fix(ocr): live busy pre-check before posting a candidate's escrow

runOcrOneShot posted the escrow before calling the supplier, so a busy
supplier's 409 stranded an already-paid escrow until reclaim - and the
buyer-side busy-retry multiplies this into several stranded escrows per
page. Add a live /status check per candidate, immediately before its
escrow posts, mirroring sessions.ts's startChat pre-check: "working"
skips to the next candidate without posting, anything else (free,
unreachable, malformed, slow) is treated as usable so the check is an
optimisation, never a gate. Bounded to a 2s timeout so a stalled
supplier cannot meaningfully delay the request.

Existing candidate-fallback, escrow-ref capture, usage recording, and
receipt labels are unchanged.

README.md

@@ -1,3 +1,36 @@
-## Local Agents Marketplace
+# Local Agents Marketplace
 
-Local Agents Marketplace MVP. See `docs/ARCHITECTURE.md` for the architecture spec. Monorepo managed with pnpm workspaces. Work-in-progress; not production-ready.
+Bonded-escrow coordination for AI agent work on **Vector**, the Apex Fusion eUTXO L2.
+
+Buyers commission work. Suppliers claim it. Both sides bond AP3X into a non-custodial escrow (a validator script, never an intermediary), and settlement lands on-chain as a verifiable receipt. Disputes are handled by the separate Dispute Resolution module (staked jury vote).
+
+**Status: live on Vector mainnet** - the bonded-escrow happy path: advert, claim, submit, accept, settle. The contracts have been through the internal audit pipeline (methodology published in [vector-ai-agents](https://github.com/Apex-Fusion/vector-ai-agents)); they have not yet undergone independent third-party audit.
+
+## Flow
+
+1. **Advert** - a buyer posts a job with escrow bonded
+2. **Claim** - a supplier claims it, bonding their side
+3. **Submit** - work is delivered
+4. **Accept** - the buyer accepts the work
+5. **Settle** - the escrow releases; the settlement is a signed on-chain receipt
+
+## Repository layout
+
+| Path | What |
+|---|---|
+| `contracts/advert`, `contracts/escrow`, `contracts/marketplace` | Aiken validators |
+| `docs/ARCHITECTURE.md` | Architecture spec |
+| `docs/buyer/`, `docs/supplier/` | Role guides |
+| `deploy/` | Deployment and simulation tooling |
+
+Monorepo managed with pnpm workspaces.
+
+## Getting started
+
+Developer docs: [Vector AI documentation](https://apex-fusion.github.io/vector-ai-documentation/). For the contract-level picture, start with [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md).
+
+## Related
+
+- [mcp-server](https://github.com/Apex-Fusion/mcp-server) - hosted MCP server for Vector (testnet and mainnet)
+- [agent-sdk-py](https://github.com/Apex-Fusion/agent-sdk-py) / [agent-sdk-ts](https://github.com/Apex-Fusion/agent-sdk-ts) - Python and TypeScript SDKs
+- [vector-agent-modules](https://github.com/Apex-Fusion/vector-agent-modules) - dispute resolution, reputation staking, self-improvement

buyer/Dockerfile

@@ -35,6 +35,8 @@ RUN apk add --no-cache python3 make g++
 RUN corepack enable && corepack prepare pnpm@10.33.2 --activate
 
 COPY pnpm-workspace.yaml pnpm-lock.yaml package.json tsconfig.base.json .npmrc ./
+# pnpm-lock.yaml's patchedDependencies hash the patch files at install time.
+COPY patches ./patches
 COPY packages/shared/package.json ./packages/shared/package.json
 COPY buyer/package.json ./buyer/package.json
 COPY supplier/package.json ./supplier/package.json

buyer/scripts/monitor-wallets.ts

@@ -83,6 +83,48 @@ const OPERATOR_SOURCES: OperatorSource[] = [
     keyVar: "SUPPLIER_PRIV_KEY_HEX",
     role: "supplier",
   },
+  {
+    label: "supplier-kimi-2@vector-marketplace",
+    host: "root@91.98.147.172",
+    envPath: "/root/agents-marketplace/supplier/.env.kimi-2",
+    keyVar: "SUPPLIER_PRIV_KEY_HEX",
+    role: "supplier",
+  },
+  {
+    label: "supplier-kimi-3@vector-marketplace",
+    host: "root@91.98.147.172",
+    envPath: "/root/agents-marketplace/supplier/.env.kimi-3",
+    keyVar: "SUPPLIER_PRIV_KEY_HEX",
+    role: "supplier",
+  },
+  {
+    label: "supplier-kimi-4@vector-marketplace",
+    host: "root@91.98.147.172",
+    envPath: "/root/agents-marketplace/supplier/.env.kimi-4",
+    keyVar: "SUPPLIER_PRIV_KEY_HEX",
+    role: "supplier",
+  },
+  {
+    label: "supplier-deepseek-flash-1@vector-marketplace",
+    host: "root@91.98.147.172",
+    envPath: "/root/agents-marketplace/supplier/.env.deepseek-flash-1",
+    keyVar: "SUPPLIER_PRIV_KEY_HEX",
+    role: "supplier",
+  },
+  {
+    label: "supplier-deepseek-flash-2@vector-marketplace",
+    host: "root@91.98.147.172",
+    envPath: "/root/agents-marketplace/supplier/.env.deepseek-flash-2",
+    keyVar: "SUPPLIER_PRIV_KEY_HEX",
+    role: "supplier",
+  },
+  {
+    label: "supplier-deepseek-flash-3@vector-marketplace",
+    host: "root@91.98.147.172",
+    envPath: "/root/agents-marketplace/supplier/.env.deepseek-flash-3",
+    keyVar: "SUPPLIER_PRIV_KEY_HEX",
+    role: "supplier",
+  },
   {
     label: "buyer@open-webui",
     host: "open-webui",

buyer/src/config.ts

@@ -36,6 +36,12 @@
  *                         Defaults to "./data/archive" relative to cwd.
  *                         When unset, archive is disabled (responses are
  *                         not persisted; /v1/responses* return 503).
+ *   GATEWAY_PUBLIC_URL  — Public base URL of the OpenAI-compatible gateway
+ *                         (e.g. https://api.marketplace.vector.apexfusion.org).
+ *                         Injected into the SPA boot block so the "Generate
+ *                         API key" page knows where to POST /signup. When
+ *                         unset, the SPA falls back to deriving "api." +
+ *                         current host.
  */
 
 const HEX64_RE = /^[0-9a-fA-F]{64}$/;
@@ -56,6 +62,7 @@ export interface BuyerConfig {
   password: string;
   sessionSecret: string;
   cookieSecure: boolean;
+  gatewayPublicUrl: string;
 }
 
 function requireField(env: Record<string, string | undefined>, name: string): string {
@@ -141,6 +148,11 @@ export function loadConfig(env: Record<string, string | undefined>): BuyerConfig
     cookieSecure = cookieSecureStr === "1";
   }
 
+  // GATEWAY_PUBLIC_URL — optional. Public base URL of the gateway, surfaced to
+  // the SPA via the boot block (no trailing slash). Empty string means the SPA
+  // derives the URL from its own host ("api." + location.host).
+  const gatewayPublicUrl = (env.GATEWAY_PUBLIC_URL ?? "").trim().replace(/\/+$/, "");
+
   return {
     privKeyHex,
     indexerUrl,
@@ -155,5 +167,6 @@ export function loadConfig(env: Record<string, string | undefined>): BuyerConfig
     password,
     sessionSecret,
     cookieSecure,
+    gatewayPublicUrl,
   };
 }

buyer/src/index.ts

@@ -133,6 +133,7 @@ export async function runMain(env: Record<string, string | undefined>): Promise<
     password: config.password,
     sessionSecret: config.sessionSecret,
     cookieSecure: config.cookieSecure,
+    gatewayPublicUrl: config.gatewayPublicUrl,
     jobStore,
     pdfCaps,
   });

buyer/src/sdk/Marketplace.ts

@@ -73,8 +73,10 @@ import {
   buildPostEscrowTx,
   buildPostTtsEscrowTx,
   buildPostChatEscrowTx,
+  buildPostOcrEscrowTx,
   ttsPromptHash,
   chatSessionPromptHash,
+  ocrPromptHash,
   buildAcceptTx,
   buildReclaimTx,
   TxConstructionError,
@@ -88,8 +90,11 @@ import type {
   SubmitPromptResult,
   SubmitTtsOptions,
   SubmitTtsResult,
+  SubmitOcrOptions,
+  SubmitOcrResult,
   StartChatOptions,
   StartChatResult,
+  ChatSettleMode,
   EndChatOptions,
   EndChatResult,
   AcceptResultOptions,
@@ -806,6 +811,278 @@ export class Marketplace extends EventEmitterBase {
     };
   }
 
+  // ─── submitOcr — full lifecycle for ocr.page.extract.<model-slug>.v1 ──
+
+  async submitOcr(opts: SubmitOcrOptions): Promise<SubmitOcrResult> {
+    const { advertRef, image_b64, mime, output_format, payment_lovelace } = opts;
+    const request = { image_b64, mime, output_format };
+
+    let advertDatum: AdvertDatum | null = null;
+    let escrowOutputRef: OutputReference | null = null;
+    let escrowRefStr = "";
+    let postedAtMs = Date.now();
+
+    const previewStr = `[ocr:${mime} ${output_format} ${image_b64.length}b64]`;
+    const recordFailure = (reason: string): void => {
+      this.historyStore.save({
+        escrow_ref: escrowRefStr || `${"0".repeat(64)}#0`,
+        supplier_pkh: advertDatum?.supplier_pkh ?? "",
+        capability_id: advertDatum?.capability_id ?? "",
+        prompt_preview: previewStr,
+        posted_at: postedAtMs,
+        status: "failed",
+        failure_reason: reason,
+      });
+    };
+
+    // ── 1. Resolve advert + post escrow (OCR prompt_hash) ─────────────
+    let escrowResult;
+    try {
+      const utxo = await this.chain.queryUtxo(advertRef);
+      if (utxo && utxo.datumHex) {
+        try {
+          advertDatum = decodeAdvertDatum(utxo.datumHex);
+        } catch { /* builder will re-throw structurally */ }
+      }
+      escrowResult = await buildPostOcrEscrowTx({
+        chain: this.chain,
+        buyerKey: this.walletKey,
+        advertRef,
+        request,
+        payment_lovelace,
+      });
+    } catch (err) {
+      if (err instanceof TxConstructionError) {
+        recordFailure(err.reason);
+        throw err;
+      }
+      this.emitProgress({ type: "chain_submit_failed", detail: (err as Error).message });
+      recordFailure((err as Error).message);
+      throw err;
+    }
+
+    escrowOutputRef = escrowResult.escrowOutputRef;
+    escrowRefStr = refToString(escrowOutputRef);
+
+    this.emitProgress({ type: "escrow_posted", escrow_ref: escrowRefStr });
+
+    // ── 2. Wait for the escrow tx to confirm ──────────────────────────
+    try {
+      await this.chain.awaitTx(escrowResult.expectedTxHash, 120_000);
+    } catch { /* mock providers / tests */ }
+
+    // ── 3. Re-fetch escrow datum to capture posted_at ─────────────────
+    try {
+      const escrowUtxo = await this.chain.queryUtxo(escrowOutputRef);
+      if (escrowUtxo && escrowUtxo.datumHex) {
+        const ed = decodeEscrowDatum(escrowUtxo.datumHex);
+        postedAtMs = ed.posted_at;
+      }
+    } catch { /* metadata best-effort */ }
+
+    if (!advertDatum) {
+      const reason = "advert datum unavailable after escrow post";
+      recordFailure(reason);
+      throw new ReceiptVerificationError(reason);
+    }
+
+    // ── 4. Call supplier /v1/ocr/extract ──────────────────────────────
+    const supplierBaseUrl = advertDatum.endpoint_url;
+    const supplierHttp = new HttpClient({
+      baseUrl: supplierBaseUrl,
+      fetch: this.fetchImpl,
+    });
+    const ocrBody = { image_b64, mime, output_format };
+
+    let ocrResult;
+    try {
+      ocrResult = await supplierHttp.postJson("/v1/ocr/extract", ocrBody, {
+        headers: { "X-Escrow-Ref": escrowRefStr },
+      });
+    } catch (err) {
+      if (err instanceof HttpError) {
+        const reason = err.kind === "timeout" ? "timeout" : "network_error";
+        const sErr = new SupplierError(reason, { message: err.message });
+        recordFailure(reason);
+        throw sErr;
+      }
+      recordFailure((err as Error).message);
+      throw err;
+    }
+
+    // 202 + poll mirror of submitTts's async handling. OCR inference for a
+    // page runs seconds; the long tail is the Submit tx confirmation.
+    if (ocrResult.status === 202 && ocrResult.body && typeof ocrResult.body === "object") {
+      const jobId = (ocrResult.body as { job_id?: string }).job_id;
+      if (typeof jobId !== "string" || jobId.length === 0) {
+        const sErr = new SupplierError("malformed_response", {
+          status: ocrResult.status,
+          message: "supplier 202 response missing job_id",
+        });
+        recordFailure("malformed_response");
+        throw sErr;
+      }
+      const POLL_INTERVAL_MS = 2_000;
+      const POLL_TIMEOUT_MS = 180_000;
+      const pollDeadline = Date.now() + POLL_TIMEOUT_MS;
+      let polled = ocrResult;
+      let firstIter = true;
+      while (Date.now() < pollDeadline) {
+        if (!firstIter) {
+          await new Promise<void>((r) => setTimeout(r, POLL_INTERVAL_MS));
+        }
+        firstIter = false;
+        try {
+          polled = await supplierHttp.getJson(`/v1/ocr/extract/${jobId}`);
+        } catch (err) {
+          if (err instanceof HttpError) {
+            const reason = err.kind === "timeout" ? "timeout" : "network_error";
+            const sErr = new SupplierError(reason, { message: err.message });
+            recordFailure(reason);
+            throw sErr;
+          }
+          throw err;
+        }
+        if (polled.status === 200) break;
+        if (polled.status === 202) continue;
+        const failBody = (polled.body && typeof polled.body === "object")
+          ? polled.body as { reason?: string; message?: string }
+          : {};
+        const sErr = new SupplierError(failBody.reason ?? "supplier_http_error", {
+          status: polled.status,
+          message: failBody.message ?? `supplier returned ${polled.status}`,
+        });
+        recordFailure(sErr.reason);
+        throw sErr;
+      }
+      if (polled.status !== 200) {
+        const sErr = new SupplierError("timeout", {
+          status: polled.status,
+          message: `supplier job ${jobId} not done within ${POLL_TIMEOUT_MS}ms`,
+        });
+        recordFailure("timeout");
+        throw sErr;
+      }
+      ocrResult = polled;
+    }
+
+    if (!ocrResult.ok) {
+      const bodyReason =
+        ocrResult.body && typeof ocrResult.body === "object"
+          ? ((ocrResult.body as { reason?: string }).reason ?? "supplier_http_error")
+          : "supplier_http_error";
+      const sErr = new SupplierError(bodyReason, {
+        status: ocrResult.status,
+        message: `supplier returned ${ocrResult.status}`,
+      });
+      recordFailure(bodyReason);
+      throw sErr;
+    }
+    if (ocrResult.parseError || !ocrResult.body || typeof ocrResult.body !== "object") {
+      const sErr = new SupplierError("malformed_response", {
+        status: ocrResult.status,
+        message: "supplier body is not valid JSON",
+      });
+      recordFailure("malformed_response");
+      throw sErr;
+    }
+
+    const responseBody = ocrResult.body as {
+      output_format?: string;
+      content?: string;
+      usage?: { prompt_tokens?: number; completion_tokens?: number; total_tokens?: number };
+      receipt?: import("@marketplace/shared/receipt").Receipt;
+      receipt_signature?: string;
+    };
+    if (!responseBody.content || !responseBody.receipt || !responseBody.receipt_signature) {
+      const sErr = new SupplierError("malformed_response", {
+        status: ocrResult.status,
+        message: "supplier response is missing content / receipt / receipt_signature",
+      });
+      recordFailure("malformed_response");
+      throw sErr;
+    }
+    const receipt = responseBody.receipt;
+    const receiptSignature = responseBody.receipt_signature;
+
+    this.emitProgress({ type: "supplier_called", escrow_ref: escrowRefStr });
+
+    // ── 5. Receipt verification ───────────────────────────────────────
+    if (receipt.supplier_pkh !== advertDatum.supplier_pkh) {
+      recordFailure("wrong_supplier");
+      throw new ReceiptVerificationError("wrong_supplier");
+    }
+    if (receipt.escrow_ref !== escrowRefStr) {
+      recordFailure("wrong_escrow_ref");
+      throw new ReceiptVerificationError("wrong_escrow_ref");
+    }
+    // The OCR prompt commitment uses the SAME canonical hash the supplier
+    // validated against the escrow datum. Recompute from the envelope we
+    // sent; a mismatch means the supplier signed a different request.
+    const expectedPromptHash = ocrPromptHash(request);
+    if (receipt.prompt_hash !== expectedPromptHash) {
+      recordFailure("prompt_hash_mismatch");
+      throw new ReceiptVerificationError("prompt_hash_mismatch");
+    }
+    if (receipt.model !== advertDatum.model) {
+      recordFailure("request_spec_hash_mismatch");
+      throw new ReceiptVerificationError("request_spec_hash_mismatch");
+    }
+    // The receipt's response_hash must commit the content we actually
+    // received in the requested output shape.
+    const expectedResponseHash = nodeCrypto.createHash("sha256")
+      .update(canonicalize({
+        content: responseBody.content,
+        output_format: responseBody.output_format ?? output_format,
+      }), "utf8")
+      .digest("hex");
+    if (receipt.response_hash !== expectedResponseHash) {
+      recordFailure("response_hash_mismatch");
+      throw new ReceiptVerificationError("response_hash_mismatch");
+    }
+    if (typeof receiptSignature !== "string" || !SIG_RE.test(receiptSignature)) {
+      recordFailure("invalid_signature");
+      throw new ReceiptVerificationError("invalid_signature");
+    }
+    if (receiptSignature === ZERO_SIGNATURE) {
+      recordFailure("invalid_signature");
+      throw new ReceiptVerificationError("invalid_signature");
+    }
+    if (!HEX64_RE.test(receipt.prompt_hash) || !HEX64_RE.test(receipt.response_hash)) {
+      recordFailure("malformed_receipt");
+      throw new ReceiptVerificationError("malformed_receipt");
+    }
+
+    this.emitProgress({ type: "receipt_verified", escrow_ref: escrowRefStr });
+
+    this.historyStore.save({
+      escrow_ref: escrowRefStr,
+      supplier_pkh: advertDatum.supplier_pkh,
+      capability_id: advertDatum.capability_id,
+      prompt_preview: previewStr,
+      posted_at: postedAtMs,
+      status: "completed",
+      response: responseBody.content.length <= 200
+        ? responseBody.content
+        : responseBody.content.slice(0, 200),
+      receipt,
+      receipt_signature: receiptSignature,
+    });
+
+    return {
+      output_format: responseBody.output_format ?? output_format,
+      content: responseBody.content,
+      usage: {
+        prompt_tokens: responseBody.usage?.prompt_tokens ?? 0,
+        completion_tokens: responseBody.usage?.completion_tokens ?? 0,
+        total_tokens: responseBody.usage?.total_tokens ?? 0,
+      },
+      receipt,
+      receiptSignature,
+      escrowRef: escrowOutputRef,
+    };
+  }
+
   // ─── startChat / endChat — multi-turn llm.chat.v1 lifecycle ────────────
 
   /**
@@ -899,7 +1176,13 @@ export class Marketplace extends EventEmitterBase {
     }
 
     this.emitProgress({ type: "chat_started", escrow_ref: escrowRefStr });
-    return { escrowRef: escrowOutputRef, sessionNonce, supplierBaseUrl: advertDatum.endpoint_url };
+    // Ticket-mode suppliers answer {status:"ticket"} (no Claim); anything else
+    // — "claimed" or an older supplier without the field — is full settlement.
+    const startStatus = startRes.body && typeof startRes.body === "object"
+      ? (startRes.body as { status?: string }).status
+      : undefined;
+    const settleMode: ChatSettleMode = startStatus === "ticket" ? "ticket" : "full";
+    return { escrowRef: escrowOutputRef, sessionNonce, supplierBaseUrl: advertDatum.endpoint_url, settleMode };
   }
 
   /**
@@ -937,10 +1220,20 @@ export class Marketplace extends EventEmitterBase {
       });
     }
     const endBody = endRes.body as {
+      status?: string;
+      settle_mode?: string;
       receipt?: import("@marketplace/shared/receipt").Receipt;
       receipt_signature?: string;
       submitted_ref?: string;
     };
+    // Ticket sessions produce no receipt and nothing to Accept — the Open
+    // escrow returns via reclaim after deliver_by. Trusting the supplier's
+    // response is incentive-safe: lying in either direction only forfeits the
+    // supplier's own payout (the buyer's sweeper reclaims unsettled escrows).
+    if (endBody.settle_mode === "ticket") {
+      this.emitProgress({ type: "chat_ended", escrow_ref: escrowRefStr });
+      return { settleMode: "ticket", escrowRef };
+    }
     if (!endBody.receipt || !endBody.receipt_signature) {
       throw new SupplierError("malformed_response", { message: "supplier end response missing receipt" });
     }
@@ -1022,7 +1315,7 @@ export class Marketplace extends EventEmitterBase {
     this.emitProgress({ type: "accept_submitted", escrow_ref: refToString(acceptedRef) });
     this.emitProgress({ type: "chat_ended", escrow_ref: escrowRefStr });
 
-    return { receipt, receiptSignature, escrowRef, acceptedRef };
+    return { settleMode: "full", receipt, receiptSignature, escrowRef, acceptedRef };
   }
 
   // ─── acceptResult / reclaim ───────────────────────────────────────────

buyer/src/sdk/index.ts

@@ -16,6 +16,9 @@ export type {
   AcceptResultOptions,
   ReclaimOptions,
   GetTaskHistoryOptions,
+  ChatSettleMode,
+  StartChatResult,
+  EndChatResult,
 } from "./types.js";
 export { ReceiptVerificationError, IndexerError, SupplierError } from "./types.js";
 export type { TaskHistoryStore } from "./history.js";

buyer/src/sdk/types.ts

@@ -187,6 +187,35 @@ export interface SubmitTtsResult {
   escrowRef: OutputReference;
 }
 
+/**
+ * SubmitOcrOptions — arguments to Marketplace.submitOcr() for model-scoped
+ * `ocr.page.extract.<model-slug>.v1` capabilities. The same `{image_b64,
+ * mime, output_format}` envelope is hashed (sha256 ∘ canonical) into the
+ * escrow datum's `prompt_hash` AND posted to the supplier's /v1/ocr/extract
+ * body, so the supplier can verify the request matches the on-chain commit.
+ */
+export interface SubmitOcrOptions {
+  advertRef: OutputReference;
+  image_b64: string;   // plain base64, no data: prefix
+  mime: string;        // image/png | image/jpeg | image/webp
+  output_format: string; // markdown | html | json
+  payment_lovelace: bigint;
+}
+
+/**
+ * SubmitOcrResult — returned from Marketplace.submitOcr() on success.
+ * `content` is the extraction in the requested output_format; the receipt's
+ * response_hash commits sha256(canonical({content, output_format})).
+ */
+export interface SubmitOcrResult {
+  output_format: string;
+  content: string;
+  usage: { prompt_tokens: number; completion_tokens: number; total_tokens: number };
+  receipt: Receipt;
+  receiptSignature: string;
+  escrowRef: OutputReference;
+}
+
 /**
  * StartChatOptions / StartChatResult — Marketplace.startChat() for the
  * `llm.chat.v1` capability. Opens the escrow (session-init prompt_hash) and
@@ -197,6 +226,11 @@ export interface StartChatOptions {
   advertRef: OutputReference;
   payment_lovelace: bigint;
 }
+/** Chain settlement mode a chat session runs under (reported by the supplier
+ * at /v1/chat/start): "full" = Claim/Submit/Accept lifecycle; "ticket" = the
+ * buyer's escrow post is the only chain op (reclaimed after deliver_by). */
+export type ChatSettleMode = "full" | "ticket";
+
 export interface StartChatResult {
   escrowRef: OutputReference;
   /** Random nonce committed into the escrow's session-init prompt_hash. The
@@ -204,6 +238,8 @@ export interface StartChatResult {
   sessionNonce: string;
   /** Supplier endpoint_url, cached by the buyer-app for message routing. */
   supplierBaseUrl: string;
+  /** How the supplier settles this session ("full" when it Claimed). */
+  settleMode: ChatSettleMode;
 }
 
 /**
@@ -220,13 +256,21 @@ export interface EndChatOptions {
   /** The browser's local transcript mirror, used to verify response_hash. */
   transcript?: import("@marketplace/shared/tx").ChatMessage[];
 }
-export interface EndChatResult {
-  receipt: Receipt;
-  receiptSignature: string;
-  escrowRef: OutputReference;
-  /** The Submitted escrow UTxO that was Accepted. */
-  acceptedRef: OutputReference;
-}
+export type EndChatResult =
+  | {
+      settleMode: "full";
+      receipt: Receipt;
+      receiptSignature: string;
+      escrowRef: OutputReference;
+      /** The Submitted escrow UTxO that was Accepted. */
+      acceptedRef: OutputReference;
+    }
+  | {
+      /** Ticket session: nothing was Submitted/Accepted — no receipt exists.
+       * The Open escrow returns to the buyer via reclaim after deliver_by. */
+      settleMode: "ticket";
+      escrowRef: OutputReference;
+    };
 
 /**
  * AcceptResultOptions — arguments to Marketplace.acceptResult().

buyer/src/server.ts

@@ -109,6 +109,11 @@ export interface AppDeps {
    * brute-force rate-limit window. Defaults to Date.now. Tests use a fake
    * clock to drive expiry and rolling-renewal assertions deterministically. */
   nowMs?: () => number;
+  /** Public base URL of the OpenAI-compatible gateway, injected into the SPA
+   * boot block so the "Generate API key" page can POST cross-origin to
+   * `${gatewayUrl}/signup`. Public, non-secret value. When empty/omitted the
+   * SPA derives "api." + its own host. */
+  gatewayPublicUrl?: string;
   /** PDF book summarizer job registry. When provided, the /v1/pdf-* routes
    * are live; when omitted they respond 503 (feature disabled). Built in
    * runMain from the same marketplace/chain/wallet/archive deps. */
@@ -680,9 +685,11 @@ export function createApp(deps: AppDeps): Express {
   const DEMO_CHAT_MODEL = "moonshotai/kimi-k2.6";
 
   function isChatMessageArray(v: unknown): v is ChatMessage[] {
+    // The free chat demo stays text-only: the `tool` role (added to
+    // ChatMessage for the gateway demo endpoint) is not accepted here.
     return Array.isArray(v) && v.length > 0 && v.every(
       (m) => m && typeof m === "object"
-        && (["system", "user", "assistant"] as const).includes((m as ChatMessage).role)
+        && ["system", "user", "assistant"].includes((m as ChatMessage).role)
         && typeof (m as ChatMessage).content === "string",
     );
   }
@@ -927,6 +934,14 @@ export function createApp(deps: AppDeps): Express {
         supplierBaseUrl: state.supplierBaseUrl,
       });
       chatRouteState.delete(rawRef);
+      if (result.settleMode === "ticket") {
+        // Ticket session: no receipt/Accept — funds return via reclaim.
+        return res.status(200).json({
+          status: "ended",
+          escrow_ref: rawRef,
+          settle_mode: "ticket",
+        });
+      }
       return res.status(200).json({
         status: "accepted",
         escrow_ref: rawRef,
@@ -1310,6 +1325,9 @@ export function createApp(deps: AppDeps): Express {
               pubKeyHash: resolved.walletKey.pubKeyHash,
               address: displayAddress,
             },
+            // Public gateway base URL for the "Generate API key" page. Empty
+            // string when unconfigured — the SPA then derives "api." + host.
+            gatewayUrl: deps.gatewayPublicUrl ?? "",
           })};</script>`
         : "";
       const indexWithBoot = cachedTemplate.replace(

buyer/src/ui/App.tsx

@@ -6,6 +6,7 @@
  *   /tasks    → <TaskHistory />
  *   /pending  → <PendingReceipts />     UX-1: buyer-side Accept button
  *   /wallet   → <Wallet />
+ *   /api-keys → <ApiKeys />             self-serve gateway API key generation
  */
 
 import { Routes, Route, Link } from "react-router-dom";
@@ -14,6 +15,7 @@ import TaskHistory from "./pages/TaskHistory.js";
 import Wallet from "./pages/Wallet.js";
 import PendingReceipts from "./pages/PendingReceipts.js";
 import BookSummarizer from "./pages/BookSummarizer.js";
+import ApiKeys from "./pages/ApiKeys.js";
 import { RequireAuth, useAuth } from "./state/AuthContext.js";
 
 function SignOutButton() {
@@ -35,10 +37,13 @@ export default function App() {
       <div className="min-h-screen bg-gray-50 text-gray-900">
         <nav className="bg-white border-b border-gray-200 px-6 py-3 flex items-center gap-6">
           <Link to="/" className="font-medium hover:underline">Dashboard</Link>
-          <Link to="/book" className="font-medium hover:underline">Book Summarizer</Link>
+          {/* Book Summarizer tab hidden for now — the /book route and all PDF
+              summarize code remain intact; restore by re-adding this link:
+          <Link to="/book" className="font-medium hover:underline">Book Summarizer</Link> */}
           <Link to="/tasks" className="font-medium hover:underline">Tasks</Link>
           <Link to="/pending" className="font-medium hover:underline">Pending</Link>
           <Link to="/wallet" className="font-medium hover:underline">Wallet</Link>
+          <Link to="/api-keys" className="font-medium hover:underline">API Keys</Link>
           <SignOutButton />
         </nav>
         <main className="p-6">
@@ -48,6 +53,7 @@ export default function App() {
             <Route path="/tasks" element={<TaskHistory />} />
             <Route path="/pending" element={<PendingReceipts />} />
             <Route path="/wallet" element={<Wallet />} />
+            <Route path="/api-keys" element={<ApiKeys />} />
           </Routes>
         </main>
       </div>

buyer/src/ui/components/ChatForm.tsx

@@ -267,7 +267,18 @@ export default function ChatForm({ advertRef, payment_lovelace }: ChatFormProps
         } catch { /* keep fallback */ }
         throw new Error(msg);
       }
-      setCharged(payment_lovelace !== undefined ? ap3x(payment_lovelace) : "0");
+      let endBody: { settle_mode?: string } = {};
+      try {
+        endBody = (await resp.json()) as { settle_mode?: string };
+      } catch { /* older buyer-app responses stay full-settle */ }
+      // Ticket sessions never charge — the escrow returns via reclaim.
+      setCharged(
+        endBody.settle_mode === "ticket"
+          ? "0 (ticket session — escrow reclaimed)"
+          : payment_lovelace !== undefined
+            ? ap3x(payment_lovelace)
+            : "0",
+      );
     } catch (err) {
       setError(err instanceof Error ? err.message : String(err));
     } finally {

buyer/src/ui/components/JobEstimate.tsx

@@ -94,7 +94,7 @@ export default function JobEstimate({
           <dd className="text-right">{est.mapCalls}</dd>
           <dt>Reduce calls</dt>
           <dd className="text-right">{est.reduceCalls}</dd>
-          <dt className="font-medium">Total paid calls</dt>
+          <dt className="font-medium">Total escrowed calls</dt>
           <dd className="text-right font-medium">{est.totalCalls}</dd>
           <dt className="font-medium">Est. max cost</dt>
           <dd className="text-right font-medium">{est.totalAp3x} AP3X</dd>
@@ -126,7 +126,7 @@ export default function JobEstimate({
         disabled={blocked || starting}
         className="rounded bg-indigo-600 px-4 py-2 text-sm font-medium text-white hover:bg-indigo-700 disabled:bg-gray-400"
       >
-        {starting ? "Starting…" : `Proceed & pay (~${est.totalAp3x} AP3X)`}
+        {starting ? "Starting…" : `Proceed (locks ~${est.totalAp3x} AP3X in escrow)`}
       </button>
     </div>
   );

buyer/src/ui/components/PdfUploadForm.tsx

@@ -50,7 +50,7 @@ export default function PdfUploadForm({ onUploaded }: { onUploaded: (info: Uploa
   return (
     <form onSubmit={handleSubmit} className="space-y-4">
       <p className="text-sm text-gray-600">
-        Upload a text-based PDF book. The buyer will chunk it and pay marketplace
+        Upload a text-based PDF book. The buyer will chunk it and commission marketplace
         suppliers to summarize each chunk, then synthesize a full-book summary.
       </p>
       <input

buyer/src/ui/gateway.ts

@@ -0,0 +1,32 @@
+/**
+ * buyer/src/ui/gateway.ts — resolve the OpenAI-compatible gateway base URL the
+ * SPA should call for the "Generate API key" page.
+ *
+ * Resolution order:
+ *   1. window.__BUYER_BOOT__.gatewayUrl — injected by the buyer server from the
+ *      GATEWAY_PUBLIC_URL env (the configured value). Present on any page load
+ *      made while the buyer_session cookie is valid.
+ *   2. Derived "api." + current host — the deploy convention is exactly
+ *      api.<buyer-host> (marketplace.… → api.marketplace.…). This covers the
+ *      gap right after a fresh login, when the SPA flips auth state WITHOUT a
+ *      page reload so the boot block (which is only injected into
+ *      server-rendered HTML) hasn't been re-fetched yet.
+ *
+ * The result never has a trailing slash.
+ */
+
+interface BuyerBootMaybeGateway {
+  gatewayUrl?: string;
+}
+
+function stripTrailingSlashes(url: string): string {
+  return url.replace(/\/+$/, "");
+}
+
+export function resolveGatewayUrl(): string {
+  const boot = (window as unknown as { __BUYER_BOOT__?: BuyerBootMaybeGateway })
+    .__BUYER_BOOT__;
+  const configured = boot?.gatewayUrl?.trim();
+  if (configured) return stripTrailingSlashes(configured);
+  return `${window.location.protocol}//api.${window.location.host}`;
+}

buyer/src/ui/pages/ApiKeys.tsx

@@ -0,0 +1,205 @@
+/**
+ * buyer/src/ui/pages/ApiKeys.tsx — self-serve "Generate API key" page.
+ *
+ * Mints a custodial gateway API key by POSTing directly to the gateway's
+ * public /signup endpoint (cross-origin; the gateway CORS-allows this origin).
+ * The raw key is shown ONCE — it is never stored server-side in retrievable
+ * form, so it cannot be listed or recovered later. The page also surfaces the
+ * funding deposit address and a ready-to-paste usage snippet.
+ *
+ * Access is gated by the SPA's existing operator login (this page only renders
+ * inside <RequireAuth>). The gateway /signup itself stays public + IP
+ * rate-limited; per-user accounts are a later feature.
+ */
+
+import { useState } from "react";
+import { resolveGatewayUrl } from "../gateway.js";
+
+interface SignupResult {
+  api_key: string;
+  key_prefix: string;
+  deposit_address: string;
+}
+
+/** Pull a human-readable message out of whatever error shape the gateway
+ * returns (OpenAI `{error:{message}}`, plain `{error|message}`, or neither). */
+function errorMessage(body: unknown, status: number, statusText: string): string {
+  if (body && typeof body === "object") {
+    const b = body as Record<string, unknown>;
+    const err = b.error;
+    if (err && typeof err === "object" && typeof (err as Record<string, unknown>).message === "string") {
+      return (err as Record<string, string>).message;
+    }
+    if (typeof err === "string") return err;
+    if (typeof b.message === "string") return b.message;
+  }
+  return `${status} ${statusText}`.trim() || "request failed";
+}
+
+function CopyButton({ value, label }: { value: string; label?: string }) {
+  const [copied, setCopied] = useState(false);
+  return (
+    <button
+      type="button"
+      onClick={() => {
+        void navigator.clipboard
+          ?.writeText(value)
+          .then(() => {
+            setCopied(true);
+            setTimeout(() => setCopied(false), 1500);
+          })
+          .catch(() => {
+            /* clipboard blocked (insecure context / permissions) — no-op */
+          });
+      }}
+      className="rounded border border-gray-300 px-2 py-0.5 text-xs text-gray-600 hover:bg-gray-50"
+    >
+      {copied ? "Copied!" : (label ?? "Copy")}
+    </button>
+  );
+}
+
+export default function ApiKeys() {
+  const [label, setLabel] = useState("");
+  const [loading, setLoading] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+  const [result, setResult] = useState<SignupResult | null>(null);
+
+  const gatewayUrl = resolveGatewayUrl();
+  const baseUrl = `${gatewayUrl}/openai/v1`;
+
+  async function generate() {
+    setLoading(true);
+    setError(null);
+    setResult(null);
+    try {
+      const res = await fetch(`${gatewayUrl}/signup`, {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify(label.trim() ? { label: label.trim() } : {}),
+      });
+      const body = (await res.json().catch(() => null)) as unknown;
+      if (!res.ok) {
+        throw new Error(errorMessage(body, res.status, res.statusText));
+      }
+      const r = body as Partial<SignupResult> | null;
+      if (!r || typeof r.api_key !== "string" || typeof r.deposit_address !== "string") {
+        throw new Error("unexpected response from gateway");
+      }
+      setResult({
+        api_key: r.api_key,
+        key_prefix: r.key_prefix ?? r.api_key.slice(0, 12),
+        deposit_address: r.deposit_address,
+      });
+    } catch (e) {
+      setError(e instanceof Error ? e.message : "could not reach the gateway");
+    } finally {
+      setLoading(false);
+    }
+  }
+
+  const curlSnippet =
+    result === null
+      ? ""
+      : [
+          `curl ${baseUrl}/chat/completions \\`,
+          `  -H "Authorization: Bearer ${result.api_key}" \\`,
+          `  -H "Content-Type: application/json" \\`,
+          `  -d '{"model": "<model>", "messages": [{"role": "user", "content": "Hello"}]}'`,
+        ].join("\n");
+
+  return (
+    <div className="max-w-3xl space-y-6">
+      <div className="space-y-1">
+        <h1 className="text-2xl font-semibold">API Keys</h1>
+        <p className="text-sm text-gray-600">
+          Generate an OpenAI-compatible API key for the marketplace gateway. Each key has its own
+          gateway-managed wallet (the gateway holds this wallet's key) - fund its deposit address with AP3X, then point any OpenAI SDK at{" "}
+          <code className="font-mono text-xs">{baseUrl}</code>.
+        </p>
+      </div>
+
+      <div className="rounded-lg border border-gray-200 bg-white p-4 space-y-3">
+        <label className="block text-sm font-medium text-gray-700">
+          Label <span className="font-normal text-gray-400">(optional)</span>
+          <input
+            type="text"
+            value={label}
+            onChange={(e) => setLabel(e.target.value)}
+            placeholder="e.g. my-laptop"
+            maxLength={120}
+            disabled={loading}
+            className="mt-1 block w-full rounded border border-gray-300 px-3 py-2 text-sm focus:border-gray-400 focus:outline-none disabled:opacity-50"
+          />
+        </label>
+        <button
+          type="button"
+          onClick={() => void generate()}
+          disabled={loading}
+          className="rounded bg-gray-900 px-4 py-2 text-sm font-medium text-white hover:bg-gray-700 disabled:opacity-50"
+          data-testid="generate-api-key"
+        >
+          {loading ? "Generating…" : "Generate API key"}
+        </button>
+        {error !== null && (
+          <p className="text-sm text-red-600" data-testid="api-key-error">
+            {error}
+          </p>
+        )}
+      </div>
+
+      {result !== null && (
+        <div className="space-y-4 rounded-lg border border-amber-300 bg-amber-50 p-4" data-testid="api-key-result">
+          <div className="rounded border border-amber-400 bg-amber-100 px-3 py-2 text-sm text-amber-900">
+            Copy your API key now — it is shown <strong>only once</strong> and cannot be recovered. If you
+            lose it, generate a new one.
+          </div>
+
+          <div>
+            <div className="flex items-center justify-between">
+              <span className="text-sm font-medium text-gray-700">API key</span>
+              <CopyButton value={result.api_key} />
+            </div>
+            <code className="mt-1 block break-all rounded bg-white px-3 py-2 font-mono text-sm">
+              {result.api_key}
+            </code>
+          </div>
+
+          <div>
+            <div className="flex items-center justify-between">
+              <span className="text-sm font-medium text-gray-700">Deposit address (fund with AP3X)</span>
+              <CopyButton value={result.deposit_address} />
+            </div>
+            <code className="mt-1 block break-all rounded bg-white px-3 py-2 font-mono text-sm">
+              {result.deposit_address}
+            </code>
+          </div>
+
+          <div>
+            <div className="flex items-center justify-between">
+              <span className="text-sm font-medium text-gray-700">Base URL</span>
+              <CopyButton value={baseUrl} />
+            </div>
+            <code className="mt-1 block break-all rounded bg-white px-3 py-2 font-mono text-sm">
+              {baseUrl}
+            </code>
+          </div>
+
+          <div>
+            <div className="flex items-center justify-between">
+              <span className="text-sm font-medium text-gray-700">Quick start</span>
+              <CopyButton value={curlSnippet} />
+            </div>
+            <pre className="mt-1 overflow-x-auto rounded bg-gray-900 px-3 py-2 font-mono text-xs text-gray-100">
+              {curlSnippet}
+            </pre>
+            <p className="mt-1 text-xs text-gray-500">
+              Replace <code className="font-mono">&lt;model&gt;</code> with a model from{" "}
+              <code className="font-mono">{baseUrl}/models</code>.
+            </p>
+          </div>
+        </div>
+      )}
+    </div>
+  );
+}

buyer/src/ui/pages/BookSummarizer.tsx

@@ -152,7 +152,7 @@ export default function BookSummarizer() {
       <div>
         <h1 className="text-2xl font-semibold">Book Summarizer</h1>
         <p className="mt-1 text-sm text-gray-500">
-          Upload a PDF → pay suppliers to summarize each chunk → get a full-book summary.
+          Upload a PDF → commission suppliers to summarize each chunk → get a full-book summary.
           Jobs keep running if you navigate away; reopen them below.
         </p>
       </div>

buyer/src/ui/pages/PendingReceipts.tsx

@@ -1,6 +1,6 @@
 /**
  * buyer/src/ui/pages/PendingReceipts.tsx — list of currently-Submitted escrows
- * for this buyer, with an "Accept & Pay" action per row.
+ * for this buyer, with an "Accept & Settle" action per row.
  *
  * Data flow:
  *   GET /v1/pending-receipts → buyer-app server proxies indexer /escrows?buyer=<pkh>
@@ -190,7 +190,7 @@ export default function PendingReceipts() {
               <th className="px-2 py-2">Submitted</th>
               <th className="px-2 py-2">Supplier</th>
               <th className="px-2 py-2">Capability</th>
-              <th className="px-2 py-2">Payment</th>
+              <th className="px-2 py-2">Escrowed</th>
               <th className="px-2 py-2">Window left</th>
               <th className="px-2 py-2">Action</th>
             </tr>
@@ -254,7 +254,7 @@ export default function PendingReceipts() {
                               : "bg-indigo-600 text-white hover:bg-indigo-700")
                         }
                       >
-                        {isAccepting ? "accepting…" : expired ? "expired" : "Accept & Pay"}
+                        {isAccepting ? "accepting…" : expired ? "expired" : "Accept & Settle"}
                       </button>
                     )}
                   </td>

buyer/src/ui/pages/TaskHistory.tsx

@@ -304,7 +304,7 @@ function LifecycleCard({ lc }: { lc: Lifecycle }): JSX.Element {
           </>
         )}
         <span className="text-gray-500">·</span>
-        <span>{fmtAda(lc.payment_lovelace)} payment</span>
+        <span>{fmtAda(lc.payment_lovelace)} escrowed</span>
         <span className="text-gray-500">·</span>
         <span>{fmtAda(lc.buyer_bond_lovelace)} bond</span>
         <span className="text-gray-500">·</span>

deploy/mainnet/docker-compose.buyer.yml

@@ -29,6 +29,11 @@ services:
       # the response archive + PDF job store (past summaries) across container
       # recreation/redeploys.
       ARCHIVE_DIR: /repo/data/archive
+      # Public base URL of the OpenAI gateway, surfaced to the SPA boot block so
+      # the "API Keys" page can POST cross-origin to ${gatewayUrl}/signup. The
+      # gateway must CORS-allow https://marketplace.vector.apexfusion.org (see
+      # GATEWAY_CORS_ORIGINS in docker-compose.gateway.yml).
+      GATEWAY_PUBLIC_URL: https://api.marketplace.vector.apexfusion.org
       # NETWORK_ID intentionally not pinned — see supplier compose comment.
     volumes:
       - buyer-mainnet-archive:/repo/data/archive

deploy/mainnet/docker-compose.gateway.yml

@@ -27,6 +27,24 @@ services:
       NETWORK_ID: "1"
       GATEWAY_PORT: "8080"
       GATEWAY_DB_DIR: /repo/data/gateway
+      # Mainnet chain genesis (slot↔wallclock). MUST match buyer/supplier or
+      # PostEscrow validity bounds land ~51 days off and Ogmios rejects (code 3118).
+      VECTOR_ZERO_TIME_MS: "1756485600000"
+      # On-chain reference-script UTxOs (escrow spending validator) used by
+      # Accept/Reclaim/Submit. Same values the buyer/suppliers use on mainnet.
+      ESCROW_REF_UTXO: c8d84c6d67ec67a1efe5e9c6c06d53020e05d1bb96d1c55ecb1eb7d5010c4d54#0
+      ADVERT_REF_UTXO: c8d84c6d67ec67a1efe5e9c6c06d53020e05d1bb96d1c55ecb1eb7d5010c4d54#1
+      # Browser Origins allowed to call the gateway cross-origin. The buyer
+      # frontend's "API Keys" page (served from marketplace.vector…) POSTs to
+      # /signup here. Comma-separated; empty disables CORS entirely.
+      GATEWAY_CORS_ORIGINS: https://marketplace.vector.apexfusion.org
+      # Ticket mode: the escrow post at session open is the only chain op;
+      # the sweeper reclaims Open escrows after deliver_by. Must match the
+      # kimi-chat suppliers' CHAT_SETTLE_MODE.
+      CHAT_SETTLE_MODE: ticket
+      # Demo janitor closes idle sessions BEFORE the suppliers' 15 min
+      # CHAT_IDLE_TIMEOUT_MS backstop.
+      DEMO_SESSION_IDLE_MS: "600000"
     volumes:
       - gateway-mainnet-data:/repo/data/gateway
     healthcheck:
@@ -41,8 +59,18 @@ services:
     labels:
       - "traefik.enable=true"
       - "traefik.docker.network=traefik-net"
+      # This host is NOT Cloudflare-proxied: traefik terminates TLS itself with
+      # a Let's Encrypt cert (resolver defined in /root/traefik/traefik.yml on
+      # the box). The :80 router only redirects; ACME HTTP-01 challenges are
+      # answered by traefik before the redirect applies.
       - "traefik.http.routers.mp-gateway-mainnet.entrypoints=web"
       - "traefik.http.routers.mp-gateway-mainnet.rule=Host(`api.marketplace.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-gateway-mainnet.middlewares=mp-gateway-mainnet-https"
+      - "traefik.http.middlewares.mp-gateway-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-gateway-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-gateway-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-gateway-mainnet-tls.rule=Host(`api.marketplace.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-gateway-mainnet-tls.tls.certresolver=letsencrypt"
       - "traefik.http.services.mp-gateway-mainnet.loadbalancer.server.port=8080"
     logging:
       driver: json-file

deploy/mainnet/docker-compose.supplier-deepseek-flash-1.yml

@@ -0,0 +1,71 @@
+# deploy/mainnet/docker-compose.supplier-deepseek-flash-1.yml — Supplier — MAINNET (OpenRouter / DeepSeek V4 Flash #1)
+#
+# Parallel-capacity supplier for deepseek/deepseek-v4-flash via OpenRouter's
+# OpenAI-compatible API. Same shape as the kimi replicas; distinct only in
+# instance name, container, env file, Traefik host/router, and advert model.
+# Shares the same OpenRouter API key as the kimi suppliers (accepted ceiling;
+# supplier has no 429/backoff). Advert caps: max_output_tokens=262144,
+# max_processing_ms=3600000; OPENAI_TIMEOUT_MS matched.
+#
+# Pre-flight:
+#   1. DNS (proxied): mp-suppliers-deepseek-flash-1.vector.apexfusion.org → host
+#   2. supplier/.env.deepseek-flash-1 on host (chmod 600): SUPPLIER_PRIV_KEY_HEX,
+#        OPENAI_API_KEY, OGMIOS_URL, NETWORK_ID=1, VECTOR_ZERO_TIME_MS,
+#        ESCROW_REF_UTXO, ADVERT_REF_UTXO, ADVERT_REF, LIVE_CHAIN=1
+#   3. Networks marketplace-mainnet-net + traefik-net already exist.
+
+name: marketplace-mainnet-supplier-deepseek-flash-1
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-deepseek-flash-1
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.deepseek-flash-1
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      # Matches the advert's max_processing_ms=3600000 (60 min).
+      OPENAI_TIMEOUT_MS: "3600000"
+      PORT: "8080"
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL come from supplier/.env.deepseek-flash-1.
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-1-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-1-mainnet.rule=Host(`mp-suppliers-deepseek-flash-1.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-1-mainnet.middlewares=mp-suppliers-deepseek-flash-1-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-deepseek-flash-1-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-deepseek-flash-1-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-1-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-1-mainnet-tls.rule=Host(`mp-suppliers-deepseek-flash-1.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-1-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-deepseek-flash-1-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-deepseek-flash-2.yml

