Compare commits

..

1 Commits

Author SHA1 Message Date
Dai Ha 4fa6553db5 CB-527/CB-528: bound AMQP prefetch and confirm publishes before claiming durable
CI / build (pull_request) Successful in 1m3s
CI / contract (pull_request) Successful in 1m15s
CB-527: basicQos(prefetch) on the consume channel before basicConsume, configurable
via broker.prefetch (default 32), so an undrained inbox backlog stays on the broker
instead of growing the JVM heap without limit.

CB-528: publish moves to its own confirm-mode channel with mandatory=true and a
return listener, so an unroutable or unconfirmed reply now throws instead of
vanishing silently. The confirm callback checks the per-message returned flag
(set by the return listener, which the broker always fires before the matching
confirm) so an acked-but-returned publish is still reported as a failure. The ack
path stays on its own channel/lock and never waits on a publish confirm.
2026-08-16 16:55:40 +02:00
14 changed files with 283 additions and 481 deletions
+6 -2
View File
@@ -434,10 +434,14 @@ guard:
# on a durable per-target queue (agent.<target>.inbox) and survive a restart — the broker
# redelivers anything the primary had not yet drained. Production default is LavinMQ; a stock
# RabbitMQ speaks the same AMQP 0-9-1, so it is a URI-only swap.
# uri → AMQP connection URI. No trailing slash ⇒ the default vhost "/"; an empty path ("/")
# is vhost "" and will NOT connect. Encode a named vhost as .../%2Fmyvhost.
# uri → AMQP connection URI. No trailing slash ⇒ the default vhost "/"; an empty path ("/")
# is vhost "" and will NOT connect. Encode a named vhost as .../%2Fmyvhost.
# prefetch → CB-527: consumer basicQos, capping how many unacked messages the inbox holds
# in-heap per owned target (the rest sits on the broker's durable queue instead of
# growing the JVM heap). Default 32 when omitted.
# broker:
# uri: amqp://guest:guest@127.0.0.1:5672
# prefetch: 32
# Active push-to-primary (CB-307 Stage 3). When a worker reply lands with no open bridge_send,
# the ReplyPushLoop injects a *drain nudge* (never the payload) into the primary's own herdr
@@ -348,8 +348,9 @@ public final class Bridged {
// connection, so keep the reference to close it in the ordered shutdown hook.
final ReplyInbox replyInbox;
if (cfg.broker() != null && cfg.broker().isConfigured()) {
replyInbox = AmqpReplyInbox.open(cfg.broker().uri());
log.info("reply inbox: AMQP broker (durable) at {}", cfg.broker().uri());
replyInbox = AmqpReplyInbox.open(cfg.broker().uri(), cfg.broker().prefetchOrDefault());
log.info("reply inbox: AMQP broker (durable) at {} (prefetch={})",
cfg.broker().uri(), cfg.broker().prefetchOrDefault());
} else {
replyInbox = new InMemoryReplyInbox();
log.info("reply inbox: in-memory (soft-state)");
@@ -5,6 +5,7 @@ import com.fasterxml.jackson.core.JsonParser;
import com.fasterxml.jackson.core.JsonToken;
import com.fasterxml.jackson.databind.ObjectMapper;
import com.fasterxml.jackson.dataformat.yaml.YAMLFactory;
import dev.ltms.bridged.msg.AmqpReplyInbox;
import dev.ltms.bridged.peer.MemberRole;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
@@ -532,16 +533,24 @@ public record BridgedConfig(
* stays soft-state. Production default is LavinMQ; a stock RabbitMQ speaks the same AMQP 0-9-1
* and is a URI-only swap.
*
* @param uri AMQP connection URI, e.g. {@code amqp://guest:guest@127.0.0.1:5672/}. Blank/{@code null}
* ⇒ the broker block is treated as absent (in-memory adapter).
* @param uri AMQP connection URI, e.g. {@code amqp://guest:guest@127.0.0.1:5672/}. Blank/
* {@code null} ⇒ the broker block is treated as absent (in-memory adapter).
* @param prefetch CB-527: the consumer's {@code basicQos} prefetch count, bounding how many
* unacked messages the AMQP inbox holds in-heap per owned target. {@code null}/
* non-positive ⇒ {@link AmqpReplyInbox#DEFAULT_PREFETCH}.
*/
@JsonIgnoreProperties(ignoreUnknown = true)
public record Broker(String uri) {
public record Broker(String uri, Integer prefetch) {
/** True when a usable broker URI is configured (an empty block does not enable AMQP). */
public boolean isConfigured() {
return uri != null && !uri.isBlank();
}
/** The prefetch to use, defaulting to {@link AmqpReplyInbox#DEFAULT_PREFETCH} when unset. */
public int prefetchOrDefault() {
return (prefetch != null && prefetch > 0) ? prefetch : AmqpReplyInbox.DEFAULT_PREFETCH;
}
}
/**
@@ -7,6 +7,7 @@ import com.rabbitmq.client.ConnectionFactory;
import com.rabbitmq.client.DeliverCallback;
import com.rabbitmq.client.Recoverable;
import com.rabbitmq.client.RecoveryListener;
import com.rabbitmq.client.Return;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
@@ -14,7 +15,13 @@ import java.io.IOException;
import java.nio.charset.StandardCharsets;
import java.util.LinkedHashMap;
import java.util.List;
import java.util.NavigableMap;
import java.util.concurrent.CompletableFuture;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.ConcurrentSkipListMap;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.TimeoutException;
/**
* AMQP-backed {@link ReplyInbox} (CB-307 Stage 2): genuine cross-restart durability behind the same
@@ -29,6 +36,27 @@ import java.util.concurrent.ConcurrentHashMap;
* leaves them on the broker — it redelivers on reconnect. That is the durability the in-memory
* adapter cannot give, with the port contract preserved.
*
* <p><strong>Prefetch bounds the held backlog (CB-527).</strong> The consumer channel calls
* {@code basicQos} with a configurable prefetch count ({@link #DEFAULT_PREFETCH} unless the caller
* passes another value to {@link #open(String, int)}) before starting any consumer. Without a bound,
* the broker pushes its entire queue into {@link #held} the instant a target is {@link #own owned},
* so an undrained primary grows the JVM heap without limit and any queue-level control
* ({@code x-max-length}, per-message TTL) never fires because the queue never actually holds a
* backlog. Prefetch keeps the backlog where it is visible — on the broker — until the owner drains it.
