fix(deps): update rust crate adk-rust to v1

[renovate]

This PR contains the following updates:

Package	Type	Update	Change
adk-rust	dependencies	major	0.9.0 → 1.0.0

---

Release Notes

zavora-ai/adk-rust (adk-rust)

v1.0.0

Compare Source

Note: 0.10.0 was an internal-only release and was never published to crates.io. All changes below were shipped as part of 1.0.0.

Breaking

• Workspace version bump to 0.10.0. This is a breaking (0.x major) release.
  All adk-* crates move from 0.9.x to 0.10.0 in lockstep.
• adk-mistralrs: Now a workspace member. Previously excluded due to git
  dependencies; now included since mistral.rs published to crates.io.
  Uses workspace rust-version (1.94.0).
• adk-core: LlmResponse and LlmRequest gained public fields. These structs
  are not #[non_exhaustive] and are constructed with struct literals downstream,
  so the additions below are breaking changes for external code that builds them
  by struct literal (use ..Default::default(), LlmResponse::new, or
  LlmRequest::new to be forward-compatible):
  • LlmResponse.interaction_id: Option<String>
  • LlmRequest.previous_response_id: Option<String>
• adk-gemini: FunctionResponse gained a public id field, and the Model
  enum (not #[non_exhaustive]) gained new variants — both breaking for external
  struct-literal / exhaustive-match consumers.

Added

• First stable release. All 39 workspace crates promoted to 1.0.0, committing to semantic versioning guarantees. This milestone marks ADK-Rust as production-ready with a stable public API.

• adk-eval: Competitive parity features — 10 new capabilities bringing the
  evaluation framework to parity with Braintrust, LangSmith, and Inspect AI:

  • StructuredJudge — typed verdicts (pass/fail/partial) with scores and
    reasoning via function-calling or JSON fallback. Lenient JSON extractor
    handles markdown fences, raw JSON, and embedded JSON in prose.
  • EmbeddingScorer (feature: embedding) — cosine similarity between
    embedding vectors using any EmbeddingProvider implementation.
  • CostTracker — token usage extraction from event streams, dollar cost
    estimation with per-model pricing tables (Gemini, OpenAI, Anthropic, DeepSeek).
  • TraceAnalyzer — detects redundant tool calls and execution loops,
    computes efficiency score (useful_calls / total_calls).
  • BaselineStore — save/load metric snapshots as .eval-baseline.json,
    detect regressions with configurable tolerance.
  • JunitReporter (feature: ci-helpers) — JUnit XML generation for native
    CI integration (GitHub Actions, Jenkins, GitLab CI).
  • AnnotationStore — JSONL export/import for human review workflows with
    case_id validation and unmatched entry warnings.
  • AbComparator (feature: statistics) — A/B agent comparison using Wilcoxon
    signed-rank test for paired statistical significance.
  • TestGenerator — LLM-driven case generation from descriptions, plus direct
    extraction from production event logs (no LLM needed).
  • ConversationScorer — multi-turn metrics: context retention, goal completion,
    coherence, and topic drift (via StructuredJudge or EmbeddingScorer).
  • CLI integration — cargo adk eval subcommand with --save-baseline,
    --check-regression, --format (table/json/junit), --concurrency, and
    non-zero exit on regression detection.
  • EvaluationResult extended — optional cost_metrics, trace_analysis, and
    verdicts fields (backward-compatible with #[serde(default)]).
  • EvalCase extended — optional metadata field for generation tracking.
  • New feature flags: embedding, ci-helpers, statistics.
  • Example crate: examples/eval_showcase/ demonstrating all features.

