Files
OmniRoute/open-sse/executors/base.ts
diegosouzapw 0c94c397db fix(review): resolve PR #2678 multi-agent review findings (#2743)
Addresses 3 critical + 4 high + 4 medium findings from the cross-agent
review of the v3.8.4 release branch.

CRITICAL
- combo: honour skipProviderBreaker in combo.ts:2452 so embedded service
  supervisor outages signalled via X-Omni-Fallback-Hint=connection_cooldown
  no longer trip the whole-provider circuit breaker. The G-02 contract was
  added to accountFallback but never honoured by its consumer.
- combo: per-model timeout now creates an AbortController, propagates its
  signal via target.modelAbortSignal, and aborts the inner request when
  the timeout wins the race. Chat.ts wraps the request via AbortSignal.any
  so downstream cooldown/breaker/usage mutations stop instead of running
  behind the routing decision's back.
- apiKey: getOrCreateApiKey now throws ServiceApiKeyDecryptError on
  decrypt failure instead of silently regenerating. Mutating embedded
  service auth without operator awareness made every subsequent request
  401 with no log trail.

HIGH
- base.ts proactive refresh: classify isUnrecoverableRefreshError before
  spreading the result so the executor doesn't send an
  unrecoverable_refresh_error sentinel object as the access token. Mark
  the connection expired via onCredentialsRefreshed and elevate the catch
  log from warn to error per the documented onPersist contract.
- kimi-coding: persist deviceId/deviceName/deviceModel/osVersion in
  providerSpecificData at login. tokenRefresh's fallback pbkdf2(refresh_token)
  rotates per refresh since Kimi rotates refresh tokens, contradicting the
  "stable deviceId" comment and tripping anti-bot detection mid-session.
- inner-ai: resolveModels throws InnerAiModelsError on non-OK (with 401/403
  invalidating the credential cache) instead of silently returning [].
  collectContent now propagates missing_credits / reached_limit /
  rate_limit_reached events via InnerAiStreamError so non-streaming
  callers get a 429 instead of HTTP 200 with an empty body.

MEDIUM
- chatCore.ts retry-after-refresh: capture and log the error at error
  level with sanitizeErrorMessage instead of a bare catch{}.
- gemini-cli.ts refreshCredentials: capture body on !response.ok and map
  invalid_grant to unrecoverable_refresh_error for parity with
  refreshGoogleToken in tokenRefresh.ts.
- usage.ts antigravity: introduce fractionReported sentinel so an
  upstream schema drift (Antigravity not reporting remainingFraction) no
  longer masquerades as "every model is exhausted".
- proxyFetch.ts vercel relay: sanitize the missing-relayAuth throw
  message (no internal [ProxyFetch] label) and pass host through
  proxyUrlForLogs for consistent redaction.

Backlog for follow-up: Inner.ai behavioural tests, tokenRefresh.ts
@ts-nocheck removal + RefreshResult discriminated union, tokenHealthCheck
tests, structural-vs-behavioural tests in token-refresh-race-comprehensive.
Tracked in #2743.
2026-05-26 12:44:22 -03:00

1067 lines
42 KiB
TypeScript

import { HTTP_STATUS, FETCH_TIMEOUT_MS } from "../config/constants.ts";
import { applyFingerprint, isCliCompatEnabled } from "../config/cliFingerprints.ts";
import { supportsXHighEffort } from "../config/providerModels.ts";
import {
getRotatingApiKey,
getValidApiKey,
resolveKeyForRequest,
} from "../services/apiKeyRotator.ts";
import type { KeyHealth } from "../services/apiKeyRotator.ts";
import { getOpenAICompatibleType, isClaudeCodeCompatible } from "../services/provider.ts";
import {
runWithOnPersist,
getRefreshLeadMs,
isUnrecoverableRefreshError,
} from "../services/tokenRefresh.ts";
import type { ProviderRequestDefaults } from "../services/providerRequestDefaults.ts";
import { signRequestBody } from "../services/claudeCodeCCH.ts";
import {
appendAnthropicBetaHeader,
CONTEXT_1M_BETA_HEADER,
modelSupportsContext1mBeta,
} from "../services/claudeCodeCompatible.ts";
import { getClaudeCodeCompatibleRequestDefaults } from "@/lib/providers/requestDefaults";
import { remapToolNamesInRequest } from "../services/claudeCodeToolRemapper.ts";
import { obfuscateInBody } from "../services/claudeCodeObfuscation.ts";
import { applySystemTransformPipeline, PROVIDER_CLAUDE } from "../services/systemTransforms.ts";
import {
fixToolPairs,
fixToolAdjacency,
stripTrailingAssistantOrphanToolUse,
} from "../services/contextManager.ts";
import { randomUUID } from "node:crypto";
import {
CLAUDE_CODE_VERSION,
CLAUDE_CODE_STAINLESS_VERSION,
buildHashFor,
buildUserIdJson,
getSessionId,
parseUpstreamMetadataUserId,
passthroughUpstreamSessionId,
resolveAccountUUID,
resolveCliUserID,
selectBetaFlags,
stainlessArch,
stainlessOS,
stainlessRuntimeVersion,
stripProxyToolPrefix,
} from "./claudeIdentity.ts";
/**
* Sanitizes a custom API path to prevent path traversal attacks.