*
* <p><strong>Publishes require a confirmed, routable delivery (CB-528).</strong> {@link #publish}
* runs on a channel separate from the consume/ack channel ({@link #channel}), so a slow or blocked
* publish confirm can never hold {@link #channelLock} and stall an ack — the ack path never waits on
* a publish confirm. That publish channel is in publisher-confirm mode and every publish sets the
* {@code mandatory} flag, so an unroutable publish (queue not declared, e.g. the owner never called
* {@link #own}) is returned by the broker instead of silently dropped. The broker sends the
* <em>return</em> for an unroutable message before the <em>confirm</em> that covers it — the ack/nack
* callback checks the returned-set at confirm time rather than assuming an ack means routed — so
* "confirmed" here means "durably queued", not merely "accepted by the broker". A returned or nacked
* (or un-confirmed within the timeout) publish surfaces as an {@link IllegalStateException} on the
* caller's thread; the caller — {@link MessageService#reply} — must not report success for a
* black-holed reply.
*
* <p><strong>Ownership is explicit.</strong> {@link #own} declares the queue and starts the consumer;
* {@link #release} cancels it. {@link #publish} sends to the queue but does <em>not</em> imply ownership
* and does not attach a consumer. This split is required by CB-308 federation, where one gateway may
@@ -54,6 +82,12 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
private static final String QUEUE_PREFIX = "agent.";
private static final String QUEUE_SUFFIX = ".inbox";
/** CB-527: the prefetch used when a caller does not pass an explicit value to {@link #open(String, int)}. */
public static final int DEFAULT_PREFETCH = 32;
/** How long {@link #publish} waits for its publisher confirm before failing the call (CB-528). */
private static final long CONFIRM_TIMEOUT_MS = 10_000L;
private final Connection connection;
private final Channel channel;
/** All channel operations (publish/declare/ack/cancel) serialize on this — a Channel is not thread-safe. */
@@ -63,39 +97,81 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
/** Targets whose queue is declared and consumer is running, mapped to their broker consumer tag. */
private final ConcurrentHashMap<String, String> consumerTags = new ConcurrentHashMap<>();
/**
* CB-528: a dedicated channel for {@link #publish}, kept separate from {@link #channel} (consume
* + ack) so a publish confirm round trip never blocks under {@link #channelLock} and stalls an ack.
*/
private final Channel publishChannel;
private final Object publishChannelLock = new Object();
/** In-flight publishes awaiting their confirm, keyed by the publish channel's sequence number. */
private final ConcurrentSkipListMap<Long, Pending> pendingBySeq = new ConcurrentSkipListMap<>();
/** The same in-flight publishes, keyed by {@code msgId} — a broker {@code Return} carries no delivery tag. */
private final ConcurrentHashMap<String, Pending> pendingByMsgId = new ConcurrentHashMap<>();
/** A message pulled off the broker but not yet acked: its delivery-tag plus the port payload. */
private record Held(long deliveryTag, InboxMessage message) {}
/** Connect to {@code uri} (e.g. {@code amqp://guest:guest@127.0.0.1:5672/}) and open the inbox. */
/** A publish awaiting its confirm; {@link #returned} records whether the broker already returned it. */
private static final class Pending {
final String msgId;
final CompletableFuture<Void> confirmed = new CompletableFuture<>();
volatile boolean returned;
Pending(String msgId) {
this.msgId = msgId;
}
}
/** Connect to {@code uri} (e.g. {@code amqp://guest:guest@127.0.0.1:5672/}) with {@link #DEFAULT_PREFETCH}. */
public static AmqpReplyInbox open(String uri) {
return open(uri, DEFAULT_PREFETCH);
}
/** As {@link #open(String)}, with an explicit consumer prefetch (CB-527: caps the held backlog per target). */
public static AmqpReplyInbox open(String uri, int prefetch) {
try {
ConnectionFactory factory = new ConnectionFactory();
factory.setUri(uri);
// Self-heal transient blips; topology recovery re-declares queues and re-attaches consumers.
factory.setAutomaticRecoveryEnabled(true);
factory.setTopologyRecoveryEnabled(true);
return new AmqpReplyInbox(factory.newConnection("bridged-reply-inbox"));
return new AmqpReplyInbox(factory.newConnection("bridged-reply-inbox"), prefetch);
} catch (Exception e) {
throw new IllegalStateException("cannot connect to AMQP broker at " + uri, e);
}
}
/** Wrap an already-open connection (injection seam for the contract test). */
/** Wrap an already-open connection with {@link #DEFAULT_PREFETCH} (injection seam for the contract test). */
AmqpReplyInbox(Connection connection) {
this(connection, DEFAULT_PREFETCH);
}
/** As above, with an explicit prefetch (injection seam for the contract test). */
AmqpReplyInbox(Connection connection, int prefetch) {
this.connection = connection;
try {
this.channel = connection.createChannel();
// CB-527: bound the held backlog per owned target — must be set before any own()/basicConsume.
this.channel.basicQos(prefetch);
this.publishChannel = connection.createChannel();
this.publishChannel.confirmSelect();
this.publishChannel.addReturnListener(this::onReturn);
this.publishChannel.addConfirmListener(this::onAck, this::onNack);
} catch (IOException e) {
throw new IllegalStateException("cannot open AMQP channel", e);
}
// On automatic recovery the broker redelivers unacked messages with FRESH delivery-tags; the
// tags we were holding are now stale. Drop the held snapshot so the re-attached consumer
// repopulates it with valid tags (dedup by msgId still prevents any double-queue).
