refactor: split Claude agent runner entrypoint (#200)

This commit is contained in:
Eyejoker
2026-05-30 01:47:41 +09:00
committed by GitHub
parent 35988d9797
commit 1531482363
8 changed files with 864 additions and 705 deletions

View File

@@ -16,23 +16,16 @@
import fs from 'fs';
import path from 'path';
import { query } from '@anthropic-ai/claude-agent-sdk';
import {
EJCLAW_ENV,
IPC_CLOSE_SENTINEL,
IPC_INPUT_SUBDIR,
IPC_POLL_MS,
} from 'ejclaw-runners-shared';
import { fileURLToPath } from 'url';
import { resolveBundledClaudeCodeExecutable } from './bundled-cli-path.js';
import {
prependRoomRoleHeader,
type RoomRoleContext,
} from './room-role-context.js';
import { prependRoomRoleHeader } from './room-role-context.js';
import {
assertReadonlyWorkspaceRepoConnectivity,
buildClaudeReadonlySandboxSettings,
buildReviewerGitGuardEnv,
getClaudeReadonlySandboxMode,
isArbiterRuntimeEnvEnabled,
@@ -40,43 +33,16 @@ import {
isReviewerRuntime,
isReviewerRuntimeEnvEnabled,
} from './reviewer-runtime.js';
import { drainIpcInput, shouldClose } from './ipc-input.js';
import { drainIpcInput } from './ipc-input.js';
import {
MessageStream,
buildCompactionOutput,
buildMultimodalContent,
compactBoundaryOutput,
extractAssistantText,
normalizeStructuredOutput,
readStdin,
type RunnerCompaction,
writeOutput,
} from './output-protocol.js';
import {
TopLevelAgentTaskTracker,
buildTaskNotificationOutput,
buildTaskProgressOutput,
buildTaskStartedOutput,
} from './task-progress-mapping.js';
import {
createPreCompactHook,
createReviewerBashGuardHook,
createSanitizeBashHook,
} from './runner-hooks.js';
import { buildClaudeSdkEnv } from './sdk-env.js';
import { buildEjclawMcpServerConfig } from './mcp-config.js';
interface RunnerInput {
prompt: string;
sessionId?: string;
groupFolder: string;
chatJid: string;
isMain: boolean;
isScheduledTask?: boolean;
assistantName?: string;
secrets?: Record<string, string>;
roomRoleContext?: RoomRoleContext;
}
import { runSessionCommand } from './session-command.js';
import { runClaudeQuery } from './claude-query-runner.js';
import type { RunnerInput } from './runner-input.js';
// Paths configurable via env vars.
const GROUP_DIR = process.env[EJCLAW_ENV.groupDir] || '/workspace/group';
@@ -94,41 +60,6 @@ function log(message: string): void {
console.error(`[agent-runner] ${message}`);
}
function compactBoundaryFromMessage(
message: unknown,
): RunnerCompaction | undefined {
if (
!(
typeof message === 'object' &&
message !== null &&
(message as { type?: string }).type === 'system' &&
(message as { subtype?: string }).subtype === 'compact_boundary'
)
) {
return undefined;
}
const meta = (
message as {
compact_metadata?: { trigger?: string; pre_tokens?: number };
}
).compact_metadata;
log(
`Compact boundary — trigger=${meta?.trigger || '?'} pre_tokens=${meta?.pre_tokens ?? '?'}`,
);
return compactBoundaryOutput(meta?.trigger);
}
// 번들 CLI binary를 명시해야 SDK가 musl/glibc 잘못 탐색하는 걸 우회함.
