feat: configure llm history context

Author: ForeverInLaw

app/assembly.go

@@ -163,6 +163,7 @@ func activateServer(
 		MetaGiveaway:             opts.Meta.Giveaway,
 		MultiLangLimit:           opts.MultiLangWords,
 		LLMConsensus:             opts.LLM.Consensus,
+		LLMHistoryContextSize:    opts.LLM.HistoryContextSize,
 		OpenAIEnabled:            opts.OpenAI.Token != "" || opts.OpenAI.APIBase != "",
 		OpenAIVeto:               opts.OpenAI.Veto,
 		OpenAIHistorySize:        opts.OpenAI.HistorySize,
@@ -332,24 +333,25 @@ func buildDetectorConfig(opts options, ruleSet rules.RuleSet) tgspam.Config {
 		casAPI = opts.CAS.API
 	}
 	cfg := tgspam.Config{
-		MaxAllowedEmoji:     ruleSet.Detection.MaxEmoji,
-		MinMsgLen:           ruleSet.Detection.MinMsgLen,
-		SimilarityThreshold: ruleSet.Detection.SimilarityThreshold,
-		MinSpamProbability:  ruleSet.Detection.MinSpamProbability,
-		CasAPI:              casAPI,
-		CasUserAgent:        opts.CAS.UserAgent,
-		HTTPClient:          &http.Client{Timeout: opts.CAS.Timeout},
-		FirstMessageOnly:    !ruleSet.Detection.ParanoidMode,
-		FirstMessagesCount:  ruleSet.Detection.FirstMessagesCount,
-		OpenAIVeto:          ruleSet.OpenAI.Veto,
-		OpenAIHistorySize:   ruleSet.OpenAI.HistorySize,
-		GeminiVeto:          ruleSet.Gemini.Veto,
-		GeminiHistorySize:   ruleSet.Gemini.HistorySize,
-		LLMMode:             tgspam.LLMMode(ruleSet.LLM.Mode),
-		LLMConsensus:        tgspam.LLMConsensusMode(ruleSet.LLM.Consensus),
-		LLMRequestTimeout:   opts.LLM.RequestTimeout,
-		MultiLangWords:      ruleSet.Detection.MultiLangWords,
-		HistorySize:         ruleSet.Detection.HistorySize,
+		MaxAllowedEmoji:       ruleSet.Detection.MaxEmoji,
+		MinMsgLen:             ruleSet.Detection.MinMsgLen,
+		SimilarityThreshold:   ruleSet.Detection.SimilarityThreshold,
+		MinSpamProbability:    ruleSet.Detection.MinSpamProbability,
+		CasAPI:                casAPI,
+		CasUserAgent:          opts.CAS.UserAgent,
+		HTTPClient:            &http.Client{Timeout: opts.CAS.Timeout},
+		FirstMessageOnly:      !ruleSet.Detection.ParanoidMode,
+		FirstMessagesCount:    ruleSet.Detection.FirstMessagesCount,
+		OpenAIVeto:            ruleSet.OpenAI.Veto,
+		OpenAIHistorySize:     ruleSet.OpenAI.HistorySize,
+		GeminiVeto:            ruleSet.Gemini.Veto,
+		GeminiHistorySize:     ruleSet.Gemini.HistorySize,
+		LLMMode:               tgspam.LLMMode(ruleSet.LLM.Mode),
+		LLMConsensus:          tgspam.LLMConsensusMode(ruleSet.LLM.Consensus),
+		LLMHistoryContextSize: ruleSet.LLM.HistoryContextSize,
+		LLMRequestTimeout:     opts.LLM.RequestTimeout,
+		MultiLangWords:        ruleSet.Detection.MultiLangWords,
+		HistorySize:           ruleSet.Detection.HistorySize,
 	}
 
 	if ruleSet.Detection.FirstMessagesCount > 0 {

app/main.go

@@ -109,9 +109,10 @@ type options struct {
 	} `group:"gemini" namespace:"gemini" env-namespace:"GEMINI"`
 
 	LLM struct {
-		Consensus      string        `long:"consensus" env:"CONSENSUS" choice:"any" choice:"all" default:"any" description:"how eligible LLMs flip the base decision"`
-		Mode           string        `long:"mode" env:"MODE" choice:"" choice:"missed" choice:"flagged" choice:"always" default:"" description:"which messages LLM checks: missed, flagged, or always; empty preserves provider veto flags"`
-		RequestTimeout time.Duration `long:"request-timeout" env:"REQUEST_TIMEOUT" default:"30s" description:"timeout for individual LLM requests"`
+		Consensus          string        `long:"consensus" env:"CONSENSUS" choice:"any" choice:"all" default:"any" description:"how eligible LLMs flip the base decision"`
+		Mode               string        `long:"mode" env:"MODE" choice:"" choice:"missed" choice:"flagged" choice:"always" default:"" description:"which messages LLM checks: missed, flagged, or always; empty preserves provider veto flags"`
+		HistoryContextSize int           `long:"history-context-size" env:"HISTORY_CONTEXT_SIZE" default:"0" description:"recent chat messages to include in LLM context, 0 disables"`
+		RequestTimeout     time.Duration `long:"request-timeout" env:"REQUEST_TIMEOUT" default:"30s" description:"timeout for individual LLM requests"`
 	} `group:"llm" namespace:"llm" env-namespace:"LLM"`
 
