docs: align release surfaces to 0.8.0 + main → 0.8.1-SNAPSHOT

CHANGELOG.md

@@ -4,6 +4,22 @@ All notable changes to Agents.KT are documented here. The format follows [Keep a
 
 ## [Unreleased]
 
+## [0.8.0] — 2026-06-15
+
+**Interoperable, multimodal agents — with capability grants.** The largest minor since 0.5.0:
+agent-to-agent interop (**A2A v1**), full **multimodal** (audio STT/TTS, vision, image generation),
+a **RAG** seam, richer **composition** (`handoff` / `firstOf` / `.speculative` / `loopUntil` /
+built-in aggregators / forum captains), **human-in-the-loop** gates, an **eval** harness, history
+compression, an **eighth model provider (Google Gemini)**, **agent.json** definition serialization,
+and the **capability-grants** DSL (`grants { allow / confirm }`). Plus the planning groundwork for
+the agentic-web standards (AGNTCY / AG-UI / x402 / NLWeb — PRD §12.6–§12.9). Additive: existing
+public API surfaces are preserved.
+
+*Deferred to 0.9.0:* the remaining Layer-2 **sandbox backends** — `DockerSandbox` (#2895), the
+network hostname-allowlist **proxy** (#2893), and **read confinement** (#4546). `WasmSandbox` (#2894)
+was closed won't-do; the rational WASM direction (agent → WASM export, #4547) is a separate
+forward-looking track.
+
 ### Added — Google Gemini provider adapter (#1917)
 
 Eighth built-in `ModelClient`: `model { gemini("gemini-2.5-flash"); apiKey = ... }` for Google's

README.md

@@ -37,7 +37,7 @@ The 0.6–0.7 line turns those boundaries into reviewable evidence: deterministi
 ```kotlin
 // build.gradle.kts
 dependencies {
-    implementation("ai.deep-code:agents-kt:0.7.24")
+    implementation("ai.deep-code:agents-kt:0.8.0")
 }
 ```
 
@@ -262,7 +262,7 @@ What the framework does **not** enforce — your responsibility:
 - **Seven LLM providers shipped** — Ollama, Anthropic, OpenAI, DeepSeek, Kimi (Moonshot AI, #2697), OpenRouter (#2701), and Perplexity (Sonar, #3675) — the last with a `perplexitySearch` web-grounded search tool (#3676 / #3677). Google (Gemini) is the main adapter still on the roadmap (Phase 2); the injectable `ModelClient` covers test stubs and your own adapters in the meantime.
 - **Synchronous agentic loop** — `runBlocking` inside the loop until the suspend refactor lands (#638). Calling agents from existing coroutine scopes works but doesn't propagate cancellation cleanly.
 - **No built-in MCP rate limiter** — use `McpServer` auth/policy plus a gateway for throttling. Agent/runtime audit events have a first-party JSONL exporter in `:agents-kt-observability`.
-- **Streaming runtime** *(shipped — v0.5.0)*. `agent.session(input): AgentSession<OUT>` exposes `events: Flow<AgentEvent<OUT>>` — bracket events (`SkillStarted` / `SkillCompleted` / `Completed<OUT>` / `Failed`) plus mid-loop `Token` / `Reasoning` / `ToolCallStarted` / `ToolCallArgumentsDelta` / `ToolCallFinished` events as the agentic loop runs. All events carry `requestId`, `sessionId`, and `manifestHash` for audit correlation (#1913). All seven providers stream at the wire — Ollama (NDJSON), Anthropic and OpenAI (native SSE), with DeepSeek / Kimi / OpenRouter / Perplexity inheriting the OpenAI-compatible SSE path; live integration tests measure 19 / 2 / 19 chunks for the original three native adapters. `SkillCompleted.tokensUsed` and `Completed.tokensUsed` carry cumulative `TokenUsage` across all turns. The underlying `LlmChunk` sealed type + `ModelClient.chatStream(messages): Flow<LlmChunk>` foundation (#1722) is what custom adapters plug into. See [docs/streaming.md](docs/streaming.md) for the full API + the [v0.5.0 streaming premortem](docs/premortem-0.5.0-streaming.md) for design rationale.
+- **Streaming runtime** *(shipped — v0.5.0)*. `agent.session(input): AgentSession<OUT>` exposes `events: Flow<AgentEvent<OUT>>` — bracket events (`SkillStarted` / `SkillCompleted` / `Completed<OUT>` / `Failed`) plus mid-loop `Token` / `Reasoning` / `ToolCallStarted` / `ToolCallArgumentsDelta` / `ToolCallFinished` events as the agentic loop runs. All events carry `requestId`, `sessionId`, and `manifestHash` for audit correlation (#1913). All eight providers stream at the wire — Ollama (NDJSON), Anthropic, OpenAI, and Gemini (native SSE), with DeepSeek / Kimi / OpenRouter / Perplexity inheriting the OpenAI-compatible SSE path; live integration tests measure 19 / 2 / 19 chunks for the original three native adapters. `SkillCompleted.tokensUsed` and `Completed.tokensUsed` carry cumulative `TokenUsage` across all turns. The underlying `LlmChunk` sealed type + `ModelClient.chatStream(messages): Flow<LlmChunk>` foundation (#1722) is what custom adapters plug into. See [docs/streaming.md](docs/streaming.md) for the full API + the [v0.5.0 streaming premortem](docs/premortem-0.5.0-streaming.md) for design rationale.
   - *Partial cancellation today.* `Flow` collection cancels promptly, and `perToolTimeout` now applies to both regular and session-aware tool calls. Synchronous skill bodies and blocking HTTP reads still are not fully coroutine-cancellable mid-call; the remaining adapter migration is the `sendAsync`/suspend-refactor track.
   - *Composition flow-through shipped (#3866).* Every composition operator exposes `session(...)`, and every `then` overload chains streaming — pipelines mixing `Parallel` / `Forum` / `Loop` / `Branch` mid-chain stream all nested agents' events through the parent session, demultiplexable by `agentId`. Remaining: pipeline-stage event types (`StageStarted` / `PipelineCompleted`).
 - **No native binary** — JVM-only (≥ JDK 21). GraalVM and `jlink` bundles are Phase 2 priorities.
@@ -318,7 +318,7 @@ Topical guides:
 