• adk-enterprise — Native Rust SDK for the ADK-Rust Enterprise Managed Agent Service. Lightweight HTTP/SSE client with zero adk-* runtime dependencies. Supports any model (Gemini, OpenAI, Anthropic, DeepSeek, Ollama), auto-reconnect SSE streaming, automatic retry with exponential backoff, idempotency keys, and self-hosted deployments. (Experimental)

• adk-graph: Functional API (feature: functional) — Write agent workflows as
  normal async Rust functions with automatic checkpointing, typed state reducers,
  and interrupt/resume support. Includes:

  • #[entrypoint] and #[task] proc macros in adk-rust-macros
  • TaskContext — runtime context with state, checkpointing, streaming, interrupts
  • ReducedValue<T> — append-only state container persisted across checkpoints
  • UntrackedValue<T> — transient state container excluded from checkpoints
  • MessagesValue — chat messages with ID-based deduplication
  • TypedReducer trait with built-in reducers (Replace, Append, Merge)
  • StateSchemaValidator — type-level validation at workflow boundaries
  • ExecutionLog — task completion tracking for resume-skip behavior
  • Loop iteration checkpoint keying ("task::iter_N")
  • 3 example crates: functional_workflow, background_runs, cron_scheduling

• adk-server: Background Runs (feature: background) — REST API for async
  workflow execution. POST /runs submits a workflow, GET /runs/{id} polls status,
  DELETE /runs/{id} cancels. Status transitions: queued → running → completed/failed/cancelled.
  Timeout enforcement, retry with checkpoint resume, BackgroundRunner orchestrator.

• adk-server: Cron Scheduling (feature: background) — Cron job management
  with REST endpoints (POST/GET/PATCH/DELETE /cron). Supports 5-field and 6-field
  cron expressions, concurrency policies (skip/allow/queue), background scheduling
  loop, pause/resume lifecycle management.

• adk-model: cancel_response() for OpenAI background responses — new method
  on OpenAIResponsesClient that calls POST /v1/responses/{id}/cancel to cancel
  a running background response. Returns an LlmResponse with cancelled status.
  Useful for deep research and other long-running background requests.

• adk-model: Six new OpenAI Responses API example crates — standalone examples
  demonstrating WebSocket transport (openai_ws_minimal), background mode
  (openai_background), Conversations API (openai_conversations), built-in
  tools (openai_builtin_tools), deep research (openai_deep_research), and
  Open Responses compatibility (openai_open_responses).

• adk-anthropic: Claude Opus 4.8 support — added claude-opus-4-8 as the
  latest flagship model. Adaptive thinking only (same as Opus 4.7).

• adk-managed: Managed Agent Runtime (feature: managed-runtime) — New crate
  providing a provider-neutral, durable, resumable agent execution engine. The
  runtime takes a declarative ManagedAgentDef, builds a runnable agent, and
  operates it as a checkpoint-resumable, event-streaming background session.
  Key capabilities:

  • ManagedAgentRuntime trait and DefaultManagedAgentRuntime implementation
  • Durable sessions: checkpoint after every event, resume from last consistent state
  • Provider-neutral event stream (SessionEvent) — identical sequences across
    Gemini, OpenAI, Anthropic, Ollama, and OpenAI-compatible providers
  • Custom tool parking with configurable timeout
  • Event replay (stream_events(from_seq)) for SSE reconnection
  • ScriptedLlm test double for deterministic offline testing ($0)
  • Golden fixture conformance tests (F-1 through F-8)
  • STABILITY: Experimental, additive, feature-gated behind managed-runtime

• adk-anthropic: Managed Agents API client — full implementation of the
  Anthropic Managed Agents API, feature-gated behind managed-agents. Includes:

  • Agent, Environment, Session CRUD with SSE streaming
  • Custom tool flow, tool confirmation (allow/deny)
  • Vaults and credentials for MCP authentication
  • Memory stores with versioning and optimistic concurrency
  • Dreams API for memory curation (Research Preview)
  • Webhook signature verification (HMAC-SHA256)
  • Multiagent orchestration with session threads
  • Self-hosted environment work queue management
  • File upload and session resource mounting
  • 5 examples: hello, custom_tools, files, memory, multiagent
  • 47 integration tests passing against live API

• adk-anthropic: Files API client — upload, download, list, get, delete files
  for use with Messages API. Feature-gated behind files. Auto MIME type inference.

