Files
Dirigent/plugin/hooks/agent-end.ts
hzhang d8ac9ee0f9 fix: correct dormancy detection — isEmptyTurn last-line + blocked_pending drain
Two bugs that prevented turn-manager dormancy from ever triggering:

1. isEmptyTurn too strict: agents output multi-line text ending with
   "NO_REPLY" on the last line, but the regex ^NO_REPLY$ required the
   entire string to match. Now checks only the last non-empty line.

2. blocked_pending counter inflation: non-speaker suppressions incremented
   the counter but their stale NO_REPLYs were discarded at the
   !isCurrentSpeaker early return without decrementing. Over a full cycle
   the counter inflated by the number of suppressions, causing the agent's
   real empty turn to be misidentified as stale when it finally arrived.
   Fix: at both early-return points in agent_end (!isCurrentSpeaker and
   !isTurnPending), drain blocked_pending when the turn is empty.

Also fixed: pollForTailMatch now uses any-message detection (instead of
tail-fingerprint content matching) with a 30 s timeout, avoiding infinite
polling when agents send concise Discord messages after verbose LLM output.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-09 15:50:19 +01:00

264 lines
11 KiB
TypeScript

import type { OpenClawPluginApi } from "openclaw/plugin-sdk";
import type { ChannelStore } from "../core/channel-store.js";
import type { IdentityRegistry } from "../core/identity-registry.js";
import { parseDiscordChannelId } from "./before-model-resolve.js";
import {
isCurrentSpeaker,
getAnchor,
advanceSpeaker,
wakeFromDormant,
isDormant,
hasSpeakers,
getDebugInfo,
isTurnPending,
clearTurnPending,
consumeBlockedPending,
type SpeakerEntry,
} from "../turn-manager.js";
import { fetchVisibleChannelBotAccountIds } from "../core/channel-members.js";
import { pollForTailMatch, sendScheduleTrigger } from "../core/moderator-discord.js";
/**
* Process-level deduplication for agent_end events.
* OpenClaw hot-reloads plugin modules (re-imports), stacking duplicate handlers.
* Using globalThis ensures the dedup Set survives module reloads and is shared
* by all handler instances in the same process.
*/
const _AGENT_END_DEDUP_KEY = "_dirigentProcessedAgentEndRunIds";
if (!(globalThis as Record<string, unknown>)[_AGENT_END_DEDUP_KEY]) {
(globalThis as Record<string, unknown>)[_AGENT_END_DEDUP_KEY] = new Set<string>();
}
const processedAgentEndRunIds: Set<string> = (globalThis as Record<string, unknown>)[_AGENT_END_DEDUP_KEY] as Set<string>;
/**
* Per-channel advance lock: prevents two concurrent agent_end handler instances
* (from different VM contexts with different runIds) from both calling advanceSpeaker
* for the same channel at the same time, which would double-advance the speaker index.
*/
const _ADVANCE_LOCK_KEY = "_dirigentAdvancingChannels";
if (!(globalThis as Record<string, unknown>)[_ADVANCE_LOCK_KEY]) {
(globalThis as Record<string, unknown>)[_ADVANCE_LOCK_KEY] = new Set<string>();
}
const advancingChannels: Set<string> = (globalThis as Record<string, unknown>)[_ADVANCE_LOCK_KEY] as Set<string>;
/** Extract plain text from agent_end event.messages last assistant entry. */
function extractFinalText(messages: unknown[]): string {
for (let i = messages.length - 1; i >= 0; i--) {
const msg = messages[i] as Record<string, unknown> | null;
if (!msg || msg.role !== "assistant") continue;
const content = msg.content;
if (typeof content === "string") return content;
if (Array.isArray(content)) {
let text = "";
for (const part of content) {
const p = part as Record<string, unknown>;
if (p?.type === "text" && typeof p.text === "string") text += p.text;
}
return text;
}
break;
}
return "";
}
function isEmptyTurn(text: string): boolean {
const t = text.trim();
if (t === "") return true;