@@ -0,0 +1,71 @@
+# deploy/mainnet/docker-compose.supplier-deepseek-flash-2.yml — Supplier — MAINNET (OpenRouter / DeepSeek V4 Flash #2)
+#
+# Parallel-capacity supplier for deepseek/deepseek-v4-flash via OpenRouter's
+# OpenAI-compatible API. Same shape as the kimi replicas; distinct only in
+# instance name, container, env file, Traefik host/router, and advert model.
+# Shares the same OpenRouter API key as the kimi suppliers (accepted ceiling;
+# supplier has no 429/backoff). Advert caps: max_output_tokens=262144,
+# max_processing_ms=3600000; OPENAI_TIMEOUT_MS matched.
+#
+# Pre-flight:
+#   1. DNS (proxied): mp-suppliers-deepseek-flash-2.vector.apexfusion.org → host
+#   2. supplier/.env.deepseek-flash-2 on host (chmod 600): SUPPLIER_PRIV_KEY_HEX,
+#        OPENAI_API_KEY, OGMIOS_URL, NETWORK_ID=1, VECTOR_ZERO_TIME_MS,
+#        ESCROW_REF_UTXO, ADVERT_REF_UTXO, ADVERT_REF, LIVE_CHAIN=1
+#   3. Networks marketplace-mainnet-net + traefik-net already exist.
+
+name: marketplace-mainnet-supplier-deepseek-flash-2
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-deepseek-flash-2
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.deepseek-flash-2
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      # Matches the advert's max_processing_ms=3600000 (60 min).
+      OPENAI_TIMEOUT_MS: "3600000"
+      PORT: "8080"
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL come from supplier/.env.deepseek-flash-2.
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-2-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-2-mainnet.rule=Host(`mp-suppliers-deepseek-flash-2.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-2-mainnet.middlewares=mp-suppliers-deepseek-flash-2-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-deepseek-flash-2-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-deepseek-flash-2-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-2-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-2-mainnet-tls.rule=Host(`mp-suppliers-deepseek-flash-2.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-2-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-deepseek-flash-2-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-deepseek-flash-3.yml

@@ -0,0 +1,71 @@
+# deploy/mainnet/docker-compose.supplier-deepseek-flash-3.yml — Supplier — MAINNET (OpenRouter / DeepSeek V4 Flash #3)
+#
+# Parallel-capacity supplier for deepseek/deepseek-v4-flash via OpenRouter's
+# OpenAI-compatible API. Same shape as the kimi replicas; distinct only in
+# instance name, container, env file, Traefik host/router, and advert model.
+# Shares the same OpenRouter API key as the kimi suppliers (accepted ceiling;
+# supplier has no 429/backoff). Advert caps: max_output_tokens=262144,
+# max_processing_ms=3600000; OPENAI_TIMEOUT_MS matched.
+#
+# Pre-flight:
+#   1. DNS (proxied): mp-suppliers-deepseek-flash-3.vector.apexfusion.org → host
+#   2. supplier/.env.deepseek-flash-3 on host (chmod 600): SUPPLIER_PRIV_KEY_HEX,
+#        OPENAI_API_KEY, OGMIOS_URL, NETWORK_ID=1, VECTOR_ZERO_TIME_MS,
+#        ESCROW_REF_UTXO, ADVERT_REF_UTXO, ADVERT_REF, LIVE_CHAIN=1
+#   3. Networks marketplace-mainnet-net + traefik-net already exist.
+
+name: marketplace-mainnet-supplier-deepseek-flash-3
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-deepseek-flash-3
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.deepseek-flash-3
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      # Matches the advert's max_processing_ms=3600000 (60 min).
+      OPENAI_TIMEOUT_MS: "3600000"
+      PORT: "8080"
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL come from supplier/.env.deepseek-flash-3.
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-3-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-3-mainnet.rule=Host(`mp-suppliers-deepseek-flash-3.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-3-mainnet.middlewares=mp-suppliers-deepseek-flash-3-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-deepseek-flash-3-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-deepseek-flash-3-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-3-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-3-mainnet-tls.rule=Host(`mp-suppliers-deepseek-flash-3.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-deepseek-flash-3-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-deepseek-flash-3-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-deepseek.yml

@@ -65,6 +65,12 @@ services:
       - "traefik.docker.network=traefik-net"
       - "traefik.http.routers.mp-suppliers-deepseek-mainnet.entrypoints=web"
       - "traefik.http.routers.mp-suppliers-deepseek-mainnet.rule=Host(`mp-suppliers-deepseek.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-deepseek-mainnet.middlewares=mp-suppliers-deepseek-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-deepseek-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-deepseek-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-deepseek-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-deepseek-mainnet-tls.rule=Host(`mp-suppliers-deepseek.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-deepseek-mainnet-tls.tls.certresolver=letsencrypt"
       - "traefik.http.services.mp-suppliers-deepseek-mainnet.loadbalancer.server.port=8080"
     logging:
       driver: json-file

deploy/mainnet/docker-compose.supplier-ds-flash-chat-1.yml

@@ -0,0 +1,80 @@
+# deploy/mainnet/docker-compose.supplier-ds-flash-chat-1.yml — Supplier — MAINNET
+#   (OpenRouter / deepseek-v4-flash — MULTI-TURN CHAT capability `llm.chat.v1`, REPLICA 1)
+#
+# Chat-session sibling of the deepseek-flash one-off suppliers. Same OpenRouter
+# backend (deepseek/deepseek-v4-flash) but CAPABILITY_KIND=chat-session so it
+# mounts /v1/chat/{start,message,end}. Ticket settle mode (matches the mainnet
+# gateway): the buyer's Open escrow post is the only chain op; the supplier
+# never Claims/Submits; the gateway sweeper reclaims after deliver_by.
+#
+# Pre-flight:
+#   1. DNS A-record: mp-suppliers-ds-flash-chat-1.vector.apexfusion.org -> host IP
+#   2. New supplier wallet (gen-keypair), funded (advert post + fee float).
+#   3. Post a fresh llm.chat.v1 advert for this wallet:
+#        --capability-id llm.chat.v1 --model deepseek/deepseek-v4-flash
+#        --max-processing-ms 1800000 --price-lovelace 200000
+#   4. supplier/.env.ds-flash-chat-1 written on host (chmod 600) with at least:
+#        SUPPLIER_PRIV_KEY_HEX, OPENAI_API_KEY, OGMIOS_URL, NETWORK_ID=1,
+#        VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO, ADVERT_REF,
+#        LIVE_CHAIN=1, OPENAI_REASONING=off, OPENAI_MAX_TOKENS=131072
+
+name: marketplace-mainnet-supplier-ds-flash-chat-1
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-ds-flash-chat-1
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.ds-flash-chat-1
+        required: false
+    environment:
+      # Multi-turn chat session capability (mounts /v1/chat/{start,message,end}).
+      CAPABILITY_KIND: chat-session
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      OPENAI_TIMEOUT_MS: "300000"
+      CHAT_IDLE_TIMEOUT_MS: "900000"
+      CHAT_SETTLE_MODE: ticket
+      MAX_CHAT_SESSIONS: "8"
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so wallet-health/consolidate can broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL come from supplier/.env.ds-flash-chat-1.
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-1-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-1-mainnet.rule=Host(`mp-suppliers-ds-flash-chat-1.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-1-mainnet.middlewares=mp-suppliers-ds-flash-chat-1-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-chat-1-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-chat-1-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-1-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-1-mainnet-tls.rule=Host(`mp-suppliers-ds-flash-chat-1.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-1-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-ds-flash-chat-1-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-ds-flash-chat-2.yml

@@ -0,0 +1,80 @@
+# deploy/mainnet/docker-compose.supplier-ds-flash-chat-2.yml — Supplier — MAINNET
+#   (OpenRouter / deepseek-v4-flash — MULTI-TURN CHAT capability `llm.chat.v1`, REPLICA 2)
+#
+# Chat-session sibling of the deepseek-flash one-off suppliers. Same OpenRouter
+# backend (deepseek/deepseek-v4-flash) but CAPABILITY_KIND=chat-session so it
+# mounts /v1/chat/{start,message,end}. Ticket settle mode (matches the mainnet
+# gateway): the buyer's Open escrow post is the only chain op; the supplier
+# never Claims/Submits; the gateway sweeper reclaims after deliver_by.
+#
+# Pre-flight:
+#   1. DNS A-record: mp-suppliers-ds-flash-chat-2.vector.apexfusion.org -> host IP
+#   2. New supplier wallet (gen-keypair), funded (advert post + fee float).
+#   3. Post a fresh llm.chat.v1 advert for this wallet:
+#        --capability-id llm.chat.v1 --model deepseek/deepseek-v4-flash
+#        --max-processing-ms 1800000 --price-lovelace 200000
+#   4. supplier/.env.ds-flash-chat-2 written on host (chmod 600) with at least:
+#        SUPPLIER_PRIV_KEY_HEX, OPENAI_API_KEY, OGMIOS_URL, NETWORK_ID=1,
+#        VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO, ADVERT_REF,
+#        LIVE_CHAIN=1, OPENAI_REASONING=off, OPENAI_MAX_TOKENS=131072
+
+name: marketplace-mainnet-supplier-ds-flash-chat-2
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-ds-flash-chat-2
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.ds-flash-chat-2
+        required: false
+    environment:
+      # Multi-turn chat session capability (mounts /v1/chat/{start,message,end}).
+      CAPABILITY_KIND: chat-session
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      OPENAI_TIMEOUT_MS: "300000"
+      CHAT_IDLE_TIMEOUT_MS: "900000"
+      CHAT_SETTLE_MODE: ticket
+      MAX_CHAT_SESSIONS: "8"
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so wallet-health/consolidate can broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL come from supplier/.env.ds-flash-chat-2.
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-2-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-2-mainnet.rule=Host(`mp-suppliers-ds-flash-chat-2.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-2-mainnet.middlewares=mp-suppliers-ds-flash-chat-2-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-chat-2-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-chat-2-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-2-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-2-mainnet-tls.rule=Host(`mp-suppliers-ds-flash-chat-2.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-2-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-ds-flash-chat-2-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-ds-flash-chat-3.yml

@@ -0,0 +1,80 @@
+# deploy/mainnet/docker-compose.supplier-ds-flash-chat-3.yml — Supplier — MAINNET
+#   (OpenRouter / deepseek-v4-flash — MULTI-TURN CHAT capability `llm.chat.v1`, REPLICA 3)
+#
+# Chat-session sibling of the deepseek-flash one-off suppliers. Same OpenRouter
+# backend (deepseek/deepseek-v4-flash) but CAPABILITY_KIND=chat-session so it
+# mounts /v1/chat/{start,message,end}. Ticket settle mode (matches the mainnet
+# gateway): the buyer's Open escrow post is the only chain op; the supplier
+# never Claims/Submits; the gateway sweeper reclaims after deliver_by.
+#
+# Pre-flight:
+#   1. DNS A-record: mp-suppliers-ds-flash-chat-3.vector.apexfusion.org -> host IP
+#   2. New supplier wallet (gen-keypair), funded (advert post + fee float).
+#   3. Post a fresh llm.chat.v1 advert for this wallet:
+#        --capability-id llm.chat.v1 --model deepseek/deepseek-v4-flash
+#        --max-processing-ms 1800000 --price-lovelace 200000
+#   4. supplier/.env.ds-flash-chat-3 written on host (chmod 600) with at least:
+#        SUPPLIER_PRIV_KEY_HEX, OPENAI_API_KEY, OGMIOS_URL, NETWORK_ID=1,
+#        VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO, ADVERT_REF,
+#        LIVE_CHAIN=1, OPENAI_REASONING=off, OPENAI_MAX_TOKENS=131072
+
+name: marketplace-mainnet-supplier-ds-flash-chat-3
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-ds-flash-chat-3
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.ds-flash-chat-3
+        required: false
+    environment:
+      # Multi-turn chat session capability (mounts /v1/chat/{start,message,end}).
+      CAPABILITY_KIND: chat-session
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      OPENAI_TIMEOUT_MS: "300000"
+      CHAT_IDLE_TIMEOUT_MS: "900000"
+      CHAT_SETTLE_MODE: ticket
+      MAX_CHAT_SESSIONS: "8"
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so wallet-health/consolidate can broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL come from supplier/.env.ds-flash-chat-3.
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-3-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-3-mainnet.rule=Host(`mp-suppliers-ds-flash-chat-3.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-3-mainnet.middlewares=mp-suppliers-ds-flash-chat-3-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-chat-3-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-chat-3-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-3-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-3-mainnet-tls.rule=Host(`mp-suppliers-ds-flash-chat-3.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-chat-3-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-ds-flash-chat-3-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-gpt.yml

@@ -61,6 +61,12 @@ services:
       - "traefik.docker.network=traefik-net"
       - "traefik.http.routers.mp-suppliers-gpt-mainnet.entrypoints=web"
       - "traefik.http.routers.mp-suppliers-gpt-mainnet.rule=Host(`mp-suppliers-gpt.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-gpt-mainnet.middlewares=mp-suppliers-gpt-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-gpt-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-gpt-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-gpt-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-gpt-mainnet-tls.rule=Host(`mp-suppliers-gpt.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-gpt-mainnet-tls.tls.certresolver=letsencrypt"
       - "traefik.http.services.mp-suppliers-gpt-mainnet.loadbalancer.server.port=8080"
     logging:
       driver: json-file

deploy/mainnet/docker-compose.supplier-kimi-2.yml

@@ -0,0 +1,73 @@
+# deploy/mainnet/docker-compose.supplier-kimi-2.yml — Supplier — MAINNET (OpenRouter / Kimi K2.6 replica #2)
+#
+# Parallel-capacity replica of supplier-kimi (OpenRouter, moonshotai/kimi-k2.6).
+# Identical to docker-compose.supplier-kimi.yml except for the instance name,
+# container name, env file, Traefik host/router, and OPENAI_TIMEOUT_MS (raised
+# to 3600000 to match this advert's max_processing_ms=3600000, max_output_tokens=262144).
+# Shares the same OpenRouter API key as the other kimi suppliers (accepted
+# concurrency ceiling; supplier has no 429/backoff).
+#
+# Pre-flight:
+#   1. DNS (proxied): mp-suppliers-kimi-2.vector.apexfusion.org → host
+#   2. supplier/.env.kimi-2 on host (chmod 600): SUPPLIER_PRIV_KEY_HEX, OPENAI_API_KEY,
+#        OGMIOS_URL, NETWORK_ID=1, VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO,
+#        ADVERT_REF, LIVE_CHAIN=1
+#   3. Networks marketplace-mainnet-net + traefik-net already exist.
+
+name: marketplace-mainnet-supplier-kimi-2
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-kimi-2
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.kimi-2
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      # Raised to match the advert's max_processing_ms=3600000 (60 min) so the
+      # upstream call isn't aborted before the advertised processing deadline.
+      OPENAI_TIMEOUT_MS: "3600000"
+      PORT: "8080"
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from supplier/.env.kimi-2 on the mainnet host (footgun avoidance).
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-kimi-2-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-kimi-2-mainnet.rule=Host(`mp-suppliers-kimi-2.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-2-mainnet.middlewares=mp-suppliers-kimi-2-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-2-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-2-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-kimi-2-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-kimi-2-mainnet-tls.rule=Host(`mp-suppliers-kimi-2.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-2-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-kimi-2-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-kimi-3.yml

@@ -0,0 +1,73 @@
+# deploy/mainnet/docker-compose.supplier-kimi-3.yml — Supplier — MAINNET (OpenRouter / Kimi K2.6 replica #3)
+#
+# Parallel-capacity replica of supplier-kimi (OpenRouter, moonshotai/kimi-k2.6).
+# Identical to docker-compose.supplier-kimi.yml except for the instance name,
+# container name, env file, Traefik host/router, and OPENAI_TIMEOUT_MS (raised
+# to 3600000 to match this advert's max_processing_ms=3600000, max_output_tokens=262144).
+# Shares the same OpenRouter API key as the other kimi suppliers (accepted
+# concurrency ceiling; supplier has no 429/backoff).
+#
+# Pre-flight:
+#   1. DNS (proxied): mp-suppliers-kimi-3.vector.apexfusion.org → host
+#   2. supplier/.env.kimi-3 on host (chmod 600): SUPPLIER_PRIV_KEY_HEX, OPENAI_API_KEY,
+#        OGMIOS_URL, NETWORK_ID=1, VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO,
+#        ADVERT_REF, LIVE_CHAIN=1
+#   3. Networks marketplace-mainnet-net + traefik-net already exist.
+
+name: marketplace-mainnet-supplier-kimi-3
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-kimi-3
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.kimi-3
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      # Raised to match the advert's max_processing_ms=3600000 (60 min) so the
+      # upstream call isn't aborted before the advertised processing deadline.
+      OPENAI_TIMEOUT_MS: "3600000"
+      PORT: "8080"
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from supplier/.env.kimi-3 on the mainnet host (footgun avoidance).
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-kimi-3-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-kimi-3-mainnet.rule=Host(`mp-suppliers-kimi-3.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-3-mainnet.middlewares=mp-suppliers-kimi-3-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-3-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-3-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-kimi-3-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-kimi-3-mainnet-tls.rule=Host(`mp-suppliers-kimi-3.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-3-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-kimi-3-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-kimi-4.yml

@@ -0,0 +1,73 @@
+# deploy/mainnet/docker-compose.supplier-kimi-4.yml — Supplier — MAINNET (OpenRouter / Kimi K2.6 replica #4)
+#
+# Parallel-capacity replica of supplier-kimi (OpenRouter, moonshotai/kimi-k2.6).
+# Identical to docker-compose.supplier-kimi.yml except for the instance name,
+# container name, env file, Traefik host/router, and OPENAI_TIMEOUT_MS (raised
+# to 3600000 to match this advert's max_processing_ms=3600000, max_output_tokens=262144).
+# Shares the same OpenRouter API key as the other kimi suppliers (accepted
+# concurrency ceiling; supplier has no 429/backoff).
+#
+# Pre-flight:
+#   1. DNS (proxied): mp-suppliers-kimi-4.vector.apexfusion.org → host
+#   2. supplier/.env.kimi-4 on host (chmod 600): SUPPLIER_PRIV_KEY_HEX, OPENAI_API_KEY,
+#        OGMIOS_URL, NETWORK_ID=1, VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO,
+#        ADVERT_REF, LIVE_CHAIN=1
+#   3. Networks marketplace-mainnet-net + traefik-net already exist.
+
+name: marketplace-mainnet-supplier-kimi-4
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-kimi-4
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.kimi-4
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      # Raised to match the advert's max_processing_ms=3600000 (60 min) so the
+      # upstream call isn't aborted before the advertised processing deadline.
+      OPENAI_TIMEOUT_MS: "3600000"
+      PORT: "8080"
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from supplier/.env.kimi-4 on the mainnet host (footgun avoidance).
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-kimi-4-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-kimi-4-mainnet.rule=Host(`mp-suppliers-kimi-4.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-4-mainnet.middlewares=mp-suppliers-kimi-4-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-4-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-4-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-kimi-4-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-kimi-4-mainnet-tls.rule=Host(`mp-suppliers-kimi-4.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-4-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-kimi-4-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-kimi-chat-2.yml

@@ -0,0 +1,84 @@
+# deploy/mainnet/docker-compose.supplier-kimi-chat-2.yml — Supplier — MAINNET
+#   (OpenRouter / Kimi K2.6 — MULTI-TURN CHAT capability `llm.chat.v1`, REPLICA 2)
+#
+# Second llm.chat.v1 replica: a chat session holds a supplier's single slot for
+# the whole conversation, so concurrent chat users = number of chat suppliers.
+# Deployed for the shared-demo-key endpoint (session-backed /chat/completions).
+# Identical to docker-compose.supplier-kimi-chat.yml except for its own wallet,
+# advert, env file and hostname.
+#
+# Pre-flight:
+#   1. DNS A-record: mp-suppliers-kimi-chat-2.vector.apexfusion.org → host IP
+#   2. New supplier wallet (gen-keypair), funded with advert bond + fee float.
+#   3. Post a fresh llm.chat.v1 advert for this wallet:
+#        --capability-id llm.chat.v1 --model moonshotai/kimi-k2.6
+#        --max-processing-ms 1800000  (same price as the primary chat advert)
+#   4. supplier/.env.kimi-chat-2 written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, OPENAI_API_KEY, OGMIOS_URL, NETWORK_ID=1,
+#        VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO, ADVERT_REF,
+#        LIVE_CHAIN=1
+
+name: marketplace-mainnet-supplier-kimi-chat-2
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-kimi-chat-2
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.kimi-chat-2
+        required: false
+    environment:
+      # Multi-turn chat session capability (mounts /v1/chat/{start,message,end}).
+      CAPABILITY_KIND: chat-session
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      # Kimi K2.6 generations can be slow; mirror the one-off Kimi compose.
+      OPENAI_TIMEOUT_MS: "300000"
+      # 2.5 min idle end: single-slot suppliers must free up quickly for the
+      # shared demo endpoint (the gateway transparently reopens on next turn).
+      CHAT_IDLE_TIMEOUT_MS: "900000"
+      CHAT_SETTLE_MODE: ticket
+      MAX_CHAT_SESSIONS: "8"
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so Claim/Submit actually broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from the operator-provided supplier/.env.kimi-chat-2 on a mainnet host.
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-2-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-2-mainnet.rule=Host(`mp-suppliers-kimi-chat-2.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-2-mainnet.middlewares=mp-suppliers-kimi-chat-2-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-chat-2-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-chat-2-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-2-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-2-mainnet-tls.rule=Host(`mp-suppliers-kimi-chat-2.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-2-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-kimi-chat-2-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-kimi-chat-3.yml

@@ -0,0 +1,84 @@
+# deploy/mainnet/docker-compose.supplier-kimi-chat-3.yml — Supplier — MAINNET
+#   (OpenRouter / Kimi K2.6 — MULTI-TURN CHAT capability `llm.chat.v1`, REPLICA 3)
+#
+# Third llm.chat.v1 replica: a chat session holds a supplier's single slot for
+# the whole conversation, so concurrent chat users = number of chat suppliers.
+# Deployed for the shared-demo-key endpoint (session-backed /chat/completions).
+# Identical to docker-compose.supplier-kimi-chat.yml except for its own wallet,
+# advert, env file and hostname.
+#
+# Pre-flight:
+#   1. DNS A-record: mp-suppliers-kimi-chat-3.vector.apexfusion.org → host IP
+#   2. New supplier wallet (gen-keypair), funded with advert bond + fee float.
+#   3. Post a fresh llm.chat.v1 advert for this wallet:
+#        --capability-id llm.chat.v1 --model moonshotai/kimi-k2.6
+#        --max-processing-ms 1800000  (same price as the primary chat advert)
+#   4. supplier/.env.kimi-chat-3 written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, OPENAI_API_KEY, OGMIOS_URL, NETWORK_ID=1,
+#        VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO, ADVERT_REF,
+#        LIVE_CHAIN=1
+
+name: marketplace-mainnet-supplier-kimi-chat-3
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-kimi-chat-3
+    restart: unless-stopped
+    env_file:
+      - path: ../../supplier/.env.kimi-chat-3
+        required: false
+    environment:
+      # Multi-turn chat session capability (mounts /v1/chat/{start,message,end}).
+      CAPABILITY_KIND: chat-session
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      OPENAI_BASE_URL: https://openrouter.ai/api
+      # Kimi K2.6 generations can be slow; mirror the one-off Kimi compose.
+      OPENAI_TIMEOUT_MS: "300000"
+      # 2.5 min idle end: single-slot suppliers must free up quickly for the
+      # shared demo endpoint (the gateway transparently reopens on next turn).
+      CHAT_IDLE_TIMEOUT_MS: "900000"
+      CHAT_SETTLE_MODE: ticket
+      MAX_CHAT_SESSIONS: "8"
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so Claim/Submit actually broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from the operator-provided supplier/.env.kimi-chat-3 on a mainnet host.
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - marketplace-mainnet-net
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-3-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-3-mainnet.rule=Host(`mp-suppliers-kimi-chat-3.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-3-mainnet.middlewares=mp-suppliers-kimi-chat-3-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-chat-3-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-chat-3-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-3-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-3-mainnet-tls.rule=Host(`mp-suppliers-kimi-chat-3.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-3-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-kimi-chat-3-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  marketplace-mainnet-net:
+    external: true
+    name: marketplace-mainnet-net
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/docker-compose.supplier-kimi-chat.yml

@@ -49,8 +49,12 @@ services:
       OPENAI_BASE_URL: https://openrouter.ai/api
       # Kimi K2.6 generations can be slow; mirror the one-off Kimi compose.
       OPENAI_TIMEOUT_MS: "300000"
-      # Auto-end an idle session after 5 min (supplier Submits the transcript).
-      CHAT_IDLE_TIMEOUT_MS: "300000"
+      # 2.5 min idle end (supplier Submits the transcript): single-slot
+      # suppliers must free up quickly for the shared demo endpoint (the
+      # gateway transparently reopens a session on the next turn).
+      CHAT_IDLE_TIMEOUT_MS: "900000"
+      CHAT_SETTLE_MODE: ticket
+      MAX_CHAT_SESSIONS: "8"
       PORT: "8080"
       # Enable LiveOgmiosProvider so Claim/Submit actually broadcast.
       LIVE_CHAIN: "1"
@@ -70,6 +74,12 @@ services:
       - "traefik.docker.network=traefik-net"
       - "traefik.http.routers.mp-suppliers-kimi-chat-mainnet.entrypoints=web"
       - "traefik.http.routers.mp-suppliers-kimi-chat-mainnet.rule=Host(`mp-suppliers-kimi-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-mainnet.middlewares=mp-suppliers-kimi-chat-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-chat-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-chat-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-mainnet-tls.rule=Host(`mp-suppliers-kimi-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-chat-mainnet-tls.tls.certresolver=letsencrypt"
       - "traefik.http.services.mp-suppliers-kimi-chat-mainnet.loadbalancer.server.port=8080"
     logging:
       driver: json-file

deploy/mainnet/docker-compose.supplier-kimi.yml

@@ -48,8 +48,10 @@ services:
       LLM_BACKEND: openai
       # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
       OPENAI_BASE_URL: https://openrouter.ai/api
-      # Kimi K2.6 generations can be slow; mirror the DeepSeek/GPT composes.
-      OPENAI_TIMEOUT_MS: "300000"
+      # Raised to match the advert's max_processing_ms=3600000 (60 min) so the
+      # upstream call isn't aborted before the advertised processing deadline
+      # (matches the kimi-2/3/4 replicas).
+      OPENAI_TIMEOUT_MS: "3600000"
       PORT: "8080"
       # Enable LiveOgmiosProvider so submitTx/awaitTx actually broadcast.
       # Without this the supplier silently falls back to ReadOnlyOgmiosProvider
@@ -73,6 +75,12 @@ services:
       - "traefik.docker.network=traefik-net"
       - "traefik.http.routers.mp-suppliers-kimi-mainnet.entrypoints=web"
       - "traefik.http.routers.mp-suppliers-kimi-mainnet.rule=Host(`mp-suppliers-kimi.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-mainnet.middlewares=mp-suppliers-kimi-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-kimi-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-kimi-mainnet-tls.rule=Host(`mp-suppliers-kimi.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-mainnet-tls.tls.certresolver=letsencrypt"
       - "traefik.http.services.mp-suppliers-kimi-mainnet.loadbalancer.server.port=8080"
     logging:
       driver: json-file

deploy/mainnet/docker-compose.supplier-tts.yml

@@ -65,6 +65,12 @@ services:
       - "traefik.docker.network=traefik-net"
       - "traefik.http.routers.mp-suppliers-tts-mainnet.entrypoints=web"
       - "traefik.http.routers.mp-suppliers-tts-mainnet.rule=Host(`mp-suppliers-tts.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-tts-mainnet.middlewares=mp-suppliers-tts-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-tts-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-tts-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-tts-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-tts-mainnet-tls.rule=Host(`mp-suppliers-tts.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-tts-mainnet-tls.tls.certresolver=letsencrypt"
       - "traefik.http.services.mp-suppliers-tts-mainnet.loadbalancer.server.port=8080"
     logging:
       driver: json-file

deploy/mainnet/docker-compose.supplier.yml

@@ -60,6 +60,12 @@ services:
       - "traefik.docker.network=traefik-net"
       - "traefik.http.routers.mp-suppliers-mainnet.entrypoints=web"
       - "traefik.http.routers.mp-suppliers-mainnet.rule=Host(`mp-suppliers.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-mainnet.middlewares=mp-suppliers-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-mainnet-tls.rule=Host(`mp-suppliers.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-mainnet-tls.tls.certresolver=letsencrypt"
       - "traefik.http.services.mp-suppliers-mainnet.loadbalancer.server.port=8080"
     logging:
       driver: json-file

docs/ARCHITECTURE.md

@@ -8,13 +8,13 @@
 
 ## 1. Goal and scope
 
-**What this is.** A two-sided marketplace where buyers (agents or humans) submit prompts to suppliers running inference on their own hardware, pay with AP3X via on-chain bonded escrow on Vector L2, and receive signed-receipt responses. The primary validation goal is **technical feasibility + lifecycle** ("can it be done end-to-end"), not demand modelling or economic stress testing.
+**What this is.** A two-sided marketplace where buyers (agents or humans) submit prompts to suppliers running inference on their own hardware, with each job settled in AP3X through on-chain bonded escrow on Vector L2, and receive signed-receipt responses. The primary validation goal is **technical feasibility + lifecycle** ("can it be done end-to-end"), not demand modelling or economic stress testing.
 
 **What v1 is NOT.**
 - Not agentic on the supplier side. Suppliers answer single prompts. Buyer agents drive multiple prompts.
 - Not LLM-only in design. Capability model is open from day 1 (whisper / kokoro / image generation are planned siblings), but the first supplier is an LLM.
 - Not multi-slot. Single-slot per supplier process. Multi-capability operators run multiple processes.
-- Not multi-chain. Vector testnet first, mainnet later. Masumi / Sokosumi / USDM is target #2 (adapter, not fork).
+- Not multi-chain. Vector only (testnet and mainnet).
 - Not disputed. M1 ships happy-path escrow only. Module-1 disputes wire in M2.
 - Not streaming. `stream: true` is rejected by suppliers.
 - Not tool-calling. `tools` / `tool_choice` / `functions` are rejected by suppliers.
@@ -24,7 +24,7 @@
 
 | # | Decision |
 |---|---|
-| 1 | Payment token: **AP3X** (Vector testnet, later mainnet). Masumi = USDM, target #2. |
+| 1 | Settlement asset: **AP3X**, bonded in escrow; jobs settle in AP3X on accepted work. |
 | 2 | v1 scope: **Option C** — happy-path escrow only, no disputes. Module-1 wiring is milestone 2. |
 | 3 | Module-1 contract source: TBD — to be located or restored before M2. Not blocking M0–M1. |
 | 4 | Evidence: **signed receipt** `{prompt_hash, response_hash, model, tokens, wallclock}` + Ed25519 supplier signature. Hash on-chain, plaintext off-chain. |
@@ -283,7 +283,6 @@ Malformed datum · escrow to wrong script · supplier claims escrow addressed to
 | **M1** | Happy-path end-to-end, Hetzner CPU supplier | 2–3 wks | `qwen2.5:0.5b` on Hetzner, buyer posts → supplier claims → submits → buyer accepts; full Vector testnet lifecycle; indexer serves discovery |
 | **M2** | Module-1 dispute wiring | 2–3 wks post-M1 | Contract source located; supplier=claimer / buyer=auditor mapping; Dispute state; evidence URI; mainnet-dry-run lifecycle |
 | **M3** | Capability expansion + local lab | 2–4 wks (may overlap M2) | `speech.transcribe.v1` via Whisper live; local-lab GPU supplier with larger model; heterogeneous discovery validated |
-| **M4** | Masumi / Sokosumi adapter | later | MIP-003 HTTP shim; MIP-002 NFT registration; USDM payment path; Sokosumi listing |
 