// (SDK 0.2.114의 W7() 헬퍼는 linux-x64-musl를 linux-x64보다 먼저 시도하므로
// glibc 호스트에서 musl 패키지가 빈 껍데기로 설치돼 있으면 실패한다.)
let cachedClaudeCliPath: string | null = null;
function getClaudeCliPath(): string {
if (cachedClaudeCliPath) return cachedClaudeCliPath;
cachedClaudeCliPath = resolveBundledClaudeCodeExecutable();
log(`Resolved bundled Claude Code CLI: ${cachedClaudeCliPath}`);
return cachedClaudeCliPath;
}
// Graceful shutdown: SIGTERM → abort SDK query, allowing cleanup
const agentAbortController = new AbortController();
process.on('SIGTERM', () => {
@@ -136,509 +67,6 @@ process.on('SIGTERM', () => {
agentAbortController.abort();
});
/**
* Run a single query and stream results via writeOutput.
* Uses MessageStream (AsyncIterable) to keep isSingleUserTurn=false,
* allowing agent teams subagents to run to completion.
* Also pipes IPC messages into the stream during the query.
*/
async function runQuery(
prompt: string,
sessionId: string | undefined,
mcpServerPath: string,
runnerInput: RunnerInput,
sdkEnv: Record<string, string | undefined>,
reviewerRuntime: boolean,
claudeReadonlyReviewerRuntime: boolean,
claudeReadonlySandboxMode: 'strict' | 'best-effort' | null,
): Promise<{
newSessionId?: string;
closedDuringQuery: boolean;
terminalResultObserved: boolean;
compaction?: RunnerCompaction;
}> {
const stream = new MessageStream((text) => buildMultimodalContent(text, log));
stream.push(prompt);
// Poll IPC for follow-up messages and _close sentinel during the query
let ipcPolling = true;
let closedDuringQuery = false;
const pollIpcDuringQuery = () => {
if (!ipcPolling) return;
if (shouldClose(IPC_INPUT_CLOSE_SENTINEL)) {
log('Close sentinel detected during query, ending stream');
// Flush any buffered text before closing — the for-await loop may not
// reach the post-loop flush code after stream.end().
if (pendingProgressText && !terminalResultObserved) {
log(
`Flushing pending text before close (${pendingProgressText.length} chars)`,
);
writeOutput({
status: 'success',
...normalizeStructuredOutput(pendingProgressText),
newSessionId,
});
pendingProgressText = null;
terminalResultObserved = true;
resultCount++;
}
closedDuringQuery = true;
stream.end();
ipcPolling = false;
return;
}
const messages = drainIpcInput(IPC_INPUT_DIR, log);
for (const text of messages) {
log(`Piping IPC message into active query (${text.length} chars)`);
stream.push(text);
}
setTimeout(pollIpcDuringQuery, IPC_POLL_MS);
};
setTimeout(pollIpcDuringQuery, IPC_POLL_MS);
let newSessionId: string | undefined;
let messageCount = 0;
let resultCount = 0;
let terminalResultObserved = false;
let pendingProgressText: string | null = null;
let compaction: RunnerCompaction | undefined;
const trackedAgentTasks = new TopLevelAgentTaskTracker();
// Discover additional directories
const extraDirs: string[] = [];
// When WORK_DIR is set, use it as cwd and include GROUP_DIR as additional directory
const effectiveCwd = WORK_DIR || GROUP_DIR;
if (WORK_DIR && WORK_DIR !== GROUP_DIR) {
extraDirs.push(GROUP_DIR);
log(
`Work directory override: ${WORK_DIR} (group dir added to additionalDirectories)`,
);
}
if (extraDirs.length > 0) {
log(`Additional directories: ${extraDirs.join(', ')}`);
}
// Model and thinking configuration from environment
const model = process.env.CLAUDE_MODEL || undefined;
const thinkingType = process.env.CLAUDE_THINKING || undefined; // 'adaptive' | 'enabled' | 'disabled'
