marona 0.13.8

Provider-neutral AI agent runtime and MCP/Skill gateway for Marona Hub connections, managed agents, and bring-your-own-agent integrations.

Changelog

0.13.8

• Documented automatic model and provider routing together with provider-neutral lifecycle hooks.
• Verified that Responses completion hooks receive the normalized output, resolved model, and routing decision.

0.13.7

• Added one provider-neutral lifecycle hook contract across Agent, streaming, Realtime, and Responses runtimes.
• Derive requested model, framework, user, and session from trusted runtime context instead of duplicate hook configuration.

0.13.6

• Constrained A2A peer skill selection to exact IDs discovered from the Agent Card.

0.13.5

• Kept the cross-SDK release aligned with the Python 3.9 compatibility correction.

0.13.4

• Added bounded Realtime connection retries and explicit provider handshake timeouts.

0.13.3

• Resume nested Agent, A2A, and FunctionTool continuations with only the current result batch.

0.13.2

• Wait for provider acknowledgement before continuing after Realtime tool-set transitions.
• Preserve exact capability selection and automatic tool restoration across Apps, Skills, and handoffs.

0.13.1

• Keep standard and realtime Agents on the same Rust-generated live App and Skill selector contract.
• Reveal only the selected capability schemas and preserve tool_choice across governed execution.
• Translate complete MCP JSON Schemas only at restrictive realtime provider boundaries while retaining full Marona validation.
• Keep A2A, multi-agent, and generated Skill callbacks on the existing Marona governance path.

0.13.0

• Added true incremental Runner.stream(...) output over the governed runtime.
• Added provider-neutral text/audio delta, tool lifecycle, agent transition, and completed events while preserving permissions, Skills, MCP, and billing.
• Kept structured-output validation at final completion and cancellation tied to the consuming stream.

0.12.2

• Added a compact provider-neutral capability discovery and execution gateway for unscoped Hub connections.
• Preserved direct exact-tool listing for explicitly selected Apps and Skills.

0.12.1

• Added complete standard, realtime, and multi-agent A2A documentation.

0.12.0

• Added provider-neutral A2A Agent Card discovery and task execution.
• Added first-class Agent.peers across standard, streaming, and realtime runs.
• Added local credential handling, payload limits, and denied-field redaction.

0.11.1

• Preserve the owning Agent's trusted tool context when managed multi-agent runs execute local function tools.

0.11.0

• Added provider-neutral FunctionTool.create and ToolContext.
• Added local function execution across managed responses, Agent runs, streaming, realtime sessions, direct providers, and local models.
• Kept executors and trusted runtime context in the developer process.

0.10.3

• Made the realtime multi-agent documentation fully self-contained, including Marona setup, connected tools, specialist Agents, manager, and session loop.

0.10.2

• Added a complete realtime multi-agent example using the same manager graph, delegation tools, handoffs, and provider-neutral event stream.

0.10.1

• Added one provider-neutral Agent graph for delegation and handoff across standard, streaming, and realtime runs.
• Added typed input/output schemas, guardrails, lifecycle hooks, graph limits, and normalized run metadata.

0.9.3

• Unified auto, required, and none tool choice across Agent, Responses, streaming, and persistent realtime sessions.
• Kept automatic tool discovery model-led while requiring exact capability execution after discovery.

0.9.2

• RealtimeRunner.connect(...) now returns a persistent session with sendMessage, sendAudio, normalized async events, and close.
• Realtime provider events and Agent tool execution are handled inside private provider adapters.

0.9.1

• Added provider-neutral realtime capability discovery and exact MCP/Skill execution through RealtimeRunner.
• Realtime tool calls now stay in the same model session without a second model request.

0.9.0

• Unify Agent run, stream, and realtime execution around one multimodal input contract.
• Make Hub tools optional and expose completed Agent results only through result.output.
• Add capability-aware model, transport, deployment, and realtime provider routing.
• Add typed approval, permission, and service-connection interruptions.

0.8.8

• Compact tool results for local models with small context windows.

0.8.7

• Document automatic model and provider routing with Marona Agent.

0.8.6

• Preserve userId when executing governed Skills so Skill and App data use the same user scope.

0.8.5

• Add marona/auto model classification and managed provider fallback.
• Add optional cost, latency, provider, model, timeout, and fallback routing constraints.
• Expose resolved model and provider routing metadata on runtime responses.

0.8.4

• Keep service connections scoped directly to the developer-supplied user_id without legacy identity tokens.

0.8.3

• Route connected Apps through Marona governance so service authentication, permissions, and OAuth connection actions work consistently.
• Return service authorization links as standard connection artifacts for chat interfaces.