• adk-mistralrs: Now publishable to crates.io — switched from git dependency
  (mistralrs = { git = "..." }) to crates.io (mistralrs = "0.8"). The crate
  is now a full workspace member and can be published alongside other ADK crates.
  Supports 50+ model architectures including Gemma 4, Qwen 3.5, Llama 4, Voxtral,
  GPT-OSS, and multimodal models with text/image/audio/video input.

• cargo-adk: managed-agents template — cargo adk new my-agent --template managed-agents
  scaffolds a complete Anthropic Managed Agents project with --provider support
  for future multi-provider managed agents.

• adk-model: Gemini OpenAI-compatible preset — OpenAICompatibleConfig::gemini(api_key, model)
  targets Gemini's OpenAI-compatibility endpoint
  (https://generativelanguage.googleapis.com/v1beta/openai), letting callers on
  the openai feature use a GEMINI_API_KEY and a Gemini model through the
  OpenAI Chat Completions wire format (chat, streaming, function calling,
  structured output, reasoning effort). Two examples added:
  gemini_openai_compat (direct client) and gemini_openai_compat_agent
  (the compat client driving a normal LlmAgent/Runner).

• adk-gemini: Interactions API (Beta) — first-class support for Google's new
  Interactions API, the forward direction for the Gemini API. Gated behind the
  new interactions feature flag (no new dependencies; additive to the existing
  generateContent surface). New adk_gemini::interactions module with:

  • Gemini::create_interaction() fluent builder — single-turn, streaming
    (step.delta events), multimodal input, tools, structured output, and
    server-side multi-turn via previous_interaction_id.
  • Gemini::get_interaction(), delete_interaction(), cancel_interaction()
    for the stored-interaction lifecycle.
  • Typed Step timeline (user input, model output, thought, function call,
    function result) with forward-compatible Step::Other for server-side tool
    steps; polymorphic Content, Tool, and ResponseFormat; Usage,
    InteractionStatus, and SSE InteractionSseEvent/StepDelta types.
  • Interaction::output_text() and pending_function_calls() convenience
    accessors mirroring the official SDKs.
  • Requests pin the Api-Revision: 2026-05-20 steps-schema contract.
  • Example: cargo run -p adk-gemini --features interactions --example interactions_basic.

• adk-gemini: Managed Agents & Environments — extends the Interactions API
  with Google's Managed Agents (Antigravity, Deep Research) and sandbox
  Environments. All code behind the existing interactions feature flag:

  • InteractionBuilder::antigravity() — one-call setup for the Antigravity
    coding agent (agent = "antigravity-preview-05-2026", store = true).
  • InteractionBuilder::deep_research(agent_id) — one-call setup for Deep
    Research agents (background = true, store = true).
  • InteractionBuilder::environment(env) — attach a sandbox (fresh "remote",
    resume by ID, or inline EnvironmentConfig with sources + network rules).
  • InteractionBuilder::agent_config(config) — attach agent-specific config
    (Deep Research thinking summaries, visualization, collaborative planning).
  • Environment, EnvironmentConfig, EnvironmentSource (Inline, Repository,
    Gcs), NetworkConfig (Disabled, Allowlist), NetworkRule, TransformMap
    wire types with full serde round-trip and #[non_exhaustive].
  • AgentConfig enum (DeepResearch + forward-compatible Other variant).
  • Client-side validation: model/agent mutual exclusivity, Antigravity
    constraints (no background, no unsupported gen params, no function tools, no
    audio/video/document), Deep Research constraints (background required).
  • Managed Agent CRUD: Gemini::create_agent() (fluent builder),
    list_agents(), get_agent(id), delete_agent(id).
  • Gemini::download_environment(env_id) — download sandbox snapshot as tar.
  • Interaction.environment_id response field for sandbox resume.
  • TransformMap custom Debug redacts credential values in logs.
  • Example: cargo run -p gemini-managed-agents (4 practical agents with
    streaming, auto-cancellation, and live progress display).

