards the flag and avoids fake defaults.
• Added regression coverage for release version inference from branch names.

Breaking

• aceai.agent.BufferedStreamingAgent is gone; importers must use AgentBase.
• AgentBase.instructions attribute was removed; consumers should rely on add_instruction + system_message.

Release v0.1.7

AceAI v0.1.7

Highlights

• Tool description precedence now prefers tool(description=...) over func.__doc__ (also reflected in tool_schema["description"]).
• AgentBase prompt handling refactor: prompt seeds an instructions list, with system_message derived from accumulated instructions.
• Tracing overhaul: agent.run → agent.step → llm.* / tool.* spans are now parented via explicit trace_ctx, avoiding async-generator context detach errors and root-span fragmentation.
• OpenAI Responses support now surfaces structured output segments (reasoning/tool/media) and richer LLMStreamEvent mapping (tool deltas + image generation).

Tracing / Observability

• trace_ctx is now threaded through AgentBase, LLMService, provider adapters, and ToolExecutor so downstream spans consistently inherit the same trace.
• Adds Langfuse-friendly span attributes: langfuse.trace.name, langfuse.trace.input, langfuse.trace.output.

Docs / DX

• Added regression coverage for trace parenting and async-generator cancellation edge cases.
• Refactored agent tests for the AgentBase API changes and expanded tool description/schema coverage.

Notes

• Breaking: AgentBase.__init__ renamed sys_prompt -> prompt; max-step runtime error message changed to “Agent exceeded maximum steps: N without answering”.
• Breaking: provider implementers and tool executors must accept/pass through trace_ctx to keep spans correctly chained.
• Tracing: ToolExecutor span attribute changed from tool.arguments.raw_len to tool.arguments (larger attribute payloads).

Release v0.1.6

AceAI v0.1.6

Highlights

• Introduced ILLMService (Protocol) and updated AgentBase to depend on the interface instead of a concrete LLMService, improving testability and injection flexibility.
• Added first-class speech-to-text support via LLMService.transcribe(...) (with tracing + timeouts) and extended the provider contract to accept prompt.
• Made AgentBase tool execution optional (executor: ToolExecutor | None): runs without tool wiring when you don’t need tools, while still failing fast if tool calls appear without an executor.

Provider / Tools

• OpenAI STT now forwards prompt to audio.transcriptions.create(...) for better transcription guidance.
• Updated tool signature DI wiring to use ididi.NodeMeta (instead of the previous marker-based approach), tightening dependency resolution semantics.

Notes

• Potentially breaking: provider stt(...) signature changed (adds prompt), LLMService now exposes transcribe, and AgentBase constructor typing expects ILLMService + optional executor.