Run the wizard checks and task runners in isolated python mode (-I)

.opencode/session-goal.md

@@ -38,6 +38,22 @@
 ## Plan (предыдущая задача — выполнена)
 1–7. Docker E2E пайплайна «install как человек → serve → визард → воркер → джоб» — выполнено, см. Progress ниже.
 
+## Progress (день: конкурентность в агенте — параллелизм через нашу архитектуру)
+- ✅ По просьбе «реализовать параллелизм через нашу архитектуру»: `WORKER_CONCURRENCY` (поле `concurrency` в конфиге визарда, шаг Machine) — агент регистрируется один раз и ведёт N циклов claim→выполнение→upload под одним worker id. N шардов обрабатываются параллельно на одной машине, используя таски координатора как единицу параллелизма; SDK не менялся.
+- ✅ Реализация: `Config.Concurrency` + env; `Daemon.RunForever` → register once → N goroutine-циклов (общий счётчик MaxTasks под мьютексом); визард: поле «Concurrent task loops» + конфиг; тест с маркер-скриптом доказывает 3 параллельных исполнения (race-тесты зелёные после мьютекса в fake).
+- ✅ Измерено на релизном коде в Docker (один воркер, 8 шардов): concurrency=1 → 12s; concurrency=4 → 4s (**3×**). Плюс прежний `similarity-search-parallel` (потоки внутри шарда) композируется с конкурентностью.
+- ✅ Релиз v1.1.0-alpha.20 (бинарники + wheel); полный гейт: race + lint 0 issues + pytest 213.
+- ✅ Бинарник worker-agent на машине пользователя обновлён до alpha.20.
+- GIL-высвобождения в RDKit нет (проверено: `RDKIT_ALLOW_GIL_RELEASE` не помогает), поэтому внутришардовые потоки не ускоряют чистый RDKit-путь — конкурентность задач это и компенсирует.
+
+## Progress (утро: similarity-search-parallel)
+- ✅ Новый workload `similarity-search-parallel@1.0.0` (отдельная версия, как просил пользователь): подкласс `SimilaritySearchSDKWorkload` + параллельное ядро `search_parallel/core.py` — fingerprinting+скоринг шарда через `ThreadPoolExecutor` (параметр `threads`, default CPU count); `pool.map` сохраняет порядок строк, поэтому merge идентичен последовательному (`_HeapEntry`) и результат **байт-в-байт** равен `similarity-search` при любом числе потоков.
+- ✅ Тесты: байт-в-байт vs эталон для threads 1/2/4 с намеренными связями (изомеры, дубликаты), executor-прогон, валидация параметров, регистрация манифеста; 213 pytest зелёные.
+- ✅ Экспортирован в каталог координатора (5 ворклоадов), Go-тесты зелёные; release v1.1.0-alpha.19 (бинарники + wheel `scimesh-1.1.0a19`).
+- ✅ Распределённый E2E в Docker: воркер с wheel a19, джоб similarity-search-parallel (threads=4) → completed → результат байт-в-байт = локальному эталону.
+- ✅ Бинарники на машине пользователя обновлены до alpha.19 (для появления ворклоада в UI нужен рестарт serve + переустановка рантайма воркера).
+- Примечание: в CPython потоки не ускоряют чистый RDKit-путь (GIL), но структура готова к ядрам, отпускающим GIL (numpy и т.п.); при желании можно добавить процесс-пул как отдельный workload в будущей версии протокола.
+
 ## Progress (ночная сессия)
 - ✅ **П.1 Пустое имя воркера**: `domain.NewWorker` нормализует/отклоняет пустое имя + `TestNewWorkerRejectsBlankName` (починен `fixedTime`→`testNow`).
 - ✅ **П.2 `--check`**: пробует managed venv (если установлен) + реальный пробинг учётки (exchange ключа / claim-пробa) — `CheckAuth` + тесты; на машине пользователя: `✓ auth: credential accepted`, venv python, scimesh installed.

