paired: decouple unsafe host mode from fresh reviewer sessions

This commit is contained in:
ejclaw
2026-04-22 06:47:52 +09:00
parent 091f34ed23
commit 2b32b72119
6 changed files with 140 additions and 2 deletions

View File

@@ -30,6 +30,7 @@ describe('config/env loading', () => {
delete process.env.DISCORD_REVIEW_BOT_TOKEN;
delete process.env.DISCORD_CODEX_REVIEW_BOT_TOKEN;
delete process.env.PAIRED_CARRY_FORWARD_LATEST_OWNER_FINAL;
delete process.env.PAIRED_FORCE_FRESH_CLAUDE_REVIEWER_SESSION;
delete process.env.SESSION_COMMAND_USER_IDS;
vi.resetModules();
});
@@ -113,6 +114,16 @@ describe('config/env loading', () => {
expect(config.PAIRED_CARRY_FORWARD_LATEST_OWNER_FINAL).toBe(true);
});
it('defaults unsafe-host Claude reviewer fresh-session forcing to disabled and honors explicit opt-in', async () => {
let config = await import('./config.js');
expect(config.PAIRED_FORCE_FRESH_CLAUDE_REVIEWER_SESSION).toBe(false);
vi.resetModules();
process.env.PAIRED_FORCE_FRESH_CLAUDE_REVIEWER_SESSION = 'true';
config = await import('./config.js');
expect(config.PAIRED_FORCE_FRESH_CLAUDE_REVIEWER_SESSION).toBe(true);
});
it('fails fast when a legacy Discord token alias is configured', async () => {
process.env.DISCORD_BOT_TOKEN = 'legacy-owner-token';

View File

@@ -80,6 +80,14 @@ export const ARBITER_SERVICE_ID = CONFIG.paired.arbiterServiceId;
export const PAIRED_CARRY_FORWARD_LATEST_OWNER_FINAL =
CONFIG.paired.carryForwardLatestOwnerFinal;
/** Whether unsafe-host Claude reviewers must always start on a fresh SDK session. Default: false. */
export const PAIRED_FORCE_FRESH_CLAUDE_REVIEWER_SESSION =
CONFIG.paired.forceFreshClaudeReviewerSessionInUnsafeHostMode;
export function shouldForceFreshClaudeReviewerSessionInUnsafeHostMode(): boolean {
return PAIRED_FORCE_FRESH_CLAUDE_REVIEWER_SESSION;
}
/** Language for agent responses. When set, a language instruction is appended to all paired room prompts. */
export const AGENT_LANGUAGE = CONFIG.paired.agentLanguage;

View File

@@ -227,6 +227,10 @@ export function loadConfig(): AppConfig {
'PAIRED_CARRY_FORWARD_LATEST_OWNER_FINAL',
false,
),
forceFreshClaudeReviewerSessionInUnsafeHostMode: readBoolean(
'PAIRED_FORCE_FRESH_CLAUDE_REVIEWER_SESSION',
false,
),
agentLanguage: readText('AGENT_LANGUAGE') ?? '',
arbiterDeadlockThreshold: readInteger('ARBITER_DEADLOCK_THRESHOLD', 2),
maxRoundTrips,

View File

@@ -54,6 +54,7 @@ export interface AppConfig {
arbiterAgentType?: AgentType;
arbiterServiceId: string | null;
carryForwardLatestOwnerFinal: boolean;
forceFreshClaudeReviewerSessionInUnsafeHostMode: boolean;
agentLanguage: string;
arbiterDeadlockThreshold: number;
maxRoundTrips: number;

View File

@@ -8,6 +8,7 @@ import {
ARBITER_AGENT_TYPE,
REVIEWER_AGENT_TYPE,
getRoleModelConfig,
shouldForceFreshClaudeReviewerSessionInUnsafeHostMode,
} from './config.js';
import {
getAllTasks,
@@ -119,7 +120,10 @@ export async function prepareMessageAgentExecutionTarget(
const isClaudeCodeAgent = effectiveAgentType === 'claude-code';
const unsafeHostPairedMode = isUnsafeHostPairedModeEnabled();
const forceFreshClaudeReviewerSession =
reviewerMode && isClaudeCodeAgent && unsafeHostPairedMode;
reviewerMode &&
isClaudeCodeAgent &&
unsafeHostPairedMode &&
shouldForceFreshClaudeReviewerSessionInUnsafeHostMode();
const shouldPersistSession =
activeRole !== 'arbiter' &&
!args.forcedAgentType &&

