smashimo/oh-my-pi
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(ai): surface Umans request exhaustion without a burst ceiling

Browse Source 
When a payload reports weighted counters but omits hard_cap, the soft/hard
split collapses to a single umans:requests row keyed off the authoritative
weighted effective-request counter, so a spent account can report exhausted
instead of being stuck at warning forever. Raw burst traffic above the limit
still never fabricates exhaustion; weighted headroom stays decisive (#7858).
This commit is contained in:
MertSoylu
2026-08-15 18:16:45 +03:00
parent a899dbc75c
commit 9f6fc1f0c7
4 changed files with 100 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/provider-quirks.md
+1 -1
View File
@@ -1456,7 +1456,7 @@ Umans AI Coding Plan is a proxy service for AI coding models, operating via the
### Auth & usage
- **Auth**: Uses `UMANS_AI_CODING_PLAN_API_KEY` environment variable or `/login umans` key prompt (`packages/ai/src/registry/umans.ts`, `packages/ai/src/registry/registry.ts`). Key validation executes a lightweight Anthropic messages call (`max_tokens: 1`) to `https://api.code.umans.ai/v1/messages`.
- **Usage endpoint**: Fetches quota and rate limit status from `GET /v1/usage` (`packages/ai/src/usage/umans.ts`) using `Authorization: Bearer <key>`.
- **Limits surfaced**: Returns a rolling 5-hour request split into a model-weighted soft cap (`umans:requests:soft`, the "effective requests" contract) and a raw burst ceiling (`umans:requests:hard`, `hard_cap`), plus an instantaneous session concurrency limit (`umans:concurrency`). The soft cap only ever warns — `exhausted` is reserved for the burst ceiling, where throttling actually starts. Legacy payloads without weighted counters fall back to a single raw `umans:requests` row. Also surfaces low-priority status notes when rate-limit bursts occur.
- **Limits surfaced**: Returns a rolling 5-hour request split into a model-weighted soft cap (`umans:requests:soft`, the "effective requests" contract) and a raw burst ceiling (`umans:requests:hard`, `hard_cap`), plus an instantaneous session concurrency limit (`umans:concurrency`). The soft cap only ever warns — `exhausted` is reserved for the burst ceiling, where throttling actually starts. Payloads without a reported burst ceiling (`hard_cap`) collapse to a single weighted `umans:requests` row that can exhaust at the effective-request limit, so request exhaustion is never unreportable; legacy payloads without weighted counters fall back to a single raw `umans:requests` row. In both single-row shapes the weighted counter (when present) stays authoritative — raw burst traffic above the limit never fabricates an exhausted state. Also surfaces low-priority status notes when rate-limit bursts occur.
### Catalog model handling
- **Descriptor & discovery**: Registered as `umans` with default model `umans-coder` (`packages/catalog/src/provider-models/descriptors.ts`). Dynamic discovery fetches model details from `GET /v1/models/info` (`packages/catalog/src/provider-models/openai-compat.ts`).
packages/ai/CHANGELOG.md
+1 -1
View File
@@ -4,7 +4,7 @@
### Fixed
- Fixed Umans usage reporting computing utilization from raw request counts instead of the model-weighted "effective requests", falsely reporting the quota as exhausted while the account still had weighted headroom. The request limit is now split into a weighted soft-cap row (warns, never exhausts) and a raw burst-ceiling row (exhausted only where throttling actually starts), and the rolling 5h window's absolute `resets_at` is surfaced as a countdown ([#7858](https://github.com/can1357/oh-my-pi/issues/7858)).
- Fixed Umans usage reporting computing utilization from raw request counts instead of the model-weighted "effective requests", falsely reporting the quota as exhausted while the account still had weighted headroom. The request limit is now split into a weighted soft-cap row (warns, never exhausts) and a raw burst-ceiling row (exhausted only where throttling actually starts), and the rolling 5h window's absolute `resets_at` is surfaced as a countdown ([#7858](https://github.com/can1357/oh-my-pi/issues/7858)). Payloads without a reported burst ceiling collapse to a single weighted `umans:requests` row that can exhaust, so request exhaustion is never unreportable.