* Valid paths must start with '/', contain no '..' segments,
* no null bytes, and be reasonable in length.
*/
function sanitizePath(path: string): boolean {
if (typeof path !== "string") return false;
if (!path.startsWith("/")) return false;
if (path.includes("\0")) return false; // null byte
if (path.includes("..")) return false; // path traversal
if (path.length > 512) return false; // sanity limit
return true;
}
type JsonRecord = Record<string, unknown>;
export type ProviderConfig = {
id?: string;
baseUrl?: string;
baseUrls?: string[];
responsesBaseUrl?: string;
chatPath?: string;
clientVersion?: string;
clientId?: string;
clientSecret?: string;
tokenUrl?: string;
refreshUrl?: string;
authUrl?: string;
headers?: Record<string, string>;
requestDefaults?: ProviderRequestDefaults;
timeoutMs?: number;
format?: string;
};
export type ProviderCredentials = {
accessToken?: string;
refreshToken?: string;
apiKey?: string;
projectId?: string | null;
expiresAt?: string;
connectionId?: string; // T07: used for API key rotation index
maxConcurrent?: number | null;
providerSpecificData?: JsonRecord;
requestEndpointPath?: string;
};
export type ExecutorLog = {
debug?: (tag: string, message: string) => void;
info?: (tag: string, message: string) => void;
warn?: (tag: string, message: string) => void;
error?: (tag: string, message: string) => void;
};
export type ExecuteInput = {
model: string;
body: unknown;
stream: boolean;
credentials: ProviderCredentials;
signal?: AbortSignal | null;
log?: ExecutorLog | null;
extendedContext?: boolean;
/** Merged after auth + CLI fingerprint headers (values override same-named defaults). */
upstreamExtraHeaders?: Record<string, string> | null;
/** Original client request headers (read-only). Executors may forward select headers upstream. */
clientHeaders?: Record<string, string> | null;
/** Callback to persist tokens that are proactively refreshed during execution.
* Accepts a partial credentials patch (e.g. `{ accessToken, refreshToken }` or
* `{ testStatus: "expired", isActive: false }`); the caller merges into the
* stored connection row. */
onCredentialsRefreshed?: (
newCredentials: Partial<ProviderCredentials> & Record<string, unknown>
) => Promise<void> | void;
/** When true, skip the intra-URL 429 retry in execute() so the caller handles fallback. */
skipUpstreamRetry?: boolean;
};
export type CountTokensInput = {
body: Record<string, unknown>;
credentials: ProviderCredentials;
log?: ExecutorLog | null;
model: string;
signal?: AbortSignal | null;
};
/** Apply model-level extra upstream headers (e.g. Authentication, X-Custom-Auth). */
export function mergeUpstreamExtraHeaders(
headers: Record<string, string>,
extra?: Record<string, string> | null
): void {
if (!extra) return;
for (const [k, v] of Object.entries(extra)) {
if (typeof k === "string" && k.length > 0 && typeof v === "string") {
if (k.toLowerCase() === "user-agent") {
setUserAgentHeader(headers, v);
continue;
}
headers[k] = v;
}
}
}
export function getCustomUserAgent(providerSpecificData?: JsonRecord | null): string | null {
const customUserAgent =
typeof providerSpecificData?.customUserAgent === "string"
? providerSpecificData.customUserAgent.trim()
: "";
return customUserAgent || null;
}
export function setUserAgentHeader(headers: Record<string, string>, userAgent: string): void {
headers["User-Agent"] = userAgent;
if ("user-agent" in headers) {
headers["user-agent"] = userAgent;
}
}
export function applyConfiguredUserAgent(
headers: Record<string, string>,
providerSpecificData?: JsonRecord | null
): void {
const customUserAgent = getCustomUserAgent(providerSpecificData);
if (customUserAgent) {
setUserAgentHeader(headers, customUserAgent);
}
}
export function mergeAbortSignals(primary: AbortSignal, secondary: AbortSignal): AbortSignal {
const controller = new AbortController();
const abortFrom = (source: AbortSignal) => {
if (!controller.signal.aborted) {
controller.abort(source.reason);
}
};
if (primary.aborted) {
abortFrom(primary);
return controller.signal;
}
if (secondary.aborted) {
abortFrom(secondary);
return controller.signal;
}
primary.addEventListener("abort", () => abortFrom(primary), { once: true });
secondary.addEventListener("abort", () => abortFrom(secondary), { once: true });
return controller.signal;
}
function hasActiveClaudeThinking(body: Record<string, unknown>): boolean {
const thinking = body.thinking as Record<string, unknown> | undefined;
return thinking?.type === "enabled" || thinking?.type === "adaptive";
}
/**
* Sanitize reasoning_effort for providers that don't accept all values.
*
* The claude→openai translator emits reasoning_effort=xhigh when the client
* sends output_config.effort=max on a Claude-shape request. Combined with
* runtime alias remapping (e.g. claude-opus-4-6 → mimo/mimo-v2.5-pro), this
* routes xhigh to OpenAI-shape providers that don't accept the value:
*
* xiaomi-mimo : low|medium|high only — 400 literal_error on xhigh
* mistral : devstral models reject reasoning_effort entirely
* github : claude/haiku/oswe models reject reasoning_effort entirely
*
* Each rejection burns a combo fallback attempt before reaching a working
* provider. Apply provider-aware sanitation here (after transformRequest, so
* reintroductions by per-provider transforms are also caught) before fetch.
