Add federation worker and canonical handoffs

This commit is contained in:
kami
2026-07-28 16:17:18 +04:00
parent 58793a5aa3
commit 2cecbc4015
22 changed files with 1429 additions and 108 deletions
+357
View File
@@ -0,0 +1,357 @@
// orchestra-worker consumes router-issued leases for one local herdr. Homesrv
// remains the scheduler and CAS authority; this process owns only local Git
// and pane operations.
package main
import (
"context"
"encoding/json"
"fmt"
"log"
"orchestra/internal/continuity"
"orchestra/internal/domain"
"orchestra/internal/federation"
"orchestra/internal/herdr"
"orchestra/internal/orchestrator"
"os"
"os/exec"
"os/signal"
"path/filepath"
"strconv"
"strings"
"syscall"
"time"
)
type worker struct {
api federation.Client
herdr *herdr.Client
harnessID, harness, repo, root, remote string
cursor uint64
tasks map[string]domain.Task
sessions map[string]herdr.Session
leases map[string]lease
statePath string
hard float64
registration federation.Worker
}
type lease struct {
HandoffRef string `json:"handoff_ref,omitempty"`
}
type workerState struct {
Cursor uint64 `json:"cursor"`
Sessions map[string]herdr.Session `json:"sessions"`
Tasks map[string]domain.Task `json:"tasks"`
Leases map[string]lease `json:"leases"`
}
func (w *worker) load() {
b, e := os.ReadFile(w.statePath)
if e == nil {
var s workerState
if json.Unmarshal(b, &s) == nil {
w.cursor = s.Cursor
w.sessions = s.Sessions
w.tasks = s.Tasks
w.leases = s.Leases
}
}
if w.sessions == nil {
w.sessions = map[string]herdr.Session{}
}
if w.tasks == nil {
w.tasks = map[string]domain.Task{}
}
if w.leases == nil {
w.leases = map[string]lease{}
}
}
func (w *worker) save() error {
b, e := json.Marshal(workerState{Cursor: w.cursor, Sessions: w.sessions, Tasks: w.tasks, Leases: w.leases})
if e != nil {
return e
}
if e := os.MkdirAll(filepath.Dir(w.statePath), 0700); e != nil {
return e
}
return os.WriteFile(w.statePath, b, 0600)
}
type artifactCAS struct{ api federation.Client }
func (c artifactCAS) PutArtifact(b []byte) (string, error) {
return c.api.PutArtifact(context.Background(), b)
}
func (c artifactCAS) Artifact(ref string) ([]byte, error) {
return c.api.Artifact(context.Background(), ref)
}
func created(e domain.Event) (domain.Task, bool) {
if e.Type != "TaskCreated" {
return domain.Task{}, false
}
var p struct {
Source string `json:"source"`
ExternalID string `json:"external_id"`
Project string `json:"project"`
Capability []string `json:"capability"`
Title string `json:"title"`
Description string `json:"description"`
}
if json.Unmarshal(e.Payload, &p) != nil || p.Source == "" || p.ExternalID == "" || p.Project == "" {
return domain.Task{}, false
}
return domain.Task{ID: e.TaskID, Source: p.Source, ExternalID: p.ExternalID, Project: p.Project, Capability: p.Capability, Title: p.Title, Description: p.Description}, true
}
func (w *worker) syncBase(ctx context.Context) error {
if out, err := exec.CommandContext(ctx, "git", "-C", w.repo, "fetch", w.remote, "--prune").CombinedOutput(); err != nil {
return fmt.Errorf("fetch base checkout: %s: %w", out, err)
}
branch, err := exec.CommandContext(ctx, "git", "-C", w.repo, "symbolic-ref", "--quiet", "--short", "HEAD").Output()
if err != nil {
return fmt.Errorf("identify base branch: %w", err)
}
branchName := strings.TrimSpace(string(branch))
if out, err := exec.CommandContext(ctx, "git", "-C", w.repo, "merge", "--ff-only", w.remote+"/"+branchName).CombinedOutput(); err != nil {
return fmt.Errorf("fast-forward base checkout: %s: %w", out, err)
}
return nil
}
func (w *worker) start(ctx context.Context, t domain.Task, ref string) error {
var wt string
var h continuity.Handoff
var err error
// Synchronize the local base before any worktree operation. A worker never
// treats a coordinator-side path as truth; the Git remote is the only
// cross-machine transport.
