docs: correct what eve extensions can contribute (channels/schedules/subagents ARE allowed)

CLAUDE.md

@@ -155,8 +155,13 @@ and `eve-extension-demo` (a minimal eve scaffold that mounts the extension).
   enable `chatHistory` and `disableTool()` the two history slots). Static memory tools are importable from
   `@upstash/agentkit-eve-extension/tools` for `toolResultFrom`/overrides; the dynamic search and
   chat-history tools are not.
-- What an extension **cannot** contribute (stays in `@upstash/agentkit-eve`): sandbox, channels/auth
-  (rate limiting), schedules, agent config. `defineCachedTool` also stays there (wraps user tools).
+- What an extension **can** contribute (eve ≥0.41, per eve's `docs/extensions.md`): tools, channels,
+  connections, skills, schedules, subagents, instruction fragments, hooks — channels, schedules and
+  subagents **are** allowed, and a contributed subagent may own its own agent config and sandbox.
+  What the extension **root** cannot declare: agent configuration, a sandbox, or nested extensions.
+  What stays in `@upstash/agentkit-eve` is therefore a packaging choice, not a framework limit: the
+  Box sandbox backend (an extension root can't declare a sandbox), the rate-limit `AuthFn` (you drop
+  it into your own channel's `auth` walk), and `defineCachedTool` (wraps user tools).
 
 ## Naming history (so you don't resurrect old names)
 - ai-sdk caching: `cacheTools` → `cachedTool`+`cachedTools` → now **`cachedTools` only** (singular `cachedTool` removed; toolName = map key, `userId` scopes).
@@ -315,8 +320,11 @@ and `eve-extension-demo` (a minimal eve scaffold that mounts the extension).
   `SandboxSession`, `SandboxNetworkPolicy`, etc. (`eve` is a devDep of `packages/eve` for these types.)
 - `ToolDefinition<TInput,TOutput>` = `{ description, inputSchema, execute(input, ctx: ToolContext), … }`.
 - **Extensions** (eve ≥0.24): agent-shaped packages mounted under `agent/extensions/<ns>.ts`; contributions
-  compose as `<ns>__<name>`. They may contribute tools/connections/skills/hooks/instructions — NOT sandbox,
-  channels, schedules, or agent config. Config binds at runtime (mount evaluation), not at discovery.
+  compose as `<ns>__<name>`. They may contribute tools, channels, connections, skills, schedules,
+  subagents, hooks and instruction fragments (eve ≥0.41; channels keep their declared route paths and
+  schedules their cron expressions). The extension **root** cannot declare agent configuration, a
+  sandbox, or nested extensions — but a contributed subagent may own its own config and sandbox.
+  Config binds at runtime (mount evaluation), not at discovery.
   Hooks are observe-only (can't inject context or short-circuit); a thrown hook fails the turn.
 - Stream events for transcripts: `message.received` (`data.message`: flattened user text) and
   `message.completed` (`data.message: string | null`, fires multiple times per turn — interim text before

packages/eve-extension/AGENTS.md

@@ -16,11 +16,15 @@ unavailable, use https://eve.dev/docs/extensions as a fallback.
   `eve/extension`. Config is optional; read bound values via the handle's
   `.config` in tools and hooks.
 - Add contributions under `extension/` the same way as in an agent:
-  `tools/`, `connections/`, `skills/`, `hooks/`, and optional instruction
-  fragments. Names come from file paths; the mount supplies the namespace, so
-  name tools for what they do (`search`, not `crm_search`).
-- An extension cannot declare `agent.ts`, `sandbox`, `schedules`, or nested
-  `extensions/` — those belong to the consuming agent.
+  `tools/`, `channels/`, `connections/`, `skills/`, `schedules/`, `subagents/`,
+  `hooks/`, and optional instruction fragments (eve ≥0.41 supports the full set;
+  channels keep their route paths and schedules their cron expressions). Names
+  come from file paths; the mount supplies the namespace, so name tools for what
+  they do (`search`, not `crm_search`).
+- The extension **root** cannot declare agent configuration (`agent.ts`), a
+  `sandbox/`, or nested `extensions/` — those belong to the consuming agent. A
+  subagent contributed under `extension/subagents/<id>/` may still own its own
+  agent config and sandbox.
 
 ## Build and publish
 

packages/eve-extension/README.md

@@ -158,10 +158,10 @@ importable as static definitions.
 ## When to use this vs `@upstash/agentkit-eve`
 
 Use the extension when you want the batteries-included bundle under one mount. Use
-[`@upstash/agentkit-eve`](../eve) when you need the pieces the extension can't carry — the Upstash
-Box **sandbox backend** and the **rate-limit channel auth** (extensions can't contribute sandbox or
-channel config) — or its `defineCachedTool` wrapper for your own tools. The two compose fine in one
-agent.
+[`@upstash/agentkit-eve`](../eve) when you need the pieces this extension doesn't ship — the Upstash
+Box **sandbox backend** (an extension root can't declare a sandbox) and the **rate-limit channel
+auth** (an `AuthFn` you drop into your own channel's `auth` walk) — or its `defineCachedTool` wrapper
+for your own tools. The two compose fine in one agent.
 
 ## Telemetry
 

packages/eve/README.md

@@ -14,8 +14,9 @@ your `agent/` tree:
 > **Prefer one mount file over per-tool files?**
 > [`@upstash/agentkit-eve-extension`](../eve-extension) packages the memory + search tools (plus
 > durable chat-history capture) as an [eve extension](https://eve.dev/docs/extensions) — mount it once
-> in `agent/extensions/`. This package remains the home of the sandbox backend, the rate-limit auth
-> gate, and `defineCachedTool`, which extensions can't carry.
+> in `agent/extensions/`. This package remains the home of the sandbox backend (an extension root
+> can't declare a sandbox), the rate-limit auth gate, and `defineCachedTool` — the pieces you wire
+> into your own `agent/` tree.
 
 Start from an eve project. Scaffold one (it installs `eve` and an AI-SDK provider for you):