ADK Skill v2.2.0 — ADK Python 2.13.0

Updates the skill from ADK Python 2.7.1 coverage to 2.13.0. APIs and package metadata were checked against official versioned sources on 2026-10-25. Release tag: v2.2.0.

Compatibility and migration

• Python examples require Python 3.11+. Model credentials are required only for model calls; deterministic function-only workflows can run offline.
• Verified reference SDKs: Python 2.13.0, Java 1.13.0, Go 2.6.0, Kotlin 1.4.0, TypeScript 2.3.0. Java and Kotlin graph support is experimental; Kotlin's core is generally available.
• Go 2.6 examples use Go 1.26.6 and the /v2 module path. Kotlin examples compile with Kotlin 2.3.21; 2.1.20 fails on transitive dependencies with 2.3 metadata.
• Document failed-node reexecution on resume, idempotency requirements, cancellation, node/tool confirmation, authentication resumption and nested task delegation. Add a persistent SQLite session setup with the [db] extra.
• Document OpenTelemetry's otel_semconv_1_44 default, removal of call_llm, and the temporary legacy-schema override.

Added and updated

• Stable SkillToolset guidance with experimental, opt-in lifecycle modes, discovery, load/unload APIs and script output artifacts.
• MCP SDK 2.x compatibility and its separate modern-protocol opt-in; native A2A task delegation, authentication and version-specific Agent Cards.
• Experimental model failover, bounded advisor consultation, classifier routing and explicit application error recovery.
• SQLite memory, experimental MongoDB sessions/memory and backend-specific memory mutation support.
• OpenAI integration imports, LiveKit/CascadeLive, deployment identity and informational efficiency metrics.
• Native SDK examples, toolchains and capability boundaries; a maintained offline Python suite and CI gate for documented examples.

Corrected

• Await artifact save/load/list operations and validate their results.
• Use AuthenticatedFunctionTool for authenticated functions, valid callback return types and actual SDK constructors/imports.
• Replace unrestricted eval with validated arithmetic inputs.
• Correct state interpolation, response-event extraction, asynchronous waiting, graph validation rules and task completion naming.
• Replace exception-blind sequential fallback guidance with explicit recovery and model failover contracts.
• Narrow activation to ADK work; retain details in references and keep the entrypoint below 500 lines.

Verification

• 23 offline Python tests passed, including routing, joins, dynamic dispatch, approval/rejection, persistent-session restart, abort, artifacts, memory isolation, skill lifecycle, A2A startup/cards, callbacks and failover.
• All documented Python fences parse; 12 evaluation JSON examples validate against ADK schemas; 29 versioned source links resolve.
• Native checks passed: Go example with Go 1.26.6; six TypeScript examples with TypeScript 7.0.2; five Java examples with JDK 17.0.21; three Kotlin source examples with Kotlin 2.3.21.
• An independent cross-process SQLite check confirmed that approval increments the acceptance counter and rejection leaves it unchanged after restart.
• SkillEvaluator: 6/6 checks passed, A — 91.2/100. SkillSpector: SAFE, zero open findings, with 12 individually reviewed documentation suppressions and no wildcard rules.

Test limits

Cloud deployment, real model conversations/evaluations, Live audio, Android and KSP generation were not exercised. Native compilation was local; CI runs Python checks. The historical effectiveness benchmark was not rerun. Offline tests do not establish model availability or production behavior.

The bundled skill-creator quick validator rejects the supported compatibility frontmatter field; repository schema and structural gates validate it. The field is retained.

See README.md for repeatable verification commands and tasks.md for phase-level evidence. Official upstream baseline: ADK Python 2.13.0.

Merged pull request: #4.

Release pipeline passed: validation and publication.