@antseed/cli 0.1.141 → 0.1.143

package/dist/proxy/buyer-proxy.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"buyer-proxy.d.ts","sourceRoot":"","sources":["../../src/proxy/buyer-proxy.ts"],"names":[],"mappings":"AAKA,OAAO,EAIL,KAAK,WAAW,EAChB,KAAK,QAAQ,EAKb,KAAK,sBAAsB,EAG5B,MAAM,eAAe,CAAA;AAgCtB,OAAO,EAAiF,KAAK,cAAc,EAAE,KAAK,WAAW,EAAsB,MAAM,wBAAwB,CAAA;AAGjL,OAAO,EAAE,8BAA8B,EAAE,KAAK,2BAA2B,EAAE,MAAM,cAAc,CAAA;AAC/F,OAAO,EAAE,sBAAsB,EAAE,8BAA8B,EAAE,MAAM,oBAAoB,CAAA;AAE3F,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,MAAM,CAAA;IACZ,IAAI,EAAE,WAAW,CAAA;IACjB,4FAA4F;IAC5F,OAAO,EAAE,MAAM,CAAA;IACf,kGAAkG;IAClG,2BAA2B,CAAC,EAAE,MAAM,CAAA;IACpC;;;;;OAKG;IACH,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB;;;;OAIG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,yFAAyF;IACzF,QAAQ,CAAC,EAAE,cAAc,CAAA;CAC1B;AAYD;;;;;;;GAOG;AACH,wBAAgB,uBAAuB,CAAC,QAAQ,EAAE,sBAAsB,GAAG,OAAO,CAQjF;AA6ID;;;;GAIG;AACH,wBAAgB,mBAAmB,CACjC,MAAM,EAAE,OAAO,EACf,KAAK,GAAE,MAAmB,EAC1B,QAAQ,GAAE,MAA6B,GACtC,QAAQ,EAAE,CAkHZ;AAED;;;;;;GAMG;AAEH,wBAAgB,iBAAiB,CAC/B,IAAI,EAAE,IAAI,CAAC,WAAW,EAAE,aAAa,CAAC,EACtC,IAAI,EAAE,QAAQ,EACd,QAAQ,EAAE,MAAM,EAChB,MAAM,EAAE,WAAW,GAClB,WAAW,CAeb;AAED,qBAAa,UAAU;IACrB,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAQ;IAChC,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAa;IACnC,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAQ;IAC9B,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAQ;IAC7C,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAQ;IACxC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAQ;IAClC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAQ;IACnC,OAAO,CAAC,kBAAkB,CAAQ;IAClC,OAAO,CAAC,WAAW,CAAe;IAClC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAC,CAAgB;IAC3C,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAmC;IAChE,OAAO,CAAC,mBAAmB,CAA6C;IAExE,OAAO,CAAC,gBAAgB,CAAmC;IAE3D,OAAO,CAAC,YAAY,CAAiB;IACrC,OAAO,CAAC,qBAAqB,CAAI;IACjC,OAAO,CAAC,mBAAmB,CAAI;IAC/B,OAAO,CAAC,mBAAmB,CAAmC;IAC9D,OAAO,CAAC,sBAAsB,CAAI;IAClC,OAAO,CAAC,6BAA6B,CAAI;IACzC,OAAO,CAAC,gBAAgB,CAA8C;IACtE,OAAO,CAAC,aAAa,CAA2C;IAChE,4EAA4E;IAC5E,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAyC;gBAE5D,MAAM,EAAE,gBAAgB;IA4C9B,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC;YA6Bd,0BAA0B;IAuBlC,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;IAmB3B,OAAO,CAAC,eAAe;YAqBT,uBAAuB;IAcrC,0FAA0F;IAC1F,OAAO,CAAC,eAAe;YAsBT,eAAe;IAe7B,OAAO,CAAC,+BAA+B;IAIvC,OAAO,CAAC,uBAAuB;IAS/B,OAAO,CAAC,aAAa;IA4CrB,OAAO,CAAC,oBAAoB;IAkE5B;;;;OAIG;IACH,OAAO,CAAC,kBAAkB;IAoB1B;;;;;;OAMG;IACH,OAAO,CAAC,gBAAgB;IAcxB;;;;;OAKG;IACH,OAAO,CAAC,uBAAuB;YASjB,yBAAyB;YASzB,gBAAgB;YA+BhB,SAAS;IA2BvB,OAAO,CAAC,+BAA+B;YAsBzB,mBAAmB;YAmNnB,cAAc;YA8Td,WAAW;IAkBzB,OAAO,CAAC,wBAAwB;IAOhC,OAAO,CAAC,uBAAuB;IAoB/B,OAAO,CAAC,6BAA6B;IA4BrC,OAAO,CAAC,YAAY;IASpB;;;;OAIG;YACW,eAAe;CA2T9B"}
