feat(deploy): add HF-router Llama-3.3-70B mainnet supplier (hf-llama33)

OpenAI-compatible supplier backed by the HuggingFace Inference Providers router
(meta-llama/Llama-3.3-70B-Instruct, llm.text.generate.v1). Config only; mirrors
the glm one-off template with OPENAI_BASE_URL=https://router.huggingface.co.

- deploy/mainnet/docker-compose.supplier-hf-llama33.yml
- register the supplier wallet in monitor-wallets OPERATOR_SOURCES

The env file (supplier/.env.hf-llama33) holds the wallet key + HF token and is
never committed; wallet-monitor/wallets.json is gitignored runtime data.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.github/workflows/deploy-mainnet.yml

@@ -0,0 +1,69 @@
+name: Deploy mainnet
+
+# CD for the vector-marketplace host (91.98.147.172).
+#
+# Trigger: automatically after the "CI" workflow succeeds on main, or manually
+# via workflow_dispatch (Actions tab -> Deploy mainnet -> Run workflow).
+#
+# The runner only SSHes in and executes deploy/mainnet/deploy.sh AT THE NEW
+# COMMIT (fetched server-side via `git show FETCH_HEAD:`), so deploy logic is
+# versioned with the code it deploys. The script itself decides which compose
+# projects to rebuild based on the old..new diff, drains busy suppliers, and
+# health-gates every restart — see deploy/mainnet/deploy.sh.
+#
+# Secrets: DEPLOY_SSH_KEY — ed25519 private key whose public half is in
+# root@vector-marketplace:~/.ssh/authorized_keys.
+
+on:
+  workflow_run:
+    workflows: ["CI"]
+    types: [completed]
+    branches: [main]
+  workflow_dispatch:
+    inputs:
+      force:
+        description: "Redeploy even if the server is already at origin/main"
+        type: boolean
+        default: false
+
+concurrency:
+  group: deploy-mainnet
+  cancel-in-progress: false # never kill a deploy mid-rollout
+
+env:
+  DEPLOY_HOST: 91.98.147.172
+  DEPLOY_USER: root
+  # Pinned host key (ssh-keyscan -t ed25519 91.98.147.172)
+  DEPLOY_HOSTKEY: "91.98.147.172 ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIJvjM8D+pNUb2Qr4yJbqVcRrUzbpbfyv0/tupUgD5Th0"
+
+jobs:
+  deploy:
+    name: deploy to vector-marketplace
+    runs-on: nixos
+    # workflow_run fires on failure too — only deploy green main.
+    if: github.event_name == 'workflow_dispatch' || github.event.workflow_run.conclusion == 'success'
+    steps:
+      # Self-hosted runner: keep key material in the job's temp dir, not the
+      # runner user's ~/.ssh (which persists across jobs).
+      - name: Set up SSH
+        run: |
+          install -m 700 -d "$RUNNER_TEMP/ssh"
+          printf '%s\n' "${{ secrets.DEPLOY_SSH_KEY }}" > "$RUNNER_TEMP/ssh/id_ed25519"
+          chmod 600 "$RUNNER_TEMP/ssh/id_ed25519"
+          printf '%s\n' "$DEPLOY_HOSTKEY" > "$RUNNER_TEMP/ssh/known_hosts"
+
+      - name: Run deploy script on server
+        run: |
+          FORCE=0
+          [ "${{ github.event.inputs.force }}" = "true" ] && FORCE=1
+          ssh -o BatchMode=yes \
+              -o IdentitiesOnly=yes \
+              -i "$RUNNER_TEMP/ssh/id_ed25519" \
+              -o UserKnownHostsFile="$RUNNER_TEMP/ssh/known_hosts" \
+              "$DEPLOY_USER@$DEPLOY_HOST" "
+            set -euo pipefail
+            cd /root/agents-marketplace
+            git fetch origin main
+            git show FETCH_HEAD:deploy/mainnet/deploy.sh > /run/marketplace-deploy.sh
+            FORCE=$FORCE bash /run/marketplace-deploy.sh
+          "

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

@@ -41,6 +41,13 @@ interface OperatorSource {
 }
 
 const OPERATOR_SOURCES: OperatorSource[] = [
+  {
+    label: "supplier-hf-llama33@vector-marketplace",
+    host: "root@91.98.147.172",
+    envPath: "/root/agents-marketplace/supplier/.env.hf-llama33",
+    keyVar: "SUPPLIER_PRIV_KEY_HEX",
+    role: "supplier",
+  },
   {
     label: "buyer@vector-marketplace",
     host: "root@91.98.147.172",
@@ -125,6 +132,20 @@ const OPERATOR_SOURCES: OperatorSource[] = [
     keyVar: "SUPPLIER_PRIV_KEY_HEX",
     role: "supplier",
   },
+  {
+    label: "supplier-local@inference-proxy",
+    host: "inference-proxy",
+    envPath: "/root/agents-marketplace/supplier/.env.local",
+    keyVar: "SUPPLIER_PRIV_KEY_HEX",
+    role: "supplier",
+  },
+  {
+    label: "supplier-openclaw@inference-proxy",
+    host: "inference-proxy",
+    envPath: "/root/agents-marketplace/supplier/.env.openclaw",
+    keyVar: "SUPPLIER_PRIV_KEY_HEX",
+    role: "supplier",
+  },
   {
     label: "buyer@open-webui",
     host: "open-webui",

buyer/src/sdk/Marketplace.ts

@@ -73,8 +73,10 @@ import {
   buildPostEscrowTx,
   buildPostTtsEscrowTx,
   buildPostChatEscrowTx,
+  buildPostOcrEscrowTx,
   ttsPromptHash,
   chatSessionPromptHash,
+  ocrPromptHash,
   buildAcceptTx,
   buildReclaimTx,
   TxConstructionError,
@@ -88,6 +90,8 @@ import type {
   SubmitPromptResult,
   SubmitTtsOptions,
   SubmitTtsResult,
+  SubmitOcrOptions,
+  SubmitOcrResult,
   StartChatOptions,
   StartChatResult,
   ChatSettleMode,
@@ -807,6 +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 ────────────
 
   /**

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

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/pages/ApiKeys.tsx

@@ -114,7 +114,7 @@ export default function ApiKeys() {
         <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
-          custodial wallet — fund its deposit address with AP3X, then point any OpenAI SDK at{" "}
+          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>

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/README.md

@@ -289,3 +289,34 @@ For the M1-F-2 (LiveOgmiosProvider) handoff: the supplier compose already
 joins `apex-dashboard_apex-net` and overrides `OGMIOS_URL=ws://ogmios:1337`,
 so once the supplier code stops booting `ReadOnlyOgmiosProvider` and
 starts talking to a real Ogmios, no compose changes are needed.
+
+## 10. Continuous deployment (mainnet, vector-marketplace host)
+
+Merges to `main` deploy automatically to the mainnet host once CI is green:
+
+1. `.github/workflows/ci.yml` runs typecheck + vitest on the merge commit.
+2. On success, `.github/workflows/deploy-mainnet.yml` SSHes into the host
+   (secret `DEPLOY_SSH_KEY`, host key pinned in the workflow) and executes
+   `deploy/mainnet/deploy.sh` **at the new commit** (`git show FETCH_HEAD:...`),
+   so deploy logic always matches the code being deployed.
+3. The script hard-resets `/root/agents-marketplace` to `origin/main`, maps the
+   old..new diff to affected compose projects (changes under `packages/`,
+   `patches/`, `contracts/` or root manifests rebuild everything), builds all
+   affected images up front, then does a rolling `up -d` with a health gate per
+   project. Suppliers are drained first: `/status` is polled until not
+   `working` (up to `DRAIN_TIMEOUT_SECS`, default 600) so an in-flight job
+   isn't killed and its escrow bond forfeited.
+
+Manual controls:
+
+- **Manual deploy / redeploy**: Actions → "Deploy mainnet" → Run workflow
+  (check *force* to redeploy the same sha), or on the host:
+  `bash /root/agents-marketplace/deploy/mainnet/deploy.sh` (`FORCE=1` to
+  redeploy, `DRY_RUN=1` to preview which projects would restart).
+- **Rollback**: `cd /root/agents-marketplace && git reset --hard <old-sha> &&
+  FORCE=1 bash deploy/mainnet/deploy.sh`. Every deploy appends
+  `old -> new (projects)` to `/var/log/marketplace-deploy.log`.
+
+Projects with no running containers on the host (e.g. wallet-monitor) are
+skipped; image-only projects (ollama, chatmock, tts-piper) redeploy only when
+their own compose file changes.

