Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
test(mcp-host): keep each liveness witness inside the test it supports (#739)
Alfredo Lopez19f68ae45169
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.jsonindex 1fafbca65..26665d124 100644--- a/mcp-host/package-lock.json+++ b/mcp-host/package-lock.json@@ -1,12 +1,12 @@{"name": "clerum-mcp-host",- "version": "0.1.424",+ "version": "0.1.425","lockfileVersion": 3,"requires": true,"packages": {"": {"name": "clerum-mcp-host",- "version": "0.1.424",+ "version": "0.1.425","license": "SEE LICENSE IN LICENSE","dependencies": {"@anthropic-ai/sdk": "^0.32.0",diff --git a/mcp-host/package.json b/mcp-host/package.jsonindex f738305ba..f3d1ea0b0 100644--- a/mcp-host/package.json+++ b/mcp-host/package.json@@ -1,6 +1,6 @@{"name": "clerum-mcp-host",- "version": "0.1.424",+ "version": "0.1.425","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.tsindex d0f66dbc4..0b9028a03 100644--- a/mcp-host/src/agent/__tests__/taskExecutor.test.ts+++ b/mcp-host/src/agent/__tests__/taskExecutor.test.ts@@ -1,4 +1,5 @@import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'+import { LIMITS as CODEX_LIMITS } from '@clerum/llm-provider-attempt-contract'import { config as appConfig } from '../../config'import { LlmPortAdapter } from '../../core/adapters/llmPortAdapter'import { makeFakeConversation } from '../../core/conversation/__testing__/makeFakeConversation'@@ -2410,21 +2411,27 @@ describe('TaskExecutor context manager message bound (#731)', () => {}it('T-R2-3c a codex-subscription task compacts past the contract message bound', async () => {- const msgs = smallTurns(1_025)+ const msgs = smallTurns(CODEX_LIMITS.maxMessages + 1)const managed = await (await loopContextManager('codex-subscription')).manage(msgs, makeFakeConversation())- expect(managed.length).toBeLessThan(1_024)+ expect(managed.length).toBeLessThan(CODEX_LIMITS.maxMessages)})it('T-R2-3d an openai task, with no attempt contract, is not bounded by message count', async () => {- const msgs = smallTurns(1_025)+ const msgs = smallTurns(CODEX_LIMITS.maxMessages + 1)const contextManager = await loopContextManager('openai')- // Liveness witness: the loop's default manager always passes through, so the- // passthrough below proves nothing unless this is the pressure manager.+ // The loop's default manager always passes through, so the passthrough+ // below proves nothing unless this is the pressure manager.expect(contextManager).toBeInstanceOf(PressureContextManager)const managed = await contextManager.manage(msgs, makeFakeConversation())expect(managed).toBe(msgs)+ // Liveness witness: the same history is compacted by a task whose provider+ // carries the bound, so the passthrough above is the missing bound.+ const bounded = await (+ await loopContextManager('codex-subscription')+ ).manage(msgs, makeFakeConversation())+ expect(bounded.length).toBeLessThan(CODEX_LIMITS.maxMessages)})})@@ -2574,14 +2581,14 @@ describe('TaskExecutor history compaction threshold follows the context window (return call[1]?.find(m => m.role === 'tool')?.content}- it('T-R2-4b a 1M window leaves a 100k-token history untouched', async () => {+ it('T-R2-4b/4c 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)+ // Liveness witness: a 100k window pre-prunes the same history, so the+ // untouched result above is the window's doing, not a compaction that+ // never ran.+ const pruned = await oldResultReachingLoop(100_000)+ expect(pruned).toBeDefined()+ expect(pruned!.length).toBeLessThan(OLD_RESULT.length / 10)})describe('R9-12 (M-D) rehydration measures with the context manager', () => {diff --git a/mcp-host/src/core/conversation/__tests__/compaction.test.ts b/mcp-host/src/core/conversation/__tests__/compaction.test.tsindex d34f8cde4..fbe72d9d6 100644--- a/mcp-host/src/core/conversation/__tests__/compaction.test.ts+++ b/mcp-host/src/core/conversation/__tests__/compaction.test.ts@@ -139,13 +139,12 @@ describe('compactConversation — pre-prune is gated at the threshold (#731)', (expect(await compactConversation(messages, undefined, 1_000_000, undefined, prePruneOn)).toBe(messages)- })-- it('T-R2-4a witness: the same history is pre-pruned above the threshold', async () => {- const messages = prunableHistory()+ // Liveness witness: the same history is pre-pruned above the threshold, so+ // the passthrough above is the threshold's doing, not a pre-prune that+ // never runs.const compacted = await compactConversation(messages, undefined, 1_000, undefined, prePruneOn)const oldResult = compacted.find(m => m.role === 'tool')- expect(oldResult?.content).not.toBe(OLD_RESULT)+ expect(oldResult).toBeDefined()expect(oldResult!.content.length).toBeLessThan(OLD_RESULT.length / 10)})})diff --git a/mcp-host/src/core/conversation/__tests__/compactionTranscriptInvariant.test.ts b/mcp-host/src/core/conversation/__tests__/compactionTranscriptInvariant.test.tsindex a078cade7..bd209cc0d 100644--- a/mcp-host/src/core/conversation/__tests__/compactionTranscriptInvariant.test.ts+++ b/mcp-host/src/core/conversation/__tests__/compactionTranscriptInvariant.test.ts@@ -84,6 +84,9 @@ describe('compaction ↔ durable transcript invariant (FU2, spec §8 "does not aconst compacted = await compactConversation(messages, 2, 0)expect(compacted).not.toBe(messages)+ // The new array is the compaction's output: two turns of the twelve messages.