 ## 9. Open follow-ups (not blocking M0 / M1)
 
@@ -310,5 +309,3 @@ Malformed datum · escrow to wrong script · supplier claims escrow addressed to
 - Apex-dashboard CBOR decoders: `apex-dashboard/server/cbor-decoder.ts`
 - Apex-dashboard test discipline: `apex-dashboard/tests/`
 - Prior Python reference (AP3X buyer/supplier): `./docs/buyer/`, `./docs/supplier/`
-- Masumi docs: https://docs.masumi.network/
-- MIP-003 spec: https://docs.masumi.network/mips/_mip-003

docs/EXECUTIVE_SUMMARY.md

@@ -1,412 +0,0 @@
----
-marp: true
-paginate: true
-size: 16:9
----
-
-<!--
-  EXECUTIVE_SUMMARY.md — slide-style overview of the Local Agents Marketplace.
-  Build PDF + PPTX with: bash docs/build-deck.sh
-  Edit this file; mermaid blocks are pre-rendered to docs/img/diagram-*.png.
--->
-
-<style>
-  /* ── Local Agents Marketplace — unified dark theme ────────────────────
-   * Every slide uses the same navy→indigo gradient as the title slide so
-   * the deck reads as one continuous document. Title and closing slides
-   * keep the centered hero treatment; content slides stay top-aligned but
-   * share the same palette and accent rules. */
-  :root {
-    --ink: #f1f5f9;          /* primary text on dark gradient */
-    --muted: #94a3b8;
-    --accent: #818cf8;       /* indigo-400, readable on the dark gradient */
-    --accent-deep: #4f46e5;  /* used for solid accents (h2 bar, table head) */
-    --accent-soft: rgba(129,140,248,0.12);
-    --gradient: linear-gradient(135deg, #0f172a 0%, #1e293b 50%, #312e81 100%);
-    --row-even: rgba(255,255,255,0.04);
-    --row-odd:  rgba(255,255,255,0.02);
-    --border:   rgba(255,255,255,0.10);
-  }
-  section {
-    background: var(--gradient);
-    color: var(--ink);
-    font-family: 'Inter', 'Segoe UI', system-ui, -apple-system, sans-serif;
-    font-size: 22px;
-    padding: 40px 60px;
-    display: flex;
-    flex-direction: column;
-    justify-content: flex-start;
-  }
-  section::after {
-    content: attr(data-marpit-pagination) ' / ' attr(data-marpit-pagination-total);
-    color: var(--muted);
-    font-size: 13px;
-  }
-  h1 {
-    font-size: 48px;
-    font-weight: 700;
-    color: #ffffff;
-    border-bottom: 4px solid var(--accent);
-    padding-bottom: 10px;
-    margin: 0 0 18px 0;
-  }
-  h2 {
-    font-size: 34px;
-    font-weight: 700;
-    color: #ffffff;
-    margin: 0 0 14px 0;
-  }
-  h2::before {
-    content: '';
-    display: inline-block;
-    width: 6px;
-    height: 28px;
-    background: var(--accent);
-    margin-right: 14px;
-    vertical-align: middle;
-    border-radius: 2px;
-  }
-  h3 { color: #c7d2fe; font-weight: 600; font-size: 24px; margin: 8px 0; }
-  p, li { margin: 6px 0; line-height: 1.45; }
-  ul, ol { margin: 8px 0 12px 0; padding-left: 28px; }
-  strong { color: #ffffff; }
-  a { color: #a5b4fc; }
-  blockquote {
-    border-left: 4px solid var(--accent);
-    background: rgba(255,255,255,0.06);
-    padding: 10px 18px;
-    color: #e0e7ff;
-    border-radius: 0 8px 8px 0;
-    margin: 10px 0;
-  }
-  code {
-    background: rgba(255,255,255,0.10);
-    color: #e0e7ff;
-    padding: 2px 6px;
-    border-radius: 4px;
-    font-size: 0.9em;
-  }
-  /* Constrain mermaid images so they always fit on a 720px-tall slide
-   * (16:9 @ 1280×720 default). Soft white card behind the diagram so the
-   * mermaid SVG (which renders on a transparent canvas with dark ink)
-   * stays readable on the navy gradient. */
-  img {
-    display: block;
-    margin: 8px auto;
-    max-width: 92%;
-    max-height: 420px;
-    object-fit: contain;
-    background: #ffffff;
-    border-radius: 10px;
-    padding: 10px;
-    box-shadow: 0 4px 16px rgba(0,0,0,0.25);
-  }
-  /* Tile grid for the local-agent versatility slide. */
-  .backends {
-    display: grid;
-    grid-template-columns: repeat(3, 1fr);
-    gap: 14px;
-    margin: 14px 0 12px 0;
-  }
-  .tile {
-    background: var(--accent-soft);
-    border: 2px solid var(--accent);
-    border-radius: 10px;
-    padding: 14px 16px;
-    text-align: center;
-    font-weight: 600;
-    color: #ffffff;
-  }
-  .tile small {
-    display: block;
-    margin-top: 4px;
-    font-weight: 400;
-    color: var(--muted);
-    font-size: 0.8em;
-  }
-  .tile.dots {
-    background: rgba(255,255,255,0.04);
-    border-style: dashed;
-    color: var(--muted);
-    font-size: 28px;
-  }
-  .pipeline {
-    text-align: center;
-    background: rgba(255,255,255,0.08);
-    border: 1px solid var(--border);
-    color: #f8fafc;
-    padding: 12px 18px;
-    border-radius: 10px;
-    margin: 6px auto 14px auto;
-    max-width: 760px;
-    font-weight: 600;
-  }
-  .pipeline span {
-    display: inline-block;
-    margin: 0 10px;
-  }
-  /* Tables sit on a solid white card so cell text reads cleanly against
-   * the dark gradient backdrop. Rounded outer corners + shadow match the
-   * mermaid card styling. `display:table` + `table-layout:fixed` are
-   * needed so the table actually consumes the full slide width — Marp's
-   * default theme leaves tables at content-natural width, which left
-   * blank gutters on either side. */
-  table {
-    display: table !important;
-    width: 100% !important;
-    table-layout: fixed;
-    border-collapse: collapse;
-    margin: 12px 0;
-    font-size: 0.82em;
-    color: #0f172a;
-    background: #ffffff;
-    border-radius: 10px;
-    overflow: hidden;
-    box-shadow: 0 4px 16px rgba(0,0,0,0.25);
-  }
-  td, th { word-wrap: break-word; overflow-wrap: anywhere; }
-  th {
-    background: var(--accent-deep);
-    color: #ffffff;
-    text-align: left;
-    padding: 9px 12px;
-    font-weight: 700;
-  }
-  td {
-    padding: 7px 12px;
-    border-bottom: 1px solid #e2e8f0;
-    color: #0f172a;
-  }
-  tr:nth-child(even) td { background: #ffffff; }
-  tr:nth-child(odd)  td { background: #f8fafc; }
-  /* code spans inside table cells need a light pill, not the dark one
-   * defined globally — otherwise they'd disappear into the white row. */
-  td code {
-    background: #eef2ff;
-    color: #1e293b;
-  }
-  /* Bold/strong inside cells must stay dark — the global rule paints
-   * <strong> white for the dark-gradient body text, which would render
-   * invisible against the white table card. */
-  td strong, th strong {
-    color: #0f172a;
-  }
-  th strong {
-    color: #ffffff;
-  }
-  /* Title slide — centered hero treatment on the same gradient. */
-  section.title {
-    text-align: center;
-    padding: 100px 80px;
-    justify-content: center;
-  }
-  section.title h1 {
-    display: inline-block;
-    font-size: 64px;
-    margin: 0 auto 20px auto;
-  }
-  section.title blockquote {
-    margin: 28px auto;
-    max-width: 900px;
-  }
-  /* Closing slide — same centered treatment. */
-  section.closing {
-    text-align: center;
-    padding: 100px 80px;
-    justify-content: center;
-  }
-  section.closing h2::before { display: none; }
-  section.closing blockquote { font-size: 26px; max-width: 900px; }
-</style>
-
-<!-- _class: title -->
-
-# Local Agents Marketplace
-
-### A trust-minimised, on-chain marketplace for LLM inference
-
-Buy and sell prompt → response with **bonded escrow** on **Vector** (UTxO).
-No platform fee. No central rate-limiter. Verifiable receipts on chain.
-
-> *Status: M1 lifecycle proven end-to-end on Vector testnet — 2026-05-05.*
-
----
-
-## The problem
-
-* AI inference is **centralised** — a handful of providers, opaque pricing, account-level rate limits.
-* Local-AI hardware is **idle most of the time** — GPUs and CPUs sit unused between user prompts.
-* Buyers and sellers can't transact directly without a **trusted intermediary** taking a cut.
-* "Verifiable" claims about which model ran are usually nothing more than the provider's word.
-
-We need a **two-sided marketplace** where:
-1. Sellers monetise spare compute,
-2. Buyers get a **cryptographic receipt** of what was actually computed,
-3. **Neither side has to trust the other** — bonds + on-chain settlement enforce honesty.
-
----
-
-## How it works — three actors, one chain
-
-```mermaid
-flowchart LR
-    subgraph CHAIN[Vector — UTxO]
-        ADV[Advert UTxO<br/>capability, price, bonds]
-        ESC[Escrow UTxO<br/>locked funds + state]
-    end
-
-    SUP[🧠 Supplier<br/>local LLM node]
-    BUY[💼 Buyer<br/>app or agent]
-    IDX[🔭 Indexer<br/>read-only bulletin board]
-
-    SUP -->|publish advert| ADV
-    BUY -->|discover| IDX
-    IDX -.->|index events| CHAIN
-    BUY -->|post escrow| ESC
-    BUY -->|prompt + escrow_ref| SUP
-    SUP -->|signed receipt| BUY
-    SUP -->|claim & submit| ESC
-    BUY -->|accept & pay| ESC
-```
-
-* **Supplier** runs the inference and publishes a price.
-* **Buyer** posts the work as a bonded escrow, sends the prompt, accepts the receipt.
-* **Indexer** is a public mirror of chain state — anyone can run one. No platform middleman.
-
----
-
-## The bonded-escrow lifecycle
-
-```mermaid
-sequenceDiagram
-    autonumber
-    participant B as Buyer
-    participant Chain as Vector
-    participant S as Supplier
-    participant LLM as Local LLM<br/>(OpenAI-compatible)
-
-    S->>Chain: PostAdvert (locks supplier_bond)
-    B->>Chain: PostEscrow → Open<br/>(locks payment + bonds)
-    B->>S: POST /v1/chat/completions
-    S->>Chain: ClaimEscrow → Claimed
-    S->>LLM: run inference
-    LLM-->>S: assistant message
-    S->>Chain: SubmitEscrow → Submitted
-    S-->>B: 200 OK + signed receipt
-    B->>Chain: AcceptEscrow — supplier paid
-```
-
-Every transition is on-chain, validated by an Aiken Plutus V3 script. The receipt is signed by the supplier's advert key; prompt and response hashes are bound on chain.
-
----
-
-## What if someone misbehaves?
-
-```mermaid
-flowchart LR
-    OPEN[Open]:::norm --> CLAIMED[Claimed]:::norm --> SUBMITTED[Submitted]:::norm --> ACCEPTED[Accepted ✓]:::good
-    OPEN -->|deliver_by passed| RECLAIM[Reclaimed<br/>buyer recovers funds]:::warn
-    CLAIMED -->|deliver_by passed| RECLAIM
-    SUBMITTED -->|accept_window passed| RELEASE[Released<br/>supplier collects]:::warn
-
-    classDef norm fill:#e0e7ff,stroke:#4f46e5
-    classDef good fill:#dcfce7,stroke:#16a34a
-    classDef warn fill:#fef3c7,stroke:#d97706
-```
-
-| Failure mode | Recovery | Loss bearer |
-|---|---|---|
-| Supplier never claims/submits | Buyer **Reclaim** after `deliver_by` | Supplier — forfeits supplier_bond |
-| Buyer disappears after Submit | Supplier **Release** after `submitted_at + 10 min` | Buyer — forfeits buyer_bond |
-
-Bonds make honesty cheaper than running away. Both sides have skin in the game; no third-party arbiter is needed for the M1 happy/timeout paths.
-
----
-
-## Local-agent versatility — any OpenAI-compatible engine
-
-<div class="pipeline">
-  <span>💼 Buyer</span> →
-  <span>🧠 Supplier (OpenAI-compatible HTTP wrapper)</span> →
-  <span>backend ↓</span>
-</div>
-
-<div class="backends">
-  <div class="tile">Ollama<br/><small>local model runner</small></div>
-  <div class="tile">vLLM<br/><small>high-throughput GPU</small></div>
-  <div class="tile">llama.cpp<br/><small>CPU · Metal · ROCm</small></div>
-  <div class="tile">OpenRouter<br/><small>cloud API mux</small></div>
-  <div class="tile">openclaw / custom<br/><small>arbitrary stack</small></div>
-  <div class="tile dots">…<br/><small>any OpenAI-shaped API</small></div>
-</div>
-
-The supplier service is a thin HTTP wrapper — it forwards `/v1/chat/completions` to whatever backend is configured. Each supplier picks its own model, hardware, and engine, then advertises a `capability_id`. Buyers shop on capability + model + price.
-
----
-
-## Capability + model = task selection
-
-| Capability ID | Example models | Use case |
-|---|---|---|
-| `llm.text.generate.v1` | Llama 3 · Qwen 2.5 · Mistral · GPT-4 | Q&A, drafts, summarisation |
-| `code.completion.v1` | Codestral · DeepSeek-Coder · StarCoder | IDE tab-complete, refactor |
-| `vision.describe.v1` | LLaVA · Qwen-VL · Pixtral | Image captioning, OCR |
-| `embedding.text.v1` | nomic-embed · BGE · ada-002 | RAG indexing |
-| `audio.transcribe.v1` | whisper-large-v3 · distil-whisper | Voice-to-text |
-| `…` | … | … |
-
-* New capabilities = **just publish a new advert**. No marketplace governance gate.
-* Today's deploy uses Ollama + Qwen 2.5; switching to vLLM with Llama 3 70B is a config change, not a code change.
-
----
-
-## OpenClaw — agentic teams on your existing OpenAI seat
-
-<div class="pipeline">
-  <span>🧠 Supplier wrapper</span> →
-  <span>🦾 OpenClaw</span> →
-  <span>🪪 OpenAI ChatGPT seat (flat-rate)</span>
-</div>
-
-* **Cost-efficient by default.** OpenClaw exposes your existing **OpenAI ChatGPT subscription** as an OpenAI-compatible API. The marketplace supplier dispatches its inference against it, so an entire **agentic team** (coders, QA automation testers, release curators…) draws from the **flat-rate seat you already pay for** — no per-token metering, no surprise invoices.
-* **Reliability layer (optional, marginal cost).** OpenClaw can fall back to the **Anthropic API** or **OpenRouter** model pool when the primary seat is rate-limited, the prompt requires a different model, or you want a **second-layer hardening** of responses (e.g. cross-provider consensus on critical outputs). You only pay metered tokens for the requests that actually fall through.
-
----
-
-## Proven on Vector testnet — 2026-05-05
-
-| Step | Tx hash |
-|---|---|
-| Buyer posts escrow | `af762561…c2f7bc5` |
-| Supplier claims | (lifecycle tx) |
-| Supplier submits receipt | (lifecycle tx) |
-| Buyer accepts | `125e3cfe…7dc100d` |
-
-Funds flowed correctly: supplier received `payment + supplier_bond` (3 ADA); buyer got `buyer_bond` back.
-
-* **Browser demo:** https://mp-buyers.vector.testnet.apexfusion.org
-* **Live event stream:** https://mp-indexer.vector.testnet.apexfusion.org
-* **Test suite:** 1098 unit tests passing, full lifecycle exercised in CI.
-
----
-
-## Why UTxO + Vector
-
-* **Bonded escrow is native** — Plutus validators enforce the state machine without a custom token or oracle.
-* **Vector settles in seconds** — the buyer→claim→submit→accept loop fits inside a single chat-completion's wall clock.
-* **AP3X token** for fees and bonds — fungible, transparent, on-chain.
-* **No custom rollup, no bridge** — runs on a production chain today.
-
----
-
-<!-- _class: closing -->
-
-## The pitch in one sentence
-
-> Local Agents Marketplace turns any OpenAI-compatible inference endpoint into a **trustlessly billable service** — buyer and supplier exchange prompt, response, and payment with on-chain proof that what was paid for is what was delivered.
-
-### Try it
-
-* **Indexer (live event stream):** https://mp-indexer.vector.testnet.apexfusion.org
-* **Buyer:** https://mp-buyers.vector.testnet.apexfusion.org
-* **Supplier:** https://mp-suppliers.vector.testnet.apexfusion.org/capability

docs/buyer/README.md

@@ -1,11 +1,11 @@
 # Vector Buyer App
 
-A self-contained app for purchasing AI compute on the Vector L2 testnet (Cardano). You submit prompts, pay with AP3X tokens via on-chain escrow, and receive LLM responses from supplier nodes.
+A self-contained app for commissioning AI compute on the Vector L2 testnet. You submit prompts, bond AP3X in on-chain escrow, and receive LLM responses from supplier nodes; the escrow settles in AP3X on accepted work.
 
 ## How It Works
 
-1. You type a **prompt** and set a **payment** amount in the web UI
-2. The app posts an **escrow transaction** on-chain, locking your payment + bond
+1. You type a **prompt** and set a **price** in the web UI
+2. The app posts an **escrow transaction** on-chain, locking price + bond
 3. The app sends the task to a **supplier node** via HTTP
 4. The supplier calls an LLM, **claims** the escrow, and **submits** the result hash on-chain
 5. You see the response in the UI and can **accept** to release payment, or **dispute**
@@ -193,8 +193,8 @@ If you accept the result, you get your 1 AP3X bond back. Total cost = payment +
 
 - **Never share your .env file** - it contains your wallet seed phrase
 - **Never share your data/ folder** - it contains task history
-- The seed phrase controls your wallet funds. Back it up securely.
-- Escrow protects you: the supplier cannot take your funds without delivering a result
+- The seed phrase controls your wallet. Back it up securely.
+- Escrow protects you: the supplier cannot take your escrowed value without delivering a result
 - You have a 10-minute dispute window after receiving a result
 
 ## Troubleshooting

docs/gateway.md

@@ -24,8 +24,8 @@ The audit confirmed 39 holes in the first draft (10 distinct themes). This docum
 | One-shot | `/openai/v1/chat/completions` → `llm.text.generate.v1`, 1 escrow/call, buffered (+ pseudo-stream) |
 | Chat session | `/openai/v1/chat/sessions*` → `llm.chat.v1`, 1 escrow/**session**, real SSE per turn, settle at close |
 | Routing | match `model` + required capability; pick an available supplier (status `free`\|`unknown`); **not cheapest**; fallback to next match on busy/offline |
-| Pricing | cost = matched supplier `price_lovelace`; **required balance = price + buyer_bond + supplier_bond + 5 ADA collateral + fee reserve** |
-| Wallet health | gateway **auto-consolidates** each wallet to `{5-ADA collateral, working}` after settle / on a tick |
+| Pricing | cost = matched supplier `price_lovelace`; **required balance = price + buyer_bond + supplier_bond + 5 AP3X collateral + fee reserve** |
+| Wallet health | gateway **auto-consolidates** each wallet to `{5-AP3X collateral, working}` after settle / on a tick |
 | Settlement | resolve the **Submitted** ref via indexer, `acceptResult`, **await confirm** before returning success; sweeper is backstop |
 | Sweeper | state-aware: Open/Claimed past `deliver_by` → reclaim; Submitted → retry Accept in-window; never blind-reclaim Submitted |
 | Response | OpenAI object `{id, object, created, model, choices, usage}` + `x_vector:{receipt, receipt_signature, escrow_ref}`; full streaming chunk transform + final usage chunk |
@@ -41,7 +41,7 @@ New standalone pnpm package `gateway/` (add to `pnpm-workspace.yaml`). Boots **o
 - SDK methods: `discoverSuppliers`, `submitPrompt` (`Marketplace.ts:242`, verifies receipt but **does not settle**), `startChat`/`endChat` (chat session; `endChat` returns `acceptedRef` and awaits Accept), `acceptResult`, `reclaim`.
 - Escrow value/bonds: buyer funds `totalLocked = price + buyer_bond + supplier_bond` at PostEscrow — `packages/shared/src/tx/escrow/postEscrow.ts:132-200`. On Accept supplier gets `payment+supplier_bond`, buyer gets `buyer_bond` back.
 - Submitted-ref resolution: the buyer `/v1/accept` queries the indexer and matches the Submitted row by `posted_at` — `buyer/src/server.ts` `/v1/accept`. Gateway must mirror this.
-- Wallet health: `planConsolidate` / `consolidateWallet` (`{collateral, working}` shape, `DEFAULT_COLLATERAL_LOVELACE = 5 ADA`) — `packages/shared/src/tx/wallet/`. Collateral assertion: `assertCollateralCandidate` — `packages/shared/src/tx/internal/liveCbor.ts`.
+- Wallet health: `planConsolidate` / `consolidateWallet` (`{collateral, working}` shape, `DEFAULT_COLLATERAL_LOVELACE = 5 AP3X`) — `packages/shared/src/tx/wallet/`. Collateral assertion: `assertCollateralCandidate` — `packages/shared/src/tx/internal/liveCbor.ts`.
 - Supplier chat-session SSE `{type:"token"|"done"|"error"}` + headers — `supplier/src/server.ts` `makeChatSessionHandlers`; buyer relay in `buyer/src/server.ts` `/v1/chat/message` (note: **verbatim** passthrough — the gateway must transform to OpenAI chunks).
 - SQLite pattern (`better-sqlite3`, WAL) — `buyer/src/db/archive.ts`. Auth helpers `timingSafeCompareStrings`, `createRateLimiter` — `buyer/src/auth.ts`.
 - Indexer `SupplierView` (`model`, `capability_id`, `price_lovelace`, `buyer_bond_lovelace`, `supplier_bond_lovelace`, `status` free/working/offline/unknown, `advert_status`) — `indexer/src/routes/suppliers.ts`.
@@ -100,7 +100,7 @@ Raw API key never persisted (only `sha256(key)`); `key_prefix` for display. `mas
 ## Request lifecycle — one-shot `/chat/completions`
 All on-chain steps run inside the per-key `mutex.run(...)`.
 1. **Route** (`selectSupplier`): indexer rows where `model` matches, `capability_id==="llm.text.generate.v1"`, `advert_status==="Active"`, live `status ∈ {free, unknown}`. Pick an available one (not cheapest). No match → 404 `model_not_found` (`type: invalid_request_error`).
-2. **Pre-flight**: required = `price + buyer_bond + supplier_bond + 5_000_000 (collateral) + ~2 AP3X fee`. Verify total balance **and** a pure-ADA UTxO ≥5 ADA exists (`assertCollateralCandidate`-style). Else 402 `insufficient_funds` + `x_vector:{required_lovelace, available_lovelace, collateral_ok, deposit_address}`.
+2. **Pre-flight**: required = `price + buyer_bond + supplier_bond + 5_000_000 (collateral) + ~2 AP3X fee`. Verify total balance **and** a pure-AP3X UTxO ≥5 AP3X exists (`assertCollateralCandidate`-style). Else 402 `insufficient_funds` + `x_vector:{required_lovelace, available_lovelace, collateral_ok, deposit_address}`.
 3. **Execute**: `submitPrompt({advertRef, messages, payment_lovelace:price, max_output_tokens})` (posts escrow → supplier → verifies receipt).
 4. **Settle**: resolve the **Submitted** escrow ref via the indexer (mirror buyer `/v1/accept`: match by buyer pkh + `posted_at`), `acceptResult(submittedRef)`, **await confirmation**. On 409/offline mid-route, retry next matching supplier (bounded). On any post-escrow failure, record + leave for the sweeper (don't block the response on reclaim).
 5. **Respond**: `{id:"chatcmpl-…", object:"chat.completion", created:<sec>, model, choices:[{index:0,message:{role:"assistant",content},finish_reason:"stop"}], usage:{prompt_tokens,completion_tokens,total_tokens}, x_vector:{receipt,receipt_signature,escrow_ref}}`. Record `usage`.
@@ -113,9 +113,15 @@ All on-chain steps run inside the per-key `mutex.run(...)`.
 - **close** → `endChat` (Submit + Accept, awaits) → settle + bond refund; record usage; wallet-health.
 - Sessions are reclaimable by the sweeper if abandoned before close.
 
+### Settle modes (`CHAT_SETTLE_MODE`, gateway + supplier, default `full`)
+- **full** — the lifecycle above: supplier Claims at start, Submits a transcript receipt at end, gateway Accepts (buyer charged).
+- **ticket** — the buyer's escrow post at open is the ONLY chain op. The supplier verifies the Open escrow as an entry ticket (`/v1/chat/start` → `{status:"ticket"}`) but never Claims/Submits; `/v1/chat/end` → `{status:"ended", settle_mode:"ticket"}` with no receipt; the sweeper Reclaims the Open escrow after `deliver_by`, returning all locked funds (net cost ≈ tx fees). Usage rows record `cost_lovelace: 0`. Sessions may outlive `deliver_by` (no hard-cap) — a live session whose ticket got reclaimed keeps running; only the idle janitor/close ends it. Close behavior is driven per-session by the supplier's `/end` response, so mixed-mode rollout windows settle correctly.
+- Ticket caveats (accepted, demo-grade): a supplier restart forgets used tickets, so an unexpired un-reclaimed escrow ref could reopen a session (bounded by `deliver_by` + the sweep); `/account.locked_in_escrow_lovelace` over-reports because the indexer never updates spent rows.
+- Suppliers additionally take `MAX_CHAT_SESSIONS` (default 1) — concurrent session slots; `/status` reports `active_sessions`/`max_sessions` and stays `free` until all slots are taken.
+
 ## Sweeper & wallet-health
 - **Sweeper** (interval ~60s): per key, classify escrows. Open/Claimed past `deliver_by` → `reclaim`. Submitted **with a verified receipt** still in accept-window → retry `acceptResult`. Submitted past accept-window → leave for supplier Release (log). Never blind-`reclaim` Submitted.
-- **Wallet-health tick**: periodically + after each settle, `planConsolidate` → `consolidateWallet` so every wallet keeps a ≥5 ADA collateral UTxO. Skip if already healthy.
+- **Wallet-health tick**: periodically + after each settle, `planConsolidate` → `consolidateWallet` so every wallet keeps a ≥5 AP3X collateral UTxO. Skip if already healthy.
 
 ## Errors (OpenAI shape `{error:{message, type, code, param:null}}`)
 no supplier → 404 `invalid_request_error`/`model_not_found`; underfunded/no-collateral → 402 `invalid_request_error`/`insufficient_funds`; bad/missing key → 401 `authentication_error`/`invalid_api_key`; rate limit → 429 `rate_limit_error` + `Retry-After`; supplier timeout → 504 `server_error`; supplier busy race (after retries) → 503 `server_error`/`overloaded`; receipt verify → 502 `server_error`/`receipt_verification_failed`; tx build → 409/502 `server_error`/`escrow_failed`; unsupported `tools`/`functions` → 400 `invalid_request_error`/`unsupported_parameter`.
@@ -139,7 +145,7 @@ Real billing/markup & fee hook; trustless/self-custody recovery; multi-funding-U
 ## Verification (testnet, end-to-end)
 1. Indexer + Ogmios + ≥1 `llm.text.generate.v1` and ≥1 `llm.chat.v1` supplier for the same `model`, Active.
 2. Boot gateway: `LIVE_CHAIN=1`, `GATEWAY_MASTER_KEY=$(openssl rand -hex 32)`, `OGMIOS_URL`, `INDEXER_URL`, `NETWORK_ID`, `GATEWAY_DB_DIR`, `GATEWAY_PORT`.
-3. `POST /signup` → key + deposit address. Fund ≥ price + both bonds + 5 ADA collateral + ~2 ADA fees; confirm via `GET /account` (`collateral_ok:true`).
+3. `POST /signup` → key + deposit address. Fund ≥ price + both bonds + 5 AP3X collateral + ~2 AP3X fees; confirm via `GET /account` (`collateral_ok:true`).
 4. One-shot non-stream via curl and via the OpenAI Python/JS SDK (`base_url=…/openai/v1`) → valid `chat.completion` + `x_vector`; confirm the escrow **settled (Accepted)** and a `usage` row exists.
 5. **Run a 2nd and 3rd request** → confirm wallet-health kept `collateral_ok` and they succeed (the audit's request-#2 failure must not reproduce).
 6. One-shot `stream:true` → chunks + final usage + `[DONE]`.

docs/local-llm-marketplace.md

@@ -1,412 +0,0 @@
----
-marp: true
-paginate: true
-size: 16:9
----
-
-<!--
-  EXECUTIVE_SUMMARY.md — slide-style overview of the Local Agents Marketplace.
-  Build PDF + PPTX with: bash docs/build-deck.sh
-  Edit this file; mermaid blocks are pre-rendered to docs/img/diagram-*.png.
--->
-
-<style>
-  /* ── Local Agents Marketplace — unified dark theme ────────────────────
-   * Every slide uses the same navy→indigo gradient as the title slide so
-   * the deck reads as one continuous document. Title and closing slides
-   * keep the centered hero treatment; content slides stay top-aligned but
-   * share the same palette and accent rules. */
-  :root {
-    --ink: #f1f5f9;          /* primary text on dark gradient */
-    --muted: #94a3b8;
-    --accent: #818cf8;       /* indigo-400, readable on the dark gradient */
-    --accent-deep: #4f46e5;  /* used for solid accents (h2 bar, table head) */
-    --accent-soft: rgba(129,140,248,0.12);
-    --gradient: linear-gradient(135deg, #0f172a 0%, #1e293b 50%, #312e81 100%);
-    --row-even: rgba(255,255,255,0.04);
-    --row-odd:  rgba(255,255,255,0.02);
-    --border:   rgba(255,255,255,0.10);
-  }
-  section {
-    background: var(--gradient);
-    color: var(--ink);
-    font-family: 'Inter', 'Segoe UI', system-ui, -apple-system, sans-serif;
-    font-size: 22px;
-    padding: 40px 60px;
-    display: flex;
-    flex-direction: column;
-    justify-content: flex-start;
-  }
-  section::after {
-    content: attr(data-marpit-pagination) ' / ' attr(data-marpit-pagination-total);
-    color: var(--muted);
-    font-size: 13px;
-  }
-  h1 {
-    font-size: 48px;
-    font-weight: 700;
-    color: #ffffff;
-    border-bottom: 4px solid var(--accent);
-    padding-bottom: 10px;
-    margin: 0 0 18px 0;
-  }
-  h2 {
-    font-size: 34px;
-    font-weight: 700;
-    color: #ffffff;
-    margin: 0 0 14px 0;
-  }
-  h2::before {
-    content: '';
-    display: inline-block;
-    width: 6px;
-    height: 28px;
-    background: var(--accent);
-    margin-right: 14px;
-    vertical-align: middle;
-    border-radius: 2px;
-  }
-  h3 { color: #c7d2fe; font-weight: 600; font-size: 24px; margin: 8px 0; }
-  p, li { margin: 6px 0; line-height: 1.45; }
-  ul, ol { margin: 8px 0 12px 0; padding-left: 28px; }
-  strong { color: #ffffff; }
-  a { color: #a5b4fc; }
-  blockquote {
-    border-left: 4px solid var(--accent);
-    background: rgba(255,255,255,0.06);
-    padding: 10px 18px;
-    color: #e0e7ff;
-    border-radius: 0 8px 8px 0;
-    margin: 10px 0;
-  }
-  code {
-    background: rgba(255,255,255,0.10);
-    color: #e0e7ff;
-    padding: 2px 6px;
-    border-radius: 4px;
-    font-size: 0.9em;
-  }
-  /* Constrain mermaid images so they always fit on a 720px-tall slide
-   * (16:9 @ 1280×720 default). Soft white card behind the diagram so the
-   * mermaid SVG (which renders on a transparent canvas with dark ink)
-   * stays readable on the navy gradient. */
-  img {
-    display: block;
-    margin: 8px auto;
-    max-width: 92%;
-    max-height: 420px;
-    object-fit: contain;
-    background: #ffffff;
-    border-radius: 10px;
-    padding: 10px;
-    box-shadow: 0 4px 16px rgba(0,0,0,0.25);
-  }
-  /* Tile grid for the local-agent versatility slide. */
-  .backends {
-    display: grid;
-    grid-template-columns: repeat(3, 1fr);
-    gap: 14px;
-    margin: 14px 0 12px 0;
-  }
-  .tile {
-    background: var(--accent-soft);
-    border: 2px solid var(--accent);
-    border-radius: 10px;
-    padding: 14px 16px;
-    text-align: center;
-    font-weight: 600;
-    color: #ffffff;
-  }
-  .tile small {
-    display: block;
-    margin-top: 4px;
-    font-weight: 400;
-    color: var(--muted);
-    font-size: 0.8em;
-  }
-  .tile.dots {
-    background: rgba(255,255,255,0.04);
-    border-style: dashed;
-    color: var(--muted);
-    font-size: 28px;
-  }
-  .pipeline {
-    text-align: center;
-    background: rgba(255,255,255,0.08);
-    border: 1px solid var(--border);
-    color: #f8fafc;
-    padding: 12px 18px;
-    border-radius: 10px;
-    margin: 6px auto 14px auto;
-    max-width: 760px;
-    font-weight: 600;
-  }
-  .pipeline span {
-    display: inline-block;
-    margin: 0 10px;
-  }
-  /* Tables sit on a solid white card so cell text reads cleanly against
-   * the dark gradient backdrop. Rounded outer corners + shadow match the
-   * mermaid card styling. `display:table` + `table-layout:fixed` are
-   * needed so the table actually consumes the full slide width — Marp's
-   * default theme leaves tables at content-natural width, which left
-   * blank gutters on either side. */
-  table {
-    display: table !important;
-    width: 100% !important;
-    table-layout: fixed;
-    border-collapse: collapse;
-    margin: 12px 0;
-    font-size: 0.82em;
-    color: #0f172a;
-    background: #ffffff;
-    border-radius: 10px;
-    overflow: hidden;
-    box-shadow: 0 4px 16px rgba(0,0,0,0.25);
-  }
-  td, th { word-wrap: break-word; overflow-wrap: anywhere; }
-  th {
-    background: var(--accent-deep);
-    color: #ffffff;
-    text-align: left;
-    padding: 9px 12px;
-    font-weight: 700;
-  }
-  td {
-    padding: 7px 12px;
-    border-bottom: 1px solid #e2e8f0;
-    color: #0f172a;
-  }
-  tr:nth-child(even) td { background: #ffffff; }
-  tr:nth-child(odd)  td { background: #f8fafc; }
-  /* code spans inside table cells need a light pill, not the dark one
-   * defined globally — otherwise they'd disappear into the white row. */
-  td code {
-    background: #eef2ff;
-    color: #1e293b;
-  }
-  /* Bold/strong inside cells must stay dark — the global rule paints
-   * <strong> white for the dark-gradient body text, which would render
-   * invisible against the white table card. */
-  td strong, th strong {
-    color: #0f172a;
-  }
-  th strong {
-    color: #ffffff;
-  }
-  /* Title slide — centered hero treatment on the same gradient. */
-  section.title {
-    text-align: center;
-    padding: 100px 80px;
-    justify-content: center;
-  }
-  section.title h1 {
-    display: inline-block;
-    font-size: 64px;
-    margin: 0 auto 20px auto;
-  }
-  section.title blockquote {
-    margin: 28px auto;
-    max-width: 900px;
-  }
-  /* Closing slide — same centered treatment. */
-  section.closing {
-    text-align: center;
-    padding: 100px 80px;
-    justify-content: center;
-  }
-  section.closing h2::before { display: none; }
-  section.closing blockquote { font-size: 26px; max-width: 900px; }
-</style>
-
-<!-- _class: title -->
-
-# Local Agents Marketplace
-
-### A trust-minimised, on-chain marketplace for LLM inference
-
-Buy and sell prompt → response with **bonded escrow** on **Vector** (UTxO).
-No platform fee. No central rate-limiter. Verifiable receipts on chain.
-
-> *Status: M1 lifecycle proven end-to-end on Vector testnet — 2026-05-05.*
-
----
-
-## The problem
-
-* AI inference is **centralised** — a handful of providers, opaque pricing, account-level rate limits.
-* Local-AI hardware is **idle most of the time** — GPUs and CPUs sit unused between user prompts.
-* Buyers and sellers can't transact directly without a **trusted intermediary** taking a cut.
-* "Verifiable" claims about which model ran are usually nothing more than the provider's word.
-
-We need a **two-sided marketplace** where:
-1. Sellers monetise spare compute,
-2. Buyers get a **cryptographic receipt** of what was actually computed,
-3. **Neither side has to trust the other** — bonds + on-chain settlement enforce honesty.
-
----
-
-## How it works — three actors, one chain
-
-```mermaid
-flowchart LR
-    subgraph CHAIN[Vector — UTxO]
-        ADV[Advert UTxO<br/>capability, price, bonds]
-        ESC[Escrow UTxO<br/>locked funds + state]
-    end
-
-    SUP[🧠 Supplier<br/>local LLM node]
-    BUY[💼 Buyer<br/>app or agent]
-    IDX[🔭 Indexer<br/>read-only bulletin board]
-
-    SUP -->|publish advert| ADV
-    BUY -->|discover| IDX
-    IDX -.->|index events| CHAIN
-    BUY -->|post escrow| ESC
-    BUY -->|prompt + escrow_ref| SUP
-    SUP -->|signed receipt| BUY
-    SUP -->|claim & submit| ESC
-    BUY -->|accept & pay| ESC
-```
-
-* **Supplier** runs the inference and publishes a price.
-* **Buyer** posts the work as a bonded escrow, sends the prompt, accepts the receipt.
-* **Indexer** is a public mirror of chain state — anyone can run one. No platform middleman.
-
----
-
-## The bonded-escrow lifecycle
-
-```mermaid
-sequenceDiagram
-    autonumber
-    participant B as Buyer
-    participant Chain as Vector
-    participant S as Supplier
-    participant LLM as Local LLM<br/>(OpenAI-compatible)
-
-    S->>Chain: PostAdvert (locks supplier_bond)
-    B->>Chain: PostEscrow → Open<br/>(locks payment + bonds)
-    B->>S: POST /v1/chat/completions
-    S->>Chain: ClaimEscrow → Claimed
-    S->>LLM: run inference
-    LLM-->>S: assistant message
-    S->>Chain: SubmitEscrow → Submitted
-    S-->>B: 200 OK + signed receipt
-    B->>Chain: AcceptEscrow — supplier paid
-```
-
-Every transition is on-chain, validated by an Aiken Plutus V3 script. The receipt is signed by the supplier's advert key; prompt and response hashes are bound on chain.
-
----
-
-## What if someone misbehaves?
-
-```mermaid
-flowchart LR
-    OPEN[Open]:::norm --> CLAIMED[Claimed]:::norm --> SUBMITTED[Submitted]:::norm --> ACCEPTED[Accepted ✓]:::good
-    OPEN -->|deliver_by passed| RECLAIM[Reclaimed<br/>buyer recovers funds]:::warn
-    CLAIMED -->|deliver_by passed| RECLAIM
-    SUBMITTED -->|accept_window passed| RELEASE[Released<br/>supplier collects]:::warn
-
-    classDef norm fill:#e0e7ff,stroke:#4f46e5
-    classDef good fill:#dcfce7,stroke:#16a34a
-    classDef warn fill:#fef3c7,stroke:#d97706
-```
-
-| Failure mode | Recovery | Loss bearer |
-|---|---|---|
-| Supplier never claims/submits | Buyer **Reclaim** after `deliver_by` | Supplier — forfeits supplier_bond |
-| Buyer disappears after Submit | Supplier **Release** after `submitted_at + 10 min` | Buyer — forfeits buyer_bond |
-
-Bonds make honesty cheaper than running away. Both sides have skin in the game; no third-party arbiter is needed for the M1 happy/timeout paths.
-
----
-
-## Local-agent versatility — any OpenAI-compatible engine
-
-<div class="pipeline">
-  <span>💼 Buyer</span> →
-  <span>🧠 Supplier (OpenAI-compatible HTTP wrapper)</span> →
-  <span>backend ↓</span>
-</div>
-
-<div class="backends">
-  <div class="tile">Ollama<br/><small>local model runner</small></div>
-  <div class="tile">vLLM<br/><small>high-throughput GPU</small></div>
-  <div class="tile">llama.cpp<br/><small>CPU · Metal · ROCm</small></div>
-  <div class="tile">OpenRouter<br/><small>cloud API mux</small></div>
-  <div class="tile">openclaw / custom<br/><small>arbitrary stack</small></div>
-  <div class="tile dots">…<br/><small>any OpenAI-shaped API</small></div>
-</div>
-
-The supplier service is a thin HTTP wrapper — it forwards `/v1/chat/completions` to whatever backend is configured. Each supplier picks its own model, hardware, and engine, then advertises a `capability_id`. Buyers shop on capability + model + price.
-
----
-
-## Capability + model = task selection
-
-| Capability ID | Example models | Use case |
-|---|---|---|
-| `llm.text.generate.v1` | Llama 3 · Qwen 2.5 · Mistral · GPT-4 | Q&A, drafts, summarisation |
-| `code.completion.v1` | Codestral · DeepSeek-Coder · StarCoder | IDE tab-complete, refactor |
-| `vision.describe.v1` | LLaVA · Qwen-VL · Pixtral | Image captioning, OCR |
-| `embedding.text.v1` | nomic-embed · BGE · ada-002 | RAG indexing |
-| `audio.transcribe.v1` | whisper-large-v3 · distil-whisper | Voice-to-text |
-| `…` | … | … |
-
-* New capabilities = **just publish a new advert**. No marketplace governance gate.
-* Today's deploy uses Ollama + Qwen 2.5; switching to vLLM with Llama 3 70B is a config change, not a code change.
-
----
-
-## OpenClaw — agentic teams on your existing OpenAI seat
-
-<div class="pipeline">
-  <span>🧠 Supplier wrapper</span> →
-  <span>🦾 OpenClaw</span> →
-  <span>🪪 OpenAI ChatGPT seat (flat-rate)</span>
-</div>
-
-* **Cost-efficient by default.** OpenClaw exposes your existing **OpenAI ChatGPT subscription** as an OpenAI-compatible API. The marketplace supplier dispatches its inference against it, so an entire **agentic team** (coders, QA automation testers, release curators…) draws from the **flat-rate seat you already pay for** — no per-token metering, no surprise invoices.
-* **Reliability layer (optional, marginal cost).** OpenClaw can fall back to the **Anthropic API** or **OpenRouter** model pool when the primary seat is rate-limited, the prompt requires a different model, or you want a **second-layer hardening** of responses (e.g. cross-provider consensus on critical outputs). You only pay metered tokens for the requests that actually fall through.
-
----
-
-## Proven on Vector testnet — 2026-05-05
-
-| Step | Tx hash |
-|---|---|
-| Buyer posts escrow | `af762561…c2f7bc5` |
-| Supplier claims | (lifecycle tx) |
-| Supplier submits receipt | (lifecycle tx) |
-| Buyer accepts | `125e3cfe…7dc100d` |
-
-Funds flowed correctly: supplier received `payment + supplier_bond` (3 ADA); buyer got `buyer_bond` back.
-
-* **Browser demo:** https://mp-buyers.vector.testnet.apexfusion.org
-* **Live event stream:** https://mp-indexer.vector.testnet.apexfusion.org
-* **Test suite:** 1098 unit tests passing, full lifecycle exercised in CI.
-
----
-
-## Why UTxO + Vector
-
-* **Bonded escrow is native** — Plutus validators enforce the state machine without a custom token or oracle.
-* **Vector settles in seconds** — the buyer→claim→submit→accept loop fits inside a single chat-completion's wall clock.
-* **AP3X token** for fees and bonds — fungible, transparent, on-chain.
-* **No custom rollup, no bridge** — runs on a production chain today.
-
----
-
-<!-- _class: closing -->
-
-## The pitch in one sentence
-
-> Local Agents Marketplace turns any OpenAI-compatible inference endpoint into a **trustlessly billable service** — buyer and supplier exchange prompt, response, and payment with on-chain proof that what was paid for is what was delivered.
-
-### Try it
-
-* **Indexer (live event stream):** https://mp-indexer.vector.testnet.apexfusion.org
-* **Buyer:** https://mp-buyers.vector.testnet.apexfusion.org
-* **Supplier:** https://mp-suppliers.vector.testnet.apexfusion.org/capability

docs/supplier/README.md

@@ -164,8 +164,8 @@ If the buyer does NOT accept within the dispute window (10 minutes), you can col
 
 - **Never share your .env file** - it contains your wallet seed phrase
 - **Never share your data/ folder** - it contains task history
-- The seed phrase controls your wallet funds. Back it up securely.
-- The node only responds to buyers who have already locked funds on-chain, so you cannot be scammed out of compute.
+- The seed phrase controls your wallet. Back it up securely.
+- The node only responds to buyers who have already locked value on-chain, so you cannot be scammed out of compute.
 
