feat: add compaction ratio setting to control partial context compaction

Allow users to configure what percentage of older messages to compact,
keeping the most recent portion intact. Default ratio is 80%, meaning
the oldest 80% of uncompacted messages are summarized while the newest
20% remain as-is for full-fidelity context.

internal/acl/service_test.go

@@ -114,10 +114,11 @@ func makeBotRow(botID, ownerUserID pgtype.UUID) *fakeRow {
 			*dest[15].(*string) = ""                 // HeartbeatPrompt
 			*dest[16].(*bool) = false                // CompactionEnabled
 			*dest[17].(*int32) = 100000              // CompactionThreshold
-			*dest[18].(*pgtype.UUID) = pgtype.UUID{} // CompactionModelID
-			*dest[19].(*[]byte) = []byte(`{}`)
-			*dest[20].(*pgtype.Timestamptz) = pgtype.Timestamptz{}
+			*dest[18].(*int32) = 80                  // CompactionRatio
+			*dest[19].(*pgtype.UUID) = pgtype.UUID{} // CompactionModelID
+			*dest[20].(*[]byte) = []byte(`{}`)
 			*dest[21].(*pgtype.Timestamptz) = pgtype.Timestamptz{}
+			*dest[22].(*pgtype.Timestamptz) = pgtype.Timestamptz{}
 			return nil
 		},
 	}

internal/bots/service_test.go

@@ -71,10 +71,11 @@ func makeBotRow(botID, ownerUserID pgtype.UUID) *fakeRow {
 			*dest[15].(*string) = ""                 // HeartbeatPrompt
 			*dest[16].(*bool) = false                // CompactionEnabled
 			*dest[17].(*int32) = 100000              // CompactionThreshold
-			*dest[18].(*pgtype.UUID) = pgtype.UUID{} // CompactionModelID
-			*dest[19].(*[]byte) = []byte(`{}`)
-			*dest[20].(*pgtype.Timestamptz) = pgtype.Timestamptz{}
+			*dest[18].(*int32) = 80                  // CompactionRatio
+			*dest[19].(*pgtype.UUID) = pgtype.UUID{} // CompactionModelID
+			*dest[20].(*[]byte) = []byte(`{}`)
 			*dest[21].(*pgtype.Timestamptz) = pgtype.Timestamptz{}
+			*dest[22].(*pgtype.Timestamptz) = pgtype.Timestamptz{}
 			return nil
 		},
 	}

internal/compaction/service.go

@@ -79,6 +79,12 @@ func (s *Service) doCompaction(ctx context.Context, logID pgtype.UUID, sessionUU
 		return nil
 	}
 
+	toCompact := splitByRatio(messages, cfg.TotalInputTokens, cfg.Ratio)
+	if len(toCompact) == 0 {
+		s.completeLog(ctx, logID, "ok", "", "", nil, pgtype.UUID{})
+		return nil
+	}
+
 	priorLogs, err := s.queries.ListCompactionLogsBySession(ctx, sessionUUID)
 	if err != nil {
 		return err
@@ -90,9 +96,9 @@ func (s *Service) doCompaction(ctx context.Context, logID pgtype.UUID, sessionUU
 		}
 	}
 
