Add warn auto-ban for repeated admin /warn commands

CLAUDE.md

@@ -90,6 +90,16 @@
 - Deletion errors are logged but don't fail the operation (messages might be already deleted or too old)
 - Design principle: When a spammer is detected, aggressively clean up ALL their spam messages, not just the triggering one
 
+### Warn Auto-Ban
+- `/warn` admin command optionally records each warning to `warnings` table and bans the user when count within `warn.window` reaches `warn.threshold`
+- Disabled by default (`warn.threshold=0`); enabling preserves the existing warn message + delete behavior and adds the count/ban path on top
+- Threshold semantics match `Report.AutoBanThreshold`: `count >= threshold` triggers ban (so threshold=2 bans on the 2nd warn)
+- Storage layer (`app/storage/warnings.go`) opportunistically prunes rows older than 1 year on each `Add`; the 1y cap is a storage bound, NOT the configured window — `CountWithin` enforces the window at query time
+- Ban does NOT delete the warning rows: subsequent `/warn` on an already-banned user re-triggers the ban path. Telegram's `BanChatMemberConfig` is idempotent, so the repeat ban is a no-op API call but produces a fresh admin-chat notification (audit visibility for repeat offenders)
+- Warn auto-ban does NOT update spam samples (`bot.UpdateSpam` is not called) — warnings reflect admin policy, not spam content
+- `executeWarnBan` mirrors `executeAutoBan` for dry/training/soft-ban handling but does not share an abstraction (the two diverge on spam-sample updates and on `From` vs `SenderChat` resolution)
+- Settings: only `Warn.Threshold` is in `zeroAwarePaths` (0=disabled, must survive merges); `Warn.Window` zero is invalid and rejected by startup validation
+
 ### LLM Checker Structure
 - Shared provider-agnostic LLM flow lives in `lib/tgspam/llm.go`
 - Keep provider-specific transport and request construction in `lib/tgspam/openai.go`, `lib/tgspam/gemini.go`, etc

README.md

@@ -388,6 +388,27 @@ The reporting system includes rate limiting to prevent abuse. Each user can subm
 
 All reports are stored in the database for audit purposes and can help identify patterns of spam or abuse over time.
 
+### Warn-Driven Auto-Ban
+
+The admin `/warn` command can optionally escalate to an automatic ban once a user has accumulated enough warnings within a sliding time window. This complements `--report.auto-ban-threshold` (which aggregates user `/report` submissions for one message) by tracking admin warnings per user across messages.
+
+The feature is disabled by default. To enable it, set `--warn.threshold=, [$WARN_THRESHOLD]` to a positive number and adjust `--warn.window=, [$WARN_WINDOW]` (default: `720h`). When enabled:
+
+1. Each `/warn` issued by an admin is recorded in the `warnings` table together with the user/channel id and timestamp
+2. After recording the warning, the bot counts how many warnings the same user has received within the configured window
+3. If the count reaches `--warn.threshold=` (i.e. `count >= threshold`), the bot bans the user immediately, respecting `--training`, `--dry`, and `--soft-ban` modes
+4. A notification is posted to the admin chat in the form `**warn auto-banned** @user (12345) after N warns within <window>`. The verb adapts to the active mode: `auto-would have banned` in dry mode, `auto-would have banned (training)` in training mode, and `auto-restricted` for users in soft-ban mode (channels still fall through to `auto-banned` because Telegram has no restrict variant for channel senders).
+
+Example: `--warn.threshold=3 --warn.window=168h` bans a user once they accumulate three warnings within a week (the third warn triggers the ban).
+
+Notes:
+
+- The default `--warn.threshold=0` preserves the original `/warn` behavior exactly: a warning message is posted and the offending message is deleted, but no warning is recorded and no auto-ban is performed.
+- Warnings issued before the window expires are counted; older rows are pruned opportunistically by the storage layer. The storage retention is capped at one year, so configuring `--warn.window` beyond `8760h` is not supported.
+- Unlike `/spam`, `/warn` does not update spam samples — warnings reflect admin policy, not spam content.
+- Repeat bans are intentional: if an already-banned user is warned again, the threshold check fires again and re-bans them. Telegram treats banning an already-banned user as a no-op, so this is safe and serves as audit visibility for repeat offenders.
+- Toggling `--warn.threshold` from `0` to a positive value (or vice versa) requires a process restart: the warnings storage is wired only at startup. Runtime changes via the settings UI are persisted but take effect only after the next restart.
+
 ### Lua Plugins Support
 
