Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
test(codex): prove direct-first guidance across catalog changes
Alfredo Lopez90fd50bc56ae
diff --git a/.github/workflows/ci-public.yml b/.github/workflows/ci-public.ymlindex 89acc163a..f0f668785 100644--- a/.github/workflows/ci-public.yml+++ b/.github/workflows/ci-public.yml@@ -585,6 +585,8 @@ jobs:run: bash scripts/tests/test-minikube-makefile.sh- name: Test local Minikube T0/T1/T2 contract+ # Includes test-minikube-docker-cli-env.sh: the real image-manifest+ # writer's hermetic sequential-build and derived-base provenance checks.run: bash scripts/tests/test-minikube-t2-contract.sh- name: Test local Minikube T2 setup handoffdiff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.jsonindex f1f03d41c..8afefb69d 100644--- a/mcp-host/package-lock.json+++ b/mcp-host/package-lock.json@@ -1,12 +1,12 @@{"name": "clerum-mcp-host",- "version": "0.1.328",+ "version": "0.1.329","lockfileVersion": 3,"requires": true,"packages": {"": {"name": "clerum-mcp-host",- "version": "0.1.328",+ "version": "0.1.329","license": "SEE LICENSE IN LICENSE","dependencies": {"@anthropic-ai/sdk": "^0.32.0",diff --git a/mcp-host/package.json b/mcp-host/package.jsonindex 5798accc1..e8e13a6e2 100644--- a/mcp-host/package.json+++ b/mcp-host/package.json@@ -1,6 +1,6 @@{"name": "clerum-mcp-host",- "version": "0.1.328",+ "version": "0.1.329","description": "MCP Host - reads Host CRD and provides LLM access via OpenAI/Claude","main": "dist/main.js","engines": {diff --git a/mcp-host/src/config.codexToolPresentation.test.ts b/mcp-host/src/config.codexToolPresentation.test.tsindex 5fac89362..b9e0f67a8 100644--- a/mcp-host/src/config.codexToolPresentation.test.ts+++ b/mcp-host/src/config.codexToolPresentation.test.ts@@ -2,11 +2,13 @@ import { afterEach, describe, expect, it, vi } from 'vitest'const originalEnv = process.env-async function loadConfig(value: string | undefined) {+async function loadConfig(value: string | undefined, discoveryBytes?: string) {vi.resetModules()process.env = { ...originalEnv }if (value === undefined) delete process.env.CODEX_TOOL_PRESENTATIONelse process.env.CODEX_TOOL_PRESENTATION = value+ if (discoveryBytes === undefined) delete process.env.CODEX_TOOL_DISCOVERY_BYTES+ else process.env.CODEX_TOOL_DISCOVERY_BYTES = discoveryBytesreturn (await import('./config')).config}@@ -15,6 +17,28 @@ afterEach(() => {vi.resetModules()})+describe('CODEX_TOOL_DISCOVERY_BYTES import-time configuration', () => {+ it('defaults to 32768 bytes when unset', async () => {+ expect((await loadConfig(undefined)).codexToolDiscoveryBytes).toBe(32_768)+ })++ it.each(['1', '32768', String(Number.MAX_SAFE_INTEGER)])(+ 'loads the positive safe integer %s',+ async value => {+ expect((await loadConfig(undefined, value)).codexToolDiscoveryBytes).toBe(Number(value))+ }+ )++ it.each(['', '0', '-1', '1.5', '01', '+1', ' 1', '1 ', '1e3', '3bytes', '9007199254740992'])(+ 'refuses to import configuration with invalid byte limit %j',+ async value => {+ await expect(loadConfig(undefined, value)).rejects.toThrow(+ 'CODEX_TOOL_DISCOVERY_BYTES must be a positive safe integer'+ )+ }+ )+})+describe('CODEX_TOOL_PRESENTATION import-time configuration', () => {it('defaults to auto when unset', async () => {expect((await loadConfig(undefined)).codexToolPresentation).toBe('auto')diff --git a/mcp-host/src/core/orchestration/__tests__/approvedToolsLifecycle.integration.test.ts b/mcp-host/src/core/orchestration/__tests__/approvedToolsLifecycle.integration.test.tsindex a8983fb9f..65168fdc5 100644--- a/mcp-host/src/core/orchestration/__tests__/approvedToolsLifecycle.integration.test.ts+++ b/mcp-host/src/core/orchestration/__tests__/approvedToolsLifecycle.integration.test.ts@@ -13,6 +13,7 @@ import { ApprovalController } from '../../extensions/approvalController'import { InLoopContextManager, PressureContextManager } from '../../extensions/contextManager'import { UnifiedApprovalGateController } from '../../extensions/mcpApprovalGateController'import type { Tool } from '../../interfaces'+import { DefaultPromptBuilder, TOOL_DISCOVERY_TEXT } from '../../reasoning/promptBuilder'import { BasicSafety } from '../../safety/safety'import { NativeToolRegistry } from '../../tools/nativeToolRegistry'import { type ChatMessage, ConversationState, type ToolCall } from '../../types'@@ -155,6 +156,68 @@ const bridgeCall = (name: string, id = 'selected-call', args = {}): ToolCall =>})describe('approved catalog across presentation and lifecycle', () => {+ it.each([83, 150, 250])(+ 'keeps direct-first guidance accurate in both prompt paths as auto grows to %i tools and shrinks',+ async count => {+ const { manager, registry, nativeNames } = await setup(1)+ const controller = new DeferrableToolController(+ new DefaultLoopController(),+ nativeNames,+ { dynamicToolsEnabled: false, dynamicToolsThreshold: 60, codexMode: 'auto' },+ { get: () => undefined, set: () => {} }+ )+ const builder = new DefaultPromptBuilder()+ const buildPrompts = async () => {+ const presented = await controller.refreshTools(registry.listDefinitions())+ const hasBridge = presented.some(tool => tool.name === 'clerum__tool_search')+ const parts = builder.buildParts({+ identityFiles: { identity: '', soul: '', agents: '', user: '' },+ dailyLogSnapshot: '',+ model: 'codex-test',+ provider: 'codex-subscription',+ platformHints: [],+ capabilities: '',+ workflowGuidance: '',+ mcpServerGuidance: '',+ toolDiscoveryGuidance: hasBridge ? TOOL_DISCOVERY_TEXT : '',+ memoryGuidance: '',+ })+ const legacy = String(builder.buildSystemPrompt(presented).content)+ for (const prompt of [legacy, parts.context]) {+ expect(prompt).toContain('Use directly listed tools when available.')