-	entries := make([]messageEntry, 0, len(messages))
-	messageIDs := make([]pgtype.UUID, 0, len(messages))
-	for _, m := range messages {
+	entries := make([]messageEntry, 0, len(toCompact))
+	messageIDs := make([]pgtype.UUID, 0, len(toCompact))
+	for _, m := range toCompact {
 		entries = append(entries, messageEntry{
 			Role:    m.Role,
 			Content: extractTextContent(m.Content),
@@ -258,3 +264,49 @@ func extractTextContent(content []byte) string {
 func joinTexts(parts []string) string {
 	return strings.Join(parts, " ")
 }
+
+// splitByRatio splits messages so that roughly the first ratio% (by token weight)
+// are returned for compaction, and the rest are kept as-is.
+// When ratio >= 100 or totalInputTokens <= 0, all messages are returned.
+func splitByRatio(messages []sqlc.ListUncompactedMessagesBySessionRow, totalInputTokens, ratio int) []sqlc.ListUncompactedMessagesBySessionRow {
+	if ratio >= 100 || ratio <= 0 || totalInputTokens <= 0 || len(messages) == 0 {
+		return messages
+	}
+
+	keepTokens := totalInputTokens * (100 - ratio) / 100
+	if keepTokens <= 0 {
+		return messages
+	}
+
+	accumulated := 0
+	cutoff := len(messages)
+	for i := len(messages) - 1; i >= 0; i-- {
+		accumulated += estimateRowTokens(messages[i])
+		if accumulated >= keepTokens {
+			cutoff = i + 1
+			break
+		}
+	}
+
+	if cutoff <= 0 {
+		return nil
+	}
+	if cutoff >= len(messages) {
+		return messages
+	}
+	return messages[:cutoff]
+}
+
+type usagePayload struct {
+	OutputTokens *int `json:"output_tokens"`
+}
+
+func estimateRowTokens(m sqlc.ListUncompactedMessagesBySessionRow) int {
+	if len(m.Usage) > 0 {
+		var u usagePayload
+		if json.Unmarshal(m.Usage, &u) == nil && u.OutputTokens != nil && *u.OutputTokens > 0 {
+			return *u.OutputTokens
+		}
+	}
+	return len(m.Content) / 4
+}

internal/compaction/types.go

@@ -28,12 +28,14 @@ type ListLogsResponse struct {
 
 // TriggerConfig holds the parameters needed to trigger a compaction.
 type TriggerConfig struct {
-	BotID          string
-	SessionID      string
-	ModelID        string
-	ClientType     string
-	APIKey         string //nolint:gosec // runtime credential, not a hardcoded secret
-	CodexAccountID string
-	BaseURL        string
-	HTTPClient     *http.Client
+	BotID            string
+	SessionID        string
+	ModelID          string
+	ClientType       string
+	APIKey           string //nolint:gosec // runtime credential, not a hardcoded secret
+	CodexAccountID   string
+	BaseURL          string
+	HTTPClient       *http.Client
+	Ratio            int
+	TotalInputTokens int
 }

internal/conversation/flow/resolver_compaction.go

@@ -31,9 +31,16 @@ func (r *Resolver) maybeCompact(ctx context.Context, req conversation.ChatReques
 		modelID = rc.model.ID
 	}
 
+	ratio := settings.CompactionRatio
+	if ratio <= 0 || ratio > 100 {
+		ratio = 80
+	}
+
 	cfg := compaction.TriggerConfig{
-		BotID:     req.BotID,
-		SessionID: req.SessionID,
+		BotID:            req.BotID,
+		SessionID:        req.SessionID,
+		Ratio:            ratio,
+		TotalInputTokens: inputTokens,
 	}
 
 	model, err := r.modelsService.GetByID(ctx, modelID)

internal/db/sqlc/bots.sql.go

@@ -94,7 +94,7 @@ func (q *Queries) DeleteBotByID(ctx context.Context, id pgtype.UUID) error {
 }
 
