Files
nanoclaw/container/agent-runner/src/poll-loop.ts
gavrielc 350d9631fa Merge pull request #2327 from glifocat/wip/compaction-destination-reminder
fix: inject destination reminder after SDK auto-compaction
2026-05-07 23:57:29 +03:00

510 lines
20 KiB
TypeScript

import { findByName, type DestinationEntry } from './destinations.js';
import { getPendingMessages, markProcessing, markCompleted, type MessageInRow } from './db/messages-in.js';
import { writeMessageOut } from './db/messages-out.js';
import { getInboundDb, touchHeartbeat, clearStaleProcessingAcks } from './db/connection.js';
import {
clearContinuation,
migrateLegacyContinuation,
setContinuation,
} from './db/session-state.js';
import { formatMessages, extractRouting, categorizeMessage, isClearCommand, isRunnerCommand, stripInternalTags, type RoutingContext } from './formatter.js';
import type { AgentProvider, AgentQuery, ProviderEvent } from './providers/types.js';
const POLL_INTERVAL_MS = 1000;
const ACTIVE_POLL_INTERVAL_MS = 500;
function log(msg: string): void {
console.error(`[poll-loop] ${msg}`);
}
function generateId(): string {
return `msg-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
}
export interface PollLoopConfig {
provider: AgentProvider;
/**
* Name of the provider (e.g. "claude", "codex", "opencode"). Used to key
* the stored continuation per-provider so flipping providers doesn't
* resurrect a stale id from a different backend.
*/
providerName: string;
cwd: string;
systemContext?: {
instructions?: string;
};
}
/**
* Main poll loop. Runs indefinitely until the process is killed.
*
* 1. Poll messages_in for pending rows
* 2. Format into prompt, call provider.query()
* 3. While query active: continue polling, push new messages via provider.push()
* 4. On result: write messages_out
* 5. Mark messages completed
* 6. Loop
*/
export async function runPollLoop(config: PollLoopConfig): Promise<void> {
// Resume the agent's prior session from a previous container run if one
// was persisted. The continuation is opaque to the poll-loop — the
// provider decides how to use it (Claude resumes a .jsonl transcript,
// other providers may reload a thread ID, etc.). Keyed per-provider so
// a Codex thread id never gets handed to Claude or vice versa.
let continuation: string | undefined = migrateLegacyContinuation(config.providerName);
if (continuation) {
log(`Resuming agent session ${continuation}`);
}
// Clear leftover 'processing' acks from a previous crashed container.
// This lets the new container re-process those messages.
clearStaleProcessingAcks();
let pollCount = 0;
while (true) {
// Skip system messages — they're responses for MCP tools (e.g., ask_user_question)
const messages = getPendingMessages().filter((m) => m.kind !== 'system');
pollCount++;
// Periodic heartbeat so we know the loop is alive
if (pollCount % 30 === 0) {
log(`Poll heartbeat (${pollCount} iterations, ${messages.length} pending)`);
}
if (messages.length === 0) {
await sleep(POLL_INTERVAL_MS);
continue;
}
// Accumulate gate: if the batch contains only trigger=0 rows
// (context-only, router-stored under ignored_message_policy='accumulate'),
// don't wake the agent. Leave them `pending` — they'll ride along the
// next time a real trigger=1 message lands via this same getPendingMessages
// query. Without this gate, a warm container keeps processing
// (and potentially responding to) every accumulate-only batch, defeating
// the "store as context, don't engage" contract. Host-side countDueMessages
// gates the same way for wake-from-cold (see src/db/session-db.ts).
if (!messages.some((m) => m.trigger === 1)) {
await sleep(POLL_INTERVAL_MS);
continue;
}
const ids = messages.map((m) => m.id);
markProcessing(ids);
const routing = extractRouting(messages);
// Command handling: the host router gates filtered and unauthorized
// admin commands before they reach the container. The only command
// the runner handles directly is /clear (session reset).
const normalMessages: MessageInRow[] = [];
const commandIds: string[] = [];
for (const msg of messages) {
if ((msg.kind === 'chat' || msg.kind === 'chat-sdk') && isClearCommand(msg)) {
log('Clearing session (resetting continuation)');
continuation = undefined;
clearContinuation(config.providerName);
writeMessageOut({
id: generateId(),
kind: 'chat',
platform_id: routing.platformId,
channel_type: routing.channelType,
thread_id: routing.threadId,
content: JSON.stringify({ text: 'Session cleared.' }),
});
commandIds.push(msg.id);
continue;
}
normalMessages.push(msg);
}
if (commandIds.length > 0) {
markCompleted(commandIds);
}
if (normalMessages.length === 0) {
const remainingIds = ids.filter((id) => !commandIds.includes(id));
if (remainingIds.length > 0) markCompleted(remainingIds);
log(`All ${messages.length} message(s) were commands, skipping query`);
continue;
}
// Pre-task scripts: for any task rows with a `script`, run it before the
// provider call. Scripts returning wakeAgent=false (or erroring) gate
// their own task row only — surviving messages still go to the agent.
