Changelog¶

All notable changes to this project will be documented in this file.

The format follows Keep a Changelog and this project adheres to Semantic Versioning.

[0.2.2] - 2026-08-05¶

anyask --version/anyask version now print anyask v<version> (e.g. anyask v0.2.2) instead of anyask <version>, matching this repo's vX.Y.Z git tag convention.

Simplified README.md: restructured around description / features / installation / supported providers / what is anyask / why do you need anyask, trimming the detailed API/Errors/CLI reference sections down to pointers at the hosted docs (docs/reference, docs/cli, docs/providers already cover that ground) instead of duplicating it.

README.md's SPEC.md links used repo-relative paths, which break on PyPI (README.md also serves as the PyPI project description, and PyPI has no repo tree to resolve a relative link against — it just 404s). Switched to absolute GitHub blob URLs, matching the pattern already used for the same file in docs/index.md.

SPEC.md — architecture, the full public API contract, and seven Mermaid diagrams (component overview, a single ask() call as a sequence diagram, lazy per-provider imports, the Provider/AskResponse/error class hierarchy, the construction-kwargs split, the reasoning=True per-provider decision flow, and the release pipeline), plus an explicit out-of-scope list.
"What anyask does (and doesn't) do" section in README.md/docs/index.md: documents that prompt/AskResponse.content are always plain strings, so there's no multi-turn conversation state, no normalized system-prompt support, no multi-modal input/output, and no streaming — spelled out explicitly rather than left implicit.

pyproject.toml's description now says "single text prompt" / "raw text response" instead of just "raw response", matching the capabilities clarification above.
Fixed the provider count from 17 to 18 everywhere (pyproject.toml, README.md, SPEC.md, CHANGELOG.md, and three docs pages) — the actual registry has 18 named vendor providers plus the internal openai_compat base, which was never meant to be counted; "17" was wrong from the first release and had just been copied forward since.

scripts/verify-version.ps1 — unreferenced anywhere in the repo, and unlike this repo's other .ps1 scripts (which exist because their bash counterparts need a native-PowerShell equivalent), verify-version.py already runs identically on Windows with no bash dependency to work around. Also a strictly weaker check (missing .bumpversion.cfg).
scripts/verify-version.py and its two references (the local pre-commit hook, the Makefile's verify-version target). Version-drift detection now relies solely on tests/unit/test_version_consistency.py, which already runs automatically in CI via tox on every push.
.pre-commit-config.yaml — not installed (missing from the dev extra), not documented, not invoked by CI. Everything it checked (ruff, ruff-format, black, mypy) is the same tooling at the same pinned versions already enforced via tox -e lint/format/typecheck.

Replaced the Makefile's tomllib/tomli-fallback version extraction (which assumed tomli was installed on Python <3.11, but never declared it as a dependency anywhere) and scripts/release.ps1's equivalent (which had no fallback at all) with a plain regex against the raw file text — no stdlib-version gate, no extra dependency, and now also supports single-quoted version = '0.1.1' in addition to double-quoted.

anyask console script: anyask --version/anyask version, anyask check <provider> (construct one provider without a network call, reporting whether its SDK is installed and credentials resolve; exits 0/1), and anyask ready (sweep every provider using only environment-variable credentials; always exits 0 - a status report, not a gate). Built on stdlib argparse, not click, to keep the bare install at 2 packages.
Consistent temperature=0.2/max_tokens=2048 defaults across all 18 providers (previously only Anthropic, Mistral, Cohere, Bedrock, and OCI set these; OpenAI, the 9 OpenAI-compatible vendors, Gemini, Vertex AI, and Azure passed calls through with no default, meaning identical ask() calls could get very different sampling behavior purely depending on provider). Skipped for OpenAI/OpenAI-compatible vendors when reasoning=True, since reasoning models on some of them reject a temperature override and use max_completion_tokens instead of max_tokens.
reasoning: bool = False on ask()/ask_async() — enables extended/ deliberate reasoning using each provider's own real mechanism (Anthropic extended thinking, OpenAI/OpenAI-compatible reasoning_effort, Gemini/Vertex AI thinking budgets, Bedrock's Claude thinking field). Providers with no such mechanism (Azure, Mistral, Cohere, OCI) raise ProviderNotFoundError rather than silently ignoring it. See docs/providers/index.md's Reasoning support table.

Initial extraction release: ask(), ask_async(), list_models(), list_models_async(), and get_provider() — one normalized AskResponse shape across 18 providers (OpenAI, Anthropic, Gemini, Vertex AI, Azure, Mistral, Bedrock, Cohere, OCI, and nine OpenAI-compatible vendors: DeepSeek, Groq, xAI, Together AI, Fireworks AI, Cerebras, Perplexity, OpenRouter, Moonshot)
Lazy per-provider SDK imports (anyask/registry.py) — import anyask never requires any provider SDK to be installed; each provider's extra is only needed once that provider is actually used
AskLLMError/ProviderNotFoundError/ProviderError/ ProviderAuthError error hierarchy — every provider call failure is raised with from exc, preserving the original SDK exception as __cause__
No routing, fallback, or retry logic — provider and model are always explicit, required arguments