 const getBotByID = `-- name: GetBotByID :one
-SELECT id, owner_user_id, display_name, avatar_url, timezone, is_active, status, language, reasoning_enabled, reasoning_effort, chat_model_id, search_provider_id, memory_provider_id, heartbeat_enabled, heartbeat_interval, heartbeat_prompt, compaction_enabled, compaction_threshold, compaction_model_id, metadata, created_at, updated_at
+SELECT id, owner_user_id, display_name, avatar_url, timezone, is_active, status, language, reasoning_enabled, reasoning_effort, chat_model_id, search_provider_id, memory_provider_id, heartbeat_enabled, heartbeat_interval, heartbeat_prompt, compaction_enabled, compaction_threshold, compaction_ratio, compaction_model_id, metadata, created_at, updated_at
 FROM bots
 WHERE id = $1
 `
@@ -118,6 +118,7 @@ type GetBotByIDRow struct {
 	HeartbeatPrompt     string             `json:"heartbeat_prompt"`
 	CompactionEnabled   bool               `json:"compaction_enabled"`
 	CompactionThreshold int32              `json:"compaction_threshold"`
+	CompactionRatio     int32              `json:"compaction_ratio"`
 	CompactionModelID   pgtype.UUID        `json:"compaction_model_id"`
 	Metadata            []byte             `json:"metadata"`
 	CreatedAt           pgtype.Timestamptz `json:"created_at"`
@@ -146,6 +147,7 @@ func (q *Queries) GetBotByID(ctx context.Context, id pgtype.UUID) (GetBotByIDRow
 		&i.HeartbeatPrompt,
 		&i.CompactionEnabled,
 		&i.CompactionThreshold,
+		&i.CompactionRatio,
 		&i.CompactionModelID,
 		&i.Metadata,
 		&i.CreatedAt,

internal/db/sqlc/conversations.sql.go

@@ -511,7 +511,7 @@ WITH updated AS (
   SET display_name = $1,
       updated_at = now()
   WHERE bots.id = $2
-  RETURNING id, owner_user_id, display_name, avatar_url, timezone, is_active, status, language, reasoning_enabled, reasoning_effort, chat_model_id, search_provider_id, memory_provider_id, heartbeat_enabled, heartbeat_interval, heartbeat_prompt, heartbeat_model_id, compaction_enabled, compaction_threshold, compaction_model_id, title_model_id, tts_model_id, browser_context_id, metadata, created_at, updated_at, acl_default_effect
+  RETURNING id, owner_user_id, display_name, avatar_url, timezone, is_active, status, language, reasoning_enabled, reasoning_effort, chat_model_id, search_provider_id, memory_provider_id, heartbeat_enabled, heartbeat_interval, heartbeat_prompt, heartbeat_model_id, compaction_enabled, compaction_threshold, compaction_ratio, compaction_model_id, title_model_id, tts_model_id, browser_context_id, metadata, created_at, updated_at, acl_default_effect
 )
 SELECT
   updated.id AS id,

internal/db/sqlc/models.go

@@ -28,6 +28,7 @@ type Bot struct {
 	HeartbeatModelID    pgtype.UUID        `json:"heartbeat_model_id"`
 	CompactionEnabled   bool               `json:"compaction_enabled"`
 	CompactionThreshold int32              `json:"compaction_threshold"`
+	CompactionRatio     int32              `json:"compaction_ratio"`
 	CompactionModelID   pgtype.UUID        `json:"compaction_model_id"`
 	TitleModelID        pgtype.UUID        `json:"title_model_id"`
 	TtsModelID          pgtype.UUID        `json:"tts_model_id"`

