fix: refresh role rules after compaction

This commit is contained in:
ejclaw
2026-05-25 22:57:52 +09:00
parent e4a0e15cf2
commit db5332e81a
9 changed files with 570 additions and 85 deletions

View File

@@ -50,6 +50,10 @@ export interface AgentOutput {
agentDone?: boolean;
newSessionId?: string;
error?: string;
compaction?: {
completed: boolean;
trigger?: string | null;
};
}
function readRoomSkillOverridesForRunner(

104
src/compact-refresh.test.ts Normal file
View File

@@ -0,0 +1,104 @@
import fs from 'fs';
import { beforeEach, describe, expect, it, vi } from 'vitest';
vi.mock('./config.js', () => ({
DATA_DIR: '/tmp/ejclaw-compact-refresh-test',
}));
import {
clearCompactRefreshIfUnchanged,
markCompactRefreshNeeded,
maybeApplyCompactRefresh,
readCompactRefreshFlag,
} from './compact-refresh.js';
const refreshDir = '/tmp/ejclaw-compact-refresh-test/compact-refresh';
describe('compact-refresh', () => {
beforeEach(() => {
fs.rmSync(refreshDir, { recursive: true, force: true });
});
it('marks and reads compact refresh flags by session folder', () => {
const flag = markCompactRefreshNeeded({
sessionFolder: 'room:reviewer',
sessionId: 'session-1',
trigger: 'auto',
});
expect(readCompactRefreshFlag('room:reviewer')).toMatchObject({
sessionFolder: 'room:reviewer',
sessionId: 'session-1',
trigger: 'auto',
compactedAt: flag.compactedAt,
});
});
it('applies a one-shot compact refresh only for owner and reviewer roles', () => {
const flag = markCompactRefreshNeeded({
sessionFolder: 'room',
sessionId: 'session-1',
});
const applied = maybeApplyCompactRefresh({
sessionFolder: 'room',
sessionId: 'session-1',
role: 'owner',
prompt: 'continue task',
});
expect(applied?.flag).toEqual(flag);
expect(applied?.prompt).toContain('EJClaw compact refresh');
expect(applied?.prompt).toContain('continue task');
clearCompactRefreshIfUnchanged({
sessionFolder: 'room',
flag,
});
expect(readCompactRefreshFlag('room')).toBeNull();
});
it('does not apply to arbiter, compact commands, fresh sessions, or stale sessions', () => {
markCompactRefreshNeeded({
sessionFolder: 'room',
sessionId: 'session-1',
});
expect(
maybeApplyCompactRefresh({
sessionFolder: 'room',
sessionId: 'session-1',
role: 'arbiter',
prompt: 'judge',
}),
).toBeNull();
expect(
maybeApplyCompactRefresh({
sessionFolder: 'room',
sessionId: 'session-1',
role: 'owner',
prompt: '/compact',
}),
).toBeNull();
expect(
maybeApplyCompactRefresh({
sessionFolder: 'room',
role: 'owner',
prompt: 'fresh',
}),
).toBeNull();
expect(
maybeApplyCompactRefresh({
sessionFolder: 'room',
sessionId: 'different-session',
role: 'owner',
prompt: 'stale',
}),
).toBeNull();
expect(readCompactRefreshFlag('room')).toBeNull();
});
});