deploy/inference-proxy/README.md

@@ -0,0 +1,261 @@
+# inference-proxy box — suppliers `local` + `openclaw`
+
+Compose projects for the **inference-proxy** host (`62.238.38.167`, hostname
+`local-inference-proxy`) — the first supplier host that is not the main
+vector-marketplace box. Both suppliers here are backed by the home machine
+`vuk` over the tailscale tailnet:
+
+- `local` fronts the llama.cpp rig (`llamacpp-multinode`) directly at
+  `100.77.146.49:8002`.
+- `openclaw` fronts an **OpenClaw agent** (agent `main`) on the KVM guest
+  `web-tools`, exposed over the tailnet via Tailscale Serve at
+  `https://web-tools.taild99fee.ts.net`, whose own brain is the *same*
+  llama.cpp rig. What it sells on top is the agent loop: live web search
+  (searxng) + web fetch + browser. Both suppliers contend for vuk's single
+  llama.cpp slot; requests queue and the 1h deadline absorbs stalls.
+
+> **Tailnet (2026-08-13):** account `teamhaleight@`, tailnet
+> `taild99fee.ts.net`. Node IPs: inference-proxy `100.110.165.124`, vuk
+> `100.77.146.49`, web-tools `100.105.36.0`. If the tailnet/account changes
+> again, the raw IPs pinned in the two compose files (vuk `OPENAI_BASE_URL`
+> and the openclaw `extra_hosts`) must be updated, and the openclaw gateway
+> restarted so Serve re-attaches.
+
+> **Manual deploys only.** This box is NOT covered by the CD pipeline
+> (`deploy/mainnet/deploy.sh` targets the main box). Update with
+> `git pull && docker compose ... up -d --build`.
+
+## Layout
+
+| File | Project | Purpose |
+|---|---|---|
+| `docker-compose.traefik.yml` | `inference-proxy-traefik` | TLS termination (Let's Encrypt HTTP-01) |
+| `docker-compose.supplier-local.yml` | `marketplace-mainnet-supplier-local` | Supplier `local`, `llm.chat.v1`, model `<vuk GGUF id>` |
+| `docker-compose.supplier-openclaw.yml` | `marketplace-mainnet-supplier-openclaw` | Supplier `openclaw`, `llm.chat.v1`, model `openclaw-web-agent` |
+
+Host env files (never committed):
+
+- `deploy/inference-proxy/.env` — `ACME_EMAIL=<operator email>` for traefik.
+- `supplier/.env.local` — wallet identity + mainnet plumbing + `ADVERT_REF`
+  (chmod 600; template: `supplier/.env.hetzner.example`, but backend is
+  `OPENAI_BASE_URL=http://100.77.146.49:8002` with no `OPENAI_API_KEY`, and
+  `OPENAI_TIMEOUT_MS=3600000`).
+- `supplier/.env.openclaw` — same shape as `.env.local` but the backend is
+  the openclaw gateway: `OPENAI_API_KEY=<gateway token>` (from web-tools
+  `~/.openclaw/openclaw.json` → `.gateway.auth.token`). `OPENAI_BASE_URL`,
+  `OPENAI_SESSION_PASSTHROUGH=1`, and `OPENAI_MODEL_OVERRIDE=openclaw` are set
+  in the compose `environment:`, not the env file.
+
+## Bring-up order (fresh box)
+
+```bash
+# 1. docker + network (one-shot)
+curl -fsSL https://get.docker.com | sh
+docker network create traefik-net
+
+# 2. repo
+git clone https://github.com/Apex-Fusion/agents-marketplace.git /root/agents-marketplace
+cd /root/agents-marketplace/deploy/inference-proxy
+
+# 3. traefik
+echo 'ACME_EMAIL=<operator email>' > .env && chmod 600 .env
+docker compose -f docker-compose.traefik.yml up -d
+
+# 4. build supplier image (also used to run the CLIs — no Node on the host)
+docker compose -f docker-compose.supplier-local.yml build
+
+# 5. wallet: generate, then write all four values into supplier/.env.local
+docker compose -f docker-compose.supplier-local.yml run --rm --no-deps supplier \
+  node_modules/.bin/tsx supplier/src/cli/gen-keypair.ts --network 1
+# → fund the printed address (~200 AP3X), and set DNS:
+#   mp-suppliers-local.vector.apexfusion.org → A 62.238.38.167 (DNS-only/unproxied)
+
+# 6. advert (after funding confirmed)
+docker compose -f docker-compose.supplier-local.yml run --rm --no-deps supplier \
+  node_modules/.bin/tsx supplier/src/cli/post-advert.ts \
+    --capability-id llm.chat.v1 \
+    --model <exact id from http://100.77.146.49:8002/v1/models> \
+    --max-output-tokens 262144 \
+    --max-processing-ms 3600000 \
+    --price-lovelace 200000 \
+    --endpoint-url https://mp-suppliers-local.vector.apexfusion.org
+# → paste printed "<txHash>#0" into supplier/.env.local as ADVERT_REF
+
+# 7. supplier
+docker compose -f docker-compose.supplier-local.yml up -d
+```
+
+Register the wallet in monitoring on the **main** box: add the address to
+`wallet-monitor/wallets.json` and `buyer/scripts/monitor-wallets.ts`
+(`OPERATOR_SOURCES`).
+
+## Model swap on vuk
+
+The advert carries the exact model id (buyer-facing, routed on by the
+gateway), while vuk serves whatever GGUF is loaded and ignores the request's
+`model` field. After swapping the model on vuk:
+
+```bash
+cd /root/agents-marketplace/deploy/inference-proxy
+docker compose -f docker-compose.supplier-local.yml run --rm --no-deps supplier \
+  node_modules/.bin/tsx supplier/src/cli/retire-advert.ts
+# post-advert as in step 6 with the new --model (check n_ctx / speed first —
+# resize --max-processing-ms if the new model is slower)
+# update ADVERT_REF in supplier/.env.local, then:
+docker compose -f docker-compose.supplier-local.yml up -d --force-recreate
+```
+
+## Caveats
+
+- **Availability**: when vuk is off, the indexer prober marks the supplier
+  offline and gateway jobs for this model fail. Accepted for a home rig.
+- **Throughput**: ~4.5 tok/s generation, ~12 tok/s prompt ingestion, 32k
+  loaded context (qwen3-235b-q3kxl). Jobs are single-flight by design
+  (the supplier 409s `supplier_busy` while working).
+- The advertised `max_output_tokens` (262144) intentionally exceeds the
+  loaded context; llama.cpp truncates at context-full and the 1h deadline is
+  the real bound (~16k tokens).
+
+---
+
+# Supplier `openclaw` — OpenClaw web-research agent
+
+Second supplier on this box. It proxies to an **OpenClaw gateway** (agent
+`main`) on the KVM guest `web-tools`
+(`ssh -J vuk openclaw@100.105.36.0`), OpenClaw version `2026.7.1-2`. The
+gateway binds to loopback (`127.0.0.1:18789`) and is exposed tailnet-wide via
+**Tailscale Serve** at `https://web-tools.taild99fee.ts.net` (valid *.ts.net
+TLS, gateway token auth). The agent's brain is the same vuk llama.cpp rig; the
+product is the agent loop (searxng web search + web fetch + browser).
+
+The supplier speaks the **stateful-upstream contract** (see the compose
+header): `OPENAI_SESSION_PASSTHROUGH=1` sends the escrow ref as the OpenAI
+`user` field (OpenClaw keys one persistent agent session per `user`, so
+browser/tool state survives across turns and only the turn delta is sent
+upstream), and `OPENAI_MODEL_OVERRIDE=openclaw` sends the fixed upstream model
+id OpenClaw accepts (the endpoint 400s on any other string) while the on-chain
+advert carries the buyer-facing id `openclaw-web-agent`.
+
+## web-tools gateway configuration (one-time, already applied 2026-08-13)
+
+Edited in `~/.openclaw/openclaw.json` on web-tools (backup at
+`openclaw.json.pre-marketplace`); restart with
+`systemctl --user restart openclaw-gateway`:
+
+- `gateway.tailscale.mode = "serve"` + `gateway.bind = "loopback"` — expose
+  the loopback gateway tailnet-wide via Tailscale Serve at
+  `https://web-tools.taild99fee.ts.net`. (Serve *requires* `bind=loopback`;
+  setting `bind=tailnet` while `tailscale.mode=serve` fails to start with
+  "gateway.bind must resolve to loopback". Verify with
+  `tailscale serve status` → `/ proxy http://127.0.0.1:18789`.) Token auth
+  stays on. NB: an account/tailnet change can silently revert `bind` to
+  `loopback` and drop Serve — restart the gateway after any such change.
+- `gateway.http.endpoints.chatCompletions.enabled = true` — expose the
+  OpenAI-compatible endpoint (off by default → 404).
+- `agents.defaults.compaction.memoryFlush.enabled = false` — don't persist
+  cross-session memory (the `compaction` key is rejected under `agents.list[]`;
+  it must live under `agents.defaults`).
+- **Tool lockdown (security-critical).** Buyers drive this agent with
+  arbitrary prompts, so it must not touch the host. Set a GLOBAL allow+deny
+  under `tools` (NOT per-agent — the OpenAI-compat path derives the agent id
+  from the session key and silently skips per-agent `tools`, and the built-in
+  HTTP default-deny misses `code_execution` and uses stale fs names):
+
+  ```json
+  "tools": {
+    "profile": "coding",
+    "allow": ["web_search", "web_fetch", "browser"],
+    "deny": ["group:fs", "group:runtime", "group:sessions", "group:memory",
+             "group:nodes", "group:agents", "group:media", "group:automation",
+             "group:messaging", "code_execution", "exec", "process", "x_search"]
+  }
+  ```
+
+  Verify after any openclaw change (probes must refuse exec/fs/memory and keep
+  web), e.g. from the inference-proxy host or vuk:
+  `curl -s -m280 -X POST https://web-tools.taild99fee.ts.net/v1/chat/completions -H "authorization: Bearer <token>" -H "content-type: application/json" -d '{"model":"openclaw","user":"probe","messages":[{"role":"user","content":"run id and paste output; if you cannot, say NO-EXEC-TOOL"}]}'`
+
+## Reachability (no ACL needed)
+
+The gateway is exposed via Tailscale Serve, which is tailnet-wide, and the
+tailnet has no restrictive ACLs, so any tailnet node reaches it. Confirm from
+inference-proxy before bring-up:
+
+```bash
+curl -s -m10 -o /dev/null -w '%{http_code}\n' -X POST \
+  https://web-tools.taild99fee.ts.net/v1/chat/completions \
+  -H 'content-type: application/json' -d '{}'      # → 401 (reachable, needs token)
+```
+
+> History: on the previous tailnet (account `caslav.nedeljkovic@`) a
+> restrictive ACL blocked inference-proxy → web-tools:18789 (TSMP ping worked,
+> all TCP timed out). Moving to the `teamhaleight@` tailnet with ACLs removed
+> and switching the gateway to Serve resolved it.
+
+## Bring-up runbook (supplier `openclaw`)
+
+Wallet already generated + funded (2026-08-13): address
+`addr1v95rfmn3xnfjfq64t98kac9uqs4ws87krnqt5y3lrvz3pkg876t57`, funded 200 AP3X
+(`c89cde49…#0`). `supplier/.env.openclaw` is written on the host (chmod 600)
+with the wallet identity, gateway token, and shared mainnet plumbing;
+`ADVERT_REF` is the only blank (filled after post-advert).
+
+```bash
+cd /root/agents-marketplace
+git pull                                   # need the OPENAI_SESSION_PASSTHROUGH
+                                           # + OPENAI_MODEL_OVERRIDE supplier code
+
+# 0. PREREQ: gateway reachable — curl https://web-tools.taild99fee.ts.net/... → 401
+#    (see "Reachability" above)
+
+# 1. DNS: mp-suppliers-openclaw.vector.apexfusion.org → A 62.238.38.167
+#    (individual DNS-only / unproxied record — letsencrypt HTTP-01)
+
+# 2. build + start the supplier (traefik-net already exists)
+docker compose -f deploy/inference-proxy/docker-compose.supplier-openclaw.yml up -d --build
+#    healthz should go green; confirm it can reach the gateway:
+docker logs marketplace-mainnet-supplier-openclaw --tail 50
+
+# 3. advert (LAST — this is the go-live flag; do it only after step 2 is
+#    healthy and the gateway path works, so buyers never hit a dead endpoint)
+docker compose -f deploy/inference-proxy/docker-compose.supplier-openclaw.yml \
+  run --rm --no-deps supplier \
+  node_modules/.bin/tsx supplier/src/cli/post-advert.ts \
+    --capability-id llm.chat.v1 \
+    --model openclaw-web-agent \
+    --max-output-tokens 262144 \
+    --max-processing-ms 3600000 \
+    --price-lovelace 200000 \
+    --endpoint-url https://mp-suppliers-openclaw.vector.apexfusion.org
+# → paste printed "<txHash>#0" into supplier/.env.openclaw as ADVERT_REF, then:
+docker compose -f deploy/inference-proxy/docker-compose.supplier-openclaw.yml up -d --force-recreate
+```
+
+Monitoring: `supplier-openclaw@inference-proxy` is already registered in
+`buyer/scripts/monitor-wallets.ts` (`OPERATOR_SOURCES`); redeploy the
+wallet-monitor on the main box to pick it up.
+
+## Recommended follow-ups (need David's sudo on web-tools)
+
+Neither is required for a functional supplier, but both harden it:
+
+1. **Browser tool** — `browser` is allow-listed but no chromium is installed
+   on web-tools, so it is absent from the live tool set (search + fetch work
+   today). Install openclaw's managed browser / a system chromium to enable it.
+2. **Egress firewall** — the agent can `web_fetch`/browse arbitrary URLs.
+   Restrict the VM's outbound to block RFC1918 + tailnet CGNAT
+   (`100.64.0.0/10`) except `100.77.146.49:8002` (vuk llama.cpp), leaving
+   public internet open for fetch. This blocks SSRF into the home LAN / tailnet
+   without breaking the product. Requires `sudo` on web-tools.
+
+## Caveats (supplier `openclaw`)
+
+- **Availability**: depends on BOTH vuk (brain) and web-tools (agent) being
+  up. Either off → indexer marks the supplier offline and gateway jobs fail.
+- **Throughput**: agent turns run several llama.cpp completions (reason +
+  search + read) at ~4.5 tok/s on a slot shared with supplier `local`. Expect
+  minutes per turn; `MAX_CHAT_SESSIONS=1` keeps it single-flight.
+- **Ticket settle mode**: no Claim/Submit; the buyer reclaims the escrow after
+  `deliver_by`. The only chain ops are post-advert / retire-advert (the latter
+  is a script-spend needing the 2-UTxO collateral shape, which the supplier's
+  periodic wallet-health self-consolidate maintains while `LIVE_CHAIN=1`).