internal/db/sqlc/settings.sql.go

@@ -21,6 +21,7 @@ SET language = 'auto',
     heartbeat_prompt = '',
     compaction_enabled = false,
     compaction_threshold = 100000,
+    compaction_ratio = 80,
     chat_model_id = NULL,
     heartbeat_model_id = NULL,
     compaction_model_id = NULL,
@@ -49,6 +50,7 @@ SELECT
   bots.heartbeat_prompt,
   bots.compaction_enabled,
   bots.compaction_threshold,
+  bots.compaction_ratio,
   chat_models.id AS chat_model_id,
   heartbeat_models.id AS heartbeat_model_id,
   compaction_models.id AS compaction_model_id,
@@ -79,6 +81,7 @@ type GetSettingsByBotIDRow struct {
 	HeartbeatPrompt     string      `json:"heartbeat_prompt"`
 	CompactionEnabled   bool        `json:"compaction_enabled"`
 	CompactionThreshold int32       `json:"compaction_threshold"`
+	CompactionRatio     int32       `json:"compaction_ratio"`
 	ChatModelID         pgtype.UUID `json:"chat_model_id"`
 	HeartbeatModelID    pgtype.UUID `json:"heartbeat_model_id"`
 	CompactionModelID   pgtype.UUID `json:"compaction_model_id"`
@@ -102,6 +105,7 @@ func (q *Queries) GetSettingsByBotID(ctx context.Context, id pgtype.UUID) (GetSe
 		&i.HeartbeatPrompt,
 		&i.CompactionEnabled,
 		&i.CompactionThreshold,
+		&i.CompactionRatio,
 		&i.ChatModelID,
 		&i.HeartbeatModelID,
 		&i.CompactionModelID,
@@ -125,17 +129,18 @@ WITH updated AS (
       heartbeat_prompt = $6,
       compaction_enabled = $7,
       compaction_threshold = $8,
-      chat_model_id = COALESCE($9::uuid, bots.chat_model_id),
-      heartbeat_model_id = COALESCE($10::uuid, bots.heartbeat_model_id),
-      compaction_model_id = COALESCE($11::uuid, bots.compaction_model_id),
-      title_model_id = COALESCE($12::uuid, bots.title_model_id),
-      search_provider_id = COALESCE($13::uuid, bots.search_provider_id),
-      memory_provider_id = COALESCE($14::uuid, bots.memory_provider_id),
-      tts_model_id = COALESCE($15::uuid, bots.tts_model_id),
-      browser_context_id = COALESCE($16::uuid, bots.browser_context_id),
+      compaction_ratio = $9,
+      chat_model_id = COALESCE($10::uuid, bots.chat_model_id),
+      heartbeat_model_id = COALESCE($11::uuid, bots.heartbeat_model_id),
+      compaction_model_id = COALESCE($12::uuid, bots.compaction_model_id),
+      title_model_id = COALESCE($13::uuid, bots.title_model_id),
+      search_provider_id = COALESCE($14::uuid, bots.search_provider_id),
+      memory_provider_id = COALESCE($15::uuid, bots.memory_provider_id),
+      tts_model_id = COALESCE($16::uuid, bots.tts_model_id),
+      browser_context_id = COALESCE($17::uuid, bots.browser_context_id),
       updated_at = now()
-  WHERE bots.id = $17
-  RETURNING bots.id, bots.language, bots.reasoning_enabled, bots.reasoning_effort, bots.heartbeat_enabled, bots.heartbeat_interval, bots.heartbeat_prompt, bots.compaction_enabled, bots.compaction_threshold, bots.chat_model_id, bots.heartbeat_model_id, bots.compaction_model_id, bots.title_model_id, bots.search_provider_id, bots.memory_provider_id, bots.tts_model_id, bots.browser_context_id
+  WHERE bots.id = $18
+  RETURNING bots.id, bots.language, bots.reasoning_enabled, bots.reasoning_effort, bots.heartbeat_enabled, bots.heartbeat_interval, bots.heartbeat_prompt, bots.compaction_enabled, bots.compaction_threshold, bots.compaction_ratio, bots.chat_model_id, bots.heartbeat_model_id, bots.compaction_model_id, bots.title_model_id, bots.search_provider_id, bots.memory_provider_id, bots.tts_model_id, bots.browser_context_id
 )
 SELECT
   updated.id AS bot_id,
@@ -147,6 +152,7 @@ SELECT
   updated.heartbeat_prompt,
   updated.compaction_enabled,
   updated.compaction_threshold,
+  updated.compaction_ratio,
   chat_models.id AS chat_model_id,
   heartbeat_models.id AS heartbeat_model_id,
   compaction_models.id AS compaction_model_id,
@@ -175,6 +181,7 @@ type UpsertBotSettingsParams struct {
 	HeartbeatPrompt     string      `json:"heartbeat_prompt"`
 	CompactionEnabled   bool        `json:"compaction_enabled"`
 	CompactionThreshold int32       `json:"compaction_threshold"`
+	CompactionRatio     int32       `json:"compaction_ratio"`
 	ChatModelID         pgtype.UUID `json:"chat_model_id"`
 	HeartbeatModelID    pgtype.UUID `json:"heartbeat_model_id"`
 	CompactionModelID   pgtype.UUID `json:"compaction_model_id"`
@@ -196,6 +203,7 @@ type UpsertBotSettingsRow struct {
 	HeartbeatPrompt     string      `json:"heartbeat_prompt"`
 	CompactionEnabled   bool        `json:"compaction_enabled"`
 	CompactionThreshold int32       `json:"compaction_threshold"`
+	CompactionRatio     int32       `json:"compaction_ratio"`
 	ChatModelID         pgtype.UUID `json:"chat_model_id"`
 	HeartbeatModelID    pgtype.UUID `json:"heartbeat_model_id"`
 	CompactionModelID   pgtype.UUID `json:"compaction_model_id"`
@@ -216,6 +224,7 @@ func (q *Queries) UpsertBotSettings(ctx context.Context, arg UpsertBotSettingsPa
 		arg.HeartbeatPrompt,
 		arg.CompactionEnabled,
 		arg.CompactionThreshold,
+		arg.CompactionRatio,
 		arg.ChatModelID,
 		arg.HeartbeatModelID,
 		arg.CompactionModelID,
@@ -237,6 +246,7 @@ func (q *Queries) UpsertBotSettings(ctx context.Context, arg UpsertBotSettingsPa
 		&i.HeartbeatPrompt,
 		&i.CompactionEnabled,
 		&i.CompactionThreshold,
+		&i.CompactionRatio,
 		&i.ChatModelID,
 		&i.HeartbeatModelID,
 		&i.CompactionModelID,

