Drop a leftover helper

.github/workflows/release.yml

@@ -110,6 +110,8 @@ jobs:
             artifacts/*
             install.sh
             install.ps1
+            uninstall.sh
+            uninstall.ps1
           # Pre-release tags (e.g. v1.1.0-alpha.1) publish as pre-releases.
           prerelease: ${{ contains(github.ref_name, '-alpha') }}
           generate_release_notes: true

.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.
@@ -59,8 +75,18 @@
 - ✅ На машине пользователя: визард alpha.15, правильный токен, venv из wheel, воркер emil-pc online, 15 пустых воркеров вычищены из БД.
 - ✅ Гейт: race + lint + pytest 208.
 
+## Progress (день 2: контроль версий + изоляция)
+- ✅ Визард сверяет установленную версию scimesh с версией бинарника: preflight показывает «installed X, worker needs Y — press Install» и снова предлагает кнопку Install (ensureMatchingScimeshVersion, dev-сборки пропускают). Проверено на машине: a21 vs a22 → красный → Install → зелёный 1.1.0a22.
+- ✅ Все проверки и task-раннеры запускаются с `-I` (isolated): каталог запуска больше не перекрывает venv (был артефакт cwd-затенения исходниками). Затронуты: check, дефолтные раннеры агента, визард (config/save/start), serve, agent_cmd.
+- ✅ Релиз v1.1.0-alpha.22; полный гейт (race + lint 0 issues + pytest 213); бинарник юзера обновлён, рантайм переустановлен (a22), воркер с concurrency=4; большой джоб юзера (1934 шарда) завершён (очередь 0).
+
 ## Completion
 COMPLETED — ночной план выполнен полностью (10 пунктов + 2 найденных бага, включая E2E quorum на релизном коде). Все гейты зелёные, релиз v1.1.0-alpha.16 опубликован.
 
+## Progress (день 2: контроль версий + изоляция)
+- ✅ Визард сверяет установленную версию scimesh с версией бинарника: preflight показывает «installed X, worker needs Y — press Install» и снова предлагает кнопку Install (ensureMatchingScimeshVersion, dev-сборки пропускают). Проверено на машине: a21 vs a22 → красный → Install → зелёный 1.1.0a22.
+- ✅ Все проверки и task-раннеры запускаются с `-I` (isolated): каталог запуска больше не перекрывает venv (был артефакт cwd-затенения исходниками). Затронуты: check, дефолтные раннеры агента, визард (config/save/start), serve, agent_cmd.
+- ✅ Релиз v1.1.0-alpha.22; полный гейт (race + lint 0 issues + pytest 213); бинарник юзера обновлён, рантайм переустановлен (a22), воркер с concurrency=4; большой джоб юзера (1934 шарда) завершён (очередь 0).
+
 ## Completion (предыдущая задача)
 COMPLETED — пайплайн доведён до рабочего состояния и проверен на релизных артефактах v1.1.0-alpha.14.