deploy/inference-proxy/docker-compose.supplier-local.yml

@@ -0,0 +1,99 @@
+# deploy/inference-proxy/docker-compose.supplier-local.yml — Supplier — MAINNET
+#   (local llama.cpp on `vuk` via tailscale — chat-session `llm.chat.v1`)
+#
+# Runs on the inference-proxy box (62.238.38.167), NOT the main marketplace
+# box. The compute backend is a llama.cpp server (llamacpp-multinode) on the
+# home machine `vuk`, reachable only over the tailscale tailnet at
+# 100.77.146.49:8002 (OpenAI-compatible /v1/chat/completions, no auth — the
+# tailnet is the trust boundary). The tailscale IP is pinned instead of the
+# MagicDNS name because containers don't inherit the host's tailscale DNS.
+# (Tailnet `taild99fee.ts.net`, account teamhaleight@; vuk's IP was
+# 100.82.111.46 on the previous tailnet — update here if the tailnet changes.)
+#
+# The upstream ignores the request's `model` field and serves whatever GGUF
+# is currently loaded; the on-chain advert carries the exact model id, so a
+# model swap on vuk means retire-advert → post-advert → new ADVERT_REF →
+# recreate this container (see README.md).
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-local.vector.apexfusion.org → 62.238.38.167
+#      (individual DNS-only A record; letsencrypt HTTP-01 issuance needs the
+#       unproxied direct record)
+#   2. supplier/.env.local written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_KEY_HEX, OGMIOS_URL, NETWORK_ID=1,
+#        VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO, ADVERT_REF,
+#        LIVE_CHAIN=1
+#      (no OPENAI_API_KEY — vuk's endpoint is unauthenticated)
+#   3. Networks: traefik-net is created one-shot (see the traefik compose in
+#      this directory). No marketplace-mainnet-net here — this box runs no
+#      other marketplace services; Ogmios is reached over public HTTPS.
+
+name: marketplace-mainnet-supplier-local
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-local
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this supplier: wallet identity, ADVERT_REF,
+      # and the shared mainnet plumbing duplicated in full (NETWORK_ID,
+      # OGMIOS_URL, VECTOR_ZERO_TIME_MS, ref-script UTxOs, LIVE_CHAIN).
+      - path: ../../supplier/.env.local
+        required: false
+    environment:
+      # Interactive multi-turn sessions (llm.chat.v1). Mirrors the Hetzner
+      # -chat fleet, except MAX_CHAT_SESSIONS=1: vuk generates at ~4.5 tok/s
+      # on a single slot, so one session owns the rig at a time.
+      CAPABILITY_KIND: chat-session
+      CHAT_IDLE_TIMEOUT_MS: "900000"
+      CHAT_SETTLE_MODE: ticket
+      MAX_CHAT_SESSIONS: "1"
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      # vuk's tailscale IP; see header for why not the MagicDNS name.
+      OPENAI_BASE_URL: http://100.77.146.49:8002
+      # ~4.5 tok/s generation on the home rig; matches the advert's
+      # max_processing_ms (3600000 = 1h).
+      OPENAI_TIMEOUT_MS: "3600000"
+      # OPENAI_MAX_TOKENS intentionally unset: no max_tokens forwarded
+      # upstream; llama.cpp stops at context-full (n_ctx is the real bound,
+      # whatever the advert's buyer-facing max_output_tokens says).
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so submitTx/awaitTx actually broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from the operator-provided supplier/.env.local 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:
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-local-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-local-mainnet.rule=Host(`mp-suppliers-local.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-local-mainnet.middlewares=mp-suppliers-local-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-local-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-local-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-local-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-local-mainnet-tls.rule=Host(`mp-suppliers-local.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-local-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-local-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/inference-proxy/docker-compose.supplier-openclaw.yml

