Skip to Content
These docs describe donkey-kit v0.2.1. Release notes
Roadmap

Roadmap

DDK is delivered in five phases, with consolidation releases between them. Each phase is a GitHub milestone with a clear goal; the progress bars and issue lists below are loaded from those milestones when you open the page. Phase 1 shipped as donkey-kit 0.1.0, and its stabilisation releases, 0.1.1, 0.1.2 and 0.1.3, followed. Phase 2 shipped as 0.2.0, and the consolidation release 0.2.1 sits on top of it before Phase 3. The newest release is the one on PyPI , with its notes on the release notes page. The status pill on each phase below shows which one is in progress.

Status synced from GitHub milestones on 10 November 2026.

Complete every issue in the milestone is closed · In progress work has landed and more is open · Planned designed, not started. Want to help move a phase forward? See Contribute.

Phase 1 — Build the MVP

Version 0.1.0 · Goal: a developer who tries DDK for fifteen minutes finds three things they cannot get from a base_url and two headers — and one of them saves them from a production incident.

The shared governed transport and the capabilities that hang off it: typed refusals, budget & pacing, the local simulator, simulate() and the conformance suite, OpenTelemetry GenAI spans, correlation IDs and cost tags. Plus a deep LangGraph adapter, connection_kwargs() for seven more frameworks, the decorators and CLI, the documentation site, and the PyPI release. The phase closed with steering the gateway’s semantic cache, surfacing cache hits and semantic-routing matches (topic and score) on last_call, and fixing model-substitution detection when the gateway strips a provider prefix.

100% complete · 135 done · 0 open

Done (135)
  • ●#9 Epic: Testing — the adapter conformance kit, contract fixtures, nightly matrix
  • ●#10 core/config: DonkeyConfig resolution order and all-at-once validation
  • ●#11 core/transport: the single shared HTTP client that injects every governed header
  • ●#12 core/auth: AuthProvider protocol, AnypointConnectedApp, StaticToken, ChainedAuth
  • ●#13 core/errors: the governed error taxonomy and classify()
  • ●#14 core/telemetry: correlation IDs, run_context, and optional OpenTelemetry spans
  • ●#16 Donkey facade: lazy per-framework adapters, curated ImportError, async and sync lifecycle
  • ●#17 core: keep core framework-free — import-linter contracts as a blocking CI gate
  • ●#18 llm: donkey.llm.client() — the framework-free governed OpenAI client (async + blocking overloads)
  • ●#24 adapter/langgraph: chat_model() returning a native ChatOpenAI (Tier 1)
  • ●#33 adapters: connection_kwargs() and module-level factories across all eight frameworks
  • ●#179 transport: add the four lifecycle hooks every Phase 1 feature attaches to
  • ●#180 donkey: Donkey.openai() returns a native openai.AsyncOpenAI
  • ●#181 errors: re-align classify() to the documented v1.11-v1.13 rejection contracts
  • ●#182 errors: require a non-empty remediation on every PolicyViolation
  • ●#183 errors: prove policy refusals are terminal and openai's built-in 429 retry never fires
  • ●#184 errors: honest fall-through for guardrail verdicts whose shape is unconfirmed
  • ●#185 budget: Budget as a first-class object parsed from x-token-* headers
  • ●#186 budget: wait_for_reset() and pace(reserve=) raising BudgetReserveReached
  • ●#187 simulator: donkey mock — a local HTTP server that replays real rejection fixtures
  • ●#188 simulator: scenario scripting for PII, budget and injection
  • ●#189 simulator: the honesty guarantee — x-donkey-simulator header and a stock-client test
  • ●#190 testing: donkey.simulate() — in-process refusal injection with no server
  • ●#191 testing: the pytest plugin and the conformance suite customers run against their own agent
  • ●#192 telemetry: OTel GenAI spans at a pinned semconv version, dual-emitted with a stable donkey.* namespace
  • ●#193 telemetry: a refused request still produces a span, and streaming produces exactly one
  • ●#194 telemetry: zero-config OTLP export, one-flag opt-out, <1ms overhead benchmarked in CI
  • ●#195 telemetry: per-call and per-run correlation IDs via donkey.run()
  • ●#196 telemetry: validated cost-attribution tags (team, project, env, enduser.id)
  • ●#197 adapters: demote seven adapters to connection_kwargs-only and cut the conformance roster to LangGraph
  • ●#198 adapter/langgraph: deepen to the one conformance-gated adapter
  • ●#199 demos: langgraph-support-triage implements Scenario A end-to-end against the simulator
  • ●#200 decorators: @donkey.governed and @donkey.tool
  • ●#201 cli: the donkey command surface — init, plus consistent flags and exit codes across all four
  • ●#202 cli: donkey doctor — tell wrong credentials from wrong URL from model-not-allowed
  • ●#203 docs: a 15-minute quickstart that needs no gateway
  • ●#204 docs: scenario pages, the refusal cookbook, and per-framework connection_kwargs pages
  • ●#205 docs: publish llms.txt so coding assistants can read the SDK docs
  • ●#206 release: publish to PyPI with semver and a changelog
  • ●#207 meta: replace the build plan with the phase-based plan and realign CLAUDE.md and the skills
  • ●#259 epic: Phase 1 — Build the MVP
  • ●#266 docs: migrate the ~497 legacy §N.N citations to BG §N.N and retire spec/archive
  • ●#267 docs: realign the docs site to the phase model and cut the retired surfaces
  • ●#270 docs: drop the Omni Gateway architecture image from the feature overview
  • ●#273 Remove demos/ after moving it to a dedicated companion repo
  • ●#277 adapters: free donkey.openai by renaming the Agents SDK adapter
  • ●#278 testing: an out-of-process gateway fixture tests can assert against
  • ●#281 Git-ignore the .playwright-mcp/ local MCP artifact directory
  • ●#285 mypy --strict fails on core/telemetry.py get_tracer under opentelemetry-api 1.44
  • ●#286 Negative Retry-After header makes the retry sleep raise ValueError
  • ●#289 errors: type the documented Regex-Prompt-Guard and Content-Safety refusals; reconcile PIIDetected.entities
  • ●#292 Realign ARCHITECTURE.md with the code and add an implement-stage skill
  • ●#295 testing: end-to-end pace→resume test against the local gateway simulator
  • ●#296 Codify "deferred scope → dedicated follow-up issue" as a stated rule in the skills
  • ●#306 telemetry: keep prompt and completion content out of spans by default
  • ●#307 budget: surface cached and reasoning token counts from the live usage payload
  • ●#339 Publish donkey-kit to PyPI and wire project URLs
  • ●#349 Wire the DDK stacked-black logo into README and the docs site
  • ●#352 budget: Budget never observes on a live 200 — the gateway emits x-llm-proxy-ratelimit, not x-token-*
  • ●#353 simulator: happy-path budget overlay emits x-token-*, but the live 200 emits x-llm-proxy-ratelimit