coordinator/cmd/coordinator/agent_cmd.go

@@ -45,7 +45,7 @@ func runAgent(args []string) error {
 		return fmt.Errorf("--coordinator-url, --token, and --work-dir are required")
 	}
 	if *taskRunner == "" {
-		*taskRunner = "python -m scimesh.worker.task"
+		*taskRunner = "python -I -m scimesh.worker.task"
 	}
 
 	logger := slog.New(slog.NewTextHandler(os.Stderr, nil))

coordinator/cmd/coordinator/serve_cmd.go

@@ -236,9 +236,9 @@ func stopAgents(agents []*exec.Cmd) {
 // system `python`.
 func defaultTaskRunner(venvPython string) string {
 	if runtimeStatus(venvPython) {
-		return venvPython + " -m scimesh.worker.task"
+		return venvPython + " -I -m scimesh.worker.task"
 	}
-	return "python -m scimesh.worker.task"
+	return "python -I -m scimesh.worker.task"
 }
 
 // ensureRuntime creates the managed venv and installs scimesh into it, unless

coordinator/internal/agent/check.go

@@ -88,8 +88,12 @@ func CheckEnvironment(ctx context.Context) CheckReport {
 // execute with.
 func CheckEnvironmentWithPython(ctx context.Context, python string) CheckReport {
 	report := CheckReport{Agent: Version, Python: CheckItem{Name: "python", OK: true, Detail: python}}
+	// The version comes from importlib.metadata, so the wizard can compare the
+	// installed package with the binary version and offer an upgrade. -I keeps
+	// the working directory out of sys.path, so a scimesh checkout in the
+	// wizard's cwd can never shadow the venv installation.
 	//nolint:gosec // G204: python is a resolved interpreter path, the argument list is constant
-	cmd := exec.CommandContext(ctx, python, "-c", "import scimesh; print(scimesh.__version__ if hasattr(scimesh, '__version__') else 'installed')")
+	cmd := exec.CommandContext(ctx, python, "-I", "-c", "import importlib.metadata as m; print(m.version('scimesh'))")
 	out, err := cmd.Output()
 	if err != nil {
 		// The worker executes workloads by spawning scimesh's task runner, so

coordinator/internal/agent/config.go

@@ -32,6 +32,10 @@ type Config struct {
 	TaskRunner     []string // command + args; defaults to python -m scimesh.worker.task
 	MaxTasks       int      // 0 = unlimited
 	ExitWhenIdle   bool
+	// Concurrency is how many claim→execute→upload loops run in parallel
+	// under one worker id: N shards processed concurrently on one machine,
+	// using the coordinator's own task pipeline as the parallel unit.
+	Concurrency int
 }
 
 func envList(name string) ([]string, error) {
@@ -108,7 +112,7 @@ func LoadConfig() (*Config, error) {
 		return nil, err
 	}
 	if len(runner) == 0 {
-		runner = []string{"python", "-m", "scimesh.worker.task"}
+		runner = []string{"python", "-I", "-m", "scimesh.worker.task"}
 	}
 	maxTasks := 0
 	if raw := os.Getenv("MAX_TASKS"); raw != "" {
@@ -145,9 +149,22 @@ func LoadConfig() (*Config, error) {
 		TaskRunner:     runner,
 		MaxTasks:       maxTasks,
 		ExitWhenIdle:   os.Getenv("EXIT_WHEN_IDLE") == "1",
+		Concurrency:    envInt("WORKER_CONCURRENCY", 1),
 	}, nil
 }
 
+func envInt(name string, fallback int) int {
+	raw := os.Getenv(name)
+	if raw == "" {
+		return fallback
+	}
+	parsed, err := strconv.Atoi(raw)
+	if err != nil || parsed < 1 {
+		return fallback
+	}
+	return parsed
+}
+
 func durationEnv(name string, fallback time.Duration) (time.Duration, error) {
 	raw := os.Getenv(name)
 	if raw == "" {

coordinator/internal/agent/configfile.go

@@ -21,6 +21,7 @@ type ConfigFile struct {
 	WorkerName     string   `json:"worker_name,omitempty"`
 	CPUCount       int      `json:"cpu_count"`
 	MemoryMB       int      `json:"memory_mb"`
+	Concurrency    int      `json:"concurrency,omitempty"`
 	TaskRunner     []string `json:"task_runner,omitempty"`
 }
 
