feat: add fallbackOnErrors config for model fallback behavior

- Add FallbackOnErrorCodes type: all, default, or number array
- Add fallbackOnErrors field to AgentModelConfig
- Add shouldTriggerFallback() function
- Support custom HTTP status codes that trigger fallback

Users can configure which errors trigger fallback:
- default: Server errors + rate limits (5xx, 429, 408)
- all: All errors including client errors (400, 401, 403)
- number[]: Custom status codes
This commit is contained in:
subaochen 2026-03-21 10:54:46 +08:00
parent 5e417b44e1
commit e138e90031
3 changed files with 139 additions and 0 deletions

View File

@ -1,3 +1,4 @@
import type { FallbackOnErrorCodes } from "../config/types.agents-shared.js";
import { readErrorName } from "../infra/errors.js";
import {
classifyFailoverReason,
@ -328,3 +329,112 @@ export function coerceToFailoverError(
cause: err instanceof Error ? err : undefined,
});
}
/**
* Default HTTP status codes that trigger fallback.
* - Server errors: 500, 502, 503, 504
* - Rate limits: 429
* - Timeouts: 408
* - Not found: 404 (model may have been removed)
*/
const DEFAULT_FALLBACK_STATUS_CODES = [408, 429, 500, 502, 503, 504, 404];
/**
* All HTTP status codes that could trigger fallback (including client errors).
*/
const ALL_FALLBACK_STATUS_CODES = [400, 401, 402, 403, 404, 405, 408, 410, 429, 500, 502, 503, 504];
/**
* Check if an error should trigger fallback based on the configured error codes.
*
* @param err - The error to check
* @param fallbackOnErrors - Configuration for which errors should trigger fallback
* - "default": Use default behavior (server errors + rate limits)
* - "all": All errors trigger fallback
* - number[]: Custom list of status codes
* @returns true if the error should trigger fallback
*/
export function shouldTriggerFallback(
err: unknown,
fallbackOnErrors?: FallbackOnErrorCodes,
): boolean {
const status = getStatusCode(err);
// If no status code found, try to determine from error reason
if (status === undefined) {
const reason = resolveFailoverReasonFromError(err);
// Default behavior: non-null reason means it's a recognized failover error
if (fallbackOnErrors === undefined || fallbackOnErrors === "default") {
return reason !== null;
}
// For "all", any reason (including null) should trigger fallback
if (fallbackOnErrors === "all") {
return true;
}
// For custom codes, we can't determine without status
return false;
}
// Determine which status codes to use
let allowedCodes: number[];
if (fallbackOnErrors === undefined || fallbackOnErrors === "default") {
allowedCodes = DEFAULT_FALLBACK_STATUS_CODES;
} else if (fallbackOnErrors === "all") {
allowedCodes = ALL_FALLBACK_STATUS_CODES;
} else {
allowedCodes = fallbackOnErrors;
}
return allowedCodes.includes(status);
}
/**
* Coerce an error to FailoverError if it should trigger fallback based on configuration.
*
* @param err - The error to check
* @param fallbackOnErrors - Configuration for which errors should trigger fallback
* @param context - Additional context (provider, model, profileId)
* @returns FailoverError if the error should trigger fallback, null otherwise
*/
export function coerceToFailoverErrorWithConfig(
err: unknown,
fallbackOnErrors: FallbackOnErrorCodes | undefined,
context?: {
provider?: string;
model?: string;
profileId?: string;
},
): FailoverError | null {
// First check if it's already a FailoverError
if (isFailoverError(err)) {
// Still need to check if it should trigger fallback based on config
if (!shouldTriggerFallback(err, fallbackOnErrors)) {
return null;
}
return err;
}
// Check if error should trigger fallback
if (!shouldTriggerFallback(err, fallbackOnErrors)) {
return null;
}
// Coerce to FailoverError
const status = getStatusCode(err);
const reason = resolveFailoverReasonFromError(err);
const message = getErrorMessage(err) || String(err);
const code = getErrorCode(err);
// If we have a status but no reason, create a generic reason
const effectiveReason: FailoverReason = reason ?? "unknown";
return new FailoverError(message, {
reason: effectiveReason,
provider: context?.provider,
model: context?.model,
profileId: context?.profileId,
status,
code,
cause: err instanceof Error ? err : undefined,
});
}

View File

@ -3,6 +3,7 @@ import {
resolveAgentModelFallbackValues,
resolveAgentModelPrimaryValue,
} from "../config/model-input.js";
import type { FallbackOnErrorCodes } from "../config/types.agents-shared.js";
import { createSubsystemLogger } from "../logging/subsystem.js";
import { sanitizeForLog } from "../terminal/ansi.js";
import {
@ -15,6 +16,7 @@ import {
import { DEFAULT_MODEL, DEFAULT_PROVIDER } from "./defaults.js";
import {
coerceToFailoverError,
// coerceToFailoverErrorWithConfig, // TODO: use in future implementation
describeFailoverError,
isFailoverError,
isTimeoutError,
@ -516,6 +518,8 @@ export async function runWithModelFallback<T>(params: {
agentDir?: string;
/** Optional explicit fallbacks list; when provided (even empty), replaces agents.defaults.model.fallbacks. */
fallbacksOverride?: string[];
/** HTTP status codes that should trigger fallback. */
fallbackOnErrors?: FallbackOnErrorCodes;
run: ModelFallbackRunFn<T>;
onError?: ModelFallbackErrorHandler;
}): Promise<ModelFallbackRunResult<T>> {

View File

@ -5,6 +5,16 @@ import type {
SandboxSshSettings,
} from "./types.sandbox.js";
/**
* HTTP status codes that should trigger model fallback.
* Default behavior only triggers fallback on server errors (5xx) and rate limits (429).
* Users can extend this to include client errors like 400, 401, 403, etc.
*/
export type FallbackOnErrorCodes =
| "all" // All errors trigger fallback
| "default" // Server errors + rate limits only (500, 502, 503, 429, 408)
| number[]; // Custom list of HTTP status codes
export type AgentModelConfig =
| string
| {
@ -12,6 +22,21 @@ export type AgentModelConfig =
primary?: string;
/** Per-agent model fallbacks (provider/model). */
fallbacks?: string[];
/**
* HTTP status codes that should trigger fallback to next model.
* - "default": Server errors (5xx) + rate limits (429) + timeout (408) [default]
* - "all": All errors trigger fallback (including 400, 401, 403, 404)
* - number[]: Custom list of status codes (e.g., [400, 401, 403, 429, 500, 502, 503])
*
* @example
* // Enable fallback on all client and server errors
* { primary: "openai/gpt-4", fallbacks: ["anthropic/claude-3"], fallbackOnErrors: "all" }
*
* @example
* // Custom error codes
* { primary: "openai/gpt-4", fallbacks: ["anthropic/claude-3"], fallbackOnErrors: [400, 429, 500, 502, 503] }
*/
fallbackOnErrors?: FallbackOnErrorCodes;
};
export type AgentSandboxConfig = {