  • ●#362 skeleton: define the last-call record once — #307 and #309 each assume a different container, and gateway identity is surfaced nowhere
  • ●#363 refusals: overriding correlation_header silently breaks DonkeyError.correlation_id read-back through classify()
  • ●#366 Quickstart Configure step omits how to get client_id/client_secret
  • ●#379 errors: a typed GatewayUnavailable for transport-level failures
  • ●#383 Move Claude config & skills to a private companion repo
  • ●#397 docs: document @donkey.governed / @donkey.tool decorators (follow-up to #396)
  • ●#400 Sandbox test suite against the provisioned LLM Gateway proxies
  • ●#410 release: add workflow_dispatch → TestPyPI dry-run path (dev snapshots aren't GitHub Releases)
  • ●#413 docs: README logo uses a relative path → broken on the PyPI project page
  • ●#415 docs: document the versioning & naming convention publicly (it lives only in a private skill)
  • ●#418 chore(release): bump version to 0.1.0.dev1 for a fresh TestPyPI dry-run
  • ●#423 Upgrade website to Nextra 4 + Next 16 (supersedes #422)
  • ●#425 Distinguish internal and fork contributors in CONTRIBUTING.md
  • ●#434 docs: build plan still says standing-invariant § anchors resolve into the deleted archive
  • ●#438 chore(website): pin @xmldom/xmldom to 0.9.12 to clear 13 Dependabot alerts
  • ●#440 Set least-privilege GITHUB_TOKEN permissions in ci, nightly-matrix, rebrand-completeness workflows
  • ●#445 Gitignore the .claude/settings.json symlink in the public repo
  • ●#448 bug(skeleton): Donkey.aclose()/close() leaks the control-plane auth's httpx client
  • ●#449 test(skeleton): no unit coverage for the AnypointConnectedApp OAuth2 token path
  • ●#450 docs(skeleton): core/auth.py narrates the (verified) OAuth2 token path as UNVERIFIED
  • ●#451 bug(refusals): classify() AuthError cites control-plane §1 for a data-plane §2 rejection
  • ●#452 bug(budget): pace()+wait_for_reset() recovery loop can livelock after window exhaustion
  • ●#453 test(simulator): packaged (pip install) fixture-resolution path is never exercised
  • ●#454 docs(testing): gateway-identity conformance exemptions asserted in code but not published as credibility
  • ●#455 docs(telemetry): telemetry.mdx banners the shipped OTel GenAI feature as "not yet shipped"
  • ●#456 docs(adapters): adapter import comments falsely claim framework class names are "verified"
  • ●#457 test(adapters): AgentFrameworkAdapter.policy_middleware() and chat_client() blocked-path are uncovered
  • ●#458 chore(cli): refused provisioning control-plane is marketed in --help and exported in top-level __all__
  • ●#459 docs: README headlines "provisioning-as-code" as a flagship capability (a "Do not build" surface)
  • ●#460 docs(cli): stale citations in provisioning/cli.py docstrings (deleted-v1 milestones; wrong BG §1.8)
  • ●#461 test(cli): provisioning spec/planner/applier/lint and the hidden CLI commands have zero coverage
  • ●#462 bug(tools): ToolSet.filter() mutates the receiver in place instead of returning an independent view
  • ●#463 bug(tools): duplicate AssetType symbols — package exports the Enum, AssetRef.type expects the Literal
  • ●#464 test(tools): no test exercises the _verify.blocked() guards in registry/ and tools/session.py
  • ●#465 docs: docs/unsupported-boundary.md is an empty template but README promises it classifies platform APIs
  • ●#471 Distinguish shipped Phase 1 cards from roadmap proposals
  • ●#484 AuthError.remediation gives data-plane advice for control-plane token rejections
  • ●#486 Replace remaining M0/M4 milestone citations outside provisioning
  • ●#487 pace() retry loop spins at zero delay when reset_at is None
  • ●#488 Correct the "one retry" claim for pace() after the window resets
  • ●#489 Inject the clock into pace() and route the e2e retry through pace()
  • ●#491 Expose fixture-lock helpers that work outside a source checkout
  • ●#496 Retire M0/M1/M4 milestone vocabulary outside python/src
  • ●#497 Clarify paired export verification messages
  • ●#501 Cover remaining verification-blocked surfaces
  • ●#504 Reconcile header-only adapters with correlation conformance
  • ●#509 Support JWT/model-wallet auth mode for the LLM proxy data plane
  • ●#510 Document apikey/bearer auth convenience and DDK's two-header stance
  • ●#517 Drop :8081 from the sandbox proxies.toml base_url guidance
  • ●#522 Verify inbound correlation and cost-tag request header names against the live gateway
  • ●#533 Group Typed refusals under the six-piece-minimum sidebar separator
  • ●#535 Docs site typography: tighten sidebar items and hero tagline size
  • ●#541 Document the proxy ingress Format (OpenAI/Anthropic/Gemini) to adapter mapping
  • ●#542 request_id is None on Bedrock routes: x-request-id is the upstream provider's header, not the gateway's
  • ●#548 classify() types a Gemini upstream 400 as a gateway PolicyViolation: list-shaped error envelope is discarded
  • ●#556 epic(docs): restructure the docs site into end-user documentation organised by capability pillars
  • ●#557 docs(intro): rewrite the Introduction around gateway awareness, with architecture and before/after sections
  • ●#558 docs(quickstart): simplify the Quickstart and mark each capability Live or Roadmap
  • ●#559 docs(feature-overview): high-level capability tour with the gateway-aware image
  • ●#560 docs(nav): restructure the sidebar from phase separators into capability pillars
  • ●#561 docs(pages): remove issue refs, phase labels and verification narration from capability pages
  • ●#562 docs(examples): add an Examples pillar built on the donkey-development-kit-demos repo
  • ●#563 docs(roadmap): phases in order with live milestone progress and the issues in each
  • ●#564 docs(community): add a Team (About us) page with Creator/Contributor pills
  • ●#565 docs(community): add a Contribute page condensed from CONTRIBUTING.md
  • ●#566 docs(site): open-source disclaimer footer, metadata and optimised images
  • ●#568 errors: flip Amazon Bedrock Guardrails refusal from documented-only to VERIFIED (LIVE) after the acceptance capture
  • ●#586 last_call.substituted is always True for provider/model names (gateway strips the routing prefix)
  • ●#587 Support gateway semantic caching: steer requests and surface hits
  • ●#590 Surface the semantic-routing match topic and score on last_call
  • ●#592 docs(roadmap): fetch milestone status live from GitHub, keep the snapshot as fallback
  • ●#593 docs(site): sync pages with the current code and the issues filed since the restructure
  • ●#595 errors: PolicyViolation's default remediation still calls content-moderation shapes under-documented
  • ●#597 mypy --strict fails under openai<3 with 4 stale ignore/cast guards
  • ●#601 docs: extend the gateway-identity example with a semantic-routing walkthrough (follow-up to #599)