@@ -103,11 +104,15 @@ func (f *ConfigFile) Config() (*Config, error) {
 	if config.MemoryMB < 0 {
 		config.MemoryMB = 0
 	}
+	config.Concurrency = f.Concurrency
+	if config.Concurrency < 1 {
+		config.Concurrency = 1
+	}
 	if len(f.TaskRunner) > 0 {
 		config.TaskRunner = f.TaskRunner
 	}
 	if len(config.TaskRunner) == 0 {
-		config.TaskRunner = []string{"python", "-m", "scimesh.worker.task"}
+		config.TaskRunner = []string{"python", "-I", "-m", "scimesh.worker.task"}
 	}
 	config.PollInterval = 2 * time.Second
 	config.RequestTimeout = 30 * time.Second

coordinator/internal/agent/daemon.go

@@ -37,15 +37,52 @@ func NewDaemon(config *Config, client *Client, runner *TaskRunner, log *slog.Log
 	return &Daemon{config: config, client: client, runner: runner, log: log}
 }
 
-// RunForever loops until interrupted, idle-exit, or max tasks.
+// RunForever registers once, then runs the claim→execute→upload loop
+// concurrently under one worker id. With Concurrency > 1, several shards are
+// processed in parallel on this machine, using the coordinator's own task
+// pipeline as the parallel unit.
 func (d *Daemon) RunForever() error {
+	if !d.registered {
+		if err := d.register(); err != nil {
+			return err
+		}
+	}
+	workers := d.config.Concurrency
+	if workers < 1 {
+		workers = 1
+	}
+	if workers == 1 {
+		return d.loop()
+	}
+	d.log.Info("agent running concurrently", "loops", workers)
+	var wg sync.WaitGroup
+	errs := make(chan error, workers)
+	for i := 0; i < workers; i++ {
+		wg.Add(1)
+		go func(loop int) {
+			defer wg.Done()
+			if err := d.loop(); err != nil {
+				errs <- err
+				return
+			}
+			errs <- nil
+		}(i)
+	}
+	wg.Wait()
+	close(errs)
+	for err := range errs {
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// loop is one claim→execute→upload cycle until interrupted, idle-exit, or
+// the shared max-tasks budget is consumed.
+func (d *Daemon) loop() error {
 	failures := 0
 	for {
-		if !d.registered {
-			if err := d.register(); err != nil {
-				return err
-			}
-		}
 		d.cleanupExpiredDirectories()
 		outcome, err := d.runOnce()
 		if err != nil {
@@ -63,8 +100,11 @@ func (d *Daemon) RunForever() error {
 		}
 		failures = 0
 		if outcome.Claimed && outcome.Completed {
+			d.mu.Lock()
 			d.completed++
-			if d.config.MaxTasks > 0 && d.completed >= d.config.MaxTasks {
+			done := d.config.MaxTasks > 0 && d.completed >= d.config.MaxTasks
+			d.mu.Unlock()
+			if done {
 				d.log.Info("max tasks reached")
 				return nil
 			}

coordinator/internal/agent/daemon_test.go

@@ -11,6 +11,7 @@ import (
 	"os"
 	"path/filepath"
 	"strings"
+	"sync"
 	"testing"
 	"time"
 )
@@ -55,6 +56,7 @@ type fakeCoordinator struct {
 	uploadSize int64
 	inputBytes []byte
 	conflict   bool // 409 on heartbeat/upload/result
+	mu         sync.Mutex
 }
 