if err := w.syncBase(ctx); err != nil {
return err
}
if ref != "" {
b, err := w.api.Artifact(ctx, ref)
if err != nil {
return err
}
h, err = continuity.Decode(b)
if err != nil {
return err
}
if out, err := exec.CommandContext(ctx, "git", "-C", w.repo, "fetch", w.remote, "--prune").CombinedOutput(); err != nil {
return fmt.Errorf("fetch pickup anchor: %s: %w", out, err)
}
wt = filepath.Join(w.root, t.ID)
if _, err := os.Stat(wt); os.IsNotExist(err) {
if out, err := exec.CommandContext(ctx, "git", "-C", w.repo, "worktree", "add", "-b", "orchestra/"+t.ID, wt, h.Anchor.GitSHA).CombinedOutput(); err != nil {
return fmt.Errorf("create pickup worktree: %s: %w", out, err)
}
}
if err = continuity.ValidatePickup(wt, h, taskHash(t)); err != nil {
return err
}
} else {
wt, err = (orchestrator.GitWorktrees{Repo: w.repo, Root: w.root}).Create(ctx, t)
if err != nil {
return err
}
}
if _, err = w.herdr.Worktree(ctx, w.repo, wt, "orchestra/"+t.ID); err != nil {
return err
}
s, err := w.herdr.StartAgent(ctx, wt, wt, "orchestra/"+t.ID, w.harness, t.ID)
if err != nil {
return err
}
p := "Begin Orchestra task " + t.ID + ".\nTitle: " + t.Title + "\nInstructions:\n" + t.Description + "\nWork only in this worktree. Do not edit TASK.md."
if ref != "" {
p += "\nA validated handoff exists. Read TASK.md and inspect local git history before continuing."
}
w.sessions[t.ID] = s
if err := w.save(); err != nil {
return err
}
// A prompt response can be lost after herdr accepted it. Persist the
// session first so the worker can reconcile/release it after restart.
return w.herdr.Prompt(ctx, s.PaneID, p, 0)
}
func taskHash(t domain.Task) string { b := continuity.RenderTaskFile(t); return domain.Hash(b) }
func (w *worker) releaseReady(ctx context.Context) {
for id, s := range w.sessions {
if report, err := os.ReadFile(filepath.Join(s.Worktree, ".orchestra-report.md")); err == nil && len(report) > 0 {
if ref, err := w.api.PutArtifact(ctx, report); err == nil {
if err = w.api.Complete(ctx, id, ref); err == nil {
delete(w.sessions, id)
delete(w.leases, id)
_ = w.save()
continue
}
log.Printf("complete %s: %v", id, err)
} else {
log.Printf("upload completion %s: %v", id, err)
}
}
if _, err := os.Stat(filepath.Join(s.Worktree, herdr.HandoffReportFile)); err != nil {
if !s.HandoffRequested {
a := herdr.CLIAdapter{Client: w.herdr, Harness: w.harness}
if occ, occErr := a.Occupancy(s); occErr == nil && occ >= w.hard {
if boundary, boundaryErr := a.AtTurnBoundary(ctx, s); boundaryErr == nil && boundary {
if err := a.RequestHandoff(ctx, s); err == nil {
s.HandoffRequested, s.HandoffReason = true, "threshold"
w.sessions[id] = s
_ = w.save()
}
}
}
}
continue
}
a := herdr.CLIAdapter{Client: w.herdr, Harness: w.harness, CAS: artifactCAS{w.api}, Remote: w.remote}
ref, err := a.Release(ctx, s)
if err != nil {
log.Printf("release %s: %v", id, err)
continue
}
sha, err := herdr.HeadSHA(s.Worktree)
if err != nil {
log.Printf("release %s anchor: %v", id, err)
continue
}
if err = w.api.Release(ctx, id, ref, sha); err != nil {
log.Printf("publish release %s: %v", id, err)
continue
}
// release_agent only removes herdr's binding. Closing the released pane
// after the handoff is durable prevents the next StartAgent from
// inheriting the predecessor's still-running terminal process.
