Files
byre/internal/agents/manager.go
T

875 lines
20 KiB
Go

package agents
import (
"context"
"crypto/rand"
"encoding/json"
"errors"
"fmt"
"log/slog"
"net/url"
"os"
"path/filepath"
"strings"
"sync"
"time"
"g.fc.run/misc/byre/internal/api"
"g.fc.run/misc/byre/internal/claude"
"g.fc.run/misc/byre/internal/git"
"g.fc.run/misc/byre/internal/mcp"
"g.fc.run/misc/byre/internal/store"
)
var (
ErrConflict = errors.New("conflict")
ErrInvalid = errors.New("invalid")
ErrNoGitIdentity = errors.New("set your git name and email (PUT /api/me) before starting agents")
)
const agentActor = "agent"
const restartNotice = "byre restarted while you were mid-turn, so your last turn was cut off. Re-check the working tree (git status and the files you were editing) before continuing where you left off."
type Manager struct {
store *store.Store
worktreeRoot string
leadRoot string
codexHome string
mu sync.Mutex
live map[string]*live
closing bool
rateLimits map[api.AgentSessionVendor]map[string]any
leadAnswered map[string]bool
paneFocused map[string]bool
deliverMu sync.Mutex
leadMu sync.Mutex
signals chan api.Event
snapMu sync.Mutex
snaps map[string]*snapState
mergeMu sync.Mutex
mergeHolders map[string]string
bad *badLines
watchStop chan struct{}
}
type live struct {
sess agentProcess
id string
dir string
model string
gitName string
gitEmail string
vendor api.AgentSessionVendor
taskID string
projectID string
owner string
stopping bool
interrupted bool
asked bool
askedBlock bool
bounced bool
compacting bool
wantCompact bool
rateLimited map[string]any
lastText string
contextModel string
contextUsed int
contextWindow int
}
func New(st *store.Store, worktreeRoot, leadRoot, codexHome string) *Manager {
m := &Manager{
store: st,
worktreeRoot: worktreeRoot,
leadRoot: leadRoot,
codexHome: codexHome,
live: map[string]*live{},
rateLimits: map[api.AgentSessionVendor]map[string]any{},
leadAnswered: map[string]bool{},
paneFocused: map[string]bool{},
signals: make(chan api.Event, 1024),
snaps: map[string]*snapState{},
mergeHolders: map[string]string{},
bad: newBadLines(),
watchStop: make(chan struct{}),
}
go m.signalLoop()
return m
}
func (m *Manager) taskFor(ctx context.Context, sess api.AgentSession) (api.Task, error) {
if sess.Owner != "" {
return api.Task{}, nil
}
return m.store.GetTask(ctx, sess.TaskId)
}
func (m *Manager) Create(ctx context.Context, task api.Task, creator api.User, prompt, model string, images ...string) (api.AgentSession, error) {
return m.create(ctx, task, creator, prompt, model, creator.Login, images)
}
func (m *Manager) create(ctx context.Context, task api.Task, creator api.User, prompt, model, actor string, images []string) (api.AgentSession, error) {
if creator.GitName == "" || creator.GitEmail == "" {
return api.AgentSession{}, ErrNoGitIdentity
}
if task.Status != api.TaskStatusActive {
return api.AgentSession{}, fmt.Errorf("%w: task %s is %s", ErrConflict, task.Id, task.Status)
}
existing, err := m.store.ListSessions(ctx, task.Id)
if err != nil {
return api.AgentSession{}, err
}
for _, s := range existing {
if s.State != api.AgentSessionStateCompleted {
return api.AgentSession{}, fmt.Errorf("%w: task %s already has session %s", ErrConflict, task.Id, s.Id)