+{"version":3,"file":"buyer-proxy.d.ts","sourceRoot":"","sources":["../../src/proxy/buyer-proxy.ts"],"names":[],"mappings":"AAKA,OAAO,EAIL,KAAK,WAAW,EAEhB,KAAK,QAAQ,EAKb,KAAK,sBAAsB,EAG5B,MAAM,eAAe,CAAA;AA+CtB,OAAO,EAAiF,KAAK,cAAc,EAAE,KAAK,WAAW,EAAsB,MAAM,wBAAwB,CAAA;AAGjL,OAAO,EAAE,8BAA8B,EAAE,KAAK,2BAA2B,EAAE,MAAM,cAAc,CAAA;AAC/F,OAAO,EAAE,sBAAsB,EAAE,8BAA8B,EAAE,0BAA0B,EAAE,kBAAkB,EAAE,MAAM,oBAAoB,CAAA;AAE3I,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,MAAM,CAAA;IACZ,IAAI,EAAE,WAAW,CAAA;IACjB,4FAA4F;IAC5F,OAAO,EAAE,MAAM,CAAA;IACf,kGAAkG;IAClG,2BAA2B,CAAC,EAAE,MAAM,CAAA;IACpC;;;;;OAKG;IACH,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB;;;;OAIG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,yFAAyF;IACzF,QAAQ,CAAC,EAAE,cAAc,CAAA;CAC1B;AAYD;;;;;;;GAOG;AACH,wBAAgB,uBAAuB,CAAC,QAAQ,EAAE,sBAAsB,GAAG,OAAO,CAQjF;AAmJD;;;;GAIG;AACH,wBAAgB,mBAAmB,CACjC,MAAM,EAAE,OAAO,EACf,KAAK,GAAE,MAAmB,EAC1B,QAAQ,GAAE,MAA6B,GACtC,QAAQ,EAAE,CAkHZ;AAED;;;;;;GAMG;AAEH,wBAAgB,iBAAiB,CAC/B,IAAI,EAAE,IAAI,CAAC,WAAW,EAAE,aAAa,CAAC,EACtC,IAAI,EAAE,QAAQ,EACd,QAAQ,EAAE,MAAM,EAChB,MAAM,EAAE,WAAW,GAClB,WAAW,CAeb;AAED,qBAAa,UAAU;IACrB,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAQ;IAChC,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAa;IACnC,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAQ;IAC9B,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAQ;IAC7C,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAQ;IACxC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAQ;IAClC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAQ;IACnC,OAAO,CAAC,kBAAkB,CAAQ;IAClC,OAAO,CAAC,WAAW,CAAe;IAClC;;;;OAIG;IACH,OAAO,CAAC,mBAAmB,CAAsB;IACjD,OAAO,CAAC,cAAc,CAAoB;IAC1C;;;;;OAKG;IACH,OAAO,CAAC,oBAAoB,CAAI;IAChC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAC,CAAgB;IAC3C,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAmC;IAChE,OAAO,CAAC,mBAAmB,CAA6C;IAExE,OAAO,CAAC,gBAAgB,CAAmC;IAE3D,OAAO,CAAC,YAAY,CAAiB;IACrC,OAAO,CAAC,qBAAqB,CAAI;IACjC,OAAO,CAAC,mBAAmB,CAAI;IAC/B,OAAO,CAAC,mBAAmB,CAAmC;IAC9D,OAAO,CAAC,sBAAsB,CAAI;IAClC,OAAO,CAAC,YAAY,CAAI;IACxB;wFACoF;IACpF,OAAO,CAAC,4BAA4B,CAAI;IACxC,OAAO,CAAC,6BAA6B,CAAI;IACzC,OAAO,CAAC,gBAAgB,CAA8C;IACtE,OAAO,CAAC,aAAa,CAA2C;IAChE,4EAA4E;IAC5E,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAyC;IAExE;;;;;;;OAOG;IACH,OAAO,CAAC,QAAQ,CAAC,qBAAqB,CAA0D;gBAEpF,MAAM,EAAE,gBAAgB;IAsDpC,yEAAyE;IACzE,OAAO,CAAC,eAAe;IAiBvB;;;;OAIG;IACH,OAAO,CAAC,yBAAyB;IAS3B,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC;YA6Bd,0BAA0B;IAuBlC,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;IAoB3B,OAAO,CAAC,eAAe;YAqBT,uBAAuB;IAkBrC,+EAA+E;IAC/E,OAAO,CAAC,kBAAkB;IAI1B,0FAA0F;IAC1F,OAAO,CAAC,eAAe;YAsBT,eAAe;IAe7B,OAAO,CAAC,+BAA+B;IAIvC,OAAO,CAAC,uBAAuB;IAS/B,OAAO,CAAC,aAAa;IA4CrB,OAAO,CAAC,oBAAoB;IAkE5B;;;;OAIG;IACH,OAAO,CAAC,kBAAkB;IAoB1B;;;;;;OAMG;IACH,OAAO,CAAC,gBAAgB;IAcxB;;;;;OAKG;IACH,OAAO,CAAC,uBAAuB;YASjB,yBAAyB;YASzB,gBAAgB;YAiChB,SAAS;IA2BvB,OAAO,CAAC,+BAA+B;YAsBzB,mBAAmB;YA4UnB,cAAc;YAqbd,WAAW;IAkBzB,OAAO,CAAC,wBAAwB;IAOhC,OAAO,CAAC,uBAAuB;IAoB/B,OAAO,CAAC,6BAA6B;IA4BrC,OAAO,CAAC,YAAY;IASpB;;;;OAIG;YACW,eAAe;CA8T9B"}

package/dist/proxy/buyer-proxy.js

@@ -5,14 +5,17 @@ import { readFile, writeFile, rename, mkdir } from 'node:fs/promises';
 import { join } from 'node:path';
 import { ANTSEED_ATTEST_PATH, computeOnChainReputationScore, decodeSweepRequest, } from '@antseed/node';
 import { createStreamingAdapter, detectRequestServiceApiProtocol, transformRequest, transformResponse, } from './service-api-adapter.js';