## [17.3.4] - 2026-08-14
packages/ai/src/usage/umans.ts
+14 -4
View File
@@ -119,10 +119,20 @@ function buildRequestsLimits(payload: UmansUsagePayload, provider: string): Usag
...(resetsAt !== undefined ? { resetsAt, resetLabel: "tick" } : {}),
};
// Payloads without weighted counters predate the soft/hard split; keep the
// legacy single row keyed off raw counts.
if (weightedUsed === undefined) {
const amount = buildAmount({ used: rawUsed, limit, remaining: rawRemaining, unit: "requests" });
// Single row: either payloads without weighted counters (legacy) or payloads
// that report weighted usage but no burst ceiling (`hard_cap`). Without a
// burst ceiling there is no hard row to defer exhaustion to, so the
// authoritative counter — weighted when available, else raw — drives the
// single row and CAN exhaust at the limit. Raw burst traffic above the
// limit still never drives exhaustion on its own: weighted headroom stays
// decisive (https://github.com/can1357/oh-my-pi/issues/7858).
if (weightedUsed === undefined || hardCap === undefined) {
const amount = buildAmount({
used: weightedUsed ?? rawUsed,
limit,
remaining: weightedUsed !== undefined ? weightedRemaining : rawRemaining,
unit: "requests",
});
return [
{
id: "umans:requests",
packages/ai/test/umans-usage.test.ts
+84
View File
@@ -174,6 +174,90 @@ describe("umans usage provider", () => {
expect(report?.limits.some(l => l.status === "exhausted")).toBe(false);
});
it("collapses to a single weighted requests row that can exhaust when no burst ceiling is reported", async () => {
// Weighted counters present but `hard_cap` absent: without a burst
// ceiling there is no hard row to defer exhaustion to, so the weighted
// effective-request budget is the operative ceiling — the single row
// must be able to report `exhausted` or a spent account could never
// trigger the usage-aware fallback.
const report = await umansUsageProvider.fetchUsage(
{
provider: "umans",
credential: { type: "api_key", apiKey: "sk-test" },
},
{
fetch: fakeFetch(
umansPayload({
limits: {
requests: { limit: 200, window_seconds: 18000 },
concurrency: { limit: 4, hard_cap: 8, burst_pct: 1.0 },
},
usage: {
requests_in_window: 400,
remaining_requests: 0,
weighted_in_window: 200,
weighted_remaining_requests: 0,
concurrent_sessions: 0,
tokens_in: 0,
tokens_out: 0,
priority: { low: false },
},
}),
),
},
);
const requests = report?.limits.find(l => l.id === "umans:requests");
expect(requests).toBeDefined();
// No soft/hard split without a reported burst ceiling.
expect(report?.limits.some(l => l.id.startsWith("umans:requests:"))).toBe(false);
// Weighted effective requests are authoritative: raw 400 overshoots the
// 200 limit, but it is the weighted 200/200 that reports exhausted.
expect(requests?.amount.used).toBe(200);
expect(requests?.amount.limit).toBe(200);
expect(requests?.amount.usedFraction).toBe(1);
expect(requests?.status).toBe("exhausted");
});
it("keeps weighted headroom decisive when no burst ceiling is reported", async () => {
// Same #7858 shape (raw usage over the soft limit, weighted headroom
// remaining) but with no `hard_cap` in the payload: the weighted counter
// must still decide, so raw burst traffic cannot fabricate an exhausted
// state even when there is no hard row to buffer it.
const report = await umansUsageProvider.fetchUsage(
{
provider: "umans",
credential: { type: "api_key", apiKey: "sk-test" },
},
{
fetch: fakeFetch(
umansPayload({
limits: {
requests: { limit: 200, window_seconds: 18000 },
concurrency: { limit: 4, hard_cap: 8, burst_pct: 1.0 },
},
usage: {
requests_in_window: 300,
remaining_requests: 0,
weighted_in_window: 100,
weighted_remaining_requests: 100,
concurrent_sessions: 0,
tokens_in: 0,
tokens_out: 0,
priority: { low: false },
},
}),
),
},
);
const requests = report?.limits.find(l => l.id === "umans:requests");
expect(requests).toBeDefined();
expect(requests?.amount.used).toBe(100);
expect(requests?.amount.remaining).toBe(100);
expect(requests?.amount.usedFraction).toBeCloseTo(0.5, 5);
expect(requests?.status).toBe("ok");
expect(report?.limits.some(l => l.status === "exhausted")).toBe(false);
});
it("reserves exhausted for the burst ceiling and warns at the soft cap", async () => {
const report = await umansUsageProvider.fetchUsage(
{