+ expect(prompt).toContain('For additional approved tools,')+ expect(prompt).not.toContain('not all listed directly')+ }+ expect(hasBridge).toBe(true)+ return { presented, parts }+ }+ const direct = await buildPrompts()+ expect(direct.presented).toEqual(registry.listDefinitions())+ expect(direct.presented.some(tool => tool.name === 'alpha__record__read_000')).toBe(true)++ remote.catalogs.set(+ 'late',+ Array.from({ length: count - 1 }, (_, i) => ({+ name: `read_${i}`,+ description: `Read record ${i}`,+ inputSchema: { type: 'object', properties: {}, additionalProperties: false },+ }))+ )+ await manager.addServer(serverInfo('late'))+ expect(manager.getAllTools()).toHaveLength(count)+ const discovery = await buildPrompts()+ expect(discovery.presented.every(tool => nativeNames.has(tool.name))).toBe(true)+ expect(discovery.parts.contextHash).toBe(direct.parts.contextHash)+ expect(discovery.parts.stableHash).toBe(direct.parts.stableHash)++ await manager.removeServer('late')+ const directAgain = await buildPrompts()+ expect(directAgain.presented).toEqual(direct.presented)+ expect(directAgain.parts.contextHash).toBe(direct.parts.contextHash)+ expect(remote.calls).not.toHaveBeenCalled()+ }+ )+for (const mode of ['direct', 'auto', 'discovery'] as const) {for (const count of [83, 90, 150, 250]) {it.each([false, true])(`${mode}: ${count} MCP + 36 natives, reversed=%s`, async reverse => {diff --git a/mcp-host/src/core/reasoning/promptBuilder.ts b/mcp-host/src/core/reasoning/promptBuilder.tsindex feae61447..932e74f02 100644--- a/mcp-host/src/core/reasoning/promptBuilder.ts+++ b/mcp-host/src/core/reasoning/promptBuilder.ts@@ -108,9 +108,9 @@ export const MCP_SERVER_SELECTION_TEXT =/*** Tool-discovery guidance (dynamic-tool-loading, Phase F4.1). Emitted whenever* the `clerum__tool_search` bridge tool is registered — i.e. when the stable- * bridge is in play. Teaches the model the 3-step discovery flow so it knows- * the deferred MCP catalog is reachable even though those tools are not listed- * directly in `tools[]`.+ * bridge is callable. Direct definitions take precedence; discovery is explained+ * for additional approved tools. This applies both below and above the auto+ * threshold without claiming that directly presented schemas are hidden.** Single source of truth, shared by the legacy and tiered paths. It is a* CONSTANT (the bridge tools are always present when active), so it lives in thediff --git a/mcp-host/src/logger.test.ts b/mcp-host/src/logger.test.tsindex 073f260ac..1cdcc20a6 100644--- a/mcp-host/src/logger.test.ts+++ b/mcp-host/src/logger.test.ts@@ -1,4 +1,4 @@-import { describe, expect, it } from 'vitest'+import { describe, expect, it, vi } from 'vitest'import { redactUnknown } from './logger.js'describe('mcp-host structured logger redaction', () => {@@ -34,4 +34,44 @@ describe('mcp-host structured logger redaction', () => {const input = { ok: true, constructor: { evil: true }, prototype: { evil: true } }expect(redactUnknown(input)).toEqual({ ok: true })})++ it('emits a structured fallback through the public logger when fields cannot be serialized', async () => {+ const previousConsole = { log: console.log, error: console.error, warn: console.warn }+ const output = vi.spyOn(console, 'log').mockImplementation(() => {})+ vi.stubEnv('LOG_LEVEL', 'info')+ vi.resetModules()+ try {+ // The module binds its sinks at import time, so capture the sink before+ // loading it and restore its console adapters after this isolated check.+ const { logger } = await import('./logger.js')+ logger.info({ taskId: 'task-1' }, 'valid fields')+ const fields = Object.defineProperty({ taskId: 'task-2' }, 'unreadable', {+ enumerable: true,+ get() {+ throw new Error('fixture field cannot be read')+ },+ })+ logger.info(fields, 'unserializable fields')++ expect(output).toHaveBeenCalledTimes(2)+ expect(JSON.parse(output.mock.calls[0]![0] as string)).toMatchObject({+ taskId: 'task-1',+ level: 'info',+ msg: 'valid fields',+ })+ expect(JSON.parse(output.mock.calls[1]![0] as string)).toEqual({+ fields: '[Unserializable]',+ timestamp: expect.any(String),+ level: 'info',+ msg: 'unserializable fields',+ })+ } finally {+ output.mockRestore()+ console.log = previousConsole.log+ console.error = previousConsole.error+ console.warn = previousConsole.warn+ vi.unstubAllEnvs()+ vi.resetModules()+ }+ })})