-import { DEBUG, log, extractRequestedService, summarizeRequestShape, summarizeErrorResponse, requestWantsStreaming, rewritePeerPinnedServiceInBody, } from './request-utils.js';
+import { DEBUG, log, extractRequestedService, summarizeRequestShape, summarizeErrorResponse, requestWantsStreaming, parsePeerPinnedService, rewritePeerPinnedServiceInBody, substituteRoutedModelAlias, overrideRoutedModelInBody, ROUTED_MODEL_ALIAS, SYSTEM_PROXY_SOURCE_HEADER, SYSTEM_ROUTED_MODEL_HEADER, } from './request-utils.js';
 import { getExplicitProviderOverride, getExplicitPeerIdOverride, resolvePeerRoutePlan, selectCandidatePeersForRouting, } from './routing.js';
 import { computeResponseTelemetry, attachAntseedTelemetryHeaders, attachStreamingAntseedHeaders, } from './telemetry.js';
 import { DEFAULT_BUYER_PEER_REFRESH_INTERVAL_MS } from '../config/defaults.js';
+import { extractConversationIdentity, extractFirstUserSnippet, isCompletionRequestPath, isTitleGenerationRequest, parseRequestBodyObject, } from './conversation-identity.js';
+import { ConversationStore } from './conversation-store.js';
+import { estimateAnthropicPromptTokens, isCountTokensPath } from './count-tokens.js';
 import { getCachedVerdict, runVerifier, verifierSupportFingerprint } from '../plugins/verifier.js';
 // Re-export for backward compatibility (used by tests and other consumers)
 export { selectCandidatePeersForRouting } from './routing.js';
-export { parsePeerPinnedService, rewritePeerPinnedServiceInBody } from './request-utils.js';
+export { parsePeerPinnedService, rewritePeerPinnedServiceInBody, substituteRoutedModelAlias, ROUTED_MODEL_ALIAS } from './request-utils.js';
 const RETRYABLE_STATUS_CODES = new Set([408, 429, 500, 502, 503, 504]);
 function isControlPlaneServicesPath(path) {
     return path.toLowerCase().startsWith('/v1/models');
@@ -48,6 +51,12 @@ export function isModelNotFoundResponse(response) {
  */
 const CARRY_FORWARD_TTL_MS = 2 * 60 * 60_000;
 const PEER_FAILURE_WINDOW_MS = 5 * 60_000;
+/**
+ * Requests kept in the spend-attribution map. Entries outlive their request on
+ * purpose (a seller-initiated auth can land after the response), so this is
+ * sized for in-flight plus recently-finished traffic, not for concurrency.
+ */
+const MAX_TRACKED_REQUEST_CONVERSATIONS = 512;
 /** Min gap between background peer refreshes triggered by model_not_found responses. */
 const MODEL_NOT_FOUND_REFRESH_THROTTLE_MS = 30_000;
 /** Verification is expensive; bound how many verdicts we retain (TTL = peer-cache TTL). */
@@ -310,6 +319,20 @@ export class BuyerProxy {
     _stateFile;
     _stateFileWatching = false;
     _pinnedPeer;
+    /**
+     * Route substituted for the `antseed` model alias (`<peerId>@<service>`).
+     * Set via `POST /_antseed/route` (the desktop keeps it on the current VPR
+     * selection) and persisted in buyer.state.json like the session peer pin.
+     */
+    _defaultRoutedModel = null;
+    _conversations;
+    /**
+     * Wall-clock of the last model-request activity (dispatch or streamed
+     * frame). Exposed on /_antseed/buyer-usage so the desktop pill can show a
+     * live "traffic" signal without depending on debug logging. A single field
+     * write per frame — negligible next to the routing/payment/stream work.
