Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

fix: forward the upstream context_length_exceeded to the Host (#731)

Alfredo Lopezd0dffe857c25
diff --git a/codex-llm-proxy/src/codexTransport.ts b/codex-llm-proxy/src/codexTransport.ts
index 32df9d60c..82dee9e4b 100644
--- a/codex-llm-proxy/src/codexTransport.ts
+++ b/codex-llm-proxy/src/codexTransport.ts
@@ -329,6 +329,12 @@ function toUpstreamPayload(
return payload
}
+// The one upstream failure code forwarded to the Host. A context-window refusal
+// is a property of this request, so retrying it cannot succeed (#731). Every
+// other upstream code stays `provider_unavailable`: an upstream string never
+// becomes a Host-visible code by itself.
+const CONTEXT_LENGTH_EXCEEDED = 'context_length_exceeded'
+
async function consumeSse(
body: ReadableStream<Uint8Array>,
onFrame: FrameSink | undefined,
@@ -370,6 +376,7 @@ async function consumeSse(
let buffer = ''
let completed = false
let failed = false
+ let contextLengthExceeded = false
let usage: SafeUsage | undefined
const maxSseBufferBytes = 1_048_576
try {
@@ -388,6 +395,7 @@ async function consumeSse(
if (mapped.usage) usage = mapped.usage
if (mapped.completed) completed = true
if (mapped.failed) failed = true
+ if (mapped.failureCode === CONTEXT_LENGTH_EXCEEDED) contextLengthExceeded = true
}
if (signal.aborted) break
}
@@ -398,6 +406,7 @@ async function consumeSse(
if (mapped.usage) usage = mapped.usage
if (mapped.completed) completed = true
if (mapped.failed) failed = true
+ if (mapped.failureCode === CONTEXT_LENGTH_EXCEEDED) contextLengthExceeded = true
}
} catch (error) {
await reader.cancel().catch(() => undefined)
@@ -413,6 +422,9 @@ async function consumeSse(
}
if (signal.aborted) return { outcome: 'canceled', usage }
if (failed) {
+ if (contextLengthExceeded) {
+ throw new CodexTransportError(CONTEXT_LENGTH_EXCEEDED, 'upstream context window exceeded')
+ }
throw new CodexTransportError('provider_unavailable', 'upstream response failed')
}
if (completed) {
@@ -479,6 +491,7 @@ function mapUpstreamEvent(
usage?: SafeUsage
completed?: boolean
failed?: boolean
+ failureCode?: string
} {
if (!event || typeof event !== 'object') return {}
const row = event as Record<string, unknown>
@@ -520,11 +533,23 @@ function mapUpstreamEvent(
return { completed: true, usage: parseSafeUsage(response.usage) }
}
if (type === 'response.failed' || type === 'response.incomplete' || type === 'error') {
- return { failed: true }
+ return { failed: true, failureCode: upstreamFailureCode(row) }
}
return {}
}
+// The upstream puts the code on `error.code` for an `error` event and on
+// `response.error.code` for `response.failed` / `response.incomplete`.
+function upstreamFailureCode(row: Record<string, unknown>): string | undefined {
+ const response = isPlainObject(row.response) ? row.response : undefined
+ const error = isPlainObject(row.error)
+ ? row.error
+ : response && isPlainObject(response.error)
+ ? response.error
+ : undefined
+ return typeof error?.code === 'string' ? error.code : undefined
+}
+
function upsertPendingTool(
pending: Map<string, PendingToolCall>,
source: Record<string, unknown> & { append?: boolean }
diff --git a/codex-llm-proxy/src/server.ts b/codex-llm-proxy/src/server.ts
index a49630b56..281e61b4f 100644
--- a/codex-llm-proxy/src/server.ts
+++ b/codex-llm-proxy/src/server.ts
@@ -433,6 +433,8 @@ export function startProxy(config: CodexLlmProxyConfig): ProxyServers {
// bounded by the proxy.
const ATTEMPT_ERROR_STATUS: Record<string, number> = {
invalid_request: 400,
+ // The upstream's own status class for a request over the context window (#731).
+ context_length_exceeded: 400,
Unauthorized: 401,
origin_denied: 403,
request_hash_mismatch: 403,
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.json
index 9333d1607..08edaafaf 100644
--- a/mcp-host/package-lock.json
+++ b/mcp-host/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.402",
+ "version": "0.1.403",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-mcp-host",
- "version": "0.1.402",
+ "version": "0.1.403",
"license": "SEE LICENSE IN LICENSE",
"dependencies": {
"@anthropic-ai/sdk": "^0.32.0",
diff --git a/mcp-host/package.json b/mcp-host/package.json
index 5f975bdd1..4f32a9e72 100644
--- a/mcp-host/package.json
+++ b/mcp-host/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.402",
+ "version": "0.1.403",
"description": "MCP Host - reads Host CRD and provides LLM access via OpenAI/Claude",
"main": "dist/main.js",
"engines": {
diff --git a/mcp-host/src/llm/codexSubscription.ts b/mcp-host/src/llm/codexSubscription.ts
index 0c25af701..a7d9959bb 100644
--- a/mcp-host/src/llm/codexSubscription.ts
+++ b/mcp-host/src/llm/codexSubscription.ts
@@ -233,8 +233,14 @@ export class CodexSubscriptionProvider implements SingleTurnProvider {
}
}
// `payload_too_large` is the proxy's 413 for an envelope over its body
- // limit: the same size refusal of this conversation, one hop later (#731).
- if (code === 'request_limit_exceeded' || code === 'payload_too_large') {
+ // limit: the same size refusal of this conversation, one hop later.
+ // `context_length_exceeded` is the upstream's refusal of a request over the
+ // model's context window, forwarded by the proxy (#731).
+ if (
+ code === 'request_limit_exceeded' ||
+ code === 'payload_too_large' ||
+ code === 'context_length_exceeded'
+ ) {
return {
code: LlmErrorCode.ContextLengthExceeded,
retryable: false,