 func newFakeCoordinator(t *testing.T, task map[string]any) *fakeCoordinator {
@@ -64,6 +66,8 @@ func newFakeCoordinator(t *testing.T, task map[string]any) *fakeCoordinator {
 	fake.uploadSize = int64(len(fake.inputBytes))
 	var server *httptest.Server
 	server = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		fake.mu.Lock()
+		defer fake.mu.Unlock()
 		switch {
 		case r.Method == http.MethodPost && r.URL.Path == "/workers/register":
 			writeJSON(w, http.StatusCreated, map[string]any{
@@ -294,3 +298,78 @@ func TestDaemonIdleClaimIsNotCompleted(t *testing.T) {
 		t.Fatalf("outcome = %+v", outcome)
 	}
 }
+
+func TestDaemonConcurrencyProcessesTasksInParallel(t *testing.T) {
+	t.Parallel()
+	marker := filepath.Join(t.TempDir(), "marker")
+	script := filepath.Join(t.TempDir(), "fake-runner.sh")
+	content := `#!/bin/sh
+out=""
+task_dir=""
+while [ "$#" -gt 0 ]; do
+  case "$1" in
+    --output) out="$2"; shift 2;;
+    --task-dir) task_dir="$2"; shift 2;;
+    *) shift;;
+  esac
+done
+echo start >> ` + marker + `
+sleep 1
+echo end >> ` + marker + `
+printf 'id,score\n1,1\n' > "$task_dir/result.csv"
+printf '{"artifact_path":"%s/result.csv","content_type":"text/csv","metrics":{"rows":1}}' "$task_dir" > "$out"
+exit 0
+`
+	if err := os.WriteFile(script, []byte(content), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	fake := newFakeCoordinator(t, validClaimedTaskPayload())
+	defer fake.close()
+	config := &Config{
+		CoordinatorURL: fake.server.URL,
+		WorkerName:     "concurrent-worker",
+		WorkerID:       "22222222-2222-4222-8222-222222222222",
+		WorkDir:        t.TempDir(),
+		CPUCount:       1,
+		PollInterval:   time.Millisecond,
+		RequestTimeout: 5 * time.Second,
+		Heartbeat:      15 * time.Second,
+		Capabilities:   []string{"similarity-search"},
+		TaskRunner:     []string{script},
+		MaxTasks:       3,
+		Concurrency:    3,
+	}
+	client := NewClient(fake.server.URL, &StaticToken{token: "test-token"}, 5*time.Second)
+	logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+	daemon := NewDaemon(config, client, NewTaskRunner(config.TaskRunner), logger)
+	if err := daemon.RunForever(); err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	raw, err := os.ReadFile(marker)
+	if err != nil {
+		t.Fatalf("marker: %v", err)
+	}
+	starts := strings.Count(string(raw), "start\n")
+	ends := strings.Count(string(raw), "end\n")
+	if starts < 3 || ends < 3 {
+		t.Fatalf("marker: %d starts / %d ends, want at least 3/3", starts, ends)
+	}
+	// With three loops sleeping 1s each, the marker proves all three ran
+	// concurrently (three starts before the first end completes a 1s sleep).
+	lines := strings.Split(strings.TrimSpace(string(raw)), "\n")
+	concurrent := 0
+	running := 0
+	for _, line := range lines {
+		if line == "start" {
+			running++
+			if running > concurrent {
+				concurrent = running
+			}
+		} else {
+			running--
+		}
+	}
+	if concurrent < 3 {
+		t.Errorf("max concurrent executions = %d, want 3", concurrent)
+	}
+}

coordinator/internal/agent/setupui/server.go

@@ -323,7 +323,7 @@ func (s *Server) ensureVenvTaskRunner() {
 		return
 	}
 	if venv := s.venvPython(); venv != "" {
-		file.TaskRunner = []string{venv, "-m", "scimesh.worker.task"}
+		file.TaskRunner = []string{venv, "-I", "-m", "scimesh.worker.task"}
 		if payload, err := json.MarshalIndent(file, "", "  "); err == nil {
 			_ = os.WriteFile(s.cfgPath, append(payload, '\n'), 0o600)
 		}
@@ -377,6 +377,7 @@ type saveConfigRequest struct {
 	WorkerName     string   `json:"worker_name"`
 	CPUCount       int      `json:"cpu_count"`
 	MemoryMB       int      `json:"memory_mb"`
+	Concurrency    int      `json:"concurrency"`
 	TaskRunner     []string `json:"task_runner"`
 }
 
