Merge pull request #3641 from fosrl/dev

1.22.0

server/db/pg/schema/privateSchema.ts

@@ -468,6 +468,9 @@ export const eventStreamingDestinations = pgTable(
         sendRequestLogs: boolean("sendRequestLogs").notNull().default(false),
         sendActionLogs: boolean("sendActionLogs").notNull().default(false),
         sendAccessLogs: boolean("sendAccessLogs").notNull().default(false),
+        sendAISessionLogs: boolean("sendAISessionLogs")
+            .notNull()
+            .default(false),
         type: varchar("type", { length: 50 }).notNull(), // e.g. "http", "kafka", etc.
         config: text("config").notNull(), // JSON string with the configuration for the destination
         enabled: boolean("enabled").notNull().default(true),

server/db/sqlite/schema/privateSchema.ts

@@ -459,6 +459,9 @@ export const eventStreamingDestinations = sqliteTable(
         sendAccessLogs: integer("sendAccessLogs", { mode: "boolean" })
             .notNull()
             .default(false),
+        sendAISessionLogs: integer("sendAISessionLogs", { mode: "boolean" })
+            .notNull()
+            .default(false),
         type: text("type").notNull(), // e.g. "http", "kafka", etc.
         config: text("config").notNull(), // JSON string with the configuration for the destination
         enabled: integer("enabled", { mode: "boolean" })

server/lib/aiCapabilities.ts

@@ -4,7 +4,7 @@ import { AI_CAPABILITIES, type AiCapability } from "@app/lib/aiCapabilities";
 export { AI_CAPABILITIES, type AiCapability };
 
 export type AiCapabilityRoute = {
-    method: "POST";
+    method: "GET" | "POST";
     path: string;
 };
 
@@ -135,6 +135,21 @@ export const AI_CAPABILITY_DEFS: Record<AiCapability, AiCapabilityDefinition> =
                 joinUpstreamUrl(base, pathFromRequest(req)),
             isStreaming: isBodyOrSseStreaming
         },
+        v1_models: {
+            id: "v1_models",
+            protocolFamily: "anthropic",
+            routes: [
+                { method: "GET", path: "/v1/models" },
+                { method: "GET", path: "/v1/models/:model" }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            // Model listings are answered from the gateway's own view of the
+            // provider allow/block lists rather than proxied upstream, so
+            // there is never a stream to detect.
+            isStreaming: () => false
+        },
         gemini_generate_content: {
             id: "gemini_generate_content",
             protocolFamily: "google",

server/lib/aiMessageNormalization.ts

@@ -471,6 +471,8 @@ const REQUEST_NORMALIZERS: Record<
     openai_chat: normalizeOpenAiChatRequest,
     openai_responses: normalizeOpenAiResponsesRequest,
     anthropic_messages: normalizeAnthropicRequest,
+    // Model discovery carries no transcript to normalize.
+    v1_models: () => null,
     gemini_generate_content: normalizeGeminiRequest,
     google_generate_content: normalizeGeminiRequest,
     google_raw_predict: normalizeBestEffortRequest,
@@ -485,6 +487,7 @@ const RESPONSE_NORMALIZERS: Record<
     openai_chat: normalizeOpenAiChatResponse,
     openai_responses: normalizeOpenAiResponsesResponse,
     anthropic_messages: normalizeAnthropicResponse,
+    v1_models: () => null,
     gemini_generate_content: normalizeGeminiResponse,
     google_generate_content: normalizeGeminiResponse,
     google_raw_predict: normalizeGoogleRawPredictResponse,

server/lib/aiModelCatalog.ts

@@ -44,6 +44,20 @@ export function getCatalogProviderForType(
     return PROVIDER_CATALOG_MAP[type];
 }
 
+/**
+ * Per-model feature flags as reported upstream. `null` means the catalog has
+ * no data for that model - deliberately distinct from `false`, so consumers
+ * can tell "unsupported" apart from "unknown".
+ */
+export type AiModelCapabilityFlags = {
+    functionCalling: boolean | null;
+    vision: boolean | null;
+    promptCaching: boolean | null;
+    reasoning: boolean | null;
+    responseSchema: boolean | null;
+    webSearch: boolean | null;
+};
+
 export type AiModelCatalogEntry = {
     provider: CatalogProvider;
     model: string;
@@ -53,8 +67,20 @@ export type AiModelCatalogEntry = {
         cache: number | null;
         reasoning: number | null;
     };
+    limits: {
+        /** Context window. */
+        input: number | null;
+        /** Cap on the output/max_tokens request parameter. */
+        output: number | null;
+    };
+    capabilities: AiModelCapabilityFlags;
 };
 
+const flag = z.boolean().nullable().optional();
+
+// limits/capabilities are optional so a catalog published before they were
+// added (or an operator's own merge_file) still parses - those entries just
+// report unknown metadata rather than failing the whole payload.
 const catalogEntrySchema = z.object({
     model: z.string(),
     provider: z.string(),
@@ -65,6 +91,22 @@ const catalogEntrySchema = z.object({
             cache: z.number().nullable().optional(),
             reasoning: z.number().nullable().optional()
         })
+        .optional(),
+    limits: z
+        .object({
+            input: z.number().nullable().optional(),
+            output: z.number().nullable().optional()
+        })
+        .optional(),
+    capabilities: z
+        .object({
+            functionCalling: flag,
+            vision: flag,
+            promptCaching: flag,
+            reasoning: flag,
+            responseSchema: flag,
+            webSearch: flag
+        })
         .optional()
 });
 
@@ -108,6 +150,18 @@ function normalizeEntry(raw: RawCatalogEntry): AiModelCatalogEntry | null {
             out: raw.pricing?.out ?? null,
             cache: raw.pricing?.cache ?? null,
             reasoning: raw.pricing?.reasoning ?? null
+        },
+        limits: {
+            input: raw.limits?.input ?? null,
+            output: raw.limits?.output ?? null
+        },
+        capabilities: {
+            functionCalling: raw.capabilities?.functionCalling ?? null,
+            vision: raw.capabilities?.vision ?? null,
+            promptCaching: raw.capabilities?.promptCaching ?? null,
+            reasoning: raw.capabilities?.reasoning ?? null,
+            responseSchema: raw.capabilities?.responseSchema ?? null,
+            webSearch: raw.capabilities?.webSearch ?? null
         }
     };
 }
@@ -284,34 +338,44 @@ export class AiModelCatalog {
 
 export const aiModelCatalog = new AiModelCatalog();
 
+/**
+ * Full catalog entries for a provider type, deduplicated by model id and
+ * sorted by id. Model discovery uses these to report real token limits and
+ * capability flags; `listCatalogModelsForType` is the id-only view of the
+ * same list.
+ */
+export function listCatalogEntriesForType(
+    type: AiProviderType,
+    query?: string
+): AiModelCatalogEntry[] {
+    const catalogProvider = getCatalogProviderForType(type);
+
+    let entries = catalogProvider ? aiModelCatalog.list(catalogProvider) : [];
+
+    if (query) {
+        const q = query.toLowerCase();
+        entries = entries.filter((e) => e.model.toLowerCase().includes(q));
+    }
+
+    const seen = new Set<string>();
+    entries = entries.filter((e) => {
+        if (seen.has(e.model)) {
+            return false;
+        }
+        seen.add(e.model);
+        return true;
+    });
+
+    return [...entries].sort((a, b) => a.model.localeCompare(b.model));
+}
+
 export function listCatalogModelsForType(
     type: AiProviderType,
     query?: string
 ): { model: string }[] {
-    const catalogProvider = getCatalogProviderForType(type);
-
-    let models = catalogProvider
-        ? aiModelCatalog.list(catalogProvider).map((entry) => ({
-              model: entry.model
-          }))
-        : [];
-
-    if (query) {
-        const q = query.toLowerCase();
-        models = models.filter((m) => m.model.toLowerCase().includes(q));
-    }
-
-    const seen = new Set<string>();
-    models = models.filter((m) => {
-        if (seen.has(m.model)) {
-            return false;
-        }
-        seen.add(m.model);
-        return true;
-    });
-
-    models.sort((a, b) => a.model.localeCompare(b.model));
-    return models;
+    return listCatalogEntriesForType(type, query).map((entry) => ({
+        model: entry.model
+    }));
 }
 
 /**

server/lib/aiModelDiscovery.ts

@@ -0,0 +1,235 @@
+import {
+    isAllowedByLists,
+    isModelKeyPattern
+} from "@server/lib/aiModelKeyMatch";
+import type { AiModelCapabilityFlags } from "@server/lib/aiModelCatalog";
+
+// Anthropic's Models API pagination: 20 per page by default, 1..1000.
+export const MODEL_PAGE_DEFAULT_LIMIT = 20;
+export const MODEL_PAGE_MAX_LIMIT = 1000;
+
+// Release dates aren't something we can know for a wildcard allow pattern or a
+// catalog entry. The Models API explicitly permits an epoch value when the
+// release date is unknown.
+const UNKNOWN_CREATED_AT = new Date(0).toISOString();
+
+/**
+ * One entry of Anthropic's `GET /v1/models` response. Only the identity fields
+ * can be filled in from a provider's model lists - token limits and
+ * per-model capability flags aren't derivable from an allow/block list, and the
+ * API schema declares all three nullable.
+ */
+export type AnthropicModelInfo = {
+    type: "model";
+    id: string;
+    display_name: string;
+    created_at: string;
+    max_input_tokens: number | null;
+    max_tokens: number | null;
+    capabilities: Record<string, unknown> | null;
+};
+
+/** A model row an administrator configured explicitly on a provider. */
+export type ConfiguredModel = { name: string; createdAt: number };
+
+/** What the pricing catalog knows about a model beyond its id. */
+export type CatalogModelMetadata = {
+    maxInputTokens: number | null;
+    maxOutputTokens: number | null;
+    capabilities: AiModelCapabilityFlags;
+};
+
+/**
+ * Translates the catalog's flat feature flags into the nested shape
+ * Anthropic's Models API uses. Best-effort by nature: the catalog carries a
+ * coarser set of flags than the Models API describes, so anything it reports
+ * as unknown (`null`) is surfaced as unsupported rather than invented.
+ */
+export function capabilitiesFromCatalog(
+    flags: AiModelCapabilityFlags
+): Record<string, unknown> {
+    const supported = (value: boolean | null) => ({
+        supported: value === true
+    });
+    // The catalog has a single `reasoning` flag and no way to distinguish
+    // adaptive from budget_tokens-style thinking, so both variants follow it.
+    const reasoning = flags.reasoning === true;
+
+    return {
+        batch: supported(null),
+        citations: supported(null),
+        code_execution: supported(null),
+        context_management: {
+            supported: false,
+            clear_thinking_20251015: null,
+            clear_tool_uses_20250919: null,
+            compact_20260112: null
+        },
+        effort: {
+            supported: reasoning,
+            low: supported(flags.reasoning),
+            medium: supported(flags.reasoning),
+            high: supported(flags.reasoning),
+            max: supported(flags.reasoning),
+            xhigh: null
+        },
+        image_input: supported(flags.vision),
+        pdf_input: supported(null),
+        structured_outputs: supported(flags.responseSchema),
+        thinking: {
+            supported: reasoning,
+            types: {
+                adaptive: { supported: reasoning },
+                enabled: { supported: reasoning }
+            }
+        }
+    };
+}
+
+/**
+ * One attached provider's contribution to a resource's model listing, with the
+ * allow/block lists already resolved for the attachment's access mode.
+ */
+export type ModelDiscoveryProvider = {
+    providerId: number;
+    allows: string[];
+    blocks: string[];
+    /**
+     * Concrete model ids the provider's type is known to serve, with whatever
+     * the catalog knows about each. This is what lets a wildcard allow such as
+     * `claude-*` enumerate into real ids; provider types with no catalog
+     * (aggregators, custom) pass an empty map and surface only their exact
+     * allow entries.
+     */
+    catalog: Map<string, CatalogModelMetadata>;
+    /** Keyed by model key, for display names and creation times. */
+    configured: Map<string, ConfiguredModel>;
+};
+
+export type ModelPage = {
+    data: AnthropicModelInfo[];
+    has_more: boolean;
+};
+
+/**
+ * Expands one provider's effective allow/block lists into concrete model ids.
+ * Two sources feed the candidate set: exact (non-wildcard) allow entries, which
+ * are already concrete ids, and the catalog for the provider's type, which is
+ * what makes wildcard allows enumerable. Every candidate is then run back
+ * through the same allow/block check the inference pipeline applies, so a block
+ * pattern hides a model here exactly as it would reject it at request time.
+ */
+export function expandProviderModels(
+    provider: ModelDiscoveryProvider
+): AnthropicModelInfo[] {
+    const candidates = new Set<string>();
+
+    for (const allow of provider.allows) {
+        if (!isModelKeyPattern(allow)) {
+            candidates.add(allow);
+        }
+    }
+    for (const modelId of provider.catalog.keys()) {
+        candidates.add(modelId);
+    }
+
+    const models: AnthropicModelInfo[] = [];
+    for (const modelKey of candidates) {
+        if (!isAllowedByLists(modelKey, provider.allows, provider.blocks)) {
+            continue;
+        }
+        const configured = provider.configured.get(modelKey);
+        const catalog = provider.catalog.get(modelKey);
+
+        models.push({
+            type: "model",
+            id: modelKey,
+            display_name: configured?.name || modelKey,
+            created_at: configured
+                ? new Date(configured.createdAt).toISOString()
+                : UNKNOWN_CREATED_AT,
+            max_input_tokens: catalog?.maxInputTokens ?? null,
+            max_tokens: catalog?.maxOutputTokens ?? null,
+            capabilities: catalog
+                ? capabilitiesFromCatalog(catalog.capabilities)
+                : null
+        });
+    }
+
+    return models;
+}
+
+/**
+ * Aggregates the permitted models across every provider attached to a
+ * resource. Unlike an inference request there is no requested model to
+ * disambiguate on, so no provider selection happens - the listing is the union
+ * of what each provider would accept, deduplicated by model id.
+ */
+export function listPermittedModels(
+    providers: ModelDiscoveryProvider[]
+): AnthropicModelInfo[] {
+    const byModelId = new Map<string, AnthropicModelInfo>();
+
+    // Sorted so a model offered by two providers always resolves to the same
+    // entry, which keeps the cursor ordering stable across requests.
+    const ordered = [...providers].sort((a, b) => a.providerId - b.providerId);
+
+    for (const provider of ordered) {
+        for (const model of expandProviderModels(provider)) {
+            if (!byModelId.has(model.id)) {
+                byModelId.set(model.id, model);
+            }
+        }
+    }
+
+    // "More recently released models are listed first" per the Models API,
+    // with the id as a tie-break so the ordering is total - cursor pagination
+    // needs it to be stable between calls.
+    return [...byModelId.values()].sort((a, b) => {
+        const byCreated = b.created_at.localeCompare(a.created_at);
+        return byCreated !== 0 ? byCreated : a.id.localeCompare(b.id);
+    });
+}
+
+/**
+ * Applies Anthropic's cursor pagination to an ordered model list. `after_id`
+ * returns the page immediately after that model, `before_id` the page
+ * immediately before it. Returns an error message for a caller mistake
+ * (both cursors, or a cursor naming a model that isn't in the list).
+ */
+export function paginateModels(
+    models: AnthropicModelInfo[],
+    limit: number,
+    cursor: { afterId?: string; beforeId?: string }
+): ModelPage | { error: string } {
+    if (cursor.afterId && cursor.beforeId) {
+        return { error: "Only one of after_id and before_id may be provided" };
+    }
+
+    const cursorId = cursor.afterId ?? cursor.beforeId;
+    if (!cursorId) {
+        return {
+            data: models.slice(0, limit),
+            has_more: models.length > limit
+        };
+    }
+
+    const index = models.findIndex((model) => model.id === cursorId);
+    if (index === -1) {
+        return { error: `Unknown cursor id "${cursorId}"` };
+    }
+
+    if (cursor.afterId) {
+        const start = index + 1;
+        return {
+            data: models.slice(start, start + limit),
+            has_more: models.length > start + limit
+        };
+    }
+
+    const start = Math.max(0, index - limit);
+    return {
+        data: models.slice(start, index),
+        has_more: start > 0
+    };
+}