// Check if the last non-empty line is NO or NO_REPLY (agents often write
// explanatory text before the final answer on the last line).
const lines = t.split("\n").map((l) => l.trim()).filter((l) => l !== "");
const last = lines[lines.length - 1] ?? "";
return /^NO$/i.test(last) || /^NO_REPLY$/i.test(last);
}
export type AgentEndDeps = {
api: OpenClawPluginApi;
channelStore: ChannelStore;
identityRegistry: IdentityRegistry;
moderatorBotToken: string | undefined;
scheduleIdentifier: string;
debugMode: boolean;
/** Called when discussion channel enters dormant — to send idle reminder. */
onDiscussionDormant?: (channelId: string) => Promise<void>;
};
/** Exposed so message-received can interrupt an in-progress tail-match wait. */
export type InterruptFn = (channelId: string) => void;
export function registerAgentEndHook(deps: AgentEndDeps): InterruptFn {
const { api, channelStore, identityRegistry, moderatorBotToken, scheduleIdentifier, debugMode, onDiscussionDormant } = deps;
const interruptedChannels = new Set<string>();
function interrupt(channelId: string): void {
interruptedChannels.add(channelId);
setTimeout(() => interruptedChannels.delete(channelId), 5000);
}
async function buildSpeakerList(channelId: string): Promise<SpeakerEntry[]> {
if (!moderatorBotToken) return [];
const agentIds = await fetchVisibleChannelBotAccountIds(api, channelId, identityRegistry);
const result: SpeakerEntry[] = [];
for (const agentId of agentIds) {
const identity = identityRegistry.findByAgentId(agentId);
if (identity) result.push({ agentId, discordUserId: identity.discordUserId });
}
return result;
}
async function triggerNextSpeaker(channelId: string, next: SpeakerEntry): Promise<void> {
if (!moderatorBotToken) return;
const msg = `<@${next.discordUserId}>${scheduleIdentifier}`;
await sendScheduleTrigger(moderatorBotToken, channelId, msg, api.logger, debugMode);
api.logger.info(`dirigent: triggered next speaker agentId=${next.agentId} channel=${channelId}`);
}
api.on("agent_end", async (event, ctx) => {
try {
// Deduplicate: skip if this runId was already processed by another handler
// instance (can happen when OpenClaw hot-reloads the plugin in the same process).
const runId = (event as Record<string, unknown>).runId as string | undefined;
if (runId) {
if (processedAgentEndRunIds.has(runId)) return;
processedAgentEndRunIds.add(runId);
// Evict old entries to prevent unbounded growth (keep last 500)
if (processedAgentEndRunIds.size > 500) {
const oldest = processedAgentEndRunIds.values().next().value;
if (oldest) processedAgentEndRunIds.delete(oldest);
}
}
const sessionKey = ctx.sessionKey;
if (!sessionKey) return;
const channelId = parseDiscordChannelId(sessionKey);
if (!channelId) return;
const mode = channelStore.getMode(channelId);
if (mode === "none" || mode === "work" || mode === "report") return;
if (!hasSpeakers(channelId)) return;
const agentId = ctx.agentId;
if (!agentId) return;
// Extract final text early so we can drain blocked_pending at every early-return point.
// This prevents the counter from staying inflated when stale NO_REPLYs are discarded
// by !isCurrentSpeaker or !isTurnPending without reaching consumeBlockedPending.
const messages = Array.isArray((event as Record<string, unknown>).messages)
? ((event as Record<string, unknown>).messages as unknown[])
: [];
const finalText = extractFinalText(messages);
const empty = isEmptyTurn(finalText);
if (!isCurrentSpeaker(channelId, agentId)) {
// Drain blocked_pending for non-speaker stale NO_REPLYs. Without this, suppressions
// that happen while this agent is not the current speaker inflate the counter and cause
// its subsequent real empty turn to be misidentified as stale.
if (empty) consumeBlockedPending(channelId, agentId);
return;
}