 ## Troubleshooting
 

gateway/.env.example

@@ -30,5 +30,7 @@ GATEWAY_DB_DIR=./data/gateway
 # KEY_RATE_MAX=60             # requests per API key per window
 # KEY_RATE_WINDOW_MS=60000
 # SWEEPER_INTERVAL_MS=60000
+# DEMO_SESSION_IDLE_MS=180000 # close demo chat sessions idle this long (frees the supplier slot)
+# CHAT_SETTLE_MODE=full       # "full" (Claim/Submit/Accept) or "ticket" (escrow post only; sweeper reclaims) — must match suppliers
 # WALLET_HEALTH_INTERVAL_MS=600000
 # SDK_REGISTRY_MAX=500        # max cached per-key Marketplace instances

gateway/Dockerfile

@@ -22,6 +22,8 @@ RUN apk add --no-cache python3 make g++
 RUN corepack enable && corepack prepare pnpm@10.33.2 --activate
 
 COPY pnpm-workspace.yaml pnpm-lock.yaml package.json tsconfig.base.json .npmrc ./
+# pnpm-lock.yaml's patchedDependencies hash the patch files at install time.
+COPY patches ./patches
 COPY packages/shared/package.json ./packages/shared/package.json
 COPY buyer/package.json ./buyer/package.json
 COPY gateway/package.json ./gateway/package.json

gateway/package.json

@@ -7,7 +7,8 @@
     "typecheck": "tsc --noEmit",
     "start": "tsx src/index.ts",
     "dev": "tsx src/index.ts",
-    "rotate-master-key": "tsx src/cli/rotate-master-key.ts"
+    "rotate-master-key": "tsx src/cli/rotate-master-key.ts",
+    "create-demo-key": "tsx src/cli/create-demo-key.ts"
   },
   "dependencies": {
     "@marketplace/buyer": "workspace:*",

gateway/src/account/routes.ts

@@ -15,7 +15,7 @@ import { seal } from "../crypto/seal.js";
 import { hashApiKey, requireKey } from "../middleware/apiKeyAuth.js";
 import { asyncHandler } from "../middleware/http.js";
 import { totalLovelace, hasCollateral } from "../onchain/preflight.js";
-import { badRequest } from "../openai/errors.js";
+import { badRequest, forbidden } from "../openai/errors.js";
 
 const ACTIVE_ESCROW_STATES = new Set(["Open", "Claimed", "Submitted"]);
 
@@ -89,7 +89,10 @@ export function makeAccountHandler(deps: GatewayDeps) {
 
     res.status(200).json({
       key_prefix: keyRow.key_prefix,
-      deposit_address: keyRow.deposit_address,
+      // The shared demo key's wallet is operator-funded; hiding the deposit
+      // address keeps strangers from confusing it with their own wallet.
+      deposit_address: keyRow.demo ? null : keyRow.deposit_address,
+      ...(keyRow.demo ? { demo: true } : {}),
       balance: {
         available_lovelace: available.toString(),
         locked_in_escrow_lovelace: lockedInEscrow.toString(),
@@ -118,6 +121,11 @@ export function makeAccountHandler(deps: GatewayDeps) {
 export function makeWithdrawHandler(deps: GatewayDeps) {
   return asyncHandler(async (req: Request, res: Response) => {
     const keyRow = requireKey(req);
+    if (keyRow.demo) {
+      // The demo key is shared publicly; without this, anyone holding it
+      // could drain the operator-funded wallet to their own address.
+      throw forbidden("demo_key_restricted", "withdrawals are disabled for the shared demo key");
+    }
     const body = (req.body ?? {}) as Record<string, unknown>;
 
     const toAddress = body.to_address;

gateway/src/cli/create-demo-key.ts

@@ -0,0 +1,76 @@
+/**
+ * gateway/src/cli/create-demo-key.ts — mint the shared demo API key.
+ *
+ * Same custodial machinery as POST /signup, but sets demo=1: withdraw is
+ * blocked, /account hides the deposit address, chat completions run through
+ * the session-backed demo executor, and per-IP rate limits apply. Public
+ * signup can never mint a demo key — this CLI is the only path.
+ *
+ * Run inside the gateway container (hits the volume-mounted DB):
+ *   docker exec <gateway> pnpm --filter @marketplace/gateway create-demo-key
+ * Or locally:
+ *   GATEWAY_MASTER_KEY=<64hex> GATEWAY_DB_DIR=./data/gateway NETWORK_ID=1 \
+ *   pnpm --filter @marketplace/gateway create-demo-key [--label shared-demo]
+ *
+ * The raw key is printed ONCE and never stored (only its sha256).
+ */
+
+import { randomBytes, randomUUID } from "crypto";
+import { GatewayStore } from "../db/store.js";
+import { genPrivKeyHex, deriveWalletKey } from "../wallet.js";
+import { seal } from "../crypto/seal.js";
+import { hashApiKey } from "../middleware/apiKeyAuth.js";
+
+const HEX64 = /^[0-9a-fA-F]{64}$/;
+
+function arg(name: string): string | undefined {
+  const i = process.argv.indexOf(`--${name}`);
+  return i >= 0 ? process.argv[i + 1] : undefined;
+}
+
+function main(): void {
+  const masterKey = process.env.GATEWAY_MASTER_KEY;
+  if (!masterKey || !HEX64.test(masterKey)) {
+    throw new Error("create-demo-key: GATEWAY_MASTER_KEY must be 64 hex chars");
+  }
+  const dbDir = arg("db") ?? process.env.GATEWAY_DB_DIR ?? "./data/gateway";
+  const networkId: 0 | 1 = process.env.NETWORK_ID === "1" ? 1 : 0;
+  const label = arg("label") ?? "shared-demo";
+
+  const privHex = genPrivKeyHex();
+  const walletKey = deriveWalletKey(privHex, networkId);
+  const sealed = seal(privHex, masterKey);
+
+  const rawKey = `vk_${networkId === 1 ? "live" : "test"}_${randomBytes(24).toString("hex")}`;
+
+  const store = new GatewayStore(dbDir);
+  store.insertKey({
+    id: randomUUID(),
+    key_hash: hashApiKey(rawKey),
+    key_prefix: rawKey.slice(0, 12),
+    label,
+    wallet_pkh: walletKey.pubKeyHash,
+    deposit_address: walletKey.address,
+    enc_priv_nonce: sealed.nonce,
+    enc_priv_ct: sealed.ct,
+    enc_priv_tag: sealed.tag,
+    master_key_version: 1,
+    created_at: Date.now(),
+    demo: 1,
+  });
+
+  // eslint-disable-next-line no-console
+  console.log(
+    [
+      `[gateway] demo key created (label=${label}, network=${networkId === 1 ? "mainnet" : "testnet"})`,
+      ``,
+      `  api_key:         ${rawKey}`,
+      `  deposit_address: ${walletKey.address}`,
+      ``,
+      `Save the api_key now — it is shown only once. Fund the deposit address`,
+      `with the demo float (withdrawals are disabled for this key).`,
+    ].join("\n"),
+  );
+}
+
+main();

gateway/src/config.ts

@@ -20,6 +20,10 @@
  *   KEY_RATE_MAX / KEY_RATE_WINDOW_MS        — per-key request limiter.
  *   SWEEPER_INTERVAL_MS / WALLET_HEALTH_INTERVAL_MS — background tick cadences.
  *   SDK_REGISTRY_MAX    — max cached per-key Marketplace instances (LRU).
+ *   GATEWAY_CORS_ORIGINS — comma-separated allowlist of browser Origins that may
+ *                         call the gateway cross-origin (e.g. the marketplace
+ *                         frontend's "Generate API key" page). Default empty =
+ *                         CORS off (no Access-Control-* headers emitted).
  */
 
 const HEX64_RE = /^[0-9a-fA-F]{64}$/;
@@ -35,9 +39,23 @@ export interface GatewayConfig {
   dbDir: string;
   signupRate: { max: number; windowMs: number };
   keyRate: { max: number; windowMs: number };
+  demoIpRate: { max: number; windowMs: number };
   sweeperIntervalMs: number;
+  /** Close demo chat sessions idle this long (frees the supplier slot).
+   * Keep below the supplier's CHAT_IDLE_TIMEOUT_MS so the gateway closes
+   * first and records usage. */
+  demoSessionIdleMs: number;
+  /** Chat-session settlement mode this deployment runs. "full" = suppliers
+   * Claim/Submit and the gateway Accepts; "ticket" = the escrow post at open
+   * is the only chain op and the sweeper Reclaims it after deliver_by.
+   * Must agree with the suppliers' CHAT_SETTLE_MODE. */
+  chatSettleMode: "full" | "ticket";
   walletHealthIntervalMs: number;
   sdkRegistryMax: number;
+  corsOrigins: string[];
+  /** Capability id the /v1/ocr/extract route escrows against. Model-scoped
+   * per the marketplace convention (one model, one id). Env: OCR_CAPABILITY_ID. */
+  ocrCapabilityId: string;
 }
 
 function requireField(env: Record<string, string | undefined>, name: string): string {
@@ -48,6 +66,14 @@ function requireField(env: Record<string, string | undefined>, name: string): st
   return v;
 }
 
+function chatSettleMode(env: Record<string, string | undefined>): "full" | "ticket" {
+  const v = env.CHAT_SETTLE_MODE ?? "full";
+  if (v !== "full" && v !== "ticket") {
+    throw new Error('loadConfig: CHAT_SETTLE_MODE must be "full" or "ticket"');
+  }
+  return v;
+}
+
 function posInt(
   env: Record<string, string | undefined>,
   name: string,
@@ -106,8 +132,19 @@ export function loadConfig(env: Record<string, string | undefined>): GatewayConf
       max: posInt(env, "KEY_RATE_MAX", 60),
       windowMs: posInt(env, "KEY_RATE_WINDOW_MS", 60 * 1000),
     },
+    demoIpRate: {
+      max: posInt(env, "DEMO_IP_RATE_MAX", 20),
+      windowMs: posInt(env, "DEMO_IP_RATE_WINDOW_MS", 60 * 1000),
+    },
     sweeperIntervalMs: posInt(env, "SWEEPER_INTERVAL_MS", 60 * 1000),
+    demoSessionIdleMs: posInt(env, "DEMO_SESSION_IDLE_MS", 3 * 60 * 1000),
+    chatSettleMode: chatSettleMode(env),
     walletHealthIntervalMs: posInt(env, "WALLET_HEALTH_INTERVAL_MS", 10 * 60 * 1000),
     sdkRegistryMax: posInt(env, "SDK_REGISTRY_MAX", 500),
+    ocrCapabilityId: env.OCR_CAPABILITY_ID ?? "ocr.page.extract.chandra-ocr-2.v1",
+    corsOrigins: (env.GATEWAY_CORS_ORIGINS ?? "")
+      .split(",")
+      .map((o) => o.trim().replace(/\/+$/, ""))
+      .filter((o) => o !== ""),
   };
 }

gateway/src/db/schema.ts

@@ -23,7 +23,8 @@ CREATE TABLE IF NOT EXISTS api_keys (
   enc_priv_tag       TEXT NOT NULL,
   master_key_version INTEGER NOT NULL DEFAULT 1,
   created_at         INTEGER NOT NULL,
-  disabled           INTEGER NOT NULL DEFAULT 0
+  disabled           INTEGER NOT NULL DEFAULT 0,
+  demo               INTEGER NOT NULL DEFAULT 0
 );
 CREATE INDEX IF NOT EXISTS idx_keys_hash ON api_keys(key_hash);
 

gateway/src/db/store.ts

@@ -24,6 +24,8 @@ export interface ApiKeyRow {
   master_key_version: number;
   created_at: number;
   disabled: number;
+  /** 1 = shared demo key: withdraw blocked, session-backed completions, per-IP limits. */
+  demo: number;
 }
 
 export interface UsageRow {
@@ -81,13 +83,20 @@ export class GatewayStore {
     this.db.pragma("journal_mode = WAL");
     this.db.exec(CREATE_TABLES_SQL);
 
+    // Additive migration for DBs created before the demo flag existed
+    // (CREATE TABLE IF NOT EXISTS won't alter an existing table).
+    const keyCols = this.db.pragma("table_info(api_keys)") as Array<{ name: string }>;
+    if (!keyCols.some((c) => c.name === "demo")) {
+      this.db.exec("ALTER TABLE api_keys ADD COLUMN demo INTEGER NOT NULL DEFAULT 0");
+    }
+
     this.sInsertKey = this.db.prepare(`
       INSERT INTO api_keys (
         id, key_hash, key_prefix, label, wallet_pkh, deposit_address,
-        enc_priv_nonce, enc_priv_ct, enc_priv_tag, master_key_version, created_at, disabled
+        enc_priv_nonce, enc_priv_ct, enc_priv_tag, master_key_version, created_at, disabled, demo
       ) VALUES (
         @id, @key_hash, @key_prefix, @label, @wallet_pkh, @deposit_address,
-        @enc_priv_nonce, @enc_priv_ct, @enc_priv_tag, @master_key_version, @created_at, 0
+        @enc_priv_nonce, @enc_priv_ct, @enc_priv_tag, @master_key_version, @created_at, 0, @demo
       )
     `);
     this.sGetKeyByHash = this.db.prepare(`SELECT * FROM api_keys WHERE key_hash = @key_hash`);
@@ -133,8 +142,8 @@ export class GatewayStore {
     this.sListOpenSessions = this.db.prepare(`SELECT * FROM sessions WHERE state = 'open'`);
   }
 
-  insertKey(row: Omit<ApiKeyRow, "disabled">): void {
-    this.sInsertKey.run({ ...row, label: row.label ?? null });
+  insertKey(row: Omit<ApiKeyRow, "disabled" | "demo"> & { demo?: number }): void {
+    this.sInsertKey.run({ ...row, label: row.label ?? null, demo: row.demo ?? 0 });
   }
 
   getKeyByHash(keyHash: string): ApiKeyRow | undefined {
@@ -193,4 +202,8 @@ export class GatewayStore {
   listOpenSessions(): SessionRow[] {
     return this.sListOpenSessions.all() as SessionRow[];
   }
+
+  listOpenSessionsByKey(keyId: string): SessionRow[] {
+    return this.listOpenSessions().filter((s) => s.key_id === keyId);
+  }
 }

gateway/src/index.ts

@@ -57,7 +57,10 @@ export function runMain(env: Record<string, string | undefined>): void {
 const invokedDirectly =
   typeof process !== "undefined" &&
   process.argv?.[1] !== undefined &&
-  /gateway\/(src|dist)\/index\.(t|j)s$/.test(process.argv[1]);
+  // Accept both path separators: on Windows process.argv[1] arrives with
+  // backslashes, so a forward-slash-only pattern silently never matches and
+  // the process exits 0 without ever starting the server.
+  /gateway[\\/](src|dist)[\\/]index\.(t|j)s$/.test(process.argv[1]);
 
 if (invokedDirectly) {
   try {

gateway/src/middleware/rateLimit.ts

@@ -60,7 +60,28 @@ export function ipRateLimit(opts: { max: number; windowMs: number }) {
   return makeLimiter(opts, (req) => req.ip ?? "unknown");
 }
 
-/** Per-API-key limiter (apply AFTER apiKeyAuth). */
+/** Per-API-key limiter (apply AFTER apiKeyAuth). Demo keys are exempt: the
+ * shared demo key would let one abusive IP exhaust everyone's budget — the
+ * per-IP demoIpRateLimit is the demo control instead. */
 export function keyRateLimit(opts: { max: number; windowMs: number }) {
-  return makeLimiter(opts, (req) => requireKey(req).id);
+  const limit = makeLimiter(opts, (req) => requireKey(req).id);
+  return (req: Request, res: Response, next: NextFunction): void => {
+    if (requireKey(req).demo) {
+      next();
+      return;
+    }
+    limit(req, res, next);
+  };
+}
+
+/** Per-IP limiter that applies ONLY to demo-key requests (after apiKeyAuth). */
+export function demoIpRateLimit(opts: { max: number; windowMs: number }) {
+  const limit = makeLimiter(opts, (req) => req.ip ?? "unknown");
+  return (req: Request, res: Response, next: NextFunction): void => {
+    if (!requireKey(req).demo) {
+      next();
+      return;
+    }
+    limit(req, res, next);
+  };
 }

gateway/src/ocr/extract.ts

@@ -0,0 +1,402 @@
+/**
+ * gateway/src/ocr/extract.ts — POST /v1/ocr/extract.
+ *
+ * One escrow per page against a model-scoped `ocr.page.extract.<model>.v1`
+ * supplier (config.ocrCapabilityId). Lifecycle mirrors the one-shot chat
+ * path, all inside the per-key mutex: parse → select supplier → preflight →
+ * per-candidate live busy pre-check → submitOcr (escrow → supplier call →
+ * receipt verify) → resolve Submitted ref → Accept (awaited) → record usage
+ * → wallet-health → respond.
+ *
+ * Response (200):
+ *   {
+ *     output_format, content,
+ *     usage: { prompt_tokens, completion_tokens, total_tokens },
+ *     x_vector: {
+ *       receipt, receipt_signature, escrow_ref,
+ *       receipt_labels: ["first-party-brokered"]
+ *     }
+ *   }
+ *
+ * Venue attribution: the caller MAY send
+ *   X-Venue:          venue slug (e.g. "apify"), ≤ 32 chars
+ *   X-Run-User-Hash:  sha256 of the venue-side runner id, 64 hex
+ * Both are logged on a structured line per settled job. The wedge-ledger
+ * emitter consumes these lines; a usage-table venue
+ * column is deliberately deferred to later work so this route stays
+ * schema-additive. payer_class is derived at ledger time from the hash
+ * (known first-party hashes vs the rest), never asserted here.
+ *
+ * Demo keys are rejected: the demo executor is chat-session-shaped and its
+ * wallet holds one escrow per session; OCR needs a funded per-key wallet.
+ */
+
+import { randomUUID } from "crypto";
+import type { Request, Response } from "express";
+import { TxConstructionError } from "@marketplace/shared/tx";
+import {
+  ALLOWED_OCR_MIMES,
+  ALLOWED_OCR_OUTPUT_FORMATS,
+  MAX_OCR_IMAGE_B64_CHARS,
+} from "@marketplace/shared/tx";
+import { SupplierError, type ProgressEvent } from "@marketplace/buyer/sdk";
+import type { GatewayDeps } from "../deps.js";
+import type { GatewayStore } from "../db/store.js";
+import { requireKey } from "../middleware/apiKeyAuth.js";
+import { asyncHandler } from "../middleware/http.js";
+import { selectCandidates, type SupplierCandidate } from "../routing/selectSupplier.js";
+import { preflight } from "../onchain/preflight.js";
+import { resolveSubmittedRef, acceptAndConfirm } from "../onchain/settle.js";
+import { ensureWalletHealthy } from "../walletHealth.js";
+import { badRequest, forbidden, notFound, paymentRequired, toGatewayError } from "../openai/errors.js";
+
+/** Operator-model v0: HAL8 operates both the custodial buyer and the
+ * supplier fleet, so every receipt this route returns is first-party
+ * brokered. Becomes per-supplier metadata when
+ * control-independent supply exists. */
+const RECEIPT_LABELS_V0 = ["first-party-brokered"] as const;
+
+const B64_RE = /^[A-Za-z0-9+/]+={0,2}$/;
+const HEX64_RE = /^[0-9a-fA-F]{64}$/;
+const VENUE_RE = /^[a-z0-9-]{1,32}$/;
+
+/** Bound on the live busy pre-check below - cheap enough that a stalled
+ * supplier never meaningfully delays the request. */
+const SUPPLIER_STATUS_CHECK_TIMEOUT_MS = 2_000;
+
+interface ParsedOcrRequest {
+  image_b64: string;
+  mime: string;
+  output_format: string;
+  /** "" = any model under the capability (single-model ids). */
+  model: string;
+}
+
+function parseOcrRequest(body: unknown): ParsedOcrRequest {
+  if (typeof body !== "object" || body === null) {
+    throw badRequest("invalid_body", "request body must be a JSON object");
+  }
+  const b = body as Record<string, unknown>;
+
+  const image = b.image_b64;
+  if (typeof image !== "string" || image.length === 0) {
+    throw badRequest("invalid_image", "`image_b64` is required and must be a non-empty base64 string");
+  }
+  if (image.length > MAX_OCR_IMAGE_B64_CHARS) {
+    throw badRequest("image_too_large", `\`image_b64\` exceeds ${MAX_OCR_IMAGE_B64_CHARS} characters`);
+  }
+  if (!B64_RE.test(image)) {
+    throw badRequest("invalid_image", "`image_b64` must be plain base64 (no data: prefix, no whitespace)");
+  }
+
+  const mime = b.mime;
+  if (typeof mime !== "string" || !ALLOWED_OCR_MIMES.has(mime)) {
+    throw badRequest("invalid_mime", `\`mime\` must be one of: ${[...ALLOWED_OCR_MIMES].join(", ")}`);
+  }
+
+  const rawFormat = b.output_format ?? "markdown";
+  if (typeof rawFormat !== "string" || !ALLOWED_OCR_OUTPUT_FORMATS.has(rawFormat)) {
+    throw badRequest(
+      "invalid_output_format",
+      `\`output_format\` must be one of: ${[...ALLOWED_OCR_OUTPUT_FORMATS].join(", ")}`,
+    );
+  }
+
+  let model = "";
+  if (b.model !== undefined && b.model !== null) {
+    if (typeof b.model !== "string" || b.model === "") {
+      throw badRequest("invalid_model", "`model` must be a non-empty string when provided");
+    }
+    model = b.model;
+  }
+
+  return { image_b64: image, mime, output_format: rawFormat, model };
+}
+
+/** Sanitized venue-attribution headers. Invalid values drop to "".
+ * Exported for direct unit testing (I3 review finding) — it is the only
+ * gate on what lands in the ocr_settled/ocr_failed structured log lines the
+ * wedge-ledger emitter reads, so it needs coverage independent of the HTTP
+ * route. */
+export function venueAttribution(req: Request): { venue: string; runUserHash: string } {
+  const venueRaw = req.header("X-Venue") ?? "";
+  const hashRaw = req.header("X-Run-User-Hash") ?? "";
+  return {
+    venue: VENUE_RE.test(venueRaw) ? venueRaw : "",
+    runUserHash: HEX64_RE.test(hashRaw) ? hashRaw.toLowerCase() : "",
+  };
+}
+
+/**
+ * Live busy pre-check, run immediately before a candidate's escrow posts.
+ * Mirrors sessions.ts's startChat pre-check: same signal ("working" on the
+ * supplier's own /status), same best-effort posture. Posting an escrow only
+ * to have the supplier reject as busy strands the escrow until reclaim and
+ * burns the tx fee, so this is worth a bounded, cheap HTTP round trip.
+ *
+ * Never a gate: the indexer already filters candidates on advert status, so
+ * a timeout, network error, non-2xx, or malformed body all resolve to "not
+ * busy" (unknown, not blocking) - a briefly-unreachable /status must never
+ * stop the candidate from being attempted.
+ */
+async function isSupplierBusy(
+  endpointUrl: string,
+  fetchFn: typeof globalThis.fetch,
+): Promise<boolean> {
+  try {
+    const res = await fetchFn(`${endpointUrl.replace(/\/+$/, "")}/status`, {
+      signal: AbortSignal.timeout(SUPPLIER_STATUS_CHECK_TIMEOUT_MS),
+    });
+    if (!res.ok) return false;
+    const body = (await res.json()) as unknown;
+    const status = body && typeof body === "object" ? (body as { status?: unknown }).status : undefined;
+    return status === "working";
+  } catch {
+    return false;
+  }
+}
+
+function recordOcrUsage(
+  store: GatewayStore,
+  p: {
+    keyId: string;
+    capabilityId: string;
+    model: string;
+    supplierPkh: string | null;
+    escrowRef: string | null;
+    costLovelace: bigint | null;
+    promptTokens: number;
+    completionTokens: number;
+    status: "completed" | "failed";
+    failureReason?: string;
+  },
+): void {
+  store.insertUsage({
+    id: randomUUID(),
+    key_id: p.keyId,
+    created_at: Date.now(),
+    kind: "ocr",
+    model: p.model,
+    capability_id: p.capabilityId,
+    supplier_pkh: p.supplierPkh,
+    escrow_ref: p.escrowRef,
+    cost_lovelace: p.costLovelace !== null ? p.costLovelace.toString() : null,
+    prompt_tokens: p.promptTokens,
+    completion_tokens: p.completionTokens,
+    status: p.status,
+    failure_reason: p.failureReason ?? null,
+  });
+}
+
+export function makeOcrExtractHandler(deps: GatewayDeps) {
+  return asyncHandler(async (req: Request, res: Response) => {
+    const keyRow = requireKey(req);
+    if (keyRow.demo) {
+      throw forbidden("demo_key_unsupported", "demo keys cannot use /v1/ocr/extract; create a funded API key");
+    }
+    const ctx = deps.registry.getContext(keyRow);
+    const parsed = parseOcrRequest(req.body);
+    const attribution = venueAttribution(req);
+    await ctx.mutex.run(() => runOcrOneShot(deps, keyRow.id, ctx, parsed, attribution, res));
+  });
+}
+
+async function runOcrOneShot(
+  deps: GatewayDeps,
+  keyId: string,
+  ctx: ReturnType<GatewayDeps["registry"]["getContext"]>,
+  parsed: ParsedOcrRequest,
+  attribution: { venue: string; runUserHash: string },
+  res: Response,
+): Promise<void> {
+  const { config, store, chain, fetchFn } = deps;
+  const capabilityId = config.ocrCapabilityId;
+
+  // Hoisted above the try so the catch block can report whatever is known
+  // even when the failure happens AFTER an escrow was posted (e.g. the
+  // supplier call or settle step fails) — these fields are the
+  // reconciliation spine that joins on-chain settlement to platform
+  // billing, and losing them makes paid-but-failed jobs unreconcilable
+  // (I2 review finding).
+  let result: Awaited<ReturnType<typeof ctx.sdk.submitOcr>> | undefined;
+  let used: SupplierCandidate | undefined;
+  let escrowRefStr: string | null = null;
+
+  try {
+    const candidates = await selectCandidates({
+      indexerUrl: config.indexerUrl,
+      model: parsed.model,
+      capabilityId,
+      fetchFn,
+    });
+    if (candidates.length === 0) {
+      throw notFound(
+        "model_not_found",
+        parsed.model === ""
+          ? `no available supplier for capability "${capabilityId}"`
+          : `no available supplier for model "${parsed.model}" under "${capabilityId}"`,
+      );
+    }
+
+    const primary = candidates[0];
+    const pf = await preflight(chain, ctx.walletKey.address, primary);
+    if (!pf.ok) {
+      throw paymentRequired(
+        pf.collateralOk
+          ? "insufficient balance for this request"
+          : "wallet has no pure-AP3X UTxO ≥ 5 AP3X for collateral; deposit a little more AP3X",
+        {
+          x_vector: {
+            required_lovelace: pf.requiredLovelace.toString(),
+            available_lovelace: pf.availableLovelace.toString(),
+            collateral_ok: pf.collateralOk,
+            deposit_address: ctx.walletKey.address,
+          },
+        },
+      );
+    }
+
+    // Submit. Fall back to the next supplier ONLY on a pre-post
+    // TxConstructionError (no escrow was posted). A SupplierError means the
+    // escrow is already posted — do not retry (the sweeper recovers it).
+    //
+    // submitOcr posts the escrow internally and only returns escrowRef when
+    // it resolves. When the supplier call AFTER the post throws
+    // (SupplierError: timeout, network failure, non-2xx, malformed
+    // response, receipt mismatch) the escrow already landed on chain but
+    // escrowRefStr below would otherwise stay null — the OCR upstream has a
+    // 120s budget, so this is likely the most common paid-but-failed shape.
+    // The SDK emits a "progress" event of type "escrow_posted" (carrying
+    // escrow_ref) the instant the tx confirms, so listen for it around each
+    // attempt and capture the ref from the event rather than only from
+    // submitOcr's return value (I2 follow-up).
+    let lastErr: unknown;
+    for (const cand of candidates) {
+      // Live pre-check, run right here (not once at the top of the request):
+      // skip a candidate that is busy right now instead of paying to find
+      // out. See isSupplierBusy's docstring for the exact signal + posture.
+      if (await isSupplierBusy(cand.endpointUrl, fetchFn)) {
+        lastErr = new SupplierError("supplier_busy", {
+          status: 409,
+          message: "supplier is busy with another job",
+        });
+        continue;
+      }
+
+      const onProgress = (ev: unknown): void => {
+        const e = ev as ProgressEvent;
+        if (e.type === "escrow_posted" && e.escrow_ref) escrowRefStr = e.escrow_ref;
+      };
+      ctx.sdk.on("progress", onProgress);
+      try {
+        result = await ctx.sdk.submitOcr({
+          advertRef: cand.advertRef,
+          image_b64: parsed.image_b64,
+          mime: parsed.mime,
+          output_format: parsed.output_format,
+          payment_lovelace: cand.priceLovelace,
+        });
+        used = cand;
+        break;
+      } catch (err) {
+        lastErr = err;
+        if (err instanceof TxConstructionError) continue;
+        throw err;
+      } finally {
+        ctx.sdk.off("progress", onProgress);
+      }
+    }
+    if (!result || !used) {
+      throw toGatewayError(lastErr ?? new Error("no supplier could be engaged"));
+    }
+
+    // Settle: resolve the live Submitted ref, Accept, await confirmation.
+    escrowRefStr = `${result.escrowRef.txHash}#${result.escrowRef.index}`;
+    const submittedRef = await resolveSubmittedRef({
+      indexerUrl: config.indexerUrl,
+      buyerPkh: ctx.walletKey.pubKeyHash,
+      originalRefStr: escrowRefStr,
+      fetchFn,
+    });
+    await acceptAndConfirm(chain, ctx.walletKey, submittedRef);
+
+    recordOcrUsage(store, {
+      keyId,
+      capabilityId,
+      model: used.model,
+      supplierPkh: used.supplierPkh,
+      escrowRef: escrowRefStr,
+      costLovelace: used.priceLovelace,
+      promptTokens: result.usage.prompt_tokens,
+      completionTokens: result.usage.completion_tokens,
+      status: "completed",
+    });
+
+    // Structured settle line for the wedge-ledger emitter. One line per
+    // settled job; venue/run_user_hash empty when the
+    // caller sent none.
+    console.log(JSON.stringify({
+      evt: "ocr_settled",
+      escrow_ref: escrowRefStr,
+      capability_id: capabilityId,
+      model: used.model,
+      supplier_pkh: used.supplierPkh,
+      cost_lovelace: used.priceLovelace.toString(),
+      receipt_labels: RECEIPT_LABELS_V0,
+      venue: attribution.venue,
+      run_user_hash: attribution.runUserHash,
+      ts: new Date().toISOString(),
+    }));
+
+    // Re-shape the wallet ({collateral, working}) so the next request keeps
+    // a valid collateral UTxO. Best-effort; serialized by the mutex.
+    await ensureWalletHealthy(chain, ctx.walletKey).catch((e) =>
+      // eslint-disable-next-line no-console
+      console.error("[gateway] wallet-health after ocr:", e instanceof Error ? e.message : e),
+    );
+
+    res.status(200).json({
+      output_format: result.output_format,
+      content: result.content,
+      usage: result.usage,
+      x_vector: {
+        receipt: result.receipt,
+        receipt_signature: result.receiptSignature,
+        escrow_ref: escrowRefStr,
+        receipt_labels: RECEIPT_LABELS_V0,
+      },
+    });
+  } catch (err) {
+    recordOcrUsage(store, {
+      keyId,
+      capabilityId,
+      model: used?.model ?? parsed.model,
+      supplierPkh: used?.supplierPkh ?? null,
+      escrowRef: escrowRefStr,
+      costLovelace: used?.priceLovelace ?? null,
+      promptTokens: result?.usage.prompt_tokens ?? 0,
+      completionTokens: result?.usage.completion_tokens ?? 0,
+      status: "failed",
+      failureReason: err instanceof Error ? err.message : String(err),
+    });
+
+    // Structured failure line mirroring ocr_settled's shape (same venue /
+    // run_user_hash / escrow_ref fields) so paid-but-failed jobs — an
+    // escrow posted on-chain, then a later step (supplier call, settle)
+    // failed — can still be reconciled against on-chain state instead of
+    // vanishing from the wedge-ledger emitter's input (I2 review finding).
+    console.log(JSON.stringify({
+      evt: "ocr_failed",
+      escrow_ref: escrowRefStr,
+      capability_id: capabilityId,
+      model: used?.model ?? parsed.model,
+      supplier_pkh: used?.supplierPkh ?? null,
+      venue: attribution.venue,
+      run_user_hash: attribution.runUserHash,
+      error: err instanceof Error ? err.message : String(err),
+      ts: new Date().toISOString(),
+    }));
+
+    throw err;
+  }
+}

gateway/src/onchain/preflight.ts

@@ -2,14 +2,20 @@
  * gateway/src/onchain/preflight.ts — balance + collateral pre-flight.
  *
  * The buyer wallet funds the FULL escrow at PostEscrow: price + buyer_bond +
- * supplier_bond (postEscrow.ts:132). Accept/Reclaim are Plutus script-spends
- * that also need a pure-ADA collateral UTxO ≥ 5 ADA (assertCollateralCandidate).
- * So required = price + buyer_bond + supplier_bond + collateral + fee reserve,
- * AND the wallet must currently hold a pure-ADA UTxO ≥ collateral. We check both
- * up front and return a clean 402 instead of failing mid-flight.
+ * supplier_bond, raised to the Submitted-state min-ada floor so long
+ * capability ids (e.g. the 33-byte OCR one) don't get bumped mid-flight —
+ * see escrowLockFloor (packages/shared/.../minAdaFloor.ts) and
+ * postOcrEscrow.ts:205, the SAME floor a POST actually locks. Accept/Reclaim
+ * are Plutus script-spends that also need a pure-AP3X collateral UTxO ≥ 5
+ * AP3X (assertCollateralCandidate). So required = escrowLockFloor(price +
+ * buyer_bond + supplier_bond) + collateral + fee reserve, AND the wallet
+ * must currently hold a pure-AP3X UTxO ≥ collateral. We check both up front
+ * and return a clean 402 instead of failing mid-flight.
  */
 
 import type { ChainProvider, Utxo } from "@marketplace/shared/chain";
+import type { EscrowDatum } from "@marketplace/shared/cbor";
+import { escrowLockFloor } from "@marketplace/shared/tx";
 
 export const COLLATERAL_MIN_LOVELACE = 5_000_000n;
 export const FEE_RESERVE_LOVELACE = 2_000_000n;
@@ -18,7 +24,7 @@ export function totalLovelace(utxos: Utxo[]): bigint {
   return utxos.reduce((sum, u) => sum + u.lovelace, 0n);
 }
 
-/** True iff some UTxO is pure-ADA (no native assets) and ≥ collateral floor. */
+/** True iff some UTxO is pure-AP3X (no native assets) and ≥ collateral floor. */
 export function hasCollateral(utxos: Utxo[]): boolean {
   return utxos.some(
     (u) => u.lovelace >= COLLATERAL_MIN_LOVELACE && Object.keys(u.assets).length === 0,
@@ -32,20 +38,48 @@ export interface PreflightResult {
   collateralOk: boolean;
 }
 
+export interface PreflightCost {
+  capabilityId: string;
+  priceLovelace: bigint;
+  buyerBondLovelace: bigint;
+  supplierBondLovelace: bigint;
+}
+
+/** Placeholder datum used only to size-estimate the Submitted-state min-ada
+ * floor. Every field except capability_id and the cost amounts is a
+ * fixed-size stand-in (real hash/pkh byte widths, a realistic 13-digit
+ * POSIX-ms timestamp) — capability_id's actual length is what drives the
+ * estimate, exactly as in escrowLockFloor's own contract. */
+function representativeEscrowDatum(cost: PreflightCost): EscrowDatum {
+  return {
+    buyer_pkh: "3b2fdaed0398485ff35887fb7d0ad2025348864407854f7831dc6a55",
+    supplier_pkh: "401af418a7e6fb106277292881fcdeaa1cc15b5b112e9c8bdb2d46c9",
+    advert_ref: { txHash: "a6".repeat(32), index: 0 },
+    capability_id: cost.capabilityId,
+    request_spec_hash: "1b".repeat(32),
+    prompt_hash: "23".repeat(32),
+    payment_lovelace: cost.priceLovelace,
+    buyer_bond_lovelace: cost.buyerBondLovelace,
+    supplier_bond_lovelace: cost.supplierBondLovelace,
+    deliver_by: 1_786_105_955_257,
+    posted_at: 1_786_105_625_257,
+    submitted_at: null,
+    result_receipt_hash: null,
+    state: "Open",
+  };
+}
+
 export async function preflight(
   chain: ChainProvider,
   address: string,
-  cost: { priceLovelace: bigint; buyerBondLovelace: bigint; supplierBondLovelace: bigint },
+  cost: PreflightCost,
 ): Promise<PreflightResult> {
   const utxos = await chain.queryUtxosByAddress(address);
   const available = totalLovelace(utxos);
   const collateralOk = hasCollateral(utxos);
-  const required =
-    cost.priceLovelace +
-    cost.buyerBondLovelace +
-    cost.supplierBondLovelace +
-    COLLATERAL_MIN_LOVELACE +
-    FEE_RESERVE_LOVELACE;
+  const economicTotal = cost.priceLovelace + cost.buyerBondLovelace + cost.supplierBondLovelace;
+  const escrowFloor = escrowLockFloor(representativeEscrowDatum(cost), economicTotal);
+  const required = escrowFloor + COLLATERAL_MIN_LOVELACE + FEE_RESERVE_LOVELACE;
   return {
     ok: available >= required && collateralOk,
     availableLovelace: available,

gateway/src/onchain/settle.ts

@@ -22,6 +22,8 @@ export interface EscrowRow {
   utxo_ref: string;
   state: string;
   posted_at: number;
+  /** Validator reclaim floor (indexer serves it; absent on very old rows). */
+  deliver_by?: number;
 }
 
 export async function fetchEscrows(

gateway/src/openai/affinity.ts

@@ -0,0 +1,116 @@
+/**
+ * gateway/src/openai/affinity.ts — transcript-prefix session matching.
+ *
+ * Stateless OpenAI clients resend the FULL message history each call; sessions
+ * are stateful. An incoming request belongs to an open session when that
+ * session's mirrored transcript is a prefix of the incoming messages — the
+ * suffix (delta) is the next turn. Client-authored messages (user/system/tool)
+ * must match byte-exactly, but clients routinely reshape ASSISTANT text when
+ * echoing it back (strip <think> reasoning, re-render whitespace), so
+ * assistant content is compared normalized. Strict (byte-exact) matches are
+ * preferred over lenient ones to disambiguate concurrent threads that share a
+ * user-message prefix. The mirror itself is NEVER rewritten with client text:
+ * the delta is forwarded verbatim and mirrored only after a successful turn,
+ * so the gateway mirror and the supplier transcript stay hash-identical for
+ * the receipt response_hash.
+ *
+ * Pure module — no I/O — so the matching rules are unit-testable.
+ */
+
+import { chatMessagesEquivalent, toolCallsEqual, type ChatMessage } from "@marketplace/shared/tx";
+
+export interface AffinityCandidate {
+  sessionId: string;
+  model: string;
+  mirror: ChatMessage[];
+  lastUsedAt: number;
+}
+
+export interface AffinityMatch {
+  sessionId: string;
+  /** Messages the session has not seen yet — forwarded as the next turn. */
+  delta: ChatMessage[];
+}
+
+/** Canonicalize assistant text the way clients mangle it when re-echoing
+ * history: drop <think>…</think> reasoning blocks, collapse whitespace runs,
+ * trim. Exported for tests. */
+export function normalizeAssistantContent(content: string): string {
+  return content
+    .replace(/<think>[\s\S]*?<\/think>/gi, "")
+    .replace(/\s+/g, " ")
+    .trim();
+}
+
+/** Lenient equality: strict for client-authored roles (the client sends the
+ * same bytes both times); assistant content is compared normalized while
+ * assistant tool_calls stay strict (ids anchor tool-result turns). */
+function chatMessagesLenientEquivalent(a: ChatMessage, b: ChatMessage): boolean {
+  if (a.role !== b.role) return false;
+  if (a.role !== "assistant") return chatMessagesEquivalent(a, b);
+  return (
+    normalizeAssistantContent(a.content) === normalizeAssistantContent(b.content) &&
+    toolCallsEqual(a.tool_calls, b.tool_calls)
+  );
+}
+
+type PrefixQuality = "strict" | "lenient";
+
+function classifyPrefix(mirror: ChatMessage[], incoming: ChatMessage[]): PrefixQuality | null {
+  if (mirror.length >= incoming.length) return null;
+  let strict = true;
+  for (let i = 0; i < mirror.length; i++) {
+    if (chatMessagesEquivalent(mirror[i], incoming[i])) continue;
+    if (!chatMessagesLenientEquivalent(mirror[i], incoming[i])) return null;
+    strict = false;
+  }
+  return strict ? "strict" : "lenient";
+}
+
+function deltaRespondable(delta: ChatMessage[]): boolean {
+  const last = delta[delta.length - 1];
+  return last !== undefined && (last.role === "user" || last.role === "tool");
+}
+
+/** Drop trailing assistant/system messages so the history ends with a
+ * respondable message (regenerate semantics for clients that send a trailing
+ * assistant turn). Empty result = nothing to answer. */
+export function trimToRespondable(messages: ChatMessage[]): ChatMessage[] {
+  let end = messages.length;
+  while (end > 0 && messages[end - 1].role !== "user" && messages[end - 1].role !== "tool") end--;
+  return messages.slice(0, end);
+}
+
+/**
+ * Pick the open session whose mirror best matches the incoming history.
+ * Strict prefixes beat lenient ones; then longest mirror wins (most
+ * conversation reuse); ties go to the most recently used session. Returns
+ * null when nothing matches — caller opens a new session and replays the full
+ * history. Note `incoming.length <= mirror.length` never matches (a client
+ * retry of an already-answered request is a miss by design: the session
+ * transcript must stay append-only).
+ */
+export function matchSessionPrefix(
+  candidates: AffinityCandidate[],
+  incoming: ChatMessage[],
+  model: string,
+): AffinityMatch | null {
+  let best: { cand: AffinityCandidate; quality: PrefixQuality } | null = null;
+  for (const cand of candidates) {
+    if (cand.model !== model) continue;
+    const quality = classifyPrefix(cand.mirror, incoming);
+    if (quality === null) continue;
+    if (!deltaRespondable(incoming.slice(cand.mirror.length))) continue;
+    if (
+      best === null ||
+      (quality === "strict" && best.quality === "lenient") ||
+      (quality === best.quality &&
+        (cand.mirror.length > best.cand.mirror.length ||
+          (cand.mirror.length === best.cand.mirror.length && cand.lastUsedAt > best.cand.lastUsedAt)))
+    ) {
+      best = { cand, quality };
+    }
+  }
+  if (!best) return null;
+  return { sessionId: best.cand.sessionId, delta: incoming.slice(best.cand.mirror.length) };
+}

gateway/src/openai/chatCompletions.ts

@@ -21,6 +21,7 @@ import { preflight } from "../onchain/preflight.js";
 import { resolveSubmittedRef, acceptAndConfirm } from "../onchain/settle.js";
 import { ensureWalletHealthy } from "../walletHealth.js";
 import { parseChatRequest, type ParsedChatRequest } from "./validate.js";
+import { runDemoChat } from "./demoChat.js";
 import {
   genId,
   buildChatCompletion,
@@ -67,8 +68,17 @@ function recordUsage(
 export function makeChatCompletionsHandler(deps: GatewayDeps) {
   return asyncHandler(async (req: Request, res: Response) => {
     const keyRow = requireKey(req);
-    const parsed = parseChatRequest(req.body);
     const ctx = deps.registry.getContext(keyRow);
+    if (keyRow.demo) {
+      // Demo keys run session-backed with tools passthrough, and never fall
+      // back to one-shot (its settle path assumes ≤1 in-flight escrow per
+      // wallet; the demo wallet holds one escrow per concurrent session).
+      // Not under the key mutex — runDemoChat takes it only for session opens.
+      const parsed = parseChatRequest(req.body, { allowTools: true });
+      await runDemoChat(deps, keyRow, ctx, parsed, res);
+      return;
+    }
+    const parsed = parseChatRequest(req.body);
     await ctx.mutex.run(() => runOneShot(deps, keyRow, ctx, parsed, res));
   });
 }
@@ -98,7 +108,7 @@ async function runOneShot(
     throw paymentRequired(
       pf.collateralOk
         ? "insufficient balance for this request"
-        : "wallet has no pure-ADA UTxO ≥ 5 ADA for collateral; deposit a little more ADA",
+        : "wallet has no pure-AP3X UTxO ≥ 5 AP3X for collateral; deposit a little more AP3X",
       {
         x_vector: {
           required_lovelace: pf.requiredLovelace.toString(),