if err := a.Kill(ctx, s); err != nil {
log.Printf("close released pane %s: %v", id, err)
}
delete(w.sessions, id)
_ = w.save()
}
}
func (w *worker) once(ctx context.Context) error {
es, _, err := w.api.Events(ctx, w.cursor)
if err != nil {
return err
}
for _, e := range es {
if t, ok := created(e); ok {
w.tasks[t.ID] = t
}
if e.Type == "TaskLeased" {
var p struct {
HarnessID string `json:"harness_id"`
HandoffRef string `json:"handoff_ref"`
}
if json.Unmarshal(e.Payload, &p) == nil && p.HarnessID == w.harnessID {
w.leases[e.TaskID] = lease{HandoffRef: p.HandoffRef}
}
}
if e.Type == "TaskReleased" || e.Type == "TaskCompleted" || e.Type == "TaskFailed" || e.Type == "TaskBlocked" {
delete(w.leases, e.TaskID)
delete(w.sessions, e.TaskID)
}
if e.Seq > w.cursor {
w.cursor = e.Seq
}
}
// State is only a cache. If a lease survived but its TaskCreated event is
// older than the worker's cursor (or the event has been compacted), hydrate
// the authoritative task projection before deciding whether to start.
for taskID := range w.leases {
if _, ok := w.tasks[taskID]; !ok {
tasks, err := w.api.Tasks(ctx)
if err != nil {
return fmt.Errorf("hydrate leased task %s: %w", taskID, err)
}
for _, task := range tasks {
w.tasks[task.ID] = task
}
break
}
}
// Project the whole batch before launching. This prevents a new worker
// from resurrecting every historical lease during its initial replay.
for taskID, l := range w.leases {
if _, started := w.sessions[taskID]; started {
continue
}
if t, ok := w.tasks[taskID]; ok {
if err := w.start(ctx, t, l.HandoffRef); err != nil {
log.Printf("lease %s: %v", t.ID, err)
}
}
}
w.releaseReady(ctx)
if err := w.save(); err != nil {
return err
}
return w.api.Ack(ctx, w.cursor)
}
// reRegisterAfterCoordinatorRestart restores the coordinator's in-memory
// worker registry. A worker must survive a server restart without operator
// intervention; its persisted cursor and sessions remain valid.
func (w *worker) reRegisterAfterCoordinatorRestart(ctx context.Context, cause error) bool {
if cause == nil || !strings.Contains(cause.Error(), "401 Unauthorized: unknown worker") {
return false
}
if err := w.api.Register(ctx, w.registration); err != nil {
log.Printf("re-register: %v", err)
return false
}
log.Printf("re-registered after coordinator restart")
return true
}
func required(k string) string {
v := os.Getenv(k)
if v == "" {
log.Fatalf("%s required", k)
}
return v
}
func main() {
id, token := required("ORCHESTRA_WORKER_ID"), required("ORCHESTRA_WORKER_TOKEN")
hard := .75
if v, err := strconv.ParseFloat(os.Getenv("ORCHESTRA_OCCUPANCY_HARD"), 64); err == nil && v > 0 && v < 1 {
hard = v
}
w := &worker{api: federation.Client{BaseURL: required("ORCHESTRA_URL"), WorkerID: id, Token: token, AdmitToken: os.Getenv("ORCHESTRA_FEDERATION_ADMIT_TOKEN")}, harnessID: required("ORCHESTRA_WORKER_HERDR_ID"), harness: required("ORCHESTRA_WORKER_HARNESS"), repo: required("ORCHESTRA_REPO"), root: required("ORCHESTRA_WORKTREE_ROOT"), remote: required("ORCHESTRA_GIT_REMOTE"), tasks: map[string]domain.Task{}, sessions: map[string]herdr.Session{}, leases: map[string]lease{}, statePath: os.Getenv("ORCHESTRA_WORKER_STATE"), hard: hard, registration: federation.Worker{ID: id, Address: os.Getenv("ORCHESTRA_WORKER_ADDRESS"), Capacity: 1}}