README.md

@@ -61,7 +61,16 @@ powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/emi
 ```
 
 Set `SCIMESH_AUTO_START=0` to install without starting anything. The old demo
-control room was removed: `/ui` is the admin console. A standalone
+control room was removed: `/ui` is the admin console.
+
+To remove a component, run the matching uninstaller (data is kept unless you
+pass `--purge`):
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/emil28092005/SciMesh/main/uninstall.sh | bash -s coordinator
+curl -fsSL https://raw.githubusercontent.com/emil28092005/SciMesh/main/uninstall.sh | bash -s worker --purge
+# Windows: irm .../uninstall.ps1 | iex  (set $env:SCIMESH_COMPONENT, -Purge deletes data)
+``` A standalone
 worker is installed the same way (`bash -s worker`, or
 `SCIMESH_COMPONENT=worker` on Windows); its installer opens the local setup
 wizard (`worker-agent setup`) in the browser automatically.

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

@@ -10,6 +10,7 @@ import (
 	"flag"
 	"fmt"
 	"log/slog"
+	"net"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -96,18 +97,16 @@ func runServe(args []string) error {
 	defer func() { _ = closeUsers() }()
 
 	// 5. Local worker agents before the server, so they can claim immediately.
-	coordinatorURL := "http://" + *addr
-	agents, err := spawnAgents(ctx, log, *dataDir, *workers, coordinatorURL, workerToken, venvPython)
+	// They always dial the loopback address: an --addr of 0.0.0.0 is not a
+	// connectable target from the same host.
+	agentURL, resolvedPublic := serveURLs(*addr, *publicURL)
+	agents, err := spawnAgents(ctx, log, *dataDir, *workers, agentURL, workerToken, venvPython)
 	if err != nil {
 		return err
 	}
 	defer stopAgents(agents)
 
 	// 6. The coordinator server itself.
-	coordinatorPublicURL := *publicURL
-	if coordinatorPublicURL == "" {
-		coordinatorPublicURL = "http://" + *addr
-	}
 	cfg := infra.Config{
 		Addr:                 *addr,
 		DatabaseEngine:       "sqlite",
@@ -115,8 +114,10 @@ func runServe(args []string) error {
 		Token:                workerToken,
 		JWTSecret:            jwtSecret,
 		UserserviceURL:       "http://" + usersAddr,
-		PublicCoordinatorURL: coordinatorPublicURL,
-		PublicUserserviceURL: "http://" + usersAddr,
+		PublicCoordinatorURL: resolvedPublic,
+		// The exchange is fronted by the coordinator's own proxy, so the UI
+		// falls back to the coordinator origin for USERSERVICE_URL.
+		PublicUserserviceURL: "",
 		LogLevel:             "info",
 		StorageDir:           filepath.Join(*dataDir, "artifacts"),
 		DocsDir:              *docsDir,
@@ -132,12 +133,13 @@ func runServe(args []string) error {
 		WorkerOfflineAfter:   1 * time.Minute,
 		AutoMigrate:          true,
 	}
+	browserURL, _ := serveURLs(*addr, *publicURL)
 	if *open {
-		openBrowser("http://" + *addr + "/ui/admin")
+		openBrowser(browserURL + "/ui/admin")
 	}
 
 	// Print the login once the server is about to start.
-	fmt.Printf("\nSciMesh is starting at http://%s/ui\n", *addr)
+	fmt.Printf("\nSciMesh is starting at %s/ui\n", browserURL)
 	fmt.Printf("  admin login: %s / %s\n", *email, *password)
 	if runtimeStatus(venvPython) {
 		fmt.Printf("  scientific runtime: ready (%s)\n", venvPython)
@@ -236,9 +238,9 @@ 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
@@ -337,3 +339,33 @@ func openBrowser(target string) {
 	// #nosec G204 -- target is the local UI URL the operator asked to open.
 	_ = exec.CommandContext(context.Background(), command, target).Start()
 }
+
+// serveURLs derives the two addresses of a serve instance from the listen
+// address and the optional --public-url flag:
+//
+//   - the agent URL is always the loopback form of the port, because spawned
+//     local workers share the host and 0.0.0.0 is not connectable from it;
+//   - the public URL is what browsers and remote workers are told. An explicit
+//     --public-url wins; a listen host that is a real address is used as-is;
+//     a wildcard host (0.0.0.0, ::, or empty) yields an empty public URL, so
+//     the UI falls back to the browser's own origin (the coordinator's LAN
+//     address as the browser sees it).
+func serveURLs(addr, publicURL string) (agentURL, resolvedPublic string) {
+	host, port, err := net.SplitHostPort(addr)
+	if err != nil {