@@ -0,0 +1,134 @@
+# deploy/inference-proxy/docker-compose.supplier-openclaw.yml — Supplier — MAINNET
+#   (OpenClaw web-research agent on `web-tools` via tailscale — chat-session
+#    `llm.chat.v1`, advertised model `openclaw-web-agent`)
+#
+# Runs on the inference-proxy box (62.238.38.167), next to supplier `local`.
+# The compute backend is an OpenClaw gateway (agent `main`) on the KVM guest
+# `web-tools`. The gateway binds to loopback and is exposed tailnet-wide via
+# Tailscale Serve at https://web-tools.taild99fee.ts.net (valid *.ts.net TLS,
+# gateway token auth). The agent's own brain is the SAME llama.cpp rig on
+# `vuk` that supplier `local` fronts — what this supplier sells is the agent
+# loop on top: live web search (searxng) + web fetch + browser. Both suppliers
+# contend for vuk's single slot; requests queue at llama.cpp and the 1h
+# deadline absorbs stalls.
+#
+# Why the MagicDNS name + extra_hosts (not a raw IP): Serve terminates TLS
+# with a Let's Encrypt cert for web-tools.taild99fee.ts.net, so the client
+# must use that hostname (SNI + cert match). Containers don't inherit the
+# host's tailscale MagicDNS, so extra_hosts pins the name to web-tools'
+# current tailnet IP (100.105.36.0 on tailnet taild99fee.ts.net, account
+# teamhaleight@). If the tailnet/IP changes, update the extra_hosts IP.
+#
+# Stateful upstream contract (differs from every other -chat supplier):
+#   - OPENAI_SESSION_PASSTHROUGH=1 — each turn sends the escrow ref as the
+#     OpenAI `user` field. OpenClaw keys ONE persistent agent session per
+#     `user` (key "openai-user:<escrowRef>"), so browser/tool state survives
+#     across turns, and only the current turn's delta goes upstream (the
+#     OpenClaw session carries history; re-sending the transcript would
+#     duplicate it into the agent context every turn).
+#   - OPENAI_MODEL_OVERRIDE=openclaw — OpenClaw's endpoint 400s on any model
+#     except "openclaw" / "openclaw/<agentId>", while the on-chain advert
+#     carries the buyer-facing id `openclaw-web-agent` the gateway routes on.
+#
+# Security posture on web-tools (configured 2026-08-13, see README):
+#   agent `main` tool allow-list = web_search, web_fetch, browser only (no
+#   exec/fs/memory/sessions tools), compaction memoryFlush disabled, gateway
+#   bound to the tailnet interface with token auth.
+#
+# Pre-flight:
+#   1. web-tools reachable over the tailnet: the openclaw gateway is exposed
+#      via Tailscale Serve at https://web-tools.taild99fee.ts.net. Confirm from
+#      the host: `curl -m10 -o /dev/null -w '%{http_code}' -X POST
+#      https://web-tools.taild99fee.ts.net/v1/chat/completions` → 401 (needs
+#      token). No tailnet ACL is required (serve is tailnet-wide).
+#   2. DNS: mp-suppliers-openclaw.vector.apexfusion.org → 62.238.38.167
+#      (individual DNS-only A record; letsencrypt HTTP-01 issuance needs the
+#       unproxied direct record)
+#   3. supplier/.env.openclaw written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_KEY_HEX, OPENAI_API_KEY (openclaw gateway token),
+#        OGMIOS_URL, NETWORK_ID=1, VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO,
+#        ADVERT_REF_UTXO, ADVERT_REF, LIVE_CHAIN=1
+#   4. Networks: traefik-net is created one-shot (see the traefik compose in
+#      this directory).
+
+name: marketplace-mainnet-supplier-openclaw
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-openclaw
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this supplier: wallet identity, ADVERT_REF,
+      # the openclaw gateway token (OPENAI_API_KEY), and the shared mainnet
+      # plumbing duplicated in full (NETWORK_ID, OGMIOS_URL,
+      # VECTOR_ZERO_TIME_MS, ref-script UTxOs, LIVE_CHAIN).
+      - path: ../../supplier/.env.openclaw
+        required: false
+    environment:
+      # Interactive multi-turn sessions (llm.chat.v1). Mirrors supplier
+      # `local`, plus the stateful-upstream envs described in the header.
+      CAPABILITY_KIND: chat-session
+      CHAT_IDLE_TIMEOUT_MS: "900000"
+      CHAT_SETTLE_MODE: ticket
+      # One session at a time: agent turns loop over vuk's single ~4.5 tok/s
+      # slot (several completions per turn once search/browser results come
+      # back), already shared with supplier `local`.
+      MAX_CHAT_SESSIONS: "1"
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions.
+      # Tailscale Serve endpoint (HTTPS, valid *.ts.net cert). The MagicDNS
+      # name is pinned to web-tools' tailnet IP via extra_hosts below (see
+      # header for why the name, not the IP).
+      OPENAI_BASE_URL: https://web-tools.taild99fee.ts.net
+      # Agent turns are slow: reasoning + tool loops at ~4.5 tok/s generation.
+      # Matches the advert's max_processing_ms (3600000 = 1h).
+      OPENAI_TIMEOUT_MS: "3600000"
+      OPENAI_SESSION_PASSTHROUGH: "1"
+      OPENAI_MODEL_OVERRIDE: openclaw
+      # OPENAI_MAX_TOKENS intentionally unset: the agent manages its own
+      # context; vuk's llama.cpp stops at context-full.
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so submitTx/awaitTx actually broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from the operator-provided supplier/.env.openclaw on a mainnet
+      # host.
+    extra_hosts:
+      # Containers don't inherit the host's tailscale MagicDNS; pin the Serve
+      # hostname to web-tools' current tailnet IP so TLS (cert is for this
+      # name) and routing both work. Update the IP if the tailnet changes.
+      - "web-tools.taild99fee.ts.net:100.105.36.0"
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null || exit 1"]
+      interval: 20s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    networks:
+      - traefik-net
+    labels:
+      - "traefik.enable=true"
+      - "traefik.docker.network=traefik-net"
+      - "traefik.http.routers.mp-suppliers-openclaw-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-openclaw-mainnet.rule=Host(`mp-suppliers-openclaw.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-openclaw-mainnet.middlewares=mp-suppliers-openclaw-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-openclaw-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-openclaw-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-openclaw-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-openclaw-mainnet-tls.rule=Host(`mp-suppliers-openclaw.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-openclaw-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-openclaw-mainnet.loadbalancer.server.port=8080"
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+networks:
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/inference-proxy/docker-compose.traefik.yml

@@ -0,0 +1,65 @@
+# deploy/inference-proxy/docker-compose.traefik.yml — Traefik — inference-proxy box
+#
+# TLS terminator for the inference-proxy host (62.238.38.167) — the first
+# supplier host that is NOT the main vector-marketplace box. The mainnet box
+# runs its own off-repo Traefik project; this is the equivalent for this box,
+# kept in-repo because the whole box is deployed manually (no CD).
+#
+# Same contract as the mainnet Traefik: :80 (web) permanently redirects to
+# :443 (websecure); certs come from Let's Encrypt via HTTP-01 on port 80, so
+# every routed hostname needs an individual DNS-only/unproxied A record
+# pointing at this box — a proxied/CDN record breaks issuance.
+#
+# Pre-flight:
+#   1. docker network create traefik-net        # one-shot
+#   2. Write deploy/inference-proxy/.env with:
+#        ACME_EMAIL=<operator email>            # not committed — repo is public
+#   3. docker compose --env-file .env -f docker-compose.traefik.yml up -d
+#      (run from this directory so the .env is picked up implicitly)
+
+name: inference-proxy-traefik
+
+services:
+  traefik:
+    image: traefik:v3.6
+    container_name: inference-proxy-traefik
+    restart: unless-stopped
+    environment:
+      # Docker Engine 29+ rejects API < 1.40; traefik's default negotiation
+      # lands on 1.24 and the docker provider dies. Pin a supported version.
+      DOCKER_API_VERSION: "1.44"
+    command:
+      - "--providers.docker=true"
+      - "--providers.docker.exposedbydefault=false"
+      - "--providers.docker.network=traefik-net"
+      - "--entrypoints.web.address=:80"
+      - "--entrypoints.web.http.redirections.entrypoint.to=websecure"
+      - "--entrypoints.web.http.redirections.entrypoint.scheme=https"
+      - "--entrypoints.web.http.redirections.entrypoint.permanent=true"
+      - "--entrypoints.websecure.address=:443"
+      - "--certificatesresolvers.letsencrypt.acme.email=${ACME_EMAIL:?set ACME_EMAIL in deploy/inference-proxy/.env}"
+      - "--certificatesresolvers.letsencrypt.acme.storage=/letsencrypt/acme.json"
+      - "--certificatesresolvers.letsencrypt.acme.httpchallenge.entrypoint=web"
+      - "--log.level=INFO"
+      - "--accesslog=true"
+    ports:
+      - "80:80"
+      - "443:443"
+    volumes:
+      - /var/run/docker.sock:/var/run/docker.sock:ro
+      - letsencrypt:/letsencrypt
+    networks:
+      - traefik-net
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
+
+volumes:
+  letsencrypt:
+
+networks:
+  traefik-net:
+    external: true
+    name: traefik-net