server/lib/aiUsageExtraction.ts

@@ -335,6 +335,8 @@ const EXTRACTORS: Record<
     openai_chat: extractOpenAiChat,
     openai_responses: extractOpenAiResponses,
     anthropic_messages: extractAnthropicMessages,
+    // Model discovery never runs a model, so there are no tokens to bill.
+    v1_models: () => null,
     gemini_generate_content: extractGoogleGenerateContent,
     google_generate_content: extractGoogleGenerateContent,
     // rawPredict is a passthrough to whatever the underlying publisher

server/lib/billing/tierMatrix.ts

@@ -9,6 +9,7 @@ export enum TierFeature {
     AccessLogs = "accessLogs", // set the retention period to none on downgrade
     ActionLogs = "actionLogs", // set the retention period to none on downgrade
     ConnectionLogs = "connectionLogs",
+    AISessionLogs = "aiSessionLogs",
     RotateCredentials = "rotateCredentials",
     MaintenancePage = "maintenancePage", // handle downgrade
     DevicePosture = "devicePosture",
@@ -37,6 +38,7 @@ export const tierMatrix: Record<TierFeature, Tier[]> = {
     [TierFeature.AccessLogs]: ["tier2", "tier3", "enterprise"],
     [TierFeature.ActionLogs]: ["tier2", "tier3", "enterprise"],
     [TierFeature.ConnectionLogs]: ["tier2", "tier3", "enterprise"],
+    [TierFeature.AISessionLogs]: ["tier2", "tier3", "enterprise"],
     [TierFeature.RotateCredentials]: ["tier1", "tier2", "tier3", "enterprise"],
     [TierFeature.MaintenancePage]: ["tier1", "tier2", "tier3", "enterprise"],
     [TierFeature.DevicePosture]: ["tier2", "tier3", "enterprise"],

server/lib/blueprints/applyBlueprint.ts

@@ -1,5 +1,6 @@
 import {
     db,
+    primaryDb,
     newts,
     blueprints,
     Blueprint,
@@ -80,93 +81,103 @@ export async function applyBlueprint({
                 trx,
                 siteId
             );
+        });
 
-            // We need to update the targets on the newts from the successfully updated information
-            for (const result of publicResourcesResults) {
-                for (const target of result.targetsToUpdate) {
-                    const [site] = await trx
-                        .select()
-                        .from(sites)
-                        .innerJoin(newts, eq(sites.siteId, newts.siteId))
-                        .where(
-                            and(
-                                eq(sites.siteId, target.siteId),
-                                eq(sites.orgId, orgId),
-                                eq(sites.type, "newt"),
-                                isNotNull(sites.pubKey)
-                            )
+        // Push updates to newts/clients only after the transaction has
+        // committed. Doing this while the transaction is still open can
+        // race with the writes (e.g. newts requesting config before the
+        // new targets/resources are actually visible), leaving them out
+        // of sync until manually toggled.
+
+        // We need to update the targets on the newts from the successfully updated information
+        for (const result of publicResourcesResults) {
+            for (const target of result.targetsToUpdate) {
+                // read from the primary: this determines whether/how we push
+                // the just-created target to the newt, so a lagging replica
+                // returning stale or missing data here would silently skip
+                // the push
+                const [site] = await primaryDb
+                    .select()
+                    .from(sites)
+                    .innerJoin(newts, eq(sites.siteId, newts.siteId))
+                    .where(
+                        and(
+                            eq(sites.siteId, target.siteId),
+                            eq(sites.orgId, orgId),
+                            eq(sites.type, "newt"),
+                            isNotNull(sites.pubKey)
                         )
-                        .limit(1);
+                    )
+                    .limit(1);
 
-                    if (site) {
-                        logger.debug(
-                            `Updating target ${target.targetId} on site ${site.sites.siteId}`
+                if (site) {
+                    logger.debug(
+                        `Updating target ${target.targetId} on site ${site.sites.siteId}`
+                    );
+
+                    // see if you can find a matching target health check from the healthchecksToUpdate array
+                    const matchingHealthcheck =
+                        result.healthchecksToUpdate.find(
+                            (hc) => hc.targetId === target.targetId
                         );
 
-                        // see if you can find a matching target health check from the healthchecksToUpdate array
-                        const matchingHealthcheck =
-                            result.healthchecksToUpdate.find(
-                                (hc) => hc.targetId === target.targetId
-                            );
-
-                        if (["http", "tcp", "udp"].includes(target.mode)) {
-                            await addProxyTargets(
-                                site.newt.newtId,
-                                [target],
-                                matchingHealthcheck
-                                    ? [matchingHealthcheck]
-                                    : [],
-                                result.proxyResource.mode === "udp"
-                                    ? "udp"
-                                    : "tcp",
-                                site.newt.version
-                            );
-                        } else if (
-                            ["ssh", "rdp", "vnc"].includes(target.mode)
-                        ) {
-                            await sendBrowserGatewayTargets(
-                                site.newt.newtId,
-                                [target],
-                                site.newt.version
-                            );
-                        }
+                    if (["http", "tcp", "udp"].includes(target.mode)) {
+                        await addProxyTargets(
+                            site.newt.newtId,
+                            [target],
+                            matchingHealthcheck
+                                ? [matchingHealthcheck]
+                                : [],
+                            result.proxyResource.mode === "udp"
+                                ? "udp"
+                                : "tcp",
+                            site.newt.version
+                        );
+                    } else if (
+                        ["ssh", "rdp", "vnc"].includes(target.mode)
+                    ) {
+                        await sendBrowserGatewayTargets(
+                            site.newt.newtId,
+                            [target],
+                            site.newt.version
+                        );
                     }
                 }
             }
+        }
 
-            logger.debug(
-                `Successfully updated public resources for org ${orgId}: ${JSON.stringify(publicResourcesResults)}`
-            );
+        logger.debug(
+            `Successfully updated public resources for org ${orgId}: ${JSON.stringify(publicResourcesResults)}`
+        );
 
-            // We need to update the targets on the newts from the successfully updated information
-            for (const result of privateResourcesResults) {
-                rebuildClientAssociationsFromSiteResource(
-                    result.newSiteResource
+        // We need to update the targets on the newts from the successfully updated information
+        for (const result of privateResourcesResults) {
+            rebuildClientAssociationsFromSiteResource(
+                result.newSiteResource
+            )
+                .then(() =>
+                    waitForSiteResourceRebuildIdle(
+                        result.newSiteResource.siteResourceId
+                    )
                 )
-                    .then(() =>
-                        waitForSiteResourceRebuildIdle(
-                            result.newSiteResource.siteResourceId
-                        )
+                .then(() =>
+                    handleMessagingForUpdatedSiteResource(
+                        result.oldSiteResource,
+                        result.newSiteResource,
+                        result.oldSites.map((s) => s.siteId),
+                        result.newSites.map((s) => s.siteId)
                     )
-                    .then(() =>
-                        handleMessagingForUpdatedSiteResource(
-                            result.oldSiteResource,
-                            result.newSiteResource,
-                            result.oldSites.map((s) => s.siteId),
-                            result.newSites.map((s) => s.siteId)
-                        )
-                    )
-                    .catch((e) => {
-                        logger.error(
-                            `Failed to rebuild and handle messaging for site resource ${result.newSiteResource.siteResourceId}. Error: ${e}`
-                        );
-                    });
-            }
+                )
+                .catch((e) => {
+                    logger.error(
+                        `Failed to rebuild and handle messaging for site resource ${result.newSiteResource.siteResourceId}. Error: ${e}`
+                    );
+                });
+        }
 
-            logger.debug(
-                `Successfully updated private resources for org ${orgId}: ${JSON.stringify(privateResourcesResults)}`
-            );
-        });
+        logger.debug(
+            `Successfully updated private resources for org ${orgId}: ${JSON.stringify(privateResourcesResults)}`
+        );
 
         blueprintSucceeded = true;
         blueprintMessage = "Blueprint applied successfully";

server/lib/exitNodes/exitNodes.ts

@@ -22,7 +22,10 @@ export async function listExitNodes(
     // Accepted for parity with the enterprise implementation (used there for
     // site-label filtering of remote exit nodes). The OSS build has no remote
     // exit nodes, so it is unused here.
-    siteId?: number
+    siteId?: number,
+    // Same as above: accepted for parity, unused since the OSS build has no
+    // remote exit nodes to exclude.
+    noRemote = false
 ) {
     // TODO: pick which nodes to send and ping better than just all of them that are not remote
     const allExitNodes = await db

server/lib/exitNodes/subnet.ts

@@ -1,20 +1,26 @@
-import { db, exitNodes, Transaction } from "@server/db";
+import { db, exitNodes, exitNodeOrgs, Transaction } from "@server/db";
 import config from "@server/lib/config";
 import { findNextAvailableCidr } from "@server/lib/ip";
 import { lockManager } from "#dynamic/lib/lock";
+import { eq } from "drizzle-orm";
 
 /**
  * Reserves the next available exit node subnet.
  *
- * Exit node subnets must never overlap with one another - regardless of
- * which org(s) they belong to - since HA exit nodes can end up routing for
- * the same org. This acquires a lock that the caller MUST release (via the
- * returned `release`) only after the chosen address has been durably
- * persisted (e.g. after the enclosing transaction commits), otherwise
- * concurrent callers can race and pick the same subnet.
+ * There isn't enough address space to give every exit node in every org a
+ * globally unique subnet, so we only guarantee uniqueness among exit nodes
+ * that already belong to the same org - that's all that actually matters,
+ * since HA only routes multiple exit nodes for a single org. Pass `orgId` to
+ * scope the search to that org's existing exit nodes; without it, the search
+ * considers every exit node (used by flows with no org context, e.g. the
+ * initial gerbil exit node bootstrap). This acquires a lock that the caller
+ * MUST release (via the returned `release`) only after the chosen address
+ * has been durably persisted (e.g. after the enclosing transaction commits),
+ * otherwise concurrent callers can race and pick the same subnet.
  */
 export async function getNextAvailableSubnet(
-    trx: Transaction | typeof db = db
+    trx: Transaction | typeof db = db,
+    orgId?: string
 ): Promise<{ value: string; release: () => Promise<void> }> {
     const lockKey = "exit-node-subnet-allocation";
     const acquired = await lockManager.acquireLockWithRetry(lockKey, 6000);
@@ -24,12 +30,19 @@ export async function getNextAvailableSubnet(
     const release = () => lockManager.releaseLock(lockKey, acquired);
 
     try {
-        // Get all existing subnets from routes table
-        const existingAddresses = await trx
-            .select({
-                address: exitNodes.address
-            })
-            .from(exitNodes);
+        // Get existing subnets, scoped to this org's exit nodes when known
+        const existingAddresses = orgId
+            ? await trx
+                  .select({ address: exitNodes.address })
+                  .from(exitNodes)
+                  .innerJoin(
+                      exitNodeOrgs,
+                      eq(exitNodeOrgs.exitNodeId, exitNodes.exitNodeId)
+                  )
+                  .where(eq(exitNodeOrgs.orgId, orgId))
+            : await trx
+                  .select({ address: exitNodes.address })
+                  .from(exitNodes);
 