const thinkingBudget = process.env.CLAUDE_THINKING_BUDGET
? parseInt(process.env.CLAUDE_THINKING_BUDGET, 10)
: undefined;
const effort =
(process.env.CLAUDE_EFFORT as 'low' | 'medium' | 'high' | 'max') ||
undefined;
const thinking =
thinkingType === 'adaptive'
? { type: 'adaptive' as const }
: thinkingType === 'enabled'
? { type: 'enabled' as const, budgetTokens: thinkingBudget }
: thinkingType === 'disabled'
? { type: 'disabled' as const }
: undefined;
if (model) log(`Using model: ${model}`);
if (thinking) log(`Thinking config: ${JSON.stringify(thinking)}`);
if (effort) log(`Effort: ${effort}`);
if (reviewerRuntime) log('Reviewer runtime restrictions enabled');
if (claudeReadonlyReviewerRuntime) {
log(
`Claude host reviewer read-only sandbox enabled (${claudeReadonlySandboxMode || 'unknown'})`,
);
}
if (claudeReadonlySandboxMode === 'best-effort') {
log(
'Claude host reviewer sandbox capability unavailable on this host, using best-effort read-only mode',
);
}
const readonlyReviewerRuntime =
reviewerRuntime || claudeReadonlyReviewerRuntime;
const allowedTools = readonlyReviewerRuntime
? [
'Bash',
'Read',
'Glob',
'Grep',
'WebSearch',
'WebFetch',
'Task',
'TaskOutput',
'TaskStop',
'TeamCreate',
'TeamDelete',
'SendMessage',
'TodoWrite',
'ToolSearch',
'Skill',
'mcp__ejclaw__*',
]
: [
'Bash',
'Read',
'Write',
'Edit',
'Glob',
'Grep',
'WebSearch',
'WebFetch',
'Task',
'TaskOutput',
'TaskStop',
'TeamCreate',
'TeamDelete',
'SendMessage',
'TodoWrite',
'ToolSearch',
'Skill',
'NotebookEdit',
'mcp__ejclaw__*',
];
const readonlyProtectedPaths = [effectiveCwd, ...extraDirs].filter(
(value): value is string => Boolean(value),
);
const claudeReadonlySandboxSettings =
claudeReadonlyReviewerRuntime && claudeReadonlySandboxMode
? buildClaudeReadonlySandboxSettings(
readonlyProtectedPaths,
undefined,
claudeReadonlySandboxMode,
)
: undefined;
for await (const message of query({
prompt: stream,
options: {
// 번들 CLI binary를 명시해야 SDK가 musl/glibc 잘못 탐색하는 걸 우회함.
pathToClaudeCodeExecutable: getClaudeCliPath(),
cwd: effectiveCwd,
model,
thinking,
effort,
additionalDirectories: extraDirs.length > 0 ? extraDirs : undefined,
resume: sessionId,
allowedTools,
env: sdkEnv,
permissionMode: 'bypassPermissions',
allowDangerouslySkipPermissions: true,
...(claudeReadonlySandboxSettings
? {
sandbox: claudeReadonlySandboxSettings,
}
: {}),
settingSources: ['project', 'user'],
abortController: agentAbortController,
mcpServers: {
ejclaw: buildEjclawMcpServerConfig(mcpServerPath, {
chatJid: runnerInput.chatJid,
groupFolder: runnerInput.groupFolder,
isMain: runnerInput.isMain,
agentType: process.env[EJCLAW_ENV.agentType] || 'claude-code',
roomRole: runnerInput.roomRoleContext?.role || '',
ipcDir: process.env[EJCLAW_ENV.ipcDir],
hostIpcDir: process.env[EJCLAW_ENV.hostIpcDir],
}),
},
hooks: {
PreCompact: [
{
hooks: [
createPreCompactHook({
assistantName: runnerInput.assistantName,
groupDir: GROUP_DIR,
groupFolder: GROUP_FOLDER,
hostTasksDir: HOST_TASKS_DIR,
log,
writeOutput,
}),
],
},
],
PreToolUse: [
{
matcher: 'Bash',
hooks: readonlyReviewerRuntime
? [createReviewerBashGuardHook(), createSanitizeBashHook()]
: [createSanitizeBashHook()],
},
],
},
agentProgressSummaries: true,
},
})) {
messageCount++;
const msgType =
message.type === 'system'
? `system/${(message as { subtype?: string }).subtype}`
: message.type;
log(`[msg #${messageCount}] type=${msgType}`);
// Flush pending intermediate text as a regular message on non-assistant events.