// repopulates it with valid tags (dedup by msgId still prevents any double-queue). Any publish
// confirm still in flight when the connection dropped is equally stale — its sequence number
// meant nothing on the old channel and means nothing on the recovered one, so fail it now
// rather than let it silently ride out CONFIRM_TIMEOUT_MS.
if (connection instanceof Recoverable recoverable) {
recoverable.addRecoveryListener(new RecoveryListener() {
@Override
public void handleRecovery(Recoverable recoverable) {
held.clear();
failPendingPublishesOnRecovery();
log.info("AMQP connection recovered; cleared held replies for fresh redelivery");
}
@@ -141,6 +217,12 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
}
}
/**
* Publish {@code content} and block until the broker's publisher confirm for it lands (CB-528).
* Throws {@link IllegalStateException} if the message is returned as unroutable, nacked, or not
* confirmed within {@link #CONFIRM_TIMEOUT_MS} — the caller must treat that as a failed publish,
* not a lost-and-forgotten one.
*/
@Override
public void publish(String target, String msgId, String content) {
AMQP.BasicProperties props = new AMQP.BasicProperties.Builder()
@@ -148,12 +230,35 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
.deliveryMode(2) // persistent — survives a broker restart
.contentType("text/plain")
.build();
try {
synchronized (channelLock) {
channel.basicPublish("", queueName(target), props, content.getBytes(StandardCharsets.UTF_8));
Pending pending = new Pending(msgId);
long seq;
synchronized (publishChannelLock) {
seq = publishChannel.getNextPublishSeqNo();
pendingBySeq.put(seq, pending);
pendingByMsgId.put(msgId, pending);
try {
publishChannel.basicPublish("", queueName(target), true, props,
content.getBytes(StandardCharsets.UTF_8));
} catch (IOException e) {
pendingBySeq.remove(seq, pending);
pendingByMsgId.remove(msgId, pending);
throw new IllegalStateException("cannot publish reply to " + queueName(target), e);
}
} catch (IOException e) {
throw new IllegalStateException("cannot publish reply to " + queueName(target), e);
}
try {
pending.confirmed.get(CONFIRM_TIMEOUT_MS, TimeUnit.MILLISECONDS);
} catch (ExecutionException e) {
Throwable cause = e.getCause();
throw cause instanceof RuntimeException re ? re : new IllegalStateException(cause);
} catch (TimeoutException e) {
throw new IllegalStateException("publish confirm for reply " + msgId + " to " + queueName(target)
+ " timed out after " + CONFIRM_TIMEOUT_MS + "ms — broker may be unreachable or overloaded", e);
} catch (InterruptedException e) {
Thread.currentThread().interrupt();
throw new IllegalStateException("interrupted awaiting publish confirm for " + msgId, e);
} finally {
pendingBySeq.remove(seq, pending);
pendingByMsgId.remove(msgId, pending);
}
}
@@ -222,6 +327,65 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
};
}
/** Broker return for an unroutable {@code mandatory} publish — arrives BEFORE its confirm (CB-528). */
private void onReturn(Return r) {
String msgId = r.getProperties() == null ? null : r.getProperties().getMessageId();
Pending pending = msgId == null ? null : pendingByMsgId.get(msgId);
if (pending != null) {
pending.returned = true;
} else {
log.warn("AMQP return for reply {} (routingKey={}, {} {}) with no matching in-flight publish"
+ " — already resolved by a prior confirm", msgId, r.getRoutingKey(), r.getReplyCode(),
r.getReplyText());
}
}
private void onAck(long seq, boolean multiple) {
resolveConfirm(seq, multiple, true);
}
private void onNack(long seq, boolean multiple) {
resolveConfirm(seq, multiple, false);
}
/**
* Resolve every pending publish covered by this confirm (a single seq, or — {@code multiple} —
* every seq up to and including it). Checks {@link Pending#returned} at confirm time: since the
* broker's return for an unroutable message always precedes its confirm, an ack that arrives after
* a return means "confirmed but never routed", not "durably queued".
*/
private void resolveConfirm(long seq, boolean multiple, boolean ack) {
NavigableMap<Long, Pending> covered = multiple
? pendingBySeq.headMap(seq, true)
: pendingBySeq.subMap(seq, true, seq, true);
for (var it = covered.entrySet().iterator(); it.hasNext(); ) {
Pending pending = it.next().getValue();
it.remove();
pendingByMsgId.remove(pending.msgId, pending);
if (ack && !pending.returned) {
pending.confirmed.complete(null);
} else if (ack) {
pending.confirmed.completeExceptionally(new IllegalStateException(
"reply " + pending.msgId + " was returned as unroutable (queue not declared/owned)"));
} else {
pending.confirmed.completeExceptionally(new IllegalStateException(
"broker nacked publish of reply " + pending.msgId));
}
}
}
/** Fail every publish still awaiting its confirm — their sequence numbers are stale after recovery. */
private void failPendingPublishesOnRecovery() {
for (var it = pendingBySeq.entrySet().iterator(); it.hasNext(); ) {
Pending pending = it.next().getValue();
it.remove();
pendingByMsgId.remove(pending.msgId, pending);
pending.confirmed.completeExceptionally(new IllegalStateException(
"AMQP connection recovered mid-publish; confirm status of reply " + pending.msgId
+ " is unknown"));
}
}
private static String queueName(String target) {
return QUEUE_PREFIX + target + QUEUE_SUFFIX;
}
@@ -233,6 +397,11 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
} catch (Exception e) {
log.debug("AMQP channel close: {}", e.toString());
}
try {
publishChannel.close();
} catch (Exception e) {
log.debug("AMQP publish channel close: {}", e.toString());
}
try {
connection.close();
} catch (Exception e) {
@@ -234,14 +234,7 @@ public final class BridgedApp {
List<Map<String, Object>> out = sessions.roster().stream()
.map(s -> SessionManager.rosterView(s, live.get(s.terminalId())))
.toList();
Map<String, Object> body = new LinkedHashMap<>();
body.put("workers", out);
// CB-586: operator visibility for the refs/wip snapshot store without shelling into the
// repo — how many snapshot refs exist and roughly what they cost. Present only once a
// worktree session has established the repo, so a never-snapshotted fleet reports nothing.
sessions.wipRefs().ifPresent(st -> body.put("wipRefs",
Map.of("count", st.count(), "costBytes", st.costBytes())));
ctx.status(200).json(body);
ctx.status(200).json(Map.of("workers", out));
}
/** The configured worker profiles and which one a no-argument spawn uses. */
@@ -12,12 +12,9 @@ import java.nio.file.Files;
import java.nio.file.Path;
import java.nio.file.StandardCopyOption;
import java.security.SecureRandom;
import java.util.ArrayList;
import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.Optional;
import java.util.Set;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicLong;
import java.util.stream.Collectors;
@@ -302,129 +299,6 @@ public final class GitWorktrees implements Worktrees {
return index;
}
/**
* One {@code refs/wip/<branch>} snapshot ref as read by {@link #listWipRefs}: its full ref name,
* the snapshot commit's sha, and that commit's committer time in unix millis (the age of the
* snapshot — a snapshot is written once and never rewritten, so the commit date is the ref's).