         const addresses = existingAddresses.map((a) => a.address);
         let subnet = findNextAvailableCidr(

server/lib/readConfigFile.ts

@@ -348,7 +348,6 @@ export const configSchema = z
                     .optional()
                     .pipe(z.string())
                     .transform((url) => url.toLowerCase()),
-                use_subdomain: z.boolean().optional().default(false),
                 subnet_group: z.string().optional().default("100.89.137.0/20"),
                 block_size: z.number().positive().gt(0).optional().default(24),
                 site_block_size: z

server/private/lib/alerts/processTestAlerts.ts

@@ -1,3 +1,16 @@
+/*
+ * This file is part of a proprietary work.
+ *
+ * Copyright (c) 2025-2026 Fossorial, Inc.
+ * All rights reserved.
+ *
+ * This file is licensed under the Fossorial Commercial License.
+ * You may not use this file except in compliance with the License.
+ * Unauthorized use, copying, modification, or distribution is strictly prohibited.
+ *
+ * This file is not licensed under the AGPLv3.
+ */
+
 import { db, userOrgRoles, users } from "@server/db";
 import logger from "@server/logger";
 import type {

server/private/lib/exitNodes/exitNodes.ts

@@ -153,7 +153,8 @@ export async function listExitNodes(
     orgId: string,
     filterOnline = false,
     noCloud = false,
-    siteId?: number
+    siteId?: number,
+    noRemote = false
 ) {
     const allExitNodes = await db
         .select({
@@ -242,7 +243,9 @@ export async function listExitNodes(
 
     let remoteExitNodesList = allExitNodes.filter(
         (node) =>
-            node.type === "remoteExitNode" && (!filterOnline || node.online)
+            node.type === "remoteExitNode" &&
+            !noRemote &&
+            (!filterOnline || node.online)
     );
     const gerbilExitNodes = allExitNodes.filter(
         (node) =>

server/private/lib/logStreaming/LogStreamingManager.ts

@@ -19,7 +19,8 @@ import {
     requestAuditLog,
     actionAuditLog,
     accessAuditLog,
-    connectionAuditLog
+    connectionAuditLog,
+    aiSessionLog
 } from "@server/db";
 import logger from "@server/logger";
 import { and, eq, gt, desc, max, sql } from "drizzle-orm";
@@ -309,6 +310,7 @@ export class LogStreamingManager {
         if (dest.sendActionLogs) enabledTypes.push("action");
         if (dest.sendAccessLogs) enabledTypes.push("access");
         if (dest.sendConnectionLogs) enabledTypes.push("connection");
+        if (dest.sendAISessionLogs) enabledTypes.push("aiSession");
 
         if (enabledTypes.length === 0) return;
 
@@ -585,6 +587,13 @@ export class LogStreamingManager {
                         .where(eq(connectionAuditLog.orgId, orgId));
                     return row?.maxId ?? 0;
                 }
+                case "aiSession": {
+                    const [row] = await logsDb
+                        .select({ maxId: max(aiSessionLog.id) })
+                        .from(aiSessionLog)
+                        .where(eq(aiSessionLog.orgId, orgId));
+                    return row?.maxId ?? 0;
+                }
             }
         } catch (err) {
             logger.warn(
@@ -670,6 +679,21 @@ export class LogStreamingManager {
                     .limit(limit)) as Array<
                     Record<string, unknown> & { id: number }
                 >;
+
+            case "aiSession":
+                return (await logsDb
+                    .select()
+                    .from(aiSessionLog)
+                    .where(
+                        and(
+                            eq(aiSessionLog.orgId, orgId),
+                            gt(aiSessionLog.id, afterId)
+                        )
+                    )
+                    .orderBy(aiSessionLog.id)
+                    .limit(limit)) as Array<
+                    Record<string, unknown> & { id: number }
+                >;
         }
     }
 
@@ -694,6 +718,14 @@ export class LogStreamingManager {
                 timestamp =
                     typeof row.startedAt === "number" ? row.startedAt : 0;
                 break;
+            case "aiSession":
+                // createdAt is stored as epoch milliseconds; normalise to
+                // epoch seconds to match the other log types.
+                timestamp =
+                    typeof row.createdAt === "number"
+                        ? Math.floor(row.createdAt / 1000)
+                        : 0;
+                break;
         }
 
         const orgId = typeof row.orgId === "string" ? row.orgId : "";

server/private/lib/logStreaming/types.ts

@@ -15,13 +15,14 @@
 // Log type identifiers
 // ---------------------------------------------------------------------------
 
-export type LogType = "request" | "action" | "access" | "connection";
+export type LogType = "request" | "action" | "access" | "connection" | "aiSession";
 
 export const LOG_TYPES: LogType[] = [
     "request",
     "action",
     "access",
-    "connection"
+    "connection",
+    "aiSession"
 ];
 
 // ---------------------------------------------------------------------------

server/private/routers/aiGateway/logAiSession.ts

@@ -0,0 +1,288 @@
+/*
+ * This file is part of a proprietary work.
+ *
+ * Copyright (c) 2025-2026 Fossorial, Inc.
+ * All rights reserved.
+ *
+ * This file is licensed under the Fossorial Commercial License.
+ * You may not use this file except in compliance with the License.
+ * Unauthorized use, copying, modification, or distribution is strictly prohibited.
+ *
+ * This file is not licensed under the AGPLv3.
+ */
+
+import { logsDb, db, orgs, aiSessionLog, type AiProvider } from "@server/db";
+import type { InferInsertModel } from "drizzle-orm";
+import logger from "@server/logger";
+import { and, eq, lt } from "drizzle-orm";
+import cache from "#private/lib/cache";
+import { calculateCutoffTimestamp } from "@server/lib/cleanupLogs";
+import { sanitizeString } from "@server/lib/sanitize";
+import type { AiCapability } from "@server/lib/aiCapabilities";
+import {
+    normalizeAiRequest,
+    normalizeAiResponse
+} from "@server/lib/aiMessageNormalization";
+
+// Caps how much of the request/response body we keep per row, so a single
+// huge multimodal payload can't blow up buffer memory or storage.
+const AI_SESSION_LOG_MAX_BODY_CHARS = 200_000;
+
+type AiSessionLogInsert = InferInsertModel<typeof aiSessionLog>;
+
+// In-memory buffer for batching AI session log inserts, mirroring the
+// approach in server/routers/badger/logRequestAudit.ts.
+const sessionLogBuffer: AiSessionLogInsert[] = [];
+
+const BATCH_SIZE = 100; // Write to DB every 100 logs
+const BATCH_INTERVAL_MS = 5000; // Or every 5 seconds, whichever comes first
+const MAX_BUFFER_SIZE = 10000; // Prevent unbounded memory growth
+let flushTimer: NodeJS.Timeout | null = null;
+let isFlushInProgress = false;
+
+/**
+ * Flush buffered logs to database
+ */
+async function flushSessionLogs() {
+    if (sessionLogBuffer.length === 0 || isFlushInProgress) {
+        return;
+    }
+
+    isFlushInProgress = true;
+
+    // Take all current logs and clear buffer
+    const logsToWrite = sessionLogBuffer.splice(0, sessionLogBuffer.length);
+
+    try {
+        // Use a transaction to ensure all inserts succeed or fail together
+        await logsDb.transaction(async (tx) => {
+            // Batch insert logs in groups of 25 to avoid overwhelming the database
+            const BATCH_DB_SIZE = 25;
+            for (let i = 0; i < logsToWrite.length; i += BATCH_DB_SIZE) {
+                const batch = logsToWrite.slice(i, i + BATCH_DB_SIZE);
+                await tx.insert(aiSessionLog).values(batch);
+            }
+        });
+        logger.debug(
+            `Flushed ${logsToWrite.length} AI session logs to database`
+        );
+    } catch (error) {
+        logger.error("Error flushing AI session logs:", error);
+        // On transaction error, put logs back at the front of the buffer to retry
+        // but only if buffer isn't too large
+        if (sessionLogBuffer.length < MAX_BUFFER_SIZE - logsToWrite.length) {
+            sessionLogBuffer.unshift(...logsToWrite);
+            logger.info(
+                `Re-queued ${logsToWrite.length} AI session logs for retry`
+            );
+        } else {
+            logger.error(
+                `Buffer full, dropped ${logsToWrite.length} AI session logs`
+            );
+        }
+    } finally {
+        isFlushInProgress = false;
+        // If buffer filled up while we were flushing, flush again
+        if (sessionLogBuffer.length >= BATCH_SIZE) {
+            flushSessionLogs().catch((err) =>
+                logger.error("Error in follow-up AI session log flush:", err)
+            );
+        }
+    }
+}
+
+/**
+ * Schedule a flush if not already scheduled
+ */
+function scheduleFlush() {
+    if (flushTimer === null) {
+        flushTimer = setTimeout(() => {
+            flushTimer = null;
+            flushSessionLogs().catch((err) =>
+                logger.error("Error in scheduled AI session log flush:", err)
+            );
+        }, BATCH_INTERVAL_MS);
+    }
+}
+
+/**
+ * Gracefully flush all pending logs (call this on shutdown)
+ */
+export async function shutdownAiSessionLogger() {
+    if (flushTimer) {
+        clearTimeout(flushTimer);
+        flushTimer = null;
+    }
+    // Force flush even if one is in progress by waiting and retrying
+    while (isFlushInProgress) {
+        await new Promise((resolve) => setTimeout(resolve, 100));
+    }
+    await flushSessionLogs();
+}
+
+async function getRetentionDays(orgId: string): Promise<number> {
+    // check cache first
+    const cached = await cache.get<number>(`org_${orgId}_aiSessionsDays`);
+    if (cached !== undefined) {
+        return cached;
+    }
+
+    const [org] = await db
+        .select({
+            settingsLogRetentionDaysAISessions:
+                orgs.settingsLogRetentionDaysAISessions
+        })
+        .from(orgs)
+        .where(eq(orgs.orgId, orgId))
+        .limit(1);
+
+    if (!org) {
+        return 0;
+    }
+
+    // store the result in cache
+    await cache.set(
+        `org_${orgId}_aiSessionsDays`,
+        org.settingsLogRetentionDaysAISessions,
+        300
+    );
+
+    return org.settingsLogRetentionDaysAISessions;
+}
+
+export async function cleanUpOldLogs(orgId: string, retentionDays: number) {
+    // calculateCutoffTimestamp returns a seconds-epoch cutoff (built for
+    // requestAuditLog.timestamp), but aiSessionLog.createdAt is ms-epoch to
+    // match aiUsageRecords - convert before comparing.
+    const cutoffTimestampMs = calculateCutoffTimestamp(retentionDays) * 1000;
+
+    try {
+        await logsDb
+            .delete(aiSessionLog)
+            .where(
+                and(
+                    lt(aiSessionLog.createdAt, cutoffTimestampMs),
+                    eq(aiSessionLog.orgId, orgId)
+                )
+            );
+    } catch (error) {
+        logger.error("Error cleaning up old AI session logs:", error);
+    }
+}
+
+function truncateBody(value: string): { value: string; truncated: boolean } {
+    if (value.length <= AI_SESSION_LOG_MAX_BODY_CHARS) {
+        return { value, truncated: false };
+    }
+    return {
+        value: value.slice(0, AI_SESSION_LOG_MAX_BODY_CHARS),
+        truncated: true
+    };
+}
+
+export function logAiSession(data: {
+    sessionId: string;
+    capability: AiCapability;
+    provider: AiProvider;
+    requestedModel: string | undefined;
+    requestBody: unknown;
+    responseText: string;
+    isStream: boolean;
+    statusCode: number;
+    orgId: string | null;
+    resourceId: number | null;
+    siteResourceId: number | null;
+    requestUserId: string | null;
+    virtualApiKeyId: string | null;
+}): void {
+    (async () => {
+        try {
+            // Check retention before buffering any logs
+            if (data.orgId) {
+                const retentionDays = await getRetentionDays(data.orgId);
+                if (retentionDays === 0) {
+                    // do not log
+                    return;
+                }
+            } else {
+                // No org resolved for this request - nothing to govern
+                // retention with, so don't log it.
+                return;
+            }
+
+            const requestBodyText = truncateBody(
+                JSON.stringify(data.requestBody ?? "")
+            );
+            const responseBodyText = truncateBody(data.responseText ?? "");
+
+            // Uniform, capability-agnostic transcript for search/display -
+            // computed from the untruncated originals so normalization sees
+            // the full content; the normalized result gets its own
+            // (typically much smaller) truncation pass below.
+            const normalizedRequestMessages = normalizeAiRequest(
+                data.capability,
+                data.requestBody
+            );
+            const normalizedResponseMessages = normalizeAiResponse(
+                data.capability,
+                data.responseText ?? "",
+                data.isStream
+            );
+            const normalizedRequestText = normalizedRequestMessages
+                ? truncateBody(JSON.stringify(normalizedRequestMessages))
+                : null;
+            const normalizedResponseText = normalizedResponseMessages
+                ? truncateBody(JSON.stringify(normalizedResponseMessages))
+                : null;
+
+            // Prevent unbounded buffer growth - drop oldest entries if buffer is too large
+            if (sessionLogBuffer.length >= MAX_BUFFER_SIZE) {
+                const dropped = sessionLogBuffer.splice(0, BATCH_SIZE);
+                logger.warn(
+                    `AI session log buffer exceeded max size (${MAX_BUFFER_SIZE}), dropped ${dropped.length} oldest entries`
+                );
+            }
+
+            sessionLogBuffer.push({
+                sessionId: data.sessionId,
+                orgId: sanitizeString(data.orgId),
+                providerId: data.provider.providerId,
+                capability: data.capability,
+                resourceId: data.resourceId ?? undefined,
+                siteResourceId: data.siteResourceId ?? undefined,
+                userId: sanitizeString(data.requestUserId ?? undefined),
+                virtualApiKeyId: sanitizeString(
+                    data.virtualApiKeyId ?? undefined
+                ),
+                requestedModel: sanitizeString(data.requestedModel),
+                isStream: data.isStream,
+                requestBody: sanitizeString(requestBodyText.value),
+                responseBody: sanitizeString(responseBodyText.value),
+                normalizedRequest: normalizedRequestText
+                    ? sanitizeString(normalizedRequestText.value)
+                    : undefined,
+                normalizedResponse: normalizedResponseText
+                    ? sanitizeString(normalizedResponseText.value)
+                    : undefined,
+                truncated:
+                    requestBodyText.truncated ||
+                    responseBodyText.truncated ||
+                    (normalizedRequestText?.truncated ?? false) ||
+                    (normalizedResponseText?.truncated ?? false),
+                statusCode: data.statusCode,
+                createdAt: Date.now()
+            });
+
+            // Flush immediately if buffer is full, otherwise schedule a flush
+            if (sessionLogBuffer.length >= BATCH_SIZE) {
+                flushSessionLogs().catch((err) =>
+                    logger.error("Error flushing AI session logs:", err)
+                );
+            } else {
+                scheduleFlush();
+            }
+        } catch (error) {
+            logger.error("Failed to log AI session", { error });
+        }
+    })();
+}

server/private/routers/billing/featureLifecycle.ts

@@ -291,6 +291,10 @@ async function disableFeature(
                 await disableConnectionLogs(orgId);
                 break;
 
