Skip to content

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

TypeScript

fix(grok): classify the tool-call and message limits as contract refusals

Alfredo Lopez76e0cbc93c38
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.json
index fb023890b..8ee570bfc 100644
--- a/mcp-host/package-lock.json
+++ b/mcp-host/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.363",
+ "version": "0.1.364",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-mcp-host",
- "version": "0.1.363",
+ "version": "0.1.364",
"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 2fc9ce550..81e89ac35 100644
--- a/mcp-host/package.json
+++ b/mcp-host/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.363",
+ "version": "0.1.364",
"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/grokSubscription.ts b/mcp-host/src/llm/grokSubscription.ts
index b3f75f40a..f28aa0cff 100644
--- a/mcp-host/src/llm/grokSubscription.ts
+++ b/mcp-host/src/llm/grokSubscription.ts
@@ -68,7 +68,7 @@ function assertTerminalGrokOutcome(result: {
outcome: 'success' | 'canceled' | 'error' | 'unknown'
}): void {
if (result.toolCalls.length > LIMITS.maxToolCalls) {
- throw new GrokProxyError('provider_unavailable', `tool calls exceed ${LIMITS.maxToolCalls}`)
+ throw new GrokProxyError('tool_call_limit_exceeded', `tool calls exceed ${LIMITS.maxToolCalls}`)
}
if (result.toolCalls.length > 0 && result.outcome !== 'success') {
throw new GrokProxyError(
@@ -184,6 +184,28 @@ export class GrokSubscriptionProvider implements SingleTurnProvider {
...(providerDispatched !== undefined ? { providerDispatched } : {}),
}
}
+ if (code === 'tool_call_limit_exceeded') {
+ // Not retryable: `retryable: true` would make this error
+ // failover-eligible and enable the workflow fallback after tool
+ // results, re-running the turn on another provider. The limit is a
+ // contract rejection of the model output, not a provider outage.
+ return {
+ code: LlmErrorCode.ToolCallLimitExceeded,
+ retryable: false,
+ message: err instanceof Error ? err.message : String(err),
+ providerCode: code,
+ ...(providerDispatched !== undefined ? { providerDispatched } : {}),
+ }
+ }
+ if (code === 'request_limit_exceeded') {
+ return {
+ code: LlmErrorCode.ContextLengthExceeded,
+ retryable: false,
+ message: err instanceof Error ? err.message : String(err),
+ providerCode: code,
+ ...(providerDispatched !== undefined ? { providerDispatched } : {}),
+ }
+ }
if (code === 'provider_unavailable' || code === 'connection_unavailable') {
return {
code: LlmErrorCode.ModelOverloaded,
@@ -264,6 +286,14 @@ export class GrokSubscriptionProvider implements SingleTurnProvider {
if (options?.signal?.aborted) {
throw new GrokProxyError('canceled', 'aborted before authorize', false)
}
+ // An over-long history is reported as a context-length failure instead of
+ // a generic invalid request; it is thrown before authorize and dispatch.
+ if (messages.length > LIMITS.maxMessages) {
+ throw new CodexAuthorizeError(
+ 'request_limit_exceeded',
+ `messages exceed ${LIMITS.maxMessages}`
+ )
+ }
// Hash and send the validated wire projection — exactly what control-api
// authorize and the proxy re-derive. Hashing the locally built object let
// shapes the parser normalizes away (an empty `generation` from a