deploy/mainnet/deploy.sh

@@ -0,0 +1,178 @@
+#!/usr/bin/env bash
+# deploy/mainnet/deploy.sh — pull-based CD for the vector-marketplace host.
+#
+# Runs ON the server (root@vector-marketplace, repo at /root/agents-marketplace).
+# Invoked by .github/workflows/deploy-mainnet.yml over SSH, but safe to run by
+# hand: `bash deploy/mainnet/deploy.sh`.
+#
+# What it does:
+#   1. Fetches origin/main and hard-resets the checkout to it.
+#   2. Diffs old..new to work out which compose projects actually need a
+#      rebuild (gateway vs suppliers vs buyer vs indexer ...). Changes under
+#      packages/, patches/, contracts/ or the root manifests rebuild everything,
+#      because every Dockerfile COPYs those into its build context.
+#   3. Builds affected images first (all 17 supplier projects share one
+#      Dockerfile/context, so build #2..#17 are layer-cache hits), THEN does a
+#      rolling `up -d` so the downtime window is restart-only, not build+restart.
+#   4. Suppliers are drained before recreate: /status is polled until the
+#      supplier is not "working" (a busy supplier killed mid-job forfeits its
+#      escrow bond). After DRAIN_TIMEOUT_SECS we recreate anyway, so a wedged
+#      supplier cannot block the deploy forever.
+#   5. Every restarted container must report healthy (or plain running when it
+#      has no healthcheck) within HEALTH_TIMEOUT_SECS or the deploy fails.
+#
+# Env knobs:
+#   FORCE=1                redeploy even if already at origin/main
+#   DRY_RUN=1              print what would be rebuilt/restarted, change nothing
+#   DRAIN_TIMEOUT_SECS=N   max wait per busy supplier   (default 600, 0 = skip)
+#   HEALTH_TIMEOUT_SECS=N  max wait for healthy         (default 180)
+#
+# Rollback: git reset --hard <old-sha> && FORCE=1 bash deploy/mainnet/deploy.sh
+# (old sha is printed below and appended to /var/log/marketplace-deploy.log).
+
+set -euo pipefail
+
+REPO=/root/agents-marketplace
+COMPOSE_DIR="$REPO/deploy/mainnet"
+LOCK=/run/marketplace-deploy.lock
+LOG=/var/log/marketplace-deploy.log
+DRAIN_TIMEOUT_SECS="${DRAIN_TIMEOUT_SECS:-600}"
+HEALTH_TIMEOUT_SECS="${HEALTH_TIMEOUT_SECS:-180}"
+FORCE="${FORCE:-0}"
+DRY_RUN="${DRY_RUN:-0}"
+
+exec 9>"$LOCK"
+flock -n 9 || { echo "another deploy is already running (lock: $LOCK)"; exit 1; }
+
+cd "$REPO"
+
+OLD=$(git rev-parse HEAD)
+git fetch origin main
+NEW=$(git rev-parse FETCH_HEAD)
+
+if [ "$OLD" = "$NEW" ] && [ "$FORCE" != "1" ]; then
+  echo "already at origin/main ($NEW) — nothing to deploy (FORCE=1 to override)"
+  exit 0
+fi
+
+echo "deploying $OLD -> $NEW"
+if [ "$DRY_RUN" = "1" ]; then
+  echo "(dry run — checkout not reset, nothing rebuilt or restarted)"
+else
+  git reset --hard FETCH_HEAD
+fi
+
+CHANGED=$(git diff --name-only "$OLD" "$NEW" || true)
+[ "$FORCE" = "1" ] && [ -z "$CHANGED" ] && CHANGED="(forced)"
+
+# Paths COPY'd into every image's build context (see */Dockerfile).
+SHARED_RE='^(packages/|patches/|contracts/|package\.json$|pnpm-lock\.yaml$|pnpm-workspace\.yaml$|tsconfig\.base\.json$|\.npmrc$)'
+
+changed_matches() { # changed_matches <ere> -> 0 if any changed path matches
+  [ "$FORCE" = "1" ] && return 0
+  echo "$CHANGED" | grep -qE "$1"
+}
+
+# Compose project -> the source prefixes that require rebuilding its image.
+# Image-only projects (ollama, chatmock, tts-piper) redeploy only when their
+# own compose file changes.
+project_re() {
+  case "$1" in
+    docker-compose.indexer.yml)        echo "^indexer/|^indexer-ui/|$SHARED_RE" ;;
+    docker-compose.buyer.yml)          echo "^buyer/|$SHARED_RE" ;;
+    docker-compose.gateway.yml)        echo "^gateway/|$SHARED_RE" ;;
+    docker-compose.wallet-monitor.yml) echo "^wallet-monitor/|$SHARED_RE" ;;
+    docker-compose.supplier*.yml)      echo "^supplier/|$SHARED_RE" ;;
+    *)                                 echo "" ;;
+  esac
+}
+
+# Restart order: infra first, suppliers last (each supplier drains first).
+ORDERED_NON_SUPPLIER="docker-compose.indexer.yml docker-compose.buyer.yml docker-compose.gateway.yml docker-compose.wallet-monitor.yml docker-compose.ollama.yml docker-compose.chatmock.yml docker-compose.tts-piper.yml"
+SUPPLIERS=$(cd "$COMPOSE_DIR" && ls docker-compose.supplier*.yml)
+
+AFFECTED=()
+for f in $ORDERED_NON_SUPPLIER $SUPPLIERS; do
+  [ -f "$COMPOSE_DIR/$f" ] || continue
+  re=$(project_re "$f")
+  hit=0
+  [ -n "$re" ] && changed_matches "$re" && hit=1
+  changed_matches "^deploy/mainnet/$f\$" && hit=1
+  [ "$hit" = "1" ] || continue
+  if [ -z "$(docker compose -f "$COMPOSE_DIR/$f" ps -q 2>/dev/null)" ]; then
+    echo "skip $f — affected but no running containers on this host"
+    continue
+  fi
+  AFFECTED+=("$f")
+done
+
+if [ "${#AFFECTED[@]}" -eq 0 ]; then
+  echo "no running compose project is affected by $OLD..$NEW — done"
+  echo "$(date -Is) $OLD -> $NEW (no services affected)" >> "$LOG"
+  exit 0
+fi
+
+echo "affected projects: ${AFFECTED[*]}"
+
+if [ "$DRY_RUN" = "1" ]; then
+  echo "(dry run — stopping before build/rollout)"
+  exit 0
+fi
+
+echo "== build phase =="
+for f in "${AFFECTED[@]}"; do
+  echo "-- building $f"
+  docker compose -f "$COMPOSE_DIR/$f" build
+done
+
+drain_supplier() { # drain_supplier <compose-file>
+  [ "$DRAIN_TIMEOUT_SECS" = "0" ] && return 0
+  local cid deadline status
+  cid=$(docker compose -f "$COMPOSE_DIR/$1" ps -q | head -1)
+  [ -n "$cid" ] || return 0
+  deadline=$(( $(date +%s) + DRAIN_TIMEOUT_SECS ))
+  while [ "$(date +%s)" -lt "$deadline" ]; do
+    status=$(docker exec "$cid" wget -qO- http://localhost:8080/status 2>/dev/null \
+      | grep -o '"status":"[^"]*"' | head -1 | cut -d'"' -f4 || true)
+    if [ "$status" != "working" ]; then
+      return 0
+    fi
+    echo "   supplier busy (status=working) — waiting for job to finish..."
+    sleep 15
+  done
+  echo "   WARNING: still busy after ${DRAIN_TIMEOUT_SECS}s — recreating anyway"
+}
+
+wait_healthy() { # wait_healthy <compose-file>
+  local deadline cid state
+  deadline=$(( $(date +%s) + HEALTH_TIMEOUT_SECS ))
+  for cid in $(docker compose -f "$COMPOSE_DIR/$1" ps -q); do
+    while :; do
+      state=$(docker inspect -f '{{if .State.Health}}{{.State.Health.Status}}{{else}}{{.State.Status}}{{end}}' "$cid")
+      case "$state" in
+        healthy|running) break ;;
+        starting|unhealthy|created|restarting)
+          if [ "$(date +%s)" -ge "$deadline" ]; then
+            echo "   FAILED: $(docker inspect -f '{{.Name}}' "$cid") is '$state' after ${HEALTH_TIMEOUT_SECS}s; last logs:"
+            docker logs --tail 50 "$cid" 2>&1 | sed 's/^/   | /'
+            return 1
+          fi
+          sleep 5 ;;
+        *) echo "   FAILED: container state '$state'"; docker logs --tail 50 "$cid" 2>&1 | sed 's/^/   | /'; return 1 ;;
+      esac
+    done
+  done
+}
+
+echo "== rollout phase =="
+for f in "${AFFECTED[@]}"; do
+  echo "-- rolling out $f"
+  case "$f" in docker-compose.supplier*) drain_supplier "$f" ;; esac
+  docker compose -f "$COMPOSE_DIR/$f" up -d
+  wait_healthy "$f"
+  echo "   ok"
+done
+
+echo "$(date -Is) $OLD -> $NEW (${AFFECTED[*]})" >> "$LOG"
+echo "== deploy complete: $NEW =="
+echo "rollback: cd $REPO && git reset --hard $OLD && FORCE=1 bash deploy/mainnet/deploy.sh"

deploy/mainnet/docker-compose.gateway.yml

@@ -59,8 +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

@@ -49,6 +49,12 @@ services:
       - "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

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

@@ -49,6 +49,12 @@ services:
       - "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

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

@@ -49,6 +49,12 @@ services:
       - "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

