refactor(agentos): drop the host function wrappers, names, and descriptions

[NathanFlurry]

Host functions are now a plain record of collections. The keys name everything, and a function needs only a schema and a handler:

hostFunctions: {
  store: {
    listOrders: {
      inputSchema: z.object({ customer: z.string() }).describe("List a customer's orders."),
      execute: ({ customer }) => db.orders.findMany({ customer }),
    },
  },
}

Removed from the public API

• hostFunction() and hostFunctions() wrappers.
• HostFunctions.name and HostFunctions.description (the collection key names it).
• The functions: nesting key (functions are the collection's own keys).
• HostFunction.description (comes from the schema's .describe()).
• MAX_HOST_FUNCTION_DESCRIPTION_LENGTH and validateHostFunctions.
• HostFunctions as an interface, replaced by HostFunctionCollection / HostFunctionCollections.
• HostFunctionDefinition and its commandAliases, which only made sense in NodeRuntime's old one-command-per-function model.

Behavior

• Keys may be camelCase or kebab-case; both resolve to one kebab-case command name, so store.listOrders is agentos-store list-orders on the CLI and store.listOrders(input) in guest JavaScript. A key that cannot become a command name, or two keys that collide once converted, throw before the VM is created.
• Descriptions are optional on both sides now. The sidecar keeps the length cap and no longer rejects a registration that omits one; the client no longer duplicates that limit.
• execute still gets its input typed from its own inputSchema, with no wrapper to hang the inference on. The options types carry a schema type parameter that AgentOs.create(), agentOS(), NodeRuntime.create() and the secure-exec entry points infer from the literal. A collection built outside the call that takes it has nothing to infer from, so its input falls back to any rather than forcing every handler to annotate.
• Mirrored in the Rust client: HostFunction loses name/description, HostFunctions becomes HostFunctionCollections (a BTreeMap), and the description is read from the JSON Schema.

Breaking change for NodeRuntime guest code

NodeRuntime used to take a flat map and register each function as its own single-callback collection, so the guest invoked it as a top-level command named after the function (add --json ...). It now takes the same two-level collections as AgentOs.create(), and the guest reaches a function as agentos-<collection> <function>. The host function module moves from packages/core down into packages/runtime-core so both layers share one implementation rather than two parallel ones.

agentos list-bindings is intentionally retained as an alias for list-host-functions.

Docs, examples, and both docs bundles are updated to the inline form.

🤖 Generated with Claude Code

[railway-app]

🚅 Environment agentos-pr-2006 in rivet-frontend has no services deployed.

[the-company-company]

🔴 1 high · 🟠 2 medium · 🔵 1 low

_{Reviewed commit 150b793.}

[the-company-company]

🔴 High · Sandbox host functions are discarded before VM registration

resolveHostFunctions runs before resolveSandboxOptions, but the latter is what adds the automatic sandbox collection. hostFunctions is then fixed to the earlier result, so AgentOs.create({ sandbox: ... }) mounts the sandbox filesystem but never registers agentos-sandbox or any process callbacks. Resolve the expanded options.hostFunctions after sandbox expansion (while validating caller-provided functions before starting the provider if needed).

[the-company-company]

🟠 Medium · Sandbox integration test still calls the removed wrapper shape

createSandboxHostFunctions() now returns the collection directly (runCommand, createProcess, and so on), so hostFunctions.hostFunctions is undefined. This test throws before exercising the sandbox, and the same stale access remains for create/list/kill below. Update these calls to the new camelCase members so the migrated integration test can run.

[the-company-company]

🟠 Medium · Removed exports leave a default test file unloadable

The non-nightly packages/core/tests/host-function-command-exec.test.ts still imports both hostFunction and hostFunctions from this entry point and constructs its fixture with them. Removing these exports without migrating that test makes Vitest fail while loading the default core suite, before any tests execute. Convert the fixture to the new keyed object API in this PR.

[the-company-company]

🔵 Low · Explicit evaluation result types disable host-function input inference

TypeScript does not infer omitted type parameters after any explicit type argument; it uses their defaults. Therefore the common evaluate<MyResult>(..., { hostFunctions: ... }) form fixes HOST_FUNCTIONS to HostFunctionSchemas, and every inline execute(input) falls back to any, defeating the new schema-derived typing. Add a type-level regression test for this call form and reshape the API so the result annotation does not prevent host-function inference.