gateway/src/openai/demoChat.ts

@@ -0,0 +1,413 @@
+/**
+ * gateway/src/openai/demoChat.ts — session-backed /chat/completions for the
+ * shared demo key.
+ *
+ * Demo requests keep the standard OpenAI wire shape (tools included) but run
+ * against real escrow-per-session marketplace conversations: transcript-prefix
+ * affinity maps each stateless request onto an open session and forwards only
+ * the unseen delta; a miss transparently opens a new session (escrow post —
+ * the only chain work, under the key mutex) and replays the full history as
+ * the first turn. Supplier idle-closes are healed by reopening; the sweeper
+ * settles the abandoned escrow.
+ *
+ * Demo keys NEVER fall back to the one-shot path: resolveSubmittedRef's
+ * lone-Submitted-row fallback assumes ≤1 in-flight escrow per wallet, and the
+ * demo wallet holds one escrow per concurrent session.
+ */
+
+import { randomUUID } from "crypto";
+import type { Response } from "express";
+import type { ChatMessage, ToolCall } from "@marketplace/shared/tx";
+import type { GatewayDeps } from "../deps.js";
+import type { ApiKeyRow, SessionRow } from "../db/store.js";
+import type { KeyContext } from "../sdk/registry.js";
+import { transcripts, getSessionLock, dropSessionState } from "./transcripts.js";
+import { matchSessionPrefix, trimToRespondable, type AffinityCandidate } from "./affinity.js";
+import {
+  CAPABILITY,
+  SessionGoneError,
+  openSessionCore,
+  streamSupplierTurn,
+  type TurnResult,
+} from "./sessions.js";
+import type { ParsedChatRequest } from "./validate.js";
+import { genId, buildChatCompletion, buildChunk, sseData, SSE_DONE, type Usage } from "./shapes.js";
+import { GatewayError, badRequest, toGatewayError, toErrorBody } from "./errors.js";
+
+interface DemoSessionMeta {
+  keyId: string;
+  model: string;
+  lastUsedAt: number;
+}
+
+/** Open sessions owned by the demo executor (sessionId → meta). Mirrors live
+ * in transcripts.ts; this map scopes affinity to demo-opened sessions. */
+const demoSessions = new Map<string, DemoSessionMeta>();
+
+/** Forget a session (closed/settled elsewhere, e.g. by the sweeper). */
+export function dropDemoSession(sessionId: string): void {
+  demoSessions.delete(sessionId);
+}
+
+function candidatesFor(keyId: string, model: string): AffinityCandidate[] {
+  const out: AffinityCandidate[] = [];
+  for (const [sessionId, meta] of demoSessions) {
+    if (meta.keyId !== keyId || meta.model !== model) continue;
+    const mirror = transcripts.get(sessionId);
+    if (!mirror) continue;
+    out.push({ sessionId, model: meta.model, mirror, lastUsedAt: meta.lastUsedAt });
+  }
+  return out;
+}
+
+/** Record the session's cost when the demo executor closes its row. Full
+ * mode: the flat session fee (supplier Submitted; the sweeper Accepts the
+ * escrow afterwards). Ticket mode: 0 — nothing settles, the sweeper reclaims
+ * the escrow and the wallet is made whole minus fees. */
+function recordSessionCost(deps: GatewayDeps, session: SessionRow): void {
+  deps.store.insertUsage({
+    id: randomUUID(),
+    key_id: session.key_id,
+    created_at: Date.now(),
+    kind: "chat_session",
+    model: session.model,
+    capability_id: CAPABILITY,
+    supplier_pkh: session.supplier_pkh,
+    escrow_ref: session.escrow_ref,
+    cost_lovelace: deps.config.chatSettleMode === "ticket" ? "0" : session.price_lovelace,
+    prompt_tokens: 0,
+    completion_tokens: 0,
+    status: "completed",
+    failure_reason: null,
+  });
+}
+
+function markSessionGone(deps: GatewayDeps, session: SessionRow): void {
+  deps.store.setSessionState(session.id, "closed", Date.now());
+  dropSessionState(session.id);
+  dropDemoSession(session.id);
+  recordSessionCost(deps, session);
+}
+
+/** Ask the supplier to settle (Submit) the session. NO gateway chain work —
+ * the sweeper Accepts the resulting Submitted escrow. The supplier releases
+ * its single slot as part of ending, so a truthy return means the supplier is
+ * free again. Best-effort: on failure the sweeper reclaims later. */
+async function requestSupplierEnd(deps: GatewayDeps, session: SessionRow): Promise<boolean> {
+  try {
+    const res = await deps.fetchFn(`${session.supplier_base_url.replace(/\/+$/, "")}/v1/chat/end`, {
+      method: "POST",
+      headers: { "content-type": "application/json", "X-Escrow-Ref": session.escrow_ref },
+      body: JSON.stringify({}),
+      signal: AbortSignal.timeout(90_000),
+    });
+    await res.text().catch(() => "");
+    return res.ok;
+  } catch {
+    return false;
+  }
+}
+
+/** Close one demo session (evicted for capacity, or idle-janitored). Runs
+ * under the session lock so an in-flight turn finishes first; takes NO key
+ * mutex (no chain work here — see requestSupplierEnd). Returns the freed
+ * supplier_pkh, or null when the session was already gone or (janitor) was
+ * used again while we waited for the lock. */
+async function closeDemoSession(
+  deps: GatewayDeps,
+  sessionId: string,
+  opts: { reason: "evicted" | "idle"; idleCutoffMs?: number; nowMs?: number },
+): Promise<string | null> {
+  if (!demoSessions.has(sessionId)) return null;
+  return getSessionLock(sessionId).run(async () => {
+    const meta = demoSessions.get(sessionId);
+    if (!meta) return null; // lost a race with another closer
+    // Re-check idleness inside the lock: a turn that won the lock first has
+    // bumped lastUsedAt past the sweep's reference clock.
+    if (opts.idleCutoffMs !== undefined && (opts.nowMs ?? Date.now()) - meta.lastUsedAt < opts.idleCutoffMs) {
+      return null;
+    }
+    const session = deps.store.getSession(sessionId);
+    if (!session || session.state !== "open") {
+      dropDemoSession(sessionId);
+      return null;
+    }
+    await requestSupplierEnd(deps, session);
+    markSessionGone(deps, session);
+    // eslint-disable-next-line no-console
+    console.log(`[gateway:demo] closed session ${sessionId} (${opts.reason}), freed supplier ${session.supplier_pkh}`);
+    return session.supplier_pkh;
+  });
+}
+
+/** Least-recently-used open demo session for (key, model), excluding prior
+ * victims. Same-model scoping doubles as a guard: a bogus model name has no
+ * demo sessions, so eviction can never loop on a genuine model_not_found. */
+function pickLruDemoSession(keyId: string, model: string, exclude: ReadonlySet<string>): string | null {
+  let lruId: string | null = null;
+  let lruAt = Infinity;
+  for (const [sessionId, meta] of demoSessions) {
+    if (meta.keyId !== keyId || meta.model !== model || exclude.has(sessionId)) continue;
+    if (meta.lastUsedAt < lruAt) {
+      lruAt = meta.lastUsedAt;
+      lruId = sessionId;
+    }
+  }
+  return lruId;
+}
+
+/** Close demo sessions idle past deps.config.demoSessionIdleMs so abandoned
+ * conversations (client gone, title-gen one-offs) release their single-slot
+ * suppliers. Called from the sweeper tick; sequential to keep supplier
+ * /v1/chat/end calls orderly. */
+export async function sweepIdleDemoSessions(deps: GatewayDeps, nowMs = Date.now()): Promise<void> {
+  const idleMs = deps.config.demoSessionIdleMs;
+  for (const [sessionId, meta] of [...demoSessions]) {
+    if (nowMs - meta.lastUsedAt < idleMs) continue;
+    try {
+      await closeDemoSession(deps, sessionId, { reason: "idle", idleCutoffMs: idleMs, nowMs });
+    } catch (err) {
+      // eslint-disable-next-line no-console
+      console.error(`[gateway:demo] idle close of ${sessionId} failed:`, err instanceof Error ? err.message : err);
+    }
+  }
+}
+
+const MAX_EVICTIONS = 2;
+
+function isNoSupplierError(err: unknown): boolean {
+  return err instanceof GatewayError && (err.code === "model_not_found" || err.code === "overloaded");
+}
+
+/** Open a session, evicting up to MAX_EVICTIONS LRU demo sessions when every
+ * supplier is pinned. Eviction frees the supplier immediately (its slot
+ * releases at Submit) but the indexer's status poll lags, so the retry passes
+ * the freed pkhs through selectCandidates' ignoreStatusFor. Lock ordering is
+ * safe: eviction holds only the session lock, the open holds only the key
+ * mutex — never nested. */
+async function openSessionWithEviction(
+  deps: GatewayDeps,
+  keyRow: ApiKeyRow,
+  ctx: KeyContext,
+  model: string,
+): Promise<SessionRow> {
+  const freedPkhs = new Set<string>();
+  const evicted = new Set<string>();
+  for (let attempt = 0; ; attempt++) {
+    try {
+      return await ctx.mutex.run(() =>
+        openSessionCore(deps, keyRow, ctx, model, freedPkhs.size > 0 ? { ignoreStatusFor: freedPkhs } : undefined),
+      );
+    } catch (err) {
+      if (attempt >= MAX_EVICTIONS || !isNoSupplierError(err)) throw err;
+      const victimId = pickLruDemoSession(keyRow.id, model, evicted);
+      if (!victimId) throw err;
+      evicted.add(victimId);
+      const pkh = await closeDemoSession(deps, victimId, { reason: "evicted" });
+      if (!pkh) throw err; // nothing actually freed — don't loop
+      freedPkhs.add(pkh);
+    }
+  }
+}
+
+type LockedTurnOutcome =
+  | { kind: "done"; result: TurnResult }
+  | { kind: "retry" } // candidate no longer valid — re-run the match
+  | { kind: "gone" }; // supplier lost the session — open a fresh one
+
+/** Streaming sink shared across turn attempts on one response. */
+interface StreamSink {
+  onToken: (token: string) => void;
+  emitToolCalls: (toolCalls: ToolCall[]) => void;
+}
+
+export async function runDemoChat(
+  deps: GatewayDeps,
+  keyRow: ApiKeyRow,
+  ctx: KeyContext,
+  parsed: ParsedChatRequest,
+  res: Response,
+): Promise<void> {
+  // Clients sometimes send histories ending in an assistant (continue/retry
+  // flows) or system message; the chat.v1 supplier can only respond to a
+  // user/tool-terminated turn. Trim trailing unrespondable messages
+  // (regenerate semantics) and reject — before any chain work — only when
+  // nothing respondable remains.
+  const trimmed = trimToRespondable(parsed.messages);
+  if (trimmed.length === 0) {
+    throw badRequest("invalid_messages", "no user or tool message to respond to");
+  }
+  const effective: ParsedChatRequest =
+    trimmed.length === parsed.messages.length ? parsed : { ...parsed, messages: trimmed };
+
+  const id = genId();
+  let keepalive: ReturnType<typeof setInterval> | undefined;
+  let roleSent = false;
+  if (parsed.stream) {
+    res.status(200);
+    res.setHeader("Content-Type", "text/event-stream");
+    res.setHeader("Cache-Control", "no-cache, no-transform");
+    res.setHeader("Connection", "keep-alive");
+    res.setHeader("X-Accel-Buffering", "no");
+    res.write(": connected\n\n");
+    keepalive = setInterval(() => {
+      try {
+        res.write(": keepalive\n\n");
+      } catch {
+        /* client gone */
+      }
+    }, 10_000);
+  }
+
+  const ensureRole = (): void => {
+    if (roleSent) return;
+    roleSent = true;
+    res.write(sseData(buildChunk({ id, model: parsed.model, delta: { role: "assistant" }, finishReason: null })));
+  };
+  const sink: StreamSink = {
+    onToken: (token) => {
+      if (!parsed.stream) return;
+      ensureRole();
+      res.write(sseData(buildChunk({ id, model: parsed.model, delta: { content: token }, finishReason: null })));
+    },
+    emitToolCalls: (toolCalls) => {
+      if (!parsed.stream) return;
+      ensureRole();
+      res.write(
+        sseData(
+          buildChunk({
+            id,
+            model: parsed.model,
+            delta: { tool_calls: toolCalls.map((tc, index) => ({ index, ...tc })) },
+            finishReason: null,
+          }),
+        ),
+      );
+    },
+  };
+
+  try {
+    const result = await runWithAffinity(deps, keyRow, ctx, effective, sink);
+    if (keepalive) clearInterval(keepalive);
+
+    const usage: Usage = {
+      prompt_tokens: result.usage.prompt_tokens,
+      completion_tokens: result.usage.completion_tokens,
+      total_tokens: result.usage.prompt_tokens + result.usage.completion_tokens,
+    };
+    if (parsed.stream) {
+      if (result.assistantMessage.tool_calls?.length) sink.emitToolCalls(result.assistantMessage.tool_calls);
+      ensureRole();
+      res.write(sseData(buildChunk({ id, model: parsed.model, delta: {}, finishReason: result.finishReason, usage })));
+      res.write(SSE_DONE);
+      res.end();
+    } else {
+      res.status(200).json(
+        buildChatCompletion({
+          id,
+          model: parsed.model,
+          content: result.assistantMessage.content,
+          toolCalls: result.assistantMessage.tool_calls,
+          finishReason: result.finishReason,
+          usage,
+        }),
+      );
+    }
+  } catch (err) {
+    if (keepalive) clearInterval(keepalive);
+    if (parsed.stream && res.headersSent) {
+      try {
+        res.write(sseData(toErrorBody(toGatewayError(err))));
+      } catch {
+        /* client gone */
+      }
+      res.end();
+      return;
+    }
+    throw err;
+  }
+}
+
+async function runWithAffinity(
+  deps: GatewayDeps,
+  keyRow: ApiKeyRow,
+  ctx: KeyContext,
+  parsed: ParsedChatRequest,
+  sink: StreamSink,
+): Promise<TurnResult> {
+  // Affinity: try matching sessions until one serves the turn. The match runs
+  // synchronously (no await between match and lock acquire, so the candidate
+  // set can't shift underneath us); the prefix is re-checked inside the lock
+  // because a concurrent request may have advanced the session first.
+  for (let attempt = 0; attempt < 3; attempt++) {
+    const match = matchSessionPrefix(candidatesFor(keyRow.id, parsed.model), parsed.messages, parsed.model);
+    if (!match) break;
+
+    const outcome: LockedTurnOutcome = await getSessionLock(match.sessionId).run(async () => {
+      const meta = demoSessions.get(match.sessionId);
+      const mirror = transcripts.get(match.sessionId);
+      const session = deps.store.getSession(match.sessionId);
+      if (!meta || !mirror || !session || session.state !== "open") {
+        dropDemoSession(match.sessionId);
+        return { kind: "retry" };
+      }
+      const recheck = matchSessionPrefix(
+        [{ sessionId: session.id, model: meta.model, mirror, lastUsedAt: meta.lastUsedAt }],
+        parsed.messages,
+        parsed.model,
+      );
+      if (!recheck) return { kind: "retry" };
+
+      try {
+        const result = await streamSupplierTurn(
+          deps,
+          session,
+          { messages: recheck.delta, tools: parsed.tools, tool_choice: parsed.toolChoice },
+          sink.onToken,
+        );
+        meta.lastUsedAt = Date.now();
+        return { kind: "done", result };
+      } catch (err) {
+        if (err instanceof SessionGoneError) {
+          // Supplier idle-closed (or restarted). Close our row, record the
+          // fee (the escrow settles via the sweeper) and reopen fresh.
+          markSessionGone(deps, session);
+          return { kind: "gone" };
+        }
+        throw err;
+      }
+    });
+
+    if (outcome.kind === "done") return outcome.result;
+    if (outcome.kind === "gone") break;
+    // retry: candidate vanished/advanced — re-match against current state.
+  }
+
+  // Miss: open a new session (chain work → key mutex; evicts an LRU demo
+  // session when every supplier is pinned) and replay the full history as its
+  // first turn.
+  const session = await openSessionWithEviction(deps, keyRow, ctx, parsed.model);
+  demoSessions.set(session.id, { keyId: keyRow.id, model: parsed.model, lastUsedAt: Date.now() });
+  return getSessionLock(session.id).run(async () => {
+    try {
+      const result = await streamSupplierTurn(
+        deps,
+        session,
+        { messages: parsed.messages, tools: parsed.tools, tool_choice: parsed.toolChoice },
+        sink.onToken,
+      );
+      // Refresh recency after the (possibly long) first stream so this live
+      // thread isn't the LRU/janitor's first victim.
+      const meta = demoSessions.get(session.id);
+      if (meta) meta.lastUsedAt = Date.now();
+      return result;
+    } catch (err) {
+      if (err instanceof SessionGoneError) {
+        // Freshly opened yet already gone (supplier crashed mid-claim).
+        // Surface as an upstream error rather than looping.
+        markSessionGone(deps, session);
+      }
+      throw err;
+    }
+  });
+}

gateway/src/openai/errors.ts

@@ -36,6 +36,8 @@ export const unauthorized = (message = "invalid API key"): GatewayError =>
   new GatewayError(401, "authentication_error", "invalid_api_key", message);
 export const paymentRequired = (message: string, extra?: Record<string, unknown>): GatewayError =>
   new GatewayError(402, "invalid_request_error", "insufficient_funds", message, extra);
+export const forbidden = (code: string, message: string): GatewayError =>
+  new GatewayError(403, "invalid_request_error", code, message);
 export const notFound = (code: string, message: string): GatewayError =>
   new GatewayError(404, "invalid_request_error", code, message);
 export const rateLimited = (message: string): GatewayError =>

gateway/src/openai/models.ts

@@ -7,12 +7,20 @@
 import type { Request, Response } from "express";
 import type { GatewayDeps } from "../deps.js";
 import { asyncHandler } from "../middleware/http.js";
+import { requireKey } from "../middleware/apiKeyAuth.js";
 import { listModels } from "../routing/selectSupplier.js";
 import { buildModelsList } from "./shapes.js";
+import { CAPABILITY as CHAT_CAPABILITY } from "./sessions.js";
 
 export function makeModelsHandler(deps: GatewayDeps) {
-  return asyncHandler(async (_req: Request, res: Response) => {
-    const models = await listModels({ indexerUrl: deps.config.indexerUrl, fetchFn: deps.fetchFn });
+  return asyncHandler(async (req: Request, res: Response) => {
+    const keyRow = requireKey(req);
+    // Demo keys only reach chat-capable suppliers (session-backed executor).
+    const models = await listModels({
+      indexerUrl: deps.config.indexerUrl,
+      capabilityId: keyRow.demo ? CHAT_CAPABILITY : undefined,
+      fetchFn: deps.fetchFn,
+    });
     res.status(200).json(buildModelsList(models));
   });
 }

gateway/src/openai/sessions.ts

@@ -15,8 +15,9 @@
 
 import { randomUUID } from "crypto";
 import type { Request, Response } from "express";
-import { TxConstructionError, type ChatMessage } from "@marketplace/shared/tx";
+import { TxConstructionError, normalizeChatMessage, type ChatMessage } from "@marketplace/shared/tx";
 import { SupplierError } from "@marketplace/buyer/sdk";
+import { transcripts, getSessionLock, dropSessionState } from "./transcripts.js";
 import type { GatewayDeps } from "../deps.js";
 import type { ApiKeyRow, GatewayStore, SessionRow } from "../db/store.js";
 import type { KeyContext } from "../sdk/registry.js";
@@ -36,10 +37,15 @@ import {
   type Usage,
 } from "./shapes.js";
 
-const CAPABILITY = "llm.chat.v1";
+export const CAPABILITY = "llm.chat.v1";
 
-/** In-memory per-session transcript mirror (sessionId → messages). */
-const transcripts = new Map<string, ChatMessage[]>();
+/** Thrown when the supplier no longer knows the session (idle-closed/restart). */
+export class SessionGoneError extends Error {
+  constructor(sessionId: string) {
+    super(`supplier no longer has session ${sessionId}`);
+    this.name = "SessionGoneError";
+  }
+}
 
 function refStr(ref: { txHash: string; index: number }): string {
   return `${ref.txHash}#${ref.index}`;
@@ -56,22 +62,29 @@ export function makeOpenSessionHandler(deps: GatewayDeps) {
       throw badRequest("invalid_model", "`model` is required");
     }
     const ctx = deps.registry.getContext(keyRow);
-    await ctx.mutex.run(() => openSession(deps, keyRow, ctx, model, res));
+    const session = await ctx.mutex.run(() => openSessionCore(deps, keyRow, ctx, model));
+    res.status(200).json({ id: session.id, object: "chat.session", model, created: nowSec() });
   });
 }
 
-async function openSession(
+/**
+ * Select a supplier, preflight the wallet, post+claim the session escrow and
+ * persist the SessionRow. Chain work — the CALLER must hold the key mutex
+ * (concurrent PostEscrow from one wallet double-spends the same UTxO).
+ */
+export async function openSessionCore(
   deps: GatewayDeps,
   keyRow: ApiKeyRow,
   ctx: KeyContext,
   model: string,
-  res: Response,
-): Promise<void> {
+  opts?: { ignoreStatusFor?: ReadonlySet<string> },
+): Promise<SessionRow> {
   const candidates = await selectCandidates({
     indexerUrl: deps.config.indexerUrl,
     model,
     capabilityId: CAPABILITY,
     fetchFn: deps.fetchFn,
+    ignoreStatusFor: opts?.ignoreStatusFor,
   });
   if (candidates.length === 0) {
     throw notFound("model_not_found", `no available chat (llm.chat.v1) supplier for model "${model}"`);
@@ -80,7 +93,7 @@ async function openSession(
   const pf = await preflight(deps.chain, ctx.walletKey.address, primary);
   if (!pf.ok) {
     throw paymentRequired(
-      pf.collateralOk ? "insufficient balance to open a session" : "wallet has no ≥5 ADA pure-ADA collateral UTxO",
+      pf.collateralOk ? "insufficient balance to open a session" : "wallet has no ≥5 AP3X pure-AP3X collateral UTxO",
       {
         x_vector: {
           required_lovelace: pf.requiredLovelace.toString(),
@@ -112,6 +125,16 @@ async function openSession(
   }
   if (!started || !used) throw toGatewayError(lastErr ?? new Error("no chat supplier could be engaged"));
 
+  // Deployment sanity: supplier and gateway must agree on the settle mode.
+  // Close behavior is driven per-session by the supplier's /end response, so
+  // a mismatch (rollout window) still settles correctly — just warn.
+  if ((started.settleMode === "ticket") !== (deps.config.chatSettleMode === "ticket")) {
+    // eslint-disable-next-line no-console
+    console.warn(
+      `[gateway] settle-mode mismatch: supplier ${used.supplierPkh.slice(0, 8)} ran "${started.settleMode}" but gateway is configured "${deps.config.chatSettleMode}"`,
+    );
+  }
+
   const sessionId = randomUUID();
   deps.store.insertSession({
     id: sessionId,
@@ -127,7 +150,9 @@ async function openSession(
   });
   transcripts.set(sessionId, []);
 
-  res.status(200).json({ id: sessionId, object: "chat.session", model, created: nowSec() });
+  const session = deps.store.getSession(sessionId);
+  if (!session) throw new Error(`session ${sessionId} vanished after insert`);
+  return session;
 }
 
 // ─── messages (one streamed turn) ────────────────────────────────────────────
@@ -144,8 +169,12 @@ export function makeSessionMessageHandler(deps: GatewayDeps) {
     const content = resolveTurnContent(session.id, body);
     if (content === "") throw badRequest("content_required", "no user content in this turn");
 
-    const ctx = deps.registry.getContext(keyRow);
-    await ctx.mutex.run(() => streamTurn(deps, session, content, stream, res));
+    // Turns are chain-free (one HTTP POST to the supplier + SSE relay), so the
+    // per-key mutex — which exists to serialize wallet UTxO spends — is NOT
+    // held here. What needs serialization is this session's transcript:
+    // concurrent turns would interleave supplier appends and corrupt both the
+    // mirror and the receipt hash. Hence a per-session lock.
+    await getSessionLock(session.id).run(() => streamTurn(deps, session, content, stream, res));
   });
 }
 
@@ -176,6 +205,88 @@ function foldSystem(sessionId: string, content: string, body: Record<string, unk
   return content;
 }
 
+/** One turn's payload to the supplier: legacy single user string, or a
+ * multi-message delta (e.g. tool results) with optional tools passthrough. */
+export type TurnPayload =
+  | { content: string }
+  | { messages: ChatMessage[]; tools?: unknown[]; tool_choice?: unknown };
+
+export interface TurnResult {
+  assistantMessage: ChatMessage;
+  finishReason: "stop" | "tool_calls";
+  usage: { prompt_tokens: number; completion_tokens: number };
+}
+
+/**
+ * Drive one off-chain turn against the session's supplier and mirror the
+ * result into the session transcript. Caller must hold the session lock.
+ * Throws SessionGoneError when the supplier no longer knows the session
+ * (idle-closed or restarted) so callers can transparently reopen.
+ */
+export async function streamSupplierTurn(
+  deps: GatewayDeps,
+  session: SessionRow,
+  payload: TurnPayload,
+  onToken?: (token: string) => void,
+): Promise<TurnResult> {
+  const upstream = await deps.fetchFn(`${session.supplier_base_url.replace(/\/+$/, "")}/v1/chat/message`, {
+    method: "POST",
+    headers: { "content-type": "application/json", accept: "text/event-stream", "X-Escrow-Ref": session.escrow_ref },
+    body: JSON.stringify(payload),
+  });
+  if (upstream.status === 404) {
+    await upstream.text().catch(() => "");
+    throw new SessionGoneError(session.id);
+  }
+  if (!upstream.ok || !upstream.body) {
+    const detail = await upstream.text().catch(() => "");
+    throw toGatewayError(new SupplierError("supplier_http_error", { status: upstream.status, message: detail.slice(0, 200) }));
+  }
+
+  let assistant = "";
+  let errored: string | undefined;
+  let doneMessage: ChatMessage | null = null;
+  let doneFinish: string | undefined;
+  let usage = { prompt_tokens: 0, completion_tokens: 0 };
+  await readSupplierSse(upstream.body as ReadableStream<Uint8Array>, (frame) => {
+    if (frame.type === "token" && typeof frame.value === "string") {
+      assistant += frame.value;
+      onToken?.(frame.value);
+    } else if (frame.type === "done") {
+      // Upgraded suppliers echo the final assistant message verbatim; mirror
+      // that exact object so both transcripts stay hash-identical.
+      doneMessage = normalizeChatMessage(frame.message);
+      if (typeof frame.finish_reason === "string") doneFinish = frame.finish_reason;
+      if (frame.usage && typeof frame.usage === "object") {
+        const u = frame.usage as Record<string, unknown>;
+        usage = {
+          prompt_tokens: typeof u.prompt_tokens === "number" ? u.prompt_tokens : 0,
+          completion_tokens: typeof u.completion_tokens === "number" ? u.completion_tokens : 0,
+        };
+      }
+    } else if (frame.type === "error") {
+      errored = typeof frame.message === "string" ? frame.message : "supplier stream error";
+    }
+  });
+
+  if (errored !== undefined) {
+    throw toGatewayError(new SupplierError("upstream_error", { message: errored }));
+  }
+
+  const assistantMessage: ChatMessage = doneMessage ?? { role: "assistant", content: assistant };
+  const finishReason: TurnResult["finishReason"] =
+    doneFinish === "tool_calls" || (assistantMessage.tool_calls?.length ?? 0) > 0 ? "tool_calls" : "stop";
+
+  // Mirror the turn for endChat's response_hash verification / affinity matching.
+  const mirror = transcripts.get(session.id) ?? [];
+  if ("content" in payload) mirror.push({ role: "user", content: payload.content });
+  else mirror.push(...payload.messages);
+  mirror.push(assistantMessage);
+  transcripts.set(session.id, mirror);
+
+  return { assistantMessage, finishReason, usage };
+}
+
 async function streamTurn(
   deps: GatewayDeps,
   session: SessionRow,
@@ -184,15 +295,6 @@ async function streamTurn(
   res: Response,
 ): Promise<void> {
   const id = genId();
-  const upstream = await deps.fetchFn(`${session.supplier_base_url.replace(/\/+$/, "")}/v1/chat/message`, {
-    method: "POST",
-    headers: { "content-type": "application/json", accept: "text/event-stream", "X-Escrow-Ref": session.escrow_ref },
-    body: JSON.stringify({ content }),
-  });
-  if (!upstream.ok || !upstream.body) {
-    const detail = await upstream.text().catch(() => "");
-    throw toGatewayError(new SupplierError("supplier_http_error", { status: upstream.status, message: detail.slice(0, 200) }));
-  }
 
   if (stream) {
     res.status(200);
@@ -203,46 +305,43 @@ async function streamTurn(
     res.write(sseData(buildChunk({ id, model: session.model, delta: { role: "assistant" }, finishReason: null })));
   }
 
-  let assistant = "";
-  let errored: string | undefined;
-  await readSupplierSse(upstream.body as ReadableStream<Uint8Array>, (frame) => {
-    if (frame.type === "token" && typeof frame.value === "string") {
-      assistant += frame.value;
-      if (stream) res.write(sseData(buildChunk({ id, model: session.model, delta: { content: frame.value }, finishReason: null })));
-    } else if (frame.type === "error") {
-      errored = typeof frame.message === "string" ? frame.message : "supplier stream error";
+  let result: TurnResult;
+  try {
+    result = await streamSupplierTurn(deps, session, { content }, (token) => {
+      if (stream) res.write(sseData(buildChunk({ id, model: session.model, delta: { content: token }, finishReason: null })));
+    });
+  } catch (err) {
+    if (err instanceof SessionGoneError) {
+      err = toGatewayError(new SupplierError("supplier_http_error", { status: 404, message: err.message }));
     }
-  });
-
-  if (errored !== undefined) {
     if (stream) {
-      res.write(sseData(toErrorBody(toGatewayError(new SupplierError("upstream_error", { message: errored })))));
+      res.write(sseData(toErrorBody(toGatewayError(err))));
       res.end();
       return;
     }
-    throw toGatewayError(new SupplierError("upstream_error", { message: errored }));
+    throw toGatewayError(err);
   }
 
-  // Mirror the transcript for endChat's response_hash verification.
-  const mirror = transcripts.get(session.id) ?? [];
-  mirror.push({ role: "user", content });
-  mirror.push({ role: "assistant", content: assistant });
-  transcripts.set(session.id, mirror);
-
   if (stream) {
-    res.write(sseData(buildChunk({ id, model: session.model, delta: {}, finishReason: "stop" })));
+    res.write(sseData(buildChunk({ id, model: session.model, delta: {}, finishReason: result.finishReason })));
     res.write(SSE_DONE);
     res.end();
   } else {
-    const usage: Usage = { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 };
-    res.status(200).json(buildChatCompletion({ id, model: session.model, content: assistant, usage }));
+    const usage: Usage = {
+      prompt_tokens: result.usage.prompt_tokens,
+      completion_tokens: result.usage.completion_tokens,
+      total_tokens: result.usage.prompt_tokens + result.usage.completion_tokens,
+    };
+    res.status(200).json(buildChatCompletion({ id, model: session.model, content: result.assistantMessage.content, usage }));
   }
 }
 
 interface SupplierFrame {
   type?: string;
   value?: string;
-  message?: string;
+  message?: unknown;
+  finish_reason?: unknown;
+  usage?: unknown;
 }
 
 /** Read a supplier SSE stream of {type:token|done|error} frames. */
@@ -308,7 +407,34 @@ async function closeSession(
   });
 
   deps.store.setSessionState(session.id, "closed", Date.now());
-  transcripts.delete(session.id);
+  dropSessionState(session.id);
+
+  if (result.settleMode === "ticket") {
+    // Nothing was Submitted/Accepted: no receipt, no charge (the escrow
+    // returns via the sweeper's reclaim), and no wallet churn to tidy up.
+    deps.store.insertUsage({
+      id: randomUUID(),
+      key_id: keyRow.id,
+      created_at: Date.now(),
+      kind: "chat_session",
+      model: session.model,
+      capability_id: CAPABILITY,
+      supplier_pkh: session.supplier_pkh,
+      escrow_ref: session.escrow_ref,
+      cost_lovelace: "0",
+      prompt_tokens: 0,
+      completion_tokens: 0,
+      status: "completed",
+      failure_reason: null,
+    });
+    res.status(200).json({
+      status: "ended",
+      escrow_ref: session.escrow_ref,
+      settle_mode: "ticket",
+      model: session.model,
+    });
+    return;
+  }
 
   const prompt = result.receipt.prompt_tokens ?? 0;
   const completion = result.receipt.completion_tokens ?? 0;

gateway/src/openai/shapes.ts

@@ -7,6 +7,7 @@
 
 import { randomBytes } from "crypto";
 import type { Receipt } from "@marketplace/shared/receipt";
+import type { ToolCall } from "@marketplace/shared/tx";
 
 export function genId(): string {
   return `chatcmpl-${randomBytes(16).toString("hex")}`;
@@ -40,6 +41,8 @@ export function buildChatCompletion(args: {
   model: string;
   content: string;
   usage: Usage;
+  toolCalls?: ToolCall[];
+  finishReason?: "stop" | "tool_calls";
   vector?: VectorReceipt;
 }): Record<string, unknown> {
   return {
@@ -50,8 +53,12 @@ export function buildChatCompletion(args: {
     choices: [
       {
         index: 0,
-        message: { role: "assistant", content: args.content },
-        finish_reason: "stop",
+        message: {
+          role: "assistant",
+          content: args.content,
+          ...(args.toolCalls?.length ? { tool_calls: args.toolCalls } : {}),
+        },
+        finish_reason: args.finishReason ?? "stop",
       },
     ],
     usage: args.usage,
@@ -63,8 +70,10 @@ export function buildChatCompletion(args: {
 export function buildChunk(args: {
   id: string;
   model: string;
-  delta: { role?: string; content?: string } | Record<string, never>;
-  finishReason: "stop" | null;
+  delta:
+    | { role?: string; content?: string; tool_calls?: Array<{ index: number } & ToolCall> }
+    | Record<string, never>;
+  finishReason: "stop" | "tool_calls" | null;
   usage?: Usage;
   vector?: VectorReceipt;
 }): Record<string, unknown> {

gateway/src/openai/transcripts.ts

@@ -0,0 +1,35 @@
+/**
+ * gateway/src/openai/transcripts.ts — in-memory per-session state shared by
+ * the session routes (sessions.ts), the demo affinity executor (demoChat.ts)
+ * and the sweeper's housekeeping.
+ *
+ * transcripts: mirror of each open session's transcript (sessionId → messages).
+ *   Kept hash-identical to the supplier's authoritative transcript (endChat's
+ *   response_hash check). Lost on restart — abandoned sessions are recovered
+ *   by the supplier idle watchdog + gateway sweeper.
+ * sessionLocks: per-session Mutex serializing turns on ONE session (concurrent
+ *   turns would interleave supplier transcript appends). Distinct from the
+ *   per-key Mutex, which only serializes on-chain wallet work.
+ */
+
+import type { ChatMessage } from "@marketplace/shared/tx";
+import { Mutex } from "../sdk/registry.js";
+
+export const transcripts = new Map<string, ChatMessage[]>();
+
+const sessionLocks = new Map<string, Mutex>();
+
+export function getSessionLock(sessionId: string): Mutex {
+  let lock = sessionLocks.get(sessionId);
+  if (!lock) {
+    lock = new Mutex();
+    sessionLocks.set(sessionId, lock);
+  }
+  return lock;
+}
+
+/** Forget all in-memory state for a session (closed/gone). */
+export function dropSessionState(sessionId: string): void {
+  transcripts.delete(sessionId);
+  sessionLocks.delete(sessionId);
+}

gateway/src/openai/validate.ts

@@ -2,7 +2,7 @@
  * gateway/src/openai/validate.ts — request body validation for chat endpoints.
  */
 
-import type { ChatMessage } from "@marketplace/shared/tx";
+import { normalizeChatMessage, type ChatMessage } from "@marketplace/shared/tx";
 import { badRequest } from "./errors.js";
 
 export interface ParsedChatRequest {
@@ -10,21 +10,30 @@ export interface ParsedChatRequest {
   messages: ChatMessage[];
   maxTokens?: number;
   stream: boolean;
+  tools?: unknown[];
+  toolChoice?: unknown;
 }
 
 const ROLES = new Set(["system", "user", "assistant"]);
 
-/** Validate an OpenAI chat.completions-style body. Rejects tools/functions;
- * silently ignores temperature/n/stop/top_p/etc (documented). */
-export function parseChatRequest(body: unknown): ParsedChatRequest {
+/** Validate an OpenAI chat.completions-style body. Default path rejects
+ * tools/functions; `allowTools` (demo keys) accepts tools/tool_choice, the
+ * `tool` role, null content and assistant tool_calls, normalizing every
+ * message into canonical mirror shape. Silently ignores temperature/n/stop/
+ * top_p/etc (documented). */
+export function parseChatRequest(body: unknown, opts?: { allowTools?: boolean }): ParsedChatRequest {
+  const allowTools = opts?.allowTools === true;
   if (typeof body !== "object" || body === null) {
     throw badRequest("invalid_body", "request body must be a JSON object");
   }
   const b = body as Record<string, unknown>;
 
-  if ("tools" in b || "tool_choice" in b || "functions" in b || "function_call" in b) {
+  if (!allowTools && ("tools" in b || "tool_choice" in b || "functions" in b || "function_call" in b)) {
     throw badRequest("unsupported_parameter", "tools and function-calling are not supported by this gateway");
   }
+  if (allowTools && ("functions" in b || "function_call" in b)) {
+    throw badRequest("unsupported_parameter", "legacy functions/function_call are not supported; use tools");
+  }
 
   const model = b.model;
   if (typeof model !== "string" || model === "") {
@@ -37,6 +46,14 @@ export function parseChatRequest(body: unknown): ParsedChatRequest {
   }
   const messages: ChatMessage[] = [];
   for (const m of rawMessages) {
+    if (allowTools) {
+      const msg = normalizeChatMessage(m);
+      if (!msg) {
+        throw badRequest("invalid_messages", "each message must be OpenAI-shaped {role, content, tool_calls?, tool_call_id?}");
+      }
+      messages.push(msg);
+      continue;
+    }
     if (
       typeof m !== "object" ||
       m === null ||
@@ -50,6 +67,13 @@ export function parseChatRequest(body: unknown): ParsedChatRequest {
     messages.push({ role: mm.role, content: mm.content });
   }
 
+  let tools: unknown[] | undefined;
+  let toolChoice: unknown;
+  if (allowTools && Array.isArray(b.tools) && b.tools.length > 0) {
+    tools = b.tools;
+    toolChoice = b.tool_choice;
+  }
+
   let maxTokens: number | undefined;
   const mt = b.max_tokens ?? b.max_completion_tokens;
   if (mt !== undefined && mt !== null) {
@@ -60,5 +84,5 @@ export function parseChatRequest(body: unknown): ParsedChatRequest {
   }
 
   const stream = b.stream === true;
-  return { model, messages, maxTokens, stream };
+  return { model, messages, maxTokens, stream, tools, toolChoice };
 }

gateway/src/routing/selectSupplier.ts

@@ -53,9 +53,17 @@ export function parseRef(ref: string): OutputReference | null {
 
 export interface SelectSupplierOpts {
   indexerUrl: string;
+  /** Exact model match; "" matches ANY model under the capability (used by
+   * single-model capabilities like the ocr.* ids, where the capability
+   * already pins the model). */
   model: string;
   capabilityId: string;
   fetchFn?: typeof globalThis.fetch;
+  /** supplier_pkhs eligible regardless of polled status — used after the demo
+   * executor frees a supplier the indexer poller (~20 s cadence) still reports
+   * busy. sdk.startChat's live /status pre-check still guards genuinely-busy
+   * suppliers. */
+  ignoreStatusFor?: ReadonlySet<string>;
 }
 
 /**
@@ -77,9 +85,11 @@ export async function selectCandidates(opts: SelectSupplierOpts): Promise<Suppli
   const candidates: SupplierCandidate[] = [];
   for (const raw of body as SupplierView[]) {
     if (raw.capability_id !== opts.capabilityId) continue;
-    if (raw.model !== opts.model) continue;
+    if (opts.model !== "" && raw.model !== opts.model) continue;
     if (raw.advert_status !== "Active") continue;
-    if (raw.status !== "free" && raw.status !== "unknown") continue;
+    if (raw.status !== "free" && raw.status !== "unknown" && !opts.ignoreStatusFor?.has(raw.supplier_pkh)) {
+      continue;
+    }
     const advertRef = parseRef(raw.utxo_ref);
     if (!advertRef) continue;
     candidates.push({
@@ -101,19 +111,27 @@ export async function selectCandidates(opts: SelectSupplierOpts): Promise<Suppli
   return candidates;
 }
 
-/** Distinct models across Active suppliers (for GET /openai/v1/models). */
+/** Distinct models across Active suppliers (for GET /openai/v1/models).
+ * Pass capabilityId to list only models under one capability (demo keys see
+ * only llm.chat.v1 models — the session-backed executor's routing target). */
 export async function listModels(opts: {
   indexerUrl: string;
+  capabilityId?: string;
   fetchFn?: typeof globalThis.fetch;
 }): Promise<string[]> {
   const fetchFn = opts.fetchFn ?? globalThis.fetch;
-  const res = await fetchFn(`${opts.indexerUrl}/suppliers`);
+  const url = opts.capabilityId
+    ? `${opts.indexerUrl}/suppliers?capability_id=${encodeURIComponent(opts.capabilityId)}`
+    : `${opts.indexerUrl}/suppliers`;
+  const res = await fetchFn(url);
   if (!res.ok) throw new Error(`indexer /suppliers returned ${res.status}`);
   const body = (await res.json()) as unknown;
   if (!Array.isArray(body)) throw new Error("indexer /suppliers did not return an array");
   const models = new Set<string>();
   for (const raw of body as SupplierView[]) {
-    if (raw.advert_status === "Active" && typeof raw.model === "string") models.add(raw.model);
+    if (raw.advert_status !== "Active" || typeof raw.model !== "string") continue;
+    if (opts.capabilityId !== undefined && raw.capability_id !== opts.capabilityId) continue;
+    models.add(raw.model);
   }
   return [...models].sort();
 }

gateway/src/server.ts

@@ -7,13 +7,14 @@
  *           POST /openai/v1/chat/sessions[/:id/messages|/close].
  */
 
-import express, { type Express, type Request, type Response } from "express";
+import express, { type Express, type NextFunction, type Request, type Response } from "express";
 import type { GatewayDeps } from "./deps.js";
 import { makeApiKeyAuth } from "./middleware/apiKeyAuth.js";
-import { ipRateLimit, keyRateLimit } from "./middleware/rateLimit.js";
+import { ipRateLimit, keyRateLimit, demoIpRateLimit } from "./middleware/rateLimit.js";
 import { sendError } from "./middleware/http.js";
 import { notFound } from "./openai/errors.js";
 import { makeChatCompletionsHandler } from "./openai/chatCompletions.js";
+import { makeOcrExtractHandler } from "./ocr/extract.js";
 import { makeModelsHandler } from "./openai/models.js";
 import {
   makeOpenSessionHandler,
@@ -23,11 +24,44 @@ import {
 import { makeSignupHandler, makeAccountHandler, makeWithdrawHandler } from "./account/routes.js";
 import { INDEX_HTML } from "./ui/page.js";
 
+/**
+ * CORS for browser clients on an allowlisted Origin (e.g. the marketplace
+ * frontend's "Generate API key" page, served from a different host than the
+ * gateway). Only echoes Access-Control-* headers when the request's Origin is
+ * in the configured allowlist; an empty allowlist disables CORS entirely.
+ * No credentials are allowed — the signup call is a plain bearer-less POST, so
+ * cookies never ride along cross-origin. Preflight OPTIONS is answered 204.
+ */
+function corsMiddleware(allowOrigins: string[]) {
+  const allow = new Set(allowOrigins);
+  return (req: Request, res: Response, next: NextFunction): void => {
+    const origin = req.headers.origin;
+    if (typeof origin === "string" && allow.has(origin)) {
+      res.setHeader("Access-Control-Allow-Origin", origin);
+      res.setHeader("Vary", "Origin");
+      res.setHeader("Access-Control-Allow-Methods", "GET, POST, OPTIONS");
+      res.setHeader("Access-Control-Allow-Headers", "Content-Type, Authorization");
+      res.setHeader("Access-Control-Max-Age", "600");
+    }
+    if (req.method === "OPTIONS") {
+      res.status(204).end();
+      return;
+    }
+    next();
+  };
+}
+
 export function createApp(deps: GatewayDeps): Express {
   const app = express();
   // Behind traefik/Cloudflare — trust the proxy chain so req.ip is the client.
   app.set("trust proxy", true);
-  app.use(express.json({ limit: "1mb" }));
+  // CORS before body-parsing so preflight OPTIONS short-circuits cheaply.
+  app.use(corsMiddleware(deps.config.corsOrigins));
+  // /v1/ocr carries a base64 page image and mounts its own larger parser on
+  // the route; every other path keeps the 1mb ceiling.
+  const jsonBody = express.json({ limit: "1mb" });
+  app.use((req, res, next) =>
+    req.path.startsWith("/v1/ocr/") ? next() : jsonBody(req, res, next));
 
   app.get("/healthz", (_req: Request, res: Response) => {
     res.json({ ok: true });
@@ -40,15 +74,23 @@ export function createApp(deps: GatewayDeps): Express {
   // Public signup, rate-limited per IP.
   app.post("/signup", ipRateLimit(deps.config.signupRate), makeSignupHandler(deps));
 
-  // Bearer-gated routes: auth → per-key rate limit → handler.
+  // Bearer-gated routes: auth → per-key rate limit (skips demo keys) →
+  // per-IP demo limit (demo keys only) → handler.
   const auth = makeApiKeyAuth(deps.store);
   const krl = keyRateLimit(deps.config.keyRate);
-  const gated = [auth, krl] as const;
+  const dirl = demoIpRateLimit(deps.config.demoIpRate);
+  const gated = [auth, krl, dirl] as const;
 
   app.get("/account", auth, makeAccountHandler(deps));
   app.post("/account/withdraw", auth, makeWithdrawHandler(deps));
 
   app.post("/openai/v1/chat/completions", ...gated, makeChatCompletionsHandler(deps));
+  // Model-scoped OCR: 16mb parser covers MAX_OCR_IMAGE_B64_CHARS
+  // (~12M chars ≈ 9 MB binary) plus JSON envelope. Gates (auth + rate
+  // limits) run BEFORE the parser so an unauthenticated caller cannot
+  // buffer/JSON.parse up to 16mb before getting a 401, and rate limiting
+  // can actually throttle the request (2026-08 review finding C2).
+  app.post("/v1/ocr/extract", ...gated, express.json({ limit: "16mb" }), makeOcrExtractHandler(deps));
   app.get("/openai/v1/models", auth, makeModelsHandler(deps));
   app.post("/openai/v1/chat/sessions", ...gated, makeOpenSessionHandler(deps));
   app.post("/openai/v1/chat/sessions/:id/messages", ...gated, makeSessionMessageHandler(deps));

gateway/src/sweeper.ts

@@ -10,26 +10,120 @@
  * state and never blind-reclaim a Submitted escrow. The on-chain validator is
  * the source of truth for deadlines, so premature/late attempts simply fail and
  * are caught here; the age thresholds just avoid racing in-flight requests.
+ *
+ * Chat sessions get special handling:
+ *   - Their deliver_by is ~30 min (vs ~10 min one-shot), so rows belonging to
+ *     an OPEN session use CHAT_RECLAIM_AFTER_MS — otherwise every open session
+ *     older than 11 min would trigger a doomed Reclaim attempt per sweep while
+ *     holding the key's mutex.
+ *   - When a session's escrow reaches a terminal state (Accepted here, spent
+ *     elsewhere, or Reclaimed), the session row is closed, its in-memory state
+ *     dropped, and a usage row recorded — abandoned demo sessions (nobody calls
+ *     /close) settle and get billed through this path.
+ *
+ * Escrow↔session lineage: the indexer keeps one row per utxo_ref per state
+ * (terminal spends update nothing), and the validator preserves posted_at
+ * across states — so a session's lineage is "rows sharing the posted_at of the
+ * session's original Open ref row".
+ *
+ * Because terminal spends never update indexer rows, every settled escrow
+ * would be retried (one doomed chain query each) on every sweep forever. The
+ * in-process `settledRefs` memory skips refs this process has already settled
+ * or found spent; a restart just causes one harmless retry burst.
+ *
+ * Ticket mode (config.chatSettleMode === "ticket"): session escrows are never
+ * Claimed — they sit Open until reclaimed here after deliver_by. Reclaiming
+ * an Open ticket does NOT close a still-open session row: nothing re-checks
+ * the escrow after /v1/chat/start, so the conversation (and its affinity
+ * mirror) lives on, bounded only by the idle janitor.
  */
 