}
}
id := newUUID()
vendor := vendorOf(model)
sess := api.AgentSession{
Id: id,
TaskId: task.Id,
Vendor: vendor,
Model: model,
VendorSessionId: initialVendorSessionID(vendor, id),
State: api.AgentSessionStateIdle,
Role: api.Implementer,
CreatedBy: creator.Login,
CreatedAt: time.Now().UTC(),
}
err = m.store.CreateSession(ctx, sess, task.ProjectId)
if err != nil {
return api.AgentSession{}, err
}
return m.turn(ctx, sess, task, actor, prompt, images)
}
func (m *Manager) Send(ctx context.Context, sessionID, actor, text, model string, images ...string) (api.AgentSession, error) {
sess, err := m.store.GetSession(ctx, sessionID)
if err != nil {
return api.AgentSession{}, err
}
if sess.State != api.AgentSessionStateIdle {
return api.AgentSession{}, fmt.Errorf("%w: session %s is %s", ErrConflict, sess.Id, sess.State)
}
if model != "" && model != sess.Model {
err = checkModel(sess, model)
if err != nil {
return api.AgentSession{}, err
}
err = m.store.SetSessionModel(ctx, sess.Id, model, actor)
if err != nil {
return api.AgentSession{}, err
}
sess.Model = model
}
task, err := m.taskFor(ctx, sess)
if err != nil {
return api.AgentSession{}, err
}
return m.turn(ctx, sess, task, actor, text, images)
}
func (m *Manager) Interrupt(ctx context.Context, sessionID string) (api.AgentSession, error) {
sess, err := m.store.GetSession(ctx, sessionID)
if err != nil {
return api.AgentSession{}, err
}
m.mu.Lock()
l := m.live[sessionID]
m.mu.Unlock()
if sess.State != api.AgentSessionStateRunning || l == nil {
return api.AgentSession{}, fmt.Errorf("%w: session %s has no turn running", ErrConflict, sess.Id)
}
m.mu.Lock()
l.interrupted = true
m.mu.Unlock()
ictx, cancel := context.WithTimeout(ctx, 10*time.Second)
defer cancel()
err = l.sess.interrupt(ictx)
if err != nil {
return api.AgentSession{}, err
}
return sess, nil
}
func (m *Manager) StopTask(ctx context.Context, taskID, actor string) error {
sessions, err := m.store.ListSessions(ctx, taskID)
if err != nil {
return err
}
for _, sess := range sessions {
if sess.State == api.AgentSessionStateCompleted {
continue
}
err = m.retire(ctx, sess, actor)
if err != nil {
return err
}
}
return nil
}
func (m *Manager) retire(ctx context.Context, sess api.AgentSession, actor string) error {
m.mu.Lock()
l := m.live[sess.Id]
if l != nil {
l.stopping = true
}
m.mu.Unlock()
if l != nil {
l.sess.kill()
}
return m.store.SetSessionState(ctx, sess.Id, api.AgentSessionStateCompleted, actor, nil)
}
func (m *Manager) Recover(ctx context.Context) error {
err := m.recoverMerges(ctx)
if err != nil {
return err
}
sessions, err := m.store.SessionsInState(ctx, api.AgentSessionStateRunning)
if err != nil {
return err
}
for _, sess := range sessions {
task, err := m.taskFor(ctx, sess)
if err != nil {
return err
}
l, err := m.ensureLive(ctx, sess, task)
if err != nil {
slog.Error("resume session", "session", sess.Id, "err", err)
m.logUndelivered(ctx, sess.Id, "resume failed")
err = m.store.SetSessionState(ctx, sess.Id, api.AgentSessionStateIdle, "byre", map[string]any{"error": err.Error()})
if err != nil {
return err
}
if sess.Owner == "" {
m.abandonMerge(task.Id, task.ProjectId)
}
continue
}
err = m.message(ctx, l, "byre", restartNotice, nil)