internal/settings/service.go

@@ -69,7 +69,7 @@ func (s *Service) UpsertBot(ctx context.Context, botID string, req UpsertRequest
 	if err != nil {
 		return Settings{}, err
 	}
-	current := normalizeBotSetting(botRow.Language, aclDefaultEffect, botRow.ReasoningEnabled, botRow.ReasoningEffort, botRow.HeartbeatEnabled, botRow.HeartbeatInterval, botRow.CompactionEnabled, botRow.CompactionThreshold)
+	current := normalizeBotSetting(botRow.Language, aclDefaultEffect, botRow.ReasoningEnabled, botRow.ReasoningEffort, botRow.HeartbeatEnabled, botRow.HeartbeatInterval, botRow.CompactionEnabled, botRow.CompactionThreshold, botRow.CompactionRatio)
 	if strings.TrimSpace(req.Language) != "" {
 		current.Language = strings.TrimSpace(req.Language)
 	}
@@ -94,6 +94,9 @@ func (s *Service) UpsertBot(ctx context.Context, botID string, req UpsertRequest
 	if req.CompactionThreshold != nil && *req.CompactionThreshold >= 0 {
 		current.CompactionThreshold = *req.CompactionThreshold
 	}
+	if req.CompactionRatio != nil && *req.CompactionRatio >= 1 && *req.CompactionRatio <= 100 {
+		current.CompactionRatio = *req.CompactionRatio
+	}
 	chatModelUUID := pgtype.UUID{}
 	if value := strings.TrimSpace(req.ChatModelID); value != "" {
 		modelID, err := s.resolveModelUUID(ctx, value)
@@ -170,6 +173,7 @@ func (s *Service) UpsertBot(ctx context.Context, botID string, req UpsertRequest
 		HeartbeatPrompt:     "",
 		CompactionEnabled:   current.CompactionEnabled,
 		CompactionThreshold: int32(current.CompactionThreshold), //nolint:gosec // bounded by non-negative setter above
+		CompactionRatio:     int32(current.CompactionRatio),     //nolint:gosec // bounded 1-100 above
 		ChatModelID:         chatModelUUID,
 		HeartbeatModelID:    heartbeatModelUUID,
 		CompactionModelID:   compactionModelUUID,
@@ -209,7 +213,7 @@ func (s *Service) Delete(ctx context.Context, botID string) error {
 	return nil
 }
 