+     */
+    _lastModelActivityAt = 0;
     _verifier;
     _verifyCache = new Map();
     _stateWatchDebounce = null;
@@ -319,11 +342,24 @@ export class BuyerProxy {
     _cacheMutationEpoch = 0;
     _peerRefreshPromise = null;
     _lastStaleCacheLogAtMs = 0;
+    _startedAtMs = 0;
+    /** After a network switch the DHT routing table can stay populated with stale
+        nodes, so repeated empty sweeps are the reachability signal — not node count. */
+    _consecutiveEmptyDiscoveries = 0;
     _lastModelNotFoundRefreshAtMs = 0;
     _bgRefreshHandle = null;
     _peerFailures = new Map();
     /** Latest relayer receipt per sweep authNonce, for CLI progress polling. */
     _sweepReceipts = new Map();
+    /**
+     * requestId -> the conversation that issued it, so the node's per-request
+     * spend events can be attributed to a chat. Only the proxy knows this
+     * mapping; the payment layer sees a seller and a request id.
+     *
+     * `counted` guards requestCount: one request can produce several spend
+     * deltas (buyer- and seller-initiated auth both advance the cumulative).
+     */
+    _requestConversations = new Map();
     constructor(config) {
         this._node = config.node;
         this._verifier = config.verifier;
@@ -332,6 +368,7 @@ export class BuyerProxy {
         this._peerCacheTtlMs = Math.max(0, config.peerCacheTtlMs ?? Math.max(6 * 60_000, this._bgRefreshIntervalMs + 60_000));
         this._stateDir = config.dataDir;
         this._stateFile = join(config.dataDir, 'buyer.state.json');
+        this._conversations = new ConversationStore(config.dataDir);
         this._pinnedPeer = config.pinnedPeerId?.toLowerCase() ?? null;
         this._server = createServer((req, res) => {
             this._handleRequest(req, res).catch((err) => {
@@ -353,6 +390,12 @@ export class BuyerProxy {
                 }
             });
         }
+        const spendEventNode = this._node;
+        if (typeof spendEventNode.on === 'function') {
+            spendEventNode.on('payment:spend', (event) => {
+                this._attributeSpend(event);
+            });
+        }
         const eventNode = this._node;
         if (typeof eventNode.on === 'function') {
             eventNode.on('peers:discovered', (peers) => {
@@ -363,18 +406,47 @@ export class BuyerProxy {
             });
         }
     }
+    /** Roll a signed spend delta into the conversation that triggered it. */
+    _attributeSpend(event) {
+        if (!event.requestId)
+            return;
+        const entry = this._requestConversations.get(event.requestId);
+        if (!entry)
+            return;
+        this._conversations.addSpend(entry.convId, {
+            amountUsdc: event.amountUsdc,
+            inputTokens: event.inputTokens,
+            cachedInputTokens: event.cachedInputTokens,
+            outputTokens: event.outputTokens,
+        }, !entry.counted);
+        entry.counted = true;
+    }
+    /**
+     * Remember which chat a request belongs to for the life of the request plus
+     * a grace window — the seller-initiated auth path can land just after the
+     * response is returned. Bounded so a leaked id can't grow the map forever.
+     */
+    _trackRequestConversation(requestId, convId) {
+        this._requestConversations.set(requestId, { convId, counted: false });
+        while (this._requestConversations.size > MAX_TRACKED_REQUEST_CONVERSATIONS) {
+            const oldest = this._requestConversations.keys().next().value;
+            if (oldest === undefined)
+                break;
+            this._requestConversations.delete(oldest);
+        }
+    }
     async start() {
+        this._startedAtMs = Date.now();
         // Hydrate the in-memory peer cache from the persisted state file BEFORE
         // the server starts accepting requests. This lets the first request after
         // startup route from the warm cache without blocking on DHT discovery.
         // The background refresh still runs to pick up fresh peers and IP changes.
         await this._hydratePeersFromStateFile();
-        // If the CLI didn't pass --peer, adopt whatever a previous
-        // `antseed buyer connection set --peer` wrote to buyer.state.json so the
-        // pin survives daemon restart.
-        if (this._pinnedPeer === null) {
-            await this._reloadSessionOverrides();
-        }
+        // Adopt persisted session overrides (peer pin, default routed model) so
+        // they survive daemon restart. A --peer CLI flag beats the persisted pin
+        // at startup; runtime `connection set` writes still take over via the
+        // state-file watcher.