+            case TierFeature.AISessionLogs:
+                await disableAISessionLogs(orgId);
+                break;
+
             case TierFeature.RotateCredentials:
                 await disableRotateCredentials(orgId);
                 break;
@@ -493,6 +497,15 @@ async function disableConnectionLogs(orgId: string): Promise<void> {
     logger.info(`Disabled connection logs for org ${orgId}`);
 }
 
+async function disableAISessionLogs(orgId: string): Promise<void> {
+    await db
+        .update(orgs)
+        .set({ settingsLogRetentionDaysAISessions: 0 })
+        .where(eq(orgs.orgId, orgId));
+
+    logger.info(`Disabled AI session logs for org ${orgId}`);
+}
+
 async function disableRotateCredentials(orgId: string): Promise<void> {}
 
 async function disablemaintenancePage(orgId: string): Promise<void> {

server/private/routers/eventStreamingDestination/createEventStreamingDestination.ts

@@ -37,7 +37,8 @@ const bodySchema = z.strictObject({
     sendConnectionLogs: z.boolean().optional().default(false),
     sendRequestLogs: z.boolean().optional().default(false),
     sendActionLogs: z.boolean().optional().default(false),
-    sendAccessLogs: z.boolean().optional().default(false)
+    sendAccessLogs: z.boolean().optional().default(false),
+    sendAISessionLogs: z.boolean().optional().default(false)
 });
 
 export type CreateEventStreamingDestinationResponse = {
@@ -122,7 +123,8 @@ export async function createEventStreamingDestination(
                 sendAccessLogs: parsedBody.data.sendAccessLogs,
                 sendActionLogs: parsedBody.data.sendActionLogs,
                 sendConnectionLogs: parsedBody.data.sendConnectionLogs,
-                sendRequestLogs: parsedBody.data.sendRequestLogs
+                sendRequestLogs: parsedBody.data.sendRequestLogs,
+                sendAISessionLogs: parsedBody.data.sendAISessionLogs
             })
             .returning();
 

server/private/routers/eventStreamingDestination/listEventStreamingDestinations.ts

@@ -60,6 +60,7 @@ export type ListEventStreamingDestinationsResponse = {
         sendRequestLogs: boolean;
         sendActionLogs: boolean;
         sendAccessLogs: boolean;
+        sendAISessionLogs: boolean;
     }[];
     pagination: {
         total: number;
@@ -83,7 +84,8 @@ const ListEventStreamingDestinationsResponseDataSchema = z.object({
             sendConnectionLogs: z.boolean(),
             sendRequestLogs: z.boolean(),
             sendActionLogs: z.boolean(),
-            sendAccessLogs: z.boolean()
+            sendAccessLogs: z.boolean(),
+            sendAISessionLogs: z.boolean()
         })
     ),
     pagination: z.object({

server/private/routers/eventStreamingDestination/updateEventStreamingDestination.ts

@@ -40,7 +40,8 @@ const bodySchema = z.strictObject({
     sendConnectionLogs: z.boolean().optional(),
     sendRequestLogs: z.boolean().optional(),
     sendActionLogs: z.boolean().optional(),
-    sendAccessLogs: z.boolean().optional()
+    sendAccessLogs: z.boolean().optional(),
+    sendAISessionLogs: z.boolean().optional()
 });
 
 export type UpdateEventStreamingDestinationResponse = {
@@ -125,7 +126,7 @@ export async function updateEventStreamingDestination(
             );
         }
 
-        const { type, config: configToUpdate, enabled, sendAccessLogs, sendActionLogs, sendConnectionLogs, sendRequestLogs } = parsedBody.data;
+        const { type, config: configToUpdate, enabled, sendAccessLogs, sendActionLogs, sendConnectionLogs, sendRequestLogs, sendAISessionLogs } = parsedBody.data;
 
         const updateData: Record<string, unknown> = {
             updatedAt: Date.now()
@@ -141,6 +142,7 @@ export async function updateEventStreamingDestination(
         if (sendActionLogs !== undefined) updateData.sendActionLogs = sendActionLogs;
         if (sendConnectionLogs !== undefined) updateData.sendConnectionLogs = sendConnectionLogs;
         if (sendRequestLogs !== undefined) updateData.sendRequestLogs = sendRequestLogs;
+        if (sendAISessionLogs !== undefined) updateData.sendAISessionLogs = sendAISessionLogs;
 
         await db
             .update(eventStreamingDestinations)

server/private/routers/external.ts

@@ -21,6 +21,10 @@ import * as auth from "#private/routers/auth";
 import * as license from "#private/routers/license";
 import * as generateLicense from "#private/routers/generatedLicense";
 import * as logs from "#private/routers/auditLogs";
+import {
+    queryAiSessionLogs,
+    exportAiSessionLogs
+} from "@server/routers/auditLogs";
 import * as misc from "#private/routers/misc";
 import * as reKey from "#private/routers/re-key";
 import * as approval from "#private/routers/approvals";
@@ -591,6 +595,25 @@ authenticated.get(
     logs.exportConnectionAuditLogs
 );
 
+authenticated.get(
+    "/org/:orgId/logs/ai",
+    verifyValidLicense,
+    verifyValidSubscription(tierMatrix.aiSessionLogs),
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.viewLogs),
+    queryAiSessionLogs
+);
+
+authenticated.get(
+    "/org/:orgId/logs/ai/export",
+    verifyValidLicense,
+    verifyValidSubscription(tierMatrix.aiSessionLogs),
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.exportLogs),
+    logActionAudit(ActionsEnum.exportLogs),
+    exportAiSessionLogs
+);
+
 authenticated.post(
     "/re-key/:clientId/regenerate-client-secret",
     verifyClientAccess, // this is first to set the org id

server/private/routers/gerbil/createExitNode.ts

@@ -34,10 +34,6 @@ export async function createExitNode(
             // TODO: eventually we will want to get the next available port so that we can multiple exit nodes
             // const listenPort = await getNextAvailablePort();
             const listenPort = config.getRawConfig().gerbil.start_port;
-            let subEndpoint = "";
-            if (config.getRawConfig().gerbil.use_subdomain) {
-                subEndpoint = await getUniqueExitNodeEndpointName();
-            }
 
             const exitNodeName =
                 config.getRawConfig().gerbil.exit_node_name ||
@@ -48,7 +44,7 @@ export async function createExitNode(
                 .insert(exitNodes)
                 .values({
                     publicKey,
-                    endpoint: `${subEndpoint}${subEndpoint != "" ? "." : ""}${config.getRawConfig().gerbil.base_endpoint}`,
+                    endpoint: config.getRawConfig().gerbil.base_endpoint,
                     address,
                     listenPort,
                     online: true,

server/private/routers/integration.ts

@@ -43,6 +43,10 @@ import {
     unauthenticated as ua,
     authenticated as a
 } from "@server/routers/integration";
+import {
+    queryAiSessionLogs,
+    exportAiSessionLogs
+} from "@server/routers/auditLogs";
 import { logActionAudit } from "#private/middlewares";
 import { tierMatrix } from "@server/lib/billing/tierMatrix";
 import { build } from "@server/build";
@@ -153,6 +157,25 @@ authenticated.get(
     logs.exportConnectionAuditLogs
 );
 
+authenticated.get(
+    "/org/:orgId/logs/ai",
+    verifyValidLicense,
+    verifyValidSubscription(tierMatrix.aiSessionLogs),
+    verifyApiKeyOrgAccess,
+    verifyApiKeyHasAction(ActionsEnum.viewLogs),
+    queryAiSessionLogs
+);
+
+authenticated.get(
+    "/org/:orgId/logs/ai/export",
+    verifyValidLicense,
+    verifyValidSubscription(tierMatrix.aiSessionLogs),
+    verifyApiKeyOrgAccess,
+    verifyApiKeyHasAction(ActionsEnum.exportLogs),
+    logActionAudit(ActionsEnum.exportLogs),
+    exportAiSessionLogs
+);
+
 authenticated.put(
     "/org/:orgId/idp/oidc",
     verifyValidLicense,

server/private/routers/newt/handleRequestLogMessage.ts

@@ -1,238 +0,0 @@
-/*
- * This file is part of a proprietary work.
- *
- * Copyright (c) 2025-2026 Fossorial, Inc.
- * All rights reserved.
- *
- * This file is licensed under the Fossorial Commercial License.
- * You may not use this file except in compliance with the License.
- * Unauthorized use, copying, modification, or distribution is strictly prohibited.
- *
- * This file is not licensed under the AGPLv3.
- */
-
-import { db } from "@server/db";
-import { MessageHandler } from "@server/routers/ws";
-import { sites, Newt, orgs, clients, clientSitesAssociationsCache } from "@server/db";
-import { and, eq, inArray } from "drizzle-orm";
-import logger from "@server/logger";
-import { inflate } from "zlib";
-import { promisify } from "util";
-import { logRequestAudit } from "@server/routers/badger/logRequestAudit";
-import { getCountryCodeForIp } from "@server/lib/geoip";
-
-export async function flushRequestLogToDb(): Promise<void> {
-    return;
-}
-
-const zlibInflate = promisify(inflate);
-
-interface HTTPRequestLogData {
-    requestId: string;
-    resourceId: number; // siteResourceId
-    timestamp: string; // ISO 8601
-    method: string;
-    scheme: string; // "http" or "https"
-    host: string;
-    path: string;
-    rawQuery?: string;
-    userAgent?: string;
-    sourceAddr: string; // ip:port
-    tls: boolean;
-}
-
-/**
- * Decompress a base64-encoded zlib-compressed string into parsed JSON.
- */
-async function decompressRequestLog(
-    compressed: string
-): Promise<HTTPRequestLogData[]> {
-    const compressedBuffer = Buffer.from(compressed, "base64");
-    const decompressed = await zlibInflate(compressedBuffer);
-    const jsonString = decompressed.toString("utf-8");
-    const parsed = JSON.parse(jsonString);
-
-    if (!Array.isArray(parsed)) {
-        throw new Error("Decompressed request log data is not an array");
-    }
-
-    return parsed;
-}
-
-export const handleRequestLogMessage: MessageHandler = async (context) => {
-    const { message, client } = context;
-    const newt = client as Newt;
-
-    if (!newt) {
-        logger.warn("Request log received but no newt client in context");
-        return;
-    }
-
-    if (!newt.siteId) {
-        logger.warn("Request log received but newt has no siteId");
-        return;
-    }
-
-    if (!message.data?.compressed) {
-        logger.warn("Request log message missing compressed data");
-        return;
-    }
-
-    // Look up the org for this site and check retention settings
-    const [site] = await db
-        .select({
-            orgId: sites.orgId,
-            orgSubnet: orgs.subnet,
-            settingsLogRetentionDaysRequest:
-                orgs.settingsLogRetentionDaysRequest
-        })
-        .from(sites)
-        .innerJoin(orgs, eq(sites.orgId, orgs.orgId))
-        .where(eq(sites.siteId, newt.siteId));
-
-    if (!site) {
-        logger.warn(
-            `Request log received but site ${newt.siteId} not found in database`
-        );
-        return;
-    }
-
-    const orgId = site.orgId;
-
-    if (site.settingsLogRetentionDaysRequest === 0) {
-        logger.debug(
-            `Request log retention is disabled for org ${orgId}, skipping`
-        );
-        return;
-    }
-
-    let entries: HTTPRequestLogData[];
-    try {
-        entries = await decompressRequestLog(message.data.compressed);
-    } catch (error) {
-        logger.error("Failed to decompress request log data:", error);
-        return;
-    }
-
-    if (entries.length === 0) {
-        return;
-    }
-
-    logger.debug(`Request log entries: ${JSON.stringify(entries)}`);
-
-    // Build a map from sourceIp → external endpoint string by joining clients
-    // with clientSitesAssociationsCache. The endpoint is the real-world IP:port
-    // of the client device and is used for GeoIP lookup.
-    const ipToEndpoint = new Map<string, string>();
-
-    const cidrSuffix = site.orgSubnet?.includes("/")
-        ? site.orgSubnet.substring(site.orgSubnet.indexOf("/"))
-        : null;
-
-    if (cidrSuffix) {
-        const uniqueSourceAddrs = new Set<string>();
-        for (const entry of entries) {
-            if (entry.sourceAddr) {
-                uniqueSourceAddrs.add(entry.sourceAddr);
-            }
-        }
-
-        if (uniqueSourceAddrs.size > 0) {
-            const subnetQueries = Array.from(uniqueSourceAddrs).map((addr) => {
-                const ip = addr.includes(":") ? addr.split(":")[0] : addr;
-                return `${ip}${cidrSuffix}`;
-            });
-
-            const matchedClients = await db
-                .select({
-                    subnet: clients.subnet,
-                    endpoint: clientSitesAssociationsCache.endpoint
-                })
-                .from(clients)
-                .innerJoin(
-                    clientSitesAssociationsCache,
-                    and(
-                        eq(
-                            clientSitesAssociationsCache.clientId,
-                            clients.clientId
-                        ),
-                        eq(clientSitesAssociationsCache.siteId, newt.siteId)
-                    )
-                )
-                .where(
-                    and(
-                        eq(clients.orgId, orgId),
-                        inArray(clients.subnet, subnetQueries)
-                    )
-                );
-
-            for (const c of matchedClients) {
-                if (c.endpoint) {
-                    const ip = c.subnet.split("/")[0];
-                    ipToEndpoint.set(ip, c.endpoint);
-                }
-            }
-        }
-    }
-
-    for (const entry of entries) {
-        if (
-            !entry.requestId ||
-            !entry.resourceId ||
-            !entry.method ||
-            !entry.scheme ||
-            !entry.host ||
-            !entry.path ||
-            !entry.sourceAddr
-        ) {
-            logger.debug(
-                `Skipping request log entry with missing required fields: ${JSON.stringify(entry)}`
-            );
-            continue;
-        }
-
-        const originalRequestURL =
-            entry.scheme +
-            "://" +
-            entry.host +
-            entry.path +
-            (entry.rawQuery ? "?" + entry.rawQuery : "");
-
-        // Resolve the client's external endpoint for GeoIP lookup.
-        // sourceAddr is the WireGuard IP (possibly ip:port), so strip the port.
-        const sourceIp = entry.sourceAddr.includes(":")
-            ? entry.sourceAddr.split(":")[0]
-            : entry.sourceAddr;
-        const endpoint = ipToEndpoint.get(sourceIp);
-        let location: string | undefined;
-        if (endpoint) {
-            const endpointIp = endpoint.includes(":")
-                ? endpoint.split(":")[0]
-                : endpoint;
-            location = await getCountryCodeForIp(endpointIp);
-        }
-
-        await logRequestAudit(
-            {
-                action: true,
-                reason: 108,
-                siteResourceId: entry.resourceId,
-                orgId,
-                location
-            },
-            {
-                path: entry.path,
-                originalRequestURL,
-                scheme: entry.scheme,
-                host: entry.host,
-                method: entry.method,
-                tls: entry.tls,
-                requestIp: entry.sourceAddr
-            }
-        );
-    }
-
-    logger.debug(
-        `Buffered ${entries.length} request log entry/entries from newt ${newt.newtId} (site ${newt.siteId})`
-    );
-};