* Models that genuinely support xhigh (registry flag supportsXHighEffort)
* pass through unchanged.
*/
const MISTRAL_NO_REASONING_EFFORT_PATTERN = /devstral/i;
const GITHUB_NO_REASONING_EFFORT_PATTERN = /(claude|haiku|oswe)/i;
export function sanitizeReasoningEffortForProvider(
body: unknown,
provider: string,
model: string | undefined,
log?: { info?: (tag: string, msg: string) => void } | null
): unknown {
if (!body || typeof body !== "object" || Array.isArray(body)) return body;
const b = body as Record<string, unknown>;
const reasoning =
b.reasoning && typeof b.reasoning === "object" && !Array.isArray(b.reasoning)
? (b.reasoning as Record<string, unknown>)
: null;
const effort = b.reasoning_effort ?? reasoning?.effort;
if (effort === undefined) return body;
const effortStr = typeof effort === "string" ? effort.toLowerCase() : "";
const modelStr = model || "";
if (effortStr === "xhigh" && !supportsXHighEffort(provider, modelStr)) {
log?.info?.(
"REASONING_SANITIZE",
`${provider}/${modelStr}: downgraded reasoning_effort xhigh → high`
);
const next: Record<string, unknown> = { ...b, reasoning_effort: "high" };
if (reasoning) {
next.reasoning = { ...reasoning, effort: "high" };
}
return next;
}
const rejecting =
(provider === "mistral" && MISTRAL_NO_REASONING_EFFORT_PATTERN.test(modelStr)) ||
(provider === "github" && GITHUB_NO_REASONING_EFFORT_PATTERN.test(modelStr));
if (rejecting) {
log?.info?.(
"REASONING_SANITIZE",
`${provider}/${modelStr}: removed unsupported reasoning_effort`
);
const next: Record<string, unknown> = { ...b };
delete next.reasoning_effort;
if (reasoning) {
const r = { ...reasoning };
delete r.effort;
if (Object.keys(r).length === 0) delete next.reasoning;
else next.reasoning = r;
}
return next;
}
return body;
}
/**
* BaseExecutor - Base class for provider executors.
* Implements the Strategy pattern: subclasses override specific methods
* (buildUrl, buildHeaders, transformRequest, etc.) for each provider.
*/
export class BaseExecutor {
provider: string;
config: ProviderConfig;
constructor(provider: string, config: ProviderConfig) {
this.provider = provider;
this.config = config;
}
getProvider() {
return this.provider;
}
getBaseUrls() {
return this.config.baseUrls || (this.config.baseUrl ? [this.config.baseUrl] : []);
}
getFallbackCount() {
return this.getBaseUrls().length || 1;
}
getTimeoutMs() {
const configured = this.config?.timeoutMs;
if (typeof configured !== "number" || !Number.isFinite(configured)) {
return FETCH_TIMEOUT_MS;
}
return Math.max(1, Math.floor(configured));
}
getCountTokensTimeoutMs() {
return this.getTimeoutMs();
}
buildUrl(
model: string,
stream: boolean,
urlIndex = 0,
credentials: ProviderCredentials | null = null
) {
void model;
void stream;
if (this.provider?.startsWith?.("openai-compatible-")) {
const psd = credentials?.providerSpecificData;
const baseUrl = typeof psd?.baseUrl === "string" ? psd.baseUrl : "https://api.openai.com/v1";
const normalized = baseUrl.replace(/\/$/, "");
// Sanitize custom path: must start with '/', no path traversal, no null bytes
const rawPath = typeof psd?.chatPath === "string" && psd.chatPath ? psd.chatPath : null;
const customPath = rawPath && sanitizePath(rawPath) ? rawPath : null;
if (customPath) return `${normalized}${customPath}`;
const path =
getOpenAICompatibleType(this.provider, psd) === "responses"
? "/responses"
: "/chat/completions";
return `${normalized}${path}`;
}
const baseUrls = this.getBaseUrls();
return baseUrls[urlIndex] || baseUrls[0] || this.config.baseUrl || "";
}
buildHeaders(
credentials: ProviderCredentials,
stream = true,
clientHeaders?: Record<string, string> | null,
model?: string,
health?: Record<string, KeyHealth>
): Record<string, string> {
void clientHeaders;
void model;
const headers: Record<string, string> = {
"Content-Type": "application/json",
...this.config.headers,
};
// Allow per-provider User-Agent override via environment variable.
// Example: CLAUDE_USER_AGENT="my-agent/2.0" overrides the default for the Claude provider.
