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>
This commit is contained in:
h z
2026-04-09 15:50:19 +01:00
parent 9e61af4a16
commit d8ac9ee0f9
3 changed files with 184 additions and 98 deletions

View File

@@ -16,10 +16,7 @@ import {
type SpeakerEntry,
} from "../turn-manager.js";
import { fetchVisibleChannelBotAccountIds } from "../core/channel-members.js";
import { pollForTailMatch, sendAndDelete } from "../core/moderator-discord.js";
const TAIL_LENGTH = 40;
const TAIL_MATCH_TIMEOUT_MS = 15_000;
import { pollForTailMatch, sendScheduleTrigger } from "../core/moderator-discord.js";
/**
* Process-level deduplication for agent_end events.
@@ -33,6 +30,17 @@ if (!(globalThis as Record<string, unknown>)[_AGENT_END_DEDUP_KEY]) {
}
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--) {
@@ -55,7 +63,12 @@ function extractFinalText(messages: unknown[]): string {
function isEmptyTurn(text: string): boolean {
const t = text.trim();
return t === "" || /^NO$/i.test(t) || /^NO_REPLY$/i.test(t);
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 = {
@@ -64,6 +77,7 @@ export type AgentEndDeps = {
identityRegistry: IdentityRegistry;
moderatorBotToken: string | undefined;
scheduleIdentifier: string;
debugMode: boolean;
/** Called when discussion channel enters dormant — to send idle reminder. */
onDiscussionDormant?: (channelId: string) => Promise<void>;
};
@@ -72,7 +86,7 @@ export type AgentEndDeps = {
export type InterruptFn = (channelId: string) => void;
export function registerAgentEndHook(deps: AgentEndDeps): InterruptFn {
const { api, channelStore, identityRegistry, moderatorBotToken, scheduleIdentifier, onDiscussionDormant } = deps;
const { api, channelStore, identityRegistry, moderatorBotToken, scheduleIdentifier, debugMode, onDiscussionDormant } = deps;
const interruptedChannels = new Set<string>();
@@ -95,7 +109,7 @@ export function registerAgentEndHook(deps: AgentEndDeps): InterruptFn {
async function triggerNextSpeaker(channelId: string, next: SpeakerEntry): Promise<void> {
if (!moderatorBotToken) return;
const msg = `<@${next.discordUserId}>${scheduleIdentifier}`;
await sendAndDelete(moderatorBotToken, channelId, msg, api.logger);
await sendScheduleTrigger(moderatorBotToken, channelId, msg, api.logger, debugMode);
api.logger.info(`dirigent: triggered next speaker agentId=${next.agentId} channel=${channelId}`);
}
@@ -127,51 +141,60 @@ export function registerAgentEndHook(deps: AgentEndDeps): InterruptFn {
const agentId = ctx.agentId;
if (!agentId) return;
if (!isCurrentSpeaker(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}`);
return;
}
// Consume a blocked-pending slot if any exists. These are NO_REPLY completions
// from before_model_resolve blocking events (non-speaker or init-suppressed) that
// fire late — after the agent became the current speaker — due to history-building
// overhead (~10s). We skip them until the counter is exhausted, at which point
// the next agent_end is the real LLM response.
if (consumeBlockedPending(channelId, agentId)) {
api.logger.info(`dirigent: agent_end skipping blocked-pending NO_REPLY agentId=${agentId} channel=${channelId}`);
return;
}
clearTurnPending(channelId, agentId);
// 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 via tail-match polling
// 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 tail = [...finalText].slice(-TAIL_LENGTH).join("");
const { matched, interrupted } = await pollForTailMatch({
const { matched: _matched, interrupted } = await pollForTailMatch({
token: moderatorBotToken,
channelId,
anchorId,
agentDiscordUserId: identity.discordUserId,
tailFingerprint: tail,
timeoutMs: TAIL_MATCH_TIMEOUT_MS,
isInterrupted: () => interruptedChannels.has(channelId),
});
@@ -187,26 +210,38 @@ export function registerAgentEndHook(deps: AgentEndDeps): InterruptFn {
// Fall through to normal advance so the turn cycle continues correctly.
api.logger.info(`dirigent: tail-match interrupted (non-dormant) channel=${channelId} — advancing normally`);
}
if (!matched) {
api.logger.warn(`dirigent: tail-match timeout channel=${channelId} agentId=${agentId} — advancing anyway`);
}
}
}
// Determine shuffle mode from current list size
const debugBefore = getDebugInfo(channelId);
const currentListSize = debugBefore.exists ? (debugBefore.speakerList?.length ?? 0) : 0;
const isShuffle = currentListSize > 2;
// In shuffle mode, pass agentId as previousLastAgentId so new list avoids it as first
const previousLastAgentId = isShuffle ? agentId : undefined;
// Guard against concurrent advanceSpeaker calls for the same channel.