0.8.2

• Forward the Marona developer key as an MCP header only for Apps that explicitly require developer authorization.

0.8.1

• Add managed marona/gpt-5-mini and marona/gpt-5-nano model aliases.

0.8.0

• Require explicit marona/... names for Marona-funded managed inference.
• Add managed GPT, Claude, Gemini, and DeepSeek model aliases.
• Require a Marona developer key for managed, direct-provider, gateway, and local routes.
• Keep direct-provider billing with the developer while managed routes use Marona balance.

0.7.4

• Add OpenRouter as a built-in provider through openrouter/... model routes.

0.7.3

• Split image and file request documentation into complete standalone examples.

0.7.2

• Add response.output as the canonical ordered text and artifact collection.
• Keep response.outputText as a text-only convenience accessor.

0.7.1

• Exposed Skill definitions and marona.skills.publish(...) publicly.
• Added public and private Skill visibility, defaulting to private.

0.7.0

• Add Marona-owned Agent, Runner, and RunResult APIs.
• Scope optional user sessions by authenticated developer, user, and session IDs.
• Keep managed agents on the existing provider-neutral Responses runtime.

0.6.5

• Let hub.connect() connect every capability available to the developer key.
• Align the runtime package version with published package metadata.

0.6.4

• Use a real listed App and Skill in every package example.

0.6.3

• Remove identity-token parameters and synthetic identity headers from App, MCP, and Skill execution.
• Expose the developer key's Workspace, Application, usage account, and resolved capabilities in connection diagnostics.

0.6.2

• Keep valid Hub tools connected when an App or Skill selector is unresolved.
• Expose structured selector validation through connection.warnings and connection.session.

0.6.1

• Add a clearly labeled OpenAI Agents SDK example to the package documentation.

0.6.0

• Return a provider-neutral MCP connection from hub.connect().
• Add responses.create() for the Marona-managed model and governed tool loop.
• Expose listTools(), callTool(), serverUrl, and session for external agent adapters.
• Remove legacy client, message, sync, pairing, identity-token, Skill-builder, and vendor-tool exports.
• Remove global model selection; choose the model per response request.

0.5.4

• Add explicit hub.install(apps: ...) preparation for offline Apps.
• Let hub.connect(apps: null) expose all Apps installed on the device.
• Give local models one lazy discovery gateway, then only matched schemas.
• Support platform package storage through local Hub install/connect adapters.
• Add conversationId to the unified client(...) request.

0.5.3

• Keep unrelated requests independent while a local Skill approval is pending.
• Resume or cancel a pending local Skill only when the model selects the matching approval action.

0.5.2

• Let the model select lightweight Skill and App functions before Marona reveals one exact capability schema.
• Keep on-device capability routing model-led without exposing the complete MCP catalog.

0.5.1

• Preserve MCP content marked display_as_provided after local Skill reasoning.
• Keep local Skill replies consistent with the managed Edge runtime.

0.5.0

• Select managed, cloud, LiteLLM, Ollama, custom, or in-process models per request.
• Keep one App, Skill, permission, approval, and MCP runtime across providers.
• Export governed Hub tools for external agents and execute them through Marona.
• Support text, image, and document input across provider adapters.

0.4.3

• Keep local capability catalogs compact for small-context and on-device models.
• Preserve App and Skill discovery without duplicating workflow metadata.

0.4.2

• Use the same provider-neutral discovery loop for downloaded, private, and hosted models.
• Let ordinary local requests complete directly without a second route-classifier call.

0.4.1

• Execute selected Skill steps and approvals in the local model runtime.
• Reject false local success and use the normal Edge fallback in hybrid mode.

0.4.0

• Use one typed step() API for reasoning, approval, and App Skill steps.
• Replace raw Skill step maps with validated SkillStep values.

0.3.18

• Let the model answer after a local capability call and execute identical calls only once per request.

0.3.17

• Normalize multi-turn tool results across local model providers.
• Discover exact App capabilities hierarchically instead of exposing every App schema.

0.3.16

• Use one provider-neutral model runtime path for endpoint and downloaded in-process models.
• Replace discovered schemas on each discovery and hide Skill-managed MCP tools from raw App exposure.

0.3.15

• Require local and custom models to discover installed apps before claiming stored data is unavailable.

0.3.14

• Remove the unused local gateway option; installed signed targets are the only local execution source.

0.3.13

• Execute only installed signed local targets and reject server deployment metadata in offline mode.
• Keep synced private memory and conversation context available to local model execution.

0.3.12

• Require authenticated state import before routing stateful offline apps.

