UNPKG

@bitkyc08/opencodex

Version:

Universal provider proxy for OpenAI Codex & Claude Code — use any LLM with Codex CLI/App/SDK and Claude Code

344 lines (315 loc) • 17.8 kB
import { parseUpstreamJsonPayload, safeUpstreamErrorString, sanitizeUpstreamErrorText } from "./upstream-http-error"; import { isLocationUnsupportedMessage } from "../lib/errors"; import { googleErrorDetails, isAntigravityTosViolationBody } from "./google-error-details"; export const ANTIGRAVITY_VALIDATION_REQUIRED_PREFIX = "Antigravity account validation required (VALIDATION_REQUIRED)"; /** Leading words of every normalized Antigravity 403 that is not a validation demand (see `classifyGoogle`). */ export const ANTIGRAVITY_ACCESS_DENIED_PREFIX = "Antigravity access denied"; /** Leading words of a normalized Antigravity refusal whose text names the credential itself. */ export const ANTIGRAVITY_AUTH_FAILED_PREFIX = "Antigravity authentication failed"; /** * Written right after a normalized prefix only when the response carried Google's structured * `error.status: "PERMISSION_DENIED"`. Message wording alone never produces it, so a reader of the * normalized string can tell a refusal Google attributed to the account from a proxy page or an * enum-less envelope that merely says "expired". */ export const GOOGLE_PERMISSION_DENIED_SUFFIX = " (PERMISSION_DENIED)"; /** Emitted only for a body `isAntigravityTosViolationBody` accepts; dispatch trusts nothing else. */ export const ANTIGRAVITY_TOS_VIOLATION_PREFIX = "Antigravity account suspended (TOS_VIOLATION)"; const ANTIGRAVITY_TOS_VIOLATION_MESSAGE = "Google disabled this account for a Terms of Service violation; appeal with Google, then sign in again"; function hasAntigravityValidationReason(payloadText: string): boolean { const payload = parseUpstreamJsonPayload(payloadText); if (!payload || typeof payload !== "object" || Array.isArray(payload)) return false; const error = (payload as { error?: unknown }).error; if (!error || typeof error !== "object" || Array.isArray(error)) return false; const details = (error as { details?: unknown }).details; return Array.isArray(details) && details.some(detail => detail !== null && typeof detail === "object" && !Array.isArray(detail) && (detail as { reason?: unknown }).reason === "VALIDATION_REQUIRED"); } /** Pull the human detail out of the Google API error envelope `{error:{message,status,code}}`. */ function googleErrorDetail(payloadText: string): { message?: string; status?: string } { const trimmed = payloadText.trim(); if (!trimmed || (!trimmed.startsWith("{") && !trimmed.startsWith("["))) { return { message: trimmed || undefined }; } const parsed = parseUpstreamJsonPayload(trimmed); if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) return {}; const err = (parsed as { error?: { message?: unknown; status?: unknown } }).error; return { message: safeUpstreamErrorString(err?.message), status: safeUpstreamErrorString(err?.status), }; } const GOOGLE_QUOTA_EXHAUSTED_NEEDLES = [ "quotafailure", "quota exceeded", "exceeded your current quota", "billing", "individual quota reached", "quota reached", "enable overages", "exhausted your capacity", "daily limit reached", "weekly limit reached", ]; // Per-minute / per-second / concurrency limits are transient rate limits: they should be // retried, not treated as hard quota exhaustion. These guards run before the needles so a // message like "Per-minute quota exceeded" stays in the retryable bucket. const GOOGLE_TRANSIENT_RATE_LIMIT_PATTERNS = [ "per minute", "per-minute", "per min", "rpm", "requests per minute", "too many requests", "rate limit", "retry after", // Upstream writes the header name both ways in prose ("retry-after: 60"); matching only the // spaced spelling let a transient 429 fall through to the exhaustion needles below. "retry-after", "concurrent request limit", ]; /** Whether the text names a transient rate limit rather than hard quota exhaustion. */ export function isGoogleTransientRateLimitText(text: string): boolean { const lower = text.toLowerCase(); return GOOGLE_TRANSIENT_RATE_LIMIT_PATTERNS.some(needle => lower.includes(needle)); } export function isGoogleQuotaExhaustedText(text: string): boolean { const lower = text.toLowerCase(); if (GOOGLE_TRANSIENT_RATE_LIMIT_PATTERNS.some(needle => lower.includes(needle))) return false; return GOOGLE_QUOTA_EXHAUSTED_NEEDLES.some(needle => lower.includes(needle)); } /** * `ErrorInfo.reason` values that make a 429 a rate/quota signal the retry ladder may wait out. * Google writes these into `error.details[]`; Cloud Code Assist answers a genuine capacity 429 * with `reason: "RATE_LIMIT_EXCEEDED"` and answers a request-shape/policy rejection with no * `details` at all. */ const GOOGLE_RETRYABLE_429_ERROR_REASONS: Record<string, true> = { RATE_LIMIT_EXCEEDED: true, QUOTA_EXCEEDED: true, RATE_LIMIT_EXCEEDED_MODEL: true, }; /** Narrow a JSON value to a plain object without asserting a shape it may not have. */ function jsonRecord(value: unknown): Record<string, unknown> | undefined { return value && typeof value === "object" && !Array.isArray(value) ? (value as Record<string, unknown>) : undefined; } /** ProtoJSON type of a detail whose `reason` field carries an `ErrorInfo` verdict. */ const GOOGLE_ERROR_INFO_TYPE = "google.rpc.ErrorInfo"; /** ProtoJSON type of a detail whose `retryDelay` field states when to retry. */ const GOOGLE_RETRY_INFO_TYPE = "google.rpc.RetryInfo"; /** * Whether a detail's declared `@type` admits `expected`. * * ProtoJSON spells a type either fully qualified (`type.googleapis.com/google.rpc.ErrorInfo`) or as * the bare message name; both name the same message, and only the last path segment carries the * identity. Suffix matching alone would accept a foreign `type.example.com/relay.google.rpc.ErrorInfo`, * so the segment is compared whole. A detail with no `@type` asserts nothing and admits both. */ function detailTypeMatches(record: Record<string, unknown>, expected: string): boolean { if (!Object.hasOwn(record, "@type")) return true; const declared = record["@type"]; if (typeof declared !== "string") return false; const segment = declared.slice(declared.lastIndexOf("/") + 1); return segment === expected; } /** * The `error` envelope of a Google error payload, or `undefined` when the text is not JSON. * * `error` is normally an object, but both the retry-delay extractor and Google's own fixtures * accept the array spelling, so every reader here unwraps `error[0]` the same way rather than * disagreeing about which payloads count as evidence. */ function googleErrorEnvelope(payloadText: string): Record<string, unknown> | undefined { const trimmed = payloadText.trim(); if (!trimmed.startsWith("{") && !trimmed.startsWith("[")) return undefined; const parsed = jsonRecord(parseUpstreamJsonPayload(trimmed)); if (!parsed) return undefined; const error = parsed.error; if (Array.isArray(error)) return jsonRecord(error[0]) ?? parsed; return jsonRecord(error) ?? parsed; } /** Every `error.details[].reason` in a Google error envelope, in payload order. */ function googleErrorReasons(payloadText: string): string[] { const details = googleErrorEnvelope(payloadText)?.details; if (!Array.isArray(details)) return []; const reasons: string[] = []; for (const detail of details) { const record = jsonRecord(detail); if (!record) continue; // `@type` names which protobuf message owns this object. A detail that declares a different // message may carry an unrelated field spelled `reason`, so only an `ErrorInfo` detail (or one // with no `@type` to contradict it) speaks for the rate-limit verdict. if (!detailTypeMatches(record, GOOGLE_ERROR_INFO_TYPE)) continue; const reason = safeUpstreamErrorString(record.reason); if (reason) reasons.push(reason); } return reasons; } function isGoogleRetryDuration(value: unknown): boolean { return typeof value === "string" && /^\s*\d+(?:\.\d+)?\s*s\s*$/i.test(value); } /** * Whether a detail states a wait, including an explicit `0s`. * * `extractGoogleRetryAfterSeconds` drops `0s` because it answers "how long should the caller * wait", and a zero there is no instruction. For eligibility the question is different: the * provider named a wait at all, which says the refusal is a rate signal rather than a verdict on * the request's shape. Local backoff still supplies the actual delay. */ function hasGoogleRetryDelayDetail(payloadText: string): boolean { const details = googleErrorEnvelope(payloadText)?.details; if (!Array.isArray(details)) return false; for (const detail of details) { const record = jsonRecord(detail); if (!record) continue; if (detailTypeMatches(record, GOOGLE_RETRY_INFO_TYPE) && isGoogleRetryDuration(record.retryDelay)) return true; const metadata = jsonRecord(record.metadata); if (detailTypeMatches(record, GOOGLE_ERROR_INFO_TYPE) && isGoogleRetryDuration(metadata?.quotaResetDelay)) return true; } return false; } /** * The text a message-level needle may match: the decoded `error.message` and enum status for * JSON, the whole payload for non-JSON. * * Reading the raw JSON would let any string inside it authorize a replay — a `BadRequest` field * violation whose description quotes "rate limit" is part of a request-shape verdict, not a rate * signal, and matching it re-enables exactly the futile replay this gate exists to stop. */ function googleMessageText(payloadText: string): string { const trimmed = payloadText.trim(); if (!trimmed.startsWith("{") && !trimmed.startsWith("[")) return trimmed; const envelope = googleErrorEnvelope(trimmed); if (!envelope) return ""; return [safeUpstreamErrorString(envelope.message), safeUpstreamErrorString(envelope.status)] .filter((part): part is string => typeof part === "string") .join(" "); } /** * Message wording that is evidence for a replay but must not change quota classification. * * Google's documented 429 for pay-as-you-go capacity is "Resource exhausted. Please try again * later." — no header, no details, no rate-limit wording. Treating it as a general transient * pattern would also veto hard-quota classification for every body that says it, so it is scoped * to the evidence predicate instead. */ const GOOGLE_RETRYABLE_429_MESSAGE_NEEDLES = [ "try again later", ]; /** * Whether a 429 says anything a replay can act on. * * The ladder replays a byte-identical body after 250 ms, doubling to 2 s. That is a sensible * probe for a burst limit that clears, and meaningless for a refusal Google already made on the * request's shape: the same bytes are refused again. Only evidence in the response separates * the two. When there is no validated Retry-After, check for a typed delay detail, a * transient rate-limit phrase, or a structured `ErrorInfo.reason` before replaying. * * Observed Cloud Code Assist shapes behind this split (same account, minutes apart): * - capacity: `RESOURCE_EXHAUSTED` / "You have exhausted your capacity on this model. Resets in * 0s." with `details[].reason: "RATE_LIMIT_EXCEEDED"` — already hard-quota classified, and * evidence-bearing if it ever is not; * - request-shape rejection: `RESOURCE_EXHAUSTED` / "Resource has been exhausted (e.g. check * quota)." with no `details`, no `Retry-After` — replayed three times per turn, refused three * times per turn. */ export function hasGoogle429BodyRetryEvidence(payloadText: string): boolean { if (hasGoogleRetryDelayDetail(payloadText)) return true; const messageText = googleMessageText(payloadText).toLowerCase(); if (isGoogleTransientRateLimitText(messageText)) return true; if (GOOGLE_RETRYABLE_429_MESSAGE_NEEDLES.some(needle => messageText.includes(needle))) return true; return googleErrorReasons(payloadText).some(reason => GOOGLE_RETRYABLE_429_ERROR_REASONS[reason] === true); } function classifyGoogle(label: string, status: number | undefined, enumStatus: string | undefined, text: string, reason?: string): string { // Structured reasons outrank prose. Quota exhaustion is the same contract on every Google // surface; RATE_LIMIT_EXCEEDED overrides quota-like prose only on Antigravity, as CPA reads it. if (reason === "QUOTA_EXHAUSTED" || reason === "INSUFFICIENT_G1_CREDITS_BALANCE") return `${label} quota exhausted`; if (reason === "RATE_LIMIT_EXCEEDED" && label === "Antigravity") return `${label} rate limit exceeded`; const lower = `${enumStatus ?? ""} ${text}`.toLowerCase(); const quotaExhausted = isGoogleQuotaExhaustedText(lower); if ((!enumStatus || enumStatus === "RESOURCE_EXHAUSTED") && quotaExhausted) return `${label} quota exhausted`; if (status === 429 || enumStatus === "RESOURCE_EXHAUSTED" || lower.includes("rate limit")) { return `${label} rate limit exceeded`; } if (status === 401 || enumStatus === "UNAUTHENTICATED" || lower.includes("unauthenticated") || lower.includes("invalid authentication") || lower.includes("expired")) { // Keep the enum here too: the access-denial recovery accepts this wording only under it. return `${label} authentication failed${enumStatus === "PERMISSION_DENIED" ? GOOGLE_PERMISSION_DENIED_SUFFIX : ""}`; } // Keep Google's explicit enum in the normalized text. Responses/combo handling receives // only this string, so