Skip to content

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

TypeScript

test(mcp-host): size J1's context window so it detects a word-count regression (#731)

Alfredo Lopezb0a78e114ab3
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.json
index 08edaafaf..0c86f97a4 100644
--- a/mcp-host/package-lock.json
+++ b/mcp-host/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.403",
+ "version": "0.1.404",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-mcp-host",
- "version": "0.1.403",
+ "version": "0.1.404",
"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 4f32a9e72..a7d89d34e 100644
--- a/mcp-host/package.json
+++ b/mcp-host/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.403",
+ "version": "0.1.404",
"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/__tests__/contextOverrunJourney.test.ts b/mcp-host/src/llm/__tests__/contextOverrunJourney.test.ts
index 1b7ecdc62..0474716d8 100644
--- a/mcp-host/src/llm/__tests__/contextOverrunJourney.test.ts
+++ b/mcp-host/src/llm/__tests__/contextOverrunJourney.test.ts
@@ -133,6 +133,13 @@ function unshrinkableHistory(finalResultBytes: number): ChatMessage[] {
*/
const J1_RESULT_BYTES = 35_000
const J1_TURNS = Math.ceil((LIMITS.maxRequestBodyBytes * 1.1) / J1_RESULT_BYTES)
+/**
+ * J1's window sits between the two estimates of that history: the byte-based
+ * count fills it about 2.3 times, the word-based count #731 replaced about 0.55
+ * times. With a 100K window both overrun it and pre-prune runs either way, so
+ * J1 could not tell a regressed tokenizer from the fixed one.
+ */
+const J1_CONTEXT_WINDOW_TOKENS = 1_000_000
function mcpHeavyHistory(turns: number, bytesPerResult: number): ChatMessage[] {
const msgs: ChatMessage[] = [{ role: 'system', content: 'You are a helpful assistant.' }]
@@ -175,18 +182,24 @@ describe('#731 context-overrun journey', () => {
// Built the way `taskExecutor` builds it: pre-prune on or off, and its
// options, come from the deployed configuration.
- const manager = new PressureContextManager(100000, undefined, undefined, undefined, {
- prePruneEnabled: appConfig.compactionPrePruneEnabled,
- prePruneOptions: {
- protectedTailTurns: appConfig.compactionPrePruneProtectedTailTurns,
- summaryThresholdTokens: appConfig.compactionPrePruneSummaryTokens,
- maxArgsBytes: appConfig.compactionPrePruneMaxArgsBytes,
- dedupEnabled: appConfig.compactionPrePruneDedup,
- oneLineSummariesEnabled: appConfig.compactionPrePruneOneLine,
- jsonSafeTruncateEnabled: appConfig.compactionPrePruneJsonTruncate,
- stripMediaEnabled: appConfig.compactionPrePruneStripMedia,
- },
- })
+ const manager = new PressureContextManager(
+ J1_CONTEXT_WINDOW_TOKENS,
+ undefined,
+ undefined,
+ undefined,
+ {
+ prePruneEnabled: appConfig.compactionPrePruneEnabled,
+ prePruneOptions: {
+ protectedTailTurns: appConfig.compactionPrePruneProtectedTailTurns,
+ summaryThresholdTokens: appConfig.compactionPrePruneSummaryTokens,
+ maxArgsBytes: appConfig.compactionPrePruneMaxArgsBytes,
+ dedupEnabled: appConfig.compactionPrePruneDedup,
+ oneLineSummariesEnabled: appConfig.compactionPrePruneOneLine,
+ jsonSafeTruncateEnabled: appConfig.compactionPrePruneJsonTruncate,
+ stripMediaEnabled: appConfig.compactionPrePruneStripMedia,
+ },
+ }
+ )
const wired = deps()
const provider = new CodexSubscriptionProvider('gpt-5.3-codex', wired as never)