@@ -395,6 +396,7 @@ func (s *Server) handleSaveConfig(w http.ResponseWriter, r *http.Request) {
 		WorkerName:     strings.TrimSpace(req.WorkerName),
 		CPUCount:       req.CPUCount,
 		MemoryMB:       req.MemoryMB,
+		Concurrency:    req.Concurrency,
 		TaskRunner:     req.TaskRunner,
 	}
 	if file.CoordinatorURL == "" {
@@ -418,12 +420,15 @@ func (s *Server) handleSaveConfig(w http.ResponseWriter, r *http.Request) {
 	if file.CPUCount < 1 {
 		file.CPUCount = 1
 	}
+	if file.Concurrency < 1 {
+		file.Concurrency = 1
+	}
 	// The wizard UI bakes the venv python into the runner after an install;
 	// an API-driven or scripted flow may not, so the server guarantees it:
 	// workloads execute through scimesh's task runner, which lives in the venv.
 	if len(file.TaskRunner) == 0 {
 		if venv := s.venvPython(); venv != "" {
-			file.TaskRunner = []string{venv, "-m", "scimesh.worker.task"}
+			file.TaskRunner = []string{venv, "-I", "-m", "scimesh.worker.task"}
 		}
 	}
 	if err := agent.SaveConfigFile(s.cfgPath, file); err != nil {
@@ -449,6 +454,9 @@ func (s *Server) handleTest(w http.ResponseWriter, r *http.Request) {
 	// checking the bare system python3 would keep reporting scimesh as
 	// missing even though the worker would run with the venv.
 	report := agent.RunCheck(r.Context(), url, s.venvPython(), req.Token, req.WorkerKey, req.UserserviceURL)
+	if report.Scimesh.OK {
+		report.Scimesh = ensureMatchingScimeshVersion(report.Scimesh)
+	}
 	writeJSON(w, http.StatusOK, report)
 }
 
@@ -623,3 +631,26 @@ func truncate(s string, n int) string {
 	}
 	return s[:n] + "…"
 }
+
+// ensureMatchingScimeshVersion flips a green scimesh check to a stale one when
+// the installed package does not match the worker-agent's own version: a
+// version-locked wheel is the only supported runtime, and a mismatch means the
+// workload catalog the worker advertises is not what it executes. The wizard
+// UI then offers the Install button again. Dev builds have no release wheel,
+// so they skip the comparison.
+func ensureMatchingScimeshVersion(item agent.CheckItem) agent.CheckItem {
+	if agent.Version == "" || agent.Version == "dev" {
+		return item
+	}
+	want := agent.NormalizePEP440(agent.Version)
+	got := strings.TrimSpace(item.Detail)
+	if got == "" || got == want {
+		return item
+	}
+	item.OK = false
+	item.Detail = fmt.Sprintf(
+		"installed scimesh %s, but this worker-agent (%s) needs %s — press Install to upgrade",
+		got, agent.Version, want,
+	)
+	return item
+}

coordinator/internal/agent/setupui/server_test.go

@@ -429,8 +429,8 @@ func TestStartPinsTheVenvTaskRunner(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if len(config.TaskRunner) != 3 || config.TaskRunner[0] != venvPython || config.TaskRunner[1] != "-m" || config.TaskRunner[2] != "scimesh.worker.task" {
-		t.Errorf("task runner = %v, want the venv python runner", config.TaskRunner)
+	if len(config.TaskRunner) != 4 || config.TaskRunner[0] != venvPython || config.TaskRunner[1] != "-I" || config.TaskRunner[2] != "-m" || config.TaskRunner[3] != "scimesh.worker.task" {
+		t.Errorf("task runner = %v, want the venv python runner with -I", config.TaskRunner)
 	}
 }
 
@@ -450,8 +450,8 @@ func TestSaveConfigPinsVenvRunnerWhenPresent(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if len(config.TaskRunner) != 3 || config.TaskRunner[0] != venvPython {
-		t.Errorf("task runner = %v, want the venv python", config.TaskRunner)
+	if len(config.TaskRunner) != 4 || config.TaskRunner[0] != venvPython || config.TaskRunner[1] != "-I" {
+		t.Errorf("task runner = %v, want the venv python with -I", config.TaskRunner)
 	}
 }
 