server/private/routers/newt/index.ts

@@ -12,4 +12,3 @@
  */
 
 export * from "./handleConnectionLogMessage";
-export * from "./handleRequestLogMessage";

server/private/routers/remoteExitNode/createRemoteExitNode.ts

@@ -191,13 +191,20 @@ export async function createRemoteExitNode(
 
         // If this remote exit node isn't already backing an exit node in
         // another org, we're about to create a brand new one. Reserve a
-        // subnet for it up front so the allocation lock is held across the
-        // whole insert - this guarantees exit node subnets never overlap,
-        // even under concurrent creation, which matters for HA setups.
+        // subnet for it up front, scoped to this org's existing exit nodes,
+        // so the allocation lock is held across the whole insert - this
+        // guarantees exit node subnets never overlap within the org, even
+        // under concurrent creation, which matters for HA setups. Subnets
+        // may still be reused across different orgs; there isn't enough
+        // address space to avoid that, and it isn't necessary since HA only
+        // routes multiple exit nodes for the same org.
         let releaseSubnetLock: (() => Promise<void>) | null = null;
         let newExitNodeAddress: string | null = null;
         if (!existingExitNode) {
-            const { value, release } = await getNextAvailableSubnet();
+            const { value, release } = await getNextAvailableSubnet(
+                db,
+                orgId
+            );
             newExitNodeAddress = value;
             releaseSubnetLock = release;
         }

server/private/routers/ws/messageHandlers.ts

@@ -18,12 +18,10 @@ import {
 import { MessageHandler } from "@server/routers/ws";
 import {
     handleConnectionLogMessage,
-    handleRequestLogMessage
 } from "#private/routers/newt";
 
 export const messageHandlers: Record<string, MessageHandler> = {
     "remoteExitNode/register": handleRemoteExitNodeRegisterMessage,
     "remoteExitNode/ping": handleRemoteExitNodePingMessage,
     "newt/access-log": handleConnectionLogMessage,
-    "newt/request-log": handleRequestLogMessage
 };

server/private/routers/ws/ws.ts

@@ -139,7 +139,7 @@ const processMessage = async (
             }
         }
     } catch (error) {
-        logger.error("Message handling error:", error);
+        logger.warn("Message handling error:", error);
         // ws.send(JSON.stringify({
         //     type: "error",
         //     data: {

server/routers/aiGateway/createAiGatewayRouter.ts

@@ -1,19 +1,37 @@
-import { Router } from "express";
+import { Router, type Request, type Response } from "express";
 import {
     AI_CAPABILITY_DEFS,
     type AiCapability
 } from "@server/lib/aiCapabilities";
 import { handleAiGatewayProxy } from "@server/routers/aiGateway/pipeline";
+import { handleV1Models } from "@server/routers/aiGateway";
+
+type CapabilityHandler = (
+    req: Request,
+    res: Response,
+    capability: AiCapability
+) => Promise<any>;
+
+// Capabilities the gateway answers itself instead of proxying upstream.
+// Everything else goes through the inference pipeline.
+const LOCAL_HANDLERS: Partial<Record<AiCapability, CapabilityHandler>> = {
+    v1_models: handleV1Models
+};
 
 export function createAiGatewayRouter() {
     const router = Router();
 
     for (const def of Object.values(AI_CAPABILITY_DEFS)) {
         const capability = def.id as AiCapability;
+        const handler = LOCAL_HANDLERS[capability] ?? handleAiGatewayProxy;
         for (const route of def.routes) {
-            router.post(route.path, (req, res) =>
-                handleAiGatewayProxy(req, res, capability)
-            );
+            const bind = (req: Request, res: Response) =>
+                handler(req, res, capability);
+            if (route.method === "GET") {
+                router.get(route.path, bind);
+            } else {
+                router.post(route.path, bind);
+            }
         }
     }
 

server/routers/aiGateway/index.ts

@@ -1,2 +1,3 @@
 export { handleAiGatewayProxy } from "./pipeline";
+export { handleV1Models } from "./v1Models";
 export { createAiGatewayRouter } from "./createAiGatewayRouter";

server/routers/aiGateway/logAiSession.ts

@@ -1,171 +1,12 @@
-import { logsDb, db, orgs, aiSessionLog, type AiProvider } from "@server/db";
-import type { InferInsertModel } from "drizzle-orm";
-import logger from "@server/logger";
-import { and, eq, lt } from "drizzle-orm";
-import cache from "#dynamic/lib/cache";
-import { calculateCutoffTimestamp } from "@server/lib/cleanupLogs";
-import { sanitizeString } from "@server/lib/sanitize";
-import type { AiCapability } from "@server/lib/aiCapabilities";
-import {
-    normalizeAiRequest,
-    normalizeAiResponse
-} from "@server/lib/aiMessageNormalization";
-
-// Caps how much of the request/response body we keep per row, so a single
-// huge multimodal payload can't blow up buffer memory or storage.
-const AI_SESSION_LOG_MAX_BODY_CHARS = 200_000;
-
-type AiSessionLogInsert = InferInsertModel<typeof aiSessionLog>;
-
-// In-memory buffer for batching AI session log inserts, mirroring the
-// approach in server/routers/badger/logRequestAudit.ts.
-const sessionLogBuffer: AiSessionLogInsert[] = [];
-
-const BATCH_SIZE = 100; // Write to DB every 100 logs
-const BATCH_INTERVAL_MS = 5000; // Or every 5 seconds, whichever comes first
-const MAX_BUFFER_SIZE = 10000; // Prevent unbounded memory growth
-let flushTimer: NodeJS.Timeout | null = null;
-let isFlushInProgress = false;
-
-/**
- * Flush buffered logs to database
- */
-async function flushSessionLogs() {
-    if (sessionLogBuffer.length === 0 || isFlushInProgress) {
-        return;
-    }
-
-    isFlushInProgress = true;
-
-    // Take all current logs and clear buffer
-    const logsToWrite = sessionLogBuffer.splice(0, sessionLogBuffer.length);
-
-    try {
-        // Use a transaction to ensure all inserts succeed or fail together
-        await logsDb.transaction(async (tx) => {
-            // Batch insert logs in groups of 25 to avoid overwhelming the database
-            const BATCH_DB_SIZE = 25;
-            for (let i = 0; i < logsToWrite.length; i += BATCH_DB_SIZE) {
-                const batch = logsToWrite.slice(i, i + BATCH_DB_SIZE);
-                await tx.insert(aiSessionLog).values(batch);
-            }
-        });
-        logger.debug(
-            `Flushed ${logsToWrite.length} AI session logs to database`
-        );
-    } catch (error) {
-        logger.error("Error flushing AI session logs:", error);
-        // On transaction error, put logs back at the front of the buffer to retry
-        // but only if buffer isn't too large
-        if (sessionLogBuffer.length < MAX_BUFFER_SIZE - logsToWrite.length) {
-            sessionLogBuffer.unshift(...logsToWrite);
-            logger.info(
-                `Re-queued ${logsToWrite.length} AI session logs for retry`
-            );
-        } else {
-            logger.error(
-                `Buffer full, dropped ${logsToWrite.length} AI session logs`
-            );
-        }
-    } finally {
-        isFlushInProgress = false;
-        // If buffer filled up while we were flushing, flush again
-        if (sessionLogBuffer.length >= BATCH_SIZE) {
-            flushSessionLogs().catch((err) =>
-                logger.error("Error in follow-up AI session log flush:", err)
-            );
-        }
-    }
-}
-
-/**
- * Schedule a flush if not already scheduled
- */
-function scheduleFlush() {
-    if (flushTimer === null) {
-        flushTimer = setTimeout(() => {
-            flushTimer = null;
-            flushSessionLogs().catch((err) =>
-                logger.error("Error in scheduled AI session log flush:", err)
-            );
-        }, BATCH_INTERVAL_MS);
-    }
-}
+import { AiCapability } from "@app/lib/aiCapabilities";
+import { AiProvider } from "@server/db";
 
 /**
  * Gracefully flush all pending logs (call this on shutdown)
  */
-export async function shutdownAiSessionLogger() {
-    if (flushTimer) {
-        clearTimeout(flushTimer);
-        flushTimer = null;
-    }
-    // Force flush even if one is in progress by waiting and retrying
-    while (isFlushInProgress) {
-        await new Promise((resolve) => setTimeout(resolve, 100));
-    }
-    await flushSessionLogs();
-}
+export async function shutdownAiSessionLogger() {}
 