*/
private record WipRef(String refName, String sha, long committerMillis) {
String branch() {
return refName.substring("refs/wip/".length());
}
}
@Override
public WipRefStats wipRefs(String repoRoot) {
List<WipRef> refs = listWipRefs(repoRoot);
long costBytes = 0;
for (WipRef ref : refs) {
costBytes += treeSize(repoRoot, ref.sha());
}
return new WipRefStats(refs.size(), costBytes);
}
@Override
public int pruneWipRefs(String repoRoot, long minAgeMillis) {
// The rule is documented on Worktrees#pruneWipRefs: delete only a snapshot whose tree
// content is already reachable from main AND that is older than minAgeMillis. Reachability
// is the floor that keeps a worker's last copy; the age floor keeps a just-written snapshot
// from being swept while a lead may still be looking at it.
List<WipRef> refs = listWipRefs(repoRoot);
if (refs.isEmpty()) {
return 0;
}
long nowMillis = System.currentTimeMillis();
// Resolve what main carries once per sweep, not once per ref.
Set<String> mainObjects = reachableObjectsFromMain(repoRoot);
int deleted = 0;
for (WipRef ref : refs) {
long ageMillis = nowMillis - ref.committerMillis();
if (ageMillis <= minAgeMillis) {
continue; // too recent — never swept, even if it looks recoverable (CB-586)
}
String tree = exec("git", "-C", repoRoot, "rev-parse", ref.sha() + "^{tree}").trim();
if (!mainObjects.contains(tree)) {
// Last copy of the snapshot's content — the worker's work exists nowhere else.
// Never delete automatically (CB-586 criterion 2).
continue;
}
exec("git", "-C", repoRoot, "update-ref", "-d", ref.refName());
deleted++;
log.info("pruned snapshot ref refs/wip/{} commit={} (age {}h): its tree is already "
+ "reachable from main, so the work is preserved; recover from reflog via "
+ "git update-ref refs/wip/{} {}",
ref.branch(), ref.sha(), TimeUnit.MILLISECONDS.toHours(ageMillis),
ref.branch(), ref.sha());
}
return deleted;
}
/**
* Every {@code refs/wip/*} ref (see {@link WipRef}). The committer date is read as a unix
* count of seconds and converted to millis. {@code %00} (NUL) separates the fields because a
* branch name may contain spaces.
*/
private List<WipRef> listWipRefs(String repoRoot) {
String out = exec("git", "-C", repoRoot, "for-each-ref",
"--format=%(refname)%00%(objectname)%00%(committerdate:unix)", "refs/wip/");
List<WipRef> refs = new ArrayList<>();
for (String line : out.split("\\R")) {
if (line.isBlank()) {
continue;
}
String[] parts = line.split("\u0000", -1);
if (parts.length == 3 && !parts[1].isBlank()) {
refs.add(new WipRef(parts[0], parts[1], Long.parseLong(parts[2]) * 1000L));
}
}
return refs;
}
/**
* The set of object shas reachable from {@code main}, or an empty set when {@code main} cannot
* be resolved. An empty set is the safe direction: the retention sweep then concludes nothing
* is recoverable, so it deletes nothing — a repo with no {@code main} must never cause a
* worker's last copy of a snapshot to be dropped on a reachability misreading.
*/
private Set<String> reachableObjectsFromMain(String repoRoot) {
if (exitCode("git", "-C", repoRoot, "rev-parse", "--verify", "main") != 0) {
log.debug("refs/wip retention: no 'main' ref in {} — treating nothing as reachable", repoRoot);
return Set.of();
}
String out = exec("git", "-C", repoRoot, "rev-list", "--objects", "main");
Set<String> objects = new HashSet<>();
for (String line : out.split("\\R")) {
if (line.isBlank()) {
continue;
}
int sp = line.indexOf(' ');
objects.add(sp < 0 ? line : line.substring(0, sp));
}
return objects;
}
/** Approximate cost of a snapshot: the sum of every blob's size in its committed tree. */
private long treeSize(String repoRoot, String sha) {
String out = exec("git", "-C", repoRoot, "ls-tree", "-r", "-l", sha);
long total = 0;
for (String line : out.split("\\R")) {
if (line.isBlank()) {
continue;
}
// ls-tree -l row: "<mode> <type> <object> <size>\t<path>"; the size is only numeric for
// blobs (trees read "-"), so gate on the type token and take the 4th whitespace field.
String[] parts = line.split("\\s+");
if (parts.length >= 4 && "blob".equals(parts[1])) {
try {
total += Long.parseLong(parts[3]);
} catch (NumberFormatException ignored) {
// a '-' size (or any anomaly) contributes nothing to the rough figure
}
}
}
return total;
}
/** Resolve the directory that will hold per-session worktree checkouts. */
private Path resolveRoot(String repoRoot) {
if (configuredRoot != null && !configuredRoot.isBlank()) {
@@ -53,14 +53,6 @@ public final class SessionManager implements TurnListener {
private final int contextCap;
private final boolean clearAfterTurn;
private volatile MemberLifecycle memberLifecycle = MemberLifecycle.NONE;
/**
* CB-586: the repo root the fleet actually works in, remembered the first time a worktree
* session is spawned (worktrees are checkouts of it). {@code refs/wip/*} live there, and this
* single cached value is what the snapshot retention sweep and the operator-visible census run
* against. The daemon is bridged into one project at a time, so "the first worktree's repo" is
* the repo; {@code null} until any worktree is spawned, meaning nothing to sweep or measure.