+        await this._reloadSessionOverrides({ preservePeerPin: this._pinnedPeer !== null });
         await new Promise((resolve, reject) => {
             this._server.once('error', reject);
             this._server.listen(this._port, '127.0.0.1', () => {
@@ -428,6 +500,7 @@ export class BuyerProxy {
             this._bgRefreshHandle = null;
         }
         await this._writeStateFile('stopped');
+        await this._conversations.flush();
         return new Promise((resolve) => {
             this._server.close(() => resolve());
         });
@@ -455,20 +528,28 @@ export class BuyerProxy {
             // watcher setup failed; non-fatal
         }
     }
-    async _reloadSessionOverrides() {
+    async _reloadSessionOverrides(opts = {}) {
         try {
             const raw = await readFile(this._stateFile, 'utf-8');
             const parsed = JSON.parse(raw);
-            const pinnedPeer = typeof parsed.pinnedPeerId === 'string' && parsed.pinnedPeerId.trim().length > 0
-                ? parsed.pinnedPeerId.trim().toLowerCase()
-                : null;
-            this._pinnedPeer = pinnedPeer;
-            log(`Session overrides reloaded: peer=${pinnedPeer ?? 'none'}`);
+            if (!opts.preservePeerPin) {
+                const pinnedPeer = typeof parsed.pinnedPeerId === 'string' && parsed.pinnedPeerId.trim().length > 0
+                    ? parsed.pinnedPeerId.trim().toLowerCase()
+                    : null;
+                this._pinnedPeer = pinnedPeer;
+            }
+            const routedModel = typeof parsed.defaultRoutedModel === 'string' ? parsed.defaultRoutedModel.trim() : '';
+            this._defaultRoutedModel = parsePeerPinnedService(routedModel) ? routedModel : null;
+            log(`Session overrides reloaded: peer=${this._pinnedPeer ?? 'none'} route=${this._defaultRoutedModel ?? 'none'}`);
         }
         catch {
             // state file unreadable; keep current values
         }
     }
+    /** Stamp the last model-request activity time (dispatch or streamed frame). */
+    _markModelActivity() {
+        this._lastModelActivityAt = Date.now();
+    }
     /** Serialised read-modify-write to buyer.state.json. Returns the queued write promise. */
     _mergeStateFile(patch) {
         this._stateWriteChain = this._stateWriteChain.then(async () => {
@@ -494,11 +575,11 @@ export class BuyerProxy {
         return this._stateWriteChain;
     }
     async _writeStateFile(state) {
-        // When stopping, preserve whatever pinnedPeerId is already
+        // When stopping, preserve whatever session overrides are already
         // in the file — the debounce may have been cancelled before
         // _reloadSessionOverrides could commit the latest CLI-written values.
         const sessionOverrides = state === 'connected'
-            ? { pinnedPeerId: this._pinnedPeer }
+            ? { pinnedPeerId: this._pinnedPeer, defaultRoutedModel: this._defaultRoutedModel }
             : {};
         await this._mergeStateFile({
             state,
@@ -695,9 +776,11 @@ export class BuyerProxy {
         this._peerRefreshPromise = (async () => {
             const peers = await this._discoverPeersFromNetwork();
             if (peers.length > 0) {
+                this._consecutiveEmptyDiscoveries = 0;
                 this._replacePeers(peers);
                 return peers;
             }
+            this._consecutiveEmptyDiscoveries += 1;
             const fallbackPeers = previousCachedPeers.length > 0 && this._cacheMutationEpoch === mutationEpochAtStart
                 ? [...previousCachedPeers]
                 : [];
@@ -765,6 +848,20 @@ export class BuyerProxy {
             res.end();
             return;
         }
+        if (path === '/_antseed/status' && method === 'GET') {
+            // Network reachability snapshot for UI diagnostics.
+            res.writeHead(200, { 'content-type': 'application/json' });
+            res.end(JSON.stringify({
+                ok: true,
+                dhtNodeCount: this._node.dhtNodeCount,
+                consecutiveEmptyDiscoveries: this._consecutiveEmptyDiscoveries,
+                peerCount: this._cachedPeers.length,
+                peersUpdatedAt: this._cacheLastUpdatedAtMs > 0 ? this._cacheLastUpdatedAtMs : null,
+                startedAt: this._startedAtMs,
+                uptimeMs: this._startedAtMs > 0 ? Date.now() - this._startedAtMs : 0,
+            }));
+            return;
+        }
         if (path === '/_antseed/peers/refresh' && method === 'POST') {
             try {
                 const peers = await this._refreshPeersNow();
@@ -797,6 +894,107 @@ export class BuyerProxy {
             res.end(JSON.stringify({ ok: true, peers: payload }));
             return;
         }
+        if (path === '/_antseed/route' && method === 'GET') {
+            res.writeHead(200, { 'content-type': 'application/json' });
+            res.end(JSON.stringify({ ok: true, model: this._defaultRoutedModel }));
+            return;
+        }
+        if (path === '/_antseed/route' && method === 'POST') {
+            const chunks = [];
+            let totalSize = 0;
+            for await (const chunk of req) {
+                totalSize += chunk.length;
+                if (totalSize > 8192) {
+                    res.writeHead(413, { 'content-type': 'application/json' });
+                    res.end(JSON.stringify({ ok: false, error: 'Request body too large' }));