const providerId = this.config?.id || this.provider;
if (providerId) {
const envKey = `${providerId.toUpperCase().replace(/[^A-Z0-9]/g, "_")}_USER_AGENT`;
const envUA = process.env[envKey]?.trim();
if (envUA) {
setUserAgentHeader(headers, envUA);
}
}
if (credentials.accessToken) {
headers["Authorization"] = `Bearer ${credentials.accessToken}`;
} else if (credentials.apiKey) {
const extraKeys =
(credentials.providerSpecificData?.extraApiKeys as string[] | undefined) ?? [];
const selectedKeyId = (
credentials.providerSpecificData as Record<string, unknown> | undefined
)?.selectedKeyId as string | undefined;
let effectiveKey = credentials.apiKey;
if (extraKeys.length > 0 && credentials.connectionId) {
const resolved = resolveKeyForRequest(
credentials.connectionId,
credentials.apiKey,
extraKeys,
selectedKeyId ?? null
);
effectiveKey = resolved?.key ?? credentials.apiKey;
if (resolved && credentials.providerSpecificData) {
(credentials.providerSpecificData as Record<string, unknown>).selectedKeyId =
resolved.keyId;
}
}
headers["Authorization"] = `Bearer ${effectiveKey}`;
}
headers["Accept"] = stream ? "text/event-stream" : "application/json";
return headers;
}
// Override in subclass for provider-specific transformations
transformRequest(
model: string,
body: unknown,
stream: boolean,
credentials: ProviderCredentials
): unknown {
void model;
void stream;
void credentials;
// Fix #1674: Remove empty string values from optional parameters
// like tool descriptions to avoid upstream validation failures.
if (body && typeof body === "object" && !Array.isArray(body)) {
const cloned = { ...body } as Record<string, unknown>;
if (Array.isArray(cloned.tools)) {
cloned.tools = cloned.tools.map((tool: unknown) => {
if (tool && typeof tool === "object" && !Array.isArray(tool)) {
const toolRecord = tool as JsonRecord;
const toolFunction = toolRecord.function;
if (toolFunction && typeof toolFunction === "object" && !Array.isArray(toolFunction)) {
const func = { ...(toolFunction as JsonRecord) };
if (func.description === "") delete func.description;
if (typeof func.name !== "string" || func.name.trim() === "") {
func.name = "unnamed_tool";
}
return { ...toolRecord, function: func };
}
}
return tool;
});
}
// Fix #1884: Cursor sends prompt_cache_retention which breaks strict upstream endpoints
delete cloned.prompt_cache_retention;
// Also clean up top level optional fields that commonly cause issues when empty
const optionalKeys = ["user", "stop", "seed", "response_format"];
for (const key of optionalKeys) {
if (cloned[key] === "") delete cloned[key];
}
return cloned;
}
return body;
}
shouldRetry(status: number, urlIndex: number) {
return status === HTTP_STATUS.RATE_LIMITED && urlIndex + 1 < this.getFallbackCount();
}
// Intra-URL retry config: retry same URL before falling back to next node
static readonly RETRY_CONFIG = { maxAttempts: 2, delayMs: 2000 };
// Timeout for receiving the initial upstream response headers. Once the response
// starts streaming, STREAM_IDLE_TIMEOUT_MS / Undici bodyTimeout handle stalls.
static FETCH_START_TIMEOUT_MS = FETCH_TIMEOUT_MS;
// Override in subclass for provider-specific refresh
async refreshCredentials(
credentials: ProviderCredentials,
log: ExecutorLog | null
): Promise<Partial<ProviderCredentials> | null> {
void credentials;
void log;
return null;
}
needsRefresh(credentials?: ProviderCredentials | null) {
if (!credentials?.expiresAt) return false;
const expiresAtMs = new Date(credentials.expiresAt).getTime();
// Use the provider-specific lead time (REFRESH_LEAD_MS) so rotating-token
// providers like Codex refresh proactively far ahead of expiry. Keeping the
// refresh_token "warm" prevents Auth0 from marking it as stale and revoking
// the token family on first use after long idle.
const lead = getRefreshLeadMs(this.provider);
return expiresAtMs - Date.now() < lead;
}
parseError(response: Response, bodyText: string) {
return { status: response.status, message: bodyText || `HTTP ${response.status}` };
}
buildCountTokensUrl(model: string, credentials: ProviderCredentials | null = null) {
void model;
void credentials;
const baseUrl = this.buildUrl(model, false, 0, credentials);
if (typeof baseUrl !== "string" || baseUrl.length === 0) return null;
if (this.config?.format !== "claude" || !baseUrl.includes("/messages")) return null;
const [path, query = ""] = baseUrl.split("?");
const normalizedPath = path.endsWith("/messages")
? `${path}/count_tokens`
: `${path}/count_tokens`;
return query ? `${normalizedPath}?${query}` : normalizedPath;
}
async countTokens({ model, body, credentials, signal, log }: CountTokensInput) {
const url = this.buildCountTokensUrl(model, credentials);
if (!url) return null;
const headers = this.buildHeaders(credentials, false);
const requestBody =
body && typeof body === "object"
? {
...body,
model,
}
: { model };
let timeoutId: ReturnType<typeof setTimeout> | null = null;
let activeSignal = signal || null;
let controller: AbortController | null = null;
const timeoutMs = this.getCountTokensTimeoutMs();
if (timeoutMs > 0) {
controller = new AbortController();
timeoutId = setTimeout(() => controller?.abort(), timeoutMs);
activeSignal = signal ? mergeAbortSignals(signal, controller.signal) : controller.signal;
}
try {
const response = await fetch(url, {
method: "POST",
headers,
body: JSON.stringify(requestBody),
signal: activeSignal || undefined,
});
const text = await response.text();
if (!response.ok) {
const parsedError = this.parseError(response, text);
throw new Error(parsedError.message);
}
const parsed = text ? JSON.parse(text) : {};
const inputTokens = Number(parsed?.input_tokens);
if (!Number.isFinite(inputTokens)) {
throw new Error("Provider count_tokens response missing input_tokens");
}
return { input_tokens: inputTokens, provider: this.provider, source: "provider" };
} catch (error) {
log?.debug?.(
"COUNT_TOKENS",
`${this.provider}/${model} real count unavailable: ${error instanceof Error ? error.message : String(error)}`
);
return null;
} finally {
if (timeoutId) clearTimeout(timeoutId);
}
}
async execute(input: ExecuteInput) {
const {
model,
body,
stream,
credentials,
signal,
log,
extendedContext,
upstreamExtraHeaders,
clientHeaders,
skipUpstreamRetry = false,
onCredentialsRefreshed,
} = input;
const fallbackCount = this.getFallbackCount();
let lastError: unknown = null;
let lastStatus = 0;
let activeCredentials = credentials;
// Track per-URL intra-retry attempts to avoid infinite loops
const retryAttemptsByUrl: Record<number, number> = {};
if (this.needsRefresh(credentials)) {
try {
// Fix A: wire onCredentialsRefreshed through runWithOnPersist so it runs
// INSIDE the per-connection mutex inside getAccessToken. Not every
// executor routes through getAccessToken (e.g. github.ts), so use a flag
// to detect whether the persist callback actually fired and fall back to
// post-refresh mutation when it didn't.