*/
private volatile String fleetRepoRoot;
/** CB-520: notified with a terminalId on every acquire; no-op until wired. */
private final List<Consumer<String>> acquireListeners = new java.util.concurrent.CopyOnWriteArrayList<>();
@@ -458,11 +450,6 @@ public final class SessionManager implements TurnListener {
// The non-worktree path always used this chain; only this branch was missed.
String repoRoot = worktrees.repoRoot(
launcher.effectiveCwd(new SpawnRequest(preResolvedProfile, requestedCwd, callerCwd)));
if (fleetRepoRoot == null) {
// CB-586: remember the repo whose worktrees the fleet spawns — its refs/wip/* are the
// snapshot store the retention sweep and the operator census operate on.
fleetRepoRoot = repoRoot;
}
String branch = "worker/" + slug(wt.ticketSlug()) + "-" + nonce();
String path = null;
PeerHandle handle;
@@ -753,26 +740,6 @@ public final class SessionManager implements TurnListener {
return registry.size();
}
/**
* CB-586: the operator-visible census of {@code refs/wip/*} in the repo the fleet works in —
* how many snapshot refs exist and roughly what they cost. Empty (no repo known) until at
* least one worktree session has been spawned, exactly so a fleet that has never snapshotted
* anything surfaces nothing new, as it did before CB-586.
*/
public Optional<Worktrees.WipRefStats> wipRefs() {
String repo = fleetRepoRoot;
return repo == null ? Optional.empty() : Optional.of(worktrees.wipRefs(repo));
}
/**
* CB-586: run the snapshot retention sweep in the fleet's repo (a no-op until a worktree has
* been spawned, which establishes the repo). Returns how many {@code refs/wip/*} it deleted.
*/
public int sweepWipRefs(long minAgeMillis) {
String repo = fleetRepoRoot;
return repo == null ? 0 : worktrees.pruneWipRefs(repo, minAgeMillis);
}
/**
* The registered session owning {@code terminalId}, or {@code null} if none does.
*
@@ -14,20 +14,12 @@ public final class SessionReaper {
private static final Logger log = LoggerFactory.getLogger(SessionReaper.class);
private static final long DEFAULT_INTERVAL_MILLIS = 5000;
/** CB-586: the refs/wip age floor — never sweep a snapshot younger than 24h (the CB-586 rule). */
private static final long WIP_MIN_AGE_MILLIS = TimeUnit.HOURS.toMillis(24);
/**
* CB-586: how often the retention sweep runs. Given the 24h age floor, running it every few
* hours means a ref is dropped within hours of becoming eligible, never within minutes.
*/
private static final long WIP_SWEEP_INTERVAL_NANOS = TimeUnit.HOURS.toNanos(6);
private final SessionManager sessions;
private final long idleTtlNanos;
private final long intervalMillis;
private volatile boolean running;
private Thread thread;
private volatile long lastWipSweepNanos = Long.MIN_VALUE;
/** Construct a reaper with the default 5-second polling interval. */
public SessionReaper(SessionManager sessions, long idleTtlSeconds) {
@@ -57,32 +49,10 @@ public final class SessionReaper {
} catch (RuntimeException e) {
log.warn("session reaper iteration failed; continuing", e);
}
maybeSweepWipRefs();
sleep();
}
}
/**
* CB-586: run the refs/wip retention sweep on a slow cadence (hours, not the per-iteration
* millisecond loop). Best-effort — a failure must never take the idle-reap loop down with it.
*/
private void maybeSweepWipRefs() {
long now = System.nanoTime();
if (now - lastWipSweepNanos < WIP_SWEEP_INTERVAL_NANOS) {
return;
}
try {
int deleted = sessions.sweepWipRefs(WIP_MIN_AGE_MILLIS);
if (deleted > 0) {
log.info("refs/wip retention sweep deleted {} snapshot ref(s) older than 24h whose "
+ "content was already reachable from main", deleted);
}
} catch (RuntimeException e) {
log.warn("refs/wip retention sweep failed; continuing", e);
}
lastWipSweepNanos = now;
}
private void sleep() {
try {
Thread.sleep(intervalMillis);
@@ -54,48 +54,4 @@ public interface Worktrees {
* tolerance — a worktree that is gone holds nothing to snapshot)
*/
Optional<String> snapshot(String worktreePath, String branch, String message);
/**
* CB-586: how many {@code refs/wip/*} snapshot refs exist in {@code repoRoot} and roughly what
* they cost. This is the operator-visible surface for the snapshot growth CB-578 stage C left
* behind — counts of refs alone hide that each one pins a whole tree for {@code git gc}.
*
* @param repoRoot the repository to scan
* @return count of snapshot refs, and {@code costBytes} = the approximate total working-tree
* size of every snapshot's committed content (summed per ref, so shared objects are
* counted once per ref that carries them)
*/
WipRefStats wipRefs(String repoRoot);
/**
* CB-586: run the {@code refs/wip/*} retention sweep and return how many refs it deleted.
*
* <p>The retention rule is <em>reachability plus an age floor</em>. A snapshot ref is deleted
* only when <strong>both</strong> hold:
* <ol>
* <li>its commit's <em>tree content</em> is already reachable from {@code main} — the work
* the snapshot preserved has been recovered, so dropping the ref loses nothing; and</li>
* <li>the ref is older than {@code minAgeMillis} — a very recent snapshot is never swept
* while a lead may still be looking at it.</li>
* </ol>
*
* <p>Reachability is the safety property. A snapshot exists precisely because the work was not
* committed anywhere else, so a snapshot whose content is <em>not</em> reachable from
* {@code main} is the <strong>last copy</strong> of a worker's work and must never be deleted
* automatically — that is the failure CB-576 and CB-578 stage C were built to stop. Age alone
* must never drive a deletion, because age-based sweeping is exactly how the last copy gets
* destroyed. (Both numbers and the rule are CB-586's decision; this method only implements it.)
*
* <p>Every deletion logs the ref name and the commit sha, so an operator who finds they lost
* the wrong thing can still recover it from git's reflog.