View File

@@ -1,4 +1,5 @@
import { afterAll, beforeEach, describe, expect, it, vi } from 'vitest';
import * as config from './config.js';
vi.mock('./agent-runner.js', () => ({
runAgentProcess: vi.fn(),
@@ -16,9 +17,11 @@ vi.mock('./config.js', () => ({
CODEX_MAIN_SERVICE_ID: 'codex-main',
CODEX_REVIEW_SERVICE_ID: 'codex-review',
DATA_DIR: '/tmp/ejclaw-test-data',
PAIRED_FORCE_FRESH_CLAUDE_REVIEWER_SESSION: false,
REVIEWER_AGENT_TYPE: 'claude-code',
ARBITER_AGENT_TYPE: undefined,
SERVICE_SESSION_SCOPE: 'claude',
shouldForceFreshClaudeReviewerSessionInUnsafeHostMode: vi.fn(() => false),
isClaudeService: vi.fn(() => true),
normalizeServiceId: vi.fn((serviceId: string) =>
serviceId === 'codex' ? 'codex-main' : serviceId,
@@ -290,6 +293,9 @@ describe('runAgentForGroup room memory', () => {
vi.resetAllMocks();
resetPairedFollowUpScheduleState();
delete process.env.EJCLAW_UNSAFE_HOST_PAIRED_MODE;
vi.mocked(
config.shouldForceFreshClaudeReviewerSessionInUnsafeHostMode,
).mockReturnValue(false);
vi.mocked(agentRunner.runAgentProcess).mockImplementation(
async (_group, _input, _onProcess, onOutput) => {
await onOutput?.({
@@ -1862,7 +1868,7 @@ describe('runAgentForGroup room memory', () => {
);
});
it('starts reviewer Claude fresh in unsafe host mode instead of resuming a stored session', async () => {
it('keeps reviewer Claude session in unsafe host mode by default', async () => {
process.env.EJCLAW_UNSAFE_HOST_PAIRED_MODE = '1';
const group = {
...makeGroup(),
@@ -1946,6 +1952,110 @@ describe('runAgentForGroup room memory', () => {
runId: 'run-review-plan-unsafe-host',
});
expect(agentRunner.runAgentProcess).toHaveBeenCalledWith(
expect.objectContaining({
folder: 'test-group',
agentType: 'claude-code',
}),
expect.objectContaining({
sessionId: 'reviewer-session',
}),
expect.any(Function),
expect.any(Function),
expect.objectContaining({
EJCLAW_UNSAFE_HOST_PAIRED_MODE: '1',
}),
);
expect(deps.clearSession).not.toHaveBeenCalled();
});
it('starts reviewer Claude fresh in unsafe host mode when explicitly forced', async () => {
process.env.EJCLAW_UNSAFE_HOST_PAIRED_MODE = '1';
vi.mocked(
config.shouldForceFreshClaudeReviewerSessionInUnsafeHostMode,
).mockReturnValue(true);
const group = {
...makeGroup(),
folder: 'test-group',
agentType: 'codex' as const,
};
const deps = {
...makeDeps(),
getSessions: () => ({ 'test-group:reviewer': 'reviewer-session' }),
};
vi.mocked(serviceRouting.getEffectiveChannelLease).mockReturnValue({
chat_jid: 'group@test',
owner_agent_type: 'codex',
reviewer_agent_type: 'claude-code',
arbiter_agent_type: null,
owner_service_id: 'codex-main',
reviewer_service_id: 'claude',
arbiter_service_id: null,
activated_at: null,
reason: null,
explicit: false,
});
vi.mocked(db.getLatestOpenPairedTaskForChat).mockReturnValue({
id: 'paired-task-review',
chat_jid: 'group@test',
group_folder: 'test-group',
owner_service_id: 'codex-main',
reviewer_service_id: 'claude',
title: null,
source_ref: 'HEAD',
plan_notes: null,
round_trip_count: 0,
review_requested_at: '2026-03-31T00:00:00.000Z',
status: 'review_ready',
arbiter_verdict: null,
arbiter_requested_at: null,
completion_reason: null,
created_at: '2026-03-31T00:00:00.000Z',
updated_at: '2026-03-31T00:00:00.000Z',
});
vi.mocked(
pairedExecutionContext.preparePairedExecutionContext,
).mockReturnValue({
task: {
id: 'paired-task-review',
chat_jid: 'group@test',
group_folder: 'test-group',
owner_service_id: 'codex-main',
reviewer_service_id: 'claude',
title: null,
source_ref: 'HEAD',
plan_notes: null,
round_trip_count: 0,
review_requested_at: '2026-03-31T00:00:00.000Z',
status: 'review_ready',
arbiter_verdict: null,
arbiter_requested_at: null,
completion_reason: null,
created_at: '2026-03-31T00:00:00.000Z',
updated_at: '2026-03-31T00:00:00.000Z',
},
workspace: null,
envOverrides: {
EJCLAW_PAIRED_TASK_ID: 'paired-task-review',
EJCLAW_PAIRED_ROLE: 'reviewer',
EJCLAW_UNSAFE_HOST_PAIRED_MODE: '1',
CLAUDE_CONFIG_DIR: '/tmp/test-group-reviewer',
},
});
vi.mocked(agentRunner.runAgentProcess).mockResolvedValue({
status: 'success',
result: 'review ok',
newSessionId: 'review-session-new',
});
await runAgentForGroup(deps, {
group,
prompt: 'please review',
chatJid: 'group@test',
runId: 'run-review-plan-unsafe-host-fresh',
});
expect(deps.clearSession).toHaveBeenCalledWith('test-group:reviewer');
expect(agentRunner.runAgentProcess).toHaveBeenCalledWith(
expect.objectContaining({