128
src/compact-refresh.ts Normal file
View File

@@ -0,0 +1,128 @@
import fs from 'fs';
import path from 'path';
import { DATA_DIR } from './config.js';
import type { PairedRoomRole } from './types.js';
export interface CompactRefreshFlag {
sessionFolder: string;
sessionId: string;
compactedAt: string;
trigger?: string | null;
}
export interface AppliedCompactRefresh {
flag: CompactRefreshFlag;
prompt: string;
}
const COMPACT_REFRESH_DIR = path.join(DATA_DIR, 'compact-refresh');
const COMPACT_REFRESH_PROMPT = `[EJClaw compact refresh]
The previous run compacted this same SDK session. Continue following the current AGENTS.md/CLAUDE.md role rules exactly: required first-line status, paired-room role boundaries, output-only task context, concise Korean responses, and verification before completion. This is not new user scope.
[/EJClaw compact refresh]`;
function getFlagPath(sessionFolder: string): string {
return path.join(
COMPACT_REFRESH_DIR,
`${encodeURIComponent(sessionFolder)}.json`,
);
}
function isRefreshableRole(role: PairedRoomRole): boolean {
return role === 'owner' || role === 'reviewer';
}
function isSessionCommand(prompt: string): boolean {
return prompt.trim() === '/compact';
}
export function readCompactRefreshFlag(
sessionFolder: string,
): CompactRefreshFlag | null {
const flagPath = getFlagPath(sessionFolder);
if (!fs.existsSync(flagPath)) return null;
try {
const parsed = JSON.parse(
fs.readFileSync(flagPath, 'utf-8'),
) as Partial<CompactRefreshFlag>;
if (
parsed.sessionFolder !== sessionFolder ||
typeof parsed.sessionId !== 'string' ||
parsed.sessionId.length === 0 ||
typeof parsed.compactedAt !== 'string'
) {
fs.rmSync(flagPath, { force: true });
return null;
}
return {
sessionFolder,
sessionId: parsed.sessionId,
compactedAt: parsed.compactedAt,
trigger:
typeof parsed.trigger === 'string' && parsed.trigger.length > 0
? parsed.trigger
: null,
};
} catch {
fs.rmSync(flagPath, { force: true });
return null;
}
}
export function markCompactRefreshNeeded(args: {
sessionFolder: string;
sessionId: string;
trigger?: string | null;
}): CompactRefreshFlag {
const flag: CompactRefreshFlag = {
sessionFolder: args.sessionFolder,
sessionId: args.sessionId,
compactedAt: new Date().toISOString(),
trigger: args.trigger ?? null,
};
fs.mkdirSync(COMPACT_REFRESH_DIR, { recursive: true });
fs.writeFileSync(
getFlagPath(args.sessionFolder),
`${JSON.stringify(flag)}\n`,
);
return flag;
}
export function clearCompactRefreshIfUnchanged(args: {
sessionFolder: string;
flag: CompactRefreshFlag;
}): void {
const current = readCompactRefreshFlag(args.sessionFolder);
if (
current &&
current.sessionId === args.flag.sessionId &&
current.compactedAt === args.flag.compactedAt
) {
fs.rmSync(getFlagPath(args.sessionFolder), { force: true });
}
}
export function maybeApplyCompactRefresh(args: {
sessionFolder: string;
sessionId?: string;
role: PairedRoomRole;
prompt: string;
}): AppliedCompactRefresh | null {
if (!args.sessionId) return null;
if (!isRefreshableRole(args.role)) return null;
if (isSessionCommand(args.prompt)) return null;
const flag = readCompactRefreshFlag(args.sessionFolder);
if (!flag) return null;
if (flag.sessionId !== args.sessionId) {
fs.rmSync(getFlagPath(args.sessionFolder), { force: true });
return null;
}
return {
flag,
prompt: `${COMPACT_REFRESH_PROMPT}\n\n${args.prompt}`,
};
}

View File

@@ -0,0 +1,114 @@
import fs from 'fs';
import { beforeEach, describe, expect, it, vi } from 'vitest';
vi.mock('./config.js', () => ({
DATA_DIR: '/tmp/ejclaw-message-compact-refresh-test',
}));
vi.mock('./agent-runner.js', () => ({
runAgentProcess: vi.fn(),
}));
vi.mock('./codex-token-rotation.js', () => ({
getCodexAccountCount: vi.fn(() => 1),
}));
vi.mock('./session-recovery.js', () => ({
shouldResetCodexSessionOnAgentFailure: vi.fn(() => false),
shouldResetSessionOnAgentFailure: vi.fn(() => false),
shouldRetryFreshCodexSessionOnAgentFailure: vi.fn(() => false),
}));
import { runAgentProcess, type AgentOutput } from './agent-runner.js';
import {
maybeApplyCompactRefresh,
readCompactRefreshFlag,
} from './compact-refresh.js';
import { runMessageAgentAttempt } from './message-agent-executor-attempt-runner.js';
import type { RegisteredGroup } from './types.js';
const dataDir = '/tmp/ejclaw-message-compact-refresh-test';
const group: RegisteredGroup = {
name: 'Room',
folder: 'room',
trigger: '@Andy',
added_at: new Date().toISOString(),
agentType: 'claude-code',
};
function makeLifecycle() {
return {
updateSummary: vi.fn(),
recordFinalOutputBeforeDelivery: vi.fn(() => true),
};
}
function makeLogger(): any {
const logger = {
info: vi.fn(),
warn: vi.fn(),
error: vi.fn(),
child: vi.fn(),
};
logger.child.mockReturnValue(logger);
return logger;
}
describe('message agent compact refresh', () => {
beforeEach(() => {
fs.rmSync(dataDir, { recursive: true, force: true });
vi.resetAllMocks();
});
it('marks compacted owner sessions so the next turn gets one minimal refresh', async () => {
const compactedOutput: AgentOutput = {
status: 'success',
result: 'done',
output: { visibility: 'public', text: 'done' },
newSessionId: 'session-1',
compaction: { completed: true, trigger: 'auto' },
};
vi.mocked(runAgentProcess).mockResolvedValue(compactedOutput);
await runMessageAgentAttempt({
provider: 'claude',
currentSessionId: 'session-1',
isClaudeCodeAgent: true,
canRetryClaudeCredentials: false,
shouldPersistSession: true,
effectiveGroup: group,
agentInput: {
prompt: 'work',
sessionId: 'session-1',
groupFolder: 'room',
chatJid: 'room@test',
runId: 'run-1',
isMain: false,
assistantName: 'Andy',
},
activeRole: 'owner',
effectiveServiceId: 'claude',
effectiveAgentType: 'claude-code',
sessionFolder: 'room',
onPersistSession: vi.fn(),
registerProcess: vi.fn(),
pairedExecutionLifecycle: makeLifecycle(),
log: makeLogger(),
});
expect(readCompactRefreshFlag('room')).toMatchObject({
sessionId: 'session-1',
trigger: 'auto',
});
const applied = maybeApplyCompactRefresh({
sessionFolder: 'room',
sessionId: 'session-1',
role: 'owner',
prompt: 'next work',
});
expect(applied?.prompt).toContain('EJClaw compact refresh');
expect(applied?.prompt).toContain('next work');
});
});