+                    return;
+                }
+                chunks.push(chunk);
+            }
+            let model;
+            try {
+                const body = JSON.parse(Buffer.concat(chunks).toString());
+                model = typeof body.model === 'string' ? body.model.trim() : '';
+            }
+            catch {
+                res.writeHead(400, { 'content-type': 'application/json' });
+                res.end(JSON.stringify({ ok: false, error: 'Invalid JSON body' }));
+                return;
+            }
+            if (model.length > 0 && !parsePeerPinnedService(model)) {
+                res.writeHead(400, { 'content-type': 'application/json' });
+                res.end(JSON.stringify({ ok: false, error: 'model must be "<peerId>@<service>" (or empty to clear)' }));
+                return;
+            }
+            this._defaultRoutedModel = model.length > 0 ? model : null;
+            await this._mergeStateFile({ defaultRoutedModel: this._defaultRoutedModel });
+            log(`Default routed model set: ${this._defaultRoutedModel ?? 'none'}`);
+            res.writeHead(200, { 'content-type': 'application/json' });
+            res.end(JSON.stringify({ ok: true, model: this._defaultRoutedModel }));
+            return;
+        }
+        if (path === '/_antseed/conversations' && method === 'GET') {
+            res.writeHead(200, { 'content-type': 'application/json' });
+            res.end(JSON.stringify({ ok: true, conversations: this._conversations.list() }));
+            return;
+        }
+        if (path === '/_antseed/conversations/update' && method === 'POST') {
+            const chunks = [];
+            let totalSize = 0;
+            for await (const chunk of req) {
+                totalSize += chunk.length;
+                if (totalSize > 8192) {
+                    res.writeHead(413, { 'content-type': 'application/json' });
+                    res.end(JSON.stringify({ ok: false, error: 'Request body too large' }));
+                    return;
+                }
+                chunks.push(chunk);
+            }
+            let parsed;
+            try {
+                parsed = JSON.parse(Buffer.concat(chunks).toString());
+            }
+            catch {
+                res.writeHead(400, { 'content-type': 'application/json' });
+                res.end(JSON.stringify({ ok: false, error: 'Invalid JSON body' }));
+                return;
+            }
+            const id = typeof parsed.id === 'string' ? parsed.id : '';
+            if (!id) {
+                res.writeHead(400, { 'content-type': 'application/json' });
+                res.end(JSON.stringify({ ok: false, error: 'id is required' }));
+                return;
+            }
+            if (parsed.delete === true) {
+                const removed = this._conversations.remove(id);
+                res.writeHead(removed ? 200 : 404, { 'content-type': 'application/json' });
+                res.end(JSON.stringify(removed ? { ok: true } : { ok: false, error: 'Unknown conversation' }));
+                return;
+            }
+            let conversation = this._conversations.get(id);
+            if (!conversation) {
+                res.writeHead(404, { 'content-type': 'application/json' });
+                res.end(JSON.stringify({ ok: false, error: 'Unknown conversation' }));
+                return;
+            }
+            if ('pinnedModel' in parsed) {
+                const pin = typeof parsed.pinnedModel === 'string' ? parsed.pinnedModel.trim() : '';
+                if (pin.length > 0 && !parsePeerPinnedService(pin)) {
+                    res.writeHead(400, { 'content-type': 'application/json' });
+                    res.end(JSON.stringify({ ok: false, error: 'pinnedModel must be "<peerId>@<service>" (or empty to clear)' }));
+                    return;
+                }
+                conversation = this._conversations.setPinnedModel(id, pin.length > 0 ? pin : null);
+                log(`Conversation ${id.slice(0, 40)} pin: ${pin || 'cleared'}`);
+            }
+            if ('label' in parsed) {
+                const label = typeof parsed.label === 'string' ? parsed.label : null;
+                conversation = this._conversations.setLabel(id, label);
+            }
+            res.writeHead(200, { 'content-type': 'application/json' });
+            res.end(JSON.stringify({ ok: true, conversation }));
+            return;
+        }
         if (path === '/_antseed/connect' && method === 'POST') {
             const chunks = [];
             let totalSize = 0;
@@ -857,7 +1055,7 @@ export class BuyerProxy {
         if (path.startsWith('/_antseed/buyer-usage') && method === 'GET') {
             const totals = this._node.getBuyerUsageTotals();
             res.writeHead(200, { 'content-type': 'application/json' });
-            res.end(JSON.stringify({ ok: true, totals }));
+            res.end(JSON.stringify({ ok: true, totals, lastActivityAt: this._lastModelActivityAt || null }));
             return;
         }
         const meteringMatch = path.match(/^\/_antseed\/metering\/(.+)$/);
@@ -868,8 +1066,11 @@ export class BuyerProxy {
             res.end(JSON.stringify(stats));
             return;
         }
-        // Broadcast a signed deposit-sweep request over the daemon's existing
-        // seller connections (see docs/protocol/spec/09-deposit-sweep.md).