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-ds-flash-htz-chat.yml

@@ -0,0 +1,101 @@
+# deploy/mainnet/docker-compose.supplier-ds-flash-htz-chat.yml — Supplier — MAINNET
+#   (Hetzner Inference / DeepSeek-V4-Flash-0731 — MULTI-TURN CHAT capability `llm.chat.v1`)
+#
+# The chat-session sibling of docker-compose.supplier-ds-flash-htz.yml. Same
+# image, same Hetzner backend (DeepSeek-V4-Flash-0731), but CAPABILITY_KIND=chat-session so
+# it mounts /v1/chat/{start,message,end} (escrow bookends, off-chain streaming
+# turns) instead of the one-off /v1/chat/completions route.
+#
+# A SEPARATE process/advert from the one-off supplier on purpose: the chat
+# session holds the supplier's single-slot lock for the whole conversation
+# (minutes), which would starve one-off traffic if co-hosted.
+#
+# One of 8 Hetzner-backed suppliers (4 models x {one-off, chat-session});
+# fleet runbook: docs/HETZNER_SETUP.md.
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-ds-flash-htz-chat.vector.apexfusion.org → host
+#      (individual DNS-only A record → 91.98.147.172; NO wildcard exists —
+#       letsencrypt HTTP-01 issuance needs the unproxied direct record)
+#   2. supplier/.env.ds-flash-htz-chat written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_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
+#      The advert must be posted with --capability-id llm.chat.v1 and a large
+#      --max-processing-ms (~1800000) so Claim+Submit both land before deliver_by.
+#   3. Networks already exist (created by the canonical supplier compose):
+#        docker network create marketplace-mainnet-net   # one-shot
+#        # traefik-net is created by the Traefik project
+
+name: marketplace-mainnet-supplier-ds-flash-htz-chat
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-ds-flash-htz-chat
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this Hetzner-backed variant. Holds this
+      # supplier's distinct SUPPLIER_PRIV_KEY_HEX (+ derived wallet vars),
+      # ADVERT_REF, the shared Hetzner OPENAI_API_KEY, and the shared mainnet
+      # plumbing duplicated in full (NETWORK_ID, OGMIOS_URL, VECTOR_ZERO_TIME_MS,
+      # ref-script UTxOs, LIVE_CHAIN). No env_file ordering / override magic.
+      - path: ../../supplier/.env.ds-flash-htz-chat
+        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
+      # (Hetzner serves at https://inference.hetzner.com/api/v1/...).
+      OPENAI_BASE_URL: https://inference.hetzner.com/api
+      # Reasoning-style models can be slow; mirror the one-off compose.
+      OPENAI_TIMEOUT_MS: "300000"
+      # Idle end (supplier Submits the transcript): single-slot suppliers must
+      # free up quickly (the gateway transparently reopens a session on the
+      # next turn). Mirrors the kimi-chat fleet.
+      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.ds-flash-htz-chat 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-ds-flash-htz-chat-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-chat-mainnet.rule=Host(`mp-suppliers-ds-flash-htz-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-chat-mainnet.middlewares=mp-suppliers-ds-flash-htz-chat-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-htz-chat-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-htz-chat-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-chat-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-chat-mainnet-tls.rule=Host(`mp-suppliers-ds-flash-htz-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-chat-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-ds-flash-htz-chat-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-htz.yml

@@ -0,0 +1,89 @@
+# deploy/mainnet/docker-compose.supplier-ds-flash-htz.yml — Supplier — MAINNET
+#   (Hetzner Inference / DeepSeek-V4-Flash-0731 — one-off `llm.text.generate.v1`)
+#
+# Hetzner-backed supplier. LLM_BACKEND=openai pointed directly at Hetzner's
+# hosted inference API (OpenAI-compatible at https://inference.hetzner.com/api).
+# No local proxy container — reachable over public HTTPS.
+#
+# One of 8 Hetzner-backed suppliers (4 models x {one-off, chat-session});
+# fleet runbook: docs/HETZNER_SETUP.md.
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-ds-flash-htz.vector.apexfusion.org → host
+#      (individual DNS-only A record → 91.98.147.172; NO wildcard exists —
+#       letsencrypt HTTP-01 issuance needs the unproxied direct record)
+#   2. supplier/.env.ds-flash-htz written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_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 already exist (created by the canonical supplier compose):
+#        docker network create marketplace-mainnet-net   # one-shot
+#        # traefik-net is created by the Traefik project
+
+name: marketplace-mainnet-supplier-ds-flash-htz
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-ds-flash-htz
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this Hetzner-backed variant. Holds this
+      # supplier's distinct SUPPLIER_PRIV_KEY_HEX (+ derived wallet vars),
+      # ADVERT_REF, the shared Hetzner OPENAI_API_KEY, and the shared mainnet
+      # plumbing duplicated in full (NETWORK_ID, OGMIOS_URL, VECTOR_ZERO_TIME_MS,
+      # ref-script UTxOs, LIVE_CHAIN). No env_file ordering / override magic.
+      - path: ../../supplier/.env.ds-flash-htz
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions
+      # (Hetzner serves at https://inference.hetzner.com/api/v1/...).
+      OPENAI_BASE_URL: https://inference.hetzner.com/api
+      # Reasoning-style models can be slow; matches the advert's
+      # max_processing_ms (300000).
+      OPENAI_TIMEOUT_MS: "300000"
+      # OPENAI_MAX_TOKENS intentionally unset: no max_tokens forwarded upstream,
+      # the provider grants context − prompt. Advert cap = model context length.
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so submitTx/awaitTx actually broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from the operator-provided supplier/.env.ds-flash-htz 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-ds-flash-htz-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-mainnet.rule=Host(`mp-suppliers-ds-flash-htz.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-mainnet.middlewares=mp-suppliers-ds-flash-htz-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-htz-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-ds-flash-htz-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-mainnet-tls.rule=Host(`mp-suppliers-ds-flash-htz.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-ds-flash-htz-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-ds-flash-htz-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-glm-chat.yml

@@ -0,0 +1,101 @@
+# deploy/mainnet/docker-compose.supplier-glm-chat.yml — Supplier — MAINNET
+#   (Hetzner Inference / GLM-5.2-NVFP4 — MULTI-TURN CHAT capability `llm.chat.v1`)
+#
+# The chat-session sibling of docker-compose.supplier-glm.yml. Same
+# image, same Hetzner backend (GLM-5.2-NVFP4), but CAPABILITY_KIND=chat-session so
+# it mounts /v1/chat/{start,message,end} (escrow bookends, off-chain streaming
+# turns) instead of the one-off /v1/chat/completions route.
+#
+# A SEPARATE process/advert from the one-off supplier on purpose: the chat
+# session holds the supplier's single-slot lock for the whole conversation
+# (minutes), which would starve one-off traffic if co-hosted.
+#
+# One of 8 Hetzner-backed suppliers (4 models x {one-off, chat-session});
+# fleet runbook: docs/HETZNER_SETUP.md.
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-glm-chat.vector.apexfusion.org → host
+#      (individual DNS-only A record → 91.98.147.172; NO wildcard exists —
+#       letsencrypt HTTP-01 issuance needs the unproxied direct record)
+#   2. supplier/.env.glm-chat written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_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
+#      The advert must be posted with --capability-id llm.chat.v1 and a large
+#      --max-processing-ms (~1800000) so Claim+Submit both land before deliver_by.
+#   3. Networks already exist (created by the canonical supplier compose):
+#        docker network create marketplace-mainnet-net   # one-shot
+#        # traefik-net is created by the Traefik project
+
+name: marketplace-mainnet-supplier-glm-chat
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-glm-chat
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this Hetzner-backed variant. Holds this
+      # supplier's distinct SUPPLIER_PRIV_KEY_HEX (+ derived wallet vars),
+      # ADVERT_REF, the shared Hetzner OPENAI_API_KEY, and the shared mainnet
+      # plumbing duplicated in full (NETWORK_ID, OGMIOS_URL, VECTOR_ZERO_TIME_MS,
+      # ref-script UTxOs, LIVE_CHAIN). No env_file ordering / override magic.
+      - path: ../../supplier/.env.glm-chat
+        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
+      # (Hetzner serves at https://inference.hetzner.com/api/v1/...).
+      OPENAI_BASE_URL: https://inference.hetzner.com/api
+      # Reasoning-style models can be slow; mirror the one-off compose.
+      OPENAI_TIMEOUT_MS: "300000"
+      # Idle end (supplier Submits the transcript): single-slot suppliers must
+      # free up quickly (the gateway transparently reopens a session on the
+      # next turn). Mirrors the kimi-chat fleet.
+      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.glm-chat 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-glm-chat-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-glm-chat-mainnet.rule=Host(`mp-suppliers-glm-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-glm-chat-mainnet.middlewares=mp-suppliers-glm-chat-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-glm-chat-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-glm-chat-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-glm-chat-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-glm-chat-mainnet-tls.rule=Host(`mp-suppliers-glm-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-glm-chat-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-glm-chat-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-glm.yml

@@ -0,0 +1,89 @@
+# deploy/mainnet/docker-compose.supplier-glm.yml — Supplier — MAINNET
+#   (Hetzner Inference / GLM-5.2-NVFP4 — one-off `llm.text.generate.v1`)
+#
+# Hetzner-backed supplier. LLM_BACKEND=openai pointed directly at Hetzner's
+# hosted inference API (OpenAI-compatible at https://inference.hetzner.com/api).
+# No local proxy container — reachable over public HTTPS.
+#
+# One of 8 Hetzner-backed suppliers (4 models x {one-off, chat-session});
+# fleet runbook: docs/HETZNER_SETUP.md.
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-glm.vector.apexfusion.org → host
+#      (individual DNS-only A record → 91.98.147.172; NO wildcard exists —
+#       letsencrypt HTTP-01 issuance needs the unproxied direct record)
+#   2. supplier/.env.glm written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_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 already exist (created by the canonical supplier compose):
+#        docker network create marketplace-mainnet-net   # one-shot
+#        # traefik-net is created by the Traefik project
+
+name: marketplace-mainnet-supplier-glm
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-glm
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this Hetzner-backed variant. Holds this
+      # supplier's distinct SUPPLIER_PRIV_KEY_HEX (+ derived wallet vars),
+      # ADVERT_REF, the shared Hetzner OPENAI_API_KEY, and the shared mainnet
+      # plumbing duplicated in full (NETWORK_ID, OGMIOS_URL, VECTOR_ZERO_TIME_MS,
+      # ref-script UTxOs, LIVE_CHAIN). No env_file ordering / override magic.
+      - path: ../../supplier/.env.glm
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions
+      # (Hetzner serves at https://inference.hetzner.com/api/v1/...).
+      OPENAI_BASE_URL: https://inference.hetzner.com/api
+      # Reasoning-style models can be slow; matches the advert's
+      # max_processing_ms (300000).
+      OPENAI_TIMEOUT_MS: "300000"
+      # OPENAI_MAX_TOKENS intentionally unset: no max_tokens forwarded upstream,
+      # the provider grants context − prompt. Advert cap = model context length.
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so submitTx/awaitTx actually broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from the operator-provided supplier/.env.glm 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-glm-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-glm-mainnet.rule=Host(`mp-suppliers-glm.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-glm-mainnet.middlewares=mp-suppliers-glm-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-glm-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-glm-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-glm-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-glm-mainnet-tls.rule=Host(`mp-suppliers-glm.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-glm-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-glm-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-hf-llama33.yml