let proactivePersistRan = false;
const proactiveOnPersist = onCredentialsRefreshed
? async (refreshResult: Record<string, unknown>) => {
proactivePersistRan = true;
activeCredentials = {
...credentials,
...(refreshResult as Partial<ProviderCredentials>),
};
await onCredentialsRefreshed(refreshResult as Partial<ProviderCredentials>);
}
: null;
const refreshed = await runWithOnPersist(proactiveOnPersist, () =>
this.refreshCredentials(credentials, log || null)
);
if (refreshed && !proactivePersistRan) {
// Classify unrecoverable sentinels before spreading. Without this guard
// an { error: "unrecoverable_refresh_error", code } object passes the
// truthy check and pollutes activeCredentials with garbage, causing the
// executor to send a non-token to upstream. Mirrors the reactive path
// in chatCore.ts and keeps the proactive path's behaviour consistent.
if (isUnrecoverableRefreshError(refreshed)) {
const refreshCode = (refreshed as Record<string, unknown>).code;
log?.error?.(
"TOKEN",
`${this.provider.toUpperCase()} | unrecoverable refresh — marking account expired (code=${String(refreshCode ?? "unknown")})`
);
if (onCredentialsRefreshed) {
await onCredentialsRefreshed({
testStatus: "expired",
isActive: false,
});
}
// Don't spread the sentinel — keep stale credentials so the next
// upstream call surfaces the real auth error to the client.
} else {
activeCredentials = {
...credentials,
...refreshed,
};
if (onCredentialsRefreshed) {
await onCredentialsRefreshed(refreshed);
}
}
}
} catch (error) {
// tokenRefresh.ts:1352 documents that onPersist throws are re-thrown so
// the caller is aware of the persistence failure. Honor that contract:
// log at error level (not warn), with sanitized message — and let the
// request continue with stale credentials so the user-visible error
// surfaces upstream rather than being silently absorbed here.
log?.error?.(
"TOKEN",
`Credential refresh failed for ${this.provider}: ${error instanceof Error ? error.message : String(error)}`
);
}
}
for (let urlIndex = 0; urlIndex < fallbackCount; urlIndex++) {
const url = this.buildUrl(model, stream, urlIndex, activeCredentials);
const headers = this.buildHeaders(activeCredentials, stream, clientHeaders, model);
applyConfiguredUserAgent(headers, activeCredentials?.providerSpecificData);
const ccRequestDefaults = isClaudeCodeCompatible(this.provider)
? getClaudeCodeCompatibleRequestDefaults(activeCredentials?.providerSpecificData)
: {};
const shouldForwardExtendedContext =
extendedContext &&
modelSupportsContext1mBeta(model) &&
!isClaudeCodeCompatible(this.provider);
const shouldForwardCcCompatibleContext1m =
isClaudeCodeCompatible(this.provider) && ccRequestDefaults.context1m === true;
if (shouldForwardExtendedContext || shouldForwardCcCompatibleContext1m) {
appendAnthropicBetaHeader(headers, CONTEXT_1M_BETA_HEADER);
}
const rawTransformedBody = await this.transformRequest(
model,
body,
stream,
activeCredentials
);
const transformedBody = sanitizeReasoningEffortForProvider(
rawTransformedBody,
this.provider,
model,
log
);
try {
// Only enforce the timeout while waiting for the initial fetch() response.