*
* @param repoRoot the repository whose {@code refs/wip/*} to sweep
* @param minAgeMillis the age floor; a ref younger than this is never touched
* @return the number of snapshot refs deleted
*/
int pruneWipRefs(String repoRoot, long minAgeMillis);
/** CB-586: the operator-visible census of {@code refs/wip/*} in one repository. */
record WipRefStats(int count, long costBytes) {
}
}
@@ -16,6 +16,7 @@ import java.util.List;
import java.util.concurrent.TimeUnit;
import static org.junit.jupiter.api.Assertions.assertEquals;
import static org.junit.jupiter.api.Assertions.assertThrows;
import static org.junit.jupiter.api.Assertions.assertTrue;
/**
@@ -166,6 +167,88 @@ class AmqpReplyInboxContractTest {
}
}
@Test
void prefetchBoundsTheHeldBacklog() throws Exception {
String target = "worker-prefetch-" + System.nanoTime();
int prefetch = 4;
int published = 10;
try (AmqpReplyInbox inbox = new AmqpReplyInbox(newConnection(), prefetch);
Connection inspect = newConnection()) {
inbox.own(target);
for (int i = 0; i < published; i++) {
inbox.publish(target, "m" + i, "payload " + i);
}
awaitHeldAtLeast(inbox, target, prefetch);
long depth;
try (Channel ch = inspect.createChannel()) {
depth = ch.queueDeclarePassive(queueName(target)).getMessageCount();
}
assertTrue(depth >= published - prefetch,
"broker should still hold at least " + (published - prefetch)
+ " undelivered messages behind a prefetch of " + prefetch + ", saw " + depth);
drainUntilEmpty(inbox, target, inspect);
}
}
@Test
void unroutablePublishReportsFailureNotSilentSuccess() throws Exception {
String target = "worker-unroutable-" + System.nanoTime();
try (AmqpReplyInbox inbox = AmqpReplyInbox.open(uri())) {
// Deliberately never own(target): the queue is never declared, so the default-exchange
// route to agent.<target>.inbox does not exist and the broker must return the publish.
IllegalStateException ex = assertThrows(IllegalStateException.class,
() -> inbox.publish(target, "m1", "nobody home"));
assertTrue(ex.getMessage() != null && ex.getMessage().toLowerCase().contains("unroutable"),
"expected an unroutable-publish failure, got: " + ex.getMessage());
}
}
@Test
void confirmedPublishDeliversNormally() throws Exception {
String target = "worker-confirm-" + System.nanoTime();
try (AmqpReplyInbox inbox = AmqpReplyInbox.open(uri())) {
inbox.own(target);
inbox.publish(target, "m1", "confirmed delivery"); // must return normally: routed and confirmed
List<ReplyInbox.InboxMessage> got = awaitPeek(inbox, target);
assertEquals(1, got.size());
assertEquals("confirmed delivery", got.getFirst().content());
inbox.ack(target, "m1");
}
}
/** Poll peek until at least {@code n} replies for {@code target} are held, or ~10s elapse. */
@SuppressWarnings("BusyWait")
private static void awaitHeldAtLeast(AmqpReplyInbox inbox, String target, int n) throws InterruptedException {
long deadline = System.nanoTime() + TimeUnit.SECONDS.toNanos(10);
while (inbox.peek(target).size() < n && System.nanoTime() < deadline) {
Thread.sleep(50);
}
}
/**
* Repeatedly ack whatever is currently held (freeing prefetch slots for the next delivery) until
* both the local snapshot and the broker's own queue depth are empty, or ~10s elapse.
*/
@SuppressWarnings("BusyWait")
private static void drainUntilEmpty(AmqpReplyInbox inbox, String target, Connection inspect) throws Exception {
long deadline = System.nanoTime() + TimeUnit.SECONDS.toNanos(10);
while (System.nanoTime() < deadline) {
for (ReplyInbox.InboxMessage msg : inbox.peek(target)) {
inbox.ack(target, msg.msgId());
}
try (Channel ch = inspect.createChannel()) {
if (ch.queueDeclarePassive(queueName(target)).getMessageCount() == 0 && inbox.peek(target).isEmpty()) {
return;
}
}
Thread.sleep(50);
}
throw new AssertionError("did not drain " + target + " to empty within the deadline");
}
/** Poll peek (broker delivery is async) until a reply for {@code target} appears or ~10s elapse. */
@SuppressWarnings("BusyWait") // deliberate poll for async broker delivery, bounded by the deadline
private static List<ReplyInbox.InboxMessage> awaitPeek(AmqpReplyInbox inbox, String target)
@@ -27,15 +27,11 @@ public final class FakeWorktrees implements Worktrees {
public record SnapshotCall(String worktreePath, String branch, String message) {
}
public record PruneCall(String repoRoot, long minAgeMillis) {
}
private final List<AddCall> addCalls = new CopyOnWriteArrayList<>();
private final List<RemoveCall> removeCalls = new CopyOnWriteArrayList<>();
private final List<OverlayCall> overlayCalls = new CopyOnWriteArrayList<>();
private final List<RepoRootCall> repoRootCalls = new CopyOnWriteArrayList<>();
private final List<SnapshotCall> snapshotCalls = new CopyOnWriteArrayList<>();
private final List<PruneCall> pruneCalls = new CopyOnWriteArrayList<>();
private final Set<String> existingPaths = ConcurrentHashMap.newKeySet();
private final Set<String> trackedPaths = ConcurrentHashMap.newKeySet();
private final AtomicLong snapshotSeq = new AtomicLong();
@@ -44,8 +40,6 @@ public final class FakeWorktrees implements Worktrees {
private volatile boolean dirty = false;
private volatile String repoRoot = "/repo";
private volatile String prefix = "/worktrees";
private volatile WipRefStats wipRefs = new WipRefStats(0, 0L);
private volatile int pruneResult = 0;
public FakeWorktrees withRepoRoot(String root) {
this.repoRoot = root;
@@ -88,18 +82,6 @@ public final class FakeWorktrees implements Worktrees {
return this;
}
/** Configure the value returned by {@link #wipRefs}. */
public FakeWorktrees withWipRefs(WipRefStats stats) {
this.wipRefs = stats;
return this;
}
/** Configure the value returned by {@link #pruneWipRefs}. */
public FakeWorktrees withPruneResult(int deleted) {