// Only process agent_ends for turns that were explicitly started by before_model_resolve.
// This prevents stale NO_REPLY completions (from initial suppression) from being counted.
if (!isTurnPending(channelId, agentId)) {
api.logger.info(`dirigent: agent_end skipping stale turn agentId=${agentId} channel=${channelId}`);
// Also drain here: stale may arrive after clearTurnPending but before markTurnStarted.
if (empty) consumeBlockedPending(channelId, agentId);
return;
}
// Consume a blocked-pending slot only for self-wakeup stales: NO_REPLY completions
// from suppressed self-wakeup before_model_resolve calls that fire while the agent
// is current speaker with a turn already in progress.
if (empty && consumeBlockedPending(channelId, agentId)) {
api.logger.info(`dirigent: agent_end skipping blocked-pending NO_REPLY agentId=${agentId} channel=${channelId}`);
return;
}
clearTurnPending(channelId, agentId);
api.logger.info(
`dirigent: agent_end channel=${channelId} agentId=${agentId} empty=${empty} text=${finalText.slice(0, 80)}`,
);
if (!empty) {
// Real turn: wait for Discord delivery before triggering next speaker.
// Anchor was set in before_model_resolve just before the LLM call, so any
// message from the agent after the anchor must be from this turn.
const identity = identityRegistry.findByAgentId(agentId);
if (identity && moderatorBotToken) {
const anchorId = getAnchor(channelId, agentId) ?? "0";
const { matched: _matched, interrupted } = await pollForTailMatch({
token: moderatorBotToken,
channelId,
anchorId,
agentDiscordUserId: identity.discordUserId,
isInterrupted: () => interruptedChannels.has(channelId),
});
if (interrupted) {
if (isDormant(channelId)) {
// Channel is dormant: a new external message woke it — restart from first speaker
api.logger.info(`dirigent: tail-match interrupted (dormant) channel=${channelId} — waking`);
const first = wakeFromDormant(channelId);
if (first) await triggerNextSpeaker(channelId, first);
return;
}
// Not dormant: interrupt was a spurious trigger (e.g. moderator bot message).
// Fall through to normal advance so the turn cycle continues correctly.
api.logger.info(`dirigent: tail-match interrupted (non-dormant) channel=${channelId} — advancing normally`);
}
}
}
// Guard against concurrent advanceSpeaker calls for the same channel.
// Two VM context instances may both reach this point with different runIds
// (when the dedup Set doesn't catch them); only the first should advance.
if (advancingChannels.has(channelId)) {
api.logger.info(`dirigent: agent_end advance already in progress, skipping channel=${channelId} agentId=${agentId}`);
return;
}
advancingChannels.add(channelId);
let next: ReturnType<typeof import("../turn-manager.js").getCurrentSpeaker> | null = null;
let enteredDormant = false;
try {
// Determine shuffle mode from current list size
const debugBefore = getDebugInfo(channelId);
const currentListSize = debugBefore.exists ? (debugBefore.speakerList?.length ?? 0) : 0;
const isShuffle = currentListSize > 2;
// In shuffle mode, pass agentId as previousLastAgentId so new list avoids it as first
const previousLastAgentId = isShuffle ? agentId : undefined;
({ next, enteredDormant } = await advanceSpeaker(
channelId,
agentId,
empty,
() => buildSpeakerList(channelId),
previousLastAgentId,
));
} finally {
advancingChannels.delete(channelId);
}
if (enteredDormant) {
api.logger.info(`dirigent: channel=${channelId} entered dormant`);
if (mode === "discussion") {
await onDiscussionDormant?.(channelId).catch((err) => {
api.logger.warn(`dirigent: onDiscussionDormant failed: ${String(err)}`);
});
}
return;
}
if (next) await triggerNextSpeaker(channelId, next);
} catch (err) {
api.logger.warn(`dirigent: agent_end hook error: ${String(err)}`);
}
});
return interrupt;
}