+		// No port in the listen address: assume the default and treat the
+		// whole string as a host (e.g. a bare wildcard).
+		host, port = addr, "8080"
+	}
+	agentURL = "http://127.0.0.1:" + port
+	if publicURL != "" {
+		return agentURL, publicURL
+	}
+	host = strings.Trim(host, "[]")
+	switch host {
+	case "", "0.0.0.0", "::":
+		return agentURL, ""
+	default:
+		return agentURL, "http://" + addr
+	}
+}

coordinator/cmd/coordinator/serve_cmd_test.go

@@ -0,0 +1,22 @@
+package main
+
+import "testing"
+
+func TestServeURLs(t *testing.T) {
+	cases := []struct {
+		addr, public, agent, resolved string
+	}{
+		{"127.0.0.1:8080", "", "http://127.0.0.1:8080", "http://127.0.0.1:8080"},
+		{"0.0.0.0:8080", "", "http://127.0.0.1:8080", ""},
+		{":8080", "", "http://127.0.0.1:8080", ""},
+		{"::", "", "http://127.0.0.1:8080", ""},
+		{"192.168.1.10:8080", "", "http://127.0.0.1:8080", "http://192.168.1.10:8080"},
+		{"0.0.0.0:8080", "http://cluster.example:8080", "http://127.0.0.1:8080", "http://cluster.example:8080"},
+	}
+	for _, c := range cases {
+		agent, resolved := serveURLs(c.addr, c.public)
+		if agent != c.agent || resolved != c.resolved {
+			t.Errorf("serveURLs(%q, %q) = (%q, %q), want (%q, %q)", c.addr, c.public, agent, resolved, c.agent, c.resolved)
+		}
+	}
+}

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/client.go

@@ -48,6 +48,12 @@ type Client struct {
 }
 
 func NewClient(baseURL string, tokens TokenProvider, timeout time.Duration) *Client {
+	// Payload transfers get a more generous budget than control calls: a large
+	// shard over a slow link easily outlives the API timeout.
+	transferTimeout := timeout * 4
+	if transferTimeout < 2*time.Minute {
+		transferTimeout = 2 * time.Minute
+	}
 	return &Client{
 		baseURL: strings.TrimRight(baseURL, "/"),
 		tokens:  tokens,
@@ -57,7 +63,7 @@ func NewClient(baseURL string, tokens TokenProvider, timeout time.Duration) *Cli
 			CheckRedirect: func(*http.Request, []*http.Request) error { return http.ErrUseLastResponse },
 		},
 		dlClient: &http.Client{
-			Timeout: timeout,
+			Timeout: transferTimeout,
 			CheckRedirect: func(req *http.Request, via []*http.Request) error {
 				if len(via) >= 10 {
 					return fmt.Errorf("too many redirects")

coordinator/internal/agent/client_test.go

@@ -223,3 +223,17 @@ func sha256Of(t *testing.T, value string) string {
 	digest := sha256.Sum256([]byte(value))
 	return fmt.Sprintf("%x", digest)
 }
+
+func TestNewClientTransferTimeoutExceedsAPITimeout(t *testing.T) {
+	c := NewClient("http://coord:8080", &StaticToken{token: "t"}, 30*time.Second)
+	if c.apiClient.Timeout != 30*time.Second {
+		t.Errorf("api timeout = %v, want 30s", c.apiClient.Timeout)
+	}
+	if c.dlClient.Timeout < 2*time.Minute {
+		t.Errorf("transfer timeout = %v, want at least 2m", c.dlClient.Timeout)
+	}
+	short := NewClient("http://coord:8080", &StaticToken{token: "t"}, 3*time.Minute)
+	if short.dlClient.Timeout != 12*time.Minute {
+		t.Errorf("transfer timeout = %v, want 4x the api timeout", short.dlClient.Timeout)
+	}
+}

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;
 }
 

uninstall.ps1

@@ -0,0 +1,66 @@
+# SciMesh uninstaller for Windows: stops the running component, removes its
+# binary, and — when asked — deletes its data directory.
+#
+#   powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/emil28092005/SciMesh/main/uninstall.ps1 | iex"
+#
+# $env:SCIMESH_COMPONENT selects the component (coordinator | worker | all,
+# default all). Data is kept unless $env:SCIMESH_PURGE -eq "1" or the
+# -Purge switch is passed (in a pipe there is no interactive prompt).
+param(
+    [switch]$Purge
+)
+$ErrorActionPreference = "Stop"
+
+$Component = if ($env:SCIMESH_COMPONENT) { $env:SCIMESH_COMPONENT } else { "all" }
+if ($Purge -or $env:SCIMESH_PURGE -eq "1") { $Purge = $true } else { $Purge = $false }
+
+$InstallDir = if ($env:SCIMESH_INSTALL_DIR) {
+    $env:SCIMESH_INSTALL_DIR
+} else {
+    Join-Path $env:LOCALAPPDATA "SciMesh"
+}
+
+function Remove-Component {
+    param([string]$Name, [string]$Binary)
+    Write-Host "Stopping $Name…"
+    Get-Process -Name $Name -ErrorAction SilentlyContinue | Stop-Process -Force -ErrorAction SilentlyContinue
+    $target = Join-Path $InstallDir $Binary
+    if (Test-Path $target) {
