v0.1.3 — Phase 1.3: Stabilize the MVP

donkey-kit 0.1.3: Phase 1.3, Stabilize the MVP. This release makes the docs and the release process as dependable as the code. Each class of fact in the docs now has one source. Release gates are automated, the support matrix is tested, and new framework releases are watched. The CI supply chain is hardened, and typed refusals now also work through ADK's model() path. It adds no Phase 2 scope.

pip install --upgrade donkey-kit==0.1.3

Breaking changes

None.

Highlights

• Typed refusals through ADK model(): a governance refusal on the LiteLLM path of model() now reaches your code as the typed DonkeyError subclass, as it already did on the other adapters (#1077).
• Control-plane region hosts fixed: the CA and JP Hyperforce hosts pointed at ca1.anypoint. and jp1.anypoint., which don't exist. They now point at ca1.platform.mulesoft.com and jp1.platform.mulesoft.com, and every region is verified (#1113).
• Docs with one source per class of fact:
  • The API reference and config tables are generated (#1080).
  • Per-framework verification status comes from the verification ledger (#1087).
  • A generated page lists the framework versions each release supports (#1095).
  • Doc snippets are tested (#1079), and the site is built, type-checked and link-checked on every PR (#1076).
  • There is a release-notes page (#1080) and baseline SEO and share metadata (#1078).
• Release gates:
  • One version source, __version__ (#1075).
  • A tag-to-version check, a clean-install wheel smoke test, and PEP 740 attestations (#1075).
  • A locked release toolchain (#1071).
  • Each TestPyPI publish now dispatches the acceptance suite against the published package and reports the result on the promotion PR (#1086, #1123).
• Tested support matrix: CI tests the declared floors, openai 1.x, and Python 3.13 and 3.14 (#1072).
• Framework release watch: each new release of a supported framework is tested on its own against the PR lock (#1100), with an AI review of its changelog's impact (#1112).
• Supply chain: Actions are pinned by SHA. Dependabot, CODEOWNERS and dependency audits are in place (#1070).
• Contributor experience:
  • One nox task runner that mirrors CI, plus pre-commit hooks (#1074).
  • ruff format is applied across python/ (#1073).
  • bump-version.sh links the milestone to the bump issue and PR (#1115).

Extras changes (§8.4)

• openai-agents: the openai floor moves from >=1.66 to >=2.45. openai-agents>=0.20 already requires that floor, so no resolve changes.
• The Python 3.13 and 3.14 classifiers are added.
• The new release and lock dependency groups are contributor tooling, not published extras. Extras stay floors only, with no upper pins.

Verification-status changes (§0.3)

In docs/verified-apis.md:

• The constructor rows of all eight framework adapters move from UNVERIFIED (signature confirmed) to VERIFIED (LIVE): LangGraph, Google ADK, MS Agent Framework, OpenAI Agents SDK, Anthropic, CrewAI, LlamaIndex and Strands. Each was verified by a live round trip on 2026-10-13 (#1124).
• The EU, CA and JP Hyperforce control-plane hosts are verified (#1113).
• No row moved from verified to blocked.

Acceptance

donkey-kit==0.1.3 from TestPyPI passed the acceptance harness against the live proxies, with all 20 proxies configured:

• Main session: 250 passed, 19 skipped.
• Packaging: 27 passed, 3 skipped.
• Per-adapter sessions: all passed.

The skips are by design: capability-gated routes, and modules removed in 0.1.2. The one exception is the Agent Kill Switch check, which skipped because the test agent was not quarantined at run time. The automated offline gate (groups 3, 4 and 5) also passed against the published package.

Full changelog: v0.1.2...v0.1.3 (promotion PR #1138)