// Without the scheduling module, the marker block is empty, `keep`
// falls back to `normalMessages`, and no gating happens.
let keep: MessageInRow[] = normalMessages;
let skipped: string[] = [];
// MODULE-HOOK:scheduling-pre-task:start
const { applyPreTaskScripts } = await import('./scheduling/task-script.js');
const preTask = await applyPreTaskScripts(normalMessages);
keep = preTask.keep;
skipped = preTask.skipped;
if (skipped.length > 0) {
markCompleted(skipped);
log(`Pre-task script skipped ${skipped.length} task(s): ${skipped.join(', ')}`);
}
// MODULE-HOOK:scheduling-pre-task:end
if (keep.length === 0) {
log(`All ${normalMessages.length} non-command message(s) gated by script, skipping query`);
continue;
}
// Format messages: passthrough commands get raw text (only if the
// provider natively handles slash commands), others get XML.
const prompt = formatMessagesWithCommands(keep, config.provider.supportsNativeSlashCommands);
log(`Processing ${keep.length} message(s), kinds: ${[...new Set(keep.map((m) => m.kind))].join(',')}`);
const query = config.provider.query({
prompt,
continuation,
cwd: config.cwd,
systemContext: config.systemContext,
});
// Process the query while concurrently polling for new messages
const skippedSet = new Set(skipped);
const processingIds = ids.filter((id) => !commandIds.includes(id) && !skippedSet.has(id));
try {
const result = await processQuery(query, routing, processingIds, config.providerName);
if (result.continuation && result.continuation !== continuation) {
continuation = result.continuation;
setContinuation(config.providerName, continuation);
}
} catch (err) {
const errMsg = err instanceof Error ? err.message : String(err);
log(`Query error: ${errMsg}`);
// Stale/corrupt continuation recovery: ask the provider whether
// this error means the stored continuation is unusable, and clear
// it so the next attempt starts fresh.
if (continuation && config.provider.isSessionInvalid(err)) {
log(`Stale session detected (${continuation}) — clearing for next retry`);
continuation = undefined;
clearContinuation(config.providerName);
}
// Write error response so the user knows something went wrong
writeMessageOut({
id: generateId(),
kind: 'chat',
platform_id: routing.platformId,
channel_type: routing.channelType,
thread_id: routing.threadId,
content: JSON.stringify({ text: `Error: ${errMsg}` }),
});
}
// Ensure completed even if processQuery ended without a result event
// (e.g. stream closed unexpectedly).
markCompleted(processingIds);
log(`Completed ${ids.length} message(s)`);
}
}
/**
* Format messages, handling passthrough commands differently.
* When the provider handles slash commands natively (Claude Code),
* passthrough commands are sent raw (no XML wrapping) so the SDK can
* dispatch them. Otherwise they fall through to standard XML formatting.
*/
function formatMessagesWithCommands(messages: MessageInRow[], nativeSlashCommands: boolean): string {
const parts: string[] = [];
const normalBatch: MessageInRow[] = [];
for (const msg of messages) {
if (nativeSlashCommands && (msg.kind === 'chat' || msg.kind === 'chat-sdk')) {
const cmdInfo = categorizeMessage(msg);
if (cmdInfo.category === 'passthrough' || cmdInfo.category === 'admin') {
// Flush normal batch first
if (normalBatch.length > 0) {
parts.push(formatMessages(normalBatch));
normalBatch.length = 0;
}
// Pass raw command text (no XML wrapping) — SDK handles it natively
parts.push(cmdInfo.text);
continue;
}
}
normalBatch.push(msg);
}
if (normalBatch.length > 0) {
parts.push(formatMessages(normalBatch));
}
return parts.join('\n\n');
}
interface QueryResult {
continuation?: string;
}
async function processQuery(
query: AgentQuery,
routing: RoutingContext,
initialBatchIds: string[],
providerName: string,
): Promise<QueryResult> {
let queryContinuation: string | undefined;
let done = false;
// Concurrent polling: push follow-ups into the active query as they arrive.