 TG-Spam supports custom spam detection through Lua plugins. This allows users to extend the spam detection capabilities without modifying the Go codebase.
@@ -647,6 +668,10 @@ report:
       --report.rate-limit=              max reports per user per period (default: 10) [$REPORT_RATE_LIMIT]
       --report.rate-period=             rate limit time period (default: 1h) [$REPORT_RATE_PERIOD]
 
+warn:
+      --warn.threshold=                 auto-ban after N warns within window (0=disabled) (default: 0) [$WARN_THRESHOLD]
+      --warn.window=                    sliding window for counting warns (default: 720h) [$WARN_WINDOW]
+
 files:
       --files.samples=                  samples data path, defaults to dynamic data path [$FILES_SAMPLES]
       --files.dynamic=                  dynamic data path (default: data) [$FILES_DYNAMIC]

app/config/settings.go

@@ -33,6 +33,7 @@ type Settings struct {
 	Duplicates    DuplicatesSettings    `json:"duplicates" yaml:"duplicates" db:"duplicates"`
 	Reactions     ReactionsSettings     `json:"reactions" yaml:"reactions" db:"reactions"`
 	Report        ReportSettings        `json:"report" yaml:"report" db:"report"`
+	Warn          WarnSettings          `json:"warn" yaml:"warn" db:"warn"`
 
 	// spam detection settings
 	SimilarityThreshold float64 `json:"similarity_threshold" yaml:"similarity_threshold" db:"similarity_threshold"`
@@ -181,6 +182,12 @@ type ReportSettings struct {
 	RatePeriod       time.Duration `json:"rate_period" yaml:"rate_period" db:"report_rate_period"`
 }
 
+// WarnSettings contains admin /warn auto-ban settings
+type WarnSettings struct {
+	Threshold int           `json:"threshold" yaml:"threshold" db:"warn_threshold"`
+	Window    time.Duration `json:"window" yaml:"window" db:"warn_window"`
+}
+
 // LuaPluginsSettings contains Lua plugins settings
 type LuaPluginsSettings struct {
 	Enabled        bool     `json:"enabled" yaml:"enabled" db:"lua_plugins_enabled"`
@@ -238,7 +245,7 @@ type TransientSettings struct {
 	// temporary auth password (used only to generate hash)
 	WebAuthPasswd string `json:"-" yaml:"-"`
 
-	// AuthFromCLI marks web auth (hash or password) as held in memory rather
+	// authFromCLI marks web auth (hash or password) as held in memory rather
 	// than authoritative in the database. It is set by applyCLIOverrides for
 	// explicit --server.auth/--server.auth-hash overrides, and by
 	// applyAutoAuthFallback for the auto-generated password safety net. When
@@ -303,6 +310,7 @@ var zeroAwarePaths = map[string]bool{
 	"Duplicates.Threshold":    true, // app/main.go:717 (> 0): 0 disables
 	"Report.AutoBanThreshold": true, // app/main.go:336, app/events/reports.go:191 (> 0): 0 disables
 	"Report.RateLimit":        true, // app/events/reports.go:154 (<= 0): 0 disables rate limiting
+	"Warn.Threshold":          true, // app/main.go, app/events/admin.go (> 0): 0 disables
 	"OpenAI.HistorySize":      true, // lib/tgspam/detector.go:409 (> 0): 0 disables history
 	"Gemini.HistorySize":      true, // lib/tgspam/detector.go:409 (> 0): 0 disables history
 	"FirstMessagesCount":      true, // app/main.go:703, lib/tgspam/detector.go:205,208 (> 0): 0 disables

app/config/settings_test.go

@@ -220,6 +220,9 @@ func newPopulatedSettings() *Settings {
 	s.Report.RateLimit = 5
 	s.Report.RatePeriod = 90 * time.Second
 