0.3.11

• Preserve standard top-level MCP result content in local responses.

0.3.10

• Normalize local model planning output and execute app tools across model protocols.

0.3.9

• Let hybrid local models fall back from an unavailable local app target to its remote target.

0.3.8

• Added model-driven prompt planning for local models without native function calling.
• Limited local tool schemas to capabilities selected by a Skill.

0.3.7

• Allow an explicit online metadata refresh while keeping offline inference local.

0.3.6

• Discover installed Apps and Skills lazily and load exact schemas only after selection.

0.3.5

• Added model context and output limits with adaptive local request compaction.

0.3.4

• Fit synced context and capability catalogs within small on-device model limits.

0.3.3

• Added model-driven App and Skill discovery for scalable local execution.
• Avoid sending every synced tool schema to local models on every request.

0.3.1

• Added validated Skill capability governance to prevent workflow bypasses.

0.3.0

• Added versioned Skill definitions and Hub publishing.
• Sync published Skills with Apps for normal model-led discovery.
• Preserve on-device model executors when switching runtime modes.

0.2.26

• Supply sanitized conversation context to the same local model inference on every turn.
• Remove the context-recall detour and deterministic tool-argument injection.

0.2.25

• Let the model recall prior conversation context only when the current request needs it.

0.2.24

• Let the configured model reason over all connected tools without deterministic pre-ranking.
• Keep conversation and memory context passive unless the current request depends on it.

0.2.23

• Fixed pub.dev static-analysis findings and enabled recommended package lints.
• Added package-level tests and synchronized release version metadata.
• Documented the primary public client API for generated package references.

0.2.22

• Current user intent is authoritative; conversation context can no longer select an unrelated tool.
• Added bounded recent_tool_activity for same-app follow-up identifiers across offline and hybrid modes.

0.2.21

0.2.20

• Added a standard local model executor hook for on-device runtimes.

0.2.19

• Resolve offline MCP execution targets from synced metadata using standard local gateway or localhost routes.

0.2.18

• Remove private repository metadata from the public package listing.

0.2.17

• Update the public package description.

0.2.16

• Shortened the developer role note to optional app behavior.

0.2.15

• Clarified that the optional developer role controls app behavior, not model configuration.

0.2.13

• Added developer instruction support for Marona.client(...) and message calls.

0.2.12

• Prefer provider-reported model IDs when a placeholder or stale local model ID is supplied.

0.2.11

• Added portable cache import/export for offline and hybrid clients.
• Standardized private synced context packets for local model execution.
• Redacted model provider secrets from exported cache snapshots.

0.2.10

• Added client(input: ...) for text, image, and file input blocks.
• Removed conversation IDs from the high-level client API and examples.

0.2.9

• Reserved for the first high-level client input API draft.

0.2.8

• Canonicalized app execution mode labels so Hybrid is displayed as one mode.
• Refreshed the README with a complete end-to-end usage example.

0.2.7

• Added standard execution mode helpers for Online, Hybrid, and Offline app availability labels.

0.2.6

• Compact local MCP tool results before sending them back to local models for final answer synthesis.
• Kept full structured MCP results available on response.toolCall.
• Avoided resending tool schemas on local model follow-up calls.

0.2.5

• Added safe local model and MCP response decoding for JSON, event-stream JSON, and non-JSON error bodies.
• Sent local tool results back to local models as valid JSON.
• Removed internal runtime/model implementation details from public docs.

0.2.4

• Replaced runtime route scanning with portable execution_targets.
• Added private context caching for paired offline/hybrid sync.
• Added standard offline_not_installed tool result for unresolved offline targets.

0.2.3

• Added local MCP tool execution for offline/hybrid local model calls.
• Added recent conversation context cache for offline continuity.
• Added cache-backed Hub app connection in offline/hybrid mode.

0.2.2

• Added true hybrid fallback from local model execution to Marona Runtime.
• Documented hosted runtime routing with private/local runtime resources.

0.2.1

• Documented standard interface names and custom interface slug rules.
• Added MaronaInterfaces and interfaceNamePattern.

0.2.0

• Added online, offline, and hybrid runtime modes.
• Added model registration, local cache, and response.text.
• Made Marona(...) the only public runtime client surface.

0.1.1

• Added client capability negotiation and generated UI response/event models.
• Added Hub integration client with Marona.hub.connect(...).
• Refreshed package documentation with step-by-step install and usage examples.

0.1.0

• Initial Marona-compatible runtime client package.
• Added messaging, app discovery, identity pairing, service connections, inbound events, attachments, and realtime session helpers.