// We do NOT force-end the stream on silence — keeping the query open avoids
// re-spawning the SDK subprocess (~few seconds) and re-loading the .jsonl
// transcript on every turn. The Anthropic prompt cache is server-side with
// a 5-min TTL keyed on prefix hash, so stream lifecycle does NOT affect
// cache lifetime — close+reopen within 5 min still gets cache hits.
// Stream liveness is decided host-side via the heartbeat file + processing
// claim age (see src/host-sweep.ts); if something is truly stuck, the host
// will kill the container and messages get reset to pending.
let pollInFlight = false;
let endedForCommand = false;
const pollHandle = setInterval(() => {
if (done || pollInFlight || endedForCommand) return;
pollInFlight = true;
void (async () => {
try {
const pending = getPendingMessages();
// Slash commands need a fresh query: /clear resets the SDK's
// resume id (fixed at sdkQuery() time); admin/passthrough commands
// (/compact, /cost, …) only dispatch when they're the first input
// of a query — pushed mid-stream they arrive as plain text and
// the SDK never runs them. End the stream and leave the rows
// pending; the outer loop handles them on next iteration via the
// canonical command path + formatMessagesWithCommands.
if (pending.some((m) => isRunnerCommand(m))) {
log('Pending slash command — ending stream so outer loop can process');
endedForCommand = true;
query.end();
return;
}
// Skip system messages (MCP tool responses).
// Thread routing is the router's concern — if a message landed in this
// session, the agent should see it. Per-thread sessions already isolate
// threads into separate containers; shared sessions intentionally merge
// everything. Filtering on thread_id here caused deadlocks when the
// initial batch and follow-ups had mismatched thread_ids (e.g. a
// host-generated welcome trigger with null thread vs a Discord DM reply).
const newMessages = pending.filter((m) => m.kind !== 'system');
if (newMessages.length === 0) return;
const newIds = newMessages.map((m) => m.id);
markProcessing(newIds);
// Run pre-task scripts on follow-ups too — without this, a task that
// arrives during an active query (e.g. a */10 monitoring cron) bypasses
// its script gate and always wakes the agent, defeating the gate.
// Mirrors the initial-batch hook above.
let keep = newMessages;
let skipped: string[] = [];
// MODULE-HOOK:scheduling-pre-task-followup:start
const { applyPreTaskScripts } = await import('./scheduling/task-script.js');
const preTask = await applyPreTaskScripts(newMessages);
keep = preTask.keep;
skipped = preTask.skipped;
if (skipped.length > 0) {
markCompleted(skipped);
log(`Pre-task script skipped ${skipped.length} follow-up task(s): ${skipped.join(', ')}`);
}
// MODULE-HOOK:scheduling-pre-task-followup:end
if (keep.length === 0) return;
// Re-check done — the outer query may have finished while the script
// was awaited. Pushing into a closed stream is wasted work; the
// claimed messages get released by the host's processing-claim sweep.
if (done) return;
const keptIds = keep.map((m) => m.id);
const prompt = formatMessages(keep);
log(`Pushing ${keep.length} follow-up message(s) into active query`);
query.push(prompt);
markCompleted(keptIds);
} catch (err) {
// Without this catch the rejection escapes the void IIFE and Node
// terminates the container on unhandled-rejection. The initial-batch
// path is wrapped by processQuery's outer try/catch; the follow-up
// path is not, so it needs its own.
const errMsg = err instanceof Error ? err.message : String(err);
log(`Follow-up poll error: ${errMsg}`);
} finally {
pollInFlight = false;
}
})();
}, ACTIVE_POLL_INTERVAL_MS);
try {
for await (const event of query.events) {
handleEvent(event, routing);
touchHeartbeat();
if (event.type === 'init') {
queryContinuation = event.continuation;
// Persist immediately so a mid-turn container crash still lets the
// next wake resume the conversation. Without this, the session id
// was only written after the full stream completed — if the
// container died between `init` and `result`, the SDK session was
// effectively orphaned and the next message started a blank
// Claude session with no prior context.
setContinuation(providerName, event.continuation);
} else if (event.type === 'result') {
// A result — with or without text — means the turn is done. Mark
// the initial batch completed now so the host sweep doesn't see
// stale 'processing' claims while the query stays open for
// follow-up pushes. The agent may have responded via MCP
// (send_message) mid-turn, or the message may not need a response
// at all — either way the turn is finished.
markCompleted(initialBatchIds);
if (event.text) {
dispatchResultText(event.text, routing);
}
} else if (event.type === 'compacted') {
// The SDK auto-compacted the conversation. After compaction the
// model often drops the learned `<message to="…">` wrapping
// discipline (the destinations are still in the system prompt,
// but the behavioral pattern is summarized away). Inject a
// reminder back into the live query so the next turn re-anchors
// on the destination model. Only do this when there's >1
// destination — single-destination groups have a fallback that
// works without wrapping. See qwibitai/nanoclaw#2325.