+	s.Warn.Threshold = 3
+	s.Warn.Window = 12 * time.Hour
+
 	s.AggressiveCleanup = true
 	s.AggressiveCleanupLimit = 50
 
@@ -239,6 +242,7 @@ func TestSettings_JSONRoundTrip_NewGroups(t *testing.T) {
 	assert.Contains(t, jsonStr, `"duplicates"`)
 	assert.Contains(t, jsonStr, `"reactions"`)
 	assert.Contains(t, jsonStr, `"report"`)
+	assert.Contains(t, jsonStr, `"warn"`)
 	assert.Contains(t, jsonStr, `"aggressive_cleanup"`)
 	assert.Contains(t, jsonStr, `"aggressive_cleanup_limit"`)
 	assert.Contains(t, jsonStr, `"contact_only"`)
@@ -253,6 +257,7 @@ func TestSettings_JSONRoundTrip_NewGroups(t *testing.T) {
 	assert.Equal(t, original.Duplicates, restored.Duplicates)
 	assert.Equal(t, original.Reactions, restored.Reactions)
 	assert.Equal(t, original.Report, restored.Report)
+	assert.Equal(t, original.Warn, restored.Warn)
 	assert.Equal(t, original.AggressiveCleanup, restored.AggressiveCleanup)
 	assert.Equal(t, original.AggressiveCleanupLimit, restored.AggressiveCleanupLimit)
 	assert.True(t, restored.Meta.ContactOnly)
@@ -272,6 +277,7 @@ func TestSettings_YAMLRoundTrip_NewGroups(t *testing.T) {
 	assert.Contains(t, yamlStr, "duplicates:")
 	assert.Contains(t, yamlStr, "reactions:")
 	assert.Contains(t, yamlStr, "report:")
+	assert.Contains(t, yamlStr, "warn:")
 	assert.Contains(t, yamlStr, "aggressive_cleanup:")
 	assert.Contains(t, yamlStr, "aggressive_cleanup_limit:")
 	assert.Contains(t, yamlStr, "contact_only:")
@@ -286,6 +292,7 @@ func TestSettings_YAMLRoundTrip_NewGroups(t *testing.T) {
 	assert.Equal(t, original.Duplicates, restored.Duplicates)
 	assert.Equal(t, original.Reactions, restored.Reactions)
 	assert.Equal(t, original.Report, restored.Report)
+	assert.Equal(t, original.Warn, restored.Warn)
 	assert.Equal(t, original.AggressiveCleanup, restored.AggressiveCleanup)
 	assert.Equal(t, original.AggressiveCleanupLimit, restored.AggressiveCleanupLimit)
 	assert.True(t, restored.Meta.ContactOnly)
@@ -396,6 +403,14 @@ func TestSettings_ApplyDefaults_SkipsZeroAware(t *testing.T) {
 			},
 			assertFn: func(t *testing.T, target *Settings) { assert.Equal(t, 0, target.Report.AutoBanThreshold) },
 		},
+		{
+			name: "Warn.Threshold",
+			setup: func(target, template *Settings) {
+				target.Warn.Threshold = 0
+				template.Warn.Threshold = 3
+			},
+			assertFn: func(t *testing.T, target *Settings) { assert.Equal(t, 0, target.Warn.Threshold) },
+		},
 		{
 			name: "Report.RateLimit",
 			setup: func(target, template *Settings) {
@@ -457,6 +472,25 @@ func TestSettings_ApplyDefaults_SkipsZeroAware(t *testing.T) {
 	}
 }
 