// Once headers arrive, active streams must not be cut off by total elapsed time;
// post-start stalls are handled separately by STREAM_IDLE_TIMEOUT_MS / bodyTimeout.
const fetchStartTimeoutMs = this.getTimeoutMs();
const timeoutController = fetchStartTimeoutMs > 0 ? new AbortController() : null;
let timeoutId: ReturnType<typeof setTimeout> | null = null;
if (timeoutController) {
timeoutId = setTimeout(() => {
const timeoutError = new Error(
`Fetch timeout after ${fetchStartTimeoutMs}ms on ${url}`
);
timeoutError.name = "TimeoutError";
timeoutController.abort(timeoutError);
}, fetchStartTimeoutMs);
}
const timeoutSignal = timeoutController?.signal ?? null;
const combinedSignal =
signal && timeoutSignal
? mergeAbortSignals(signal, timeoutSignal)
: signal || timeoutSignal;
const isClaudeCodeClient =
clientHeaders?.["x-app"] === "cli" ||
(clientHeaders?.["user-agent"] &&
clientHeaders["user-agent"].toLowerCase().includes("claude-code")) ||
(clientHeaders?.["user-agent"] &&
clientHeaders["user-agent"].toLowerCase().includes("claude-cli"));
// Anthropic's user:sessions:claude_code OAuth scope expects CLI-shaped
// traffic. Apply the cloak whenever we have an OAuth token, regardless
// of upstream client.
const hasClaudeOAuthToken =
typeof activeCredentials?.accessToken === "string" &&
activeCredentials.accessToken.startsWith("sk-ant-oat") &&
!activeCredentials?.apiKey;
if (
this.provider === "claude" &&
(isClaudeCodeClient || hasClaudeOAuthToken) &&
typeof transformedBody === "object" &&
transformedBody !== null
) {
const tb = transformedBody as Record<string, unknown>;
stripProxyToolPrefix(tb);
remapToolNamesInRequest(tb);
obfuscateInBody(tb);
// NOTE (issue #2260): This is the native `claude` provider OAuth path.
// It is intentionally NOT routed through applyCcBridgeTransformPipeline.
// The native OAuth path already prepends its own billing line + sentinel
// (see lines ~744-773 below, dayStamp-based, cc_entrypoint=cli, cch=00000
// placeholder, signed at body level). The CC bridge transforms DSL is
// wired into buildAndSignClaudeCodeRequest (claudeCodeCompatible.ts step 5b)
// which is the anthropic-compatible-cc-* relay path — a different,
// separately classified surface. Do not double-prepend here.
// Real CLI never sets cache_control on tools.
if (Array.isArray(tb.tools)) {
for (const t of tb.tools as Array<Record<string, unknown>>) {
delete t.cache_control;
}
}
// Per-request behavior overrides via custom client headers.
// x-omniroute-effort: low | medium | high | xhigh | off
// x-omniroute-thinking: adaptive | off
// A header value applies only when the corresponding body field is
// not already set; "off" force-strips the field.
const headerEffort = (
clientHeaders?.["x-omniroute-effort"] ?? clientHeaders?.["X-OmniRoute-Effort"]
)
?.trim()
.toLowerCase();
const headerThinking = (
clientHeaders?.["x-omniroute-thinking"] ?? clientHeaders?.["X-OmniRoute-Thinking"]
)
?.trim()
.toLowerCase();
let appliedEffort: string | null = null;
let appliedThinking: string | null = null;
if (headerEffort === "off") {
if (tb.output_config && typeof tb.output_config === "object") {
delete (tb.output_config as Record<string, unknown>).effort;
}
appliedEffort = "off";
} else if (headerEffort && ["low", "medium", "high", "xhigh"].includes(headerEffort)) {
const oc =
tb.output_config && typeof tb.output_config === "object"
? (tb.output_config as Record<string, unknown>)
: {};
if (oc.effort === undefined) {
oc.effort = headerEffort;
tb.output_config = oc;
appliedEffort = headerEffort;
}
}
if (headerThinking === "adaptive") {
if (tb.thinking === undefined) {
tb.thinking = { type: "adaptive" };
appliedThinking = "adaptive";
}
if (tb.context_management === undefined) {
tb.context_management = {
edits: [{ type: "clear_thinking_20251015", keep: "all" }],
};
}
} else if (headerThinking === "off") {
delete tb.thinking;
delete tb.context_management;
appliedThinking = "off";
} else if (!headerThinking && !headerEffort) {
// Default CC logic when no override headers are present
const isHaiku = typeof tb.model === "string" && tb.model.includes("haiku");
if (isHaiku) {
// Keep tb.thinking — real Claude Desktop keeps thinking enabled for Haiku
// (issue #2454). Only strip output_config (effort) which Haiku rejects;
// context_management is re-paired with the preserved thinking below.
delete tb.output_config;
delete tb.context_management;
} else if (tb.thinking === undefined && tb.output_config === undefined) {
tb.thinking = { type: "adaptive" };
tb.context_management = {
edits: [{ type: "clear_thinking_20251015", keep: "all" }],
};
tb.output_config = { effort: "high" };
}
}
// Real CLI always pairs context_management with thinking. Mirror
// that invariant so long sessions don't accumulate thinking blocks
// toward the context cap.
if (hasActiveClaudeThinking(tb) && !tb.context_management) {
tb.context_management = {
edits: [{ type: "clear_thinking_20251015", keep: "all" }],
};
}
const seed = activeCredentials?.accessToken || activeCredentials?.apiKey || "anon";
const psd = activeCredentials?.providerSpecificData as
| Record<string, unknown>
| undefined;
let identitySource:
| "upstream-metadata"
| "upstream-header"
| "synthesized"
| "synthesized-cloaked" = "synthesized";
let sessionId: string;
let deviceId: string;
let accountUUID: string;
// For any Claude OAuth request, ignore client-supplied metadata.user_id /
// X-Claude-Code-Session-Id and synthesize per-account: the CC device_id from
// ~/.claude.json is shared across every account on one machine, which lets
// Anthropic correlate accounts behind one OmniRoute.