if w.statePath == "" {
w.statePath = filepath.Join(w.root, ".orchestra-worker-state.json")
}
w.load()
w.herdr = herdr.New(required("ORCHESTRA_WORKER_HERDR"))
if id != w.harnessID {
log.Fatal("ORCHESTRA_WORKER_ID must equal ORCHESTRA_WORKER_HERDR_ID so leases and offline recovery have one owner")
}
if err := w.api.Register(context.Background(), w.registration); err != nil {
log.Fatal(err)
}
ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
defer stop()
ticker := time.NewTicker(5 * time.Second)
defer ticker.Stop()
for {
if err := w.api.Heartbeat(ctx); err != nil {
log.Printf("heartbeat: %v", err)
if w.reRegisterAfterCoordinatorRestart(ctx, err) {
continue
}
}
if err := w.once(ctx); err != nil {
log.Printf("poll: %v", err)
w.reRegisterAfterCoordinatorRestart(ctx, err)
}
select {
case <-ctx.Done():
return
case <-ticker.C:
}
}
}
+151
View File
@@ -0,0 +1,151 @@
package main
import (
"bufio"
"context"
"encoding/json"
"errors"
"net"
"net/http"
"net/http/httptest"
"orchestra/internal/domain"
"orchestra/internal/federation"
"orchestra/internal/herdr"
"os"
"os/exec"
"path/filepath"
"testing"
)
func TestWorkerReregistersAfterCoordinatorForgetsIt(t *testing.T) {
registered := false
s := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
if r.Method != http.MethodPost || r.URL.Path != "/v1/federation/workers" {
t.Fatalf("unexpected %s %s", r.Method, r.URL.Path)
}
registered = true
w.WriteHeader(http.StatusCreated)
}))
defer s.Close()
w := &worker{api: federation.Client{BaseURL: s.URL, WorkerID: "h", Token: "t"}, registration: federation.Worker{ID: "h", Capacity: 1}}
if !w.reRegisterAfterCoordinatorRestart(context.Background(), errors.New("federation: 401 Unauthorized: unknown worker")) {
t.Fatal("worker did not identify a coordinator restart")
}
if !registered {
t.Fatal("worker did not register")
}
}
func TestInitialReplayDoesNotResurrectReleasedLease(t *testing.T) {
s := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
switch r.URL.Path {
case "/v1/federation/events":
_, _ = w.Write([]byte(`{"cursor":0,"events":[{"seq":1,"id":"c","type":"TaskCreated","task_id":"t","version":1,"payload":{"source":"s","external_id":"x","project":"p"},"surface":"system"},{"seq":2,"id":"l","type":"TaskLeased","task_id":"t","version":2,"payload":{"harness_id":"h"},"surface":"system"},{"seq":3,"id":"r","type":"TaskReleased","task_id":"t","version":3,"payload":{"reason":"expired"},"surface":"system"}]}`))
case "/v1/federation/events/ack":
w.WriteHeader(http.StatusNoContent)
default:
t.Fatalf("unexpected %s", r.URL.Path)
}
}))
defer s.Close()
w := &worker{api: federation.Client{BaseURL: s.URL, WorkerID: "h", Token: "t"}, harnessID: "h", tasks: map[string]domain.Task{}, sessions: map[string]herdr.Session{}, leases: map[string]lease{}, statePath: t.TempDir() + "/state.json", hard: .75}
if err := w.once(context.Background()); err != nil {
t.Fatal(err)
}
if len(w.leases) != 0 || len(w.sessions) != 0 {
t.Fatalf("replayed lease survived: leases=%v sessions=%v", w.leases, w.sessions)
}
if w.cursor != 3 {
t.Fatalf("cursor=%d want 3", w.cursor)
}
}
func TestWorkerHydratesLeaseMissingTaskCache(t *testing.T) {
s := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
switch r.URL.Path {
case "/v1/federation/events":