• adk-model / adk-rust: gemini-interactions feature — forwards the
  Interactions API surface up the stack. adk_model::gemini::interactions
  re-exports the module when enabled.

• adk-model: Gemini Interactions API runtime transport (Beta) — a transport
  toggle on GeminiModel routes the standard LlmAgent/Runner/tool loop
  through Google's Interactions API instead of generateContent, with no new
  agent type or rewritten agent setup. Gated behind the gemini-interactions
  feature; generateContent remains the default and recommended path. Highlights:

  • GeminiModel builder use_interactions_api(true) selects the transport;
    interaction_options(...) configures store, stateful continuation,
    background mode, and poll interval (InteractionOptions/BackgroundMode).
  • Faithful API defaults: store=true, server-side stateful continuation (via
    previous_interaction_id), and background=true for agent targets only.
  • Restricted target allowlist (InteractionTarget) with InvalidInput errors
    naming the supported models/agents for unsupported targets.
  • bypass_multi_tools_limit: built-in tools (Google Search, URL context, File
    Search) can be converted to function-calling tools
    (with_bypass_multi_tools_limit / the BypassMultiToolsLimit trait) so they
    coexist with custom function tools under the Interactions API.
  • Streaming (SSE) and background-poll completion surfaced through the same
    response stream the runner consumes.
  • Example: examples/gemini_interactions_agent/.

• adk-core: provider-neutral interaction continuity fields — new fields that
  carry conversation continuity across providers (see Breaking above):

  • LlmResponse.interaction_id: Option<String> plus an Event::interaction_id()
    accessor (mirrors ADK-Python's event.interaction_id).
  • LlmRequest.previous_response_id: Option<String>, populated by LlmAgent
    from the most recent event's interaction_id; the Gemini Interactions
    transport maps it to previous_interaction_id, falling back to transcript
    input transparently when a stored interaction has expired.

• adk-gemini: May 2026 GA models — Added Model variants for models that
  shipped or replaced previews in May 2026:

  • Gemini31FlashLite (gemini-3.1-flash-lite, GA) — replaces the preview,
    which was shut down May 25, 2026.
  • Gemini31FlashImage (gemini-3.1-flash-image, Nano Banana 2, GA).
  • Gemini3ProImage (gemini-3-pro-image, Nano Banana Pro, GA).
  • GeminiEmbedding2 (gemini-embedding-2, GA multimodal embeddings).

• adk-gemini: pricing for Gemini 3.5 Flash — GeminiPricing::GEMINI_35_FLASH
  ($1.50/MTok input, $9.00/MTok output incl. thinking).

• adk-gemini: GeminiPricing::for_model(&Model) — maps a Model to its
  standard per-token pricing, returning None for Custom models.

• adk-gemini: FunctionResponse.id + FunctionResponse::with_id() — echo the
  originating FunctionCall id to satisfy Gemini 3.x strict response matching
  (id + name + count). adk-model's Gemini provider now forwards the call id
  automatically.

Deprecated

• adk-gemini: Model::Gemini31FlashLitePreview — shut down May 25, 2026;
  use Model::Gemini31FlashLite.
• adk-gemini: Model::Gemini3ProImagePreview — shuts down June 25, 2026;
  use Model::Gemini3ProImage.

Changed

• adk-gemini: ThinkingLevel docs — clarified that Medium is the default
  for Gemini 3.5 Flash while High remains the default for Gemini 3 Flash
  Preview and Gemini 3.1 Pro, and that temperature/top_p/top_k are no
  longer recommended for Gemini 3.x.

---

Configuration

📅 Schedule: (UTC)

• Branch creation
  • At any time (no schedule defined)
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.

---

• If you want to rebase/retry this PR, check this box

---

This PR was generated by Mend Renovate. View the repository job log.