-async function getRetentionDays(orgId: string): Promise<number> {
-    // check cache first
-    const cached = await cache.get<number>(`org_${orgId}_aiSessionsDays`);
-    if (cached !== undefined) {
-        return cached;
-    }
-
-    const [org] = await db
-        .select({
-            settingsLogRetentionDaysAISessions:
-                orgs.settingsLogRetentionDaysAISessions
-        })
-        .from(orgs)
-        .where(eq(orgs.orgId, orgId))
-        .limit(1);
-
-    if (!org) {
-        return 0;
-    }
-
-    // store the result in cache
-    await cache.set(
-        `org_${orgId}_aiSessionsDays`,
-        org.settingsLogRetentionDaysAISessions,
-        300
-    );
-
-    return org.settingsLogRetentionDaysAISessions;
-}
-
-export async function cleanUpOldLogs(orgId: string, retentionDays: number) {
-    // calculateCutoffTimestamp returns a seconds-epoch cutoff (built for
-    // requestAuditLog.timestamp), but aiSessionLog.createdAt is ms-epoch to
-    // match aiUsageRecords - convert before comparing.
-    const cutoffTimestampMs = calculateCutoffTimestamp(retentionDays) * 1000;
-
-    try {
-        await logsDb
-            .delete(aiSessionLog)
-            .where(
-                and(
-                    lt(aiSessionLog.createdAt, cutoffTimestampMs),
-                    eq(aiSessionLog.orgId, orgId)
-                )
-            );
-    } catch (error) {
-        logger.error("Error cleaning up old AI session logs:", error);
-    }
-}
-
-function truncateBody(value: string): { value: string; truncated: boolean } {
-    if (value.length <= AI_SESSION_LOG_MAX_BODY_CHARS) {
-        return { value, truncated: false };
-    }
-    return {
-        value: value.slice(0, AI_SESSION_LOG_MAX_BODY_CHARS),
-        truncated: true
-    };
-}
+export async function cleanUpOldLogs(orgId: string, retentionDays: number) {}
 
 export function logAiSession(data: {
     sessionId: string;
@@ -181,95 +22,4 @@ export function logAiSession(data: {
     siteResourceId: number | null;
     requestUserId: string | null;
     virtualApiKeyId: string | null;
-}): void {
-    (async () => {
-        try {
-            // Check retention before buffering any logs
-            if (data.orgId) {
-                const retentionDays = await getRetentionDays(data.orgId);
-                if (retentionDays === 0) {
-                    // do not log
-                    return;
-                }
-            } else {
-                // No org resolved for this request - nothing to govern
-                // retention with, so don't log it.
-                return;
-            }
-
-            const requestBodyText = truncateBody(
-                JSON.stringify(data.requestBody ?? "")
-            );
-            const responseBodyText = truncateBody(data.responseText ?? "");
-
-            // Uniform, capability-agnostic transcript for search/display -
-            // computed from the untruncated originals so normalization sees
-            // the full content; the normalized result gets its own
-            // (typically much smaller) truncation pass below.
-            const normalizedRequestMessages = normalizeAiRequest(
-                data.capability,
-                data.requestBody
-            );
-            const normalizedResponseMessages = normalizeAiResponse(
-                data.capability,
-                data.responseText ?? "",
-                data.isStream
-            );
-            const normalizedRequestText = normalizedRequestMessages
-                ? truncateBody(JSON.stringify(normalizedRequestMessages))
-                : null;
-            const normalizedResponseText = normalizedResponseMessages
-                ? truncateBody(JSON.stringify(normalizedResponseMessages))
-                : null;
-
-            // Prevent unbounded buffer growth - drop oldest entries if buffer is too large
-            if (sessionLogBuffer.length >= MAX_BUFFER_SIZE) {
-                const dropped = sessionLogBuffer.splice(0, BATCH_SIZE);
-                logger.warn(
-                    `AI session log buffer exceeded max size (${MAX_BUFFER_SIZE}), dropped ${dropped.length} oldest entries`
-                );
-            }
-
-            sessionLogBuffer.push({
-                sessionId: data.sessionId,
-                orgId: sanitizeString(data.orgId),
-                providerId: data.provider.providerId,
-                capability: data.capability,
-                resourceId: data.resourceId ?? undefined,
-                siteResourceId: data.siteResourceId ?? undefined,
-                userId: sanitizeString(data.requestUserId ?? undefined),
-                virtualApiKeyId: sanitizeString(
-                    data.virtualApiKeyId ?? undefined
-                ),
-                requestedModel: sanitizeString(data.requestedModel),
-                isStream: data.isStream,
-                requestBody: sanitizeString(requestBodyText.value),
-                responseBody: sanitizeString(responseBodyText.value),
-                normalizedRequest: normalizedRequestText
-                    ? sanitizeString(normalizedRequestText.value)
-                    : undefined,
-                normalizedResponse: normalizedResponseText
-                    ? sanitizeString(normalizedResponseText.value)
-                    : undefined,
-                truncated:
-                    requestBodyText.truncated ||
-                    responseBodyText.truncated ||
-                    (normalizedRequestText?.truncated ?? false) ||
-                    (normalizedResponseText?.truncated ?? false),
-                statusCode: data.statusCode,
-                createdAt: Date.now()
-            });
-
-            // Flush immediately if buffer is full, otherwise schedule a flush
-            if (sessionLogBuffer.length >= BATCH_SIZE) {
-                flushSessionLogs().catch((err) =>
-                    logger.error("Error flushing AI session logs:", err)
-                );
-            } else {
-                scheduleFlush();
-            }
-        } catch (error) {
-            logger.error("Failed to log AI session", { error });
-        }
-    })();
-}
+}): void {}

server/routers/aiGateway/pipeline.ts

@@ -86,7 +86,7 @@ import {
     type AiUsage
 } from "@server/lib/aiUsageExtraction";
 import { streamAiGatewayResponse } from "@server/routers/aiGateway/streamAiGatewayResponse";
-import { logAiSession } from "@server/routers/aiGateway/logAiSession";
+import { logAiSession } from "#dynamic/routers/aiGateway/logAiSession";
 
 const EXIT_NODE_RANGES_CACHE_KEY = "aiGateway:exitNodeRanges";
 const EXIT_NODE_RANGES_TTL_SEC = 6000;
@@ -137,7 +137,7 @@ async function findClientByIp(ip: string): Promise<CachedClient> {
     return result;
 }
 
-type ProviderAttachment = {
+export type ProviderAttachment = {
     provider: AiProvider;
     accessMode: AccessMode;
 };
@@ -149,12 +149,12 @@ type ResourceModelPattern = {
     enabled: boolean;
 };
 
-type ProviderPatternLists = {
+export type ProviderPatternLists = {
     allows: string[];
     blocks: string[];
 };
 
-type ResolvedTarget = {
+export type ResolvedTarget = {
     resourceId: number | null;
     siteResourceId: number | null;
     orgId: string | null;
@@ -362,7 +362,7 @@ function getRequestHeader(req: Request, name: string): string | undefined {
 // request came through, per the trust middleware's resource-type header -
 // falls back to checking both (public preferred on overlap) only when that
 // header is absent, e.g. a request that reached the gateway outside Traefik.
-async function resolveTarget(
+export async function resolveTarget(
     host: string,
     resourceType: AiGatewayResourceType | null
 ): Promise<ResolvedTarget | null> {
@@ -728,7 +728,9 @@ export function recordAiGatewayCompletion(args: {
     let cost: ReturnType<typeof calculateAiCost> = null;
 
     if (upstreamSucceeded) {
-        usage = extractUsage(capability, responseText, isStream, headers) ?? emptyUsage();
+        usage =
+            extractUsage(capability, responseText, isStream, headers) ??
+            emptyUsage();
         if (isUsageEmpty(usage)) {
             usage = estimateUsage(
                 JSON.stringify(requestBody ?? ""),
@@ -810,6 +812,17 @@ export function recordAiGatewayCompletion(args: {
     });
 }
 
+// p-host is only used sometimes when overriding the host header for some
+// middleware proxy. Shared with the model-discovery endpoint so both resolve
+// the inference resource off the same hostname.
+export function resolveGatewayHost(req: Request): string {
+    return (
+        (req.headers["p-host"] as string | undefined) ||
+        req.headers.host ||
+        ""
+    ).split(":")[0];
+}
+
 export async function handleAiGatewayProxy(
     req: Request,
     res: Response,
@@ -818,11 +831,7 @@ export async function handleAiGatewayProxy(
     try {
         const def = AI_CAPABILITY_DEFS[capability];
 
-        const host = (
-            (req.headers["p-host"] as string | undefined) || // p-host is only used sometimes when overriding the host header for some middleware proxy
-            req.headers.host ||
-            ""
-        ).split(":")[0];
+        const host = resolveGatewayHost(req);
         if (!host) {
             return res
                 .status(HttpCode.BAD_REQUEST)

server/routers/aiGateway/targetRouting.ts

@@ -99,6 +99,7 @@ async function fetchProviderTargets(
             method: targets.method,
             exitNodeSubnet: sites.exitNodeSubnet,
             reachableAt: exitNodes.reachableAt,
+            exitNodeType: exitNodes.type,
             hcHealth: targetHealthCheck.hcHealth
         })
         .from(targets)
@@ -119,6 +120,12 @@ async function fetchProviderTargets(
         if (!row.exitNodeSubnet || !row.reachableAt) {
             continue;
         }
+        // Sites connected to a remote exit node aren't reachable via a
+        // gerbil sidecar's /router/* proxy - only "gerbil" type exit nodes
+        // run that endpoint.
+        if (row.exitNodeType !== "gerbil") {
+            continue;
+        }
         // A target with an active health check that's currently failing is
         // taken out of rotation. No health check (null) or "unknown" (check
         // hasn't run yet / hcEnabled is off) still routes normally, matching

server/routers/aiGateway/v1Models.ts

@@ -0,0 +1,308 @@
+import { Request, Response } from "express";
+import { inArray } from "drizzle-orm";
+import { z } from "zod";
+import { aiModels, db } from "@server/db";
+import {
+    providerHasCapability,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import {
+    buildAiCapabilityErrorBody,
+    type AiCapabilityErrorKind
+} from "@server/lib/aiGatewayAuthError";
+import {
+    getAiGatewayResourceType,
+    isAiGatewayTrustHeaderValid
+} from "@server/lib/aiGatewayTrust";
+import { resolveEffectiveLists } from "@server/lib/aiInferenceResource";
+import { listCatalogEntriesForType } from "@server/lib/aiModelCatalog";
+import {
+    listPermittedModels,
+    paginateModels,
+    MODEL_PAGE_DEFAULT_LIMIT,
+    MODEL_PAGE_MAX_LIMIT,
+    type CatalogModelMetadata,
+    type ConfiguredModel,
+    type ModelDiscoveryProvider
+} from "@server/lib/aiModelDiscovery";
+import type { AiProviderType } from "@server/lib/aiProviderDefaults";
+import {
+    resolveGatewayHost,
+    resolveTarget,
+    type ProviderAttachment,
+    type ProviderPatternLists
+} from "@server/routers/aiGateway/pipeline";
+import logger from "@server/logger";
+import HttpCode from "@server/types/HttpCode";
+
+const CAPABILITY: AiCapability = "v1_models";
+
+const querySchema = z.object({
+    limit: z.coerce.number().int().min(1).max(MODEL_PAGE_MAX_LIMIT).optional(),
+    after_id: z.string().min(1).optional(),
+    before_id: z.string().min(1).optional()
+});
+
+type ProviderModelLists = {
+    allowsByProvider: Map<number, string[]>;
+    blocksByProvider: Map<number, string[]>;
+    configuredByProvider: Map<number, Map<string, ConfiguredModel>>;
+};
+
+function errorResponse(
+    res: Response,
+    status: number,
+    kind: AiCapabilityErrorKind,
+    message: string
+) {
+    return res
+        .status(status)
+        .json(buildAiCapabilityErrorBody(CAPABILITY, kind, message, status));
+}
+
+// Provider-level allow/block lists, plus the display name and creation time of
+// every catalog row, so explicitly configured models are reported with the name
+// the administrator gave them rather than a bare model id.
+async function loadProviderModelLists(
+    providerIds: number[]
+): Promise<ProviderModelLists> {
+    const lists: ProviderModelLists = {
+        allowsByProvider: new Map(),
+        blocksByProvider: new Map(),
+        configuredByProvider: new Map()
+    };
+
+    if (providerIds.length === 0) {
+        return lists;
+    }
+
+    const rows = await db
+        .select({
+            providerId: aiModels.providerId,
+            modelKey: aiModels.modelKey,
+            name: aiModels.name,
+            listType: aiModels.listType,
+            enabled: aiModels.enabled,
+            createdAt: aiModels.createdAt
+        })
+        .from(aiModels)
+        .where(inArray(aiModels.providerId, providerIds));
+
+    for (const row of rows) {
+        if (!row.enabled) {
+            continue;
+        }
+        const targetMap =
+            row.listType === "allow"
+                ? lists.allowsByProvider
+                : lists.blocksByProvider;
+        const existing = targetMap.get(row.providerId) ?? [];
+        existing.push(row.modelKey);
+        targetMap.set(row.providerId, existing);
+
+        let configured = lists.configuredByProvider.get(row.providerId);
+        if (!configured) {
+            configured = new Map();
+            lists.configuredByProvider.set(row.providerId, configured);
+        }
+        configured.set(row.modelKey, {
+            name: row.name,
+            createdAt: row.createdAt
+        });
+    }
+
+    return lists;
+}
+
+function catalogMetadataForType(
+    type: AiProviderType
+): Map<string, CatalogModelMetadata> {
+    const metadata = new Map<string, CatalogModelMetadata>();
+    for (const entry of listCatalogEntriesForType(type)) {
+        metadata.set(entry.model, {
+            maxInputTokens: entry.limits.input,
+            maxOutputTokens: entry.limits.output,
+            capabilities: entry.capabilities
+        });
+    }
+    return metadata;
+}
+
+function buildDiscoveryProviders(
+    attachments: ProviderAttachment[],
+    resourceListsByProvider: Map<number, ProviderPatternLists>,
+    lists: ProviderModelLists
+): ModelDiscoveryProvider[] {
+    return attachments.map((attachment) => {
+        const providerId = attachment.provider.providerId;
+        const resourceLists = resourceListsByProvider.get(providerId);
+        const { allows, blocks } = resolveEffectiveLists({
+            accessMode: attachment.accessMode,
+            providerAllows: lists.allowsByProvider.get(providerId) ?? [],
+            providerBlocks: lists.blocksByProvider.get(providerId) ?? [],
+            resourceAllows: resourceLists?.allows ?? [],
+            resourceBlocks: resourceLists?.blocks ?? []
+        });
+
+        return {
+            providerId,
+            allows,
+            blocks,
+            catalog: catalogMetadataForType(
+                attachment.provider.type as AiProviderType
+            ),
+            configured: lists.configuredByProvider.get(providerId) ?? new Map()
+        };
+    });
+}
+
+/**
+ * Serves Anthropic's model-discovery endpoints (`GET /v1/models` and
+ * `GET /v1/models/{id}`) for an inference resource. The gateway answers these
+ * itself rather than proxying: upstream providers either don't expose a model
+ * list at all or would expose models the resource's allow/block lists forbid,
+ * so the response is built from the same effective lists that gate inference.
+ */
+export async function handleV1Models(
+    req: Request,
+    res: Response
+): Promise<any> {
+    try {
+        const host = resolveGatewayHost(req);
+        if (!host) {
+            return errorResponse(
+                res,
+                HttpCode.BAD_REQUEST,
+                "invalid_request",
+                "Missing Host header"
+            );
+        }
+
+        const resourceType = getAiGatewayResourceType(
+            req.headers as Record<string, string>
+        );
+        const target = await resolveTarget(host, resourceType);
+        if (!target) {
+            return errorResponse(
+                res,
+                HttpCode.NOT_FOUND,
+                "not_found",
+                "No inference resource found for this host"
+            );
+        }
+
+        // Same gate as the inference pipeline: public inference must pass
+        // Badger verify-session first, which is what stamps the trust header.
+        if (
+            target.resourceId != null &&
+            !isAiGatewayTrustHeaderValid(req.headers as Record<string, string>)
+        ) {
+            return errorResponse(
+                res,
+                HttpCode.UNAUTHORIZED,
+                "authentication",
+                "Request must be authenticated via the inference resource"
+            );
+        }
+
+        if (target.attachments.length === 0) {
+            return errorResponse(
+                res,
+                HttpCode.FORBIDDEN,
+                "permission",
+                "No AI providers configured for this resource"
+            );
+        }
+
+        const capableAttachments = target.attachments.filter((a) =>
+            providerHasCapability(a.provider.capabilities, CAPABILITY)
+        );
+        if (capableAttachments.length === 0) {
+            return errorResponse(
+                res,
+                HttpCode.FORBIDDEN,
+                "permission",
+                `No AI provider on this resource supports ${CAPABILITY}`
+            );
+        }
+
+        const lists = await loadProviderModelLists(
+            capableAttachments.map((a) => a.provider.providerId)
+        );
+        const models = listPermittedModels(
+            buildDiscoveryProviders(
+                capableAttachments,
+                target.resourceListsByProvider,
+                lists
+            )
+        );
+
+        // `GET /v1/models/{id}` - a single model, 404 when this resource
+        // doesn't permit it.
+        const requestedModel = req.params?.model;
+        if (typeof requestedModel === "string" && requestedModel.length > 0) {
+            const model = models.find((m) => m.id === requestedModel);
+            if (!model) {
+                return errorResponse(
+                    res,
+                    HttpCode.NOT_FOUND,
+                    "not_found",
+                    `Model "${requestedModel}" is not available on this resource`
+                );
+            }
+            return res.status(HttpCode.OK).json(model);
+        }
+
+        const parsedQuery = querySchema.safeParse(req.query);
+        if (!parsedQuery.success) {
+            return errorResponse(
+                res,
+                HttpCode.BAD_REQUEST,
+                "invalid_request",
+                parsedQuery.error.issues[0]?.message ??
+                    "Invalid pagination parameters"
+            );
+        }
+
+        const page = paginateModels(
+            models,
+            parsedQuery.data.limit ?? MODEL_PAGE_DEFAULT_LIMIT,
+            {
+                afterId: parsedQuery.data.after_id,
+                beforeId: parsedQuery.data.before_id
+            }
+        );
+        if ("error" in page) {
+            return errorResponse(
+                res,
+                HttpCode.BAD_REQUEST,
+                "invalid_request",
+                page.error
+            );
+        }
+
+        logger.debug("AI gateway model discovery", {
+            host,
+            resourceId: target.resourceId,
+            siteResourceId: target.siteResourceId,
+            providers: capableAttachments.length,
+            total: models.length,
+            returned: page.data.length
+        });
+
+        return res.status(HttpCode.OK).json({
+            data: page.data,
+            has_more: page.has_more,
+            first_id: page.data[0]?.id ?? null,
+            last_id: page.data[page.data.length - 1]?.id ?? null
+        });
+    } catch (error) {
+        logger.error(error);
+        return errorResponse(
+            res,
+            HttpCode.INTERNAL_SERVER_ERROR,
+            "internal",
+            "Failed to list models"
+        );
+    }
+}