+func TestSettings_ApplyDefaults_FillsWarnWindow(t *testing.T) {
+	// warn.Window is NOT zero-aware: zero is invalid (caught by startup
+	// validation), so the regular merge semantics must fill it from the
+	// template when the persisted blob has zero.
+	target := New()
+	target.Warn.Threshold = 5 // non-zero (preserved either way)
+	template := &Settings{
+		Warn: WarnSettings{
+			Threshold: 1,
+			Window:    720 * time.Hour,
+		},
+	}
+
+	target.ApplyDefaults(template)
+
+	assert.Equal(t, 5, target.Warn.Threshold, "non-zero target threshold preserved")
+	assert.Equal(t, 720*time.Hour, target.Warn.Window, "zero target window filled from template")
+}
+
 func TestSettings_ApplyDefaults_NestedStructs(t *testing.T) {
 	target := New()
 	template := &Settings{

app/events/admin.go

@@ -16,6 +16,14 @@ import (
 	"github.com/umputun/tg-spam/app/bot"
 )
 
+//go:generate moq --out mocks/warnings.go --pkg mocks --with-resets --skip-ensure . Warnings
+
+// Warnings is an interface for admin /warn records storage used by the warn auto-ban feature
+type Warnings interface {
+	Add(ctx context.Context, userID int64, userName string) error
+	CountWithin(ctx context.Context, userID int64, window time.Duration) (int, error)
+}
+
 // admin is a helper to handle all admin-group related stuff, created by listener
 // public methods kept public (on a private struct) to be able to recognize the api
 type admin struct {
@@ -31,6 +39,9 @@ type admin struct {
 	warnMsg                string
 	aggressiveCleanup      bool
 	aggressiveCleanupLimit int
+	warnings               Warnings      // storage for /warn records, used by DirectWarnReport auto-ban path
+	warnThreshold          int           // auto-ban after N /warn within warnWindow (0 disables auto-ban)
+	warnWindow             time.Duration // sliding window for counting warns
 }
 
 const (
@@ -343,27 +354,142 @@ func (a *admin) DirectWarnReport(update tbapi.Update) error {
 	}
 
 	// make a warning message and replay to origMsg.MessageID
-	warnTarget := "@" + origMsg.From.UserName
+	warnTargetName := "@" + origMsg.From.UserName
 	if origMsg.SenderChat != nil && origMsg.SenderChat.ID != 0 && origMsg.SenderChat.ID != a.primChatID {
 		chName := a.channelDisplayName(origMsg.SenderChat)
 		if origMsg.SenderChat.UserName != "" {
-			warnTarget = "@" + chName
+			warnTargetName = "@" + chName
 		} else {
-			warnTarget = chName
+			warnTargetName = chName
 		}
 	}
 	warnMsg := fmt.Sprintf("warning from %s\n\n%s %s", update.Message.From.UserName,
-		warnTarget, a.warnMsg)
+		warnTargetName, a.warnMsg)
 	if err := send(tbapi.NewMessage(a.primChatID, escapeMarkDownV1Text(warnMsg)), a.tbAPI); err != nil {
 		errs = multierror.Append(errs, fmt.Errorf("failed to send warning to main chat: %w", err))
 	}
 
+	if banErr := a.trackWarnAndMaybeBan(origMsg); banErr != nil {
+		errs = multierror.Append(errs, banErr)
+	}
+
 	if err := errs.ErrorOrNil(); err != nil {
 		return fmt.Errorf("direct warn report failed: %w", err)
 	}
 	return nil
 }
 