if (message.type !== 'assistant' && pendingProgressText) {
writeOutput({
status: 'success',
phase: 'intermediate',
...normalizeStructuredOutput(pendingProgressText),
newSessionId,
});
pendingProgressText = null;
}
if (message.type === 'system' && message.subtype === 'init') {
newSessionId = message.session_id;
log(`Session initialized: ${newSessionId}`);
}
compaction = compactBoundaryFromMessage(message) ?? compaction;
if (
message.type === 'system' &&
(message as { subtype?: string }).subtype === 'task_notification'
) {
const tn = message as {
task_id: string;
tool_use_id?: string;
status: string;
summary: string;
};
log(
`Task notification: task=${tn.task_id} status=${tn.status} summary=${tn.summary}`,
);
const mapped = buildTaskNotificationOutput(
trackedAgentTasks,
tn,
newSessionId,
);
if (mapped) {
writeOutput(mapped);
}
}
if (
message.type === 'system' &&
(message as { subtype?: string }).subtype === 'task_progress'
) {
const tp = message as Record<string, unknown>;
const description =
typeof tp.description === 'string' ? tp.description : '';
const mapped = buildTaskProgressOutput(
trackedAgentTasks,
tp,
newSessionId,
);
if (mapped) {
writeOutput(mapped);
} else if (description) {
// Long AI summary → skip (too long for progress sub-line)
log(
`Skipping long task_progress description (${description.length} chars)`,
);
}
}
if (
message.type === 'system' &&
(message as { subtype?: string }).subtype === 'task_started'
) {
const ts = message as { task_id: string; description?: string };
const desc = ts.description || '';
log(`Subagent started: task=${ts.task_id} desc=${desc.slice(0, 200)}`);
const mapped = buildTaskStartedOutput(
trackedAgentTasks,
ts,
newSessionId,
);
if (mapped) {
writeOutput(mapped);
}
}
if (message.type === 'tool_progress') {
const tp = message as {
tool_name: string;
elapsed_time_seconds: number;
};
const label = `${tp.tool_name} (${Math.round(tp.elapsed_time_seconds)}s)`;
log(`Tool progress: ${label}`);
const normalized = normalizeStructuredOutput(label);
writeOutput({
status: 'success',
phase: 'progress',
...normalized,
newSessionId,
});
}
if (message.type === 'tool_use_summary') {
const ts = message as { summary: string };
log(`Tool use summary: ${ts.summary.slice(0, 200)}`);
const normalized = normalizeStructuredOutput(ts.summary);
writeOutput({
status: 'success',
phase: 'progress',
...normalized,
newSessionId,
});
}
if (message.type === 'result') {
resultCount++;
let textResult =
'result' in message ? (message as { result?: string }).result : null;
const isError = message.subtype?.startsWith('error');
// Discard pending progress if it matches the final result (prevent duplicate)
if (
pendingProgressText &&
textResult &&
pendingProgressText === textResult
) {
log(`Discarding pending progress (matches result)`);
pendingProgressText = null;
} else if (pendingProgressText) {
// If the result has no text, promote pending progress to the result
// so it gets delivered as the final output instead of being lost.
if (!textResult) {
log(
`Promoting pending progress text to result (${pendingProgressText.length} chars)`,
);
textResult = pendingProgressText;
} else {
writeOutput({
status: 'success',
phase: 'intermediate',
result: pendingProgressText,
newSessionId,
});
}
pendingProgressText = null;
}
log(
`Result #${resultCount}: subtype=${message.subtype}${textResult ? ` text=${textResult.slice(0, 200)}` : ''}`,
);
if (isError) {
// Log full error details for debugging
const msg = message as Record<string, unknown>;
const sdkErrors = Array.isArray(msg.errors)
? (msg.errors as string[])
: [];
const errorDetail = JSON.stringify({
subtype: message.subtype,
result: textResult?.slice(0, 500),
errors: sdkErrors,
stop_reason: msg.stop_reason,
duration_ms: msg.duration_ms,
duration_api_ms: msg.duration_api_ms,
session_id: msg.session_id,
});
log(`Error result detail: ${errorDetail}`);
// Pass SDK errors through so host can detect session issues
const errorText =
sdkErrors.length > 0 ? sdkErrors.join('; ') : undefined;
writeOutput({
status: 'error',
result: textResult || null,
newSessionId,
error: errorText || `Agent error: ${message.subtype}`,
...buildCompactionOutput(compaction),
});
} else {
const normalized = normalizeStructuredOutput(textResult || null);
writeOutput({
status: 'success',
...normalized,
newSessionId,
...buildCompactionOutput(compaction),
});
}
// Single-turn runtimes must terminate the query after the first
// terminal result. Leaving the message stream open can keep the SDK
// query alive indefinitely, which pins the host queue after a reply.