Phase 1.1 — Stabilize the MVP

Version 0.1.1 · Goal: make the 0.1.0 surface safe and dependable to build on, without adding Phase 2 scope.

Credentials scoped to the plane and the config source they belong to, endpoint trust (no credentials sent to unchecked or non-loopback http:// endpoints), and secrets kept out of printed output and refusal messages. Adapter and transport fixes across the frameworks, a typed AgentKilled refusal for the Agent Kill Switch, an all extra that installs, a native donkey.adk.gemini() model for Gemini-format proxies, and a pass over the docs. Upgrading from 0.1.0? See the migration guide .

100% complete · 83 done · 0 open

Done (83)
  • ●#630 Adapter factories raise TypeError when a governed kwarg is overridden
  • ●#631 LangGraph agent example misdescribes last_call and budget after the loop
  • ●#658 langgraph: surface the gateway's served provider/model per message, not just on last_call
  • ●#659 docs/langgraph: explain model_provider vs the served provider, and how to read it per call and per message
  • ●#660 docs(contributing): codify live-capture, attribution and multi-PR guidelines
  • ●#661 docs: state that Omni/Flex Gateway Local Mode is not supported; the simulator is the local dev loop
  • ●#663 Add a PyPI version badge to the docs site top bar
  • ●#664 Add a PyPI version badge to the org homepage README
  • ●#669 reject.injection-protection fixture has an empty body; live capture returned the real 79-byte message
  • ●#670 reject.injection-protection fixture headers capture only 3 of ~10 live response headers
  • ●#671 Add a PyPI version badge to the package README and drop the hardcoded version
  • ●#674 Split release.yml into publish-pypi and publish-testpypi workflows
  • ●#675 Remove the rebrand-completeness CI guard and migration scaffolding
  • ●#680 Track §8 verification-ledger follow-ups from PR #679 review
  • ●#681 Reconcile §8 constructor-status vocabulary in verified-apis.md
  • ●#682 Fix stale §4 cross-reference for the /responses data plane (should be §2)
  • ●#683 Deduplicate the CrewAI return-class finding across §8 and §8.1
  • ●#684 donkey.crewai.llm() is annotated -> crewai.LLM but returns OpenAICompletion
  • ●#687 Refresh docs/releasing.md wording that predates the 0.1.0 publish
  • ●#688 docs: document how to configure the local simulator's port (CLI flag, serve(), no env var)
  • ●#691 ADK: donkey.adk.gemini() — native Gemini model bound to a Format=Gemini proxy
  • ●#692 Add production-only Cloudflare Web Analytics beacon to the docs site
  • ●#694 classify() mis-types the Agent Kill Switch 403 as UpstreamRequestError
  • ●#697 pip install donkey-kit[all] fails with resolution-too-deep
  • ●#701 donkey.anthropic.client() raises TypeError with anthropic>=1.0 (httpx2)
  • ●#715 bug(errors): export every documented error and give every DonkeyError a remediation
  • ●#716 bug(typing): make governed() and the adapter API type-check for downstream mypy --strict users
  • ●#718 docs(src): stop restating verification status in docstrings and fix the drifted ones
  • ●#725 bug(adapters): back module-level adapter factories with one process-default governed runtime
  • ●#733 bug(adapters): never hand the owning shared client to a framework (Strands closes it after one call)
  • ●#734 bug(adapters): turn off framework retries so a policy refusal is sent exactly once
  • ●#735 bug(adk): guard gemini() against google-adk versions that silently drop the governed client
  • ●#736 bug(langgraph): route sync invoke() through the governed transport and define sync support per adapter
  • ●#737 bug(conformance): never pass, and never reach the network, when the agent bypasses the Donkey client
  • ●#738 bug(anthropic): keep correlation and call IDs on refusals that pass through the httpx2 bridge
  • ●#739 bug(agent-framework): fix or withdraw policy_middleware()
  • ●#741 bug(adapters): raise the same curated missing-dependency error from every form
  • ●#743 bug(deps): raise the extras floors to the lowest verified versions
  • ●#773 docs(config): remove non-existent from_env() kwargs and document the real construction and precedence
  • ●#775 docs: reconcile adapter verification status across README, ARCHITECTURE, the site and verified-apis §8
  • ●#777 docs(migration): fix the conformance flag, the CLI list and the never-emitted telemetry names
  • ●#779 docs(site): label roadmap features on the home page and README extras, and align the discovery sketch
  • ●#781 docs(config): update the header and attribution docs to the verified gateway facts
  • ●#783 docs(conformance): publish the missing JWT-refresh exemption and correct the JWT-capable adapter list
  • ●#785 docs(contributing): remove links to untracked files, dead docs-sync targets and a harness that doesn't exist
  • ●#787 docs(site): fix broken links and anchors, make the PyPI README links absolute, and point site links at main
  • ●#789 docs(architecture): fix the ARCHITECTURE.md divergences from the code
  • ●#791 docs(refusals): fix the typed-refusals example whose except clauses can never fire
  • ●#793 docs(examples): use DonkeyConfig.validated() in the examples and fix the verification contradictions
  • ●#801 bug(simulator): simulate(), donkey mock and the conformance harness must stay offline under HTTP(S)_PROXY
  • ●#803 bug(telemetry): stop binding one correlation ID to the whole async context outside donkey.run()
  • ●#805 bug(telemetry): classify streamed refusals before recording the span
  • ●#807 bug(transport): make the shared async client safe to reuse across event loops
  • ●#809 bug(config): validate config values and raise ConfigError instead of builtin exceptions
  • ●#811 bug(cli): make doctor honour --config/--env, route the entry point through main(), and avoid SystemExit at import
  • ●#813 bug(errors): raise typed errors on the documented surfaces (doctor outage detection, token fetch, closed client)
  • ●#815 bug(config): read [targets] from the XDG config, and never return a negative retry_after
  • ●#817 bug(telemetry): a failing span recorder silently drops streamed usage from donkey.last_call
  • ●#825 bug(refusals): stop typing the bare-model-name 400 as a PolicyViolation
  • ●#826 bug(agent-framework): chat_client() hits /responses; Azure routes 404
  • ●#827 bug(anthropic): last_call misses request-id, cache and streamed tokens
  • ●#828 bug(adapters): jwt mode sends the api_key placeholder as the bearer
  • ●#829 bug(llamaindex): prefixed names get temperature 0.1 and a 3,900 window
  • ●#830 upstream: Gemini-routed chat streams one chat.completion, not chunk deltas
  • ●#833 bug(transport): send LLM-proxy-only headers only on data-plane requests
  • ●#834 bug(ci): install the local and otel extras in the nightly langgraph job so its example smoke step can pass
  • ●#835 bug(adapters): raise ConfigError from the header-only adapter forms in jwt mode
  • ●#836 feat(adapters): add a first-class bearer-token option for LLM-proxy model calls
  • ●#837 bug(config): read ~/.config/.donkey-kit.toml when XDG_CONFIG_HOME is unset
  • ●#838 chore(release): close the pending security advisories after 0.1.1
  • ●#850 donkey.last_call stays UNOBSERVED after ChatOpenAI.ainvoke()
  • ●#865 LangGraph interrupt() in async nodes fails on Python 3.10
  • ●#881 develop is red: CLI entrypoint test asserts the pre-#857 region error
  • ●#882 CLI entrypoint test asserts the pre-#857 unknown-region message, failing CI on every PR
  • ●#883 test_cli_entrypoint asserts the pre-#857 region error, so CI on develop is red
  • ●#891 Make ModelNotRoutable simulatable via simulate() and donkey mock
  • ●#899 Check whether non-LangGraph adapters leave last_call UNOBSERVED after a single async call
  • ●#901 test_main_reports_donkey_error_without_traceback fails on develop after #857 reworded the region error
  • ●#903 Teach the shared-client test helpers about anthropic>=1's bridged httpx2 client
  • ●#907 docs(examples): resync demo 08's verification wording in the framework-objects example
  • ●#909 Serve the SDK docs site from the docs.donkey-kit.dev custom domain
  • ●#915 Update release docs and version references for the 0.1.1 release
  • ●#916 Enable the Cloudflare Web Analytics beacon on docs.donkey-kit.dev

Phase 1.2 — Stabilize the MVP

Version 0.1.2 · Goal: act on a full code review of the 0.1.1 surface, so later phases build on settled foundations.

One transport core shared by the sync and async clients, the documented config precedence, a single adapter contract and roster, and typed refusals recovered from whatever a framework re-raises, including inside an ExceptionGroup. Telemetry no longer installs the global OpenTelemetry provider unless you ask it to, the public API surface is defined, and budget pacing also honours request-count rate limits. Around them: security and packaging hygiene, architecture decision records, and a stricter test suite. No Phase 2 scope. Upgrading from 0.1.1? See the migration guide .