server/routers/certificates/createCertificate.ts

@@ -23,6 +23,12 @@ export async function createCertificate(
         throw new Error(`Domain with ID ${domainId} not found`);
     }
 
+    // Note: certificates.domain has a global UNIQUE constraint (it is not
+    // scoped per-domainId), so existence must be checked by domain value
+    // alone. Filtering on domainId here as well can cause this check to
+    // miss an existing cert (e.g. if it was stored under a different but
+    // still-valid domainId), leading to an INSERT that then fails on the
+    // unique constraint.
     let existing: Certificate[] = [];
     if (domainRecord.type == "ns" || domainRecord.type == "wildcard") {
         const domainLevelDown = domain.split(".").slice(1).join(".");
@@ -32,16 +38,13 @@ export async function createCertificate(
             .select()
             .from(certificates)
             .where(
-                and(
-                    eq(certificates.domainId, domainId),
-                    or(
-                        eq(certificates.domain, domain),
-                        and(
-                            eq(certificates.wildcard, true),
-                            or(
-                                eq(certificates.domain, domainLevelDown),
-                                eq(certificates.domain, wildcardPrefixed)
-                            )
+                or(
+                    eq(certificates.domain, domain),
+                    and(
+                        eq(certificates.wildcard, true),
+                        or(
+                            eq(certificates.domain, domainLevelDown),
+                            eq(certificates.domain, wildcardPrefixed)
                         )
                     )
                 )
@@ -51,12 +54,7 @@ export async function createCertificate(
         existing = await trx
             .select()
             .from(certificates)
-            .where(
-                and(
-                    eq(certificates.domainId, domainId),
-                    eq(certificates.domain, domain) // exact match for non-NS domains
-                )
-            );
+            .where(eq(certificates.domain, domain)); // exact match for non-NS domains
     }
 
     if (existing.length > 0) {
@@ -87,16 +85,22 @@ export async function createCertificate(
         }
     }
 
-    // No cert found, create a new one in pending state
-    await trx.insert(certificates).values({
-        domain: domainToWrite,
-        domainId,
-        wildcard:
-            domainRecord.type == "ns" ||
-            (domainRecord.type == "wildcard" &&
-                domainRecord.preferWildcardCert), // we can only create wildcard certs for NS domains
-        status: "pending",
-        updatedAt: Math.floor(Date.now() / 1000),
-        createdAt: Math.floor(Date.now() / 1000)
-    });
+    // No cert found, create a new one in pending state. onConflictDoNothing
+    // guards against the domain having been inserted concurrently (or under
+    // a different domainId) between the existence check above and this
+    // insert, since certificates.domain is globally unique.
+    await trx
+        .insert(certificates)
+        .values({
+            domain: domainToWrite,
+            domainId,
+            wildcard:
+                domainRecord.type == "ns" ||
+                (domainRecord.type == "wildcard" &&
+                    domainRecord.preferWildcardCert), // we can only create wildcard certs for NS domains
+            status: "pending",
+            updatedAt: Math.floor(Date.now() / 1000),
+            createdAt: Math.floor(Date.now() / 1000)
+        })
+        .onConflictDoNothing();
 }

server/routers/external.ts

@@ -1490,21 +1490,6 @@ authenticated.get(
     logs.exportRequestAuditLogs
 );
 
-authenticated.get(
-    "/org/:orgId/logs/ai",
-    verifyOrgAccess,
-    verifyUserHasAction(ActionsEnum.viewLogs),
-    logs.queryAiSessionLogs
-);
-
-authenticated.get(
-    "/org/:orgId/logs/ai/export",
-    verifyOrgAccess,
-    verifyUserHasAction(ActionsEnum.exportLogs),
-    logActionAudit(ActionsEnum.exportLogs),
-    logs.exportAiSessionLogs
-);
-
 authenticated.get(
     "/org/:orgId/logs/ai/usage/filters",
     verifyOrgAccess,

server/routers/gerbil/createExitNode.ts

@@ -15,13 +15,7 @@ export async function createExitNode(
     if (!exitNodeQuery) {
         const { value: address, release } = await getNextAvailableSubnet();
         try {
-            // TODO: eventually we will want to get the next available port so that we can multiple exit nodes
-            // const listenPort = await getNextAvailablePort();
             const listenPort = config.getRawConfig().gerbil.start_port;
-            let subEndpoint = "";
-            if (config.getRawConfig().gerbil.use_subdomain) {
-                subEndpoint = await getUniqueExitNodeEndpointName();
-            }
 
             const exitNodeName =
                 config.getRawConfig().gerbil.exit_node_name ||
@@ -32,7 +26,7 @@ export async function createExitNode(
                 .insert(exitNodes)
                 .values({
                     publicKey,
-                    endpoint: `${subEndpoint}${subEndpoint != "" ? "." : ""}${config.getRawConfig().gerbil.base_endpoint}`,
+                    endpoint: config.getRawConfig().gerbil.base_endpoint,
                     address,
                     online: true,
                     listenPort,

server/routers/integration.ts

@@ -1532,21 +1532,6 @@ authenticated.get(
     logs.exportRequestAuditLogs
 );
 
-authenticated.get(
-    "/org/:orgId/logs/ai",
-    verifyApiKeyOrgAccess,
-    verifyApiKeyHasAction(ActionsEnum.viewLogs),
-    logs.queryAiSessionLogs
-);
-
-authenticated.get(
-    "/org/:orgId/logs/ai/export",
-    verifyApiKeyOrgAccess,
-    verifyApiKeyHasAction(ActionsEnum.exportLogs),
-    logActionAudit(ActionsEnum.exportLogs),
-    logs.exportAiSessionLogs
-);
-
 authenticated.get(
     "/org/:orgId/logs/ai/usage/filters",
     verifyApiKeyOrgAccess,

server/routers/newt/handleRequestLogMessage.ts

@@ -1,9 +1,248 @@
+/*
+ * This file is part of a proprietary work.
+ *
+ * Copyright (c) 2025-2026 Fossorial, Inc.
+ * All rights reserved.
+ *
+ * This file is licensed under the Fossorial Commercial License.
+ * You may not use this file except in compliance with the License.
+ * Unauthorized use, copying, modification, or distribution is strictly prohibited.
+ *
+ * This file is not licensed under the AGPLv3.
+ */
+
+import { db } from "@server/db";
 import { MessageHandler } from "@server/routers/ws";
+import { sites, Newt, orgs, clients, clientSitesAssociationsCache, users } from "@server/db";
+import { and, eq, inArray } from "drizzle-orm";
+import logger from "@server/logger";
+import { inflate } from "zlib";
+import { promisify } from "util";
+import { logRequestAudit } from "@server/routers/badger/logRequestAudit";
+import { getCountryCodeForIp } from "@server/lib/geoip";
 