+// warnTarget identifies the entity (user or channel) that a warn applies to.
+// channelID is 0 for plain users; for channel posts it equals the SenderChat ID.
+type warnTarget struct {
+	userID    int64
+	userName  string
+	channelID int64
+}
+
+// resolveWarnTarget extracts the warn target from the original message.
+// returns (target, true) for plain users and channel posts; (target, false) for
+// anonymous admin posts (SenderChat == group itself, From is shared GroupAnonymousBot)
+// and updates with no resolvable identity.
+func (a *admin) resolveWarnTarget(origMsg *tbapi.Message) (warnTarget, bool) {
+	if origMsg.SenderChat != nil && origMsg.SenderChat.ID != 0 {
+		// anonymous admin posts have SenderChat.ID == primChatID; From identity is the
+		// shared GroupAnonymousBot user. tracking warns against either is meaningless
+		// (banning the group itself or a shared bot id), so skip entirely.
+		if origMsg.SenderChat.ID == a.primChatID {
+			return warnTarget{}, false
+		}
+		return warnTarget{
+			userID:    origMsg.SenderChat.ID,
+			userName:  a.channelDisplayName(origMsg.SenderChat),
+			channelID: origMsg.SenderChat.ID,
+		}, true
+	}
+	if origMsg.From != nil && origMsg.From.ID != 0 {
+		return warnTarget{userID: origMsg.From.ID, userName: origMsg.From.UserName}, true
+	}
+	return warnTarget{}, false
+}
+
+// trackWarnAndMaybeBan records the warning and triggers an auto-ban when the
+// configured threshold is reached within the sliding window. it is a no-op when
+// the feature is disabled (threshold == 0), warnings storage is unwired, or the
+// target cannot be resolved (anonymous admin posts, missing From/SenderChat).
+// returns nil unless the ban itself fails - storage failures are logged but not propagated
+// because the warning message has already been posted (best-effort).
+func (a *admin) trackWarnAndMaybeBan(origMsg *tbapi.Message) error {
+	if a.warnThreshold == 0 || a.warnings == nil {
+		return nil
+	}
+	target, ok := a.resolveWarnTarget(origMsg)
+	if !ok {
+		return nil
+	}
+	ctx := context.TODO()
+	if err := a.warnings.Add(ctx, target.userID, target.userName); err != nil {
+		log.Printf("[WARN] failed to record warn for %q (%d): %v", target.userName, target.userID, err)
+		return nil
+	}
+	count, err := a.warnings.CountWithin(ctx, target.userID, a.warnWindow)
+	if err != nil {
+		log.Printf("[WARN] failed to count warns for %q (%d): %v", target.userName, target.userID, err)
+		return nil
+	}
+	if count < a.warnThreshold {
+		return nil
+	}
+	return a.executeWarnBan(target, count)
+}
+
+// executeWarnBan bans a user or channel after the warn-threshold is reached within warnWindow.
+// it respects dry, training, and softBan modes, and posts an admin-chat notification.
+// it does not update spam samples - a warn is not necessarily spam content.
+func (a *admin) executeWarnBan(target warnTarget, count int) error {
+	log.Printf("[INFO] warn auto-ban triggered for %q (%d): %d warns within %v",
+		target.userName, target.userID, count, a.warnWindow)
+
+	banReq := banRequest{
+		duration:  bot.PermanentBanDuration,
+		userID:    target.userID,
+		channelID: target.channelID,
+		chatID:    a.primChatID,
+		tbAPI:     a.tbAPI,
+		dry:       a.dry,
+		training:  a.trainingMode,
+		userName:  target.userName,
+		restrict:  a.softBan,
+	}
+	if err := banUserOrChannel(banReq); err != nil {
+		return fmt.Errorf("failed to auto-ban %q (%d) after %d warns: %w",
+			target.userName, target.userID, count, err)
+	}
+
+	if a.adminChatID == 0 {
+		return nil
+	}
+
+	action := "banned"
+	switch {
+	case a.dry:
+		action = "would have banned"
+	case a.trainingMode:
+		action = "would have banned (training)"
+	case a.softBan && target.channelID == 0:
+		action = "restricted"
+	}
+
+	displayName := target.userName
+	if target.channelID == 0 && target.userName != "" {
+		displayName = "@" + target.userName
+	}
+	notification := fmt.Sprintf("**warn auto-%s** %s (%d) after %d warns within %v",
+		action, escapeMarkDownV1Text(displayName), target.userID, count, a.warnWindow)
+	if err := send(tbapi.NewMessage(a.adminChatID, notification), a.tbAPI); err != nil {
+		return fmt.Errorf("failed to send warn auto-ban notification: %w", err)
+	}
+	return nil
+}
+
 // returns the user ID and username from the tg update if's forwarded message,
 // or just username in case sender is hidden user
 func (a *admin) getForwardUsernameAndID(update tbapi.Update) (fwdID int64, username string) {