+import { randomUUID } from "crypto";
 import type { GatewayDeps } from "./deps.js";
-import { fetchEscrows, acceptAndConfirm, reclaimAndConfirm } from "./onchain/settle.js";
+import type { SessionRow } from "./db/store.js";
+import { fetchEscrows, acceptAndConfirm, reclaimAndConfirm, type EscrowRow } from "./onchain/settle.js";
 import { parseRef } from "./routing/selectSupplier.js";
+import { dropSessionState } from "./openai/transcripts.js";
+import { dropDemoSession, sweepIdleDemoSessions } from "./openai/demoChat.js";
+import { CAPABILITY as CHAT_CAPABILITY } from "./openai/sessions.js";
 
 // Don't touch escrows newer than this — an in-flight request is settling them.
 const MIN_AGE_MS = 90_000;
 // Reclaim Open/Claimed escrows older than this (past deliver_by ~10m + buffer).
 const RECLAIM_AFTER_MS = 11 * 60 * 1000;
+// Chat-session escrows: deliver_by ≈ posted_at + 30 min, so wait longer.
+const CHAT_RECLAIM_AFTER_MS = 35 * 60 * 1000;
+// Reclaim only fires once the validator allows it (validity lower bound >=
+// deliver_by); attempting earlier is a guaranteed-doomed tx build.
+const DELIVER_BY_BUFFER_MS = 60_000;
+
+/** Refs this process already settled (Accepted/Reclaimed) or found spent.
+ * The indexer never updates terminal rows, so without this every historical
+ * escrow is retried each sweep. Bounded FIFO; restart = one retry burst. */
+const settledRefs = new Set<string>();
+const SETTLED_REFS_MAX = 5_000;
+
+function markSettled(ref: string): void {
+  settledRefs.add(ref);
+  if (settledRefs.size > SETTLED_REFS_MAX) {
+    const first = settledRefs.values().next().value;
+    if (first !== undefined) settledRefs.delete(first);
+  }
+}
 
 function log(msg: string): void {
   // eslint-disable-next-line no-console
   console.log(`[gateway:sweeper] ${msg}`);
 }
 
-export async function runSweepOnce(deps: GatewayDeps): Promise<void> {
+function closeSessionRow(
+  deps: GatewayDeps,
+  session: SessionRow,
+  outcome: { status: "completed" | "reclaimed"; costLovelace: string | null },
+): void {
+  // Re-read: an in-flight close (route handler or demo executor) may have won.
+  const current = deps.store.getSession(session.id);
+  if (!current || current.state !== "open") return;
+  deps.store.setSessionState(session.id, "closed", Date.now());
+  dropSessionState(session.id);
+  dropDemoSession(session.id);
+  deps.store.insertUsage({
+    id: randomUUID(),
+    key_id: session.key_id,
+    created_at: Date.now(),
+    kind: "chat_session",
+    model: session.model,
+    capability_id: CHAT_CAPABILITY,
+    supplier_pkh: session.supplier_pkh,
+    escrow_ref: session.escrow_ref,
+    cost_lovelace: outcome.costLovelace,
+    prompt_tokens: 0,
+    completion_tokens: 0,
+    status: outcome.status,
+    failure_reason: null,
+  });
+  log(`closed abandoned session ${session.id} (${outcome.status})`);
+}
+
+/** Chain-settle functions, injectable for tests. */
+export interface SweepSettleFns {
+  accept: typeof acceptAndConfirm;
+  reclaim: typeof reclaimAndConfirm;
+}
+
+export async function runSweepOnce(
+  deps: GatewayDeps,
+  settle: SweepSettleFns = { accept: acceptAndConfirm, reclaim: reclaimAndConfirm },
+): Promise<void> {
+  // Idle demo sessions first: closing them makes the supplier Submit, so this
+  // sweep (or the next) can Accept the escrow right after.
+  await sweepIdleDemoSessions(deps).catch((err) =>
+    log(`demo idle janitor: ${err instanceof Error ? err.message : String(err)}`),
+  );
+
   const now = Date.now();
   for (const keyRow of deps.store.listKeys()) {
-    let rows;
+    let rows: EscrowRow[];
     try {
       rows = await fetchEscrows(deps.config.indexerUrl, keyRow.wallet_pkh, deps.fetchFn);
     } catch (err) {
@@ -37,20 +131,71 @@ export async function runSweepOnce(deps: GatewayDeps): Promise<void> {
       continue;
     }
 
+    // Lineage-match open sessions: the session stores its ORIGINAL Open ref;
+    // rows sharing that row's posted_at are the same escrow in later states.
+    const sessionByPostedAt = new Map<number, SessionRow>();
+    for (const session of deps.store.listOpenSessionsByKey(keyRow.id)) {
+      const origin = rows.find((r) => r.utxo_ref === session.escrow_ref);
+      if (origin) sessionByPostedAt.set(origin.posted_at, session);
+    }
+
     const ctx = deps.registry.getContext(keyRow);
     for (const row of rows) {
       const ref = parseRef(row.utxo_ref);
       if (!ref) continue;
+      if (settledRefs.has(row.utxo_ref)) continue;
       const ageMs = now - (row.posted_at ?? 0);
       if (ageMs < MIN_AGE_MS) continue;
+      const chatSession = sessionByPostedAt.get(row.posted_at);
 
       try {
         if (row.state === "Submitted") {
-          await ctx.mutex.run(() => acceptAndConfirm(deps.chain, ctx.walletKey, ref));
-          log(`accepted stranded Submitted escrow ${row.utxo_ref}`);
-        } else if ((row.state === "Open" || row.state === "Claimed") && ageMs > RECLAIM_AFTER_MS) {
-          await ctx.mutex.run(() => reclaimAndConfirm(deps.chain, ctx.walletKey, ref));
-          log(`reclaimed stranded ${row.state} escrow ${row.utxo_ref}`);
+          try {
+            await ctx.mutex.run(() => settle.accept(deps.chain, ctx.walletKey, ref));
+            log(`accepted stranded Submitted escrow ${row.utxo_ref}`);
+            markSettled(row.utxo_ref);
+            if (chatSession) {
+              closeSessionRow(deps, chatSession, { status: "completed", costLovelace: chatSession.price_lovelace });
+            }
+          } catch (err) {
+            // Accept fails routinely when the UTxO is already spent (settled
+            // by an in-flight request, an earlier sweep, or supplier Release).
+            // For session lineages, confirm the spend on-chain before closing
+            // the row — an Accept that failed for another reason (timeout)
+            // must leave the session open for the next sweep.
+            if ((await deps.chain.queryUtxo(ref)) === null) {
+              markSettled(row.utxo_ref);
+              if (chatSession) {
+                closeSessionRow(deps, chatSession, { status: "completed", costLovelace: chatSession.price_lovelace });
+              }
+            } else if (!chatSession) {
+              // Preserve the historical quiet-skip for one-shot escrows.
+              void err;
+            }
+          }
+        } else if (row.state === "Open" || row.state === "Claimed") {
+          const reclaimAfterMs = chatSession ? CHAT_RECLAIM_AFTER_MS : RECLAIM_AFTER_MS;
+          // The validator only allows Reclaim at/after deliver_by — skip
+          // attempts that cannot succeed yet (early-closed ticket sessions).
+          const pastDeliverBy =
+            typeof row.deliver_by === "number" ? now > row.deliver_by + DELIVER_BY_BUFFER_MS : true;
+          if (ageMs > reclaimAfterMs && pastDeliverBy) {
+            await ctx.mutex.run(() => settle.reclaim(deps.chain, ctx.walletKey, ref));
+            log(`reclaimed stranded ${row.state} escrow ${row.utxo_ref}`);
+            markSettled(row.utxo_ref);
+            if (chatSession) {
+              const ticketOpen = deps.config.chatSettleMode === "ticket" && row.state === "Open";
+              if (ticketOpen) {
+                // Ticket escrow reclaimed out from under a LIVE session —
+                // by design. Keep the session row + mirror; only the idle
+                // janitor / close paths end ticket sessions.
+                log(`ticket mode: escrow reclaimed, session ${chatSession.id} stays open`);
+              } else {
+                // Funds returned — the session never settled; record no cost.
+                closeSessionRow(deps, chatSession, { status: "reclaimed", costLovelace: null });
+              }
+            }
+          }
         }
       } catch {
         // Expected when the validator rejects (wrong window/state) or the UTxO
@@ -58,9 +203,6 @@ export async function runSweepOnce(deps: GatewayDeps): Promise<void> {
       }
     }
   }
-
-  // Mark closed/settled chat sessions whose escrow is no longer active.
-  // (Best-effort housekeeping; the escrow recovery above is the load-bearing part.)
 }
 
 /** Start the periodic sweeper. Returns a stop() function. */

gateway/src/ui/page.ts

@@ -8,7 +8,7 @@ export const INDEX_HTML = `<!doctype html>
 <head>
 <meta charset="utf-8" />
 <meta name="viewport" content="width=device-width, initial-scale=1" />
-<title>Vector Marketplace — OpenAI Gateway</title>
+<title>Vector Marketplace - OpenAI Gateway</title>
 <style>
   body { font-family: system-ui, sans-serif; max-width: 760px; margin: 2rem auto; padding: 0 1rem; line-height: 1.5; }
   h1 { font-size: 1.5rem; } h2 { font-size: 1.1rem; margin-top: 2rem; }
@@ -22,7 +22,7 @@ export const INDEX_HTML = `<!doctype html>
 </style>
 </head>
 <body>
-<h1>Vector Marketplace — OpenAI-compatible Gateway</h1>
+<h1>Vector Marketplace - OpenAI-compatible Gateway</h1>
 <p class="muted">Point any OpenAI SDK at <code id="base"></code> using your API key as the bearer token.</p>
 
 <h2>1. Create an API key</h2>
@@ -33,7 +33,7 @@ export const INDEX_HTML = `<!doctype html>
 <div id="signupOut"></div>
 
 <h2>2. Fund your wallet</h2>
-<p class="muted">Send AP3X to your deposit address (you need price + bonds + ~5 ADA collateral + fees).</p>
+<p class="muted">Send AP3X to your deposit address (you need price + bonds + ~5 AP3X collateral + fees).</p>
 
 <h2>3. Account</h2>
 <div class="row">

gateway/src/walletHealth.ts

@@ -2,8 +2,8 @@
  * gateway/src/walletHealth.ts — keep each custodial wallet usable.
  *
  * After every settled request (and on a periodic tick) we re-shape the wallet to
- * {≥5 ADA collateral UTxO, working UTxO} via the shared consolidate builder, so
- * the NEXT request's Accept/Reclaim script-spend always finds a pure-ADA
+ * {≥5 AP3X collateral UTxO, working UTxO} via the shared consolidate builder, so
+ * the NEXT request's Accept/Reclaim script-spend always finds a pure-AP3X
  * collateral candidate. No-op when the wallet is already healthy.
  */
 

gateway/test/affinity.test.ts

@@ -0,0 +1,143 @@
+/**
+ * affinity.test.ts — transcript-prefix session matching rules.
+ */
+
+import { describe, it, expect } from "vitest";
+import {
+  matchSessionPrefix,
+  normalizeAssistantContent,
+  trimToRespondable,
+  type AffinityCandidate,
+} from "../src/openai/affinity.js";
+import type { ChatMessage } from "@marketplace/shared/tx";
+
+const u = (content: string): ChatMessage => ({ role: "user", content });
+const a = (content: string): ChatMessage => ({ role: "assistant", content });
+const TOOL_CALL = { id: "call_1", type: "function" as const, function: { name: "get_time", arguments: "{}" } };
+
+function cand(sessionId: string, mirror: ChatMessage[], lastUsedAt = 0, model = "m"): AffinityCandidate {
+  return { sessionId, model, mirror, lastUsedAt };
+}
+
+describe("matchSessionPrefix", () => {
+  it("matches an exact prefix and returns the delta", () => {
+    const got = matchSessionPrefix([cand("s1", [u("hi"), a("hello")])], [u("hi"), a("hello"), u("more")], "m");
+    expect(got).toEqual({ sessionId: "s1", delta: [u("more")] });
+  });
+
+  it("accepts a multi-message tool delta ending with a tool result", () => {
+    const mirror = [u("hi"), { role: "assistant" as const, content: "", tool_calls: [TOOL_CALL] }];
+    const incoming = [...mirror, { role: "tool" as const, content: "3pm", tool_call_id: "call_1" }];
+    const got = matchSessionPrefix([cand("s1", mirror)], incoming, "m");
+    expect(got?.sessionId).toBe("s1");
+    expect(got?.delta).toHaveLength(1);
+    expect(got?.delta[0].role).toBe("tool");
+  });
+
+  it("rejects deltas that end with a non-respondable role", () => {
+    const got = matchSessionPrefix([cand("s1", [u("hi")])], [u("hi"), a("stray assistant")], "m");
+    expect(got).toBeNull();
+  });
+
+  it("never matches when incoming is not longer than the mirror (retry of answered request)", () => {
+    const mirror = [u("hi"), a("hello")];
+    expect(matchSessionPrefix([cand("s1", mirror)], mirror, "m")).toBeNull();
+    expect(matchSessionPrefix([cand("s1", mirror)], [u("hi")], "m")).toBeNull();
+  });
+
+  it("rejects diverged histories (prefix mismatch)", () => {
+    const got = matchSessionPrefix([cand("s1", [u("hi"), a("hello")])], [u("hi"), a("DIFFERENT"), u("more")], "m");
+    expect(got).toBeNull();
+  });
+
+  it("requires the model to match", () => {
+    expect(matchSessionPrefix([cand("s1", [u("hi")], 0, "other")], [u("hi"), u("more")], "m")).toBeNull();
+  });
+
+  it("prefers the longest mirror, then most recently used", () => {
+    const incoming = [u("hi"), a("hello"), u("more")];
+    const longest = matchSessionPrefix(
+      [cand("short", [u("hi")], 99), cand("long", [u("hi"), a("hello")], 0)],
+      incoming,
+      "m",
+    );
+    expect(longest?.sessionId).toBe("long");
+
+    const tie = matchSessionPrefix(
+      [cand("old", [u("hi"), a("hello")], 1), cand("recent", [u("hi"), a("hello")], 2)],
+      incoming,
+      "m",
+    );
+    expect(tie?.sessionId).toBe("recent");
+  });
+
+  it("matches an empty mirror (fresh session) against any respondable history", () => {
+    const incoming = [{ role: "system" as const, content: "be nice" }, u("hi")];
+    const got = matchSessionPrefix([cand("fresh", [])], incoming, "m");
+    expect(got).toEqual({ sessionId: "fresh", delta: incoming });
+  });
+
+  it("treats content null-normalization as equivalent (tool-call assistant echo)", () => {
+    const mirror = [u("hi"), { role: "assistant" as const, content: "", tool_calls: [TOOL_CALL] }];
+    // Client echoes the assistant message back with content:"" (normalized from null upstream).
+    const incoming = [u("hi"), { role: "assistant" as const, content: "", tool_calls: [{ ...TOOL_CALL }] }, { role: "tool" as const, content: "x", tool_call_id: "call_1" }];
+    expect(matchSessionPrefix([cand("s1", mirror)], incoming, "m")?.sessionId).toBe("s1");
+  });
+
+  it("matches leniently when the client reshapes the assistant echo (think-strip + whitespace)", () => {
+    const mirror = [u("hi"), a("<think>plan it</think>Hello  there\n")];
+    const got = matchSessionPrefix([cand("s1", mirror)], [u("hi"), a("Hello there"), u("more")], "m");
+    expect(got).toEqual({ sessionId: "s1", delta: [u("more")] });
+  });
+
+  it("stays strict on user/tool divergence even with lenient assistant matching", () => {
+    const mirror = [u("hi "), a("Hello")];
+    // User content differs by whitespace — client-authored roles must be byte-exact.
+    expect(matchSessionPrefix([cand("s1", mirror)], [u("hi"), a("Hello"), u("more")], "m")).toBeNull();
+  });
+
+  it("rejects assistant tool_calls divergence under lenient matching", () => {
+    const mirror = [u("hi"), { role: "assistant" as const, content: "x", tool_calls: [TOOL_CALL] }];
+    const changed = { ...TOOL_CALL, function: { name: "get_time", arguments: '{"tz":"utc"}' } };
+    const incoming = [u("hi"), { role: "assistant" as const, content: "x ", tool_calls: [changed] }, { role: "tool" as const, content: "t", tool_call_id: "call_1" }];
+    expect(matchSessionPrefix([cand("s1", mirror)], incoming, "m")).toBeNull();
+  });
+
+  it("prefers a strict match over a lenient one regardless of recency", () => {
+    const incoming = [u("hi"), a("Hello"), u("more")];
+    const got = matchSessionPrefix(
+      [
+        cand("lenient", [u("hi"), a("Hello   ")], 99),
+        cand("strict", [u("hi"), a("Hello")], 1),
+      ],
+      incoming,
+      "m",
+    );
+    expect(got?.sessionId).toBe("strict");
+  });
+});
+
+describe("normalizeAssistantContent", () => {
+  it("strips think blocks, collapses whitespace and trims", () => {
+    expect(normalizeAssistantContent("<think>a\nb</think> Hi\n\nthere  friend ")).toBe("Hi there friend");
+    expect(normalizeAssistantContent("<THINK>x</THINK>ok")).toBe("ok");
+    expect(normalizeAssistantContent("plain")).toBe("plain");
+  });
+});
+
+describe("trimToRespondable", () => {
+  it("drops trailing assistant/system messages", () => {
+    expect(trimToRespondable([u("hi"), a("partial")])).toEqual([u("hi")]);
+    expect(trimToRespondable([u("hi"), a("x"), { role: "system" as const, content: "nudge" }])).toEqual([u("hi")]);
+  });
+
+  it("returns empty when nothing respondable remains", () => {
+    expect(trimToRespondable([a("hello")])).toEqual([]);
+    expect(trimToRespondable([])).toEqual([]);
+  });
+
+  it("keeps already-respondable histories unchanged", () => {
+    const msgs = [u("hi"), a("x"), u("more")];
+    expect(trimToRespondable(msgs)).toEqual(msgs);
+  });
+});

gateway/test/http.test.ts

@@ -1,23 +1,37 @@
-import { describe, it, expect } from "vitest";
+import { describe, it, expect, vi } from "vitest";
 import { tmpdir } from "os";
 import { join } from "path";
-import { randomUUID } from "crypto";
+import { randomBytes, randomUUID } from "crypto";
 import request from "supertest";
 import { createApp } from "../src/server.js";
+import { sweepIdleDemoSessions } from "../src/openai/demoChat.js";
 import { GatewayStore } from "../src/db/store.js";
 import { SdkRegistry } from "../src/sdk/registry.js";
+import { genPrivKeyHex, deriveWalletKey } from "../src/wallet.js";
+import { seal } from "../src/crypto/seal.js";
+import { hashApiKey } from "../src/middleware/apiKeyAuth.js";
 import type { GatewayConfig } from "../src/config.js";
 import type { GatewayDeps } from "../src/deps.js";
 
+const MASTER = "ab".repeat(32);
+
 const SUPPLIERS = [
   { utxo_ref: "aa".repeat(32) + "#0", supplier_pkh: "s1", capability_id: "llm.text.generate.v1", model: "qwen", max_output_tokens: 256, price_lovelace: "1000000", supplier_bond_lovelace: "1000000", buyer_bond_lovelace: "1000000", endpoint_url: "http://sup", advert_status: "Active", status: "free" },
+  { utxo_ref: "bb".repeat(32) + "#0", supplier_pkh: "s2", capability_id: "llm.chat.v1", model: "kimi", max_output_tokens: 256, price_lovelace: "1000000", supplier_bond_lovelace: "1000000", buyer_bond_lovelace: "1000000", endpoint_url: "http://sup", advert_status: "Active", status: "free" },
 ];
 
-function makeDeps(fetchFn?: typeof globalThis.fetch): GatewayDeps {
+function makeDeps(
+  fetchFn?: typeof globalThis.fetch,
+  overrides?: {
+    demoIpRate?: { max: number; windowMs: number };
+    chain?: GatewayDeps["chain"];
+    chatSettleMode?: "full" | "ticket";
+  },
+): GatewayDeps {
   const dbDir = join(tmpdir(), `gw-test-${randomUUID()}`);
   const store = new GatewayStore(dbDir);
   const config: GatewayConfig = {
-    masterKeyHex: "ab".repeat(32),
+    masterKeyHex: MASTER,
     indexerUrl: "http://ix",
     ogmiosUrl: "http://og",
     networkId: 0,
@@ -26,11 +40,16 @@ function makeDeps(fetchFn?: typeof globalThis.fetch): GatewayDeps {
     dbDir,
     signupRate: { max: 1000, windowMs: 60_000 },
     keyRate: { max: 1000, windowMs: 60_000 },
+    demoIpRate: overrides?.demoIpRate ?? { max: 1000, windowMs: 60_000 },
     sweeperIntervalMs: 60_000,
+    demoSessionIdleMs: 180_000,
+    chatSettleMode: overrides?.chatSettleMode ?? "full",
     walletHealthIntervalMs: 600_000,
     sdkRegistryMax: 100,
+    ocrCapabilityId: "ocr.page.extract.chandra-ocr-2.v1",
+    corsOrigins: [],
   };
-  const chain = { queryUtxosByAddress: async () => [] } as unknown as GatewayDeps["chain"];
+  const chain = overrides?.chain ?? ({ queryUtxosByAddress: async () => [] } as unknown as GatewayDeps["chain"]);
   const registry = new SdkRegistry({
     chain,
     indexerUrl: config.indexerUrl,
@@ -103,3 +122,450 @@ describe("gateway HTTP", () => {
     expect(res.body.error.code).toBe("unsupported_parameter");
   });
 });
+
+// ─── Shared demo key (session-backed completions) ───────────────────────────
+
+const TOOLS = [{ type: "function", function: { name: "get_time", parameters: {} } }];
+const TOOL_CALL = { id: "call_1", type: "function", function: { name: "get_time", arguments: "{}" } };
+
+/** A funded wallet: covers price+bonds+collateral+fee and has pure-ADA ≥ 5. */
+const fundedChain = {
+  queryUtxosByAddress: async () => [
+    { ref: { txHash: "ff".repeat(32), index: 0 }, address: "a", lovelace: 20_000_000n, assets: {}, datumHex: null, scriptRef: null },
+  ],
+} as unknown as GatewayDeps["chain"];
+
+function supplierSse(frames: unknown[]): Response {
+  const body = frames.map((f) => `data: ${JSON.stringify(f)}\n\n`).join("");
+  return new Response(body, { status: 200, headers: { "content-type": "text/event-stream" } });
+}
+
+function doneFrame(message: Record<string, unknown>, finishReason: string, usage = { prompt_tokens: 3, completion_tokens: 2 }) {
+  return { type: "done", message, finish_reason: finishReason, usage };
+}
+
+/** fetchFn serving the indexer supplier list and scripted /v1/chat/message
+ * responses. Pass `suppliers` to serve a mutable list (eviction tests flip
+ * status mid-test); /v1/chat/end calls are recorded and 200-acked. */
+function demoFetch(messageResponses: Array<() => Response>, suppliers?: () => unknown[]) {
+  const messageCalls: Array<Record<string, unknown>> = [];
+  const endCalls: string[] = [];
+  let i = 0;
+  const fetchFn = (async (url: unknown, init?: { body?: string; headers?: Record<string, string> }) => {
+    const u = String(url);
+    if (u.includes("/v1/chat/message")) {
+      messageCalls.push(init?.body ? (JSON.parse(init.body) as Record<string, unknown>) : {});
+      const make = messageResponses[Math.min(i, messageResponses.length - 1)];
+      i += 1;
+      return make();
+    }
+    if (u.includes("/v1/chat/end")) {
+      endCalls.push(init?.headers?.["X-Escrow-Ref"] ?? "");
+      return new Response(JSON.stringify({ status: "submitted" }), { status: 200 });
+    }
+    if (u.includes("/suppliers")) return new Response(JSON.stringify(suppliers ? suppliers() : SUPPLIERS), { status: 200 });
+    return new Response("[]", { status: 200 });
+  }) as unknown as typeof globalThis.fetch;
+  return { fetchFn, messageCalls, endCalls };
+}
+
+/** Insert a demo-flagged key directly (public signup can't mint one) and stub
+ * the SDK's chain-touching startChat on the pre-warmed per-key context. */
+function setupDemoKey(deps: GatewayDeps) {
+  const rawKey = `vk_test_${randomBytes(24).toString("hex")}`;
+  const privHex = genPrivKeyHex();
+  const walletKey = deriveWalletKey(privHex, 0);
+  const sealed = seal(privHex, MASTER);
+  deps.store.insertKey({
+    id: randomUUID(),
+    key_hash: hashApiKey(rawKey),
+    key_prefix: rawKey.slice(0, 12),
+    label: "shared-demo",
+    wallet_pkh: walletKey.pubKeyHash,
+    deposit_address: walletKey.address,
+    enc_priv_nonce: sealed.nonce,
+    enc_priv_ct: sealed.ct,
+    enc_priv_tag: sealed.tag,
+    master_key_version: 1,
+    created_at: Date.now(),
+    demo: 1,
+  });
+  const keyRow = deps.store.getKeyByHash(hashApiKey(rawKey))!;
+  const ctx = deps.registry.getContext(keyRow);
+  const startChat = vi.fn(async () => ({
+    escrowRef: { txHash: randomBytes(32).toString("hex"), index: 0 },
+    sessionNonce: "nonce",
+    supplierBaseUrl: "http://sup",
+  }));
+  (ctx.sdk as { startChat: unknown }).startChat = startChat;
+  return { rawKey, keyRow, startChat };
+}
+
+describe("gateway demo key", () => {
+  it("runs a session-backed completion with tools and reuses the session via affinity", async () => {
+    const { fetchFn, messageCalls } = demoFetch([
+      () => supplierSse([doneFrame({ role: "assistant", content: "", tool_calls: [TOOL_CALL] }, "tool_calls")]),
+      () => supplierSse([{ type: "token", value: "It is 3pm" }, doneFrame({ role: "assistant", content: "It is 3pm" }, "stop")]),
+    ]);
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const { rawKey, startChat } = setupDemoKey(deps);
+    const app = createApp(deps);
+
+    // Turn 1: model asks for a tool call.
+    const history: Array<Record<string, unknown>> = [{ role: "user", content: "what time is it?" }];
+    const res1 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: history, tools: TOOLS });
+    expect(res1.status).toBe(200);
+    expect(res1.body.choices[0].message.tool_calls).toEqual([TOOL_CALL]);
+    expect(res1.body.choices[0].finish_reason).toBe("tool_calls");
+    expect(startChat).toHaveBeenCalledTimes(1);
+    // First turn replays the full history and forwards tools.
+    expect((messageCalls[0].messages as unknown[]).length).toBe(1);
+    expect(messageCalls[0].tools).toEqual(TOOLS);
+
+    // Turn 2: client echoes the assistant tool_calls + tool result.
+    history.push({ role: "assistant", content: "", tool_calls: [TOOL_CALL] });
+    history.push({ role: "tool", content: "3pm", tool_call_id: "call_1" });
+    const res2 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: history, tools: TOOLS });
+    expect(res2.status).toBe(200);
+    expect(res2.body.choices[0].message.content).toBe("It is 3pm");
+    expect(res2.body.choices[0].finish_reason).toBe("stop");
+    // Affinity hit: no new session, only the unseen delta is forwarded.
+    expect(startChat).toHaveBeenCalledTimes(1);
+    const delta = messageCalls[1].messages as Array<{ role: string }>;
+    expect(delta).toHaveLength(1);
+    expect(delta[0].role).toBe("tool");
+  });
+
+  it("streams tool_calls chunks with finish_reason tool_calls", async () => {
+    const { fetchFn } = demoFetch([
+      () => supplierSse([doneFrame({ role: "assistant", content: "", tool_calls: [TOOL_CALL] }, "tool_calls")]),
+    ]);
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const { rawKey } = setupDemoKey(deps);
+    const res = await request(createApp(deps))
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "user", content: "time?" }], tools: TOOLS, stream: true });
+    expect(res.status).toBe(200);
+    const chunks = res.text
+      .split("\n\n")
+      .filter((l) => l.startsWith("data: ") && !l.includes("[DONE]"))
+      .map((l) => JSON.parse(l.slice(6)));
+    const toolChunk = chunks.find((c) => c.choices?.[0]?.delta?.tool_calls);
+    expect(toolChunk.choices[0].delta.tool_calls[0].function.name).toBe("get_time");
+    expect(chunks.at(-1).choices[0].finish_reason).toBe("tool_calls");
+    expect(res.text).toContain("data: [DONE]");
+  });
+
+  it("transparently reopens when the supplier lost the session (idle-close)", async () => {
+    const { fetchFn, messageCalls } = demoFetch([
+      () => supplierSse([{ type: "token", value: "Hi" }, doneFrame({ role: "assistant", content: "Hi" }, "stop")]),
+      () => new Response("not found", { status: 404 }),
+      () => supplierSse([{ type: "token", value: "Sure" }, doneFrame({ role: "assistant", content: "Sure" }, "stop")]),
+    ]);
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const { rawKey, startChat } = setupDemoKey(deps);
+    const app = createApp(deps);
+
+    const res1 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "user", content: "hi" }] });
+    expect(res1.status).toBe(200);
+
+    // Extended history matches the session, but the supplier 404s the turn →
+    // the executor closes the session and reopens, replaying full history.
+    const history = [
+      { role: "user", content: "hi" },
+      { role: "assistant", content: "Hi" },
+      { role: "user", content: "more please" },
+    ];
+    const res2 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: history });
+    expect(res2.status).toBe(200);
+    expect(res2.body.choices[0].message.content).toBe("Sure");
+    expect(startChat).toHaveBeenCalledTimes(2);
+    expect((messageCalls[2].messages as unknown[]).length).toBe(3);
+    // The dead session's fee was recorded for spend audit.
+    const usage = deps.store.listUsage(deps.store.listAllKeys().find((k) => k.demo === 1)!.id, 10);
+    expect(usage.some((u) => u.kind === "chat_session" && u.cost_lovelace === "1000000")).toBe(true);
+  });
+
+  it("rejects histories with nothing respondable before any chain work", async () => {
+    const deps = makeDeps(undefined, { chain: fundedChain });
+    const { rawKey, startChat } = setupDemoKey(deps);
+    const res = await request(createApp(deps))
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "assistant", content: "hello" }] });
+    expect(res.status).toBe(400);
+    expect(res.body.error.code).toBe("invalid_messages");
+    expect(startChat).not.toHaveBeenCalled();
+  });
+
+  it("answers a trailing-assistant history by trimming to the last user message (regenerate)", async () => {
+    const { fetchFn, messageCalls } = demoFetch([
+      () => supplierSse([{ type: "token", value: "Again" }, doneFrame({ role: "assistant", content: "Again" }, "stop")]),
+    ]);
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const { rawKey, startChat } = setupDemoKey(deps);
+    const res = await request(createApp(deps))
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({
+        model: "kimi",
+        messages: [
+          { role: "user", content: "hi" },
+          { role: "assistant", content: "partial answer" },
+        ],
+      });
+    expect(res.status).toBe(200);
+    expect(res.body.choices[0].message.content).toBe("Again");
+    expect(startChat).toHaveBeenCalledTimes(1);
+    // The trailing assistant message is dropped from the replay.
+    expect((messageCalls[0].messages as Array<{ role: string }>).map((m) => m.role)).toEqual(["user"]);
+  });
+
+  it("keeps the session when the client reshapes the assistant echo (lenient affinity)", async () => {
+    const { fetchFn, messageCalls } = demoFetch([
+      () => supplierSse([{ type: "token", value: "<think>hm</think>Hi  there" }, doneFrame({ role: "assistant", content: "<think>hm</think>Hi  there" }, "stop")]),
+      () => supplierSse([{ type: "token", value: "Sure" }, doneFrame({ role: "assistant", content: "Sure" }, "stop")]),
+    ]);
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const { rawKey, startChat } = setupDemoKey(deps);
+    const app = createApp(deps);
+
+    const res1 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "user", content: "hi" }] });
+    expect(res1.status).toBe(200);
+
+    // Client echoes the assistant reply with the think block stripped and
+    // whitespace collapsed — still the same session, delta only.
+    const res2 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({
+        model: "kimi",
+        messages: [
+          { role: "user", content: "hi" },
+          { role: "assistant", content: "Hi there" },
+          { role: "user", content: "more" },
+        ],
+      });
+    expect(res2.status).toBe(200);
+    expect(res2.body.choices[0].message.content).toBe("Sure");
+    expect(startChat).toHaveBeenCalledTimes(1);
+    expect((messageCalls[1].messages as unknown[]).length).toBe(1);
+  });
+
+  it("evicts the LRU demo session when every supplier is pinned", async () => {
+    const suppliers = [{ ...SUPPLIERS[1] }]; // single chat supplier, mutable status
+    const { fetchFn, messageCalls, endCalls } = demoFetch(
+      [
+        () => supplierSse([{ type: "token", value: "Hi" }, doneFrame({ role: "assistant", content: "Hi" }, "stop")]),
+        () => supplierSse([{ type: "token", value: "New" }, doneFrame({ role: "assistant", content: "New" }, "stop")]),
+      ],
+      () => suppliers,
+    );
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const { rawKey, startChat } = setupDemoKey(deps);
+    const app = createApp(deps);
+
+    const res1 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "user", content: "hi" }] });
+    expect(res1.status).toBe(200);
+
+    // The lone supplier is now pinned by session 1 (indexer sees it working).
+    suppliers[0] = { ...suppliers[0], status: "working" };
+
+    // Unrelated history → affinity miss → no free supplier → evict session 1
+    // (supplier /v1/chat/end) and retry, bypassing the stale indexer status.
+    const res2 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "user", content: "different topic" }] });
+    expect(res2.status).toBe(200);
+    expect(res2.body.choices[0].message.content).toBe("New");
+    expect(startChat).toHaveBeenCalledTimes(2);
+    expect(endCalls).toHaveLength(1);
+    expect((messageCalls[1].messages as unknown[]).length).toBe(1);
+    // Evicted session closed + billed.
+    const keyId = deps.store.listAllKeys().find((k) => k.demo === 1)!.id;
+    const usage = deps.store.listUsage(keyId, 10);
+    expect(usage.some((u) => u.kind === "chat_session" && u.cost_lovelace === "1000000")).toBe(true);
+    expect(deps.store.listOpenSessionsByKey(keyId)).toHaveLength(1); // only session 2
+  });
+
+  it("surfaces the original no-supplier error when there is nothing to evict", async () => {
+    const suppliers = [{ ...SUPPLIERS[1], status: "working" }];
+    const { fetchFn, endCalls } = demoFetch([() => new Response("unused", { status: 500 })], () => suppliers);
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const { rawKey, startChat } = setupDemoKey(deps);
+    const res = await request(createApp(deps))
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "user", content: "hi" }] });
+    expect(res.status).toBe(404);
+    expect(res.body.error.code).toBe("model_not_found");
+    expect(startChat).not.toHaveBeenCalled();
+    expect(endCalls).toHaveLength(0);
+  });
+
+  it("ticket mode: session close settles nothing and bills zero", async () => {
+    const { fetchFn } = demoFetch([
+      () => supplierSse([{ type: "token", value: "Hi" }, doneFrame({ role: "assistant", content: "Hi" }, "stop")]),
+    ]);
+    const deps = makeDeps(fetchFn, { chain: fundedChain, chatSettleMode: "ticket" });
+    const { rawKey, keyRow } = setupDemoKey(deps);
+    const ctx = deps.registry.getContext(keyRow);
+    const endChat = vi.fn(async (opts: { escrowRef: unknown }) => ({
+      settleMode: "ticket" as const,
+      escrowRef: opts.escrowRef,
+    }));
+    (ctx.sdk as { endChat: unknown }).endChat = endChat;
+    const app = createApp(deps);
+
+    const res1 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "user", content: "hi" }] });
+    expect(res1.status).toBe(200);
+
+    const session = deps.store.listOpenSessionsByKey(keyRow.id)[0];
+    const close = await request(app)
+      .post(`/openai/v1/chat/sessions/${session.id}/close`)
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({});
+    expect(close.status).toBe(200);
+    expect(close.body.settle_mode).toBe("ticket");
+    expect(close.body.accepted_ref).toBeUndefined();
+    expect(endChat).toHaveBeenCalledTimes(1);
+
+    const usage = deps.store.listUsage(keyRow.id, 10);
+    expect(usage.some((u) => u.kind === "chat_session" && u.cost_lovelace === "0")).toBe(true);
+  });
+
+  it("ticket mode: janitor-closed demo sessions bill zero", async () => {
+    const { fetchFn, endCalls } = demoFetch([
+      () => supplierSse([{ type: "token", value: "Hi" }, doneFrame({ role: "assistant", content: "Hi" }, "stop")]),
+    ]);
+    const deps = makeDeps(fetchFn, { chain: fundedChain, chatSettleMode: "ticket" });
+    const { rawKey, keyRow } = setupDemoKey(deps);
+    const app = createApp(deps);
+
+    const res = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "user", content: "hi" }] });
+    expect(res.status).toBe(200);
+
+    await sweepIdleDemoSessions(deps, Date.now() + deps.config.demoSessionIdleMs + 1);
+    expect(endCalls.length).toBeGreaterThan(0);
+    const usage = deps.store.listUsage(keyRow.id, 10);
+    expect(usage.some((u) => u.kind === "chat_session" && u.cost_lovelace === "0")).toBe(true);
+  });
+
+  it("janitor closes idle demo sessions so the next request opens fresh", async () => {
+    const { fetchFn, messageCalls, endCalls } = demoFetch([
+      () => supplierSse([{ type: "token", value: "Hi" }, doneFrame({ role: "assistant", content: "Hi" }, "stop")]),
+      () => supplierSse([{ type: "token", value: "Fresh" }, doneFrame({ role: "assistant", content: "Fresh" }, "stop")]),
+    ]);
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const { rawKey, startChat } = setupDemoKey(deps);
+    const app = createApp(deps);
+
+    const res1 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ model: "kimi", messages: [{ role: "user", content: "hi" }] });
+    expect(res1.status).toBe(200);
+
+    // Pretend demoSessionIdleMs elapsed: the sweeper tick closes the session.
+    await sweepIdleDemoSessions(deps, Date.now() + deps.config.demoSessionIdleMs + 1);
+    expect(endCalls).toHaveLength(1);
+    const keyId = deps.store.listAllKeys().find((k) => k.demo === 1)!.id;
+    expect(deps.store.listOpenSessionsByKey(keyId)).toHaveLength(0);
+
+    // Continuing the thread now misses affinity and reopens with full replay.
+    const res2 = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({
+        model: "kimi",
+        messages: [
+          { role: "user", content: "hi" },
+          { role: "assistant", content: "Hi" },
+          { role: "user", content: "more" },
+        ],
+      });
+    expect(res2.status).toBe(200);
+    expect(startChat).toHaveBeenCalledTimes(2);
+    expect((messageCalls[1].messages as unknown[]).length).toBe(3);
+  });
+
+  it("blocks withdraw and hides the deposit address for demo keys", async () => {
+    const deps = makeDeps();
+    const { rawKey } = setupDemoKey(deps);
+    const app = createApp(deps);
+
+    const wd = await request(app)
+      .post("/account/withdraw")
+      .set("authorization", `Bearer ${rawKey}`)
+      .send({ to_address: "addr_test1abc" });
+    expect(wd.status).toBe(403);
+    expect(wd.body.error.code).toBe("demo_key_restricted");
+
+    const acct = await request(app).get("/account").set("authorization", `Bearer ${rawKey}`);
+    expect(acct.status).toBe(200);
+    expect(acct.body.deposit_address).toBeNull();
+    expect(acct.body.demo).toBe(true);
+  });
+
+  it("lists only chat-capable models for demo keys", async () => {
+    const fetchFn = (async (url: unknown) =>
+      new Response(JSON.stringify(String(url).includes("/suppliers") ? SUPPLIERS.filter((s) => !String(url).includes("capability_id") || s.capability_id === "llm.chat.v1") : []), { status: 200 })) as unknown as typeof globalThis.fetch;
+    const deps = makeDeps(fetchFn);
+    const { rawKey } = setupDemoKey(deps);
+    const app = createApp(deps);
+
+    const demoModels = await request(app).get("/openai/v1/models").set("authorization", `Bearer ${rawKey}`);
+    expect(demoModels.body.data.map((m: { id: string }) => m.id)).toEqual(["kimi"]);
+
+    const signup = await request(app).post("/signup").send({});
+    const normalModels = await request(app).get("/openai/v1/models").set("authorization", `Bearer ${signup.body.api_key as string}`);
+    expect(normalModels.body.data.map((m: { id: string }) => m.id).sort()).toEqual(["kimi", "qwen"]);
+  });
+
+  it("rate-limits demo requests per IP (and not per key)", async () => {
+    const deps = makeDeps(undefined, { demoIpRate: { max: 1, windowMs: 60_000 } });
+    const { rawKey } = setupDemoKey(deps);
+    const app = createApp(deps);
+
+    // Model with no chat suppliers → cheap 404 AFTER passing the limiter.
+    const body = { model: "nope", messages: [{ role: "user", content: "hi" }] };
+    const first = await request(app).post("/openai/v1/chat/completions").set("authorization", `Bearer ${rawKey}`).send(body);
+    expect(first.status).toBe(404);
+    const second = await request(app).post("/openai/v1/chat/completions").set("authorization", `Bearer ${rawKey}`).send(body);
+    expect(second.status).toBe(429);
+
+    // Non-demo keys are untouched by the demo IP limiter.
+    const signup = await request(app).post("/signup").send({});
+    const normal = await request(app)
+      .post("/openai/v1/chat/completions")
+      .set("authorization", `Bearer ${signup.body.api_key as string}`)
+      .send(body);
+    expect(normal.status).not.toBe(429);
+  });
+});

gateway/test/migration.test.ts

@@ -0,0 +1,79 @@
+/**
+ * migration.test.ts — the demo-column migration on pre-existing databases.
+ *
+ * GatewayStore has no migration framework (pure CREATE TABLE IF NOT EXISTS),
+ * so opening a DB created before the `demo` column must ALTER it in place.
+ */
+
+import { describe, it, expect } from "vitest";
+import { tmpdir } from "os";
+import { join } from "path";
+import { mkdirSync } from "fs";
+import { randomUUID } from "crypto";
+import Database from "better-sqlite3";
+import { GatewayStore } from "../src/db/store.js";
+
+const LEGACY_API_KEYS_SQL = `
+CREATE TABLE IF NOT EXISTS api_keys (
+  id                 TEXT PRIMARY KEY,
+  key_hash           TEXT NOT NULL UNIQUE,
+  key_prefix         TEXT NOT NULL,
+  label              TEXT,
+  wallet_pkh         TEXT NOT NULL,
+  deposit_address    TEXT NOT NULL,
+  enc_priv_nonce     TEXT NOT NULL,
+  enc_priv_ct        TEXT NOT NULL,
+  enc_priv_tag       TEXT NOT NULL,
+  master_key_version INTEGER NOT NULL DEFAULT 1,
+  created_at         INTEGER NOT NULL,
+  disabled           INTEGER NOT NULL DEFAULT 0
+);
+`;
+
+function keyRow(id: string) {
+  return {
+    id,
+    key_hash: `hash-${id}`,
+    key_prefix: "vk_test_xxxx",
+    label: null,
+    wallet_pkh: "pkh",
+    deposit_address: "addr_test1x",
+    enc_priv_nonce: "n",
+    enc_priv_ct: "c",
+    enc_priv_tag: "t",
+    master_key_version: 1,
+    created_at: Date.now(),
+  };
+}
+
+describe("GatewayStore demo-column migration", () => {
+  it("adds the demo column to a legacy DB and defaults existing keys to 0", () => {
+    const dir = join(tmpdir(), `gw-mig-${randomUUID()}`);
+    mkdirSync(dir, { recursive: true });
+    const raw = new Database(join(dir, "gateway.db"));
+    raw.exec(LEGACY_API_KEYS_SQL);
+    raw
+      .prepare(
+        `INSERT INTO api_keys (id, key_hash, key_prefix, label, wallet_pkh, deposit_address,
+         enc_priv_nonce, enc_priv_ct, enc_priv_tag, master_key_version, created_at, disabled)
+         VALUES (@id, @key_hash, @key_prefix, @label, @wallet_pkh, @deposit_address,
+         @enc_priv_nonce, @enc_priv_ct, @enc_priv_tag, @master_key_version, @created_at, 0)`,
+      )
+      .run(keyRow("legacy"));
+    raw.close();
+
+    const store = new GatewayStore(dir);
+    const legacy = store.getKeyByHash("hash-legacy");
+    expect(legacy?.demo).toBe(0);
+
+    store.insertKey({ ...keyRow("demo-key"), demo: 1 });
+    expect(store.getKeyByHash("hash-demo-key")?.demo).toBe(1);
+  });
+
+  it("persists the demo flag on fresh databases (default 0)", () => {
+    const dir = join(tmpdir(), `gw-fresh-${randomUUID()}`);
+    const store = new GatewayStore(dir);
+    store.insertKey(keyRow("plain"));
+    expect(store.getKeyByHash("hash-plain")?.demo).toBe(0);
+  });
+});