@@ -462,7 +462,7 @@ func TestTestProbesTheVenvPythonAfterInstall(t *testing.T) {
 	// a fake scimesh version so the preflight goes green through the venv.
 	venvPython := filepath.Join(server.dir, "venv", "bin", "python")
 	_ = os.MkdirAll(filepath.Dir(venvPython), 0o755)
-	_ = os.WriteFile(venvPython, []byte("#!/bin/sh\nif [ \"$1\" = \"-c\" ]; then echo 9.9.9-test; exit 0; fi\nexit 0\n"), 0o755)
+	_ = os.WriteFile(venvPython, []byte("#!/bin/sh\nfor a in \"$@\"; do if [ \"$a\" = \"-c\" ]; then echo 9.9.9-test; exit 0; fi; done\nexit 0\n"), 0o755)
 
 	req, _ := http.NewRequestWithContext(context.Background(), http.MethodPost, base+"/api/test", strings.NewReader(`{"coordinator_url":"http://127.0.0.1:1"}`))
 	req.Header.Set("Content-Type", "application/json")
@@ -497,3 +497,26 @@ time=6 level=WARN msg="agent cycle failed" error="boom"
 		t.Errorf("stats = %+v, want registered claimed=2 completed=1 failed=1", stats)
 	}
 }
+
+func testCheckScimeshVersion(t *testing.T, installed, binary string, wantOK bool, wantDetail string) {
+	t.Helper()
+	old := agent.Version
+	agent.Version = binary
+	t.Cleanup(func() { agent.Version = old })
+	item := ensureMatchingScimeshVersion(agent.CheckItem{Name: "scimesh", OK: true, Detail: installed})
+	if item.OK != wantOK {
+		t.Errorf("installed=%s binary=%s: ok=%v, want %v (%s)", installed, binary, item.OK, wantOK, item.Detail)
+	}
+	if wantDetail != "" && !strings.Contains(item.Detail, wantDetail) {
+		t.Errorf("detail = %q, want it to contain %q", item.Detail, wantDetail)
+	}
+}
+
+func TestEnsureMatchingScimeshVersion(t *testing.T) {
+	testCheckScimeshVersion(t, "1.1.0a20", "1.1.0-alpha.20", true, "")
+	testCheckScimeshVersion(t, "1.1.0a17", "1.1.0-alpha.20", false, "press Install to upgrade")
+	testCheckScimeshVersion(t, "1.1.0a16.dev7+gea0fb8c59.d20260803", "1.1.0-alpha.20", false, "needs 1.1.0a20")
+	// Dev builds and unknown versions never block.
+	testCheckScimeshVersion(t, "anything", "dev", true, "")
+	testCheckScimeshVersion(t, "1.1.0a20", "", true, "")
+}

coordinator/internal/agent/setupui/template.html

@@ -141,6 +141,7 @@ code{font-family:var(--mono);font-size:.86em}
           </div>
         </div>
         <div class="field" id="cpu-field" style="display:none"><label>CPU count</label><input id="in-cpu" type="number" min="1" value="1"></div>
+        <div class="field"><label>Concurrent task loops</label><input id="in-conc" type="number" min="1" max="64" value="1"><p class="hint">Process this many shards in parallel on this machine. Each loop runs its own task runner subprocess.</p></div>
         <div class="actions"><button class="btn btn-ghost" id="b2b">← Back</button><button class="btn btn-primary" id="b2">Continue →</button></div>
       </div>
     </div>
@@ -228,9 +229,10 @@ function draftConfig(){
     userservice_url:state.mode==='key'?$('in-users').value.trim():'',
     work_dir:$('in-dir').value.trim(),
     worker_name:$('in-name').value.trim(),
-    cpu_count:state.cpu==='custom'?parseInt($('in-cpu').value||'1',10):0
+    cpu_count:state.cpu==='custom'?parseInt($('in-cpu').value||'1',10):0,
+    concurrency:parseInt($('in-conc').value||'1',10)
   };
-  if(state.venvPython)cfg.task_runner=[state.venvPython,'-m','scimesh.worker.task'];
+  if(state.venvPython)cfg.task_runner=[state.venvPython,'-I','-m','scimesh.worker.task'];
   return cfg;
 }