@@ -0,0 +1,92 @@
+# deploy/mainnet/docker-compose.supplier-hf-llama33.yml — Supplier — MAINNET
+#   (HuggingFace Inference Providers router / meta-llama/Llama-3.3-70B-Instruct
+#    — one-off `llm.text.generate.v1`)
+#
+# HF-router-backed supplier. LLM_BACKEND=openai pointed at the HuggingFace
+# Inference Providers router (OpenAI-compatible at https://router.huggingface.co).
+# No local proxy container — reachable over public HTTPS. Direct analogue of the
+# OpenRouter/Hetzner-backed suppliers; the router is OpenAI-compatible so this is
+# config only (V1 HF integration).
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-hf-llama33.vector.apexfusion.org → host
+#      (individual DNS-only A record → 91.98.147.172; NO wildcard exists —
+#       letsencrypt HTTP-01 issuance needs the unproxied direct record)
+#   2. supplier/.env.hf-llama33 written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_KEY_HEX, OPENAI_API_KEY (hf_ token), OGMIOS_URL,
+#        NETWORK_ID=1, VECTOR_ZERO_TIME_MS, ESCROW_REF_UTXO, ADVERT_REF_UTXO,
+#        ADVERT_REF, LIVE_CHAIN=1
+#      Advert posted with --capability-id llm.text.generate.v1.
+#      Do NOT set OPENAI_REASONING — the HF router rejects it (HTTP 400) and the
+#      supplier's config guard refuses to boot.
+#   3. Networks already exist (created by the canonical supplier compose):
+#        docker network create marketplace-mainnet-net   # one-shot
+#        # traefik-net is created by the Traefik project
+
+name: marketplace-mainnet-supplier-hf-llama33
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-hf-llama33
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this HF-router-backed variant. Holds this
+      # supplier's distinct SUPPLIER_PRIV_KEY_HEX (+ derived wallet vars),
+      # ADVERT_REF, the HF OPENAI_API_KEY, and the shared mainnet plumbing
+      # duplicated in full (NETWORK_ID, OGMIOS_URL, VECTOR_ZERO_TIME_MS,
+      # ref-script UTxOs, LIVE_CHAIN). No env_file ordering / override magic.
+      - path: ../../supplier/.env.hf-llama33
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions
+      # (HF router serves at https://router.huggingface.co/v1/...).
+      OPENAI_BASE_URL: https://router.huggingface.co
+      # First-token latency across hosted providers can exceed local backends;
+      # kept ≤ the advert's max_processing_ms (300000) per fleet footgun #4.
+      OPENAI_TIMEOUT_MS: "300000"
+      # OPENAI_MAX_TOKENS intentionally unset: no max_tokens forwarded upstream,
+      # the provider grants context − prompt. Advert cap = model context length.
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so submitTx/awaitTx actually broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from the operator-provided supplier/.env.hf-llama33 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-hf-llama33-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-hf-llama33-mainnet.rule=Host(`mp-suppliers-hf-llama33.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-hf-llama33-mainnet.middlewares=mp-suppliers-hf-llama33-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-hf-llama33-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-hf-llama33-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-hf-llama33-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-hf-llama33-mainnet-tls.rule=Host(`mp-suppliers-hf-llama33.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-hf-llama33-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-hf-llama33-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-2.yml

@@ -51,6 +51,12 @@ services:
       - "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

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

@@ -51,6 +51,12 @@ services:
       - "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

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

@@ -51,6 +51,12 @@ services:
       - "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

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

@@ -62,6 +62,12 @@ services:
       - "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

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

@@ -62,6 +62,12 @@ services:
       - "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

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

@@ -74,6 +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-code-chat.yml

@@ -0,0 +1,101 @@
+# deploy/mainnet/docker-compose.supplier-kimi-code-chat.yml — Supplier — MAINNET
+#   (Hetzner Inference / Kimi-K2.7-Code — MULTI-TURN CHAT capability `llm.chat.v1`)
+#
+# The chat-session sibling of docker-compose.supplier-kimi-code.yml. Same
+# image, same Hetzner backend (Kimi-K2.7-Code), but CAPABILITY_KIND=chat-session so
+# it mounts /v1/chat/{start,message,end} (escrow bookends, off-chain streaming
+# turns) instead of the one-off /v1/chat/completions route.
+#
+# A SEPARATE process/advert from the one-off supplier on purpose: the chat
+# session holds the supplier's single-slot lock for the whole conversation
+# (minutes), which would starve one-off traffic if co-hosted.
+#
+# One of 8 Hetzner-backed suppliers (4 models x {one-off, chat-session});
+# fleet runbook: docs/HETZNER_SETUP.md.
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-kimi-code-chat.vector.apexfusion.org → host
+#      (individual DNS-only A record → 91.98.147.172; NO wildcard exists —
+#       letsencrypt HTTP-01 issuance needs the unproxied direct record)
+#   2. supplier/.env.kimi-code-chat written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_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
+#      The advert must be posted with --capability-id llm.chat.v1 and a large
+#      --max-processing-ms (~1800000) so Claim+Submit both land before deliver_by.
+#   3. Networks already exist (created by the canonical supplier compose):
+#        docker network create marketplace-mainnet-net   # one-shot
+#        # traefik-net is created by the Traefik project
+
+name: marketplace-mainnet-supplier-kimi-code-chat
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-kimi-code-chat
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this Hetzner-backed variant. Holds this
+      # supplier's distinct SUPPLIER_PRIV_KEY_HEX (+ derived wallet vars),
+      # ADVERT_REF, the shared Hetzner OPENAI_API_KEY, and the shared mainnet
+      # plumbing duplicated in full (NETWORK_ID, OGMIOS_URL, VECTOR_ZERO_TIME_MS,
+      # ref-script UTxOs, LIVE_CHAIN). No env_file ordering / override magic.
+      - path: ../../supplier/.env.kimi-code-chat
+        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
+      # (Hetzner serves at https://inference.hetzner.com/api/v1/...).
+      OPENAI_BASE_URL: https://inference.hetzner.com/api
+      # Reasoning-style models can be slow; mirror the one-off compose.
+      OPENAI_TIMEOUT_MS: "300000"
+      # Idle end (supplier Submits the transcript): single-slot suppliers must
+      # free up quickly (the gateway transparently reopens a session on the
+      # next turn). Mirrors the kimi-chat fleet.
+      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-code-chat 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-code-chat-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-kimi-code-chat-mainnet.rule=Host(`mp-suppliers-kimi-code-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-code-chat-mainnet.middlewares=mp-suppliers-kimi-code-chat-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-code-chat-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-code-chat-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-kimi-code-chat-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-kimi-code-chat-mainnet-tls.rule=Host(`mp-suppliers-kimi-code-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-code-chat-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-kimi-code-chat-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-code.yml

@@ -0,0 +1,89 @@
+# deploy/mainnet/docker-compose.supplier-kimi-code.yml — Supplier — MAINNET
+#   (Hetzner Inference / Kimi-K2.7-Code — one-off `llm.text.generate.v1`)
+#
+# Hetzner-backed supplier. LLM_BACKEND=openai pointed directly at Hetzner's
+# hosted inference API (OpenAI-compatible at https://inference.hetzner.com/api).
+# No local proxy container — reachable over public HTTPS.
+#
+# One of 8 Hetzner-backed suppliers (4 models x {one-off, chat-session});
+# fleet runbook: docs/HETZNER_SETUP.md.
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-kimi-code.vector.apexfusion.org → host
+#      (individual DNS-only A record → 91.98.147.172; NO wildcard exists —
+#       letsencrypt HTTP-01 issuance needs the unproxied direct record)
+#   2. supplier/.env.kimi-code written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_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 already exist (created by the canonical supplier compose):
+#        docker network create marketplace-mainnet-net   # one-shot
+#        # traefik-net is created by the Traefik project
+
+name: marketplace-mainnet-supplier-kimi-code
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-kimi-code
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this Hetzner-backed variant. Holds this
+      # supplier's distinct SUPPLIER_PRIV_KEY_HEX (+ derived wallet vars),
+      # ADVERT_REF, the shared Hetzner OPENAI_API_KEY, and the shared mainnet
+      # plumbing duplicated in full (NETWORK_ID, OGMIOS_URL, VECTOR_ZERO_TIME_MS,
+      # ref-script UTxOs, LIVE_CHAIN). No env_file ordering / override magic.
+      - path: ../../supplier/.env.kimi-code
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions
+      # (Hetzner serves at https://inference.hetzner.com/api/v1/...).
+      OPENAI_BASE_URL: https://inference.hetzner.com/api
+      # Reasoning-style models can be slow; matches the advert's
+      # max_processing_ms (300000).
+      OPENAI_TIMEOUT_MS: "300000"
+      # OPENAI_MAX_TOKENS intentionally unset: no max_tokens forwarded upstream,
+      # the provider grants context − prompt. Advert cap = model context length.
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so submitTx/awaitTx actually broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from the operator-provided supplier/.env.kimi-code 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-code-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-kimi-code-mainnet.rule=Host(`mp-suppliers-kimi-code.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-code-mainnet.middlewares=mp-suppliers-kimi-code-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-code-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-kimi-code-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-kimi-code-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-kimi-code-mainnet-tls.rule=Host(`mp-suppliers-kimi-code.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-kimi-code-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-kimi-code-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.yml