if err != nil {
return err
}
slog.Info("resumed session", "session", sess.Id, "task", task.Id, "owner", sess.Owner)
m.logUndelivered(ctx, sess.Id, "resumed with restart notice")
}
idle, err := m.store.SessionsInState(ctx, api.AgentSessionStateIdle)
if err != nil {
return err
}
for _, sess := range idle {
m.logUndelivered(ctx, sess.Id, "idle at startup")
}
go m.deliverAllLeads()
go m.startHeldMerges(context.Background())
return nil
}
func (m *Manager) logUndelivered(ctx context.Context, sessionID, reason string) {
pending, err := m.store.Undelivered(ctx, sessionID)
if err != nil {
slog.Error("list undelivered responses", "session", sessionID, "err", err)
return
}
if len(pending) == 0 {
return
}
slog.Info("undelivered responses", "session", sessionID, "reason", reason, "responses", responseIDs(pending))
}
func (m *Manager) Close() {
close(m.watchStop)
m.mu.Lock()
m.closing = true
all := []*live{}
for _, l := range m.live {
all = append(all, l)
}
m.mu.Unlock()
for _, l := range all {
l.sess.kill()
}
err := m.parkLeads(context.Background())
if err != nil {
slog.Error("park lead sessions", "err", err)
}
}
func (m *Manager) LiveCount() int {
m.mu.Lock()
defer m.mu.Unlock()
return len(m.live)
}
type Usage struct {
Running []api.TurnRef
RateLimits map[string]map[string]any
}
func (m *Manager) Usage(ctx context.Context) (Usage, error) {
u := Usage{
Running: []api.TurnRef{},
RateLimits: map[string]map[string]any{},
}
running, err := m.store.SessionsInState(ctx, api.AgentSessionStateRunning)
if err != nil {
return u, err
}
for _, sess := range running {
if sess.Owner != "" {
continue
}
u.Running = append(u.Running, api.TurnRef{SessionId: sess.Id, TaskId: sess.TaskId, Vendor: string(sess.Vendor)})
}
m.mu.Lock()
defer m.mu.Unlock()
for vendor, info := range m.rateLimits {
u.RateLimits[string(vendor)] = info
}
return u, nil
}
func (m *Manager) turn(ctx context.Context, sess api.AgentSession, task api.Task, actor, text string, images []string) (api.AgentSession, error) {
err := transition(ctx, m.store, sess.Id, api.AgentSessionStateIdle, api.AgentSessionStateRunning, actor)
if err != nil {
return api.AgentSession{}, err
}
l, err := m.ensureLive(ctx, sess, task)
if err == nil {
err = m.syncModel(ctx, l, sess.Model)
}
if err != nil {
serr := m.store.SetSessionState(ctx, sess.Id, api.AgentSessionStateIdle, actor, map[string]any{"error": err.Error()})
if serr != nil {
slog.Error("record failed start", "session", sess.Id, "err", serr)
}
return api.AgentSession{}, err
}
err = m.message(ctx, l, actor, text, images)
if err != nil {
return api.AgentSession{}, err
}
return m.store.GetSession(ctx, sess.Id)
}
func (m *Manager) syncModel(ctx context.Context, l *live, model string) error {
if model == l.model {
return nil
}
cctx, cancel := context.WithTimeout(ctx, 10*time.Second)
defer cancel()
err := l.sess.setModel(cctx, model)
if err != nil {
return fmt.Errorf("set model %s: %w", model, err)
}
l.model = model
return nil
}
func transition(ctx context.Context, st *store.Store, id string, from, to api.AgentSessionState, actor string) error {
err := st.TransitionSession(ctx, id, from, to, actor)
if errors.Is(err, store.ErrStateChanged) {