this.pruneResult = deleted;
return this;
}
@Override
public String add(String repoRoot, String branch, String baseRef) {
addCalls.add(new AddCall(repoRoot, branch, baseRef));
@@ -153,17 +135,6 @@ public final class FakeWorktrees implements Worktrees {
return Optional.of("wip" + snapshotSeq.incrementAndGet());
}
@Override
public WipRefStats wipRefs(String repoRoot) {
return wipRefs;
}
@Override
public int pruneWipRefs(String repoRoot, long minAgeMillis) {
pruneCalls.add(new PruneCall(repoRoot, minAgeMillis));
return pruneResult;
}
public List<AddCall> addCalls() {
return List.copyOf(addCalls);
}
@@ -196,10 +167,6 @@ public final class FakeWorktrees implements Worktrees {
return List.copyOf(snapshotCalls);
}
public List<PruneCall> pruneCalls() {
return List.copyOf(pruneCalls);
}
public SnapshotCall lastSnapshot() {
return snapshotCalls.isEmpty() ? null : snapshotCalls.getLast();
}
@@ -3,7 +3,6 @@ package dev.ltms.bridged.session;
import org.junit.jupiter.api.Test;
import org.junit.jupiter.api.io.TempDir;
import java.nio.charset.StandardCharsets;
import java.nio.file.Files;
import java.nio.file.Path;
import java.util.HashSet;
@@ -139,53 +138,6 @@ class GitWorktreesTest {
return out;
}
/** Write {@code content} as a blob into the object database; returns its sha. */
private static String blobOf(Path cwd, String content) throws Exception {
Process p = new ProcessBuilder("git", "-C", cwd.toString(), "hash-object", "-w", "--stdin")
.redirectErrorStream(true).start();
p.getOutputStream().write(content.getBytes(StandardCharsets.UTF_8));
p.getOutputStream().close();
String out = new String(p.getInputStream().readAllBytes()).trim();
assertTrue(p.waitFor(30, TimeUnit.SECONDS), "git hash-object timed out");
assertEquals(0, p.exitValue(), "git hash-object failed:\n" + out);
return out;
}
/** Build a single-file tree object from {@code blob}; returns the tree's sha. */
private static String treeOf(Path cwd, String path, String blob) throws Exception {
Process p = new ProcessBuilder("git", "-C", cwd.toString(), "mktree")
.redirectErrorStream(true).start();
p.getOutputStream().write(("100644 blob " + blob + "\t" + path + "\n").getBytes(StandardCharsets.UTF_8));
p.getOutputStream().close();
String out = new String(p.getInputStream().readAllBytes()).trim();
assertTrue(p.waitFor(30, TimeUnit.SECONDS), "git mktree timed out");
assertEquals(0, p.exitValue(), "git mktree failed:\n" + out);
return out;
}
/** {@code git commit-tree} rooted at {@code tree} with a chosen committer date; returns the sha. */
private static String commitTree(Path cwd, String tree, String parent, String committerDate,
String message) throws Exception {
ProcessBuilder pb = new ProcessBuilder("git", "-C", cwd.toString(), "commit-tree",
tree, "-p", parent, "-m", message);
pb.environment().put("GIT_COMMITTER_DATE", committerDate);
Process p = pb.redirectErrorStream(true).start();
String out = new String(p.getInputStream().readAllBytes()).trim();
assertTrue(p.waitFor(30, TimeUnit.SECONDS), "git commit-tree timed out");
assertEquals(0, p.exitValue(), "git commit-tree failed:\n" + out);
return out;
}
/** {@code git update-ref <ref> <sha>} — create the snapshot ref directly. */
private static void updateRef(Path cwd, String ref, String sha) throws Exception {
git(cwd, "update-ref", ref, sha);
}
/** True when {@code ref} exists in the repo (for-each-ref on a missing ref is empty, not an error). */
private static boolean refExists(Path cwd, String ref) throws Exception {
return !forEachRef(cwd, ref).trim().isEmpty();
}
/**
* The heart of CB-525: a provisioned worktree must not inherit the primary's MCP servers. Without
* the isolation step the checked-out {@code .mcp.json} carries them in, and a worker navigating
@@ -522,104 +474,4 @@ class GitWorktreesTest {
assertThrows(WorktreeException.class, () -> gitWorktrees.snapshot(wt, branch, "test snapshot"),
"an unresolvable real index must fail loudly, not silently snapshot from an empty index");
}
/**
* CB-586, criterion 2. A snapshot whose content is NOT reachable from {@code main} is the last
* copy of a worker's work, and must never be deleted automatically — even when it is old and
* even when the caller passes a zero age floor. Uses the real snapshot path on a dirty worktree,
* so the unreachable tree is exactly the shape CB-576/CB-578 stage C exist to protect.
*/
@Test
void anUnreachableSnapshotIsNeverPruned(@TempDir Path tmp) throws Exception {
Path repo = initRepo(tmp.resolve("repo"));
GitWorktrees gitWorktrees = new GitWorktrees(tmp.resolve("wts").toString());
String branch = "cb-586-unreachable";
String wt = gitWorktrees.add(repo.toString(), branch, "HEAD");
Files.writeString(Path.of(wt).resolve("worker-draft.txt"), "work that exists nowhere else\n");
Optional<String> ref = gitWorktrees.snapshot(wt, branch, "snapshot with unreachable content");
assertTrue(ref.isPresent());
// Age floor 0 makes age a non-issue: only reachability can save it — and it must.
assertEquals(0, gitWorktrees.pruneWipRefs(repo.toString(), 0),
"the unreachable snapshot is the last copy and must not be pruned");
assertTrue(refExists(repo, "refs/wip/" + branch),
"an unreachable snapshot must survive the sweep");
}
/**
* CB-586, criterion 1 (the reachable half). A snapshot whose tree content IS already reachable
* from {@code main} and which is older than the age floor is pure duplication — the work is
* recovered — so it must be pruned.
*/
@Test
void aReachableSnapshotOlderThanTheFloorIsPruned(@TempDir Path tmp) throws Exception {
Path repo = initRepo(tmp.resolve("repo"));
GitWorktrees gitWorktrees = new GitWorktrees(tmp.resolve("wts").toString());
// A snapshot whose tree is exactly main's current tree: fully reachable from main.
