feat: persist and dispatch client command input

This commit is contained in:
2026-09-06 10:40:23 +00:00
parent 8155e5f81f
commit ad564de23f
16 changed files with 836 additions and 22 deletions
+66 -2
View File
@@ -141,13 +141,14 @@ func (server *AgentServer) serveConnection(parent context.Context, connection *w
capacity := &CapacityShadow{}
var capacityMu sync.Mutex
reservations := make(map[string]DispatchLane)
stdinSent := make(map[string]struct{})
reconciled := make(chan struct{})
var reconcileOnce sync.Once
if !capacity.UpdateAdvertised(0, 0, hello.GetMaxRunningCommands(), hello.GetMaxQueuedCommands()) {
server.close(connection, websocket.StatusPolicyViolation, "invalid initial client capacity")
return
}
go server.dispatchLoop(sessionContext, queue, handle.DispatchWake(), reconciled, &capacityMu, capacity, reservations, hello.GetClientId(), hello.GetPlatform(), encodeSessionID(sessionID), registration.Generation, cancel)
go server.dispatchLoop(sessionContext, queue, handle.DispatchWake(), reconciled, &capacityMu, capacity, reservations, stdinSent, hello.GetClientId(), hello.GetPlatform(), encodeSessionID(sessionID), registration.Generation, cancel)
encodedSessionID := encodeSessionID(sessionID)
welcome, err := proto.Marshal(&rvboxv1.AgentEnvelope{
SessionId: encodedSessionID, SessionGeneration: registration.Generation,
@@ -265,6 +266,13 @@ func (server *AgentServer) serveConnection(parent context.Context, connection *w
server.close(connection, websocket.StatusInternalError, "could not acknowledge command event")
return
}
if stdinAck := event.GetStdinAck(); stdinAck != nil {
issue, _ := domain.ParseUUIDv7(event.GetIssueUuid())
capacityMu.Lock()
delete(stdinSent, stdinIntentKey(issue, stdinAck.GetWriteSeq()))
capacityMu.Unlock()
handle.SignalDispatch()
}
}
}
}
@@ -366,10 +374,58 @@ func (server *AgentServer) enqueueNextDispatch(ctx context.Context, queue *Write
return candidate.IssueUUID, true, nil
}
// enqueueNextStdin exposes one durable input intent on the essential control
// lane. Session-local sent tracking suppresses duplicate frames while a live
// connection remains usable; reconnecting naturally replays unacknowledged
// writes from storage.
func (server *AgentServer) enqueueNextStdin(ctx context.Context, queue *WriterQueue, clientID, sessionID string, generation uint64, sent map[string]struct{}, sentMu *sync.Mutex) (bool, error) {
intents, err := server.Store.PendingStdin(ctx, clientID)
if err != nil {
return false, err
}
for _, intent := range intents {
key := stdinIntentKey(intent.IssueUUID, intent.WriteSeq)
sentMu.Lock()
_, alreadySent := sent[key]
if !alreadySent {
sent[key] = struct{}{}
}
sentMu.Unlock()
if alreadySent {
continue
}
envelope := &rvboxv1.AgentEnvelope{SessionId: sessionID, SessionGeneration: generation}
if intent.Close {
envelope.Payload = &rvboxv1.AgentEnvelope_CloseStdin{CloseStdin: &rvboxv1.CloseStdin{IssueUuid: intent.IssueUUID.String(), WriteSeq: intent.WriteSeq}}
} else {
envelope.Payload = &rvboxv1.AgentEnvelope_StdinWrite{StdinWrite: &rvboxv1.StdinWrite{IssueUuid: intent.IssueUUID.String(), WriteSeq: intent.WriteSeq, Data: append([]byte(nil), intent.Data...), AppendNewline: intent.AppendNewline}}
}
encoded, marshalErr := proto.Marshal(envelope)
if marshalErr != nil {
sentMu.Lock()
delete(sent, key)
sentMu.Unlock()
return false, marshalErr
}
if enqueueErr := queue.EnqueueControl(Frame{Kind: FrameControl, Payload: encoded}); enqueueErr != nil {
sentMu.Lock()
delete(sent, key)
sentMu.Unlock()
return false, enqueueErr
}
return true, nil
}
return false, nil
}
func stdinIntentKey(issue domain.UUID, writeSeq uint64) string {
return issue.String() + ":" + fmt.Sprint(writeSeq)
}
// dispatchLoop is the per-session serialized dispatcher. It waits for a
// complete reconciliation result before consuming queued work, then coalesces
// wakeups from local control RPCs, capacity advertisements, and acceptances.
func (server *AgentServer) dispatchLoop(ctx context.Context, queue *WriterQueue, wake <-chan struct{}, reconciled <-chan struct{}, capacityMu *sync.Mutex, capacity *CapacityShadow, reservations map[string]DispatchLane, clientID string, platform rvboxv1.Platform, sessionID string, generation uint64, cancel context.CancelFunc) {
func (server *AgentServer) dispatchLoop(ctx context.Context, queue *WriterQueue, wake <-chan struct{}, reconciled <-chan struct{}, capacityMu *sync.Mutex, capacity *CapacityShadow, reservations map[string]DispatchLane, stdinSent map[string]struct{}, clientID string, platform rvboxv1.Platform, sessionID string, generation uint64, cancel context.CancelFunc) {
select {
case <-reconciled:
case <-ctx.Done():
@@ -377,6 +433,14 @@ func (server *AgentServer) dispatchLoop(ctx context.Context, queue *WriterQueue,
}
for {
for {
stdinQueued, stdinErr := server.enqueueNextStdin(ctx, queue, clientID, sessionID, generation, stdinSent, capacityMu)
if stdinErr != nil {
server.closeForDispatchFailure(cancel)
return
}
if stdinQueued {
continue
}
capacityMu.Lock()
lane := capacity.Reserve()
capacityMu.Unlock()
+13
View File
@@ -12,6 +12,7 @@ import (
rvboxv1 "github.com/rvbox/rvbox/gen/go/rvbox/v1"
"github.com/rvbox/rvbox/internal/domain"
"google.golang.org/protobuf/proto"
)
const (
@@ -235,6 +236,18 @@ retained_compressed_bytes = retained_compressed_bytes + ?, output_charged_bytes
}
}
}
if err == nil && event.EventType == 7 {
// Stdin acknowledgements are command events, but their durable delivery
// cursor lives in stdin_writes. Mark the cumulative prefix in the same
// transaction as the event so a crash cannot make an acknowledged write
// replay forever or release it before its event is durable.