View File

@@ -3,6 +3,7 @@ import type { Logger } from 'pino';
import { createEvaluatedOutputHandler } from './agent-attempt.js';
import type { AttemptStreamedTrigger } from './agent-attempt-retry.js';
import { runAgentProcess, type AgentOutput } from './agent-runner.js';
import { markCompactRefreshNeeded } from './compact-refresh.js';
import { getCodexAccountCount } from './codex-token-rotation.js';
import type { PreparedPairedExecutionContext } from './paired-execution-context.js';
import {
@@ -34,6 +35,35 @@ interface AgentInput {
roomRoleContext?: RoomRoleContext;
}
function maybeMarkCompactRefreshForOutput(args: {
output: AgentOutput;
activeRole: string;
sessionFolder: string;
}): void {
if (
(args.activeRole !== 'owner' && args.activeRole !== 'reviewer') ||
args.output.compaction?.completed !== true ||
!args.output.newSessionId
) {
return;
}
markCompactRefreshNeeded({
sessionFolder: args.sessionFolder,
sessionId: args.output.newSessionId,
trigger: args.output.compaction.trigger ?? null,
});
}
function createProviderLog(
log: Logger,
provider: 'claude' | 'codex',
agentType: AgentType,
): Logger {
const providerLog = log.child({ provider, agentType });
providerLog.info('Using provider');
return providerLog;
}
export async function runMessageAgentAttempt(args: {
provider: 'claude' | 'codex';
currentSessionId: string | undefined;
@@ -111,6 +141,7 @@ export async function runMessageAgentAttempt(args: {
structuredOutput,
evaluation,
}) => {
maybeMarkCompactRefreshForOutput({ output, activeRole, sessionFolder });
const outputPhase = output.phase ?? 'final';
if (outputPhase !== 'final') {
log.info(
@@ -252,11 +283,7 @@ export async function runMessageAgentAttempt(args: {
await streamedOutputHandler.handleOutput(output);
};
const providerLog = log.child({
provider,
agentType: effectiveAgentType,
});
providerLog.info('Using provider');
const providerLog = createProviderLog(log, provider, effectiveAgentType);
try {
const output = await runAgentProcess(
@@ -273,6 +300,7 @@ export async function runMessageAgentAttempt(args: {
if (output.newSessionId && shouldPersistSession) {
onPersistSession(output.newSessionId);
}
maybeMarkCompactRefreshForOutput({ output, activeRole, sessionFolder });
providerLog.info(
{