terminalResultObserved = true;
ipcPolling = false;
stream.end();
log('Terminal result observed, ending query stream');
break;
}
if (message.type === 'assistant') {
trackedAgentTasks.rememberAssistantMessage(message);
const stopReason = (message as { stop_reason?: string }).stop_reason;
const textResult = extractAssistantText(message);
// Only log when there's something interesting (text or terminal)
if (textResult || stopReason === 'end_turn') {
log(
`Assistant: stop=${stopReason} text=${textResult ? textResult.length + ' chars' : 'null'}`,
);
}
if (stopReason === 'end_turn' && textResult) {
resultCount++;
log(
`Terminal assistant turn observed without result event (${textResult.length} chars), ending query stream`,
);
writeOutput({
status: 'success',
...normalizeStructuredOutput(textResult),
newSessionId,
...buildCompactionOutput(compaction),
});
terminalResultObserved = true;
ipcPolling = false;
stream.end();
break;
}
// Intermediate assistant text between tool calls → buffer as pending progress.
// Don't emit immediately — if the next message is a result with the same text,
// this would cause a duplicate. The pending text is flushed when the next
// non-result message arrives, or discarded if result matches.
if (stopReason !== 'end_turn' && textResult) {
// Flush previous pending as a regular message (not progress heading)
if (pendingProgressText) {
const normalized = normalizeStructuredOutput(pendingProgressText);
writeOutput({
status: 'success',
phase: 'intermediate',
...normalized,
newSessionId,
});
}
pendingProgressText = textResult;
log(
`Intermediate assistant text buffered (${textResult.length} chars, stop=${stopReason})`,
);
}
}
}
// Flush any remaining buffered text that was never followed by a result event.
// This happens when the agent produces a short response without a formal end_turn.
if (pendingProgressText && !terminalResultObserved) {
log(
`Flushing remaining pending progress text as final output (${pendingProgressText.length} chars)`,
);
writeOutput({
status: 'success',
...normalizeStructuredOutput(pendingProgressText),
newSessionId,
...buildCompactionOutput(compaction),
});
terminalResultObserved = true;
resultCount++;
}
ipcPolling = false;
log(
`Query done. Messages: ${messageCount}, results: ${resultCount}, closedDuringQuery: ${closedDuringQuery}`,
);
return {
newSessionId,
closedDuringQuery,
terminalResultObserved,
compaction,
};
}
async function main(): Promise<void> {
let runnerInput: RunnerInput;
@@ -723,126 +151,18 @@ async function main(): Promise<void> {
if (isSessionSlashCommand) {
log(`Handling session command: ${trimmedPrompt}`);
let slashSessionId: string | undefined;
let compactBoundarySeen = false;
let slashCompaction: RunnerCompaction | undefined;
let hadError = false;
let resultEmitted = false;
try {
for await (const message of query({
prompt: trimmedPrompt,
options: {
// 번들 CLI binary를 명시해야 SDK가 musl/glibc 잘못 탐색하는 걸 우회함.
