Skip to content

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

TypeScript

test(mcp-host): require pre-prune to wait for the window-derived threshold (#731)

Alfredo Lopez72af06c08144
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.json
index 400669dc5..ed84de275 100644
--- a/mcp-host/package-lock.json
+++ b/mcp-host/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.386",
+ "version": "0.1.387",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-mcp-host",
- "version": "0.1.386",
+ "version": "0.1.387",
"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 b27bfac6c..8a5bc9f79 100644
--- a/mcp-host/package.json
+++ b/mcp-host/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.386",
+ "version": "0.1.387",
"description": "MCP Host - reads Host CRD and provides LLM access via OpenAI/Claude",
"main": "dist/main.js",
"engines": {
diff --git a/mcp-host/src/agent/__tests__/taskExecutor.test.ts b/mcp-host/src/agent/__tests__/taskExecutor.test.ts
index 0bcb99ea7..e469d0c2f 100644
--- a/mcp-host/src/agent/__tests__/taskExecutor.test.ts
+++ b/mcp-host/src/agent/__tests__/taskExecutor.test.ts
@@ -2339,3 +2339,75 @@ describe('TaskExecutor context manager message bound (#731)', () => {
expect(managed).toBe(msgs)
})
})
+
+describe('TaskExecutor history compaction threshold follows the context window (#731)', () => {
+ // ~100k tokens by either count (tiktoken reads one token per `word `, the
+ // byte heuristic 125k): above the 80k literal default, below 0.8 of a 1M
+ // window. The old tool result sits outside the protected tail of three turns.
+ const OLD_RESULT = 'word '.repeat(100_000)
+ function prunableHistory(): ChatMessage[] {
+ return [
+ { role: 'system', content: 'sys' },
+ { role: 'user', content: 'q0' },
+ {
+ role: 'assistant',
+ content: '',
+ tool_calls: [{ id: 'tc0', name: 'fetch', arguments: { url: 'https://example.test' } }],
+ },
+ { role: 'tool', tool_call_id: 'tc0', name: 'fetch', content: OLD_RESULT },
+ { role: 'user', content: 'q1' },
+ { role: 'assistant', content: 'a1' },
+ { role: 'user', content: 'q2' },
+ { role: 'assistant', content: 'a2' },
+ { role: 'user', content: 'q3' },
+ { role: 'assistant', content: 'a3' },
+ ]
+ }
+
+ // The config mock of this file carries no pre-prune fields; these are the
+ // defaults `config.ts` deploys (T-B pins the master switch).
+ const DEPLOYED_PRE_PRUNE = {
+ compactionPrePruneEnabled: true,
+ compactionPrePruneDedup: true,
+ compactionPrePruneOneLine: true,
+ compactionPrePruneJsonTruncate: true,
+ compactionPrePruneStripMedia: true,
+ compactionPrePruneMaxArgsBytes: 4096,
+ compactionPrePruneSummaryTokens: 200,
+ compactionPrePruneProtectedTailTurns: 3,
+ }
+ const mutableConfig = appConfig as unknown as Record<string, unknown>
+ let previousConfig: Record<string, unknown>
+ beforeEach(() => {
+ previousConfig = Object.fromEntries(
+ Object.keys(DEPLOYED_PRE_PRUNE).map(key => [key, mutableConfig[key]])
+ )
+ Object.assign(mutableConfig, DEPLOYED_PRE_PRUNE)
+ })
+ afterEach(() => {
+ Object.assign(mutableConfig, previousConfig)
+ })
+
+ // Runs a task whose rehydrated history is `prunableHistory()` and returns the
+ // old tool result as it reached the loop.
+ async function oldResultReachingLoop(contextWindowTokens: number): Promise<string | undefined> {
+ vi.mocked(runToolUseLoop).mockResolvedValueOnce({ type: 'response', content: 'ok' } as any)
+ const conversationManager = new ConversationManager()
+ vi.spyOn(conversationManager, 'buildMessageHistory').mockReturnValue(prunableHistory())
+ const deps = createDeps({ conversationManager, contextWindowTokens })
+ await new TaskExecutor(createTask('hi'), deps).run()
+ const call = vi.mocked(runToolUseLoop).mock.calls.at(-1)
+ if (!call) throw new Error('Expected runToolUseLoop to be called')
+ return call[1]?.find(m => m.role === 'tool')?.content
+ }
+
+ it('T-R2-4b a 1M window leaves a 100k-token history untouched', async () => {
+ expect(await oldResultReachingLoop(1_000_000)).toBe(OLD_RESULT)
+ })
+
+ it('T-R2-4c witness: a 100k window pre-prunes the same history', async () => {
+ const reached = await oldResultReachingLoop(100_000)
+ expect(reached).toBeDefined()
+ expect(reached!.length).toBeLessThan(OLD_RESULT.length / 10)
+ })
+})
diff --git a/mcp-host/src/core/conversation/__tests__/compaction.test.ts b/mcp-host/src/core/conversation/__tests__/compaction.test.ts
index bf88ca66b..5a8ac51b4 100644
--- a/mcp-host/src/core/conversation/__tests__/compaction.test.ts
+++ b/mcp-host/src/core/conversation/__tests__/compaction.test.ts
@@ -108,3 +108,44 @@ describe('compactConversation', () => {
expect(() => validateToolLinkages(compacted)).not.toThrow()
})
})
+
+describe('compactConversation — pre-prune is gated at the threshold (#731)', () => {
+ // One old turn whose tool result is far over the pre-prune summary threshold
+ // (200 tokens), then three short turns the pre-prune protects. Four turns
+ // stay under the default `maxTurns` of 5, so only the pre-prune can change it.
+ const OLD_RESULT = 'x'.repeat(40_000)
+ function prunableHistory(): ChatMessage[] {
+ return [
+ { role: 'system', content: 'sys' },
+ { role: 'user', content: 'q0' },
+ {
+ role: 'assistant',
+ content: '',
+ tool_calls: [{ id: 'tc0', name: 'fetch', arguments: { url: 'https://example.test' } }],
+ },
+ { role: 'tool', tool_call_id: 'tc0', name: 'fetch', content: OLD_RESULT },
+ { role: 'user', content: 'q1' },
+ { role: 'assistant', content: 'a1' },
+ { role: 'user', content: 'q2' },
+ { role: 'assistant', content: 'a2' },
+ { role: 'user', content: 'q3' },
+ { role: 'assistant', content: 'a3' },
+ ]
+ }
+ const prePruneOn = { enabled: true }
+
+ it('T-R2-4a returns the input untouched below the threshold even with pre-prune on', () => {
+ const messages = prunableHistory()
+ expect(compactConversation(messages, undefined, 1_000_000, undefined, prePruneOn)).toBe(
+ messages
+ )
+ })
+
+ it('T-R2-4a witness: the same history is pre-pruned above the threshold', () => {
+ const messages = prunableHistory()
+ const compacted = compactConversation(messages, undefined, 1_000, undefined, prePruneOn)
+ const oldResult = compacted.find(m => m.role === 'tool')
+ expect(oldResult?.content).not.toBe(OLD_RESULT)
+ expect(oldResult!.content.length).toBeLessThan(OLD_RESULT.length / 10)
+ })
+})