_, _ = w.Write([]byte(`{"cursor":2,"events":[{"seq":2,"id":"l","type":"TaskLeased","task_id":"t","version":2,"payload":{"harness_id":"h"},"surface":"system"}]}`))
case "/v1/tasks":
_, _ = w.Write([]byte(`[{"id":"t","source":"s","external_id":"x","project":"p","state":"leased"}]`))
case "/v1/federation/events/ack":
w.WriteHeader(http.StatusNoContent)
default:
t.Fatalf("unexpected %s", r.URL.Path)
}
}))
defer s.Close()
w := &worker{api: federation.Client{BaseURL: s.URL, WorkerID: "h", Token: "t"}, harnessID: "h", cursor: 1, tasks: map[string]domain.Task{}, sessions: map[string]herdr.Session{"t": {}}, leases: map[string]lease{}, statePath: t.TempDir() + "/state.json", hard: .75}
if err := w.once(context.Background()); err != nil {
t.Fatal(err)
}
if _, ok := w.tasks["t"]; !ok || w.cursor != 2 {
t.Fatalf("task was not hydrated: tasks=%v cursor=%d", w.tasks, w.cursor)
}
}
func TestWorkerStartsRouterIssuedLeaseInLocalGitWorktree(t *testing.T) {
remote := filepath.Join(t.TempDir(), "remote.git")
if out, err := exec.Command("git", "init", "--bare", remote).CombinedOutput(); err != nil {
t.Fatalf("remote: %v %s", err, out)
}
seed := t.TempDir()
for _, a := range [][]string{{"init", seed}, {"-C", seed, "config", "user.email", "t@t"}, {"-C", seed, "config", "user.name", "t"}, {"-C", seed, "commit", "--allow-empty", "-m", "init"}, {"-C", seed, "remote", "add", "origin", remote}, {"-C", seed, "push", "-u", "origin", "HEAD:master"}} {
if out, err := exec.Command("git", a...).CombinedOutput(); err != nil {
t.Fatalf("git %v: %v %s", a, err, out)
}
}
repo := filepath.Join(t.TempDir(), "repo")
if out, err := exec.Command("git", "clone", remote, repo).CombinedOutput(); err != nil {
t.Fatal(string(out))
}
ln, err := net.Listen("tcp", "127.0.0.1:0")
if err != nil {
t.Fatal(err)
}
defer ln.Close()
go func() {
for {
c, e := ln.Accept()
if e != nil {
return
}
go func() {
defer c.Close()
var r herdr.Request
if json.NewDecoder(bufio.NewReader(c)).Decode(&r) != nil {
return
}
var result string
switch r.Method {
case "worktree.create":
result = `{"path":"` + filepath.Join(filepath.Dir(repo), "worktrees", "task") + `","root_pane":{"pane_id":"p1"}}`
case "agent.start":
result = `{}`
case "pane.get":
result = `{"pane":{"agent":"opencode","agent_status":"idle"}}`
case "pane.read":
result = `{"read":{"text":""}}`
case "agent.prompt":
result = `{}`
default:
result = `{}`
}
_ = json.NewEncoder(c).Encode(herdr.Response{ID: r.ID, Result: json.RawMessage(result)})
}()
}
}()
root := filepath.Join(filepath.Dir(repo), "worktrees")
w := &worker{herdr: &herdr.Client{Path: ln.Addr().String()}, repo: repo, root: root, remote: "origin", harness: "opencode", tasks: map[string]domain.Task{}, sessions: map[string]herdr.Session{}, leases: map[string]lease{}, statePath: filepath.Join(t.TempDir(), "state.json"), hard: .75}
// New() is needed for its pane map; override the address for the fake.
w.herdr = herdr.New(ln.Addr().String())
task := domain.Task{ID: "task", Source: "s", ExternalID: "x", Project: "p", Title: "test", Description: "do work"}
if err := w.start(context.Background(), task, ""); err != nil {
t.Fatal(err)
}
if _, ok := w.sessions[task.ID]; !ok {
t.Fatal("session not persisted")
}
if _, err := os.Stat(filepath.Join(root, "task", "TASK.md")); err != nil {
t.Fatalf("TASK.md: %v", err)
}
}