v0.5.0

Minor release.

Highlights

• Serving helpers (P23) — agentix.serving turns an Agent into a streaming HTTP endpoint: dependency-free SSE/NDJSON serializers (event_to_dict, sse_events, ndjson_events, outcome_to_payload) plus a lazy FastAPI/Starlette adapter (sse_response / ndjson_response, install agentix-toolkit[serving]). Pairs with suspend/resume for web human-in-the-loop. See examples/30_serving_fastapi.py and the Serving guide.
• Friendlier wording — plainer, beginner-first package summary and README intro (no API changes).

Install: pip install "agentix-toolkit[anthropic]" (import as agentix).

Full notes: CHANGELOG.md.

v0.4.1

Patch release — metadata only, no functional code change. Re-publishes so the
work merged after 0.4.0 shows up on PyPI:

• PyPI project links — the project page now shows Documentation (the docs
  site), Changelog, and Issues (0.4.0 only had Homepage/Repository).
• Richer keywords & classifiers (AI, AsyncIO, Typed, …) for package search.
• The documentation site (https://skwijeratne.github.io/agentix-toolkit/) and
  an llms.txt index so AI coding assistants can read the docs cleanly.

Install: pip install "agentix-toolkit[anthropic]" (import as agentix).

v0.4.0

Minor release — all changes since 0.3.0 are additive and backward compatible.

Highlights

• Token-accurate context (P19) — TokenCounter + count_tokens + FitContextWindow trim the transcript to a real token budget (pluggable tokenizer; tiktoken/provider counters drop in).
• Pluggable memory (P20) — a Memory protocol (recall/write) + dependency-free InMemoryMemory; Agent(memory=…, remember_exchange=…) recalls cross-session context (bring your own vector DB behind the interface).
• First-class structured output (P21) — Agent(response_model=…) (Pydantic class or JSON-Schema dict) wires validation + retry, a schema prompt, and native provider enforcement (with_response_format) in one knob; outcome.parsed is typed.
• Rate-limit-aware retries (P22) — RetryModel honors a provider's Retry-After (default_retry_after, max_sleep, on_retry) instead of blind exponential backoff.
• Tier C polish — eval JSONL/CSV loaders (load_cases), record/replay cassettes (CassetteModel) for deterministic real-model tests, subagent cost roll-up (child spend → parent totals), and one-call tracing (instrument(agent)).

Install: pip install "agentix-toolkit[anthropic]" (import as agentix).

Full notes: see CHANGELOG.md.

v0.3.0

Minor release — all changes since 0.2.1 are additive and backward compatible.

Highlights

• Multi-provider adapters (P14) — OpenAI (+ any OpenAI-compatible URL), Gemini, Bedrock (Converse), Ollama (local), and a LiteLLM bridge. Each behind its own extra; all drop-in ModelFns.
• Multimodal content (P15) — Message.content is now str | list[ContentPart] (TextPart/ImagePart/DocumentPart/AudioPart) with a Message.text view and per-adapter translation. Plain-string content stays fully backward compatible.
• Sandboxed executor (P16) — SubprocessExecutor runs untrusted / model-generated code in a separate process: denies network when network_allowlist is empty (Linux netns, fail-closed), plus CPU/memory/file-size/process rlimits, an isolated temp cwd, a scrubbed env, and a hard timeout.
• Suspendable human-in-the-loop (P18) — suspend_on_confirm pauses at a confirmation, persists, and resume(run_id, decisions=…) continues (web/serverless-friendly; resumes even in a new process).

Install: pip install "agentix-toolkit[anthropic]" (import as agentix).

Full notes: see CHANGELOG.md.

agentix-toolkit 0.1.0

First public release. pip install agentix-toolkit / uv add agentix-toolkit (import as agentix).

Includes the async agent loop, @tool decorator, opt-in security guard pipeline, Anthropic adapter, streaming, persistence/resume, MCP client support, context management, subagents, USD cost tracking, and interrupts. See CHANGELOG.md.

PyPI: https://pypi.org/project/agentix-toolkit/0.1.0/