Skip to content

CHANGELOG.md

Latest commit

 

History

History
230 lines (185 loc) · 14 KB

CHANGELOG.md

File metadata and controls

230 lines (185 loc) · 14 KB

Changelog

All notable changes to this project will be documented in this file. Format follows Keep a Changelog. Versions use Semantic Versioning.

[Unreleased]

Development work lands here before the next versioned release.

[0.4.0] - 2026-04-27

This release adopts the awaken crate name on crates.io. Versions 0.1–0.3 of awaken belonged to a separate, now-archived crate maintained by @brayniac, who generously transferred the name — thank you. The codebase continues the line that previously shipped as awaken-agent; the 0.4 jump exists only to skip past the prior versions already published under the awaken name. There is no 0.3 release of this codebase. Rust imports remain awaken either way.

The headline work since awaken-agent 0.2.1: streaming LLM calls now survive mid-stream failures and can resume across processes; threads carry a durable parent-child lineage with explicit child-delete strategies; provider keys and admin tokens redact themselves through RedactedString.

Added

  • New AgentEvent::ToolCallCancel and AgentEvent::StreamReset variants so consumers can drop partial deltas during stream recovery without ambiguity.
  • StreamCheckpointStore contract for cross-process stream resume (the in-process retry loop already covered same-process recovery).
  • ProviderSpec.adapter_options: a non-secret BTreeMap<String, Value> that adapters can read for things like custom headers on OpenAI-compatible proxies; unknown keys are accepted by the schema but ignored at build time.
  • Filtered thread queries (ThreadQuery/ThreadPage) and message paging (MessageQuery/MessagePage) on ThreadStore, consumed by the AG-UI and AI SDK v6 protocol routes.
  • Thread parent-child lineage: Thread.parent_thread_id, RunRequestSnapshot.parent_thread_id, and ThreadStore::delete_thread_with_strategy (reject/detach/cascade), with backend-native overrides on the file, PostgreSQL, and NATS-buffered stores.
  • AdminApiConfig.expose_config_routes toggle and ConfigRuntimeManager::with_min_apply_interval for hardening and debouncing the admin/config plane. Provider executors are reused across applies whose ProviderSpec is unchanged.
  • awaken::RedactedString re-export from the facade so secret-handling code no longer needs a direct awaken-contract dependency.

Changed

  • ProviderSpec.api_key, AdminApiConfig.bearer_token, and ServerConfig.a2a_extended_card_bearer_token are now Option<RedactedString>. JSON wire format is unchanged. Code that reads these fields must call .expose_secret(); logging that relied on Debug/Display now prints ***.
  • InferenceExecutionError is #[non_exhaustive] and splits into retryable (Provider, RateLimited, Overloaded, Timeout, StreamInterrupted), permanent (ContextOverflow, InvalidRequest, Unauthorized, ModelNotFound, ContentFiltered), and fail-fast (AllModelsUnavailable, Cancelled) classes. RateLimited and Overloaded carry an optional retry_after parsed from Retry-After. Prefer is_retryable(), counts_toward_circuit_breaker(), and retry_after() over matching specific variants.
  • Context compaction runs on a background task with single-flight semantics and swaps the summary back through the owner inbox; the synchronous compact_with_llm path is removed.

Fixed

  • NATS buffered thread store now quarantines poison WAL messages with bounded NAK and a stable hash, so an unrecoverable entry no longer stalls the WAL.
  • Background-task spawn commit failures surface through SpawnError and a metric, rather than silently dropping the spawn.
  • Self-cancel commands now cascade across runs in the same lineage.
  • Thread-query history ordering hardened across server CI surfaces.

Compatibility

  • Existing serialized config keeps loading; RedactedString serializes and deserializes as a plain JSON string.
  • The 0.4 jump is solely a name-collision skip; the Rust API additions and changes are limited to the items above.
  • 0.2.0 mailbox and server-facing types remain compatible. The NATS backend is additive behind the awaken-stores/nats feature.

[0.2.1] - 2026-04-21

Added

  • NATS-backed MailboxStore and ThreadRunStore implementations through the awaken-stores/nats feature, including JetStream dispatch wakeups, KV-backed dispatch state, buffered checkpoint WAL, and read-your-writes overlays.
  • Live mailbox steering for active runs, with durable dispatch fallback when no live subscriber acknowledges the command.
  • NATS mailbox operational metrics, delayed dispatch-signal NAK backoff, sweeper republish TTL, configurable signal-loop tuning, and stress coverage for multi-node contention and large KV buckets.

