Infinite loading guard: heartbeats + stall watchdogs on the model runs

A user got stuck on the wizard's loading screen: a hung model meant
silence, and silence was indistinguishable from work - the server's own
hard deadline is 2x OLLAMA_TIMEOUT_MS (ten minutes by default), and the
client waited on a dead stream indefinitely.

Both generate routes (deck, homework) now send an NDJSON heartbeat
every 15s, and the client guards the two quiet phases: no response at
all for 45s, and no bytes for 90s mid-stream, both surfacing the
regular error state with a Retry. The caller's abort still maps to the
friendly "cancelled" state.

Verified with a stalled route and Playwright's fake clock in the wizard
suite (93 checks), plus contrast 0 findings, mobile 124/124, chat
101/101, regression 44/44, i18n 13/13, native 33/33.
Version 0.4.0, SW -> freezmo-v69.
This commit is contained in:
Martin Dvořák 2026-10-13 18:06:25 +00:00
commit 20bb08b0e7
8 changed files with 125 additions and 19 deletions

View file

@ -30,10 +30,17 @@ export type GenerateProgress = {
message: string;
};
/** No response at all for this long: the run is stuck, not slow. */
const RESPONSE_GUARD_MS = 45_000;
/** No bytes for this long while streaming: the model stopped talking. The
* routes heartbeat every 15s, so 90s of silence really is silence. */
const STREAM_GUARD_MS = 90_000;
type StreamEvent =
| { type: "stage"; stage: "vision" | "generation" | "done"; message: string }
| { type: "deck"; deck: StreamedDeck; meta: Record<string, unknown> }
| { type: "error"; error: string; kind?: string };
| { type: "error"; error: string; kind?: string }
| { type: "ping" };
export class ApiError extends Error {
kind: string;
@ -52,11 +59,13 @@ export class ApiError extends Error {
export async function consumeNdjson(
response: Response,
handleEvent: (event: unknown) => void,
options: { stallMs?: number } = {},
): Promise<void> {
if (!response.body) throw new ApiError(t("api.empty"), "bad_response");
const reader = response.body.getReader();
const decoder = new TextDecoder();
let buffer = "";
const stallMs = options.stallMs ?? 0;
const flush = (line: string) => {
if (!line.trim()) return;
try {
@ -66,9 +75,25 @@ export async function consumeNdjson(
}
};
while (true) {
const { done, value } = await reader.read();
if (done) break;
buffer += decoder.decode(value, { stream: true });
let chunk: ReadableStreamReadResult<Uint8Array>;
if (stallMs > 0) {
let timer: ReturnType<typeof setTimeout> | null = null;
const stalled = new Promise<never>((_, reject) => {
timer = setTimeout(() => reject(new ApiError(t("api.stalled"), "timeout")), stallMs);
});
try {
chunk = await Promise.race([reader.read(), stalled]);
} catch (error) {
void reader.cancel().catch(() => {});
throw error;
} finally {
if (timer) clearTimeout(timer);
}
} else {
chunk = await reader.read();
}
if (chunk.done) break;
buffer += decoder.decode(chunk.value, { stream: true });
let newline = buffer.indexOf("\n");
while (newline !== -1) {
flush(buffer.slice(0, newline));
@ -87,19 +112,35 @@ export async function generateDeck(
input: GenerateInput,
options: { onProgress?: (progress: GenerateProgress) => void; signal?: AbortSignal } = {},
): Promise<{ deck: StreamedDeck; meta: Record<string, unknown> }> {
// A run that goes quiet is worse than one that fails: the wizard used to sit
// on its loading screen until the server's own 2x deadline (ten minutes by
// default) when the model hung. The caller's signal is forwarded, and a
// no-response guard covers the "nothing at all" phase.
const controller = new AbortController();
const forwarded = options.signal;
let headersTimedOut = false;
if (forwarded) {
if (forwarded.aborted) controller.abort();
else forwarded.addEventListener("abort", () => controller.abort(), { once: true });
}
const headersTimer = setTimeout(() => {
headersTimedOut = true;
controller.abort();
}, RESPONSE_GUARD_MS);
let response: Response;
try {
response = await fetch("/api/generate", {
method: "POST",
headers: { "Content-Type": "application/json", ...authHeaders(), ...localeHeader() },
body: JSON.stringify(input),
signal: options.signal,
signal: controller.signal,
});
} catch (error) {
if (error instanceof Error && error.name === "AbortError") {
throw new ApiError(t("api.cancelled"), "aborted");
}
if (forwarded?.aborted) throw new ApiError(t("api.cancelled"), "aborted");
if (headersTimedOut) throw new ApiError(t("api.stalled"), "timeout");
throw new ApiError(t("api.unreachable"), "network");
} finally {
clearTimeout(headersTimer);
}
if (!response.ok) {
@ -124,7 +165,7 @@ export async function generateDeck(
} else if (event.type === "error") {
streamError = new ApiError(event.error, event.kind);
}
});
}, { stallMs: STREAM_GUARD_MS });
if (streamError) throw streamError;
if (!deck) {
@ -183,7 +224,8 @@ export type StreamedHomework =
type HomeworkStreamEvent =
| { type: "stage"; stage: "vision" | "generation" | "done"; message: string }
| { type: "homework"; homework: StreamedHomework; meta: Record<string, unknown> }
| { type: "error"; error: string; kind?: string };
| { type: "error"; error: string; kind?: string }
| { type: "ping" };
/**
* Calls POST /api/homework and consumes the NDJSON progress stream. Same shape
@ -193,19 +235,32 @@ export async function generateHomework(
input: HomeworkInput,
options: { onProgress?: (progress: HomeworkProgress) => void; signal?: AbortSignal } = {},
): Promise<{ homework: StreamedHomework; meta: Record<string, unknown> }> {
// Same guards as generateDeck: no infinite loading screen on a hung model.
const controller = new AbortController();
const forwarded = options.signal;
let headersTimedOut = false;
if (forwarded) {
if (forwarded.aborted) controller.abort();
else forwarded.addEventListener("abort", () => controller.abort(), { once: true });
}
const headersTimer = setTimeout(() => {
headersTimedOut = true;
controller.abort();
}, RESPONSE_GUARD_MS);
let response: Response;
try {
response = await fetch("/api/homework", {
method: "POST",
headers: { "Content-Type": "application/json", ...authHeaders(), ...localeHeader() },
body: JSON.stringify(input),
signal: options.signal,
signal: controller.signal,
});
} catch (error) {
if (error instanceof Error && error.name === "AbortError") {
throw new ApiError(t("api.cancelled"), "aborted");
}
if (forwarded?.aborted) throw new ApiError(t("api.cancelled"), "aborted");
if (headersTimedOut) throw new ApiError(t("api.stalled"), "timeout");
throw new ApiError(t("api.unreachable"), "network");
} finally {
clearTimeout(headersTimer);
}
if (!response.ok) {
@ -228,7 +283,7 @@ export async function generateHomework(
} else if (event.type === "error") {
streamError = new ApiError(event.error, event.kind);
}
});
}, { stallMs: STREAM_GUARD_MS });
if (streamError) throw streamError;
if (!homework) {