return fmt.Errorf("%w: %w", ErrConflict, err)
}
return err
}
func (m *Manager) message(ctx context.Context, l *live, actor, text string, images []string) error {
imgs := []image{}
for _, id := range images {
b, data, err := m.store.ReadBlob(ctx, id)
if err != nil {
return err
}
imgs = append(imgs, image{mediaType: b.ContentType, data: data})
}
payload := map[string]any{"text": text}
if l.owner == "" {
head, err := git.Head(l.dir)
if err != nil {
return err
}
payload["head"] = head
}
if len(images) > 0 {
payload["images"] = images
}
compacting := isHandoffPrompt(actor, text)
if compacting {
payload["handoff"] = true
}
if isContinuation(actor, text) {
payload["continuation"] = true
}
err := m.emit(ctx, l, actor, "user.message", payload)
if err != nil {
return err
}
m.mu.Lock()
l.interrupted = false
l.asked = false
l.askedBlock = false
l.bounced = false
l.wantCompact = false
l.compacting = compacting
l.rateLimited = nil
l.lastText = ""
m.mu.Unlock()
return l.sess.send(ctx, text, imgs)
}
func (m *Manager) ensureLive(ctx context.Context, sess api.AgentSession, task api.Task) (*live, error) {
m.mu.Lock()
defer m.mu.Unlock()
if m.closing {
return nil, errors.New("shutting down")
}
l := m.live[sess.Id]
if l != nil {
return l, nil
}
creator, err := m.store.GetUser(ctx, sess.CreatedBy)
if err != nil {
return nil, err
}
l = &live{
id: sess.Id,
dir: task.WorktreePath,
model: sess.Model,
gitName: creator.GitName,
gitEmail: creator.GitEmail,
vendor: sess.Vendor,
taskID: task.Id,
projectID: task.ProjectId,
owner: sess.Owner,
}
g, err := m.taskGuidance(ctx, task)
if err != nil {
return nil, err
}
tools := []mcp.Tool{m.askTool(l), m.cancelQuestionTool(l), m.createTaskTool(l), m.setMergeTitleTool(l), m.exposePortTool(l)}
env := portEnv(task)
if l.owner != "" {
l.dir = filepath.Join(m.leadRoot, url.PathEscape(l.owner))
err = os.MkdirAll(l.dir, 0o755)
if err != nil {
return nil, err
}
g = leadGuidance
tools = m.leadTools(l)
env = nil
}
servers := []*mcp.Server{{Name: "byre", Version: "0.0.0", Tools: tools}}
sctx, cancel := context.WithTimeout(ctx, 60*time.Second)
defer cancel()
switch l.vendor {
case api.Codex:
l.sess, err = m.startCodex(sctx, sess, l, g, servers, env)
default:
l.sess, err = m.startClaude(sctx, sess, l, g, servers, env)
}
if err != nil {
return nil, err
}
m.live[sess.Id] = l
go m.pump(l)
return l, nil
}
func (m *Manager) pump(l *live) {
ctx := context.Background()
l.sess.drive(m, l)
m.mu.Lock()
delete(m.live, l.id)
quiet := m.closing || l.stopping
m.mu.Unlock()
if quiet {
m.logUndelivered(ctx, l.id, "agent process ended while closing or stopping")
return
}
sess, err := m.store.GetSession(ctx, l.id)
if err != nil {
slog.Error("load exited session", "session", l.id, "err", err)
return
}
if sess.State != api.AgentSessionStateRunning {
return
}
m.logUndelivered(ctx, l.id, "agent process exited mid-turn")
err = m.store.SetSessionState(ctx, l.id, api.AgentSessionStateIdle, agentActor, map[string]any{"error": "agent process exited mid-turn"})
if err != nil {
slog.Error("record session exit", "session", l.id, "err", err)
}
if l.owner == "" {
m.abandonMerge(l.taskID, l.projectID)
m.fileError(l, "The agent process exited mid-turn. Send a message to resume it.")