Fixed

  • Preserved 0.2.0 public API compatibility for mailbox and server-facing types while keeping new dispatch signal behavior additive.
  • Hardened distributed mailbox scheduling around thread claim guards, available_at retry windows, dispatch epoch checks, stale claim rejection, dedupe lock reconciliation, and terminal claim-field cleanup.
  • Fixed NATS buffered thread flushing so WAL messages are ACKed only after both the inner checkpoint and flushed_seq watermark write succeed.
  • Wired NATS credentials config into both mailbox and buffered thread store connections.
  • Normalized NATS KV Delete/Purge tombstones across claim guards, dedupe locks, thread indexes, epochs, and dispatch control paths.
  • Made terminal dispatch GC authoritative and changed expired-lease reclaim to use thread-claim guard records instead of scanning all dispatch records.

Changed

  • Updated genai to 0.6.0-beta.17.
  • Aligned Rust toolchain metadata with the current workspace MSRV.

Compatibility

  • cargo semver-checks passes against v0.2.0; this release is intended as a non-breaking 0.2 patch release.
  • The NATS backend is additive and behind the awaken-stores/nats feature.

[0.2.0] - 2026-04-11

Breaking Changes

  • Agent and model configuration now uses the canonical model_id -> ModelBinding { provider_id, upstream_model } chain. Legacy model, provider, model_name, and fallback_models fields are rejected in managed Awaken config.
  • A2A HTTP routes and payloads use the v1 message:send, message:stream, task wrapper, supportedInterfaces, and enum naming shapes. Older tasks/send and top-level AgentCard id/url shapes are not emitted.
  • Tool interception should move to ToolGateHook via PluginRegistrar::register_tool_gate_hook(). BeforeToolExecute now represents execution-time hooks only.
  • Remote agent endpoints use the canonical RemoteEndpoint shape: backend, base_url, auth, target, timeout_ms, and options. Legacy A2A endpoint fields are accepted only as an isolated migration input and cannot be mixed with canonical fields.

Added

  • Canonical ExecutionBackend contract for local and remote agent execution, including root execution, delegate execution, abort, remote state continuation, input/auth waits, and output capability reporting.
  • Managed configuration API for agents, models, providers, and mcp-servers, backed by ConfigStore and runtime snapshot validation.
  • Admin Console app for editing runtime configuration through the Config API.
  • SQLite mailbox store via SqliteMailboxStore and the awaken-stores/sqlite feature.
  • A2A streaming, task subscription, push notification config routes, extended agent cards, and official protocol interop coverage.
  • MCP Streamable HTTP session lifecycle, including MCP-Session-Id, strict initialize/session validation, streaming tools/call, and DELETE /v1/mcp session termination.
  • OpenUI chat example and expanded generative UI support for A2UI, JSON Render, and OpenUI Lang integrations.

Changed

  • The runtime lifecycle now includes a pure ToolGate decision point before tool execution, making permission and interception behavior independent from execution-time side effects.
  • Runtime-managed config changes compile and validate a candidate registry snapshot before it replaces the active runtime snapshot.
  • A2A remote execution preserves backend task lifecycle state across polling, streaming, interruption, cancellation, and continuation.
  • Provider/model retry and inference overrides now operate on upstream model names for the already resolved provider.
  • Documentation and examples now use awaken-agent = 0.2 dependency snippets and canonical model/provider APIs.

Compatibility

  • Existing Rust imports continue to use awaken even though the crate is published as awaken-agent.
  • awaken_runtime::extensions::a2a still re-exports compatibility aliases such as AgentBackend, AgentBackendFactory, and DelegateRunResult, but new code should use ExecutionBackend, ExecutionBackendFactory, and BackendRunResult.
  • RemoteEndpoint deserializes legacy bearer_token, agent_id, and poll_interval_ms only when no canonical fields are present. New serialized config should use auth, target, and options.