 	LuaPlugins struct {

app/main_runtime_assembly_test.go

@@ -253,6 +253,7 @@ func Test_makeDetector(t *testing.T) {
 				HistorySize:        5,
 				CheckShortMessages: true,
 			},
+			LLM: rules.LLMCommonRules{HistoryContextSize: 3},
 		}
 
 		res := makeDetectorWithRuleSet(opts, ruleSet)
@@ -268,6 +269,7 @@ func Test_makeDetector(t *testing.T) {
 		assert.Equal(t, 4, res.OpenAIHistorySize)
 		assert.True(t, res.GeminiVeto)
 		assert.Equal(t, 5, res.GeminiHistorySize)
+		assert.Equal(t, 3, res.LLMHistoryContextSize)
 		assert.Empty(t, res.LLMMode)
 	})
 }
@@ -437,7 +439,7 @@ func TestAssembleRuntimeUsesActiveRuleSet(t *testing.T) {
 	require.NoError(t, err)
 	_, err = db.ExecContext(ctx,
 		"INSERT INTO rule_set_versions (workspace_id, gid, tenant_id, version, source, payload, created_at) VALUES (?, ?, ?, ?, ?, ?, ?)",
-		"gr1", opts.InstanceID, opts.InstanceID, 2, "test", `{"workspace_id":"gr1","version":2,"source":"test","meta":{"links_limit":5,"mentions_limit":4},"duplicates":{"threshold":4,"window":60000000000},"abnormal_spacing":{"enabled":true,"space_ratio_threshold":0.25,"short_word_ratio_threshold":0.6,"short_word_len":5,"min_words":8},"moderation":{"first_strike":900000000000,"second_strike":21600000000000,"soft_ban":true,"dry_run":true},"reports":{"enabled":true,"threshold":5,"auto_ban_threshold":6,"rate_limit":2,"rate_period":120000000000},"openai":{"enabled":false,"veto":true,"model":"gpt-active","history_size":3,"check_short_messages":true},"gemini":{"enabled":false,"veto":false,"model":"gemini-active","history_size":7,"check_short_messages":false}}`,
+		"gr1", opts.InstanceID, opts.InstanceID, 2, "test", `{"workspace_id":"gr1","version":2,"source":"test","meta":{"links_limit":5,"mentions_limit":4},"duplicates":{"threshold":4,"window":60000000000},"abnormal_spacing":{"enabled":true,"space_ratio_threshold":0.25,"short_word_ratio_threshold":0.6,"short_word_len":5,"min_words":8},"moderation":{"first_strike":900000000000,"second_strike":21600000000000,"soft_ban":true,"dry_run":true},"reports":{"enabled":true,"threshold":5,"auto_ban_threshold":6,"rate_limit":2,"rate_period":120000000000},"llm":{"history_context_size":6},"openai":{"enabled":false,"veto":true,"model":"gpt-active","history_size":3,"check_short_messages":true},"gemini":{"enabled":false,"veto":false,"model":"gemini-active","history_size":7,"check_short_messages":false}}`,
 		time.Now().UTC(),
 	)
 	require.NoError(t, err)
@@ -457,6 +459,7 @@ func TestAssembleRuntimeUsesActiveRuleSet(t *testing.T) {
 	assert.Equal(t, 4, assembly.Detector.DuplicateDetection.Threshold)
 	assert.True(t, assembly.Detector.AbnormalSpacing.Enabled)
 	assert.Equal(t, 3, assembly.Detector.OpenAIHistorySize)
+	assert.Equal(t, 6, assembly.Detector.LLMHistoryContextSize)
 
 	tbAPI := &tbapi.BotAPI{Self: tbapi.User{UserName: "bot"}}
 	listener := assembly.makeTelegramListener(opts, tbAPI)
@@ -607,7 +610,7 @@ func TestBuildDetectorConfig_ReadsDetectionFromRuleSet(t *testing.T) {
 			FirstMessagesCount:  1,
 			ParanoidMode:        false,
 		},
-		LLM: rules.LLMCommonRules{Mode: "flagged", Consensus: "any"},
+		LLM: rules.LLMCommonRules{Mode: "flagged", Consensus: "any", HistoryContextSize: 4},
 	}
 
 	cfg := buildDetectorConfig(opts, rs)
@@ -620,6 +623,7 @@ func TestBuildDetectorConfig_ReadsDetectionFromRuleSet(t *testing.T) {
 	assert.Equal(t, 1, cfg.FirstMessagesCount)
 	assert.Equal(t, tgspam.LLMMode("flagged"), cfg.LLMMode)
 	assert.Equal(t, tgspam.LLMConsensusMode("any"), cfg.LLMConsensus)
+	assert.Equal(t, 4, cfg.LLMHistoryContextSize)
 	assert.Equal(t, "https://api.cas.chat", cfg.CasAPI)
 }
 

app/rules/ruleset.go

@@ -4,7 +4,7 @@ import "time"
 