}
}
func (m *Manager) handle(ctx context.Context, l *live, ev claude.Event) error {
switch ev.Type {
case "assistant":
blocks, err := ev.Blocks()
if err != nil {
return err
}
parent, err := ev.ParentToolUseID()
if err != nil {
return err
}
uuid, err := ev.UUID()
if err != nil {
return err
}
for _, b := range blocks {
err = m.handleAssistantBlock(ctx, l, b, parent, uuid)
if err != nil {
return err
}
}
if parent != "" {
return nil
}
usage, err := ev.ContextUsage()
if err != nil {
return err
}
if usage.Tokens == 0 {
return nil
}
l.contextModel = usage.Model
return m.updateContext(ctx, l, usage.Tokens, l.contextWindow)
case "user":
blocks, err := ev.Blocks()
if err != nil {
return err
}
parent, err := ev.ParentToolUseID()
if err != nil {
return err
}
uuid, err := ev.UUID()
if err != nil {
return err
}
for _, b := range blocks {
if b.Type != "tool_result" {
continue
}
content, err := decode(b.Content)
if err != nil {
return err
}
err = m.emit(ctx, l, agentActor, "agent.tool_result", withParent(map[string]any{
"tool_use_id": b.ToolUseID,
"is_error": b.IsError,
"content": content,
}, parent, uuid))
if err != nil {
return err
}
}
return nil
case "result":
res, err := ev.Result()
if err != nil {
return err
}
return m.turnEnded(ctx, l, endResult{
subtype: res.Subtype,
isError: res.IsError,
result: res.Result,
payload: store.ToPayload(res),
window: res.ModelUsage[l.contextModel].ContextWindow,
})
case "stream_event":
text, err := ev.TextDelta()
if err != nil || text == "" {
return err
}
m.store.Broadcast(m.event(l, agentActor, "agent.delta", map[string]any{"text": text}))
return nil
case "rate_limit_event":
info := struct {
RateLimitInfo map[string]any `json:"rate_limit_info"`
}{}
err := json.Unmarshal(ev.Raw, &info)
if err != nil {
return err
}
m.mu.Lock()
m.rateLimits[l.vendor] = info.RateLimitInfo
if info.RateLimitInfo["status"] == "rejected" {
l.rateLimited = info.RateLimitInfo
}
m.mu.Unlock()
case "system":
if ev.Subtype == "init" {
return nil
}
}
raw := map[string]any{}
err := json.Unmarshal(ev.Raw, &raw)
if err != nil {
return err
}
return m.emit(ctx, l, agentActor, "agent.vendor", raw)
}
type endResult struct {
subtype string
isError bool
result string
payload map[string]any
window int
}
func (m *Manager) turnEnded(ctx context.Context, l *live, res endResult) error {
if l.owner != "" {
return m.leadTurnEnded(ctx, l, res)
}
m.mu.Lock()
interrupted := l.interrupted
rateLimited := l.rateLimited
compacting := l.compacting
askedBlock := l.askedBlock
m.mu.Unlock()
finished := !res.isError && !interrupted && strings.TrimSpace(res.result) != ""
handoff := compacting && finished
payload := res.payload
sha, err := git.CommitAll(l.dir, commitMessage(res), l.gitName, l.gitEmail)
if err != nil {
slog.Error("commit turn", "session", l.id, "err", err)
}
if sha != "" {
payload["commit"] = sha
}
switch {
case handoff:
payload["handoff"] = true
case finished && !compacting && !askedBlock:
payload["inbox_item_id"] = m.fileSummary(l, res.result)
}
err = m.emit(ctx, l, agentActor, "turn.ended", payload)
if err != nil {
return err
}
m.snapshotSoon(l, api.SnapshotTriggerTurnEnd, 0)
if res.window > 0 {
err = m.updateContext(ctx, l, l.contextUsed, res.window)
if err != nil {
return err
}
}
failure := fmt.Sprintf("The turn ended with an error (%s): %s", res.subtype, res.result)
switch {
case !res.isError || interrupted:
case rateLimited != nil:
m.fileRateLimited(l, rateLimited, failure)
default:
m.fileError(l, failure)
}
err = m.store.SetSessionState(ctx, l.id, api.AgentSessionStateIdle, agentActor, nil)
if err != nil {
return err
}
if handoff {
m.logUndelivered(ctx, l.id, "turn ended in handoff")
go m.continueFromHandoff(l.id, res.result)
return nil
}
mergeBy := m.takeMerge(l.taskID)
switch {
case mergeBy != "" && finished:
m.logUndelivered(ctx, l.id, "turn ended in merge")
go m.finishMerge(l, mergeBy)
return nil
case mergeBy != "":
go m.releaseMerge(l.projectID, l.taskID)
default:
go m.nextMerge(l.projectID)
}
err = m.Deliver(ctx, l.id)
if err != nil {
slog.Error("deliver inbox responses", "session", l.id, "err", err)
}
return nil
}
func commitMessage(res endResult) string {
line, _, _ := strings.Cut(strings.TrimSpace(res.result), "\n")
sentence, _, found := strings.Cut(line, ". ")
if found {
sentence += "."
