Skip to content

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

TypeScript

test(host): witness the page collapse in the loop test and cover Grok

Alfredo Lopez6504865dcdde
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.json
index 8549293fa..5e766e243 100644
--- a/mcp-host/package-lock.json
+++ b/mcp-host/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.494",
+ "version": "0.1.495",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-mcp-host",
- "version": "0.1.494",
+ "version": "0.1.495",
"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 0076aed62..17741ccc8 100644
--- a/mcp-host/package.json
+++ b/mcp-host/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.494",
+ "version": "0.1.495",
"description": "MCP Host - reads Host CRD and provides LLM access via OpenAI/Claude",
"main": "dist/main.js",
"engines": {
diff --git a/mcp-host/src/core/orchestration/__tests__/loopInjectedProviderProjection.test.ts b/mcp-host/src/core/orchestration/__tests__/loopInjectedProviderProjection.test.ts
index 34f91e4d4..6ff2efed3 100644
--- a/mcp-host/src/core/orchestration/__tests__/loopInjectedProviderProjection.test.ts
+++ b/mcp-host/src/core/orchestration/__tests__/loopInjectedProviderProjection.test.ts
@@ -11,6 +11,7 @@
import { describe, expect, it, vi } from 'vitest'
import { ClaudeProvider } from '../../../llm/claude'
import { CodexSubscriptionProvider } from '../../../llm/codexSubscription'
+import { GrokSubscriptionProvider } from '../../../llm/grokSubscription'
import { OpenAIProvider } from '../../../llm/openai'
import type { ChatMessage } from '../../types'
import { isLoopInjectedMessage, loopInjectedUserMessage } from '../toolUseLoopMessages'
@@ -90,6 +91,35 @@ describe('LM1 the loopInjected marker never reaches a provider request body', ()
expectNoMarker(streamed)
})
+ it('Grok subscription authorize request and proxy stream request', async () => {
+ const authorize = vi.fn().mockResolvedValue({
+ providerAttemptId: 'attempt-1',
+ requestHash: 'a'.repeat(64),
+ executionTicket: 'ticket-123456',
+ expiresAt: '2026-08-20T10:00:00.000Z',
+ })
+ const stream = vi.fn().mockResolvedValue({ text: 'ok', toolCalls: [], outcome: 'success' })
+ const provider = new GrokSubscriptionProvider('grok-4.6', {
+ authorizer: { authorize },
+ proxy: { stream },
+ attemptContext: () => ({ policyRevision: 1, policyHash: 'b'.repeat(64), hostRef: 'chatllm' }),
+ } as never)
+
+ await provider.completeSingleTurnWithTools(conversation(), TOOLS)
+
+ expect(authorize).toHaveBeenCalledTimes(1)
+ expect(stream).toHaveBeenCalledTimes(1)
+ const authorized = authorize.mock.calls[0]![0] as { request: { messages: unknown[] } }
+ expect(authorized.request.messages).toEqual([
+ { role: 'user', content: GENUINE_TEXT },
+ { role: 'assistant', content: ASSISTANT_TEXT },
+ { role: 'user', content: INJECTED_TEXT },
+ ])
+ expectNoMarker(authorized)
+ const { signal: _signal, ...streamed } = stream.mock.calls[0]![0] as Record<string, unknown>
+ expectNoMarker(streamed)
+ })
+
it('Claude messages', async () => {
const create = vi.fn().mockResolvedValue({
content: [{ type: 'text', text: 'ok' }],
diff --git a/mcp-host/src/core/orchestration/__tests__/toolUseLoop.loopInjectedTurn.test.ts b/mcp-host/src/core/orchestration/__tests__/toolUseLoop.loopInjectedTurn.test.ts
index 313cdc5bb..f8aa81c41 100644
--- a/mcp-host/src/core/orchestration/__tests__/toolUseLoop.loopInjectedTurn.test.ts
+++ b/mcp-host/src/core/orchestration/__tests__/toolUseLoop.loopInjectedTurn.test.ts
@@ -5,7 +5,9 @@
* The loop runs for real: a page is read, the loop injects a `role: 'user'`
* message, a second page is read, and the context manager applies `prePrune`
* under qualifying pressure on every iteration. Both pages belong to the same
- * user turn, so neither may be collapsed before the model answers.
+ * user turn, so neither may be collapsed before the model answers. A page read
+ * in an earlier user turn is collapsed in the same run, which shows the pass
+ * was active.
*/
import { describe, expect, it, vi } from 'vitest'
import { makeFakeConversation } from '../../conversation/__testing__/makeFakeConversation'
@@ -162,7 +164,17 @@ async function runScenario(scenario: Scenario) {
contextManager: { manage },
})
+ // An earlier user turn already read a page. It is the collapse candidate the
+ // pass must act on in this same run, under the same pressure.
const outcome = await runToolUseLoop(config, [
+ { role: 'user', content: 'Read the start of the attached file.' },
+ {
+ role: 'assistant',
+ content: '',
+ tool_calls: [{ id: 'tc_p0', name: READ_TOOL, arguments: { attachmentId: 'att_1' } }],
+ },
+ { role: 'tool', name: READ_TOOL, tool_call_id: 'tc_p0', content: nativePage(0) },
+ { role: 'assistant', content: 'Read the first page.' },
{ role: 'user', content: 'Summarize the attached file.' },
])
return { outcome, sent, manage, collapseRuns }
@@ -174,7 +186,7 @@ describe('LM1 loop-injected user messages keep the current turn for page collaps
['a nudge from onTextRejected', nudgeScenario],
])('keeps both same-turn pages across %s under pressure', async (_, build) => {
const scenario = build()
- const { outcome, sent, manage } = await runScenario(scenario)
+ const { outcome, sent, manage, collapseRuns } = await runScenario(scenario)
// Witnesses: the loop ran the whole script, the context manager ran on
// every iteration, and the injected user message sits between the pages.
@@ -182,6 +194,13 @@ describe('LM1 loop-injected user messages keep the current turn for page collaps
expect(sent).toHaveLength(4)
expect(manage.mock.calls.length).toBeGreaterThanOrEqual(4)
const final = sent[3]!
+
+ // Witness that the collapse pass ran in this run: the earlier turn's page
+ // reached the model collapsed.
+ expect(collapseRuns.flat()).toContain('attachment_page_collapse')
+ const earlierPage = final.findIndex(m => m.role === 'tool' && m.tool_call_id === 'tc_p0')
+ expect(earlierPage).toBeGreaterThan(0)
+ expect(pageText(final[earlierPage]!)).toBe(COLLAPSE_MARKER)
const firstPage = final.findIndex(m => m.role === 'tool' && m.tool_call_id === 'tc_p1')
const injected = final.findIndex(m => m.role === 'user' && m.content === scenario.injected)
const secondPage = final.findIndex(m => m.role === 'tool' && m.tool_call_id === 'tc_p2')