100% complete · 77 done · 0 open

Done (77)
  • ●#706 epic(conventions): code review — coding conventions & code quality
  • ●#707 epic(architecture): code review — architecture
  • ●#708 epic(adapters): code review — framework integrations
  • ●#709 epic(packaging): code review — packaging & dependencies
  • ●#710 epic(testing): code review — testing
  • ●#713 epic(robustness): code review — robustness bugs
  • ●#714 epic(security): code review — security & repository hygiene
  • ●#717 feat(logging): adopt a logging convention and stop swallowing failures silently
  • ●#719 feat(api): define the public API surface and remove cross-package private access
  • ●#720 chore(core): remove dead code and single-source duplicated definitions
  • ●#721 chore(lint): encode the house style in ruff and mypy config
  • ●#722 docs(src): add docstrings to the main public entry points
  • ●#723 feat(conventions): ADR for value objects — frozen dataclasses, pydantic only at external-schema boundaries
  • ●#724 feat(refusals): a framework-agnostic typed-refusal bridge
  • ●#726 feat(adapters): formalise the adapter contract and keep the roster in one place
  • ●#727 feat(config): implement the specified config precedence with a declarative field table
  • ●#728 refactor(transport): split core/transport.py around a sans-IO policy shared by the sync and async clients
  • ●#729 feat(lint-imports): architecture rules as code — extend the import contracts and test each invariant
  • ●#730 refactor(cli): move the supported CLI to donkey_kit/cli and quarantine legacy provisioning
  • ●#731 feat(adr): introduce an ADR process and seed the architecture decisions from this review
  • ●#732 feat(telemetry): don't install the global OTel TracerProvider implicitly
  • ●#740 feat(adapters): use transport injection for LlamaIndex, MAF and CrewAI
  • ●#742 test(adapters): add an integration checklist and an adapter contract suite parametrised from ADAPTERS
  • ●#744 build(deps): move dev tooling to dependency groups and stop publishing extras with no code behind them
  • ●#745 build(pkg): modernise the license and classifier metadata and publish a Python support policy
  • ●#746 build(pkg): ship simulator fixtures as package data and make the pytest11 plugin cheap and collision-safe
  • ●#747 test: make the suite hermetic — clear proxy/OTEL/DONKEY env, block real sockets, cap test time
  • ●#748 test(ci): run framework-dependent tests against real framework installs
  • ●#749 test(conformance): implement or retire the internal 20-scenario conformance matrix
  • ●#750 test: reset module-level state between tests and randomise test order in CI
  • ●#751 test(ci): measure coverage in CI and enforce a floor that only goes up
  • ●#752 test(fixtures): extend the integrity lock to every captured fixture
  • ●#753 test: scheduled live-contract check against the fixtures, plus benchmark trending
  • ●#755 ci: run CI on pushes to develop and alert on nightly failures
  • ●#761 ci: add job timeouts, concurrency cancellation, pip caching and fail-fast: false
  • ●#819 chore(security): commit ignore rules for local secret-bearing config files and add secret scanning
  • ●#821 chore(security): scrub tenant identifiers and live endpoints from fixtures and the published wheel
  • ●#823 chore(security): add SECURITY.md and enable private vulnerability reporting
  • ●#832 chore(governance): apply the endpoint rule to [targets.*] base URLs before target calls are enabled
  • ●#895 simulator: serve POST /chat/completions
  • ●#896 llm/cli: check donkey.llm and donkey doctor on routes that 404 /responses
  • ●#920 Keep the PyPI version badge current right after a release
  • ●#922 Move the PyPI badge before the search field in the docs-site top bar
  • ●#932 docs: document how to turn on the SDK's DEBUG logging
  • ●#933 `classify(exc.response)` fails mypy --strict on openai>=3 and anthropic>=1
  • ●#934 docs: add the 0.1.2 deprecations, removals and renames to MIGRATION.md
  • ●#935 Group auto-generated release notes by label and always include a Breaking changes section
  • ●#936 chore(deps): gate new direct dependencies behind a reviewed allowlist and a PyPI existence check
  • ●#948 chore(ci): resolve PR CI dependencies from a lock (ADR 0007 rule 3)
  • ●#949 fix(runtime): close the standalone adapter's own blocking client (ADR 0003 gap)
  • ●#950 feat(refusals): translate refusals inside ExceptionGroup / TaskGroup
  • ●#951 fix(strands): don't let Strands retry a terminal 429 token-budget refusal
  • ●#953 fix(transport): stamp x-should-retry: false when a 502/504 model POST is not retried
  • ●#955 test(refusals): typed-refusal bridge conformance for openai-agents, LlamaIndex, ADK gemini()
  • ●#956 fix(tests): make test_adk_last_call_status_does_not_depend_on_factory_order order-independent
  • ●#957 Give the seven framework examples a build() for the conformance kit
  • ●#958 Route CrewAI through the shared transport via interceptor hooks, or record why not
  • ●#970 docs: fix 0.1.2 doc drift in MIGRATION, READMEs, CONTRIBUTING, ARCHITECTURE and ADRs
  • ●#972 Ignore .venv* variants in .gitignore
  • ●#974 budget: pace on request-count rate limits, not only the token window
  • ●#983 test(conformance): cover every typed_refusals=True factory in the bridge suite
  • ●#984 docs(readme): correct the ADK model() typed-refusal exemption wording
  • ●#985 chore(strands): drop the retry_strategy=None workaround now that #951 landed
  • ●#996 Add docs/ci.md: CI/CD conventions and workflow map
  • ●#999 docs: document live-contract-check setup (live-sandbox environment + secrets) in docs/ci.md
  • ●#1004 ci: pin one Python patch version per minor so pip cache keys match
  • ●#1005 ci: save pip caches only from main/develop pushes; PRs restore only
  • ●#1006 ci-ok and nightly alert jobs lack timeout-minutes, failing develop CI
  • ●#1009 ci: make GitHub Actions workflow names uniform across DDK repos
  • ●#1010 test: test_the_secrets_warning_names_the_callers_file is flaky
  • ●#1014 Nightly framework matrix failing
  • ●#1015 docs(ci): add Mermaid scenario maps for feature, hotfix, and recovery
  • ●#1043 feat(adapters): default agent-framework and LangGraph to Chat Completions
  • ●#1044 chore(testing): replay captured route fixtures in contract_gateway.py
  • ●#1050 docs: quickstart install omits the cli extra, so donkey mock fails
  • ●#1057 Update release docs and version references for the 0.1.2 release
  • ●#1060 ci: benchmark job fails on main pushes because the hermetic env strips DONKEY_BENCHMARK_JSON

Phase 1.3 — Stabilize the MVP

Version 0.1.3 · Goal: make the docs and the release process as dependable as the code.

Documentation with one source of truth per class of fact, including generated reference tables, a framework version-compatibility page and a version banner on every page. Automated release gates, a tested support matrix across Python and dependency versions, a watch on supported-framework releases, a hardened build supply chain, and a documented hotfix process. Typed refusals are now carried across ADK’s model() (LiteLLM) path. No Phase 2 scope.

100% complete · 39 done · 0 open

Done (39)
  • ●#711 epic(ci): code review — CI/CD & developer experience
  • ●#712 epic(docs): code review — documentation, examples & website
  • ●#757 ci(security): harden the Actions supply chain — SHA pins, Dependabot, CODEOWNERS, least-privilege Pages
  • ●#759 ci(site): build, type-check and link-check the docs site on PRs, and move off Node 20
  • ●#763 ci(deps): pin the dev and release toolchain with a lock, and make mypy independent of installed extras
  • ●#765 chore(dx): one task runner that mirrors CI, pre-commit, and an honest CONTRIBUTING gate section
  • ●#767 ci(release): automated release gates — one version source, tag check, wheel smoke test, reviewers, attestations
  • ●#769 ci: test the declared support matrix — lowest-direct per extra, openai<2, Python 3.13/3.14, co-installability
  • ●#771 ci(site): pin manual docs deploys to main
  • ●#795 docs: one source of truth per class of fact — generated tables, API reference, release notes, version banner
  • ●#797 docs: make doc snippets tests, add a frontmatter status vocabulary, and add a path-based docs-sync reminder
  • ●#799 feat(site): add baseline SEO and share metadata
  • ●#969 feat(refusals): bridge typed refusals through ADK model() (LiteLLM)
  • ●#1031 Add mermaid diagrams for release scenarios to docs/releasing.md
  • ●#1033 docs(verified-apis): reflect Strands stream=False default and Gemini streaming caveat
  • ●#1035 docs(architecture): add mermaid diagrams for client wiring and classify()
  • ●#1038 Document the hotfix release process in docs/releasing.md
  • ●#1042 Add Phase 1.2 and 1.3 milestones to the docs/releasing.md version table
  • ●#1048 Reject placeholder git identities in PR commits
  • ●#1052 Nightly framework matrix failing
  • ●#1053 docs(site): add mermaid diagrams for classify() order and credential checks
  • ●#1063 docs: add a table of contents to the root and docs/ markdown files
  • ●#1065 epic(ci): watch supported-framework releases, assess their impact, and publish supported versions
  • ●#1066 ci: detect new framework releases and test each one in isolation against the PR lock
  • ●#1067 ci: add an AI changelog impact review to the framework release watch
  • ●#1068 docs: publish a generated framework version-compatibility page
  • ●#1081 docs: single-source per-framework verification status and drift-check it
  • ●#1082 ci(release): dispatch acceptance after TestPyPI publish, report on PR
  • ●#1088 ci(deps): stop Dependabot proposing @types/node and TypeScript majors the site can't take
  • ●#1090 docs(examples): stop calling the connection_kwargs-only examples "conformance-tested"
  • ●#1091 chore(repo): PR template test plan lists the pre-nox raw gate
  • ●#1093 chore(website): override katex to 0.18.10 to clear Dependabot alert #49
  • ●#1114 chore(release): let bump-version.sh link a milestone to the bump issue and PR
  • ●#1116 framework-release: crewai 1.15.23 → 1.15.27
  • ●#1117 chore(release): bump version to 0.1.3.dev0
  • ●#1120 Remove the Agent Fabric → DDK rebrand references from the docs
  • ●#1122 ci(release): TestPyPI acceptance job fails to start (no checkout for the python working directory)
  • ●#1133 Update release docs and version references for the 0.1.3 release
  • ●#1134 chore(release): bump version to 0.1.3

Phase 2 — Differentiate, go beyond

Version 0.2.0 · Goal: capabilities no generic LLM client offers, because they depend on the platform behind the gateway.