gateway/test/ocr.test.ts

@@ -0,0 +1,572 @@
+/**
+ * gateway/test/ocr.test.ts — POST /v1/ocr/extract route coverage.
+ *
+ * Focus: auth gating, demo-key rejection, body validation, capability-based
+ * supplier lookup, and the enlarged body limit. The settle path (submitOcr →
+ * resolveSubmittedRef → acceptAndConfirm) reuses the one-shot chat plumbing
+ * covered by http.test.ts + the SDK suites; end-to-end runs live on testnet.
+ */
+
+import { describe, it, expect, vi } from "vitest";
+import { tmpdir } from "os";
+import { join } from "path";
+import { randomBytes, randomUUID } from "crypto";
+import request from "supertest";
+import { createApp } from "../src/server.js";
+import { GatewayStore } from "../src/db/store.js";
+import { SdkRegistry } from "../src/sdk/registry.js";
+import { genPrivKeyHex, deriveWalletKey } from "../src/wallet.js";
+import { seal } from "../src/crypto/seal.js";
+import { hashApiKey } from "../src/middleware/apiKeyAuth.js";
+import { venueAttribution } from "../src/ocr/extract.js";
+import type { GatewayConfig } from "../src/config.js";
+import type { GatewayDeps } from "../src/deps.js";
+import { SupplierError } from "@marketplace/buyer/sdk";
+
+// The one-shot settle step (resolveSubmittedRef → acceptAndConfirm) talks to
+// the live indexer + chain; the happy-path route test below only needs to
+// prove that a successful submitOcr reaches the response with the
+// first-party-brokered disclosure label (I3), so it stubs just those two
+// chain-touching steps and keeps every other export real.
+vi.mock("../src/onchain/settle.js", async (importOriginal) => {
+  const actual = await importOriginal<typeof import("../src/onchain/settle.js")>();
+  return {
+    ...actual,
+    resolveSubmittedRef: vi.fn(async () => ({ txHash: "e".repeat(64), index: 0 })),
+    acceptAndConfirm: vi.fn(async () => "f".repeat(64)),
+  };
+});
+
+const MASTER = "ab".repeat(32);
+const IMG = Buffer.from("page bytes").toString("base64");
+
+/** A funded wallet: covers price+bonds+collateral+fee and has pure-ADA ≥ 5. */
+const fundedChain = {
+  queryUtxosByAddress: async () => [
+    { ref: { txHash: "ff".repeat(32), index: 0 }, address: "a", lovelace: 20_000_000n, assets: {}, datumHex: null, scriptRef: null },
+  ],
+} as unknown as GatewayDeps["chain"];
+
+const OCR_SUPPLIER = {
+  utxo_ref: `${"cc".repeat(32)}#0`,
+  supplier_pkh: "s3",
+  capability_id: "ocr.page.extract.chandra-ocr-2.v1",
+  model: "chandra-ocr-2",
+  max_output_tokens: 4096,
+  price_lovelace: "200000",
+  supplier_bond_lovelace: "1000000",
+  buyer_bond_lovelace: "1000000",
+  endpoint_url: "http://sup",
+  advert_status: "Active",
+  status: "free",
+};
+
+function makeDeps(fetchFn?: typeof globalThis.fetch, overrides?: { chain?: GatewayDeps["chain"] }): GatewayDeps {
+  const dbDir = join(tmpdir(), `gw-ocr-test-${randomUUID()}`);
+  const store = new GatewayStore(dbDir);
+  const config: GatewayConfig = {
+    masterKeyHex: MASTER,
+    indexerUrl: "http://ix",
+    ogmiosUrl: "http://og",
+    networkId: 0,
+    liveChain: true,
+    port: 0,
+    dbDir,
+    signupRate: { max: 1000, windowMs: 60_000 },
+    keyRate: { max: 1000, windowMs: 60_000 },
+    demoIpRate: { max: 1000, windowMs: 60_000 },
+    sweeperIntervalMs: 60_000,
+    demoSessionIdleMs: 180_000,
+    chatSettleMode: "full",
+    walletHealthIntervalMs: 600_000,
+    sdkRegistryMax: 100,
+    ocrCapabilityId: "ocr.page.extract.chandra-ocr-2.v1",
+    corsOrigins: [],
+  };
+  const chain = overrides?.chain ?? ({ queryUtxosByAddress: async () => [] } as unknown as GatewayDeps["chain"]);
+  const registry = new SdkRegistry({
+    chain,
+    indexerUrl: config.indexerUrl,
+    networkId: config.networkId,
+    masterKeyHex: config.masterKeyHex,
+    max: config.sdkRegistryMax,
+  });
+  const defaultFetch = (async () => new Response("[]", { status: 200 })) as unknown as typeof globalThis.fetch;
+  return { config, store, chain, registry, fetchFn: fetchFn ?? defaultFetch };
+}
+
+/** Insert a key row (demo or funded) and return the raw bearer token. */
+function makeKey(deps: GatewayDeps, demo: 0 | 1): string {
+  const rawKey = `vk-${randomBytes(24).toString("hex")}`;
+  const privHex = genPrivKeyHex();
+  const walletKey = deriveWalletKey(privHex, 0);
+  const sealed = seal(privHex, MASTER);
+  deps.store.insertKey({
+    id: randomUUID(),
+    key_hash: hashApiKey(rawKey),
+    key_prefix: rawKey.slice(0, 12),
+    label: demo ? "demo" : "funded",
+    wallet_pkh: walletKey.pubKeyHash,
+    deposit_address: walletKey.address,
+    enc_priv_nonce: sealed.nonce,
+    enc_priv_ct: sealed.ct,
+    enc_priv_tag: sealed.tag,
+    master_key_version: 1,
+    created_at: Date.now(),
+    demo,
+  });
+  return rawKey;
+}
+
+function validBody() {
+  return { image_b64: IMG, mime: "image/png", output_format: "markdown" };
+}
+
+describe("POST /v1/ocr/extract — auth + demo gating", () => {
+  it("401 without a bearer key, OpenAI error shape", async () => {
+    const res = await request(createApp(makeDeps()))
+      .post("/v1/ocr/extract").send(validBody());
+    expect(res.status).toBe(401);
+    expect(res.body.error?.type).toBe("authentication_error");
+  });
+
+  it("403 demo_key_unsupported for demo keys", async () => {
+    const deps = makeDeps();
+    const rawKey = makeKey(deps, 1);
+    const res = await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(validBody());
+    expect(res.status).toBe(403);
+    expect(res.body.error?.code).toBe("demo_key_unsupported");
+  });
+});
+
+describe("POST /v1/ocr/extract — body validation", () => {
+  async function post(body: unknown) {
+    const deps = makeDeps();
+    const rawKey = makeKey(deps, 0);
+    return request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(body as object);
+  }
+
+  it("400 invalid_image on missing/empty/data-URL image_b64", async () => {
+    for (const image_b64 of [undefined, "", `data:image/png;base64,${IMG}`]) {
+      const res = await post({ ...validBody(), image_b64 });
+      expect(res.status).toBe(400);
+      expect(res.body.error?.code).toBe("invalid_image");
+    }
+  });
+
+  it("400 invalid_mime on unsupported mime", async () => {
+    const res = await post({ ...validBody(), mime: "image/gif" });
+    expect(res.status).toBe(400);
+    expect(res.body.error?.code).toBe("invalid_mime");
+  });
+
+  it("400 invalid_output_format on unsupported format", async () => {
+    const res = await post({ ...validBody(), output_format: "yaml" });
+    expect(res.status).toBe(400);
+    expect(res.body.error?.code).toBe("invalid_output_format");
+  });
+
+  it("output_format defaults to markdown (reaches supplier lookup)", async () => {
+    const res = await post({ image_b64: IMG, mime: "image/png" });
+    // Passes validation, then the empty indexer answer → 404 model_not_found.
+    expect(res.status).toBe(404);
+    expect(res.body.error?.code).toBe("model_not_found");
+  });
+});
+
+describe("POST /v1/ocr/extract — supplier lookup + body limit", () => {
+  it("404 model_not_found when no supplier advertises the capability", async () => {
+    const deps = makeDeps();
+    const rawKey = makeKey(deps, 0);
+    const res = await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(validBody());
+    expect(res.status).toBe(404);
+    expect(res.body.error?.code).toBe("model_not_found");
+  });
+
+  it("queries the indexer with the configured ocr capability id", async () => {
+    let queried = "";
+    const fetchFn = (async (url: unknown) => {
+      queried = String(url);
+      return new Response("[]", { status: 200 });
+    }) as unknown as typeof globalThis.fetch;
+    const deps = makeDeps(fetchFn);
+    const rawKey = makeKey(deps, 0);
+    await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(validBody());
+    expect(queried).toContain("capability_id=ocr.page.extract.chandra-ocr-2.v1");
+  });
+
+  it("accepts a 2MB body (over the 1mb global limit — OCR route has its own parser)", async () => {
+    const deps = makeDeps();
+    const rawKey = makeKey(deps, 0);
+    const big = "A".repeat(2_000_000);
+    const res = await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send({ ...validBody(), image_b64: big });
+    expect(res.status).not.toBe(413);
+    // Valid base64 + no suppliers → 404 model_not_found.
+    expect(res.status).toBe(404);
+  });
+
+  it("chat completions keep the 1mb ceiling (413 on a 2MB body)", async () => {
+    const deps = makeDeps();
+    const rawKey = makeKey(deps, 0);
+    const big = "A".repeat(2_000_000);
+    const res = await request(createApp(deps))
+      .post("/openai/v1/chat/completions")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send({ model: "qwen", messages: [{ role: "user", content: big }] });
+    expect(res.status).toBe(413);
+  });
+});
+
+// ─── venueAttribution — pure-function header sanitization (I3) ──────────────
+//
+// grep for "first-party-brokered" used to return only source lines: nothing
+// exercised venueAttribution's regex gates or RECEIPT_LABELS_V0 reaching the
+// wire. These two blocks close that gap.
+
+function fakeReq(headers: Record<string, string>): Parameters<typeof venueAttribution>[0] {
+  return { header: (name: string) => headers[name] } as unknown as Parameters<typeof venueAttribution>[0];
+}
+
+describe("venueAttribution — header sanitization", () => {
+  it("passes through a valid venue slug and a valid lowercase hash", () => {
+    const hash = "a".repeat(64);
+    const out = venueAttribution(fakeReq({ "X-Venue": "apify", "X-Run-User-Hash": hash }));
+    expect(out).toEqual({ venue: "apify", runUserHash: hash });
+  });
+
+  it("lowercases a valid but uppercase hex hash", () => {
+    const out = venueAttribution(fakeReq({ "X-Venue": "apify", "X-Run-User-Hash": "A".repeat(64) }));
+    expect(out.runUserHash).toBe("a".repeat(64));
+  });
+
+  it("drops an invalid venue to empty string (uppercase, bad chars, oversized)", () => {
+    for (const bad of ["Apify", "apify!", "apify_venue", "a".repeat(33)]) {
+      const out = venueAttribution(fakeReq({ "X-Venue": bad, "X-Run-User-Hash": "a".repeat(64) }));
+      expect(out.venue).toBe("");
+    }
+  });
+
+  it("drops an invalid run-user-hash to empty string (wrong length, non-hex)", () => {
+    for (const bad of ["a".repeat(63), "a".repeat(65), "z".repeat(64), "not-hex-at-all"]) {
+      const out = venueAttribution(fakeReq({ "X-Venue": "apify", "X-Run-User-Hash": bad }));
+      expect(out.runUserHash).toBe("");
+    }
+  });
+
+  it("defaults both fields to empty string when headers are missing", () => {
+    const out = venueAttribution(fakeReq({}));
+    expect(out).toEqual({ venue: "", runUserHash: "" });
+  });
+});
+
+// ─── Happy path — receipt disclosure label reaches the response (I3) ────────
+
+describe("POST /v1/ocr/extract — happy path", () => {
+  it("200 response x_vector.receipt_labels contains first-party-brokered", async () => {
+    const fetchFn = (async (url: unknown) => {
+      const u = String(url);
+      if (u.includes("/suppliers")) {
+        return new Response(JSON.stringify([OCR_SUPPLIER]), { status: 200 });
+      }
+      return new Response("[]", { status: 200 });
+    }) as unknown as typeof globalThis.fetch;
+
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const rawKey = makeKey(deps, 0);
+    const keyRow = deps.store.getKeyByHash(hashApiKey(rawKey))!;
+
+    // Stub only the chain-touching supplier call; escrow ref shape matches
+    // what resolveSubmittedRef's mock (module-level, top of file) expects.
+    const ctx = deps.registry.getContext(keyRow);
+    const submitOcr = vi.fn(async () => ({
+      output_format: "markdown",
+      content: "extracted text",
+      usage: { prompt_tokens: 10, completion_tokens: 5, total_tokens: 15 },
+      receipt: {
+        prompt_hash: "1b".repeat(32),
+        response_hash: "2c".repeat(32),
+        model: "chandra-ocr-2",
+        prompt_tokens: 10,
+        completion_tokens: 5,
+        wallclock_ms: Date.now(),
+        supplier_pkh: "s3",
+        escrow_ref: `${"d".repeat(64)}#0`,
+      },
+      receiptSignature: "sig".repeat(20),
+      escrowRef: { txHash: "d".repeat(64), index: 0 },
+    }));
+    (ctx.sdk as { submitOcr: unknown }).submitOcr = submitOcr;
+
+    const res = await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(validBody());
+
+    expect(res.status).toBe(200);
+    expect(submitOcr).toHaveBeenCalledTimes(1);
+    expect(res.body.x_vector.receipt_labels).toContain("first-party-brokered");
+    // escrow_ref in the response comes from submitOcr's returned escrowRef
+    // (the ORIGINAL Open ref), not the mocked resolveSubmittedRef's answer —
+    // see settle.ts's docstring on why those two refs differ on-chain.
+    expect(res.body.x_vector.escrow_ref).toBe(`${"d".repeat(64)}#0`);
+  });
+});
+
+// ─── Failure path — escrow ref survives a post-escrow SupplierError (I2) ────
+//
+// submitOcr posts the escrow, THEN calls the supplier. A SupplierError
+// (timeout, network failure, non-2xx, malformed response, receipt mismatch)
+// means the escrow already landed on chain even though submitOcr rejects —
+// per the OCR route's 120s supplier budget, this is likely the MOST common
+// paid-but-failed shape. Before this fix, escrowRefStr was only assigned
+// from submitOcr's *return value*, so this exact case recorded escrow_ref:
+// null. The route now also listens for the SDK's "escrow_posted" progress
+// event around each attempt, so the ref survives regardless of whether
+// submitOcr goes on to resolve or reject.
+
+describe("POST /v1/ocr/extract — escrow ref survives a post-escrow SupplierError", () => {
+  it("records the escrow_ref from the escrow_posted progress event, not null, when submitOcr posts an escrow then rejects", async () => {
+    const fetchFn = (async (url: unknown) => {
+      const u = String(url);
+      if (u.includes("/suppliers")) {
+        return new Response(JSON.stringify([OCR_SUPPLIER]), { status: 200 });
+      }
+      return new Response("[]", { status: 200 });
+    }) as unknown as typeof globalThis.fetch;
+
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const rawKey = makeKey(deps, 0);
+    const keyRow = deps.store.getKeyByHash(hashApiKey(rawKey))!;
+
+    // Same monkeypatch pattern as the happy-path test above: fetch the real
+    // per-key Marketplace instance from the registry BEFORE the request so
+    // its real EventEmitter (on/off/emit) stays live, then replace only
+    // submitOcr. The stub mirrors what Marketplace.submitOcr actually does
+    // on this failure shape: emit "escrow_posted" the instant the tx
+    // confirms, then throw once the supplier call fails.
+    const ctx = deps.registry.getContext(keyRow);
+    const knownEscrowRef = `${"9".repeat(64)}#0`;
+    const submitOcr = vi.fn(async () => {
+      ctx.sdk.emit("progress", { type: "escrow_posted", escrow_ref: knownEscrowRef });
+      throw new SupplierError("timeout", { message: "supplier did not respond within budget" });
+    });
+    (ctx.sdk as { submitOcr: unknown }).submitOcr = submitOcr;
+
+    const logSpy = vi.spyOn(console, "log").mockImplementation(() => {});
+    const res = await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(validBody());
+    // Snapshot the recorded calls BEFORE restoring — mockRestore() also
+    // clears .mock.calls (it runs mockReset()/mockClear() first), so
+    // reading logSpy.mock.calls after restore always sees [].
+    const loggedLines = logSpy.mock.calls.map((args) => String(args[0]));
+    logSpy.mockRestore();
+
+    expect(submitOcr).toHaveBeenCalledTimes(1);
+    // toGatewayError maps SupplierError("timeout") to 504 supplier_timeout —
+    // confirms the request actually took the failure path this test targets.
+    expect(res.status).toBe(504);
+    expect(res.body.error?.code).toBe("supplier_timeout");
+
+    // The reconciliation spine: the usage row recorded in the catch block
+    // must carry the real ref, not null.
+    const rows = deps.store.listUsage(keyRow.id, 10);
+    expect(rows).toHaveLength(1);
+    expect(rows[0].status).toBe("failed");
+    expect(rows[0].escrow_ref).toBe(knownEscrowRef);
+
+    // The structured ocr_failed log line (wedge-ledger emitter input) must
+    // carry the same ref.
+    const failedLine = loggedLines
+      .filter((line) => line.includes('"evt":"ocr_failed"'))
+      .map((line) => JSON.parse(line) as { escrow_ref: string | null });
+    expect(failedLine).toHaveLength(1);
+    expect(failedLine[0].escrow_ref).toBe(knownEscrowRef);
+  });
+});
+
+// ─── Busy pre-check - never post an escrow to a candidate that is busy ──────
+//
+// runOcrOneShot now checks each candidate's live /status immediately before
+// calling submitOcr for it (mirrors sessions.ts's startChat pre-check):
+// "working" means busy and moves on to the next candidate WITHOUT posting an
+// escrow; anything else (free, unreachable, malformed, slow) is treated as
+// usable. submitOcr is the escrow-posting path in these tests (same
+// monkeypatch pattern as the suites above), so asserting its call count is
+// exactly how "no escrow was posted" is observed.
+
+const OCR_SUPPLIER_A = {
+  ...OCR_SUPPLIER,
+  utxo_ref: `${"aa".repeat(32)}#0`,
+  supplier_pkh: "sA",
+  endpoint_url: "http://sup-a",
+};
+const OCR_SUPPLIER_B = {
+  ...OCR_SUPPLIER,
+  utxo_ref: `${"bb".repeat(32)}#0`,
+  supplier_pkh: "sB",
+  endpoint_url: "http://sup-b",
+};
+
+/** Indexer /suppliers lookup plus a per-endpoint /status responder, on one
+ * fetchFn - everything else falls back to the suite's usual empty array. */
+function makeFetchWithStatus(
+  suppliers: Array<typeof OCR_SUPPLIER>,
+  statusByEndpoint: Record<string, () => Response>,
+): typeof globalThis.fetch {
+  return (async (url: unknown) => {
+    const u = String(url);
+    if (u.includes("/suppliers")) {
+      return new Response(JSON.stringify(suppliers), { status: 200 });
+    }
+    for (const [endpoint, respond] of Object.entries(statusByEndpoint)) {
+      if (u === `${endpoint}/status`) return respond();
+    }
+    return new Response("[]", { status: 200 });
+  }) as unknown as typeof globalThis.fetch;
+}
+
+/** Same successful submitOcr shape as the happy-path suite above, reused
+ * across the busy pre-check tests. */
+function makeSubmitOcrStub() {
+  return vi.fn(async (_opts: { advertRef: { txHash: string; index: number } }) => ({
+    output_format: "markdown",
+    content: "extracted text",
+    usage: { prompt_tokens: 10, completion_tokens: 5, total_tokens: 15 },
+    receipt: {
+      prompt_hash: "1b".repeat(32),
+      response_hash: "2c".repeat(32),
+      model: "chandra-ocr-2",
+      prompt_tokens: 10,
+      completion_tokens: 5,
+      wallclock_ms: Date.now(),
+      supplier_pkh: "s3",
+      escrow_ref: `${"d".repeat(64)}#0`,
+    },
+    receiptSignature: "sig".repeat(20),
+    escrowRef: { txHash: "d".repeat(64), index: 0 },
+  }));
+}
+
+describe("POST /v1/ocr/extract - busy pre-check", () => {
+  it("a busy supplier is never sent an escrow; the route returns the busy response", async () => {
+    const fetchFn = makeFetchWithStatus([OCR_SUPPLIER], {
+      "http://sup": () => new Response(JSON.stringify({ status: "working" }), { status: 200 }),
+    });
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const rawKey = makeKey(deps, 0);
+    const keyRow = deps.store.getKeyByHash(hashApiKey(rawKey))!;
+
+    const ctx = deps.registry.getContext(keyRow);
+    const submitOcr = makeSubmitOcrStub();
+    (ctx.sdk as { submitOcr: unknown }).submitOcr = submitOcr;
+
+    const res = await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(validBody());
+
+    // The escrow-posting path was never invoked.
+    expect(submitOcr).not.toHaveBeenCalled();
+    // toGatewayError maps SupplierError("supplier_busy") to 503 overloaded -
+    // the same busy response a post-escrow 409 would have produced.
+    expect(res.status).toBe(503);
+    expect(res.body.error?.code).toBe("overloaded");
+
+    const rows = deps.store.listUsage(keyRow.id, 10);
+    expect(rows).toHaveLength(1);
+    expect(rows[0].status).toBe("failed");
+    expect(rows[0].escrow_ref).toBeNull();
+  });
+
+  it("a free supplier proceeds to post and settle exactly as before", async () => {
+    const fetchFn = makeFetchWithStatus([OCR_SUPPLIER], {
+      "http://sup": () => new Response(JSON.stringify({ status: "free" }), { status: 200 }),
+    });
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const rawKey = makeKey(deps, 0);
+    const keyRow = deps.store.getKeyByHash(hashApiKey(rawKey))!;
+
+    const ctx = deps.registry.getContext(keyRow);
+    const submitOcr = makeSubmitOcrStub();
+    (ctx.sdk as { submitOcr: unknown }).submitOcr = submitOcr;
+
+    const res = await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(validBody());
+
+    expect(submitOcr).toHaveBeenCalledTimes(1);
+    expect(res.status).toBe(200);
+    expect(res.body.x_vector.receipt_labels).toContain("first-party-brokered");
+  });
+
+  it("falls back to the second candidate when the first is busy, posting exactly one escrow", async () => {
+    const fetchFn = makeFetchWithStatus([OCR_SUPPLIER_A, OCR_SUPPLIER_B], {
+      "http://sup-a": () => new Response(JSON.stringify({ status: "working" }), { status: 200 }),
+      "http://sup-b": () => new Response(JSON.stringify({ status: "free" }), { status: 200 }),
+    });
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const rawKey = makeKey(deps, 0);
+    const keyRow = deps.store.getKeyByHash(hashApiKey(rawKey))!;
+
+    const ctx = deps.registry.getContext(keyRow);
+    const submitOcr = makeSubmitOcrStub();
+    (ctx.sdk as { submitOcr: unknown }).submitOcr = submitOcr;
+
+    const res = await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(validBody());
+
+    expect(res.status).toBe(200);
+    // Exactly one escrow posted, and it is the second (free) candidate's -
+    // the busy first candidate was skipped, not retried after failing.
+    expect(submitOcr).toHaveBeenCalledTimes(1);
+    expect(submitOcr.mock.calls[0][0].advertRef).toEqual({ txHash: "bb".repeat(32), index: 0 });
+  });
+
+  it("a status check that errors does not block the attempt", async () => {
+    const fetchFn = makeFetchWithStatus([OCR_SUPPLIER], {
+      "http://sup": () => {
+        // Simulates both a network error and an aborted (timed-out) request -
+        // isSupplierBusy's catch block treats every failure mode the same
+        // way: unknown, not busy, proceed. A real AbortSignal timeout is not
+        // exercised here to keep the suite fast; the outcome is identical.
+        throw new Error("simulated network failure reaching /status");
+      },
+    });
+    const deps = makeDeps(fetchFn, { chain: fundedChain });
+    const rawKey = makeKey(deps, 0);
+    const keyRow = deps.store.getKeyByHash(hashApiKey(rawKey))!;
+
+    const ctx = deps.registry.getContext(keyRow);
+    const submitOcr = makeSubmitOcrStub();
+    (ctx.sdk as { submitOcr: unknown }).submitOcr = submitOcr;
+
+    const res = await request(createApp(deps))
+      .post("/v1/ocr/extract")
+      .set("Authorization", `Bearer ${rawKey}`)
+      .send(validBody());
+
+    // The candidate was still tried despite the failed pre-check.
+    expect(submitOcr).toHaveBeenCalledTimes(1);
+    expect(res.status).toBe(200);
+  });
+});

gateway/test/sweeper.test.ts

@@ -0,0 +1,187 @@
+/**
+ * sweeper.test.ts — reclaim/close decoupling + settled-refs memory.
+ *
+ * Ticket mode: reclaiming a live session's Open escrow must NOT close the
+ * session row or drop its transcript mirror (the conversation continues on a
+ * spent ticket). Full mode keeps close-on-reclaim. Claimed rows close even
+ * under ticket config (they can only come from full-mode suppliers). The
+ * settled-refs memory prevents re-attempting refs already settled, and the
+ * deliver_by gate skips reclaims the validator would reject.
+ */
+
+import { describe, it, expect, vi } from "vitest";
+import { tmpdir } from "os";
+import { join } from "path";
+import { randomBytes, randomUUID } from "crypto";
+import { GatewayStore } from "../src/db/store.js";
+import { SdkRegistry } from "../src/sdk/registry.js";
+import { genPrivKeyHex, deriveWalletKey } from "../src/wallet.js";
+import { seal } from "../src/crypto/seal.js";
+import { hashApiKey } from "../src/middleware/apiKeyAuth.js";
+import { runSweepOnce } from "../src/sweeper.js";
+import { transcripts } from "../src/openai/transcripts.js";
+import type { GatewayConfig } from "../src/config.js";
+import type { GatewayDeps } from "../src/deps.js";
+import type { EscrowRow } from "../src/onchain/settle.js";
+
+const MASTER = "ab".repeat(32);
+
+function makeSweepDeps(opts: { chatSettleMode: "full" | "ticket"; escrows: EscrowRow[] }) {
+  const dbDir = join(tmpdir(), `gw-sweep-${randomUUID()}`);
+  const store = new GatewayStore(dbDir);
+  const config: GatewayConfig = {
+    masterKeyHex: MASTER,
+    indexerUrl: "http://ix",
+    ogmiosUrl: "http://og",
+    networkId: 0,
+    liveChain: true,
+    port: 0,
+    dbDir,
+    signupRate: { max: 1000, windowMs: 60_000 },
+    keyRate: { max: 1000, windowMs: 60_000 },
+    demoIpRate: { max: 1000, windowMs: 60_000 },
+    sweeperIntervalMs: 60_000,
+    demoSessionIdleMs: 180_000,
+    chatSettleMode: opts.chatSettleMode,
+    walletHealthIntervalMs: 600_000,
+    sdkRegistryMax: 100,
+    ocrCapabilityId: "ocr.page.extract.chandra-ocr-2.v1",
+    corsOrigins: [],
+  };
+  const chain = { queryUtxo: async () => null } as unknown as GatewayDeps["chain"];
+  const registry = new SdkRegistry({
+    chain,
+    indexerUrl: config.indexerUrl,
+    networkId: config.networkId,
+    masterKeyHex: config.masterKeyHex,
+    max: config.sdkRegistryMax,
+  });
+  const fetchFn = (async (url: unknown) => {
+    if (String(url).includes("/escrows")) return new Response(JSON.stringify(opts.escrows), { status: 200 });
+    return new Response("[]", { status: 200 });
+  }) as unknown as typeof globalThis.fetch;
+  const deps: GatewayDeps = { config, store, chain, registry, fetchFn };
+
+  const rawKey = `vk_test_${randomBytes(24).toString("hex")}`;
+  const privHex = genPrivKeyHex();
+  const walletKey = deriveWalletKey(privHex, 0);
+  const sealed = seal(privHex, MASTER);
+  store.insertKey({
+    id: randomUUID(),
+    key_hash: hashApiKey(rawKey),
+    key_prefix: rawKey.slice(0, 12),
+    label: "sweep-test",
+    wallet_pkh: walletKey.pubKeyHash,
+    deposit_address: walletKey.address,
+    enc_priv_nonce: sealed.nonce,
+    enc_priv_ct: sealed.ct,
+    enc_priv_tag: sealed.tag,
+    master_key_version: 1,
+    created_at: Date.now(),
+    demo: 1,
+  });
+  const keyRow = store.getKeyByHash(hashApiKey(rawKey))!;
+  return { deps, keyRow };
+}
+
+function insertOpenSession(deps: GatewayDeps, keyId: string, escrowRef: string): string {
+  const id = randomUUID();
+  deps.store.insertSession({
+    id,
+    key_id: keyId,
+    escrow_ref: escrowRef,
+    session_nonce: "nonce",
+    supplier_base_url: "http://sup",
+    supplier_pkh: "s1",
+    model: "kimi",
+    price_lovelace: "200000",
+    state: "open",
+    opened_at: Date.now() - 40 * 60 * 1000,
+  });
+  return id;
+}
+
+function openRow(refChar: string, opts?: Partial<EscrowRow>): EscrowRow {
+  return {
+    utxo_ref: `${refChar.repeat(64)}#0`,
+    state: "Open",
+    posted_at: Date.now() - 40 * 60 * 1000,
+    deliver_by: Date.now() - 5 * 60 * 1000,
+    ...opts,
+  };
+}
+
+const settleOk = () => ({
+  accept: vi.fn(async () => undefined) as never,
+  reclaim: vi.fn(async () => undefined) as never,
+});
+
+describe("sweeper — ticket-mode reclaim/close decoupling", () => {
+  it("reclaims a live session's Open ticket but keeps the session (and mirror) alive", async () => {
+    const row = openRow("1");
+    const { deps, keyRow } = makeSweepDeps({ chatSettleMode: "ticket", escrows: [row] });
+    const sessionId = insertOpenSession(deps, keyRow.id, row.utxo_ref);
+    transcripts.set(sessionId, [{ role: "user", content: "hi" }]);
+    const settle = settleOk();
+
+    await runSweepOnce(deps, settle);
+
+    expect(settle.reclaim).toHaveBeenCalledTimes(1);
+    expect(deps.store.getSession(sessionId)!.state).toBe("open"); // NOT closed
+    expect(transcripts.get(sessionId)).toBeDefined(); // mirror intact
+    expect(deps.store.listUsage(keyRow.id, 10)).toHaveLength(0); // nothing billed
+
+    transcripts.delete(sessionId);
+  });
+
+  it("full mode: reclaim closes the session row", async () => {
+    const row = openRow("2");
+    const { deps, keyRow } = makeSweepDeps({ chatSettleMode: "full", escrows: [row] });
+    const sessionId = insertOpenSession(deps, keyRow.id, row.utxo_ref);
+    const settle = settleOk();
+
+    await runSweepOnce(deps, settle);
+
+    expect(settle.reclaim).toHaveBeenCalledTimes(1);
+    expect(deps.store.getSession(sessionId)!.state).toBe("closed");
+    const usage = deps.store.listUsage(keyRow.id, 10);
+    expect(usage.some((u) => u.status === "reclaimed")).toBe(true);
+  });
+
+  it("ticket config still closes sessions behind Claimed rows (full-mode suppliers)", async () => {
+    const row = openRow("3", { state: "Claimed" });
+    const { deps, keyRow } = makeSweepDeps({ chatSettleMode: "ticket", escrows: [row] });
+    const sessionId = insertOpenSession(deps, keyRow.id, row.utxo_ref);
+    const settle = settleOk();
+
+    await runSweepOnce(deps, settle);
+
+    expect(settle.reclaim).toHaveBeenCalledTimes(1);
+    expect(deps.store.getSession(sessionId)!.state).toBe("closed");
+  });
+});
+
+describe("sweeper — reclaim gating + settled-refs memory", () => {
+  it("skips reclaim while deliver_by has not passed", async () => {
+    const row = openRow("4", { deliver_by: Date.now() + 10 * 60 * 1000 });
+    const { deps, keyRow } = makeSweepDeps({ chatSettleMode: "ticket", escrows: [row] });
+    insertOpenSession(deps, keyRow.id, row.utxo_ref);
+    const settle = settleOk();
+
+    await runSweepOnce(deps, settle);
+
+    expect(settle.reclaim).not.toHaveBeenCalled();
+  });
+
+  it("does not re-attempt refs it already settled (in-process memory)", async () => {
+    const row = openRow("5");
+    const { deps, keyRow } = makeSweepDeps({ chatSettleMode: "ticket", escrows: [row] });
+    insertOpenSession(deps, keyRow.id, row.utxo_ref);
+    const settle = settleOk();
+
+    await runSweepOnce(deps, settle);
+    await runSweepOnce(deps, settle);
+
+    expect(settle.reclaim).toHaveBeenCalledTimes(1);
+  });
+});

gateway/test/unit.test.ts

@@ -84,10 +84,31 @@ describe("routing/selectSupplier", () => {
     expect(await listModels({ indexerUrl: "http://ix", fetchFn: jsonResponse(rows) })).toEqual(["m", "other"]);
   });
 
+  it("lists only one capability's models when capabilityId is given", async () => {
+    expect(await listModels({ indexerUrl: "http://ix", capabilityId: "llm.chat.v1", fetchFn: jsonResponse(rows) })).toEqual(["m"]);
+  });
+
   it("parseRef", () => {
     expect(parseRef("aa".repeat(32) + "#3")).toEqual({ txHash: "aa".repeat(32), index: 3 });
     expect(parseRef("nope")).toBeNull();
   });
+
+  it("ignoreStatusFor readmits a busy supplier by pkh (stale indexer status)", async () => {
+    const busy = [
+      { ...rows[0], status: "working" },
+      { ...rows[1], status: "offline" },
+    ];
+    const base = { indexerUrl: "http://ix", model: "m", capabilityId: "llm.text.generate.v1" };
+    const excluded = await selectCandidates({ ...base, fetchFn: jsonResponse(busy) });
+    expect(excluded).toEqual([]);
+
+    const readmitted = await selectCandidates({
+      ...base,
+      fetchFn: jsonResponse(busy),
+      ignoreStatusFor: new Set(["s1"]),
+    });
+    expect(readmitted.map((c) => c.supplierPkh)).toEqual(["s1"]);
+  });
 });
 
 describe("openai/validate", () => {
@@ -104,6 +125,37 @@ describe("openai/validate", () => {
     expect(() => parseChatRequest({ model: "m", messages: [] })).toThrow();
     expect(() => parseChatRequest({ messages: [{ role: "user", content: "x" }] })).toThrow();
   });
+
+  const TOOL_CALL = { id: "call_1", type: "function", function: { name: "get_time", arguments: "{}" } };
+
+  it("allowTools accepts tools, tool role, null content and assistant tool_calls", () => {
+    const p = parseChatRequest(
+      {
+        model: "m",
+        messages: [
+          { role: "user", content: "hi" },
+          { role: "assistant", content: null, tool_calls: [TOOL_CALL] },
+          { role: "tool", content: "3pm", tool_call_id: "call_1" },
+        ],
+        tools: [{ type: "function", function: { name: "get_time", parameters: {} } }],
+        tool_choice: "auto",
+      },
+      { allowTools: true },
+    );
+    expect(p.tools).toHaveLength(1);
+    expect(p.toolChoice).toBe("auto");
+    expect(p.messages[1]).toEqual({ role: "assistant", content: "", tool_calls: [TOOL_CALL] });
+    expect(p.messages[2]).toEqual({ role: "tool", content: "3pm", tool_call_id: "call_1" });
+  });
+
+  it("allowTools still rejects legacy functions and the default path still rejects tool role", () => {
+    expect(() =>
+      parseChatRequest({ model: "m", messages: [{ role: "user", content: "x" }], functions: [] }, { allowTools: true }),
+    ).toThrow(GatewayError);
+    expect(() =>
+      parseChatRequest({ model: "m", messages: [{ role: "tool", content: "x", tool_call_id: "c" }] }),
+    ).toThrow(GatewayError);
+  });
 });
 
 describe("openai/shapes", () => {
@@ -121,6 +173,24 @@ describe("openai/shapes", () => {
     expect(ch.object).toBe("chat.completion.chunk");
     expect((ch.choices as any)[0].delta.content).toBe("tok");
   });
+  it("tool_calls chunk + finish_reason tool_calls", () => {
+    const tc = { index: 0, id: "call_1", type: "function" as const, function: { name: "f", arguments: "{}" } };
+    const ch = buildChunk({ id: "id", model: "m", delta: { tool_calls: [tc] }, finishReason: "tool_calls" });
+    expect((ch.choices as any)[0].delta.tool_calls).toEqual([tc]);
+    expect((ch.choices as any)[0].finish_reason).toBe("tool_calls");
+  });
+  it("completion with tool_calls and finish_reason", () => {
+    const cc = buildChatCompletion({
+      id: "id",
+      model: "m",
+      content: "",
+      usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+      toolCalls: [{ id: "call_1", type: "function", function: { name: "f", arguments: "{}" } }],
+      finishReason: "tool_calls",
+    });
+    expect((cc.choices as any)[0].message.tool_calls).toHaveLength(1);
+    expect((cc.choices as any)[0].finish_reason).toBe("tool_calls");
+  });
   it("renderMessages folds roles", () => {
     expect(renderMessages([{ role: "system", content: "be nice" }, { role: "user", content: "hi" }])).toBe("System: be nice\n\nUser: hi");
   });
@@ -138,8 +208,16 @@ describe("onchain/preflight", () => {
   });
 
   it("preflight requires balance AND collateral", async () => {
-    const cost = { priceLovelace: 2_000_000n, buyerBondLovelace: 1_000_000n, supplierBondLovelace: 1_000_000n };
-    // required = 2 + 1 + 1 + 5 (collateral) + 2 (fee) = 11 ADA
+    const cost = {
+      capabilityId: "llm.text.generate.v1",
+      priceLovelace: 2_000_000n,
+      buyerBondLovelace: 1_000_000n,
+      supplierBondLovelace: 1_000_000n,
+    };
+    // required = escrowLockFloor(2 + 1 + 1) + 5 (collateral) + 2 (fee) = 11 ADA
+    // — at this economic total (4 ADA) the Submitted-state min-ada floor for a
+    // short capability id (~2.2 ADA) stays well under the raw total, so the
+    // floor is a no-op here; tx-minada-floor.test.ts covers the lifted case.
     const fakeChain = { queryUtxosByAddress: async () => [pure(20_000_000n)] } as any;
     const ok = await preflight(fakeChain, "a", cost);
     expect(ok.ok).toBe(true);

indexer/Dockerfile

@@ -29,6 +29,8 @@ RUN apk add --no-cache python3 make g++
 RUN corepack enable && corepack prepare pnpm@10.33.2 --activate
 
 COPY pnpm-workspace.yaml pnpm-lock.yaml package.json tsconfig.base.json .npmrc ./
+# pnpm-lock.yaml's patchedDependencies hash the patch files at install time.
+COPY patches ./patches
 COPY packages/shared/package.json ./packages/shared/package.json
 COPY indexer/package.json ./indexer/package.json
 COPY indexer-ui/package.json ./indexer-ui/package.json

package.json

@@ -37,5 +37,10 @@
     "tsx": "^4.19.2",
     "typescript": "^5.7.2",
     "vitest": "^2.1.8"
+  },
+  "pnpm": {
+    "patchedDependencies": {
+      "libsodium-wrappers-sumo@0.7.16": "patches/libsodium-wrappers-sumo@0.7.16.patch"
+    }
   }
 }

packages/shared/src/tx/chatMessage.ts

@@ -0,0 +1,88 @@
+/**
+ * tx/chatMessage.ts — canonical ChatMessage normalization + equivalence.
+ *
+ * The supplier's session transcript and the gateway's in-memory mirror must be
+ * hash-identical (receipt response_hash = sha256(canonicalize(transcript))),
+ * so both sides build messages through normalizeChatMessage: optional fields
+ * are truly absent when empty, content is never null, and vendor extras are
+ * stripped. chatMessagesEquivalent is the tolerant comparison used by the
+ * demo-endpoint transcript-prefix affinity matcher.
+ */
+
+import type { ChatMessage, ToolCall } from "./types.js";
+
+const ROLES = new Set(["system", "user", "assistant", "tool"]);
+
+function normalizeToolCall(raw: unknown): ToolCall | null {
+  if (typeof raw !== "object" || raw === null) return null;
+  const tc = raw as Record<string, unknown>;
+  const fn = tc.function as Record<string, unknown> | undefined;
+  if (typeof tc.id !== "string" || tc.id === "") return null;
+  if (tc.type !== "function") return null;
+  if (typeof fn !== "object" || fn === null) return null;
+  if (typeof fn.name !== "string" || fn.name === "") return null;
+  if (typeof fn.arguments !== "string") return null;
+  return { id: tc.id, type: "function", function: { name: fn.name, arguments: fn.arguments } };
+}
+
+/**
+ * Validate and canonicalize an untrusted message-shaped value.
+ * Returns null when the value is not a usable ChatMessage.
+ */
+export function normalizeChatMessage(raw: unknown): ChatMessage | null {
+  if (typeof raw !== "object" || raw === null) return null;
+  const m = raw as Record<string, unknown>;
+  if (typeof m.role !== "string" || !ROLES.has(m.role)) return null;
+  const role = m.role as ChatMessage["role"];
+
+  let content: string;
+  if (typeof m.content === "string") content = m.content;
+  else if (m.content === null || m.content === undefined) content = "";
+  else return null;
+
+  const msg: ChatMessage = { role, content };
+
+  if (m.tool_calls !== undefined && m.tool_calls !== null) {
+    if (role !== "assistant" || !Array.isArray(m.tool_calls)) return null;
+    const calls: ToolCall[] = [];
+    for (const raw of m.tool_calls) {
+      const tc = normalizeToolCall(raw);
+      if (!tc) return null;
+      calls.push(tc);
+    }
+    if (calls.length > 0) msg.tool_calls = calls;
+  }
+
+  if (m.tool_call_id !== undefined && m.tool_call_id !== null) {
+    if (role !== "tool" || typeof m.tool_call_id !== "string" || m.tool_call_id === "") return null;
+    msg.tool_call_id = m.tool_call_id;
+  }
+  if (role === "tool" && msg.tool_call_id === undefined) return null;
+
+  return msg;
+}
+
+export function toolCallsEqual(a?: ToolCall[], b?: ToolCall[]): boolean {
+  const la = a ?? [];
+  const lb = b ?? [];
+  if (la.length !== lb.length) return false;
+  return la.every((tc, i) =>
+    tc.id === lb[i].id &&
+    tc.function.name === lb[i].function.name &&
+    tc.function.arguments === lb[i].function.arguments,
+  );
+}
+
+/**
+ * Tolerant equality for affinity prefix matching. Both sides are expected to
+ * already be normalized; comparison ignores incidental key order and treats
+ * missing tool fields as empty.
+ */
+export function chatMessagesEquivalent(a: ChatMessage, b: ChatMessage): boolean {
+  return (
+    a.role === b.role &&
+    a.content === b.content &&
+    (a.tool_call_id ?? "") === (b.tool_call_id ?? "") &&
+    toolCallsEqual(a.tool_calls, b.tool_calls)
+  );
+}

packages/shared/src/tx/escrow/postChatEscrow.ts

@@ -42,6 +42,7 @@ import { loadBlueprint } from "../blueprint.js";
 import { encodeTxBody, sha256Hex } from "../internal/testTxBody.js";
 import { mockSlotToWallclockMs, NETWORK_BUFFER_MS } from "../internal/constants.js";
 import { detectCborBackend } from "../internal/cborBackend.js";
+import { escrowLockFloor } from "../internal/minAdaFloor.js";
 import type { LiveOgmiosProvider } from "../../chain/LiveOgmiosProvider.js";
 
 export interface PostChatEscrowParams {
@@ -132,7 +133,7 @@ export async function buildPostChatEscrowTx(
   const requestSpecHash = sha256Utf8Hex(requestSpecCanonical);
   const promptHash = chatSessionPromptHash({ session_nonce });
 
