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

fix(ai): scope the Copilot 8-attempt retry budget to model flaps

Browse Source 
The 3 -> 8 bump in callWithCopilotModelRetry was shared by the generic
retryable branch, so a persistent status-less transport blip on Copilot
would ramp across 8 attempts (~11.2s of dead time) instead of the 3 it
took before, and a repeated Retry-After 429 could stretch the same way
on top of the transport's own fetchWithRetry budget.

Derive the budget from the failure kind: model-availability 400s keep the
8-attempt reroll, everything else caps at the previous 3.

Also read COPILOT_TRANSIENT_MODEL_CODES with Object.hasOwn — `code` is
provider-controlled, so a 400 body whose code was `__proto__` or
`toString` classified as transient through the prototype chain.
This commit is contained in:
Magicien
2026-08-04 21:51:35 +01:00
parent d33f2a1658
commit f3ad5ede6b
4 changed files with 64 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/ai/CHANGELOG.md
+1 -1
View File
@@ -4,7 +4,7 @@
### Fixed
- Fixed GitHub Copilot requests failing with a raw `HTTP 400 model_not_available_for_integrator` on roughly half of all turns for recently rolled-out models. Copilot's fleet is not uniform — part of it rejects models that `/models` advertises on the same host — and the transient classifier matched only the older `model_not_supported` code at a fixed envelope depth, so these rejections surfaced as terminal errors instead of entering the existing retry path. Model-availability 400s are now recognized at any envelope depth and retried on a flat delay, with the OpenAI-transport retry budget raised from 3 to 8 attempts.
- Fixed GitHub Copilot requests failing with a raw `HTTP 400 model_not_available_for_integrator` on roughly half of all turns for recently rolled-out models. Copilot's fleet is not uniform — part of it rejects models that `/models` advertises on the same host — and the transient classifier matched only the older `model_not_supported` code at a fixed envelope depth, so these rejections surfaced as terminal errors instead of entering the existing retry path. Model-availability 400s are now recognized at any envelope depth and rerolled on a flat delay with a dedicated 8-attempt budget on the OpenAI transports; every other retryable failure keeps its previous backoff and attempt count.
## [17.2.7] - 2026-08-03
packages/ai/src/error/flags.ts
+3 -1
View File
@@ -493,7 +493,9 @@ function providerErrorCode(error: object): string | undefined {
export function isCopilotTransientModelError(error: unknown): boolean {
if (!error || typeof error !== "object" || status(error) !== 400) return false;
const code = providerErrorCode(error);
if (code !== undefined && COPILOT_TRANSIENT_MODEL_CODES[code]) return true;
// `Object.hasOwn`, not a bare index: `code` is provider-controlled, and a
// prototype key (`__proto__`, `toString`, …) would otherwise read truthy.
if (code !== undefined && Object.hasOwn(COPILOT_TRANSIENT_MODEL_CODES, code)) return true;
const message: unknown = "message" in error ? error.message : undefined;
return typeof message === "string" && COPILOT_MODEL_UNAVAILABLE_PATTERN.test(message);
}
packages/ai/src/utils/retry.ts
+9 -1
View File
@@ -15,6 +15,9 @@ export { isCopilotTransientModelError };
// delay keep the residual near 5% at p=0.7 and under 1% at p=0.5, bounded at
// ~2.8s of dead time in the pathological case.
const COPILOT_MODEL_RETRY_MAX_ATTEMPTS = 8;
// Transport blips and status-bearing failures keep the pre-flap budget: they are
// not coin flips, so a longer ramp only delays surfacing a persistent fault.
const COPILOT_GENERIC_RETRY_MAX_ATTEMPTS = 3;
const COPILOT_MODEL_RETRY_BASE_DELAY_MS = 400;
/** Longest server-requested backoff we are willing to sit out before giving up. */
const COPILOT_RETRY_AFTER_MAX_WAIT_MS = 30_000;
@@ -46,7 +49,12 @@ export async function callWithCopilotModelRetry<T>(
if (options.signal?.aborted) throw error;
const transientModelError = isCopilotTransientModelError(error);
if (!transientModelError && !isRetryableError(error)) throw error;
if (attempt === COPILOT_MODEL_RETRY_MAX_ATTEMPTS - 1) break;
// Budget is per failure kind, counted over attempts already spent: the