// Two VM context instances may both reach this point with different runIds
// (when the dedup Set doesn't catch them); only the first should advance.
if (advancingChannels.has(channelId)) {
api.logger.info(`dirigent: agent_end advance already in progress, skipping channel=${channelId} agentId=${agentId}`);
return;
}
advancingChannels.add(channelId);
const { next, enteredDormant } = await advanceSpeaker(
channelId,
agentId,
empty,
() => buildSpeakerList(channelId),
previousLastAgentId,
);
let next: ReturnType<typeof import("../turn-manager.js").getCurrentSpeaker> | null = null;
let enteredDormant = false;
try {
// Determine shuffle mode from current list size
const debugBefore = getDebugInfo(channelId);
const currentListSize = debugBefore.exists ? (debugBefore.speakerList?.length ?? 0) : 0;
const isShuffle = currentListSize > 2;
// In shuffle mode, pass agentId as previousLastAgentId so new list avoids it as first
const previousLastAgentId = isShuffle ? agentId : undefined;
({ next, enteredDormant } = await advanceSpeaker(
channelId,
agentId,
empty,
() => buildSpeakerList(channelId),
previousLastAgentId,
));
} finally {
advancingChannels.delete(channelId);
}
if (enteredDormant) {
api.logger.info(`dirigent: channel=${channelId} entered dormant`);

View File

@@ -1,8 +1,8 @@
import type { OpenClawPluginApi } from "openclaw/plugin-sdk";
import type { ChannelStore } from "../core/channel-store.js";
import type { IdentityRegistry } from "../core/identity-registry.js";
import { isCurrentSpeaker, setAnchor, hasSpeakers, isDormant, setSpeakerList, markTurnStarted, incrementBlockedPending, getInitializingChannels, type SpeakerEntry } from "../turn-manager.js";
import { getLatestMessageId, sendAndDelete } from "../core/moderator-discord.js";
import { isCurrentSpeaker, isTurnPending, setAnchor, hasSpeakers, isDormant, setSpeakerList, markTurnStarted, incrementBlockedPending, getInitializingChannels, type SpeakerEntry } from "../turn-manager.js";
import { getLatestMessageId, sendScheduleTrigger } from "../core/moderator-discord.js";
import { fetchVisibleChannelBotAccountIds } from "../core/channel-members.js";
/** Extract Discord channel ID from sessionKey like "agent:home-developer:discord:channel:1234567890". */
@@ -16,9 +16,10 @@ type Deps = {
channelStore: ChannelStore;
identityRegistry: IdentityRegistry;
moderatorBotToken: string | undefined;
noReplyModel: string;
noReplyProvider: string;
scheduleIdentifier: string;
debugMode: boolean;
noReplyProvider: string;
noReplyModel: string;
};
/**
@@ -34,9 +35,7 @@ if (!(globalThis as Record<string, unknown>)[_BMR_DEDUP_KEY]) {
const processedBeforeModelResolveEvents: WeakSet<object> = (globalThis as Record<string, unknown>)[_BMR_DEDUP_KEY] as WeakSet<object>;
export function registerBeforeModelResolveHook(deps: Deps): void {
const { api, channelStore, identityRegistry, moderatorBotToken, noReplyModel, noReplyProvider, scheduleIdentifier } = deps;
const NO_REPLY = { model: noReplyModel, provider: noReplyProvider, noReply: true } as const;
const { api, channelStore, identityRegistry, moderatorBotToken, scheduleIdentifier, debugMode, noReplyProvider, noReplyModel } = deps;
/** Shared init lock — see turn-manager.ts getInitializingChannels(). */
const initializingChannels = getInitializingChannels();
@@ -57,19 +56,23 @@ export function registerBeforeModelResolveHook(deps: Deps): void {
const mode = channelStore.getMode(channelId);
// dead mode: suppress all responses
if (mode === "report" || mode === "dead" as string) return NO_REPLY;
// concluded discussion: suppress all agent responses (auto-reply handled by message_received)
if (mode === "discussion") {
const rec = channelStore.getRecord(channelId);
if (rec.discussion?.concluded) return NO_REPLY;
// dead/report mode: suppress all via no-reply model