String mainTree = revParse(repo, "main^{tree}");
String old = commitTree(repo, mainTree, revParse(repo, "HEAD"), "2020-01-01T00:00:00", "snapshot");
updateRef(repo, "refs/wip/recovered", old);
assertEquals(1, gitWorktrees.pruneWipRefs(repo.toString(), TimeUnit.HOURS.toMillis(24)),
"an old, main-reachable snapshot must be pruned");
assertFalse(refExists(repo, "refs/wip/recovered"),
"the reachable snapshot's ref must be gone after the sweep");
}
/**
* CB-586, the age floor. A snapshot whose content IS reachable from {@code main} but which is
* younger than the age floor must not be swept — a lead may still be looking at it.
*/
@Test
void aReachableButRecentSnapshotIsNotPruned(@TempDir Path tmp) throws Exception {
Path repo = initRepo(tmp.resolve("repo"));
GitWorktrees gitWorktrees = new GitWorktrees(tmp.resolve("wts").toString());
// Reachable from main, but committed "now" — a fresh snapshot. The 24h floor must protect it.
String mainTree = revParse(repo, "main^{tree}");
String fresh = commitTree(repo, mainTree, revParse(repo, "HEAD"),
"2038-01-01T00:00:00", "snapshot just taken");
updateRef(repo, "refs/wip/fresh", fresh);
assertEquals(0, gitWorktrees.pruneWipRefs(repo.toString(), TimeUnit.HOURS.toMillis(24)),
"a recent snapshot must be kept even when reachable");
assertTrue(refExists(repo, "refs/wip/fresh"),
"the recent reachable snapshot must survive the sweep");
}
/**
* CB-586, criterion 5. A fleet that has never snapshotted anything has no {@code refs/wip/*},
* so a sweep is a no-op and the census reports none — identical to before CB-586 existed.
*/
@Test
void aFleetWithNoSnapshotsPrunesNothingAndReportsNothing(@TempDir Path tmp) throws Exception {
Path repo = initRepo(tmp.resolve("repo"));
GitWorktrees gitWorktrees = new GitWorktrees(tmp.resolve("wts").toString());
assertEquals(0, gitWorktrees.pruneWipRefs(repo.toString(), 0),
"no snapshot refs means nothing to prune");
Worktrees.WipRefStats stats = gitWorktrees.wipRefs(repo.toString());
assertEquals(0, stats.count(), "a never-snapshotted fleet has zero refs/wip refs");
assertEquals(0L, stats.costBytes(), "a never-snapshotted fleet costs zero bytes");
}
/** CB-586, criterion 4: the census reports how many refs exist and roughly what they cost. */
@Test
void wipRefsReportsCountAndCost(@TempDir Path tmp) throws Exception {
Path repo = initRepo(tmp.resolve("repo"));
GitWorktrees gitWorktrees = new GitWorktrees(tmp.resolve("wts").toString());
String blob = blobOf(repo, "a recoverable snapshot's worth of content");
String tree = treeOf(repo, "snapshot.txt", blob);
updateRef(repo, "refs/wip/one", commitTree(repo, tree, revParse(repo, "HEAD"),
"2020-01-01T00:00:00", "snapshot"));
updateRef(repo, "refs/wip/two", commitTree(repo, tree, revParse(repo, "HEAD"),
"2020-01-02T00:00:00", "snapshot"));
Worktrees.WipRefStats stats = gitWorktrees.wipRefs(repo.toString());
assertEquals(2, stats.count(), "two snapshot refs are reported");
assertTrue(stats.costBytes() > 0, "the cost of the snapshots is a positive byte count");
}
}
@@ -138,16 +138,6 @@ class SessionManagerTest {
return java.util.Optional.of("wip" + snapshotSeq.incrementAndGet());
}
@Override
public WipRefStats wipRefs(String repoRoot) {
return new WipRefStats(0, 0L);
}
@Override
public int pruneWipRefs(String repoRoot, long minAgeMillis) {
return 0;
}
List<String> removeCalls() {
return List.copyOf(removeCalls);
}
@@ -444,37 +444,4 @@ class WorktreeSessionManagerTest {
"a failed snapshot leaves no ref to report");
}
/**
* CB-586, criteria 4 and 1. Once a worktree session is spawned the repo is known, so the
* operator census and the retention sweep delegate to that repo's {@code refs/wip/*}.
*/
@Test
void wipRefsAndSweepDelegateToTheFleetRepoOnceKnown() {
FakeHerdr herdr = new FakeHerdr();
FakeWorktrees worktrees = new FakeWorktrees().withRepoRoot("/repo").withPrefix("/wt")
.withWipRefs(new Worktrees.WipRefStats(3, 42L)).withPruneResult(2);
SessionManager sessions = new SessionManager(workerService(herdr), worktrees);
assertTrue(sessions.wipRefs().isEmpty(),
"no worktree spawned yet means no repo is known and nothing to report");
assertEquals(0, sessions.sweepWipRefs(TimeUnit.HOURS.toMillis(24)),
"no worktree spawned yet means the sweep is a no-op");
sessions.acquire("ltms-local", null, "/caller/proj", null,
new WorktreeRequest("cb-586", null));
Worktrees.WipRefStats stats = sessions.wipRefs().orElseThrow();
assertEquals(3, stats.count(), "the census comes from the fleet repo");
assertEquals(42L, stats.costBytes(), "the cost comes from the fleet repo");
assertEquals(2, sessions.sweepWipRefs(TimeUnit.HOURS.toMillis(24)),
"the sweep runs against the fleet repo");
List<FakeWorktrees.PruneCall> prunes = worktrees.pruneCalls();
assertEquals(1, prunes.size(), "the no-op short-circuits before reaching the seam, so only "
+ "the repo-known sweep issues a call");
assertEquals("/repo", prunes.getFirst().repoRoot(), "the sweep targets the fleet repo");
assertEquals(TimeUnit.HOURS.toMillis(24), prunes.getFirst().minAgeMillis(),
"the caller's age floor is passed through");
}
}