 // CurrentSchemaVersion is the RuleSet payload schema version. Bump it whenever new
 // fields are added so older persisted rulesets can be detected and backfilled.
-const CurrentSchemaVersion = 2
+const CurrentSchemaVersion = 3
 
 // RuleSet is the single-tenant moderation configuration snapshot.
 type RuleSet struct {
@@ -97,7 +97,8 @@ type DetectionRules struct {
 
 // LLMCommonRules holds LLM settings shared across providers.
 type LLMCommonRules struct {
-	Mode         string `json:"mode"`      // "" | missed | flagged | always
-	Consensus    string `json:"consensus"` // any | all
-	VisionPrompt string `json:"vision_prompt"`
+	Mode               string `json:"mode"`      // "" | missed | flagged | always
+	Consensus          string `json:"consensus"` // any | all
+	HistoryContextSize int    `json:"history_context_size"`
+	VisionPrompt       string `json:"vision_prompt"`
 }

app/rules/ruleset_test.go

@@ -53,16 +53,17 @@ func TestRuleSet_LegacyPayloadDecodesNewFieldsAsZero(t *testing.T) {
 }
 
 func TestCurrentSchemaVersion(t *testing.T) {
-	assert.Equal(t, 2, CurrentSchemaVersion)
+	assert.Equal(t, 3, CurrentSchemaVersion)
 }
 
-func TestRuleSet_VisionPromptJSONRoundTrip(t *testing.T) {
-	rs := RuleSet{LLM: LLMCommonRules{Mode: "flagged", VisionPrompt: "scan this image"}}
+func TestRuleSet_LLMCommonJSONRoundTrip(t *testing.T) {
+	rs := RuleSet{LLM: LLMCommonRules{Mode: "flagged", HistoryContextSize: 5, VisionPrompt: "scan this image"}}
 
 	data, err := json.Marshal(rs)
 	require.NoError(t, err)
 
 	var got RuleSet
 	require.NoError(t, json.Unmarshal(data, &got))
+	assert.Equal(t, 5, got.LLM.HistoryContextSize)
 	assert.Equal(t, "scan this image", got.LLM.VisionPrompt)
 }

app/runtime_assembly.go

@@ -379,8 +379,9 @@ func bootstrapRuleSet(opts options) rules.RuleSet {
 			ParanoidMode:        opts.ParanoidMode,
 		},
 		LLM: rules.LLMCommonRules{
-			Mode:      opts.LLM.Mode,
-			Consensus: opts.LLM.Consensus,
+			Mode:               opts.LLM.Mode,
+			Consensus:          opts.LLM.Consensus,
+			HistoryContextSize: opts.LLM.HistoryContextSize,
 		},
 		OpenAI: rules.LLMRules{
 			Enabled:            opts.OpenAI.Token != "" || opts.OpenAI.APIBase != "",
@@ -416,14 +417,30 @@ func backfillRuleSetSchema(rs rules.RuleSet, opts options) (rules.RuleSet, bool)
 	seed := bootstrapRuleSet(opts)
 	rs.SchemaVersion = rules.CurrentSchemaVersion
 	rs.Detection = seed.Detection
-	rs.LLM = seed.LLM
+	rs.LLM = backfillLLMCommonRules(rs.LLM, seed.LLM)
 	rs.OpenAI.Prompt = seed.OpenAI.Prompt
 	rs.OpenAI.VisionModel = seed.OpenAI.VisionModel
 	rs.Gemini.Prompt = seed.Gemini.Prompt
 	rs.Gemini.VisionModel = seed.Gemini.VisionModel
 	return rs, true
 }
 