const cloakIdentity = isClaudeCodeClient || hasClaudeOAuthToken;
const upstreamUserId = cloakIdentity ? null : parseUpstreamMetadataUserId(tb);
if (upstreamUserId) {
sessionId = upstreamUserId.session_id;
deviceId = upstreamUserId.device_id;
accountUUID = upstreamUserId.account_uuid;
identitySource = "upstream-metadata";
} else {
const headerSid = cloakIdentity
? null
: passthroughUpstreamSessionId(
clientHeaders as Record<string, string | undefined> | undefined
);
sessionId = headerSid ?? getSessionId(seed);
deviceId = resolveCliUserID(psd, seed);
accountUUID = resolveAccountUUID(psd, seed, activeCredentials?.accessToken);
identitySource = headerSid
? "upstream-header"
: cloakIdentity
? "synthesized-cloaked"
: "synthesized";
}
// system[0] (billing) and system[1] (sentinel) must not carry
// cache_control — that belongs on upstream prompt blocks at [2..].
const dayStamp = new Date().toISOString().slice(0, 10);
const buildHash = buildHashFor(CLAUDE_CODE_VERSION, dayStamp);
const billingLine = `x-anthropic-billing-header: cc_version=${CLAUDE_CODE_VERSION}.${buildHash}; cc_entrypoint=cli; cch=00000;`;
const SENTINEL = "You are Claude Code, Anthropic's official CLI for Claude.";
const sysBlocks: Array<Record<string, unknown>> = Array.isArray(tb.system)
? (tb.system as Array<Record<string, unknown>>)
: typeof tb.system === "string"
? [{ type: "text", text: tb.system }]
: [];
// Strip any pre-existing billing/sentinel before re-prepending — keeps
// retries idempotent and avoids stacking that breaks prompt-cache prefix
// matching (see issue #1712).
for (let i = sysBlocks.length - 1; i >= 0; i--) {
const t = sysBlocks[i]?.text;
if (typeof t === "string" && t.startsWith("x-anthropic-billing-header:")) {
sysBlocks.splice(i, 1);
}
}
for (let i = sysBlocks.length - 1; i >= 0; i--) {
const t = sysBlocks[i]?.text;
if (typeof t === "string" && t.startsWith(SENTINEL)) {
sysBlocks.splice(i, 1);
}
}
sysBlocks.unshift({ type: "text", text: billingLine }, { type: "text", text: SENTINEL });
tb.system = sysBlocks;
// Run the configurable system-transforms pipeline for the native
// `claude` provider (issue #2260 / comment 4459544580). The default
// claude pipeline runs cosmetic ops only (Open WebUI paragraph
// anchors, identity-prefix paragraph drop, ZWJ obfuscation of
// sensitive words). It deliberately does NOT include
// `inject_billing_header` — billing + sentinel are already
// prepended above. Users can extend the pipeline via Settings UI.
{
const transformResult = applySystemTransformPipeline(PROVIDER_CLAUDE, tb);
if (transformResult.appliedOpKinds.length > 0) {
console.log(
`[SystemTransforms] claude-native: ${transformResult.appliedOpKinds.join(", ")}`
);
}
}
if (!tb.metadata || typeof tb.metadata !== "object") tb.metadata = {};
(tb.metadata as Record<string, unknown>).user_id = buildUserIdJson({
deviceId,
accountUUID,
sessionId,
});
// Headers. Accept stays application/json even on streams (Stainless
// convention; SSE decoding is gated on body.stream). anthropic-beta
// is selected per request shape; the full set on a quota probe is
// itself a fingerprint.
const ccHeaders: Record<string, string> = {
Accept: "application/json",
"anthropic-version": "2023-06-01",
"anthropic-beta": selectBetaFlags(tb),
"anthropic-dangerous-direct-browser-access": "true",
"x-app": "cli",
"User-Agent": `claude-cli/${CLAUDE_CODE_VERSION} (external, cli)`,
"X-Stainless-Package-Version": CLAUDE_CODE_STAINLESS_VERSION,
"X-Stainless-Timeout": "600",
"accept-encoding": "gzip, deflate, br, zstd",
connection: "keep-alive",
"x-client-request-id": randomUUID(),
"X-Claude-Code-Session-Id": sessionId,
};
// Drop case variants of the same header name before merging — undici
// would otherwise concatenate them (issue #1454).
const ccKeysLower = new Set(Object.keys(ccHeaders).map((k) => k.toLowerCase()));
for (const key of Object.keys(headers)) {
if (ccKeysLower.has(key.toLowerCase())) delete headers[key];
}
Object.assign(headers, ccHeaders);
delete headers["X-Stainless-Helper-Method"];
// Stainless OS/Arch/Runtime are host-derived (Stainless SDK does the
// same at runtime). Hardcoding them was a unique-per-deployment tell.
headers["X-Stainless-Arch"] = stainlessArch();
headers["X-Stainless-Lang"] = "js";
headers["X-Stainless-OS"] = stainlessOS();
headers["X-Stainless-Runtime"] = "node";
headers["X-Stainless-Runtime-Version"] = stainlessRuntimeVersion();
headers["X-Stainless-Retry-Count"] = "0";
delete headers["X-Stainless-Os"];
const overrideTag =
appliedEffort || appliedThinking
? ` overrides=effort:${appliedEffort ?? "-"},thinking:${appliedThinking ?? "-"}`
: "";
log?.debug?.(
"CLAUDE",
`identity=${identitySource} sid=${sessionId.slice(0, 8)} dev=${deviceId.slice(0, 8)} acct=${accountUUID.slice(0, 8)}${overrideTag}`
);
}
// CLI fingerprint ordering — always-on for native Claude OAuth, opt-in
// for other providers. Header + body field order is itself a fingerprint.