+        Write-Host "Removing $target…"
+        Remove-Item -Force $target
+    }
+}
+
+function Remove-Data {
+    param([string]$Dir, [string]$Label)
+    if (-not (Test-Path $Dir)) { return }
+    if ($Purge) {
+        Remove-Item -Recurse -Force $Dir
+        Write-Host "Deleted $Dir"
+        return
+    }
+    # No interactive prompt in a pipe: keep the data by default.
+    Write-Host "Keeping $Dir ($Label). Pass -Purge to delete it."
+}
+
+switch ($Component) {
+    "coordinator" {
+        Remove-Component -Name "coordinator" -Binary "coordinator.exe"
+        Remove-Data -Dir (Join-Path $HOME ".scimesh") -Label "secrets, databases, artifacts, users"
+    }
+    "worker" {
+        Remove-Component -Name "worker-agent" -Binary "worker-agent.exe"
+        Remove-Data -Dir (Join-Path $HOME ".scimesh-worker") -Label "worker config, runtime venv, logs"
+    }
+    "all" {
+        Remove-Component -Name "coordinator" -Binary "coordinator.exe"
+        Remove-Component -Name "worker-agent" -Binary "worker-agent.exe"
+        Remove-Data -Dir (Join-Path $HOME ".scimesh") -Label "secrets, databases, artifacts, users"
+        Remove-Data -Dir (Join-Path $HOME ".scimesh-worker") -Label "worker config, runtime venv, logs"
+    }
+    default { throw "unknown component: $Component (use 'coordinator', 'worker' or 'all')" }
+}
+
+Write-Host ""
+Write-Host "SciMesh $Component uninstalled."
+Write-Host "Pass -Purge to also delete the data directories."

uninstall.sh

@@ -0,0 +1,83 @@
+#!/usr/bin/env bash
+# SciMesh uninstaller: stops the running component, removes its binary, and —
+# when asked — deletes its data directory.
+#
+#   curl -fsSL https://raw.githubusercontent.com/emil28092005/SciMesh/main/uninstall.sh | bash -s coordinator
+#   curl -fsSL https://raw.githubusercontent.com/emil28092005/SciMesh/main/uninstall.sh | bash -s worker
+#   curl -fsSL https://raw.githubusercontent.com/emil28092005/SciMesh/main/uninstall.sh         # both
+#
+# Data (jobs, artifacts, users, secrets, the managed venv) is kept by default.
+# Pass --purge (or set SCIMESH_PURGE=1) to delete it without asking; without
+# it the script prompts interactively. Because `curl | bash` pipes have no
+# interactive stdin, the default is always "keep data".
+set -eu
+
+COMPONENT="${1:-all}"
+PURGE=0
+for arg in "$@"; do
+  case "$arg" in
+    --purge) PURGE=1 ;;
+  esac
+done
+[ "${SCIMESH_PURGE:-0}" = "1" ] && PURGE=1
+
+INSTALL_DIR="${SCIMESH_INSTALL_DIR:-$HOME/.local/bin}"
+
+remove_component() {
+  case "$1" in
+    coordinator)
+      echo "Stopping the coordinator (serve)…"
+      pkill -x coordinator 2>/dev/null || true
+      echo "Removing $INSTALL_DIR/coordinator…"
+      rm -f "$INSTALL_DIR/coordinator"
+      ;;
+    worker)
+      echo "Stopping the worker agent and its setup wizard…"
+      pkill -x worker-agent 2>/dev/null || true
+      echo "Removing $INSTALL_DIR/worker-agent…"
+      rm -f "$INSTALL_DIR/worker-agent"
+      ;;
+    *) echo "unknown component: $1 (use 'coordinator', 'worker' or 'all')" >&2; exit 1 ;;
+  esac
+}
+
+remove_data() {
+  local dir="$1" label="$2"
+  if [ ! -d "$dir" ]; then
+    return 0
+  fi
+  if [ "$PURGE" = "1" ]; then
+    rm -rf "$dir"
+    echo "Deleted $dir"
+    return 0
+  fi
+  # In a pipe (curl | bash) stdin is exhausted, so the prompt defaults to keep.
+  printf "Delete %s (%s)? [y/N] " "$dir" "$label"
+  read -r answer || answer=""
+  case "$answer" in
+    y|Y|yes|YES) rm -rf "$dir"; echo "Deleted $dir" ;;
+    *) echo "Keeping $dir" ;;
+  esac
+}
+
+case "$COMPONENT" in
+  coordinator)
+    remove_component coordinator
+    remove_data "$HOME/.scimesh" "secrets, databases, artifacts, users, managed venv"
+    ;;
+  worker)
+    remove_component worker
+    remove_data "$HOME/.scimesh-worker" "worker config, runtime venv, logs"
+    ;;
+  all)
+    remove_component coordinator
+    remove_component worker
+    remove_data "$HOME/.scimesh" "secrets, databases, artifacts, users, managed venv"
+    remove_data "$HOME/.scimesh-worker" "worker config, runtime venv, logs"
+    ;;
+  *) echo "unknown component: $COMPONENT (use 'coordinator', 'worker' or 'all')" >&2; exit 1 ;;
+esac
+
+echo
+echo "SciMesh $COMPONENT uninstalled."
+echo "Pass --purge to also delete the data directories without asking."