 export async function flushRequestLogToDb(): Promise<void> {
     return;
 }
 
+const zlibInflate = promisify(inflate);
+
+interface HTTPRequestLogData {
+    requestId: string;
+    resourceId: number; // siteResourceId
+    timestamp: string; // ISO 8601
+    method: string;
+    scheme: string; // "http" or "https"
+    host: string;
+    path: string;
+    rawQuery?: string;
+    userAgent?: string;
+    sourceAddr: string; // ip:port
+    tls: boolean;
+}
+
+/**
+ * Decompress a base64-encoded zlib-compressed string into parsed JSON.
+ */
+async function decompressRequestLog(
+    compressed: string
+): Promise<HTTPRequestLogData[]> {
+    const compressedBuffer = Buffer.from(compressed, "base64");
+    const decompressed = await zlibInflate(compressedBuffer);
+    const jsonString = decompressed.toString("utf-8");
+    const parsed = JSON.parse(jsonString);
+
+    if (!Array.isArray(parsed)) {
+        throw new Error("Decompressed request log data is not an array");
+    }
+
+    return parsed;
+}
+
 export const handleRequestLogMessage: MessageHandler = async (context) => {
-    return;
-};
+    const { message, client } = context;
+    const newt = client as Newt;
+
+    if (!newt) {
+        logger.warn("Request log received but no newt client in context");
+        return;
+    }
+
+    if (!newt.siteId) {
+        logger.warn("Request log received but newt has no siteId");
+        return;
+    }
+
+    if (!message.data?.compressed) {
+        logger.warn("Request log message missing compressed data");
+        return;
+    }
+
+    // Look up the org for this site and check retention settings
+    const [site] = await db
+        .select({
+            orgId: sites.orgId,
+            orgSubnet: orgs.subnet,
+            settingsLogRetentionDaysRequest:
+                orgs.settingsLogRetentionDaysRequest
+        })
+        .from(sites)
+        .innerJoin(orgs, eq(sites.orgId, orgs.orgId))
+        .where(eq(sites.siteId, newt.siteId));
+
+    if (!site) {
+        logger.warn(
+            `Request log received but site ${newt.siteId} not found in database`
+        );
+        return;
+    }
+
+    const orgId = site.orgId;
+
+    if (site.settingsLogRetentionDaysRequest === 0) {
+        logger.debug(
+            `Request log retention is disabled for org ${orgId}, skipping`
+        );
+        return;
+    }
+
+    let entries: HTTPRequestLogData[];
+    try {
+        entries = await decompressRequestLog(message.data.compressed);
+    } catch (error) {
+        logger.error("Failed to decompress request log data:", error);
+        return;
+    }
+
+    if (entries.length === 0) {
+        return;
+    }
+
+    logger.debug(`Request log entries: ${JSON.stringify(entries)}`);
+
+    // Build a map from sourceIp → external endpoint string by joining clients
+    // with clientSitesAssociationsCache. The endpoint is the real-world IP:port
+    // of the client device and is used for GeoIP lookup.
+    const ipToEndpoint = new Map<string, string>();
+    // Build a map from sourceIp → the user associated with the client (if any)
+    const ipToUser = new Map<string, { username: string; userId: string }>();
+
+    const cidrSuffix = site.orgSubnet?.includes("/")
+        ? site.orgSubnet.substring(site.orgSubnet.indexOf("/"))
+        : null;
+
+    if (cidrSuffix) {
+        const uniqueSourceAddrs = new Set<string>();
+        for (const entry of entries) {
+            if (entry.sourceAddr) {
+                uniqueSourceAddrs.add(entry.sourceAddr);
+            }
+        }
+
+        if (uniqueSourceAddrs.size > 0) {
+            const subnetQueries = Array.from(uniqueSourceAddrs).map((addr) => {
+                const ip = addr.includes(":") ? addr.split(":")[0] : addr;
+                return `${ip}${cidrSuffix}`;
+            });
+
+            const matchedClients = await db
+                .select({
+                    subnet: clients.subnet,
+                    endpoint: clientSitesAssociationsCache.endpoint,
+                    username: users.username,
+                    userId: users.userId
+                })
+                .from(clients)
+                .innerJoin(
+                    clientSitesAssociationsCache,
+                    and(
+                        eq(
+                            clientSitesAssociationsCache.clientId,
+                            clients.clientId
+                        ),
+                        eq(clientSitesAssociationsCache.siteId, newt.siteId)
+                    )
+                )
+                .leftJoin(users, eq(clients.userId, users.userId))
+                .where(
+                    and(
+                        eq(clients.orgId, orgId),
+                        inArray(clients.subnet, subnetQueries)
+                    )
+                );
+
+            for (const c of matchedClients) {
+                const ip = c.subnet.split("/")[0];
+                if (c.endpoint) {
+                    ipToEndpoint.set(ip, c.endpoint);
+                }
+                if (c.userId && c.username) {
+                    ipToUser.set(ip, { userId: c.userId, username: c.username });
+                }
+            }
+        }
+    }
+
+    for (const entry of entries) {
+        if (
+            !entry.requestId ||
+            !entry.resourceId ||
+            !entry.method ||
+            !entry.scheme ||
+            !entry.host ||
+            !entry.path ||
+            !entry.sourceAddr
+        ) {
+            logger.debug(
+                `Skipping request log entry with missing required fields: ${JSON.stringify(entry)}`
+            );
+            continue;
+        }
+
+        const originalRequestURL =
+            entry.scheme +
+            "://" +
+            entry.host +
+            entry.path +
+            (entry.rawQuery ? "?" + entry.rawQuery : "");
+
+        // Resolve the client's external endpoint for GeoIP lookup.
+        // sourceAddr is the WireGuard IP (possibly ip:port), so strip the port.
+        const sourceIp = entry.sourceAddr.includes(":")
+            ? entry.sourceAddr.split(":")[0]
+            : entry.sourceAddr;
+        const endpoint = ipToEndpoint.get(sourceIp);
+        let location: string | undefined;
+        if (endpoint) {
+            const endpointIp = endpoint.includes(":")
+                ? endpoint.split(":")[0]
+                : endpoint;
+            location = await getCountryCodeForIp(endpointIp);
+        }
+        const user = ipToUser.get(sourceIp);
+
+        await logRequestAudit(
+            {
+                action: true,
+                reason: 108,
+                siteResourceId: entry.resourceId,
+                orgId,
+                location,
+                user
+            },
+            {
+                path: entry.path,
+                originalRequestURL,
+                scheme: entry.scheme,
+                host: entry.host,
+                method: entry.method,
+                tls: entry.tls,
+                requestIp: entry.sourceAddr
+            }
+        );
+    }
+
+    logger.debug(
+        `Buffered ${entries.length} request log entry/entries from newt ${newt.newtId} (site ${newt.siteId})`
+    );
+};

server/routers/olm/handleOlmExitNodesRequestMessage.ts

@@ -42,7 +42,8 @@ export const handleOlmExitNodesRequestMessage: MessageHandler = async (
         client.orgId,
         true,
         noCloud || false,
-        olm.clientId
+        olm.clientId,
+        true // don't select remote exit nodes for clients
     ); // filter for only the online ones
 
     let lastExitNodeId = null;

server/routers/org/updateOrg.ts

@@ -147,6 +147,42 @@ export async function updateOrg(
             parsedBody.data.settingsEnableGlobalNewtAutoUpdate = false; // force it off
         }
 
+        // Check access logs feature
+        const hasAccessLogsFeature = await isLicensedOrSubscribed(
+            orgId,
+            tierMatrix[TierFeature.AccessLogs]
+        );
+        if (!hasAccessLogsFeature) {
+            parsedBody.data.settingsLogRetentionDaysAccess = undefined;
+        }
+
+        // Check action logs feature
+        const hasActionLogsFeature = await isLicensedOrSubscribed(
+            orgId,
+            tierMatrix[TierFeature.ActionLogs]
+        );
+        if (!hasActionLogsFeature) {
+            parsedBody.data.settingsLogRetentionDaysAction = undefined;
+        }
+
+        // Check connection logs feature
+        const hasConnectionLogsFeature = await isLicensedOrSubscribed(
+            orgId,
+            tierMatrix[TierFeature.ConnectionLogs]
+        );
+        if (!hasConnectionLogsFeature) {
+            parsedBody.data.settingsLogRetentionDaysConnection = undefined;
+        }
+
+        // Check AI session logs feature
+        const hasAISessionLogsFeature = await isLicensedOrSubscribed(
+            orgId,
+            tierMatrix[TierFeature.AISessionLogs]
+        );
+        if (!hasAISessionLogsFeature) {
+            parsedBody.data.settingsLogRetentionDaysAISessions = undefined;
+        }
+
         if (build == "saas") {
             const { tier } = await getOrgTierData(orgId);
 

server/routers/resource/listResourceUsers.ts

@@ -1,7 +1,7 @@
 import { Request, Response, NextFunction } from "express";
 import { z } from "zod";
 import { db } from "@server/db";
-import { idp, userResources, users } from "@server/db"; // Assuming these are the correct tables
+import { idp, resources, userPolicies, userResources, users } from "@server/db"; // Assuming these are the correct tables
 import { eq } from "drizzle-orm";
 import response from "@server/lib/response";
 import HttpCode from "@server/types/HttpCode";
@@ -14,7 +14,23 @@ const listResourceUsersSchema = z.strictObject({
     resourceId: z.coerce.number().int().positive()
 });
 
-async function queryUsers(resourceId: number) {
+async function queryUsers(resourceId: number, policyId: number | null) {
+    if (policyId !== null) {
+        return await db
+            .select({
+                userId: userPolicies.userId,
+                username: users.username,
+                type: users.type,
+                idpName: idp.name,
+                idpId: users.idpId,
+                email: users.email
+            })
+            .from(userPolicies)
+            .innerJoin(users, eq(userPolicies.userId, users.userId))
+            .leftJoin(idp, eq(users.idpId, idp.idpId))
+            .where(eq(userPolicies.resourcePolicyId, policyId));
+    }
+
     return await db
         .select({
             userId: userResources.userId,
@@ -104,7 +120,26 @@ export async function listResourceUsers(
 
         const { resourceId } = parsedParams.data;
 
-        const resourceUsersList = await queryUsers(resourceId);
+        const [resource] = await db
+            .select()
+            .from(resources)
+            .where(eq(resources.resourceId, resourceId))
+            .limit(1);
+
+        if (!resource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Resource not found")
+            );
+        }
+
+        const isInlinePolicy =
+            resource.resourcePolicyId === null &&
+            resource.defaultResourcePolicyId !== null;
+
+        const resourceUsersList = await queryUsers(
+            resourceId,
+            isInlinePolicy ? resource.defaultResourcePolicyId! : null
+        );
 
         return response<ListResourceUsersResponse>(res, {
             data: {

server/routers/siteResource/addClientToSiteResource.ts

@@ -64,7 +64,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "post",
-    path: "/private-resource/{siteResourceId}/clients/add",
+    path: "/private-resource/{resourceId}/clients/add",
     description:
         "Add a single client to a site resource. Clients with a userId cannot be added.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.Client],

server/routers/siteResource/addRoleToSiteResource.ts

@@ -64,7 +64,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "post",
-    path: "/private-resource/{siteResourceId}/roles/add",
+    path: "/private-resource/{resourceId}/roles/add",
     description: "Add a single role to a site resource.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.Role],
     request: {

server/routers/siteResource/addUserToSiteResource.ts

@@ -64,7 +64,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "post",
-    path: "/private-resource/{siteResourceId}/users/add",
+    path: "/private-resource/{resourceId}/users/add",
     description: "Add a single user to a site resource.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.User],
     request: {

server/routers/siteResource/listSiteResourceClients.ts

@@ -63,7 +63,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "get",
-    path: "/private-resource/{siteResourceId}/clients",
+    path: "/private-resource/{resourceId}/clients",
     description: "List all clients for a site resource.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.Client],
     request: {

server/routers/siteResource/listSiteResourceRoles.ts

@@ -64,7 +64,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "get",
-    path: "/private-resource/{siteResourceId}/roles",
+    path: "/private-resource/{resourceId}/roles",
     description: "List all roles for a site resource.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.Role],
     request: {

server/routers/siteResource/listSiteResourceUsers.ts

@@ -67,7 +67,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "get",
-    path: "/private-resource/{siteResourceId}/users",
+    path: "/private-resource/{resourceId}/users",
     description: "List all users for a site resource.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.User],
     request: {

server/routers/siteResource/removeClientFromSiteResource.ts

@@ -64,7 +64,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "post",
-    path: "/private-resource/{siteResourceId}/clients/remove",
+    path: "/private-resource/{resourceId}/clients/remove",
     description:
         "Remove a single client from a site resource. Clients with a userId cannot be removed.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.Client],

server/routers/siteResource/removeRoleFromSiteResource.ts

@@ -64,7 +64,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "post",
-    path: "/private-resource/{siteResourceId}/roles/remove",
+    path: "/private-resource/{resourceId}/roles/remove",
     description: "Remove a single role from a site resource.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.Role],
     request: {

server/routers/siteResource/removeUserFromSiteResource.ts

@@ -64,7 +64,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "post",
-    path: "/private-resource/{siteResourceId}/users/remove",
+    path: "/private-resource/{resourceId}/users/remove",
     description: "Remove a single user from a site resource.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.User],
     request: {

server/routers/siteResource/setSiteResourceClients.ts

@@ -64,7 +64,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "post",
-    path: "/private-resource/{siteResourceId}/clients",
+    path: "/private-resource/{resourceId}/clients",
     description:
         "Set clients for a site resource. This will replace all existing clients. Clients with a userId cannot be added.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.Client],

server/routers/siteResource/setSiteResourceRoles.ts

@@ -65,7 +65,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "post",
-    path: "/private-resource/{siteResourceId}/roles",
+    path: "/private-resource/{resourceId}/roles",
     description:
         "Set roles for a site resource. This will replace all existing roles.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.Role],

server/routers/siteResource/setSiteResourceUsers.ts

@@ -66,7 +66,7 @@ registry.registerPath({
 
 registry.registerPath({
     method: "post",
-    path: "/private-resource/{siteResourceId}/users",
+    path: "/private-resource/{resourceId}/users",
     description:
         "Set users for a site resource. This will replace all existing users.",
     tags: [OpenAPITags.PrivateResource, OpenAPITags.User],

server/routers/ws/messageHandlers.ts

@@ -7,7 +7,8 @@ import {
     handleNewtExitNodesRequestMessage,
     handleApplyBlueprintMessage,
     handleNewtPingMessage,
-    handleNewtDisconnectingMessage
+    handleNewtDisconnectingMessage,
+    handleRequestLogMessage
 } from "../newt";
 import {
     handleOlmRegisterMessage,
@@ -46,5 +47,6 @@ export const messageHandlers: Record<string, MessageHandler> = {
     "newt/ping/request": handleNewtExitNodesRequestMessage,
     "newt/blueprint/apply": handleApplyBlueprintMessage,
     "newt/healthcheck/status": handleHealthcheckStatusMessage,
+    "newt/request-log": handleRequestLogMessage,
     "ws/round-trip/complete": handleRoundTripMessage
 };

server/routers/ws/ws.ts

@@ -388,7 +388,7 @@ const setupConnection = async (
                 }
             }
         } catch (error) {
-            logger.error("Message handling error:", error);
+            logger.warn("Message handling error:", error);
             ws.send(
                 JSON.stringify({
                     type: "error",

server/setup/scriptsPg/1.22.0.ts

@@ -345,6 +345,9 @@ export default async function migration() {
         await db.execute(
             sql`ALTER TABLE "virtualApiKeys" ADD CONSTRAINT "virtualApiKeys_createdByUserId_user_id_fk" FOREIGN KEY ("createdByUserId") REFERENCES "public"."user"("id") ON DELETE set null ON UPDATE no action;`
         );
+        await db.execute(
+            sql`ALTER TABLE "eventStreamingDestinations" ADD "sendAISessionLogs" boolean DEFAULT false NOT NULL;`
+        );
         await db.execute(
             sql`CREATE INDEX "idx_ai_budget_breach_events_budget_created" ON "aiBudgetBreachEvents" USING btree ("budgetId","createdAt");`
         );

server/setup/scriptsSqlite/1.22.0.ts

@@ -402,6 +402,9 @@ export default async function migration() {
             db.prepare(
                 `ALTER TABLE 'siteResources' ADD 'requiresExitNodeConnection' integer DEFAULT false NOT NULL;`
             ).run();
+            db.prepare(
+                `ALTER TABLE 'eventStreamingDestinations' ADD 'sendAISessionLogs' integer DEFAULT false NOT NULL;`
+            ).run();
 
             const insertRoleAction = db.prepare(`
                 INSERT INTO 'roleActions' ("roleId", "actionId", "orgId")