let finalHeaders = headers;
// Strip internal sentinel fields set by remapToolNamesInRequest before
// serializing — Anthropic rejects unknown top-level fields (issue #2260).
delete (transformedBody as Record<string, unknown>)[
"_claudeCodeRequiresLowercaseToolNames"
];
// Guard against orphan tool_use / tool_result pairs. Clients can ship
// truncated histories mid-tool-call which Anthropic rejects with
// `messages.N: tool_use ids were found without tool_result blocks
// immediately after: toolu_...`. fixToolPairs strips orphans, then
// stripTrailingAssistantOrphanToolUse catches the case where the
// request body itself ends on an unmatched assistant(tool_use) —
// invalid for an upstream-send turn since the body must end on a
// user message. Both are idempotent on clean histories.
{
const tb = transformedBody as Record<string, unknown>;
if (Array.isArray(tb?.messages)) {
const fixed = fixToolPairs(tb.messages as Record<string, unknown>[]);
// fixToolAdjacency enforces Claude's strict adjacency rule
// (tool_result must be in immediately next message).
// Only apply for Claude/Claude-compatible — OpenAI allows results
// spread across multiple subsequent messages.
const isClaude = this.provider === "claude" || isClaudeCodeCompatible(this.provider);
// For Claude, fixToolAdjacency may strip tool_use blocks whose
// tool_result isn't in the next message; re-run fixToolPairs to
// drop any tool_result orphaned by that strip (discussion #2410).
const adjacent = isClaude ? fixToolPairs(fixToolAdjacency(fixed)) : fixed;
tb.messages = stripTrailingAssistantOrphanToolUse(adjacent);
}
}
let bodyString = JSON.stringify(transformedBody);
const shouldFingerprint =
isCliCompatEnabled(this.provider) ||
(this.provider === "claude" && (isClaudeCodeClient || hasClaudeOAuthToken));
if (shouldFingerprint) {
const fingerprinted = applyFingerprint(this.provider, headers, transformedBody);
finalHeaders = fingerprinted.headers;
bodyString = fingerprinted.bodyString;
}
// CCH signing — replaces the cch=00000 placeholder in the billing
// header with an xxHash64 integrity token over the serialized body.
if (isClaudeCodeCompatible(this.provider) || this.provider === "claude") {
bodyString = await signRequestBody(bodyString);
}
mergeUpstreamExtraHeaders(finalHeaders, upstreamExtraHeaders);
const fetchOptions: RequestInit = {
method: "POST",
headers: finalHeaders,
body: bodyString,
};
if (combinedSignal) fetchOptions.signal = combinedSignal;
let response;
try {
response = await fetch(url, fetchOptions);
} finally {
if (timeoutId) {
clearTimeout(timeoutId);
timeoutId = null;
}
}
// Intra-URL retry: if 429 and we haven't exhausted per-URL retries, wait and retry the same URL
if (
!skipUpstreamRetry &&
response.status === HTTP_STATUS.RATE_LIMITED &&
(retryAttemptsByUrl[urlIndex] ?? 0) < BaseExecutor.RETRY_CONFIG.maxAttempts
) {
retryAttemptsByUrl[urlIndex] = (retryAttemptsByUrl[urlIndex] ?? 0) + 1;
const attempt = retryAttemptsByUrl[urlIndex];
log?.debug?.(
"RETRY",
`429 intra-retry ${attempt}/${BaseExecutor.RETRY_CONFIG.maxAttempts} on ${url} — waiting ${BaseExecutor.RETRY_CONFIG.delayMs}ms`
);
await new Promise((resolve) => setTimeout(resolve, BaseExecutor.RETRY_CONFIG.delayMs));
urlIndex--; // re-run this urlIndex on the next loop iteration
continue;
}
// T07: Handle 401 authentication errors — log and continue to fallback
if (response.status === 401 && credentials.connectionId && credentials.apiKey) {
log?.warn?.("AUTH", `401 on ${url} - API key may be invalid`);
}
if (!skipUpstreamRetry && this.shouldRetry(response.status, urlIndex)) {
log?.debug?.("RETRY", `${response.status} on ${url}, trying fallback ${urlIndex + 1}`);
lastStatus = response.status;
continue;
}
return { response, url, headers: finalHeaders, transformedBody };
} catch (error) {
// Distinguish timeout errors from other abort errors
const err = error instanceof Error ? error : new Error(String(error));
if (err.name === "TimeoutError") {
log?.warn?.("TIMEOUT", `Fetch timeout after ${this.getTimeoutMs()}ms on ${url}`);
}
lastError = err;
if (!skipUpstreamRetry && urlIndex + 1 < fallbackCount) {
log?.debug?.("RETRY", `Error on ${url}, trying fallback ${urlIndex + 1}`);
continue;
}
throw err;
}
}
throw lastError || new Error(`All ${fallbackCount} URLs failed with status ${lastStatus}`);
}
}
export default BaseExecutor;