dropping it would let location wording override the authoritative // permission reason during downstream classification. if (enumStatus === "PERMISSION_DENIED") { return `${label} access denied${GOOGLE_PERMISSION_DENIED_SUFFIX}`; } if (status === 403 || lower.includes("permission_denied") || lower.includes("permission denied") || lower.includes("access denied")) { return `${label} access denied`; } // Google rejects unsupported geographic / datacenter locations with HTTP 400 // FAILED_PRECONDITION. The payload is not malformed, so it must not fall through to // "invalid request" (#3467). Only the observed 400/precondition envelope permits // this inference; other explicit enums and server statuses remain authoritative. if (status === 400 && (!enumStatus || enumStatus === "FAILED_PRECONDITION") && isLocationUnsupportedMessage(lower)) { return `${label} location not supported`; } if (status === 503 || enumStatus === "UNAVAILABLE" || lower.includes("overloaded") || lower.includes("unavailable")) { return `${label} server overloaded`; } if (status === 400 || status === 404 || enumStatus === "INVALID_ARGUMENT" || enumStatus === "NOT_FOUND" || lower.includes("invalid") || lower.includes("not found") || lower.includes("malformed")) { return `${label} invalid request`; } return `${label} upstream error`; } /** * Normalize a Google/Vertex/Antigravity HTTP error body into a short, classified, secret-redacted * message. Mirrors `kiro-errors.ts`. `label` is the provider-facing prefix ("Vertex AI", * "Antigravity"). */ export function safeGoogleHttpErrorMessage(label: string, status: number, payloadText: string): string { // A suspension carries a fixed message: Google's text names an appeal path and the account. if (label === "Antigravity" && status === 403 && isAntigravityTosViolationBody(payloadText)) { return `${ANTIGRAVITY_TOS_VIOLATION_PREFIX}: ${ANTIGRAVITY_TOS_VIOLATION_MESSAGE}`; } const { message, status: enumStatus } = googleErrorDetail(payloadText); const prefix = label === "Antigravity" && status === 403 && hasAntigravityValidationReason(payloadText) ? ANTIGRAVITY_VALIDATION_REQUIRED_PREFIX : classifyGoogle(label, status, enumStatus, [message, enumStatus].filter(Boolean).join(" "), googleErrorDetails(payloadText).reason); const detail = message ? sanitizeUpstreamErrorText(message).slice(0, 500) : `HTTP ${status}`; return `${prefix}: ${detail}`; } /** Vertex AI HTTP error message (label = "Vertex AI"). */ export function safeVertexHttpErrorMessage(status: number, payloadText: string): string { return safeGoogleHttpErrorMessage("Vertex AI", status, payloadText); } /** Antigravity (Cloud Code Assist) HTTP error message (label = "Antigravity"). */ export function safeAntigravityHttpErrorMessage(status: number, payloadText: string): string { return safeGoogleHttpErrorMessage("Antigravity", status, payloadText); } /** Google-family retryable HTTP set (mirrors Kiro). Quota-exhausted is classified above and not retried. */ export function retryableGoogleStatus(status: number): boolean { return status === 429 || status === 500 || status === 502 || status === 503 || status === 504; } /** * True when a 429 body indicates hard quota exhaustion (not a transient rate limit). Quota * exhaustion is generally not expected to recover for hours (AIP-194), so it must NOT be retried — * unlike a plain rate limit. The HTTP status alone can't distinguish the two, so the retry layer * inspects the body with this. */ export function isQuotaExhaustedBody(payloadText: string, label?: string): boolean { const { reason } = googleErrorDetails(payloadText); if (reason === "QUOTA_EXHAUSTED" || reason === "INSUFFICIENT_G1_CREDITS_BALANCE") return true; if (reason === "RATE_LIMIT_EXCEEDED" && label === "Antigravity") return false; const { message, status } = googleErrorDetail(payloadText); if (status && status !== "RESOURCE_EXHAUSTED") return false; return isGoogleQuotaExhaustedText(message ?? payloadText); } /** * A server-stated retry delay in whole seconds, from RetryInfo.retryDelay, ErrorInfo.metadata. * quotaResetDelay or a "retry after <duration>" message, in protobuf or Go duration syntax. */ export function extractGoogleRetryAfterSeconds(payloadText: string): string | undefined { const ms = googleErrorDetails(payloadText).retryDelayMs; return ms !== undefined && ms > 0 ? String(Math.ceil(ms / 1000)) : undefined; }