View File

@@ -1,5 +1,6 @@
import { AgentOutput } from './agent-runner.js';
import { GroupQueue } from './group-queue.js';
import type { Logger } from 'pino';
import type { AgentTriggerReason } from './agent-error-detection.js';
import { getMoaConfig } from './config.js';
import { createPairedExecutionLifecycle } from './message-agent-executor-paired.js';
@@ -12,9 +13,82 @@ import { collectMoaReferences, formatMoaReferencesForPrompt } from './moa.js';
import { readArbiterPrompt } from './platform-prompts.js';
import { shouldRetryFreshSessionOnAgentFailure } from './session-recovery.js';
import type { AgentType, PairedRoomRole, RegisteredGroup } from './types.js';
import {
clearCompactRefreshIfUnchanged,
maybeApplyCompactRefresh,
type AppliedCompactRefresh,
} from './compact-refresh.js';
// ── Main executor ─────────────────────────────────────────────────
function buildPromptWithCompactRefresh(args: {
prompt: string;
sessionFolder: string;
sessionId?: string;
role: PairedRoomRole;
}): { effectivePrompt: string; compactRefresh: AppliedCompactRefresh | null } {
const compactRefresh = maybeApplyCompactRefresh({
sessionFolder: args.sessionFolder,
sessionId: args.sessionId,
role: args.role,
prompt: args.prompt,
});
return {
effectivePrompt: compactRefresh?.prompt ?? args.prompt,
compactRefresh,
};
}
function clearAppliedCompactRefreshAfterSuccess(args: {
result: 'success' | 'error';
sessionFolder: string;
compactRefresh: AppliedCompactRefresh | null;
}): void {
if (args.result !== 'success' || !args.compactRefresh) return;
clearCompactRefreshIfUnchanged({
sessionFolder: args.sessionFolder,
flag: args.compactRefresh.flag,
});
}
async function enrichArbiterPromptWithMoa(args: {
prompt: string;
enabled: boolean;
log: Pick<Logger, 'info' | 'warn'>;
}): Promise<string> {
const moaConfig = getMoaConfig();
if (!args.enabled || !moaConfig.enabled) return args.prompt;
args.log.info(
{ models: moaConfig.referenceModels.map((m) => m.name) },
'MoA: collecting reference opinions before arbiter',
);
const systemPrompt =
readArbiterPrompt(process.cwd()) || 'You are an arbiter.';
try {
const references = await collectMoaReferences({
config: moaConfig,
systemPrompt,
contextPrompt: args.prompt,
});
const moaSection = formatMoaReferencesForPrompt(references);
if (!moaSection) return args.prompt;
args.log.info(
{
successCount: references.filter((r) => !r.error).length,
totalCount: references.length,
},
'MoA: injected reference opinions into arbiter prompt',
);
return args.prompt + '\n' + moaSection;
} catch (err) {
args.log.warn(
{ err, models: moaConfig.referenceModels.map((m) => m.name) },
'MoA: failed to collect reference opinions; continuing without enrichment',
);
return args.prompt;
}
}
export interface MessageAgentExecutorDeps {
assistantName: string;
queue: Pick<GroupQueue, 'registerProcess' | 'enqueueMessageCheck'> &
@@ -81,54 +155,18 @@ export async function runAgentForGroup(
clearRoleSdkSessions();
}
// ── MoA prompt enrichment ─────────────────────────────────────
// When MoA is enabled and we're in arbiter mode, query external API
// models (Kimi, GLM, etc.) in parallel for their opinions, then inject
// those opinions into the arbiter's prompt. The SDK-based arbiter
// agent naturally aggregates all perspectives.
let moaEnrichedPrompt = prompt;
const moaConfig = getMoaConfig();
if (arbiterMode && moaConfig.enabled && pairedExecutionContext) {
log.info(
{
models: moaConfig.referenceModels.map((m) => m.name),
},
'MoA: collecting reference opinions before arbiter',
);
const moaEnrichedPrompt = await enrichArbiterPromptWithMoa({
prompt,
enabled: arbiterMode && Boolean(pairedExecutionContext),
log,
});
const systemPrompt =
readArbiterPrompt(process.cwd()) || 'You are an arbiter.';
try {
const references = await collectMoaReferences({
config: moaConfig,
systemPrompt,
contextPrompt: prompt,
});
const moaSection = formatMoaReferencesForPrompt(references);
if (moaSection) {
moaEnrichedPrompt = prompt + '\n' + moaSection;
log.info(
{
successCount: references.filter((r) => !r.error).length,
totalCount: references.length,
},
'MoA: injected reference opinions into arbiter prompt',
);
}
} catch (err) {
log.warn(
{
err,
models: moaConfig.referenceModels.map((m) => m.name),
},
'MoA: failed to collect reference opinions; continuing without enrichment',
);
}
}
const effectivePrompt = moaEnrichedPrompt;
const { effectivePrompt, compactRefresh } = buildPromptWithCompactRefresh({
prompt: moaEnrichedPrompt,
sessionFolder,
sessionId: currentSessionId,
role: activeRole,
});
const pairedExecutionLifecycle = createPairedExecutionLifecycle({
pairedExecutionContext,
pairedTurnIdentity: runtimePairedTurnIdentity,
@@ -237,7 +275,7 @@ export async function runAgentForGroup(
});
try {
return await executeMessageAgentAttemptLifecycle({
const result = await executeMessageAgentAttemptLifecycle({
provider,
runAttempt,
isClaudeCodeAgent,
@@ -260,6 +298,12 @@ export async function runAgentForGroup(
},
log,
});
clearAppliedCompactRefreshAfterSuccess({
result,
sessionFolder,
compactRefresh,
});
return result;
} finally {
await pairedExecutionLifecycle.asyncFinalize();
}