Governed tool access (MCP discovery and binding), A2A serve / expose / dev, on-behalf-of identity, human-in-the-loop, scan & publish to the registry with a GitHub Action, declarative refusal handlers and a classification registry for custom policies, typed federated guardrail verdicts, kill-switch awareness with the kill reason, and a second deep framework adapter chosen by demand.

Around them: an inbound correlation ID carried across MCP and A2A hops, a stated concurrency contract for budget and telemetry under parallel fan-out, a per-request business group for shared multi-tenant clients, bring-your-own provider keys, and a token and registry cache. Also in this phase: governed access to TypeSafe Jev , a System One decision model, with the same typed refusals, spans, budget and simulator support as LLM calls, plus worked patterns for confidence-gated review, tool ranking and budget-aware routing. Now that the package is published, the phase also picks up archiving daily PyPI adoption metrics.

100% complete · 131 done · 0 open

Done (131)
  • ●#15 core/cache: TTLCache for tokens and registry lookups, with DONKEY_NO_CACHE escape hatch
  • ●#42 tools: ToolSet.filter() — allow/deny/predicate filtering and descriptor token accounting
  • ●#208 refusals: declarative reaction handlers registered once, applied at the transport
  • ●#209 refusals: enforce that a handler can react but never un-refuse
  • ●#210 classify: a registry so custom gateway policies become typed exceptions
  • ●#211 hitl: one vocabulary for pause-and-ask-a-human
  • ●#212 hitl: map the vocabulary onto LangGraph interrupt() and Command(resume=)
  • ●#213 hitl: mappings for OpenAI Agents, ADK, Strands and MCP elicitation
  • ●#214 hitl: carry the gateway's step-up challenges (OBO + CIBA) as typed approvals
  • ●#215 identity: donkey.as_user() — RFC 8693 token exchange and header placement
  • ●#216 identity: a missing or expired user token fails typed, never falls back to service identity
  • ●#217 scanner: donkey scan derives a deterministic scan model of agents, tools and MCP servers
  • ●#218 scanner: donkey publish registers the manifest with Exchange / Agent Registry
  • ●#219 scanner: the scan-and-publish GitHub Action
  • ●#220 killswitch: recognise the kill-switch rejection and stop cleanly
  • ●#221 tools: donkey.tools.discover() returns governed MCP tools as native framework tools
  • ●#222 tools: allow/deny filtering and descriptor token accounting
  • ●#223 adapters: a second deep adapter, chosen by Phase 1 demand
  • ●#224 a2a: donkey.serve() — an A2A listener in one line, wrapping the official a2a-sdk
  • ●#225 a2a: governance-aware inbound — correlation, tags, spans and typed refusals on served tasks
  • ●#226 a2a: donkey expose — the gateway instance, policies and registry entry for an A2A agent or MCP server
  • ●#227 a2a: donkey dev — a gateway in front of the laptop, real or simulated
  • ●#260 epic: Phase 2 — Differentiate, go beyond
  • ●#308 telemetry: adopt an inbound correlation ID and propagate it across MCP and A2A hops
  • ●#309 skeleton: surface gateway-side routing and fallback, and never double-retry it
  • ●#310 skeleton: per-request business group for shared, multi-tenant gateway clients
  • ●#311 budget + telemetry: state and test the concurrency contract under parallel fan-out
  • ●#312 tools: a typed refusal for a tool outside the consumer's MCP access profile
  • ●#313 refusals: type the Akamai federated guardrail verdict
  • ●#314 killswitch: carry the kill reason — manual, anomaly, or eval drift
  • ●#322 testing: --donkey-mode server path for the conformance plugin
  • ●#367 skeleton: bring-your-own-key via a configurable proxy key header (Dynamic key)
  • ●#444 conformance: wire the Phase 2–2.8 scenario bodies as their features land
  • ●#503 scanner: define how a @donkey.tool is attributed to an MCP server / Exchange asset
  • ●#952 feat(cli): wire donkey --config to DonkeyConfig.resolve(path=) for doctor
  • ●#954 feat(telemetry): public flush for the scoped TracerProvider
  • ●#991 test(conformance): restore the 13 retired scenarios once their surfaces are verified
  • ●#1025 budget: read the upstream provider's x-ratelimit-*-requests window as a separate source
  • ●#1137 ci(release): finish the setup for the automated acceptance run after TestPyPI
  • ●#1147 Add status and framework badges to the SDK README and org homepage
  • ●#1150 tools: decide how discover() and binding treat MCP servers the gateway does not front
  • ●#1153 hitl: approval= on a discovered ToolSet, not just @donkey.tool
  • ●#1176 refusals: type the CrowdStrike AIDR Guardrails rejection
  • ●#1178 refusals: type the Google Model Armor Guardrails rejection
  • ●#1183 Report the content-safety provider and categories on the exception and span
  • ●#1193 scanner: capture MCP server transport and endpoint for publish (follow-up to #1192)
  • ●#1218 adk: bridge last_call for LiteLlm/Gemini built from connection_kwargs()
  • ●#1219 adk docs: confirm or drop "usage lands on the last one" for streaming
  • ●#1226 chore(release): let bump-version.sh file the bump's tracking issue
  • ●#1228 chore(release): bump version to 0.2.0.dev2
  • ●#1231 chore(website): bump zod override to 4.6.7 to clear Dependabot alert #57
  • ●#1234 hitl: pause Strands tool calls with interrupt() on BeforeToolCallEvent
  • ●#1235 hitl: map approval onto ADK FunctionTool(require_confirmation=...)
  • ●#1236 docs(hitl): record MCP elicitation as unsupported for client-side approval
  • ●#1237 fix(strands): typed refusals raised inside a tool don't leave the agent
  • ●#1238 fix(openai_agents): approvals miss namespaced and deferred tools
  • ●#1257 tools: log descriptor cost on bind and warn above a threshold
  • ●#1260 scanner: donkey publish publishes agent assets to Exchange
  • ●#1261 chore(release): bump version to 0.2.0.dev3
  • ●#1267 ADK model() agents with tools fail with "No module named 'mcp'" on litellm >= 1.106
  • ●#1268 `donkey --json publish --dry-run` does not print plan issues to stderr
  • ●#1269 `donkey publish`: publish provider, protocol version, capabilities and a documentation page so Agent Fabric renders the asset fully
  • ●#1271 Exchange page: list what changed since the previous version
  • ●#1273 Exchange page: an Overview section and each tool's approval and risk
  • ●#1281 `donkey publish`: apply a changed name/description to an existing Exchange asset
  • ●#1284 a2a: donkey dev Option B — simulated A2A ingress in front of donkey.serve()
  • ●#1294 chore(release): bump version to 0.2.0.dev4
  • ●#1301 Make the fan-out budget concurrency test deterministic instead of sleep-timed
  • ●#1312 tools: verify discover() and binding against a Mule-hosted MCP server
  • ●#1315 docs: build plan still lists donkey dev Option A as pending
  • ●#1318 ci: require the langgraph, llamaindex, anthropic and crewai MCP bindings in their adapter-contract legs
  • ●#1320 classify: type the OBO policy's CIBA authorization_pending instead of AuthError
  • ●#1323 feat(tools): a discover() span with cache hit/miss counters
  • ●#1324 feat(tools): donkey.tools.warm() to pre-fill the discovery cache
  • ●#1325 test(descriptor): fixture comparison fails on Python 3.13+ (docstring dedent)
  • ●#1326 chore(release): bump version to 0.2.0.dev5
  • ●#1331 chore(release): bump version to 0.2.0.dev6
  • ●#1348 chore: move Gemini defaults and examples from gemini-2.5-flash to gemini-3.8-flash
  • ●#1350 identity: make donkey.as_user() forward the user token instead of exchanging it
  • ●#1352 cli: wire donkey verify to Publication.verify()
  • ●#1364 chore(release): bump version to 0.2.0.dev7
  • ●#1366 identity: carry the donkey.as_user() token on MCP ToolSet bindings
  • ●#1370 ci(publish-testpypi): acceptance lookup can match a stale run for the same SHA and version
  • ●#1377 chore(release): bump version to 0.2.0.dev8
  • ●#1380 Flip CIBA tools/call and 502 rows to VERIFIED (LIVE) in verified-apis.md
  • ●#1414 epic(release): 0.2.0 pre-release code review
  • ●#1415 tools: stop MCP framework bindings forwarding client_secret on redirect
  • ●#1416 hitl: a resolved approval runs as the service, not the requesting user
  • ●#1417 tools: require https on MCP and A2A endpoints resolved from the gateway
  • ●#1418 identity: refuse a per-tool MCP binding called under a different user
  • ●#1419 tools: refresh() and warm() return stale governance from the registry cache
  • ●#1420 identity: rename as_user(id_token=) and drop the unreleased token_exchange_path
  • ●#1421 tools: rename ToolSet.openai() to ToolSet.openai_agents()
  • ●#1422 adapters: drop internal hooks from __all__ and unify the approval helpers
  • ●#1423 hitl: settle the Approvals API — typed errors, naming and scope
  • ●#1424 tools: confirm or revisit discover()'s unfiltered default before 0.2.0
  • ●#1425 release: settle the stability tier of the new 0.2.0 public surface
  • ●#1426 cli: donkey publish --json reports per-asset outcomes when a publish fails
  • ●#1427 cli: version every --json document and emit one on declaration issues
  • ●#1428 cli: add donkey --version and settle the Action's outputs before release
  • ●#1429 refusals: stop typing CrowdStrike AIDR until its shape is captured live
  • ●#1430 docs: HITL page marks unverified framework approvals Live and omits step-up
  • ●#1431 docs: add the 0.1.3 → 0.2.0 section to MIGRATION.md
  • ●#1432 docs: flip shipped 0.2.0 features from Roadmap to their real status
  • ●#1433 docs: correct stale blocked/verified claims in README, ARCHITECTURE and the ledger
  • ●#1434 docs: warn that Agents SDK tracing exports prompts to OpenAI by default
  • ●#1435 release: 0.2.0 readiness — rc1 on TestPyPI, acceptance sync, release notes
  • ●#1436 identity: give UserScope a sync with, and include mcp in [all]
  • ●#1437 hitl: cancel sibling CIBA waits when one server's listing fails
  • ●#1438 strands: consume an approval after one call, like ADK and OpenAI Agents
  • ●#1439 identity: apply the ADR 0005 endpoint check to the as_user() token
  • ●#1440 adk: raise the install-hint ImportError from the new ADK helpers
  • ●#1441 testing: back the README's every-other-adapter claims with conformance runs
  • ●#1442 ci: harden the Action and release workflows
  • ●#1443 a2a: harden donkey dev, serve() and scanned endpoints for non-loopback use
  • ●#1444 budget: an out-of-range reset header crashes send() with OverflowError
  • ●#1445 testing: add typecheck contracts for the new 0.2.0 headline API
  • ●#1446 refactor: remove the registry→tools reach-in and decouple Strands refusals
  • ●#1447 chore: 0.2.0 review hygiene sweep (CLI, simulator, ADK, ledger rows)
  • ●#1476 chore: ignore Cursor's machine-local .cursor/sandbox.json
  • ●#1478 core: refuse a user name or password in configured endpoint URLs
  • ●#1479 tools: apply the ADR 0005 endpoint check to the gateway credential on discovered servers
  • ●#1480 docs: classify publish's and the gateway reads' contracts on the unsupported-boundary page
  • ●#1481 scanner: a publish rerun writes the Exchange page without "Changes since"
  • ●#1489 chore(release): bump version to 0.2.0.dev9
  • ●#1499 api: ship ToolsFacade.live_tools and live_card as Provisional
  • ●#1502 langgraph: raise the install-hint ImportError from approve(), reject() and approval_interrupt()
  • ●#1503 agent_framework: policy_middleware() raises NotImplementedError instead of the install hint when agent_framework is missing
  • ●#1504 chore(release): bump version to 0.2.0rc1
  • ●#1506 chore(security): allowlist the synthetic test password so the Secret scan passes
  • ●#1520 chore(release): bump version to 0.2.0