}
sentence = strings.TrimSpace(strings.ReplaceAll(sentence, "**", ""))
if sentence == "" {
return fmt.Sprintf("Work from a turn that ended without a summary (%s)", res.subtype)
}
runes := []rune(sentence)
if len(runes) > 100 {
return string(runes[:99]) + "…"
}
return sentence
}
func (m *Manager) updateContext(ctx context.Context, l *live, used, window int) error {
if used == l.contextUsed && window == l.contextWindow {
return nil
}
l.contextUsed = used
l.contextWindow = window
return m.emit(ctx, l, agentActor, "session.context", map[string]any{"model": l.contextModel, "used": used, "window": window})
}
func (m *Manager) handleAssistantBlock(ctx context.Context, l *live, b claude.Block, parent, uuid string) error {
switch b.Type {
case "text":
return m.emit(ctx, l, agentActor, "agent.message", withParent(map[string]any{"text": b.Text}, parent, uuid))
case "thinking":
if b.Thinking == "" {
return nil
}
return m.emit(ctx, l, agentActor, "agent.thinking", withParent(map[string]any{"text": b.Thinking}, parent, uuid))
case "tool_use":
input, err := decode(b.Input)
if err != nil {
return err
}
return m.emit(ctx, l, agentActor, "agent.tool_call", withParent(map[string]any{
"id": b.ID,
"name": b.Name,
"input": input,
}, parent, uuid))
}
return m.emit(ctx, l, agentActor, "agent.vendor", withParent(store.ToPayload(b), parent, uuid))
}
func withParent(payload map[string]any, parent, uuid string) map[string]any {
if parent != "" {
payload["parent_tool_use_id"] = parent
return payload
}
if uuid != "" {
payload["vendor_uuid"] = uuid
}
return payload
}
func (m *Manager) emit(ctx context.Context, l *live, actor, kind string, payload map[string]any) error {
return m.store.AppendEvent(ctx, m.event(l, actor, kind, payload))
}
func (m *Manager) event(l *live, actor, kind string, payload map[string]any) api.Event {
ev := api.Event{
Kind: kind,
AgentSessionId: &l.id,
Actor: actor,
Payload: payload,
}
if l.owner != "" {
ev.Owner = &l.owner
return ev
}
ev.ProjectId = &l.projectID
ev.TaskId = &l.taskID
return ev
}
func decode(raw json.RawMessage) (any, error) {
if len(raw) == 0 {
return nil, nil
}
var v any
err := json.Unmarshal(raw, &v)
return v, err
}
func newUUID() string {
b := make([]byte, 16)
rand.Read(b)
b[6] = b[6]&0x0f | 0x40
b[8] = b[8]&0x3f | 0x80
return fmt.Sprintf("%x-%x-%x-%x-%x", b[0:4], b[4:6], b[6:8], b[8:10], b[10:])
}