 ## Current Release
 
-The latest published release is `0.7.24` — **Perplexity + truth surface.** (Between releases, `main` carries unreleased work under a `-SNAPSHOT` version; see the [CHANGELOG](CHANGELOG.md) *Unreleased* section for what's landed since.) Perplexity becomes the **seventh model provider** (`model { perplexity("sonar") }`, #3675) and any agent on any model can register the **`perplexitySearch`** tool for live, web-grounded answers with citations that reach both the model context and the JSONL audit row (#3676/#3677) — `untrustedOutput = true`, so results arrive flagged as data, not instructions. The release also lands the **truth-surface pass** an external 0.7.23 review called for: SECURITY.md / production-hardening / skill-routing / HITL docs now match the shipped runtime, `main` adopts a `-SNAPSHOT` between-releases version policy, and a new `DocsConsistencyTest` pins provider counts, `Decision` variants, and routing claims to the code. Dependency line: Kotlin 2.4.0, jline 4, detekt 1.23.8, ksp 2.3.9. Additive only — drop-in on the 0.7.x line.
+The latest published release is `0.8.0` — **interoperable, multimodal agents, with capability grants** (the largest minor since 0.5.0). (Between releases, `main` carries unreleased work under a `-SNAPSHOT` version; see the [CHANGELOG](CHANGELOG.md) for what's landed since.) Highlights: **A2A v1** (agents are A2A servers + typed clients), full **multimodal** (vision across Claude/OpenAI/Ollama, audio STT/TTS tools with self-hosted Whisper/Qwen, image generation), an **eighth model provider — Google Gemini** (`model { gemini("gemini-2.5-flash") }`, a full from-scratch adapter with native SSE, function calling, `responseJsonSchema` decoding, and thought-summary reasoning), **capability grants** (`grants { allow(...); confirm(...) }` — `confirm` tools need the granting agent's authorization, fail-closed), **agent.json** serialization, the **RAG** seam, and richer **composition** (`handoff` / `firstOf` / `.speculative` / `loopUntil` / aggregators / forum captains) plus HITL gates and an eval harness. The Layer-2 **sandbox backends** (Docker / egress proxy / read confinement) move to **0.9.0**. Additive only — drop-in on the 0.x line. Dependency line: Kotlin 2.4.0.
 
 **0.7.23 — maintainability + a model-error policy.** A behavior-preserving, drop-in release that closes the bulk of the code-smell remediation epic (#2790) and finishes the **AgenticLoop decomposition** begun in 0.7.21: the `Agent` god class splits into `InterceptorChain` + `ListenerRegistry` (#2793); `McpServer` into HTTP intake + a transport-agnostic `McpDispatcher` (#2795); the five composition operators' duplicated streaming-session scaffold collapses into one `agentSessionScope` (#2797); `LiveShow`'s banner + thread-shadowing spinner become their own units (#2798); `Forum`/`Branch` lose their dual-path duplication (#2802); a typed `GenerableCodec` seam collapses the `@Generable` casts to one boundary (#2803); and `executeAgentic`'s last setup block extracts as `resolveAllowedTools` (#3423). The one **new public API** is **`onLLMError`** (#3508): when a model is configured, a failed model call in the agentic loop fails fast and loud by default, with `onLLMError { e -> RespondWith(fallback) | Rethrow }` as the opt-in recovery hook. The detekt-baseline ratchet fell 423 → 415 and `@Suppress("UNCHECKED_CAST")` 42 → 30 across the release. Only #2791 (the turn-loop core of `executeAgentic`) remains open in the epic.