+        // Offer a signed deposit-sweep request to the daemon's connected relayers
+        // one at a time (see docs/protocol/spec/09-deposit-sweep.md) — sequential
+        // dispatch keeps relayers from racing the same nonce and burning gas on
+        // reverted transactions. Responds only after a relayer accepts, every
+        // candidate declines, or the offer round runs out of peers.
         if (path === '/_antseed/sweep' && method === 'POST') {
             const chunks = [];
             let totalSize = 0;
@@ -885,10 +1086,10 @@ export class BuyerProxy {
             try {
                 // Re-validate through the wire codec — same rules as inbound P2P frames.
                 const payload = decodeSweepRequest(new Uint8Array(Buffer.concat(chunks)));
-                const sent = this._node.broadcastSweepRequest(payload);
-                log(`Sweep request ${payload.nonce.slice(0, 10)}... broadcast to ${sent} peer(s)`);
+                const { offered, accepted } = await this._node.dispatchSweepRequest(payload);
+                log(`Sweep request ${payload.nonce.slice(0, 10)}... offered to ${offered} peer(s), accepted=${accepted}`);
                 res.writeHead(200, { 'content-type': 'application/json' });
-                res.end(JSON.stringify({ ok: true, sent }));
+                res.end(JSON.stringify({ ok: true, sent: offered, accepted }));
             }
             catch (err) {
                 const message = err instanceof Error ? err.message : String(err);
@@ -979,6 +1180,17 @@ export class BuyerProxy {
             chunks.push(chunk);
         }
         const body = Buffer.concat(chunks);
+        // `/v1/messages/count_tokens` sits under the completion prefix but is not
+        // a turn — Anthropic tools call it before most turns to size the context.
+        // Answering it locally keeps it off the wire: routed to a seller it costs
+        // a full inference and answers in a shape the caller cannot read.
+        if (method === 'POST' && isCountTokensPath(normalizedPath)) {
+            const inputTokens = estimateAnthropicPromptTokens(new Uint8Array(body));
+            log(`count_tokens answered locally: ${inputTokens} tokens`);
+            res.writeHead(200, { 'content-type': 'application/json' });
+            res.end(JSON.stringify({ input_tokens: inputTokens }));
+            return;
+        }
         // Build serialized request
         const headers = {};
         for (const [key, value] of Object.entries(req.headers)) {
@@ -991,6 +1203,11 @@ export class BuyerProxy {
         }
         // Remove host header (points to localhost, not the seller)
         delete headers['host'];
+        // Internal marker from the system proxy: the body's model was assigned
+        // by the proxy's route rewrite, not chosen by the tool. Stripped here so
+        // it never reaches a seller.
+        const systemRoutedModel = headers[SYSTEM_ROUTED_MODEL_HEADER] === '1';
+        delete headers[SYSTEM_ROUTED_MODEL_HEADER];
         let serializedReq = {
             requestId: randomUUID(),
             method,
@@ -998,9 +1215,106 @@ export class BuyerProxy {
             headers,
             body: new Uint8Array(body),
         };
-        // Snapshot the session peer pin before any await so a concurrent
+        // Snapshot the session overrides before any await so a concurrent
         // _reloadSessionOverrides() cannot change routing mid-request.
         const effectivePinnedPeer = this._pinnedPeer;
+        // Per-chat routing: completion requests carry a stable per-conversation
+        // identity (see conversation-identity.ts). A chat pinned to a model
+        // overrides the session default when resolving the `antseed` alias;
+        // subagent sessions inherit their parent chat's pin. The default route
+        // only steers a chat's first request — the model that serves it becomes
+        // the chat's own pin (ConversationStore.touch), so changing the default
+        // later applies to new chats only.
+        const isConversationRequest = method === 'POST' && isCompletionRequestPath(path);
+        const conversationBody = isConversationRequest
+            ? parseRequestBodyObject(serializedReq.body, serializedReq.headers)
+            : null;
+        const conversationIdentity = isConversationRequest
+            ? extractConversationIdentity(serializedReq.headers, conversationBody)
+            : null;
+        // Internal marker from the system proxy: source profile used only for
+        // local conversation attribution. Stripped here so it never reaches a seller.
+        delete serializedReq.headers[SYSTEM_PROXY_SOURCE_HEADER];
+        const chatPinnedModel = conversationIdentity
+            ? this._conversations.getPinnedModel(conversationIdentity.tool, conversationIdentity.sessionKey)
+                ?? (conversationIdentity.parentSessionKey
+                    ? this._conversations.getPinnedModel(conversationIdentity.tool, conversationIdentity.parentSessionKey)
+                    : null)
+            : null;
+        const effectiveRoutedModel = chatPinnedModel ?? this._defaultRoutedModel;
+        // Resolve the `antseed` model alias to the session's default route first,
+        // so the regular `<peerId>@<service>` pin rewrite below picks up the
+        // substituted value. Tool configs written by the desktop carry the alias
+        // so route changes apply to running sessions without config rewrites.