+func backfillLLMCommonRules(current, seed rules.LLMCommonRules) rules.LLMCommonRules {
+	if current.Mode == "" {
+		current.Mode = seed.Mode
+	}
+	if current.Consensus == "" {
+		current.Consensus = seed.Consensus
+	}
+	if current.HistoryContextSize == 0 {
+		current.HistoryContextSize = seed.HistoryContextSize
+	}
+	if current.VisionPrompt == "" {
+		current.VisionPrompt = seed.VisionPrompt
+	}
+	return current
+}
+
 func makeApprovedUsersStore(
 	ctx context.Context, dataDB *engine.SQL, detector *tgspam.Detector,
 ) (int, *storage.ApprovedUsers, error) {

app/settings_precedence.go

@@ -199,6 +199,9 @@ func applyExplicitDetectionOverrides(rs *rules.RuleSet, opts options) {
 	if configured("llm.consensus", "LLM_CONSENSUS") {
 		rs.LLM.Consensus = opts.LLM.Consensus
 	}
+	if configured("llm.history-context-size", "LLM_HISTORY_CONTEXT_SIZE") {
+		rs.LLM.HistoryContextSize = opts.LLM.HistoryContextSize
+	}
 }
 
 // envPinnedKey maps a RuleSet JSON path to its CLI flag and env var.
@@ -219,6 +222,7 @@ var envPinnedRegistry = map[string]envPinnedKey{
 	"detection.cas_enabled":          {"cas.api", "CAS_API"},
 	"llm.mode":                       {"llm.mode", "LLM_MODE"},
 	"llm.consensus":                  {"llm.consensus", "LLM_CONSENSUS"},
+	"llm.history_context_size":       {"llm.history-context-size", "LLM_HISTORY_CONTEXT_SIZE"},
 	"openai.veto":                    {"openai.veto", "OPENAI_VETO"},
 	"openai.model":                   {"openai.model", "OPENAI_MODEL"},
 	"openai.history_size":            {"openai.history-size", "OPENAI_HISTORY_SIZE"},

app/settings_precedence_test.go

@@ -21,19 +21,22 @@ func TestConfigured_NonEmptyEnvIsSet(t *testing.T) {
 func TestApplyExplicitOverrides_DetectionFields(t *testing.T) {
 	t.Setenv("MAX_EMOJI", "9")
 	t.Setenv("LLM_MODE", "always")
+	t.Setenv("LLM_HISTORY_CONTEXT_SIZE", "4")
 
 	var opts options
 	opts.MaxEmoji = 9
 	opts.LLM.Mode = "always"
+	opts.LLM.HistoryContextSize = 4
 
 	rs := rules.RuleSet{
 		Detection: rules.DetectionRules{MaxEmoji: 2},
-		LLM:       rules.LLMCommonRules{Mode: "flagged"},
+		LLM:       rules.LLMCommonRules{Mode: "flagged", HistoryContextSize: 9},
 	}
 	applyExplicitRuleSetOverrides(&rs, opts)
 
 	assert.Equal(t, 9, rs.Detection.MaxEmoji, "MAX_EMOJI env must override the ruleset")
 	assert.Equal(t, "always", rs.LLM.Mode, "LLM_MODE env must override the ruleset")
+	assert.Equal(t, 4, rs.LLM.HistoryContextSize, "LLM_HISTORY_CONTEXT_SIZE env must override the ruleset")
 }
 
 func TestApplyExplicitOverrides_DetectionFieldsUntouchedWithoutEnv(t *testing.T) {

app/webapi/assets/settings.html

@@ -92,6 +92,7 @@ <h2 class="page-title">Application Settings</h2>
                             <tr><th>First Messages Count</th><td>{{.FirstMessagesCount}}</td></tr>
                             <tr><th>Multi Lingual Words</th><td>{{.MultiLangLimit}}</td></tr>
                             <tr><th>LLM Consensus</th><td>{{.LLMConsensus}}</td></tr>
+                            <tr><th>LLM History Context Size</th><td>{{.LLMHistoryContextSize}}</td></tr>
                             <tr><th>Abnormal Spacing Enabled</th><td>{{.AbnormalSpacingEnabled}}</td></tr>
                             <tr><th>History Size</th><td>{{.HistorySize}}</td></tr>
                             <tr><th>Paranoid Mode</th><td>{{.ParanoidMode}}</td></tr>

app/webapi/assets/settings_edit.html

@@ -39,6 +39,7 @@ <h2 class="page-title">Edit Settings</h2>
                 <div class="card-body">
                     {{template "settings_enum" dict "Key" "llm.mode" "Label" "LLM mode" "Val" .RuleSet.LLM.Mode "Options" .LLMModes "Pinned" $pinned}}
                     {{template "settings_enum" dict "Key" "llm.consensus" "Label" "LLM consensus" "Val" .RuleSet.LLM.Consensus "Options" .LLMConsensus "Pinned" $pinned}}
+                    {{template "settings_int" dict "Key" "llm.history_context_size" "Label" "History context size" "Val" .RuleSet.LLM.HistoryContextSize "Pinned" $pinned}}
                     {{template "settings_text" dict "Key" "llm.vision_prompt" "Label" "Vision prompt" "Val" .RuleSet.LLM.VisionPrompt "Pinned" $pinned}}
                 </div>
             </div>