// eight-attempt allowance only covers the cheap model-availability reroll.
const maxAttempts = transientModelError
? COPILOT_MODEL_RETRY_MAX_ATTEMPTS
: COPILOT_GENERIC_RETRY_MAX_ATTEMPTS;
if (attempt >= maxAttempts - 1) break;
// Reroll the model flap on a flat delay: a ramp only adds dead time to a
// coin flip the next attempt is equally likely to win. Generic retryable
// failures (429/5xx/transport) keep the linear backoff below.
packages/ai/test/copilot-retry.test.ts
+51 -9
View File
@@ -12,13 +12,18 @@ type ErrorShape = {
code?: string;
error?: { code?: string; message?: string } | { error: { code?: string; message?: string } };
message: string;
headers?: Record<string, string>;
};
function copilotError({ status, code, error, message }: ErrorShape): Error {
function copilotError({ status, code, error, message, headers }: ErrorShape): Error {
const err = new Error(message);
(err as unknown as ErrorShape).status = status;
if (code !== undefined) (err as unknown as ErrorShape).code = code;
if (error !== undefined) (err as unknown as ErrorShape).error = error;
// Single sanctioned assertion point: `Error` carries no provider fields, and
// every test reads them back through the real classifier.
const shaped = err as unknown as ErrorShape;
shaped.status = status;
if (code !== undefined) shaped.code = code;
if (error !== undefined) shaped.error = error;
if (headers !== undefined) shaped.headers = headers;
return err;
}
@@ -75,6 +80,13 @@ describe("isCopilotTransientModelError", () => {
expect(isCopilotTransientModelError(err)).toBe(false);
});
it("does not match 400 codes that collide with Object.prototype keys", () => {
for (const code of ["__proto__", "constructor", "toString", "hasOwnProperty"]) {
const err = copilotError({ status: 400, code, message: "bad request" });
expect(isCopilotTransientModelError(err)).toBe(false);
}
});
it("does not match 401/403/500 regardless of code", () => {
for (const status of [401, 403, 500]) {
const err = copilotError({
@@ -176,9 +188,7 @@ describe("callWithCopilotModelRetry", () => {
async () => {
calls += 1;
if (calls === 1) {
const err = copilotError({ status: 429, message: "rate limited" });
(err as unknown as { headers: Record<string, string> }).headers = { "retry-after": "0.01" };
throw err;
throw copilotError({ status: 429, message: "rate limited", headers: { "retry-after": "0.01" } });
}
return "ok" as const;
},
@@ -188,6 +198,21 @@ describe("callWithCopilotModelRetry", () => {
expect(calls).toBe(2);
});
it("does not stretch a persistent Retry-After 429 across the flap budget", async () => {
let calls = 0;
const err = copilotError({ status: 429, message: "rate limited", headers: { "retry-after": "0.01" } });
await expect(
callWithCopilotModelRetry(
async () => {
calls += 1;
throw err;
},
{ provider: "github-copilot", retryBaseDelayMs: 0 },
),
).rejects.toBe(err);
expect(calls).toBe(3);
});
it("still retries status-less transport blips with the linear backoff", async () => {
let calls = 0;
const result = await callWithCopilotModelRetry(
@@ -206,7 +231,24 @@ describe("callWithCopilotModelRetry", () => {
expect(calls).toBe(2);
});
it("keeps the model-flap delay flat while generic retryable failures ramp", async () => {
it("caps persistent generic retryable failures at the pre-flap budget", async () => {
let calls = 0;
const err = new Error(
'HTTP2StreamReset fetching "https://api.example.com/x". For more information, pass `verbose: true` in the second argument to fetch()',
);
await expect(
callWithCopilotModelRetry(
async () => {
calls += 1;
throw err;
},
{ provider: "github-copilot", retryBaseDelayMs: 0 },
),
).rejects.toBe(err);
expect(calls).toBe(3);
});
it("keeps the flat delay and the larger budget scoped to model flaps", async () => {
const flatWaits: number[] = [];
const rampWaits: number[] = [];
const record = (into: number[]) => {
@@ -242,7 +284,7 @@ describe("callWithCopilotModelRetry", () => {
vi.restoreAllMocks();
expect(flatWaits).toEqual([100, 100, 100, 100, 100, 100, 100]);
expect(rampWaits).toEqual([100, 200, 300, 400, 500, 600, 700]);
expect(rampWaits).toEqual([100, 200]);
});
it("stops retrying when the caller aborts during backoff", async () => {