mirror of
https://github.com/hanzo-templates/flutter-chat.git
synced 2026-10-30 20:52:58 +00:00
The catalogue shipped 71 web templates and one EMPTY `mobile` repo, so "build an app" had no starting point on the platform people actually ship to. This is the iOS lane — Expo/React Native and Flutter for cross-platform, SwiftUI for native. The constraint that shaped every file here: an iOS archive needs macOS + Xcode + a signing identity, and our arc fleet has none. hanzoai has 519 registered runners and every one reports `os: linux`; luxfi has 151, same story; zooai has none. The only macOS builds in the org today (hanzoai/world build-desktop.yml) run on GitHub-HOSTED macos-14, which is exactly what the house CI rule says not to do. So `hanzo.yml` declares only what a Linux arc runner can genuinely execute — typecheck + web bundle for Expo, analyze + test + web for Flutter, `swiftc -parse` for SwiftUI — and the macOS archive lives in a separate `workflow_dispatch`-only `ios.yml`. A job pinned to a label no runner advertises would sit queued forever if it fired on push; parking it is honest, a permanently-yellow check is not. The README of every template states the gap and the one command that closes it. Same reasoning on the demos. Expo and Flutter both have a real web target, so `<slug>.hanzo.app` serves the SAME source the simulator runs. SwiftUI has no web target, so those two templates ship with no demo link rather than a screenshot pretending to be one. Upstreams are permissive only and recorded per template: expo/expo templates (MIT), flutter/samples (BSD-3-Clause), XcodeGen (MIT). Second measured gap, found while wiring this up: `hanzo-templates` is not in arcd's `orgs:` list either (hanzoai, luxfi, zooai, parsdao, zenlm), so the Linux jobs would queue with no runner too. Both gaps are one installer re-run, no code change; both are spelled out in every README rather than left to be discovered on the first red build. Third gap, found by probing rather than assuming, and FIXED here: every run in this org died as `startup_failure` with no logs and no annotation. A bare `runs-on: [self-hosted, linux]` probe queued fine while the same job importing hanzoai/ci failed identically at @v1 and at a pinned SHA — so it was never the tag. The org defaulted GITHUB_TOKEN to `read`, and the reusable declares `permissions: packages: write`; a caller cannot exceed the org maximum, so GitHub kills the run before creating a job. hanzo-templates now defaults to `write` like hanzoai/luxfi/zooai, and runs reach `queued` — the honest state, waiting on gap 2.
60 lines
2.0 KiB
Dart
60 lines
2.0 KiB
Dart
import 'dart:async';
|
|
import 'dart:convert';
|
|
|
|
import 'package:http/http.dart' as http;
|
|
|
|
/// The whole Hanzo integration: one streaming call to /v1/chat/completions.
|
|
///
|
|
/// Config is read at RUNTIME (Settings) rather than baked in with
|
|
/// --dart-define, because a compiled-in key ships inside every copy of the app.
|
|
/// Ship the app, let the user bring the key.
|
|
class Msg {
|
|
const Msg(this.role, this.content);
|
|
final String role;
|
|
final String content;
|
|
Map<String, String> toJson() => {'role': role, 'content': content};
|
|
}
|
|
|
|
class Hanzo {
|
|
Hanzo._();
|
|
static final Hanzo instance = Hanzo._();
|
|
|
|
String base = 'https://api.hanzo.ai/v1';
|
|
String model = 'zen-omni';
|
|
String key = '';
|
|
|
|
/// Yields the assistant reply token by token. The caller appends each chunk;
|
|
/// nothing here knows about widgets.
|
|
Stream<String> complete(List<Msg> messages) async* {
|
|
if (key.isEmpty) {
|
|
throw StateError('No API key. Open Settings and paste a Hanzo key.');
|
|
}
|
|
|
|
final req = http.Request('POST', Uri.parse('$base/chat/completions'))
|
|
..headers.addAll({'Content-Type': 'application/json', 'Authorization': 'Bearer $key'})
|
|
..body = jsonEncode({
|
|
'model': model,
|
|
'messages': messages.map((m) => m.toJson()).toList(),
|
|
'stream': true,
|
|
});
|
|
|
|
final res = await req.send();
|
|
if (res.statusCode != 200) {
|
|
throw http.ClientException('${res.statusCode} ${await res.stream.bytesToString()}');
|
|
}
|
|
|
|
// LineSplitter reassembles SSE frames that arrive split across chunks.
|
|
await for (final line in res.stream.transform(utf8.decoder).transform(const LineSplitter())) {
|
|
if (!line.startsWith('data:')) continue;
|
|
final payload = line.substring(5).trim();
|
|
if (payload == '[DONE]') return;
|
|
try {
|
|
final delta = (jsonDecode(payload)['choices'] as List).first['delta']['content'];
|
|
if (delta is String && delta.isNotEmpty) yield delta;
|
|
} catch (_) {
|
|
// keep-alive comment or a frame we do not model — skip it
|
|
}
|
|
}
|
|
}
|
|
}
|