-  const totalLocked =
+  const economicTotal =
     advertDatum.price_lovelace +
     advertDatum.buyer_bond_lovelace +
     advertDatum.supplier_bond_lovelace;
@@ -154,6 +155,11 @@ export async function buildPostChatEscrowTx(
     state: "Open",
   };
 
+  // Lock enough that the escrow output satisfies min-ada in its LARGEST
+  // future state (Submitted) - value_equal on Claim/Submit forbids any
+  // later bump (min-ada floor).
+  const totalLocked = escrowLockFloor(escrowDatum, economicTotal);
+
   // Live path: reuse the capability-agnostic chat builder (messages = []).
   // The on-chain commitment is the escrowDatum (carrying our session-init
   // prompt_hash); messages are only used for the mock cbor body, which we omit.

packages/shared/src/tx/escrow/postEscrow.ts

@@ -38,6 +38,7 @@ import { loadBlueprint } from "../blueprint.js";
 import { encodeTxBody, sha256Hex } from "../internal/testTxBody.js";
 import { mockSlotToWallclockMs, NETWORK_BUFFER_MS } from "../internal/constants.js";
 import { detectCborBackend } from "../internal/cborBackend.js";
+import { escrowLockFloor } from "../internal/minAdaFloor.js";
 import type { LiveOgmiosProvider } from "../../chain/LiveOgmiosProvider.js";
 
 export interface PostEscrowParams {
@@ -129,7 +130,7 @@ export async function buildPostEscrowTx(
   // supplier-side validation in M1-C can verify role/content/system integrity.
   const promptHash = sha256Utf8Hex(canonicalize(messages));
 
-  const totalLocked =
+  const economicTotal =
     advertDatum.price_lovelace +
     advertDatum.buyer_bond_lovelace +
     advertDatum.supplier_bond_lovelace;
@@ -151,6 +152,11 @@ export async function buildPostEscrowTx(
     state: "Open",
   };
 
+  // Lock enough that the escrow output satisfies min-ada in its LARGEST
+  // future state (Submitted) - value_equal on Claim/Submit forbids any
+  // later bump (min-ada floor).
+  const totalLocked = escrowLockFloor(escrowDatum, economicTotal);
+
   // Live path: produce real Cardano CBOR via lucid-evolution.
   if (detectCborBackend(chain) === "live") {
     const liveCborPath = "../internal/liveCbor.js";

packages/shared/src/tx/escrow/postOcrEscrow.ts

@@ -0,0 +1,257 @@
+/**
+ * tx/escrow/postOcrEscrow.ts — PostEscrow tx builder for model-scoped OCR
+ * capabilities (`ocr.page.extract.<model-slug>.v1`, one page per job).
+ *
+ * Mirrors `postTtsEscrow.ts` but commits an OCR-shaped prompt_hash over the
+ * canonicalised request envelope `{image_b64, mime, output_format}` — the
+ * SAME object the supplier hashes when validating the incoming
+ * POST /v1/ocr/extract body. Any mismatch on either side fails the
+ * supplier-side prompt_mismatch gate.
+ *
+ * The image travels off-chain (buyer → supplier HTTP body); the chain only
+ * ever carries prompt_hash. Never place image bytes or extracted text in a
+ * datum.
+ *
+ * Off-chain invariants (mirror TTS's, with OCR-shaped body validation):
+ *   1. advert UTxO exists at advertRef
+ *   2. advert datum.status === "Active"
+ *   3. payment_lovelace === advert.price_lovelace
+ *   4. buyerKey.pubKeyHash !== advert.supplier_pkh
+ *   5. body.image_b64 is a non-empty plain-base64 string within size bounds
+ *      (no data-URL prefix — the mime rides in its own field)
+ *   6. body.mime ∈ ALLOWED_OCR_MIMES
+ *   7. body.output_format ∈ ALLOWED_OCR_OUTPUT_FORMATS
+ */
+
+import * as nodeCrypto from "crypto";
+import type { ChainProvider, OutputReference } from "../../chain/ChainProvider.js";
+import type { EscrowDatum } from "../../cbor/types.js";
+import { decodeAdvertDatum } from "../../cbor/AdvertDatum.js";
+import { encodeEscrowDatum } from "../../cbor/EscrowDatum.js";
+import { canonicalize } from "../../cbor/canonical.js";
+import type { WalletKey, PostEscrowBuildResult } from "../types.js";
+import { TxConstructionError } from "../types.js";
+import { loadBlueprint } from "../blueprint.js";
+import { encodeTxBody, sha256Hex } from "../internal/testTxBody.js";
+import { mockSlotToWallclockMs, NETWORK_BUFFER_MS } from "../internal/constants.js";
+import { detectCborBackend } from "../internal/cborBackend.js";
+import { escrowLockFloor } from "../internal/minAdaFloor.js";
+import type { LiveOgmiosProvider } from "../../chain/LiveOgmiosProvider.js";
+
+/** OCR request envelope. The buyer-side hash and the supplier-side hash
+ * MUST agree on this exact set of keys, in this exact JCS-canonical
+ * encoding, or the supplier will reject with `prompt_mismatch`. */
+export interface OcrRequest {
+  /** Plain base64 of the page image bytes. No `data:` prefix. */
+  image_b64: string;
+  /** Image content type. */
+  mime: string;
+  /** Requested extraction output shape. */
+  output_format: string;
+}
+
+export const ALLOWED_OCR_MIMES = new Set([
+  "image/png", "image/jpeg", "image/webp",
+]);
+export const ALLOWED_OCR_OUTPUT_FORMATS = new Set(["markdown", "html", "json"]);
+
+/**
+ * Upper bound on the base64 payload, in characters (~9 MB of image bytes).
+ * A 300-DPI page scan lands around 0.5–3 MB binary; this cap protects the
+ * supplier's JSON body parser and the off-chain transport, not the chain
+ * (the chain only sees the 32-byte hash).
+ */
+export const MAX_OCR_IMAGE_B64_CHARS = 12_000_000;
+
+/** Plain base64 (standard alphabet, optional padding). Rejects data URLs,
+ * whitespace and base64url — one canonical encoding, one hash. */
+const BASE64_RE = /^[A-Za-z0-9+/]+={0,2}$/;
+
+export interface PostOcrEscrowParams {
+  chain: ChainProvider;
+  buyerKey: WalletKey;
+  advertRef: OutputReference;
+  request: OcrRequest;
+  payment_lovelace: bigint;
+}
+
+function sha256Utf8Hex(s: string): string {
+  return nodeCrypto.createHash("sha256").update(s, "utf8").digest("hex");
+}
+
+/** Hash function shared between buyer (here) and supplier (server.ts).
+ * Exported so unit tests on either side can pin the canonicalisation. */
+export function ocrPromptHash(req: OcrRequest): string {
+  return sha256Utf8Hex(canonicalize({
+    image_b64: req.image_b64,
+    mime: req.mime,
+    output_format: req.output_format,
+  }));
+}
+
+/** Body validation shared conceptually with the supplier route. Throws
+ * TxConstructionError with a machine-readable reason. */
+export function validateOcrRequest(request: OcrRequest): void {
+  if (typeof request?.image_b64 !== "string" || request.image_b64.length === 0) {
+    throw new TxConstructionError("image required",
+      "request.image_b64 must be a non-empty base64 string");
+  }
+  if (request.image_b64.length > MAX_OCR_IMAGE_B64_CHARS) {
+    throw new TxConstructionError("image too large",
+      `request.image_b64 exceeds ${MAX_OCR_IMAGE_B64_CHARS} chars`);
+  }
+  if (!BASE64_RE.test(request.image_b64)) {
+    throw new TxConstructionError("image not base64",
+      "request.image_b64 must be plain base64 (no data: prefix, no whitespace)");
+  }
+  if (!ALLOWED_OCR_MIMES.has(request.mime)) {
+    throw new TxConstructionError("mime invalid",
+      `mime must be one of: ${[...ALLOWED_OCR_MIMES].join(", ")}`);
+  }
+  if (!ALLOWED_OCR_OUTPUT_FORMATS.has(request.output_format)) {
+    throw new TxConstructionError("output_format invalid",
+      `output_format must be one of: ${[...ALLOWED_OCR_OUTPUT_FORMATS].join(", ")}`);
+  }
+}
+
+export async function buildPostOcrEscrowTx(
+  params: PostOcrEscrowParams,
+): Promise<PostEscrowBuildResult> {
+  const { chain, buyerKey, advertRef, request, payment_lovelace } = params;
+
+  // 5–7. Body validation before any chain access.
+  validateOcrRequest(request);
+
+  // 1. Advert UTxO must exist.
+  const advertUtxo = await chain.queryUtxo(advertRef);
+  if (advertUtxo === null) {
+    throw new TxConstructionError(
+      "advert ref not on chain",
+      `no UTxO at ${advertRef.txHash}#${advertRef.index}`,
+    );
+  }
+  if (!advertUtxo.datumHex) {
+    throw new TxConstructionError(
+      "advert datum missing",
+      `UTxO ${advertRef.txHash}#${advertRef.index} has no inline datum`,
+    );
+  }
+
+  const advertDatum = decodeAdvertDatum(advertUtxo.datumHex);
+
+  // 2. Advert must be Active.
+  if (advertDatum.status !== "Active") {
+    throw new TxConstructionError(
+      "advert is retired",
+      `advert.status is ${advertDatum.status}, expected Active`,
+    );
+  }
+
+  // 3. Payment must equal advertised price.
+  if (payment_lovelace !== advertDatum.price_lovelace) {
+    throw new TxConstructionError(
+      "payment must equal advertised price",
+      `payment ${payment_lovelace} != advert.price ${advertDatum.price_lovelace}`,
+    );
+  }
+
+  // 4. Buyer cannot be supplier.
+  if (buyerKey.pubKeyHash === advertDatum.supplier_pkh) {
+    throw new TxConstructionError(
+      "buyer cannot be supplier",
+      `buyerKey pkh ${buyerKey.pubKeyHash} equals advert.supplier_pkh`,
+    );
+  }
+
+  // Time source: same convention as chat/TTS.
+  const tipSlot = await chain.tip();
+  const isLive = detectCborBackend(chain) === "live";
+  const postedAt = isLive ? Date.now() : mockSlotToWallclockMs(tipSlot);
+  const deliverBy = postedAt + advertDatum.max_processing_ms + NETWORK_BUFFER_MS;
+
+  const requestSpecCanonical = canonicalize({
+    capability_id: advertDatum.capability_id,
+    max_output_tokens: advertDatum.max_output_tokens,
+    model: advertDatum.model,
+  });
+  const requestSpecHash = sha256Utf8Hex(requestSpecCanonical);
+  const promptHash = ocrPromptHash(request);
+
+  const economicTotal =
+    advertDatum.price_lovelace +
+    advertDatum.buyer_bond_lovelace +
+    advertDatum.supplier_bond_lovelace;
+
+  const escrowDatum: EscrowDatum = {
+    buyer_pkh: buyerKey.pubKeyHash,
+    supplier_pkh: advertDatum.supplier_pkh,
+    advert_ref: advertRef,
+    capability_id: advertDatum.capability_id,
+    request_spec_hash: requestSpecHash,
+    prompt_hash: promptHash,
+    payment_lovelace: advertDatum.price_lovelace,
+    buyer_bond_lovelace: advertDatum.buyer_bond_lovelace,
+    supplier_bond_lovelace: advertDatum.supplier_bond_lovelace,
+    deliver_by: deliverBy,
+    posted_at: postedAt,
+    submitted_at: null,
+    result_receipt_hash: null,
+    state: "Open",
+  };
+
+  // Lock enough that the escrow output satisfies min-ada in its LARGEST
+  // future state (Submitted) - value_equal on Claim/Submit forbids any
+  // later bump (2026-08-07 min-ada incident).
+  const totalLocked = escrowLockFloor(escrowDatum, economicTotal);
+
+  // Live path: reuse the capability-agnostic live builder (messages is
+  // void-marked there; the on-chain commitment is the escrowDatum, which
+  // carries our OCR prompt_hash).
+  if (isLive) {
+    const liveCborPath = "../internal/liveCbor.js";
+    const { buildLiveTxForEscrow } = await import(/* @vite-ignore */ liveCborPath);
+    return buildLiveTxForEscrow({
+      chain: chain as LiveOgmiosProvider,
+      buyerKey,
+      advertRef,
+      messages: [],
+      escrowDatum,
+      totalLocked,
+      deliverBy,
+      postedAt,
+    });
+  }
+
+  // Mock backend path (synthetic JSON-in-hex tx body).
+  const blueprint = loadBlueprint();
+  const escrowAddress = blueprint.escrowScriptAddress(0);
+
+  const body = {
+    type: "post-escrow",
+    inputs: [],
+    outputs: [
+      {
+        ref: { txHash: "$self", index: 0 },
+        address: escrowAddress,
+        lovelace: totalLocked,
+        assets: {},
+        datumHex: encodeEscrowDatum(escrowDatum),
+        scriptRef: null,
+      },
+    ],
+    requiredSigners: [buyerKey.pubKeyHash],
+    validityRange: { lowerBoundMs: postedAt, upperBoundMs: deliverBy },
+    meta: { script_hash: blueprint.escrowScriptHash, advert_ref: advertRef },
+  };
+
+  const txCborHex = encodeTxBody(body);
+  const expectedTxHash = sha256Hex(txCborHex);
+
+  await chain.submitTx(txCborHex);
+
+  return {
+    txCborHex,
+    expectedTxHash,
+    escrowOutputRef: { txHash: expectedTxHash, index: 0 },
+  };
+}

packages/shared/src/tx/escrow/postTtsEscrow.ts

@@ -31,6 +31,7 @@ import { loadBlueprint } from "../blueprint.js";
 import { encodeTxBody, sha256Hex } from "../internal/testTxBody.js";
 import { mockSlotToWallclockMs, NETWORK_BUFFER_MS } from "../internal/constants.js";
 import { detectCborBackend } from "../internal/cborBackend.js";
+import { escrowLockFloor } from "../internal/minAdaFloor.js";
 import type { LiveOgmiosProvider } from "../../chain/LiveOgmiosProvider.js";
 
 /** TTS request envelope. The buyer-side hash and the supplier-side hash
@@ -149,7 +150,7 @@ export async function buildPostTtsEscrowTx(
   const requestSpecHash = sha256Utf8Hex(requestSpecCanonical);
   const promptHash = ttsPromptHash(request);
 
-  const totalLocked =
+  const economicTotal =
     advertDatum.price_lovelace +
     advertDatum.buyer_bond_lovelace +
     advertDatum.supplier_bond_lovelace;
@@ -171,6 +172,11 @@ export async function buildPostTtsEscrowTx(
     state: "Open",
   };
 
+  // Lock enough that the escrow output satisfies min-ada in its LARGEST
+  // future state (Submitted) - value_equal on Claim/Submit forbids any
+  // later bump (min-ada floor).
+  const totalLocked = escrowLockFloor(escrowDatum, economicTotal);
+
   // Live path: reuse the chat builder. The live cbor path is already
   // capability-agnostic — `messages` is `void`-marked there. We pass an
   // empty array so the type checker is satisfied; the actual on-chain

packages/shared/src/tx/index.ts

@@ -2,10 +2,12 @@
  * tx/index.ts — public exports for M1-B tx builders.
  */
 
-export type { ChatMessage, WalletKey, BuildResult, PostAdvertBuildResult, PostEscrowBuildResult } from "./types.js";
+export type { ChatMessage, ToolCall, WalletKey, BuildResult, PostAdvertBuildResult, PostEscrowBuildResult } from "./types.js";
 export { TxConstructionError } from "./types.js";
+export { normalizeChatMessage, chatMessagesEquivalent, toolCallsEqual } from "./chatMessage.js";
 export { mockSlotToWallclockMs, mockWallclockMsToSlot, NETWORK_BUFFER_MS } from "./internal/constants.js";
 export { detectCborBackend, type CborBackend } from "./internal/cborBackend.js";
+export { escrowLockFloor, minAdaForEscrowDatum } from "./internal/minAdaFloor.js";
 export type { Signer } from "./signer.js";
 export { MockSigner } from "./signer.js";
 export type { Blueprint } from "./blueprint.js";
@@ -36,6 +38,15 @@ export {
 } from "./escrow/postTtsEscrow.js";
 export type { PostChatEscrowParams } from "./escrow/postChatEscrow.js";
 export { buildPostChatEscrowTx, chatSessionPromptHash } from "./escrow/postChatEscrow.js";
+export type { PostOcrEscrowParams, OcrRequest } from "./escrow/postOcrEscrow.js";
+export {
+  buildPostOcrEscrowTx,
+  ocrPromptHash,
+  validateOcrRequest,
+  ALLOWED_OCR_MIMES,
+  ALLOWED_OCR_OUTPUT_FORMATS,
+  MAX_OCR_IMAGE_B64_CHARS,
+} from "./escrow/postOcrEscrow.js";
 export type { ClaimParams } from "./escrow/claim.js";
 export { buildClaimTx } from "./escrow/claim.js";
 export type { SubmitParams } from "./escrow/submit.js";

packages/shared/src/tx/internal/liveCbor.ts

@@ -358,6 +358,45 @@ function rethrowAsTxError(err: unknown, fallback: string): never {
   throw new TxConstructionError(fallback, msg);
 }
 
+/** Choose PostEscrow inputs that PRESERVE a pure-ADA collateral UTxO for the
+ * later Accept/Reclaim script-spend. PostEscrow is NOT a script spend, so lucid
+ * is otherwise free to consume the dedicated ≥5-ADA collateral UTxO as an
+ * ordinary input — leaving the wallet with no collateral candidate and
+ * stranding the escrow when Accept later runs (the fragmentation bug observed
+ * on back-to-back requests).
+ *
+ * Reserve the SMALLEST pure-ADA UTxO ≥ collateral floor and fund the escrow
+ * from the remaining UTxOs. Returns null (→ default lucid coin selection) when
+ * reservation isn't possible or needed:
+ *   - a single UTxO (nothing to reserve),
+ *   - no qualifying pure-ADA collateral UTxO, or
+ *   - excluding it would underfund the escrow. The large-single-UTxO case is
+ *     safe under default selection anyway: the change output is itself a
+ *     pure-ADA UTxO well above the collateral floor. */
+export function reserveCollateralInputs(
+  utxos: LucidUTxO[],
+  lockedLovelace: bigint,
+): LucidUTxO[] | null {
+  if (utxos.length < 2) return null;
+  const candidates = utxos
+    .filter(
+      (u) =>
+        u.assets.lovelace >= COLLATERAL_MIN_LOVELACE &&
+        Object.keys(u.assets).every((k) => k === "lovelace"),
+    )
+    .sort((a, b) => (a.assets.lovelace < b.assets.lovelace ? -1 : 1));
+  if (candidates.length === 0) return null;
+  const reserved = candidates[0];
+  const remaining = utxos.filter(
+    (u) => !(u.txHash === reserved.txHash && u.outputIndex === reserved.outputIndex),
+  );
+  if (remaining.length === 0) return null;
+  const remainingLovelace = remaining.reduce((s, u) => s + u.assets.lovelace, 0n);
+  // Cover the locked output + tx fee + a min-change cushion; else fall back.
+  if (remainingLovelace < lockedLovelace + 2_000_000n) return null;
+  return remaining;
+}
+
 // ─── PostEscrow ──────────────────────────────────────────────────────
 
 export async function buildLiveTxForEscrow(
@@ -382,6 +421,12 @@ export async function buildLiveTxForEscrow(
 
   const lockedLovelace = totalLocked < MIN_UTXO_LOVELACE ? MIN_UTXO_LOVELACE : totalLocked;
 
+  // Preserve a pure-ADA collateral UTxO so the buyer's later Accept/Reclaim
+  // script-spend can find collateral. Without this, PostEscrow coin selection
+  // can consume the 5-ADA collateral UTxO and strand the escrow at Accept.
+  const walletUtxos = await lucid.wallet().getUtxos();
+  const reservedInputs = reserveCollateralInputs(walletUtxos, lockedLovelace);
+
   let signed;
   try {
     const txBuilder = lucid
@@ -404,7 +449,12 @@ export async function buildLiveTxForEscrow(
       // becomes change. Same fix as Claim/Submit/Accept paths.
       .setMinFee(500_000n);
 
-    const completed = await txBuilder.complete();
+    // When we can reserve collateral, restrict coin selection to the remaining
+    // (working) UTxOs so the collateral candidate survives for Accept. Else
+    // fall back to lucid's default selection (single/large-UTxO cases are safe).
+    const completed = reservedInputs
+      ? await txBuilder.complete({ presetWalletInputs: reservedInputs, localUPLCEval: false })
+      : await txBuilder.complete();
     signed = await completed.sign.withWallet().complete();
   } catch (err) {
     rethrowAsTxError(err, "post-escrow build failed");

packages/shared/src/tx/internal/minAdaFloor.ts

@@ -0,0 +1,67 @@
+/**
+ * minAdaFloor.ts — min-ada floor for the escrow's LARGEST future datum.
+ *
+ * The escrow validator requires value_equal(continuing, in) on Claim and
+ * Submit, so the lovelace locked at POST time must already satisfy min-ada
+ * for every datum the output will ever carry. The largest is the Submitted
+ * state: submitted_at None -> Some(POSIX ms) and result_receipt_hash
+ * None -> Some(32 bytes) add ~43 bytes over Open/Claimed. If the posted
+ * lovelace sits below that state's min-ada, lucid bumps the Submit
+ * continuing output to min-ada, value_equal fails, and the validator
+ * errors with no traces (the 2026-08-07 mainnet incident: the 33-byte
+ * "ocr.page.extract.chandra-ocr-2.v1" capability_id pushed the Submitted
+ * output 11,030 lovelace over a 2,200,000 lock; the 20-byte chat
+ * capability stayed 49,310 under, which is why only OCR Submits failed).
+ *
+ * escrowLockFloor() returns the lovelace to lock: the economic total
+ * (payment + both bonds), raised to the Submitted-state min-ada plus a
+ * safety margin when the total alone is too small.
+ */
+
+import type { EscrowDatum } from "../../cbor/types.js";
+import { encodeEscrowDatum } from "../../cbor/EscrowDatum.js";
+
+/** Conway/Babbage utxoCostPerByte on Vector mainnet (protocol params). */
+const UTXO_COST_PER_BYTE = 4310n;
+
+/** Ledger constant: min-ada = (160 + serialized output bytes) * costPerByte. */
+const OUTPUT_OVERHEAD_BYTES = 160n;
+
+/** Covers CBOR framing drift between this estimate and the ledger's exact
+ * serialization (address, coin width, datum wrapper). Generous on purpose:
+ * 50k lovelace = 0.05 AP3X, dwarfed by per-leg tx fees. */
+const SAFETY_MARGIN_LOVELACE = 50_000n;
+
+/** Serialized-output byte estimate for an enterprise script address output
+ * carrying `lovelace` and an inline datum of `datumBytes`. Mirrors the
+ * arithmetic validated against mainnet on 2026-08-07 (minada-calc). */
+function estimateOutputBytes(datumBytes: number, lovelace: bigint): bigint {
+  const addr = 1n + 29n;
+  const coin = 1n + (lovelace > 0xffffffffn ? 8n : 4n);
+  const datumField = 1n + 2n + 3n + BigInt(datumBytes);
+  const mapOverhead = 3n;
+  return addr + coin + datumField + mapOverhead;
+}
+
+/** Min-ada for the escrow output when it carries `datum`. */
+export function minAdaForEscrowDatum(datum: EscrowDatum, lovelace: bigint): bigint {
+  const datumBytes = encodeEscrowDatum(datum).length / 2;
+  const size = estimateOutputBytes(datumBytes, lovelace);
+  return (OUTPUT_OVERHEAD_BYTES + size) * UTXO_COST_PER_BYTE;
+}
+
+/** The lovelace a POST must lock so the escrow output satisfies min-ada in
+ * every reachable state, including Submitted. `economicTotal` is
+ * payment + buyer_bond + supplier_bond. */
+export function escrowLockFloor(openDatum: EscrowDatum, economicTotal: bigint): bigint {
+  const submittedShape: EscrowDatum = {
+    ...openDatum,
+    // Representative worst-case Submit fields: a 13-digit POSIX ms stamp
+    // and a 32-byte receipt hash — sizes, not values, matter here.
+    submitted_at: openDatum.deliver_by,
+    result_receipt_hash: "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff",
+    state: "Submitted",
+  };
+  const floor = minAdaForEscrowDatum(submittedShape, economicTotal) + SAFETY_MARGIN_LOVELACE;
+  return floor > economicTotal ? floor : economicTotal;
+}

packages/shared/src/tx/internal/reserveCollateral.test.ts

@@ -0,0 +1,42 @@
+import { describe, it, expect } from "vitest";
+import { reserveCollateralInputs } from "./liveCbor.js";
+
+// Minimal LucidUTxO-shaped fixtures (the helper only reads txHash/outputIndex/assets).
+const pure = (id: string, lovelace: bigint) =>
+  ({ txHash: id, outputIndex: 0, assets: { lovelace } }) as never;
+const mixed = (id: string, lovelace: bigint) =>
+  ({ txHash: id, outputIndex: 0, assets: { lovelace, "policy.tok": 1n } }) as never;
+
+const LOCKED = 2_200_000n; // price + buyer_bond + supplier_bond
+const ids = (r: ReturnType<typeof reserveCollateralInputs>) =>
+  (r ?? []).map((u: { txHash: string }) => u.txHash);
+
+describe("reserveCollateralInputs", () => {
+  it("reserves a 5-ADA collateral UTxO and funds from the rest", () => {
+    const r = reserveCollateralInputs([pure("collat", 5_000_000n), pure("work", 6_000_000n)], LOCKED);
+    expect(ids(r)).toEqual(["work"]); // collat (smallest ≥5) reserved
+  });
+
+  it("reserves the SMALLEST qualifying candidate", () => {
+    const r = reserveCollateralInputs([pure("big", 9_000_000n), pure("small", 5_000_000n)], LOCKED);
+    expect(ids(r)).toEqual(["big"]); // small reserved as collateral
+  });
+
+  it("only reserves PURE-ADA UTxOs (ignores ones carrying native assets)", () => {
+    const r = reserveCollateralInputs([mixed("mixed", 10_000_000n), pure("purecol", 6_000_000n)], LOCKED);
+    expect(ids(r)).toEqual(["mixed"]); // purecol reserved, fund from the mixed UTxO
+  });
+
+  it("returns null for a single UTxO (large change stays as collateral under default selection)", () => {
+    expect(reserveCollateralInputs([pure("only", 12_000_000n)], LOCKED)).toBeNull();
+  });
+
+  it("returns null when no UTxO qualifies as collateral", () => {
+    expect(reserveCollateralInputs([pure("a", 4_000_000n), pure("b", 4_000_000n)], LOCKED)).toBeNull();
+  });
+
+  it("returns null when reserving collateral would underfund the escrow", () => {
+    // {5 collateral, 3 working}: working can't cover locked + cushion → fall back.
+    expect(reserveCollateralInputs([pure("collat", 5_000_000n), pure("work", 3_000_000n)], LOCKED)).toBeNull();
+  });
+});

packages/shared/src/tx/types.ts

@@ -11,14 +11,29 @@
 
 import type { OutputReference } from "../chain/ChainProvider.js";
 
+/**
+ * ToolCall — OpenAI-compatible function tool call emitted by an assistant turn.
+ * `function.arguments` is the raw JSON string as produced by the model.
+ */
+export interface ToolCall {
+  id: string;
+  type: "function";
+  function: { name: string; arguments: string };
+}
+
 /**
  * ChatMessage — OpenAI-compatible chat message.
- * v1 supports user/system/assistant roles. The wire shape is intentionally
- * narrow; future capabilities (tools, multimodal) will extend or replace it.
+ * `content` is always a string ("" for tool-call-only assistant messages —
+ * OpenAI's `content: null` is normalized on ingest). `tool_calls` /
+ * `tool_call_id` MUST be omitted (not set to undefined) when absent: the
+ * receipt hash canonicalizes the transcript objects, so field presence is
+ * part of the identity shared between supplier transcript and gateway mirror.
  */
 export interface ChatMessage {
-  role: "system" | "user" | "assistant";
+  role: "system" | "user" | "assistant" | "tool";
   content: string;
+  tool_calls?: ToolCall[];
+  tool_call_id?: string;
 }
 
 /**

packages/shared/src/tx/wallet/planConsolidate.ts

@@ -2,9 +2,22 @@
  * planConsolidate.ts — pure pre-flight planner for the wallet-consolidate flow.
  *
  * Given a wallet's current UTxO set + a target collateral lovelace, decide
- * whether to consolidate (many → 2), split (1 → 2), or skip (already at
- * {≥collateral, working}). The same downstream tx-builder handles
- * consolidate and split — only the "reason" tag differs for logging.
+ * whether the wallet needs reshaping or is already usable for script spends.
+ *
+ * Healthy is a STRUCTURAL predicate, not a UTxO count (F7, 2026-08-08):
+ *   (a) some pure-lovelace UTxO >= collateralLovelace exists (the collateral
+ *       candidate lucid's setCollateral needs), AND
+ *   (b) at least MIN_WORKING_LOVELACE sits OUTSIDE that candidate (fee and
+ *       spend money that is not the candidate), AND
+ *   (c) fragmentation is bounded (utxos.length <= maxUtxos).
+ *
+ * The earlier revision returned `already-healthy` only for EXACTLY 2 UTxOs.
+ * The consolidate tx itself lands 3 (collateral + working + the fee-reserve
+ * remainder as change), so the shape consolidation produced could never
+ * satisfy its own health check — the wallet-health ticker re-consolidated
+ * every tick and burned 0.5 AP3X per pass (116 AP3X in one 24h soak).
+ * The health check also runs BEFORE the low-balance guard now: a wallet that
+ * can already transact must never be told "balance too low to consolidate".
  *
  * No lucid / WASM deps so this is unit-testable from a vanilla vitest.
  */
@@ -26,12 +39,24 @@ export interface PlannedConsolidate {
 
 export const DEFAULT_COLLATERAL_LOVELACE = 5_000_000n;
 export const DEFAULT_FEE_RESERVE = 2_000_000n;
+/** Reshape only when fragmentation passes this; receipt/change UTxOs
+ *  accumulating between jobs are fine as long as a candidate exists. */
+export const DEFAULT_MAX_UTXOS = 6;
 const MIN_WORKING_LOVELACE = 1_000_000n;
 
+/** Lovelace of a UTxO iff it carries no native assets; null otherwise. */
+function pureLovelace(u: UtxoLike): bigint | null {
+  for (const [unit, amount] of Object.entries(u.assets)) {
+    if (unit !== "lovelace" && (amount ?? 0n) > 0n) return null;
+  }
+  return u.assets.lovelace ?? 0n;
+}
+
 export function planConsolidate(
   utxos: UtxoLike[],
   collateralLovelace: bigint = DEFAULT_COLLATERAL_LOVELACE,
   feeReserve: bigint = DEFAULT_FEE_RESERVE,
+  maxUtxos: number = DEFAULT_MAX_UTXOS,
 ): PlannedConsolidate {
   if (utxos.length === 0) {
     throw new Error("wallet has no UTxOs — fund the address first");
@@ -42,6 +67,32 @@ export function planConsolidate(
     0n,
   );
 
+  // Structural health check first — reshaping a usable wallet wastes a fee.
+  if (utxos.length <= maxUtxos) {
+    // Smallest qualifying candidate: leave larger UTxOs as working funds.
+    let candidate: bigint | null = null;
+    for (const u of utxos) {
+      const pure = pureLovelace(u);
+      if (pure !== null && pure >= collateralLovelace) {
+        if (candidate === null || pure < candidate) candidate = pure;
+      }
+    }
+    if (candidate !== null) {
+      const workingOutside = totalLovelaceIn - candidate;
+      if (workingOutside >= MIN_WORKING_LOVELACE) {
+        return {
+          inputCount: utxos.length,
+          totalLovelaceIn,
+          collateralOutput: candidate,
+          workingOutput: workingOutside,
+          alreadyHealthy: true,
+          reason: "already-healthy",
+        };
+      }
+    }
+  }
+
+  // Reshape needed — only an actual reshape has a funds threshold.
   const required = collateralLovelace + feeReserve + MIN_WORKING_LOVELACE;
   if (totalLovelaceIn < required) {
     throw new Error(
@@ -51,23 +102,6 @@ export function planConsolidate(
     );
   }
 
-  if (utxos.length === 2) {
-    const lov0 = utxos[0].assets.lovelace ?? 0n;
-    const lov1 = utxos[1].assets.lovelace ?? 0n;
-    const smaller = lov0 < lov1 ? lov0 : lov1;
-    const larger = lov0 < lov1 ? lov1 : lov0;
-    if (smaller >= collateralLovelace) {
-      return {
-        inputCount: 2,
-        totalLovelaceIn,
-        collateralOutput: smaller,
-        workingOutput: larger,
-        alreadyHealthy: true,
-        reason: "already-healthy",
-      };
-    }
-  }
-
   const workingOutput = totalLovelaceIn - collateralLovelace - feeReserve;
 
   return {

pnpm-lock.yaml

@@ -4,6 +4,11 @@ settings:
   autoInstallPeers: true
   excludeLinksFromLockfile: false
 
+patchedDependencies:
+  libsodium-wrappers-sumo@0.7.16:
+    hash: 761a204004f889b65fdf148159f4bd90733703d63a0ff00eb0671c92ea085096
+    path: patches/libsodium-wrappers-sumo@0.7.16.patch
+
 importers:
 
   .:
@@ -3166,7 +3171,7 @@ snapshots:
       '@cardano-sdk/util': 0.16.0
       blake2b: 2.1.4
       i: 0.3.7
-      libsodium-wrappers-sumo: 0.7.16
+      libsodium-wrappers-sumo: 0.7.16(patch_hash=761a204004f889b65fdf148159f4bd90733703d63a0ff00eb0671c92ea085096)
       lodash: 4.18.1
       npm: 9.9.4
       pbkdf2: 3.1.5
@@ -4743,7 +4748,7 @@ snapshots:
 
   libsodium-sumo@0.7.16: {}
 
-  libsodium-wrappers-sumo@0.7.16:
+  libsodium-wrappers-sumo@0.7.16(patch_hash=761a204004f889b65fdf148159f4bd90733703d63a0ff00eb0671c92ea085096):
     dependencies:
       libsodium-sumo: 0.7.16
 

pnpm-workspace.yaml

@@ -6,3 +6,6 @@ packages:
   - "indexer"
   - "indexer-ui"
   - "wallet-monitor"
+
+onlyBuiltDependencies:
+  - better-sqlite3

supplier/.env.chandra.example

@@ -0,0 +1,83 @@
+# supplier/.env.chandra.example — Chandra OCR-2 supplier (ocr capability)
+#
+# Copy to supplier/.env.chandra on the host, chmod 600, fill the blanks.
+# TESTNET FIRST: run the full lifecycle (advert → escrow → claim → submit →
+# accept) on testnet before any mainnet value enters this file. Mainnet ops
+# (funding, advert post, container start) are operator-executed.
+#
+# The upstream is any OpenAI-compatible VISION endpoint serving
+# datalab-to/chandra-ocr-2 (vLLM recommended; `chandra_vllm` docker or
+# `vllm serve datalab-to/chandra-ocr-2 --revision <sha>`).
+# LICENSING: the model weights and the hosted API each carry their own
+# terms. Confirm your own licensing position with the model provider before
+# serving commercially.
+
+# ── Identity / chain ──────────────────────────────────────────────────────
+SUPPLIER_PRIV_KEY_HEX=            # gen-keypair output; NEW wallet, working float only
+NETWORK_ID=                       # 0 testnet | 1 mainnet — set explicitly, never default
+OGMIOS_URL=                       # wss://ogmios.vector.testnet... | mainnet equivalent
+ADVERT_REF=                       # <txHash>#<index> from tx:post-advert (below)
+LIVE_CHAIN=                       # 1 only when submitting for real
+VECTOR_ZERO_TIME_MS=              # per deploy/README (mainnet 1756485600000)
+
+# ── Capability ────────────────────────────────────────────────────────────
+CAPABILITY_KIND=ocr
+
+# ── Upstream selection ────────────────────────────────────────────────────
+# "datalab" = Datalab hosted document API (v0 deployment, decided 2026-08-06)
+# "openai-vision" = self-hosted vLLM (the attestation upgrade path)
+OCR_UPSTREAM=datalab
+
+# ── Datalab upstream (OCR_UPSTREAM=datalab) ───────────────────────────────
+# TERMS: the upstream provider's end-user terms govern how you may use and
+# redistribute output, and they vary by account and plan. Confirm your own
+# position with the provider before serving third parties.
+DATALAB_API_KEY=                  # provider account key; never commit
+DATALAB_BASE_URL=https://www.datalab.to
+DATALAB_MODE=accurate             # fast | balanced | accurate
+DATALAB_TIMEOUT_MS=120000         # submit + poll budget; keep < advert max_processing_ms
+
+# ── OpenAI-compatible vision upstream (OCR_UPSTREAM=openai-vision) ────────
+# No /v1 suffix — supplier/src/ocrVision.ts appends /v1/chat/completions.
+# OPENAI_BASE_URL=                # e.g. http://chandra-vllm:8000
+# OPENAI_API_KEY=                 # empty for a local vLLM; token for hosted endpoints
+OPENAI_TIMEOUT_MS=120000          # page inference is seconds; keep < advert max_processing_ms
+# OPENAI_MAX_TOKENS=8192          # optional output ceiling (advert cap applies regardless)
+
+# ── OCR extras ────────────────────────────────────────────────────────────
+# Datalab upstream: the pre-Claim probe defaults to per_job liveness against
+# {DATALAB_BASE_URL}/api/v1/health - no further config needed. Set
+# REVISION_PROBE_MODE=off only if you want claims without the health gate.
+#
+# openai-vision upstream only (attestation upgrade path):
+# OCR_PROMPT_OVERRIDE=            # exact serving prompt, pinned at integration
+# HF_MODEL_REVISION=              # 40-hex checkpoint SHA; mismatch refuses Claim
+# REVISION_PROBE_URL=             # e.g. http://chandra-vllm:8000/info
+
+# ── Service ───────────────────────────────────────────────────────────────
+PORT=8080
+
+# ── Advert posting (run once per network, testnet first) ──────────────────
+# Model string convention: datalab upstream = "datalab/convert-accurate"
+# (honest: hosted engine + mode); self-hosted = "datalab-to/chandra-ocr-2@<sha7>".
+# Endpoint host: your deployed supplier's public HTTPS endpoint (deploy
+# target tracked outside this repo, not in this file).
+#
+# pnpm --filter @marketplace/supplier tx:post-advert -- \
+#   --ogmios-url $OGMIOS_URL \
+#   --priv-key $SUPPLIER_PRIV_KEY_HEX \
+#   --capability-id ocr.page.extract.chandra-ocr-2.v1 \
+#   --model "datalab/convert-accurate" \
+#   --max-output-tokens 8192 \
+#   --max-processing-ms 150000 \
+#   --price-lovelace 200000 \
+#   --endpoint-url https://<your-supplier-host> \
+#   --detail-uri https://<your-supplier-host>/chandra-detail.json \
+#   --detail-hash <sha256 of the detail JSON>
+#
+# max_processing_ms 150000 covers DATALAB_TIMEOUT_MS (120s) + chain legs.
+# Price 200000 (0.2 AP3X) is the fleet convention placeholder — a future
+# pricing decision supersedes it. Bonds default to 1 AP3X each.
+# The detail JSON template lives at supplier/chandra-detail.example.json.
+# Testnet: DNS is not required — a direct http://<host-ip>:<port> endpoint
+# works for the testnet advert (retire it after the drills).

supplier/.env.example

@@ -126,3 +126,9 @@ ADVERT_DETAIL_HASH=
 # `docker compose down && up -d` to apply.
 # ESCROW_REF_UTXO=<txhash>#0
 # ADVERT_REF_UTXO=<txhash>#1
+
+# ─── Optional: chat-session tuning (CAPABILITY_KIND=chat-session only) ─────
+# CHAT_IDLE_TIMEOUT_MS=300000  # auto-end sessions idle this long
+# MAX_CHAT_SESSIONS=1          # concurrent session slots (default 1)
+# CHAT_SETTLE_MODE=full        # "full" (Claim/Submit) or "ticket" (no supplier
+#                              # chain ops; buyer reclaims the Open escrow)

supplier/Dockerfile

@@ -27,6 +27,8 @@ RUN corepack enable && corepack prepare pnpm@10.33.2 --activate
 
 # Workspace manifests first to maximise layer cache reuse.
 COPY pnpm-workspace.yaml pnpm-lock.yaml package.json tsconfig.base.json .npmrc ./
+# pnpm-lock.yaml's patchedDependencies hash the patch files at install time.
+COPY patches ./patches
 COPY packages/shared/package.json ./packages/shared/package.json
 COPY supplier/package.json ./supplier/package.json
 # Buyer + indexer manifests are needed because pnpm-lock.yaml indexes the

supplier/chandra-detail.example.json

@@ -0,0 +1,21 @@
+{
+  "capability_id": "ocr.page.extract.chandra-ocr-2.v1",
+  "weights_source": "hf:datalab-to/chandra-ocr-2",
+  "model_revision": "<40-hex commit SHA; self-hosted (OCR_UPSTREAM=openai-vision) deployments only — the Datalab-hosted default (OCR_UPSTREAM=datalab) is a third-party API with no pinnable checkpoint, so this is unset there>",
+  "license": "modified OpenRAIL-M (weights); Apache-2.0 (code)",
+  "request_shape": {
+    "route": "POST /v1/ocr/extract",
+    "body": { "image_b64": "plain base64, no data: prefix", "mime": "image/png | image/jpeg | image/webp", "output_format": "markdown | html | json" },
+    "one_page_per_job": true
+  },
+  "receipt": {
+    "schema": "v1 (8 fields, Ed25519-signed)",
+    "response_hash": "sha256(canonical({content, output_format}))",
+    "verify": "GET /capability → pub_key_hex; verify signature over canonical(receipt)"
+  },
+  "revision_probe": {
+    "route": "<runtime info route reporting the loaded checkpoint; self-hosted (openai-vision) upstream only>",
+    "policy": "self-hosted (OCR_UPSTREAM=openai-vision) upstream: probe before every Claim, mismatch refuses the job. Datalab-hosted default (OCR_UPSTREAM=datalab): per-job liveness probe only, no checkpoint pin."
+  },
+  "operator_disclosure": "first-party (HAL8-operated supplier)"
+}