Phase 2.1 — Consolidate Phase 2

Version 0.2.1 · Goal: settle the Phase 2 surface and finish the work that slipped from it, without adding Phase 3 scope.

A tool lockfile (donkey.lock) for governed MCP tools, a governed outbound A2A call path with typed refusals, simulate_mcp() and simulate_a2a() to inject named refusals offline, and a JSON-RPC refusal matcher shared by MCP and A2A. The gateway’s MCP and A2A policy refusals are verified and documented, PII detection and the content-safety blocks are typed, and raw PII echoed in refusal bodies is redacted. Schema-validation and tool-poisoning refusals are captured but stay untyped, because the gateway marks no policy on them.

100% complete · 87 done · 0 open

Done (87)
  • ●#1339 Get DDK referenced in LangGraph / LangChain docs
  • ●#1344 Get DDK referenced in Microsoft Agent Framework docs
  • ●#1345 Get DDK referenced in OpenAI Agents SDK docs
  • ●#1346 Get DDK referenced in Anthropic SDK docs
  • ●#1382 epic: verify, type and document the Omni Gateway MCP policies
  • ●#1383 epic: verify, type and document the Omni Gateway A2A policies
  • ●#1385 refusals: a JSON-RPC refusal matcher shared by MCP and A2A
  • ●#1386 refusals: type the MCP PII Detector rejection as PIIDetected
  • ●#1389 tools: support MCP Progressive Disclosure and Tool Mapping in discovery, binding, lockfile and approvals
  • ●#1390 tools: verify and document MCP Payload Optimization results
  • ●#1391 testing: MCP policy refusal fixtures and simulate() scenarios
  • ●#1392 docs: MCP gateway policies page, errors.mdx rows and verified-apis.md §9 rows
  • ●#1394 refusals: classify A2A instance policy refusals (v1 ErrorInfo, x-flex guardrail headers, x-token 429)
  • ●#1395 a2a: a governed outbound A2A call path with typed refusals
  • ●#1396 hitl: type the A2A In-Task Authorization Code auth-required challenge
  • ●#1397 simulator: donkey dev replays A2A-instance rejection shapes, not LLM-proxy ones
  • ●#1398 refusals: redact raw PII echoed in gateway refusal bodies before messages, logs and spans
  • ●#1399 telemetry: record A2A quality-evaluation headers and X-AGENT-ID connection telemetry
  • ●#1400 docs: A2A gateway policies page and verified-apis.md §4/§6 rows
  • ●#1404 verify: model-based-routing 1.0.5 native Anthropic/Gemini multi-vendor routing live
  • ●#1405 verify: x-llm-proxy-llm-provider carries groupId:vendor on model-based-routing 1.0.5; fix substitution false positive
  • ●#1408 registry: recognise the Exchange llm asset type
  • ●#1411 verify: publish a custom LLM provider asset and capture its llm-metadata file
  • ●#1413 docs: refresh native-ingress, provider and route-matrix rows for the Oct-2026 policy updates
  • ●#1482 docs: check hand-written website status badges against the ledger
  • ●#1483 tools: a discovery cache expiry still reuses older registry lookups
  • ●#1484 core: a scope object entered by two tasks at once leaks or breaks its binding
  • ●#1512 ci: the secret scan fails every PR when any branch has a finding
  • ●#1515 cli: decide expose/preflight's --path vs a positional path argument
  • ●#1516 action: decide whether the Action installs donkey-kit[cli] against a constraints lock
  • ●#1517 framework-release: langchain-core 1.7.1 → 1.7.3
  • ●#1523 framework-release: agent-framework 1.23.0 → 1.24.0
  • ●#1524 framework-release: langgraph 1.2.18 → 1.2.19
  • ●#1525 framework-release: openai 3.41.0 → 3.42.0
  • ●#1526 framework-release: strands-agents 1.62.0 → 1.63.0
  • ●#1533 registry: read the llm-metadata file of an Exchange llm asset
  • ●#1534 framework-release: crewai 1.15.33 → 1.15.34
  • ●#1535 framework-release: openai 3.41.0 → 3.42.1
  • ●#1542 Resolve minimatch denial-of-service alert in docs dependencies
  • ●#1547 Type the gateway's missing-model 400 as `ModelNotRoutable`, not an unconfirmed refusal
  • ●#1548 Note the Azure OpenAI streaming 502 gateway gap in the transport
  • ●#1551 Async socket fixture restores a stale pytest-socket guard
  • ●#1554 framework-release: anthropic 2.3.0 → 2.4.0
  • ●#1555 framework-release: langchain-core 1.7.3 → 1.7.4
  • ●#1562 simulator: native Anthropic and Gemini ingress routes for donkey mock
  • ●#1572 Mention the upstream LangChain listing on the LangGraph page
  • ●#1578 Fix the Python 3.10 LangGraph approval fallback test
  • ●#1581 framework-release: litellm 1.106.6 → 1.108.0
  • ●#1582 framework-release: crewai 1.15.34 → 1.15.35
  • ●#1583 framework-release: openai 3.42.1 → 3.45.0
  • ●#1601 Add protocol-version selection to A2A ingress simulation
  • ●#1605 framework-release: langchain-core 1.7.5 → 1.7.6
  • ●#1606 framework-release: langgraph 1.2.19 → 1.2.20
  • ●#1607 framework-release: openai 3.45.0 → 3.46.0
  • ●#1611 Send the negotiated A2A version on outbound calls
  • ●#1621 epic: type and document the remaining Omni Gateway MCP policy refusals (Phase 2.3)
  • ●#1622 chore(release): bump version to 0.2.1.dev0
  • ●#1627 framework-release: litellm 1.108.0 → 1.108.1
  • ●#1628 framework-release: anthropic 2.5.0 → 2.6.0
  • ●#1629 framework-release: crewai 1.15.35 → 1.15.36
  • ●#1630 refusals: type the A2A v1 Bedrock Guardrails content block as ContentSafetyBlocked
  • ●#1633 fix(registry): resolve_agent() should accept a2a_v1 instances
  • ●#1635 framework-release: google-adk 2.13.0 → 2.14.0
  • ●#1636 framework-release: litellm 1.108.1 → 1.108.2
  • ●#1637 framework-release: crewai 1.15.36 → 1.15.37
  • ●#1652 testing: add donkey.simulate_a2a() to inject named A2A refusals offline
  • ●#1653 Mention the upstream Agent Framework listing on the Agent Framework page
  • ●#1658 framework-release: litellm 1.108.2 → 1.108.3
  • ●#1659 framework-release: agent-framework 1.24.0 → 1.25.0
  • ●#1660 docs: document typed MCP Schema Validation and tool-poisoning refusals
  • ●#1661 docs: document typed MCP policy refusals on all ToolSet bindings
  • ●#1662 docs: document MCP Progressive Disclosure and Tool Mapping support
  • ●#1667 cli: keep donkey expose/preflight SERVICE working in 0.2.1
  • ●#1668 docs(release): prepare 0.2.1 docs for promotion to main
  • ●#1669 docs(roadmap): add Phase 2.1/2.2 to the Roadmap page and snapshot
  • ●#1670 refusals: type CIBA denial and kill-switch 403s on MCP binding responses
  • ●#1671 simulator: honor or refuse locked=True inside simulate_mcp()
  • ●#1672 docs(simulator): document the A2A SDK tracing gate in the simulate_a2a example
  • ●#1673 docs(api): mark ToolLockDrift provisional in the generated API reference
  • ●#1674 core: out-of-order scope exit leaves a stale entry in _scope_entry
  • ●#1675 tools: write_lock should keep donkey.lock readable
  • ●#1676 docs: small 0.2.1 contributor-doc fixes
  • ●#1687 chore(release): bump version to 0.2.1.dev1
  • ●#1691 docs(release): 0.2.1 MIGRATION misses the narrower refusal types; roadmap overstates schema/poisoning typing
  • ●#1693 Remediate security-review findings before the 0.2.1 release
  • ●#1696 chore(release): bump version to 0.2.1.dev2
  • ●#1698 Bump package version to 0.2.1