if (mode === "report" || mode === "dead" as string) {
return { modelOverride: noReplyModel, providerOverride: noReplyProvider };
}
// disabled modes: let agents respond freely
// concluded discussion: suppress via no-reply model
if (mode === "discussion") {
const rec = channelStore.getRecord(channelId);
if (rec.discussion?.concluded) {
return { modelOverride: noReplyModel, providerOverride: noReplyProvider };
}
}
// disabled modes: no turn management
if (mode === "none" || mode === "work") return;
// discussion / chat: check turn
// chat / discussion (active): check turn
const agentId = ctx.agentId;
if (!agentId) return;
@@ -78,7 +81,9 @@ export function registerBeforeModelResolveHook(deps: Deps): void {
// Only one concurrent initializer per channel (Node.js single-threaded: this is safe)
if (initializingChannels.has(channelId)) {
api.logger.info(`dirigent: before_model_resolve init in progress, suppressing agentId=${agentId} channel=${channelId}`);
return NO_REPLY;
// incrementBlockedPending so agent_end knows to expect a stale NO_REPLY completion later
incrementBlockedPending(channelId, agentId);
return { modelOverride: noReplyModel, providerOverride: noReplyProvider };
}
initializingChannels.add(channelId);
try {
@@ -95,12 +100,13 @@ export function registerBeforeModelResolveHook(deps: Deps): void {
const first = speakers[0];
api.logger.info(`dirigent: initialized speaker list channel=${channelId} first=${first.agentId} all=${speakers.map(s => s.agentId).join(",")}`);
// If this agent is NOT the first speaker, trigger first speaker and suppress this one
// If this agent is NOT the first speaker, trigger first speaker and suppress self
if (first.agentId !== agentId && moderatorBotToken) {
await sendAndDelete(moderatorBotToken, channelId, `<@${first.discordUserId}>${scheduleIdentifier}`, api.logger);
return NO_REPLY;
await sendScheduleTrigger(moderatorBotToken, channelId, `<@${first.discordUserId}>${scheduleIdentifier}`, api.logger, debugMode);
incrementBlockedPending(channelId, agentId);
return { modelOverride: noReplyModel, providerOverride: noReplyProvider };
}
// If this agent IS the first speaker, fall through to normal turn logic
// Fall through — this agent IS the first speaker
} else {
// No registered agents visible — let everyone respond freely
return;
@@ -113,13 +119,25 @@ export function registerBeforeModelResolveHook(deps: Deps): void {
}
}
// If channel is dormant: suppress all agents
if (isDormant(channelId)) return NO_REPLY;
// Channel is dormant: suppress via no-reply model
if (isDormant(channelId)) {
api.logger.info(`dirigent: before_model_resolve suppressing dormant agentId=${agentId} channel=${channelId}`);
incrementBlockedPending(channelId, agentId);
return { modelOverride: noReplyModel, providerOverride: noReplyProvider };
}
if (!isCurrentSpeaker(channelId, agentId)) {
api.logger.info(`dirigent: before_model_resolve blocking non-speaker session=${sessionKey} agentId=${agentId} channel=${channelId}`);
api.logger.info(`dirigent: before_model_resolve suppressing non-speaker agentId=${agentId} channel=${channelId}`);
incrementBlockedPending(channelId, agentId);
return NO_REPLY;
return { modelOverride: noReplyModel, providerOverride: noReplyProvider };
}
// If a turn is already in progress for this agent, this is a duplicate wakeup
// (e.g. agent woke itself via a message-tool send). Suppress it.
if (isTurnPending(channelId, agentId)) {
api.logger.info(`dirigent: before_model_resolve turn already in progress, suppressing self-wakeup agentId=${agentId} channel=${channelId}`);
incrementBlockedPending(channelId, agentId);
return { modelOverride: noReplyModel, providerOverride: noReplyProvider };
}
// Mark that this is a legitimate turn (guards agent_end against stale NO_REPLY completions)