+        const aliasResult = substituteRoutedModelAlias(serializedReq.body, serializedReq.headers, effectiveRoutedModel);
+        if (aliasResult.aliasRequested && !aliasResult.substituted) {
+            log(`Request rejected: model alias "${ROUTED_MODEL_ALIAS}" with no default route set`);
+            res.writeHead(400, { 'content-type': 'application/json' });
+            res.end(JSON.stringify({
+                error: {
+                    type: 'no_default_route',
+                    code: 'no_default_route',
+                    message: `Model "${ROUTED_MODEL_ALIAS}" routes to the model selected in AntStation, but no route is set. `
+                        + 'Pick a model in the desktop app, or request "<peerId>@<model>" explicitly.',
+                    param: 'model',
+                },
+            }));
+            return;
+        }
+        if (aliasResult.substituted) {
+            serializedReq = { ...serializedReq, body: aliasResult.body, headers: aliasResult.headers };
+            log(`Model alias applied: ${ROUTED_MODEL_ALIAS} -> ${effectiveRoutedModel}${chatPinnedModel ? ' (chat pin)' : ''}`);
+        }
+        // System-proxy-intercepted tools can't carry the alias — the proxy
+        // rewrites their upstream model names to its connect-time route and
+        // marks the request. A chat pin must still win there, exactly as it
+        // does for alias-carrying configs; otherwise pinning a chat in the
+        // desktop silently does nothing for intercepted apps.
+        let chatPinOverrideApplied = false;
+        if (systemRoutedModel && !aliasResult.aliasRequested && chatPinnedModel) {
+            const pinOverride = overrideRoutedModelInBody(serializedReq.body, serializedReq.headers, chatPinnedModel);
+            if (pinOverride.overridden) {
+                serializedReq = { ...serializedReq, body: pinOverride.body, headers: pinOverride.headers };
+                chatPinOverrideApplied = true;
+                log(`Chat pin applied to system-routed model: -> ${chatPinnedModel}`);
+            }
+        }
+        // Track the conversation (subagent traffic rolls up into the parent
+        // chat). The snippet is only extracted the first time a chat is seen.
+        // A thread the tool opened for its own housekeeping is routed and paid
+        // for normally, but is not a chat — Codex titles every new chat from one.
+        if (conversationIdentity?.isUserThread) {
+            const rawModel = typeof conversationBody?.['model'] === 'string' ? conversationBody['model'] : '';
+            const resolvedModel = aliasResult.substituted
+                ? effectiveRoutedModel
+                : chatPinOverrideApplied
+                    ? chatPinnedModel
+                    : (parsePeerPinnedService(rawModel) ? rawModel : null);
+            const trackedKey = conversationIdentity.parentSessionKey ?? conversationIdentity.sessionKey;
+            const known = this._conversations.get(`${conversationIdentity.tool}:${trackedKey}`);
+            // A title turn runs on the tool's own small model and races ahead of the
+            // first real turn, so it must not become the chat's model: that pin
+            // outranks the default route for every later turn, which would strand
+            // the whole session on a model the user never picked.
+            const titleTurn = isTitleGenerationRequest(conversationBody);
+            const snippet = known?.snippet ? null : extractFirstUserSnippet(conversationBody);
+            // Some tools (T3/Claude) fire title-only requests when a new chat opens.
+            // Route them normally, but do not create a blank AntSeed conversation row
+            // until the first genuine user turn arrives.
+            if (known || !titleTurn) {
+                const tracked = this._conversations.touch({
+                    tool: conversationIdentity.tool,
+                    sessionKey: trackedKey,
+                    // Extract until a label sticks: a tool's title-generation request can
+                    // race ahead of the first real turn and create the row snippet-less.
+                    snippet,
+                    lastModel: titleTurn ? null : resolvedModel,
+                });
+                // Bind the request to the chat so its cost can be attributed when the
+                // payment layer signs for it (see _attributeSpend).
+                this._trackRequestConversation(serializedReq.requestId, tracked.id);
+            }
+        }
         const { body: servicePinBody, headers: servicePinHeaders, pinnedPeerId: bodyPinnedPeer, } = rewritePeerPinnedServiceInBody(serializedReq.body, serializedReq.headers);
         if (servicePinBody !== serializedReq.body) {
             serializedReq = { ...serializedReq, body: servicePinBody, headers: servicePinHeaders };
@@ -1335,6 +1649,7 @@ export class BuyerProxy {
             log(`Outbound request shape: ${summarizeRequestShape(requestForPeer)}`);
         }
         log(`Routing to peer ${selectedPeer.peerId.slice(0, 12)}...`);
+        this._markModelActivity();
         // Forward through P2P
         const wantsStreaming = clientWantsStreaming;
         const startTime = Date.now();
@@ -1364,6 +1679,7 @@ export class BuyerProxy {
                     onResponseChunk: (chunk) => {
                         if (!streamed)
                             return;
+                        this._markModelActivity();
                         const adaptedChunks = streamResponseAdapter
                             ? streamResponseAdapter.adaptChunk(chunk)
                             : [chunk];
@@ -1447,6 +1763,7 @@ export class BuyerProxy {
                     response = inject402PeerId(response, selectedPeer.peerId);
                 }
                 const latencyMs = Date.now() - startTime;
+                this._markModelActivity();
                 log(`Response: ${response.statusCode} (${latencyMs}ms, ${response.body.length} bytes)`);
                 if (response.statusCode >= 400) {
                     const prefix = adaptResponse ? 'Upstream adapted error detail' : 'Upstream error detail';