Phase 3 — Platform capabilities

Version 0.3.0 · Goal: agents that know the rules before they call, not only after they are refused.

The policy handshake (read the in-force policy set, advisory only), policy updates pushed to the code at run boundaries, a governed structured-output path, a model catalog with honest model resolution, evaluation hooks on the run span, monetary spend and wallet warnings, custom cost dimensions, and run-level cost rollups. Much of this phase depends on new gateway endpoints tracked under Upstream gaps.

0% complete · 0 done · 11 open

Open issues (11)
  • ○#19 llm/catalog: ModelHandle, resolve(), and an honest list_models() with no /models endpoint
  • ○#228 policies: the policy handshake client — read the in-force policy set
  • ○#229 policies: advisory only — the gateway always wins, and the client learns from the refusal
  • ○#230 policies: to-the-code push — refresh at run boundaries, never mid-run
  • ○#231 llm: keep the .parse() structured-output path governed, and support it in the simulator
  • ○#232 telemetry: evaluation hooks that attach a score to the run span
  • ○#261 epic: Phase 3 — Platform capabilities
  • ○#315 budget: monetary spend and the wallet threshold warning
  • ○#316 telemetry: arbitrary validated cost dimensions beyond the fixed four
  • ○#317 telemetry: a run-scoped rollup answering "what did this task cost"
  • ○#1110 scanner: publish REST APIs (OpenAPI / RAML) through the self-managed scanner

Phase 4 — Enterprise readiness

Version 0.4.0 · Goal: everything a security, compliance and platform team asks for before rolling DDK out broadly.

Independent security review and supply-chain hardening, a latency and overhead budget enforced in CI, a full pass over every error message, the public API contract and deprecation policy, compliance evidence mapping (EU AI Act Art. 12, ISO 42001, OWASP LLM Top 10), log shipping, data residency, workload identity, air-gapped operation, opt-in gateway failover, a clear boundary against the platform’s infrastructure-as-code, and a support model with a release cadence.

13% complete · 2 done · 14 open

Open issues (14)
  • ○#233 security: independent security review and supply-chain hardening
  • ○#234 perf: a latency and overhead budget enforced in CI
  • ○#235 errors: a full pass over every remediation string and error message
  • ○#236 docs: migration guide, deprecation policy, and the public API contract
  • ○#237 compliance: map correlation and span data to EU AI Act Art. 12 and ISO 42001 evidence
  • ○#238 telemetry: structured log shipping to the control plane
  • ○#239 config: data-residency and region routing hints
  • ○#240 identity: evaluate workload identity (SPIFFE / Entra Agent ID) for agent-to-gateway auth
  • ○#241 ops: support model, issue templates and release cadence
  • ○#262 epic: Phase 4 — Enterprise readiness
  • ○#318 refusals: map the typed refusal taxonomy to the OWASP LLM Top 10
  • ○#319 provisioning: state the boundary against the platform's LLM Gateway IaC
  • ○#320 ops: air-gapped and private-cloud operation, proved by an egress-denied CI job
  • ○#380 skeleton: opt-in failover to a second Connected-Mode gateway endpoint
Done (2)
  • ●#1279 preflight: read-only check that an agent's gateway infrastructure is in place
  • ●#1528 Harden MCP redirect helpers against HTTP response splitting

Phase 5 — Complete rollout

Version 1.0.0 · Goal: a stable, multi-language SDK with guarantees.

A TypeScript port of the core capabilities with conformance scenarios shared across Python and TypeScript, the remaining framework adapters brought to the full bar by demand, framework-docs partnerships and launch channels, and the 1.0 release with stability guarantees.

0% complete · 0 done · 6 open

Open issues (6)
  • ○#242 typescript: port the six-piece minimum
  • ○#243 typescript: shared conformance scenarios across Python and TypeScript
  • ○#244 adapters: bring back remaining frameworks by demand, each to the full bar
  • ○#245 gtm: framework-docs partner PRs and launch channels
  • ○#246 release: 1.0.0 with stability guarantees
  • ○#263 epic: Phase 5 — Complete rollout

Cross-cutting tracks

These milestones run alongside every phase.

VerificationIn progress

Every endpoint, header, class name and constructor argument DDK relies on is checked against the real platform and the installed framework packages before features are built on it. The semantic-caching and semantic-routing header contracts are now live-captured; next up is scoping mTLS as a consumer auth option for the LLM proxy.