-func normalizeBotSetting(language string, aclDefaultEffect string, reasoningEnabled bool, reasoningEffort string, heartbeatEnabled bool, heartbeatInterval int32, compactionEnabled bool, compactionThreshold int32) Settings {
+func normalizeBotSetting(language string, aclDefaultEffect string, reasoningEnabled bool, reasoningEffort string, heartbeatEnabled bool, heartbeatInterval int32, compactionEnabled bool, compactionThreshold int32, compactionRatio int32) Settings {
 	settings := Settings{
 		Language:            strings.TrimSpace(language),
 		AclDefaultEffect:    strings.TrimSpace(aclDefaultEffect),
@@ -219,6 +223,7 @@ func normalizeBotSetting(language string, aclDefaultEffect string, reasoningEnab
 		HeartbeatInterval:   int(heartbeatInterval),
 		CompactionEnabled:   compactionEnabled,
 		CompactionThreshold: int(compactionThreshold),
+		CompactionRatio:     int(compactionRatio),
 	}
 	if settings.Language == "" {
 		settings.Language = DefaultLanguage
@@ -235,6 +240,9 @@ func normalizeBotSetting(language string, aclDefaultEffect string, reasoningEnab
 	if settings.CompactionThreshold < 0 {
 		settings.CompactionThreshold = 0
 	}
+	if settings.CompactionRatio < 1 || settings.CompactionRatio > 100 {
+		settings.CompactionRatio = 80
+	}
 	return settings
 }
 
@@ -256,6 +264,7 @@ func normalizeBotSettingsReadRow(row sqlc.GetSettingsByBotIDRow) Settings {
 		row.HeartbeatInterval,
 		row.CompactionEnabled,
 		row.CompactionThreshold,
+		row.CompactionRatio,
 		row.ChatModelID,
 		row.HeartbeatModelID,
 		row.CompactionModelID,
@@ -276,6 +285,7 @@ func normalizeBotSettingsWriteRow(row sqlc.UpsertBotSettingsRow) Settings {
 		row.HeartbeatInterval,
 		row.CompactionEnabled,
 		row.CompactionThreshold,
+		row.CompactionRatio,
 		row.ChatModelID,
 		row.HeartbeatModelID,
 		row.CompactionModelID,
@@ -295,6 +305,7 @@ func normalizeBotSettingsFields(
 	heartbeatInterval int32,
 	compactionEnabled bool,
 	compactionThreshold int32,
+	compactionRatio int32,
 	chatModelID pgtype.UUID,
 	heartbeatModelID pgtype.UUID,
 	compactionModelID pgtype.UUID,
@@ -304,7 +315,7 @@ func normalizeBotSettingsFields(
 	ttsModelID pgtype.UUID,
 	browserContextID pgtype.UUID,
 ) Settings {
-	settings := normalizeBotSetting(language, "", reasoningEnabled, reasoningEffort, heartbeatEnabled, heartbeatInterval, compactionEnabled, compactionThreshold)
+	settings := normalizeBotSetting(language, "", reasoningEnabled, reasoningEffort, heartbeatEnabled, heartbeatInterval, compactionEnabled, compactionThreshold, compactionRatio)
 	if chatModelID.Valid {
 		settings.ChatModelID = uuid.UUID(chatModelID.Bytes).String()
 	}

internal/settings/types.go

@@ -22,6 +22,7 @@ type Settings struct {
 	TitleModelID        string `json:"title_model_id"`
 	CompactionEnabled   bool   `json:"compaction_enabled"`
 	CompactionThreshold int    `json:"compaction_threshold"`
+	CompactionRatio     int    `json:"compaction_ratio"`
 	CompactionModelID   string `json:"compaction_model_id,omitempty"`
 }
 
@@ -41,5 +42,6 @@ type UpsertRequest struct {
 	TitleModelID        string  `json:"title_model_id,omitempty"`
 	CompactionEnabled   *bool   `json:"compaction_enabled,omitempty"`
 	CompactionThreshold *int    `json:"compaction_threshold,omitempty"`
+	CompactionRatio     *int    `json:"compaction_ratio,omitempty"`
 	CompactionModelID   *string `json:"compaction_model_id,omitempty"`
 }