+ expect(compacted.length).toBeLessThan(messages.length)+ expect(compacted.at(-1)).toEqual(messages.at(-1))expect(messages).toEqual(inputSnapshot) // input array left intact})})diff --git a/mcp-host/src/core/tokenizer/__tests__/contextManagerDryRun.test.ts b/mcp-host/src/core/tokenizer/__tests__/contextManagerDryRun.test.tsindex e97f85bde..80cef95ed 100644--- a/mcp-host/src/core/tokenizer/__tests__/contextManagerDryRun.test.ts+++ b/mcp-host/src/core/tokenizer/__tests__/contextManagerDryRun.test.ts@@ -34,6 +34,20 @@ function tinyMessages(): ChatMessage[] {]}+/**+ * `count` short user/assistant turns. The emergency tier keeps the last three,+ * so a compaction of this history is visible in its length, not only in a new+ * array reference.+ */+function turns(count: number): ChatMessage[] {+ const msgs: ChatMessage[] = []+ for (let i = 1; i <= count; i++) {+ msgs.push({ role: 'user', content: `question ${i} a b c d e f` })+ msgs.push({ role: 'assistant', content: `answer ${i} g h i j k l` })+ }+ return msgs+}+function getMismatchCount(from: string, to: string): number {const samples = (tokenizerDryrunTierMismatchTotal as unknown as {@@ -78,7 +92,7 @@ describe('PressureContextManager dry-run', () => {})it('lets the counter drive the tier when dryRun=false', async () => {- const msgs = tinyMessages()+ const msgs = turns(10)const heuristic = heuristicCount(msgs)// Same setup as above (heuristic would passthrough) but with dryRun// disabled the counter's high value should trigger compaction.@@ -89,6 +103,9 @@ describe('PressureContextManager dry-run', () => {})const result = await manager.manage(msgs, makeFakeConversation())expect(result).not.toBe(msgs)+ // The emergency tier ran: the last three turns remain, the newest intact.+ expect(result.filter(m => m.role === 'user')).toHaveLength(3)+ expect(result.at(-1)?.content).toBe(msgs.at(-1)?.content)})it('T-E1b a failing dry-run counter is reported through the service logger (#731)', async () => {@@ -149,7 +166,7 @@ describe('PressureContextManager pressure includes the tool schemas', () => {}it('T-R2-2a tools push a conversation below 0.8 into compaction on the heuristic path', async () => {- const msgs = tinyMessages()+ const msgs = turns(10)const tools = bulkyTools()const maxTokens = Math.ceil(heuristicCount(msgs) / 0.7)// Witness the arithmetic, so the two outcomes below can only differ by the tools term.@@ -160,11 +177,15 @@ describe('PressureContextManager pressure includes the tool schemas', () => {const manager = new PressureContextManager(maxTokens)expect(await manager.manage(msgs, makeFakeConversation())).toBe(msgs)- expect(await manager.manage(msgs, makeFakeConversation(), { tools })).not.toBe(msgs)+ const result = await manager.manage(msgs, makeFakeConversation(), { tools })+ expect(result).not.toBe(msgs)+ // The emergency tier ran: the last three turns remain, the newest intact.+ expect(result.filter(m => m.role === 'user')).toHaveLength(3)+ expect(result.at(-1)?.content).toBe(msgs.at(-1)?.content)})it('T-R2-2b dry-run decides from messages plus tools and hands the tools to the counter', async () => {- const msgs = tinyMessages()+ const msgs = turns(10)const tools = bulkyTools()const maxTokens = Math.ceil(heuristicCount(msgs) / 0.7)const counter = makeCounter(0) // would pass through if it decided@@ -175,6 +196,7 @@ describe('PressureContextManager pressure includes the tool schemas', () => {const result = await manager.manage(msgs, makeFakeConversation(), { tools })expect(result).not.toBe(msgs)+ expect(result.filter(m => m.role === 'user')).toHaveLength(3)expect(counter.count).toHaveBeenCalledWith(msgs, tools)})@@ -200,14 +222,6 @@ describe('PressureContextManager pressure includes the tool schemas', () => {// guidance. It is not part of `messages`, so the gauge must add it in every// branch of `computePressure`, as it adds the tools (R9-14, L-6).describe('PressureContextManager pressure includes the system prompt', () => {- function turns(count: number): ChatMessage[] {- const msgs: ChatMessage[] = []- for (let i = 1; i <= count; i++) {- msgs.push({ role: 'user', content: `question ${i} a b c d e f` })- msgs.push({ role: 'assistant', content: `answer ${i} g h i j k l` })- }- return msgs- }const asSystemMessage = (content: string): ChatMessage => ({ role: 'system', content })// Messages at 0.7 of the window; the system prompt alone is about one window.