mirror of
https://github.com/hanzoai/js-sdk.git
synced 2026-10-30 15:02:31 +00:00
Every Hanzo SDK offers the same six capabilities in the same six words. Go and
Python hang them off a client that already existed; TypeScript had none, so it
gains one — `Client`, holding an IAM client id and secret and minting its own
access token (client_secret_basic, RFC 8707 `resource`, held until 60s before
expiry, re-minted once on a 401). It takes no bearer and reads no
HANZO_API_KEY: a credential the SDK is handed is one nobody rotates and one
that says nothing about who is calling.
A refused budget and a refused policy are ANSWERS. `Answer<T>` is a union
discriminated on a literal tag — ok{value} | denied{code, reason, product?,
cures[]} | held{id, clause, reason}, all three carrying the x-request-id that
joins a call to its audit row — so `a.value` does not typecheck until the union
is narrowed and a switch that forgets an arm does not compile. The mapping from
HTTP to arm lives in one function; no capability can grow a rule of its own.
Exceptions stay for outcomes with no decision in them.
The layer is hand-written because the routes it needs are the ones the document
states as an address and not as a shape: /v1/authz/check declares neither a
request body nor a response, /v1/billing/balance and /usage declare no
response. Those three are modelled from what the served handlers read and
write, measured. It sits in src/ beside the generated tree because .generated
is the driver's manifest of what it owns — a file it does not name is this
repo's own and a regeneration never touches it — and the package entry moves to
src/hanzo.ts, which re-exports the generated src/index.ts and adds the client
beside it. Capability types are namespaced (budget.Allowance, search.Hit,
audit.Event, graph.Fact) because thirteen of the contract's names already
belong to generated models.
policy.check sends {subject, verb, path}, which is what
hanzoai/[email protected] serve/use.go reads, not the {sub, obj, act} the route's
own openapi.Describe prose states. The caller still says (sub, act, obj). The
prose and the handler are one fact and should be made to agree upstream.
Gates: `npm run wire` runs the six against a scripted fetch, asserting the
exact request each sends and decoding a realistic answer including both of
cloud's 402 bodies, a 202 hold and a 403 that is not a refusal. examples/six
drives all six in one pass against the live API and is type-checked with the
rest. The agent and tools flows are repaired for the tag names the current
document singularized (AgentsApi -> AgentApi, getTools -> getTool).
Co-authored-by: Hanzo Dev <[email protected]>
27 lines
1.1 KiB
TypeScript
27 lines
1.1 KiB
TypeScript
// tools — what this key can call.
|
|
//
|
|
// GET /v1/tools (get_tools): every tool the caller may see, from every source —
|
|
// connector, function, zap-service, agent, skill, mcp — deduplicated by name.
|
|
// `source` and `activated` narrow it.
|
|
//
|
|
// This flow has moved twice, both times for the same reason, and the rule is
|
|
// worth stating once: it calls a route the DOCUMENT declares. POST /v1/mcp is
|
|
// the fleet's JSON-RPC MCP endpoint and is live, but it is not in the document, and
|
|
// hand-rolling an HTTP request inside a generated client is the drift these SDKs
|
|
// exist to prevent. /v1/tools is declared and typed, so the flow sits here.
|
|
import { ToolApi } from 'hanzoai';
|
|
import { config, fail } from '../client';
|
|
|
|
async function main() {
|
|
const { data } = await new ToolApi(config()).getTool();
|
|
|
|
const tools = data.tools ?? [];
|
|
console.log(`${tools.length} tool(s) visible to this key`);
|
|
for (const t of tools.slice(0, 10)) {
|
|
console.log(` ${t.name ?? '(unnamed)'} [${t.source ?? '?'}] — ${t.description ?? ''}`);
|
|
}
|
|
if (tools.length > 10) console.log(` … and ${tools.length - 10} more`);
|
|
}
|
|
|
|
main().catch(fail);
|