const destinations = getAllDestinations();
if (destinations.length > 1) {
const names = destinations.map((d) => d.name).join(', ');
query.push(
`[system] Context was just compacted. Reminder: you have ${destinations.length} destinations (${names}). ` +
`Use <message to="name"> blocks to address them. Bare text goes to the scratchpad fallback only.`,
);
}
}
}
} finally {
done = true;
clearInterval(pollHandle);
}
return { continuation: queryContinuation };
}
function handleEvent(event: ProviderEvent, _routing: RoutingContext): void {
switch (event.type) {
case 'init':
log(`Session: ${event.continuation}`);
break;
case 'result':
log(`Result: ${event.text ? event.text.slice(0, 200) : '(empty)'}`);
break;
case 'error':
log(`Error: ${event.message} (retryable: ${event.retryable}${event.classification ? `, ${event.classification}` : ''})`);
break;
case 'progress':
log(`Progress: ${event.message}`);
break;
case 'compacted':
log(`Compacted: ${event.text}`);
break;
}
}
/**
* Parse the agent's final text for <message to="name">...</message> blocks
* and dispatch each one to its resolved destination. Text outside of blocks
* (including <internal>...</internal>) is scratchpad — logged but not sent.
*
* The agent must always wrap output in <message to="name">...</message>
* blocks, even with a single destination. Bare text is scratchpad only.
*/
function dispatchResultText(text: string, routing: RoutingContext): void {
const MESSAGE_RE = /<message\s+to="([^"]+)"\s*>([\s\S]*?)<\/message>/g;
let match: RegExpExecArray | null;
let sent = 0;
let lastIndex = 0;
const scratchpadParts: string[] = [];
while ((match = MESSAGE_RE.exec(text)) !== null) {
if (match.index > lastIndex) {
scratchpadParts.push(text.slice(lastIndex, match.index));
}
const toName = match[1];
const body = match[2].trim();
lastIndex = MESSAGE_RE.lastIndex;
const dest = findByName(toName);
if (!dest) {
log(`Unknown destination in <message to="${toName}">, dropping block`);
scratchpadParts.push(`[dropped: unknown destination "${toName}"] ${body}`);
continue;
}
sendToDestination(dest, body, routing);
sent++;
}
if (lastIndex < text.length) {
scratchpadParts.push(text.slice(lastIndex));
}
const scratchpad = stripInternalTags(scratchpadParts.join(''));
if (scratchpad) {
log(`[scratchpad] ${scratchpad.slice(0, 500)}${scratchpad.length > 500 ? '…' : ''}`);
}
if (sent === 0 && text.trim()) {
log(`WARNING: agent output had no <message to="..."> blocks — nothing was sent`);
}
}
function sendToDestination(dest: DestinationEntry, body: string, routing: RoutingContext): void {
const platformId = dest.type === 'channel' ? dest.platformId! : dest.agentGroupId!;
const channelType = dest.type === 'channel' ? dest.channelType! : 'agent';
// Resolve thread_id per-destination from the most recent inbound message
// that came from this same channel+platform. In agent-shared sessions,
// different destinations have different thread contexts — using a single
// routing.threadId would stamp one channel's thread onto another.
const destRouting = resolveDestinationThread(channelType, platformId);
writeMessageOut({
id: generateId(),
in_reply_to: destRouting?.inReplyTo ?? routing.inReplyTo,
kind: 'chat',
platform_id: platformId,
channel_type: channelType,
thread_id: destRouting?.threadId ?? null,
content: JSON.stringify({ text: body }),
});
}
/**
* Find the thread_id and message id from the most recent inbound message
* matching the given channel+platform. Returns null if no match found.
*/
function resolveDestinationThread(
channelType: string,
platformId: string,
): { threadId: string | null; inReplyTo: string | null } | null {
try {
const db = getInboundDb();
const row = db
.prepare(
`SELECT thread_id, id FROM messages_in
WHERE channel_type = ? AND platform_id = ?
ORDER BY seq DESC LIMIT 1`,
)
.get(channelType, platformId) as { thread_id: string | null; id: string } | undefined;
if (row) return { threadId: row.thread_id, inReplyTo: row.id };
} catch (err) {
log(`resolveDestinationThread error: ${err instanceof Error ? err.message : String(err)}`);
}
return null;
}
function sleep(ms: number): Promise<void> {
return new Promise((resolve) => setTimeout(resolve, ms));
}