68% complete · 49 done · 23 open

Open issues (23)
  • ○#1 Epic: M0 — verify every endpoint, header and class name before building on it
  • ○#79 Verify Exchange publication mechanics: metadata/tags for the digest, lifecycle states, docs pages
  • ○#88 Verify the MCP Bridge provisioning API, or confirm the Terraform/CLI fallback
  • ○#101 meta: verified-apis.md as the living source of truth, and the _verify.blocked discipline
  • ○#114 Verify Exchange draft/unlisted publication states and provenance metadata for scanner-derived assets
  • ○#123 Verify Flex Gateway Connected Mode registration from a laptop or CI runner, and the per-developer instance cost
  • ○#135 Verify PDK capabilities before any custom policy is designed (§6.9.7)
  • ○#257 verify: whether Anypoint Monitoring or Agent Visualizer ingests OTLP GenAI spans
  • ○#279 verify: whether the local simulator can be hosted on a2d-ai.com
  • ○#302 verify: upstream error envelopes for Bedrock, Gemini and Microsoft Foundry
  • ○#305 verify: federated guardrail verdict shapes for CrowdStrike and Google Armor
  • ○#607 Scope and verify mTLS as a consumer auth option for the LLM proxy
  • ○#653 verify(fastmcp): native versus protocol tool descriptors (§10)
  • ○#704 verify: model-wallet Spend (USD) budget refusal shape
  • ○#1039 verify: whether the platform's End-to-End Correlation ID uses a header other than x-correlation-id
  • ○#1101 verify: self-managed scanner is GA, and its push/commit/unstage contract
  • ○#1184 verify: capture the CrowdStrike AIDR Guardrails rejection live
  • ○#1186 verify: capture the F5 Guardrails block (HTTP 200) live
  • ○#1251 Support auto:live descriptor derivation against an MCP server URL
  • ○#1384 verify: capture the MCP policy rejection shapes live (PII, schema validation, tool poisoning, progressive disclosure, ABAC)
  • ○#1486 verify: run the framework tool-approval round trips live and flip their §8 rows
  • ○#1487 verify: capture the portal draft of a version with no page and with unpublished edits
  • ○#1546 Verify MCP Bridge transcoding through discovery and bindings
Done (49)
  • ●#21 Verify the business-group attribution header name (highest-priority M0 unknown)
  • ●#22 Verify Anypoint control-plane region hosts for EU, CA and JP Hyperforce
  • ●#23 Verify connected-app scopes and admin-context requirements per operation
  • ●#34 Verify the eight framework constructor signatures against installed packages (§8)
  • ●#53 Verify the Exchange asset search API and whether wildcard query is server-side
  • ●#54 Verify whether applied policies can be fetched in bulk per environment
  • ●#65 Verify Local Mode capabilities: can it run the LLM Proxy and MCP Bridge, and does it need a licence artifact
  • ●#253 verify: re-confirm the v1.11-v1.13 rejection contracts against current docs and a sandbox
  • ●#254 verify: the A2A proxy provisioning API, or confirm there is no supported path
  • ●#255 verify: the Trusted Agent Identity token-exchange endpoint and header
  • ●#256 verify: the Agent Kill Switch rejection shape
  • ●#258 verify: self-managed gateway image availability and licensing for local development
  • ●#300 verify: whether the gateway echoes the client x-correlation-id or mints its own
  • ●#301 verify: model-wallet spend headers and the threshold-warning signal
  • ●#303 verify: the MCP access-profile denial shape
  • ●#304 verify: the Anthropic-native Messages route on the LLM proxy
  • ●#354 docs/verify: answer #253 — the live 200 does carry budget state, as prose in x-llm-proxy-ratelimit
  • ●#372 verify: model-wallet JWT ingress auth (X-Client-Id + client_id claim, IdP/JWKS)
  • ●#520 agent_framework.chat_client() passes an unsupported model_id kwarg
  • ●#540 verify: whether the LLM proxy exposes a Gemini-native ingress route
  • ●#588 verify: the semantic-caching policy header contract on the LLM proxy
  • ●#589 verify: the semantic-routing success header (topic + score) on the LLM proxy
  • ●#636 verify(adk): installed constructor, MCP binding and descriptor APIs (§§8–10)
  • ●#637 verify(strands): installed constructor, MCP binding and descriptor APIs (§§8–10)
  • ●#638 verify(openai_agents): installed constructor, MCP binding and descriptor APIs (§§8–10)
  • ●#639 verify(anthropic): installed constructor, MCP binding and descriptor APIs (§§8–10)
  • ●#640 verify(crewai): installed constructor, MCP binding and descriptor APIs (§§8–10)
  • ●#641 verify(llamaindex): installed constructor, MCP binding and descriptor APIs (§§8–10)
  • ●#642 verify(langgraph): installed constructor, MCP binding and descriptor APIs (§§8–10)
  • ●#650 verify(local): registration prerequisite and stock-image policy probes
  • ●#652 verify(agent-framework): MCP binding and descriptor APIs (§§9–10)
  • ●#894 verification: capture /chat/completions and /responses per upstream as §2 fixtures
  • ●#1024 Verify the per-framework descriptor-derivation attributes (live introspection)
  • ●#1248 Decide auto:static descriptor derivation without a schema converter
  • ●#1278 verify: create an MCP server instance on Omni Gateway, its policies and Exchange entry
  • ●#1297 verify: resolve an Exchange MCP asset to its governed gateway endpoint
  • ●#1298 verify: read an MCP instance's governed state so registry.explain() reports real checks
  • ●#1337 docs: drop Akamai from the federated guardrails line in verified-apis
  • ●#1354 verify: read back a published A2A agent's card and its live card through the gateway
  • ●#1373 verify: as_user() on governed model calls through a JWT Validation LLM proxy
  • ●#1393 verify: capture the A2A instance policy shapes live (v1 PII, schema validation, rate limit, guardrails, quality headers)
  • ●#1625 verify: capture the A2A v1 Bedrock Guardrails content block live
  • ●#1639 verify: capture MCP Payload Optimization base64 and null-removal transforms
  • ●#1640 verify: capture MCP Payload Optimization TOON encoding
  • ●#1641 verify: capture MCP Payload Optimization LLM trimming via expected_info
  • ●#1642 verify: settle the lockfile contract under Progressive Disclosure and Tool Mapping
  • ●#1643 verify: ADK and CrewAI offline MCP injection lifecycle for simulate_mcp
  • ●#1644 docs: track the tested version of verified policies in the ledger
  • ●#1646 feat(tools): implement donkey.tools.lock() and discover(locked=True) to the #1642 contract
Upstream gapsIn progress

Requests to the Omni Gateway team for platform capabilities DDK needs — such as a budget-query endpoint, policy discovery, a dry-run mode and richer guardrail verdicts. They land whenever the gateway ships them.

7% complete · 1 done · 13 open

Open issues (13)
  • ○#247 upstream: 429 carries no Retry-After, only x-token-reset
  • ○#248 upstream: no budget-query endpoint — remaining budget is only visible in-band
  • ○#249 upstream: no policy-discovery endpoint — clients cannot ask what is in force
  • ○#250 upstream: no dry-run mode — policy cannot be evaluated without blocking
  • ○#251 upstream: guardrail-verdict error contract is under-documented
  • ○#252 upstream: no outbound-connect or tunnel path for A2A ingress
  • ○#381 upstream: no highly available or multi-region LLM proxy endpoint
  • ○#1272 upstream: the Agent Kill Switch rejection carries no kill reason
  • ○#1387 upstream: MCP Schema Validation and tool-poisoning rejections carry no policy discriminator
  • ○#1388 upstream: typed MCP policy refusals cannot reach agent code on all eight ToolSet bindings until the gateway names the policy
  • ○#1410 upstream: keep a client-visible served-provider and routing signal on llm-gw-v2 proxies
  • ○#1618 upstream: MCP PII Detector SSE rejection shape is undocumented (docs say HTTP 403, source says in-stream error event)
  • ○#1620 upstream: release mcp-pii-detector 1.0.3 (SSE streaming and elicitation decline rewrite)
Done (1)
  • ●#1401 upstream: MCP/A2A policy refusals lack a policy discriminator; A2A PII echoes raw values

What DDK will not build

At each of these boundaries the job is to make the platform’s own capability reachable and typed, not to reproduce it:

  • Client-side policy enforcement — the gateway is the enforcement point.
  • Client-side semantic caching — DDK steers the gateway’s semantic cache and reports hits, but never caches responses itself.
  • A provisioning control plane competing with API Manager or Terraform. The one thing DDK creates is the gateway instance in front of your own agent or MCP server (donkey expose); everything else stays with your platform’s infrastructure-as-code.
  • Re-implementations of Agent Scanners, Kill Switch or Trusted Agent Identity.
  • An approval UI or queue.
  • An evaluation framework.
  • The gateway inside your agent process.
  • A home-grown A2A protocol implementation — the official a2a-sdk is used.
Last updated on