[0.1.0] - 2026-04-03

Core Runtime

  • Phase pipeline with 8 lifecycle hooks: RunStart, StepStart, BeforeInference, AfterInference, BeforeToolExecute, AfterToolExecute, StepEnd, RunEnd
  • Plugin system with typed effect handlers, scheduled actions, and commit hooks for state mutation
  • AgentResolver trait for dynamic agent resolution with composite and config-backed registry implementations
  • StopPolicy trait with built-in policies: max rounds, token budget, wall-clock timeout, consecutive errors
  • Agent handoff extension for same-thread dynamic agent switching without run termination
  • Background task extension for spawning, tracking, and cancelling long-running tasks across runs
  • Sub-agent delegation via AgentTool with local and remote (A2A) backends
  • Context compaction with configurable summarization and artifact truncation to manage long conversations
  • LLM retry policy with exponential backoff, per-model circuit breaker, and ordered fallback model list
  • AgentRuntimeBuilder for ergonomic runtime assembly with plugins, tools, and registry wiring
  • StateStore with typed slot access, snapshot, and batched mutation commits

Contract Types

  • LlmExecutor trait for provider-neutral streaming inference with genai as the default backend
  • Tool and TypedTool traits with JSON Schema descriptors and ToolCallContext
  • ToolBehaviorBundle for grouping tools and plugin references by bundle ID
  • Suspension and ResumeDecisionAction for human-in-the-loop tool approval flows
  • MailboxJob, MailboxStore, and MailboxInterrupt for persistent run queuing with lease semantics
  • AgentSpec and RemoteEndpoint types for declarative agent configuration in awaken.toml
  • AgentCard with auth descriptor for A2A discovery
  • Phase enum with is_run_level / is_step_level classification helpers
  • InferenceRequestTransform trait for pre-inference prompt mutation
  • ToolInterceptPayload for plugin-level tool call interception before execution

Protocols

  • AI SDK v6 HTTP+SSE protocol (/v1/ai-sdk/...) for Vercel AI SDK frontend integration
  • AG-UI protocol (/v1/ag-ui/...) with CopilotKit frontend support
  • Google A2A protocol (/v1/a2a/...) including agent card discovery, task send/status/cancel, and message:send
  • ACP protocol with both HTTP and stdio transports
  • MCP Streamable HTTP transport (POST /v1/mcp, GET /v1/mcp) and stdio transport for tool exposure
  • SSE reconnection replay buffer so clients can catch up on missed events after a disconnect
  • Thread and run REST API (/v1/threads, /v1/runs) with pagination, metadata patch, and cancel endpoints
  • HITL decision endpoints: POST /v1/threads/:id/decision and POST /v1/runs/:id/decision
  • Mailbox dispatch endpoint (/v1/threads/:id/mailbox) for enqueuing background jobs
  • Health check endpoint (/health) with per-component store liveness reporting

Storage

  • InMemoryStore for threads, messages, and runs — suitable for testing and ephemeral deployments
  • FileStore with crash-safe atomic writes and .checkpoint_pending recovery on restart
  • PostgresStore for production thread, message, and run persistence
  • InMemoryMailboxStore for lease-based job queuing in tests and local development
  • ProfileStore trait for per-agent profile persistence with ProfileOwner scoping

Extensions / Plugins

  • PermissionPlugin: declarative allow/deny/ask rules with glob and regex matching on tool names and arguments, firewall-style priority (Deny > Allow > Ask)
  • ObservabilityPlugin: per-inference and per-tool metrics aligned with OpenTelemetry GenAI semantic conventions; pluggable MetricsSink with in-memory, batching, persistent, composite, and OTel backends
  • DeferredToolsPlugin: lazy tool loading with ToolSearch tool for runtime capability discovery
  • SkillsPlugin: filesystem and compile-time embedded skills with markdown frontmatter, FsSkill discovery, and optional MCP bridge
  • ReminderPlugin: declarative post-tool-execution context injection rules with input/output pattern matching
  • A2uiPlugin (generative UI): streaming sub-agent UI rendering via OpenUI Lang with JSON render and data model updates
  • CompactionPlugin and ContextTransformPlugin: automatic context window management with configurable summarization LLM and artifact compaction thresholds
  • HandoffPlugin: zero-cost in-loop agent variant switching with per-variant system prompt, model, and tool overlays
  • BackgroundTaskPlugin: long-lived task lifecycle with status queries and cancellation exposed as agent tools

Developer Experience

  • awaken facade crate with full feature flag and granular opt-in features: permission, observability, mcp, skills, reminder, server, generative-ui
  • awaken::prelude re-export for ergonomic imports
  • Example: generative UI parent/sub-agent delegation pipeline
  • Example: AI SDK starter with Next.js frontend wiring
  • Example: CopilotKit AG-UI starter
  • awaken-doctest crate for compile-checked documentation examples
  • TOML-based agent configuration (awaken.toml) with JSON Schema for editor validation