var wire rvboxv1.CommandEvent
if unmarshalErr := proto.Unmarshal(event.Payload, &wire); unmarshalErr != nil || wire.GetStdinAck() == nil || wire.GetStdinAck().GetWriteSeq() == 0 {
err = ErrInvalidSegmentRecord
} else {
_, err = tx.ExecContext(ctx, `UPDATE stdin_writes SET acknowledged = 1 WHERE issue_uuid = ? AND write_seq <= ?`, event.IssueUUID[:], wire.GetStdinAck().GetWriteSeq())
}
}
if err == nil {
var update sql.Result
update, err = tx.ExecContext(ctx, `UPDATE clients SET charged_bytes = ? WHERE client_id = ? AND charged_bytes = ?`, reservation.ClientTotalCharged, clientID, clientCharged)
+80
View File
@@ -1,6 +1,7 @@
package store
import (
"bytes"
"context"
"crypto/sha256"
"database/sql"
@@ -9,6 +10,7 @@ import (
"math"
"time"
"github.com/klauspost/compress/zstd"
"github.com/rvbox/rvbox/internal/domain"
)
@@ -33,6 +35,84 @@ type StdinWriteResult struct {
Duplicate bool
}
// StdinIntent is a server-owned input record ready for delivery to a live
// client session. It is returned in command/write order and contains a copy of
// the decompressed payload; callers must not mutate storage-owned buffers.
type StdinIntent struct {
IssueUUID domain.UUID
WriteSeq uint64
Data []byte
AppendNewline bool
Close bool
}
// PendingStdin returns unacknowledged input intents for one client. Delivery is
// intentionally tracked by the session dispatcher, not by this durable query:
// a lost connection simply causes the next session to replay the same write
// sequence, which the client acknowledges idempotently.
func (store *Store) PendingStdin(ctx context.Context, clientID string) ([]StdinIntent, error) {
if clientID == "" {
return nil, errors.New("client ID is required")
}
database, err := store.openDatabase()
if err != nil {
return nil, err
}
rows, err := database.QueryContext(ctx, `SELECT w.issue_uuid, w.write_seq, w.payload, w.raw_bytes, w.stored_bytes, w.compression, w.sha256, w.append_newline, w.close_intent
FROM stdin_writes w JOIN commands c ON c.issue_uuid = w.issue_uuid
WHERE c.client_id = ? AND c.lifecycle BETWEEN 1 AND 4 AND w.acknowledged = 0
ORDER BY c.issue_time, c.issue_uuid, w.write_seq`, clientID)
if err != nil {
return nil, err
}
defer rows.Close()
var result []StdinIntent
for rows.Next() {
var encodedIssue, stored, digest []byte
var writeSeq, rawBytes, storedBytes uint64
var compression uint32
var appendNewline, closeIntent bool
if err := rows.Scan(&encodedIssue, &writeSeq, &stored, &rawBytes, &storedBytes, &compression, &digest, &appendNewline, &closeIntent); err != nil {
return nil, err
}
if len(encodedIssue) != len(domain.UUID{}) || writeSeq == 0 || compression != 2 || storedBytes != uint64(len(stored)) || len(digest) != sha256.Size || rawBytes > maxStoredExecutionSpecBytes {
return nil, ErrInvalidSegmentRecord
}
var data []byte
data, err = decompressStdinPayload(stored, rawBytes)
if err != nil || (!closeIntent && rawBytes == 0) || (closeIntent && rawBytes != 0) {
return nil, ErrInvalidSegmentRecord
}
computed := sha256.Sum256(data)
if string(computed[:]) != string(digest) {
return nil, ErrInvalidSegmentRecord
}
var issue domain.UUID
copy(issue[:], encodedIssue)
result = append(result, StdinIntent{IssueUUID: issue, WriteSeq: writeSeq, Data: data, AppendNewline: appendNewline, Close: closeIntent})
}
if err := rows.Err(); err != nil {
return nil, err
}
return result, nil
}
func decompressStdinPayload(stored []byte, expected uint64) ([]byte, error) {
if expected > maxStoredExecutionSpecBytes {
return nil, ErrInvalidSegmentRecord
}
decoder, err := zstd.NewReader(bytes.NewReader(stored), zstd.WithDecoderConcurrency(1), zstd.WithDecoderMaxMemory(maxStoredExecutionSpecBytes+1))
if err != nil {
return nil, ErrInvalidSegmentRecord
}
defer decoder.Close()
data, err := decoder.DecodeAll(stored, nil)
if err != nil || uint64(len(data)) != expected {
return nil, ErrInvalidSegmentRecord
}
return data, nil
}
// AppendStdin durably records one ordered stdin intent. It does not claim
// delivery to a process; an agent acknowledgement is a later protocol event.
// Reusing RequestUUID with the same immutable hash returns the original write