docs: correct three unverified capability claims (#78)

The README comparison table and feature list described behaviour the node
does not have.

Session length "up to 8 hours" cited limits belonging to other resources.
LifecycleConfiguration.maxLifetime (default 28800s) is the AgentCore Runtime
instance lifetime, and the 28800s ceiling on
SessionConfiguration.sessionTimeoutInSeconds is for MCP Gateway sessions. The
only duration this node sends is the per-invocation timeoutSeconds, default
600, so no user could reach eight hours through it.

Streaming was listed as shipped. consumeStream() accumulates the event stream
and returns a completed result, and the node never calls n8n's sendChunk, so
nothing reaches the streaming UI. Renamed to "Structured responses" and moved
streaming to the later roadmap row.

"Firecracker microVM per session" named a hypervisor AWS public documentation
does not name. The table and docs/SPEC.md now say "isolated microVM", matching
the AgentCore docs.

Same class of error as the AI Agent comparison corrected in #77.

Co-authored-by: Sundar Raghavan <sdraghav@amazon.com>

Author: sundargthb

CHANGELOG.md

@@ -117,6 +117,21 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **npm package description** now says the node is verified, and `verified` was
   added to `keywords`. npm always publishes `README.md` regardless of the `files`
   field, so this release is what surfaces the verified status on the npm page.
+- **Three capability claims corrected in the README comparison table and feature
+  list.** All three described behaviour the node does not have. (1) "Session length:
+  up to 8 hours" cited a limit that belongs to other resources: `maxLifetime` on
+  `LifecycleConfiguration` is the AgentCore *Runtime* instance lifetime, and the
+  28800-second ceiling on `SessionConfiguration.sessionTimeoutInSeconds` is for MCP
+  *Gateway* sessions. The only duration this node sends is the per-invocation
+  `timeoutSeconds` (default 600), so no user could reach eight hours through it.
+  (2) "Streaming responses" was listed as a shipped feature, but `consumeStream()`
+  accumulates the event stream and returns a completed result; the node never calls
+  n8n's `sendChunk`, so nothing streams to the UI. Renamed to "Structured
+  responses", which is what it delivers, and moved streaming to the `later` roadmap
+  row. (3) "Firecracker microVM per session" named a hypervisor that AWS public
+  documentation does not, so the table and `docs/SPEC.md` now say "isolated microVM"
+  as the AgentCore docs do. Follows the same class of error as the AI Agent
+  comparison corrected in the previous entry.
 
 ## [0.4.0] - 2026-07-27
 

README.md

@@ -19,8 +19,8 @@ n8n’s AI Agent node covers a lot of ground, and for many workflows it is all y
 |Memory        |Simple Memory in process; Postgres, Redis and other stores available as sub-nodes you operate|Managed store provisioned for you, scoped per actor, with semantic, summary and user-preference strategies|
 |Code execution|Code node runs JS or Python in the n8n process  |Python, Node and shell inside an isolated microVM|
 |Browser       |Community nodes such as Puppeteer, self-operated |Managed cloud browser|
-|Session length|Bounded by your execution timeout               |Long-running sessions, up to 8 hours      |
-|Isolation     |Shared n8n process                              |Firecracker microVM per session           |
+|Session length|Bounded by your execution timeout               |Long-running sessions, with a per-invocation timeout you set|
+|Isolation     |Shared n8n process                              |Isolated microVM per session              |
 
 The distinction is not what n8n can do, it is what you have to run and secure
 yourself to do it.
@@ -36,7 +36,7 @@ yourself to do it.
 - **VPC, custom containers, and filesystem mounts** (v0.2) — run in your VPC, bring a linux/arm64 ECR image, mount session storage / EFS / S3 Files
 - **OAuth Bearer invoke** (v0.2) — invoke inbound-OAuth harnesses with a JWT stored on the AgentCore API credential
 - **Versions & endpoints** (v0.2) — list immutable versions, pin named endpoints, invoke by qualifier
-- **Streaming responses** with structured tool-use trace, token usage, and latency metadata
+- **Structured responses** — tool-use trace, stop reason, token usage, and latency metadata on every invocation (the node consumes the AgentCore event stream and returns the completed result; it does not yet forward partial output to n8n's streaming UI)
 - **Session persistence** — pass the same session ID across executions for multi-turn conversations
 - **Execution limits** — max iterations, max tokens, timeout
 
@@ -472,10 +472,10 @@ zero runtime dependencies, and **is verified**. It is listed at
 
 |Version           |Capability                                                                                       |
 |------------------|-------------------------------------------------------------------------------------------------|
-|**v0.1**          |Run Agent, Invoke Existing, MCP / Browser / Code Interpreter / Gateway tools, streaming, sessions|
+|**v0.1**          |Run Agent, Invoke Existing, MCP / Browser / Code Interpreter / Gateway tools, sessions|
 |**v0.2**          |Multi-provider models, managed memory, VPC, custom containers, filesystem mounts, skills, inline functions, versions & endpoints|
 |**v0.4** (current)|OAuth Bearer Token moved to the credential, Add Tools / Add Skills toggles, AWS calls through n8n's HTTP helper, node `typeVersion` 2|
-|**later**         |ExecuteCommand (shell) with Bearer, custom Browser/Code Interpreter resources, CloudFormation quick-create, Export to Code, Step Functions|
+|**later**         |Streaming partial output to n8n's streaming UI, ExecuteCommand (shell) with Bearer, custom Browser/Code Interpreter resources, CloudFormation quick-create, Export to Code, Step Functions|
 
 ## Limitations (v0.2)
 

docs/SPEC.md

@@ -14,7 +14,7 @@
 
 `@aws/n8n-nodes-agentcore` is an n8n community node that integrates Amazon Bedrock AgentCore harness into n8n workflows. It lets workflow developers run production-grade AI agents — with cross-session memory, real cloud browser, real microVM code execution, multi-provider model choice, skills, and long-running sessions — without writing infrastructure code or agent code.
 
-The node auto-provisions a harness on first execution, reuses it on subsequent runs, and updates it when configuration drifts. AWS credentials are read from n8n’s encrypted credential vault per execution, never persisted. All risky execution (model inference, browser, code interpreter) happens inside AWS-managed Firecracker microVMs, not inside the n8n process.
+The node auto-provisions a harness on first execution, reuses it on subsequent runs, and updates it when configuration drifts. AWS credentials are read from n8n’s encrypted credential vault per execution, never persisted. All risky execution (model inference, browser, code interpreter) happens inside AWS-managed isolated microVMs, not inside the n8n process.
 
 ## 2. Goals and non-goals
 
@@ -86,7 +86,7 @@ Three audiences, in priority order:
 │  │ - ListHarnesses      │                   ▼                   │
 │  │ - DeleteHarness      │     ┌─────────────────────────────┐   │
 │  └──────────┬───────────┘     │  Harness microVM            │   │
-│             │                 │  (Firecracker, isolated)    │   │
+│             │                 │  (isolated per session)     │   │
 │             │ assumes         │                             │   │
 │             ▼                 │  - Bedrock model inference  │   │
 │  ┌──────────────────────┐     │  - Browser tool             │   │