pathToClaudeCodeExecutable: getClaudeCliPath(),
cwd: mainEffectiveCwd,
resume: sessionId,
systemPrompt: undefined,
allowedTools: [],
env: guardedSdkEnv,
permissionMode: 'bypassPermissions' as const,
allowDangerouslySkipPermissions: true,
settingSources: ['project', 'user'] as const,
abortController: agentAbortController,
hooks: {
PreCompact: [
{
hooks: [
createPreCompactHook({
assistantName: runnerInput.assistantName,
groupDir: GROUP_DIR,
groupFolder: GROUP_FOLDER,
hostTasksDir: HOST_TASKS_DIR,
log,
writeOutput,
}),
],
},
],
},
},
})) {
const msgType =
message.type === 'system'
? `system/${(message as { subtype?: string }).subtype}`
: message.type;
log(`[slash-cmd] type=${msgType}`);
if (message.type === 'system' && message.subtype === 'init') {
slashSessionId = message.session_id;
log(`Session after slash command: ${slashSessionId}`);
}
const observedCompaction = compactBoundaryFromMessage(message);
if (observedCompaction) {
compactBoundarySeen = true;
slashCompaction = observedCompaction;
}
if (message.type === 'result') {
const resultSubtype = (message as { subtype?: string }).subtype;
const textResult =
'result' in message
? (message as { result?: string }).result
: null;
if (resultSubtype?.startsWith('error')) {
hadError = true;
writeOutput({
status: 'error',
result: null,
error: textResult || 'Session command failed.',
newSessionId: slashSessionId,
...buildCompactionOutput(slashCompaction),
});
} else {
writeOutput({
status: 'success',
result: textResult || 'Conversation compacted.',
newSessionId: slashSessionId,
...buildCompactionOutput(slashCompaction),
});
}
resultEmitted = true;
}
}
} catch (err) {
hadError = true;
const errorMsg = err instanceof Error ? err.message : String(err);
log(`Slash command error: ${errorMsg}`);
writeOutput({ status: 'error', result: null, error: errorMsg });
}
log(
`Slash command done. compactBoundarySeen=${compactBoundarySeen}, hadError=${hadError}`,
);
// Warn if compact_boundary was never observed — compaction may not have occurred
if (!hadError && !compactBoundarySeen) {
log(
'WARNING: compact_boundary was not observed. Compaction may not have completed.',
);
}
// Only emit final session marker if no result was emitted yet and no error occurred
if (!resultEmitted && !hadError) {
writeOutput({
status: 'success',
result: compactBoundarySeen
? 'Conversation compacted.'
: 'Compaction requested but compact_boundary was not observed.',
newSessionId: slashSessionId,
...buildCompactionOutput(slashCompaction),
});
} else if (!hadError) {
// Emit session-only marker so host updates session tracking
writeOutput({
status: 'success',
result: null,
newSessionId: slashSessionId,
...buildCompactionOutput(slashCompaction),
});
}
await runSessionCommand({
prompt: trimmedPrompt,
sessionId,
cwd: mainEffectiveCwd,
sdkEnv: guardedSdkEnv,
abortController: agentAbortController,
assistantName: runnerInput.assistantName,
groupDir: GROUP_DIR,
groupFolder: GROUP_FOLDER,
hostTasksDir: HOST_TASKS_DIR,
log,
});
return;
}
// --- End slash command handling ---
@@ -850,16 +170,26 @@ async function main(): Promise<void> {
try {
log(`Starting query (session: ${sessionId || 'new'})...`);
const queryResult = await runQuery(
const queryResult = await runClaudeQuery({
prompt,
sessionId,
mcpServerPath,
runnerInput,
guardedSdkEnv,
sdkEnv: guardedSdkEnv,
reviewerRuntime,
claudeReadonlyReviewerRuntime,
claudeReadonlySandboxMode,
);
abortController: agentAbortController,
paths: {
groupDir: GROUP_DIR,
groupFolder: GROUP_FOLDER,
hostTasksDir: HOST_TASKS_DIR,
ipcInputCloseSentinel: IPC_INPUT_CLOSE_SENTINEL,
ipcInputDir: IPC_INPUT_DIR,
workDir: WORK_DIR,
},
log,
});
if (queryResult.newSessionId) {
sessionId = queryResult.newSessionId;
}