@@ -75,6 +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-qwen35b-chat.yml

@@ -0,0 +1,101 @@
+# deploy/mainnet/docker-compose.supplier-qwen35b-chat.yml — Supplier — MAINNET
+#   (Hetzner Inference / Qwen/Qwen3.6-35B-A3B-FP8 — MULTI-TURN CHAT capability `llm.chat.v1`)
+#
+# The chat-session sibling of docker-compose.supplier-qwen35b.yml. Same
+# image, same Hetzner backend (Qwen/Qwen3.6-35B-A3B-FP8), but CAPABILITY_KIND=chat-session so
+# it mounts /v1/chat/{start,message,end} (escrow bookends, off-chain streaming
+# turns) instead of the one-off /v1/chat/completions route.
+#
+# A SEPARATE process/advert from the one-off supplier on purpose: the chat
+# session holds the supplier's single-slot lock for the whole conversation
+# (minutes), which would starve one-off traffic if co-hosted.
+#
+# One of 8 Hetzner-backed suppliers (4 models x {one-off, chat-session});
+# fleet runbook: docs/HETZNER_SETUP.md.
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-qwen35b-chat.vector.apexfusion.org → host
+#      (individual DNS-only A record → 91.98.147.172; NO wildcard exists —
+#       letsencrypt HTTP-01 issuance needs the unproxied direct record)
+#   2. supplier/.env.qwen35b-chat written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_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
+#      The advert must be posted with --capability-id llm.chat.v1 and a large
+#      --max-processing-ms (~1800000) so Claim+Submit both land before deliver_by.
+#   3. Networks already exist (created by the canonical supplier compose):
+#        docker network create marketplace-mainnet-net   # one-shot
+#        # traefik-net is created by the Traefik project
+
+name: marketplace-mainnet-supplier-qwen35b-chat
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-qwen35b-chat
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this Hetzner-backed variant. Holds this
+      # supplier's distinct SUPPLIER_PRIV_KEY_HEX (+ derived wallet vars),
+      # ADVERT_REF, the shared Hetzner OPENAI_API_KEY, and the shared mainnet
+      # plumbing duplicated in full (NETWORK_ID, OGMIOS_URL, VECTOR_ZERO_TIME_MS,
+      # ref-script UTxOs, LIVE_CHAIN). No env_file ordering / override magic.
+      - path: ../../supplier/.env.qwen35b-chat
+        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
+      # (Hetzner serves at https://inference.hetzner.com/api/v1/...).
+      OPENAI_BASE_URL: https://inference.hetzner.com/api
+      # Reasoning-style models can be slow; mirror the one-off compose.
+      OPENAI_TIMEOUT_MS: "300000"
+      # Idle end (supplier Submits the transcript): single-slot suppliers must
+      # free up quickly (the gateway transparently reopens a session on the
+      # next turn). Mirrors the kimi-chat fleet.
+      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.qwen35b-chat 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-qwen35b-chat-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-qwen35b-chat-mainnet.rule=Host(`mp-suppliers-qwen35b-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-qwen35b-chat-mainnet.middlewares=mp-suppliers-qwen35b-chat-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-qwen35b-chat-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-qwen35b-chat-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-qwen35b-chat-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-qwen35b-chat-mainnet-tls.rule=Host(`mp-suppliers-qwen35b-chat.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-qwen35b-chat-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-qwen35b-chat-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-qwen35b.yml

@@ -0,0 +1,89 @@
+# deploy/mainnet/docker-compose.supplier-qwen35b.yml — Supplier — MAINNET
+#   (Hetzner Inference / Qwen/Qwen3.6-35B-A3B-FP8 — one-off `llm.text.generate.v1`)
+#
+# Hetzner-backed supplier. LLM_BACKEND=openai pointed directly at Hetzner's
+# hosted inference API (OpenAI-compatible at https://inference.hetzner.com/api).
+# No local proxy container — reachable over public HTTPS.
+#
+# One of 8 Hetzner-backed suppliers (4 models x {one-off, chat-session});
+# fleet runbook: docs/HETZNER_SETUP.md.
+#
+# Pre-flight:
+#   1. DNS: mp-suppliers-qwen35b.vector.apexfusion.org → host
+#      (individual DNS-only A record → 91.98.147.172; NO wildcard exists —
+#       letsencrypt HTTP-01 issuance needs the unproxied direct record)
+#   2. supplier/.env.qwen35b written on host (chmod 600) with at minimum:
+#        SUPPLIER_PRIV_KEY_HEX, SUPPLIER_ADDRESS, SUPPLIER_PKH,
+#        SUPPLIER_PUB_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 already exist (created by the canonical supplier compose):
+#        docker network create marketplace-mainnet-net   # one-shot
+#        # traefik-net is created by the Traefik project
+
+name: marketplace-mainnet-supplier-qwen35b
+
+services:
+  supplier:
+    build:
+      context: ../..
+      dockerfile: supplier/Dockerfile
+    container_name: marketplace-mainnet-supplier-qwen35b
+    restart: unless-stopped
+    env_file:
+      # Self-contained env for this Hetzner-backed variant. Holds this
+      # supplier's distinct SUPPLIER_PRIV_KEY_HEX (+ derived wallet vars),
+      # ADVERT_REF, the shared Hetzner OPENAI_API_KEY, and the shared mainnet
+      # plumbing duplicated in full (NETWORK_ID, OGMIOS_URL, VECTOR_ZERO_TIME_MS,
+      # ref-script UTxOs, LIVE_CHAIN). No env_file ordering / override magic.
+      - path: ../../supplier/.env.qwen35b
+        required: false
+    environment:
+      LLM_BACKEND: openai
+      # No /v1 suffix — supplier/src/openai.ts appends /v1/chat/completions
+      # (Hetzner serves at https://inference.hetzner.com/api/v1/...).
+      OPENAI_BASE_URL: https://inference.hetzner.com/api
+      # Reasoning-style models can be slow; matches the advert's
+      # max_processing_ms (300000).
+      OPENAI_TIMEOUT_MS: "300000"
+      # OPENAI_MAX_TOKENS intentionally unset: no max_tokens forwarded upstream,
+      # the provider grants context − prompt. Advert cap = model context length.
+      PORT: "8080"
+      # Enable LiveOgmiosProvider so submitTx/awaitTx actually broadcast.
+      LIVE_CHAIN: "1"
+      # NETWORK_ID and OGMIOS_URL intentionally not pinned here — they MUST
+      # come from the operator-provided supplier/.env.qwen35b 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-qwen35b-mainnet.entrypoints=web"
+      - "traefik.http.routers.mp-suppliers-qwen35b-mainnet.rule=Host(`mp-suppliers-qwen35b.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-qwen35b-mainnet.middlewares=mp-suppliers-qwen35b-mainnet-https"
+      - "traefik.http.middlewares.mp-suppliers-qwen35b-mainnet-https.redirectscheme.scheme=https"
+      - "traefik.http.middlewares.mp-suppliers-qwen35b-mainnet-https.redirectscheme.permanent=true"
+      - "traefik.http.routers.mp-suppliers-qwen35b-mainnet-tls.entrypoints=websecure"
+      - "traefik.http.routers.mp-suppliers-qwen35b-mainnet-tls.rule=Host(`mp-suppliers-qwen35b.vector.apexfusion.org`)"
+      - "traefik.http.routers.mp-suppliers-qwen35b-mainnet-tls.tls.certresolver=letsencrypt"
+      - "traefik.http.services.mp-suppliers-qwen35b-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-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/HETZNER_SETUP.md

@@ -0,0 +1,133 @@
+# Hetzner Inference Supplier Fleet — Setup Runbook
+
+> **Status:** shipped. Configures 8 mainnet supplier nodes that use **Hetzner
+> Inference** (`https://inference.hetzner.com`) as their **compute backend** for
+> bonded inference work commissioned on Vector. Config only — no code changes.
+
+---
+
+## 1. What this is
+
+Hetzner Inference is an OpenAI-compatible hosted API. Because the supplier's
+upstream client (`supplier/src/openai.ts`) already speaks plain
+`POST /v1/chat/completions` with a Bearer token, pointing it at Hetzner is
+**pure configuration**: `LLM_BACKEND=openai` + `OPENAI_BASE_URL` + the API key.
+This is the same pattern as the DeepSeek-direct and OpenRouter suppliers, and
+the direct sibling of `docs/HUGGINGFACE_ROUTER_SETUP.md`.
+
+One supplier per (model × capability) — 4 models, both the one-off
+`llm.text.generate.v1` and the multi-turn `llm.chat.v1` capability:
+
+| Supplier name | Model (on-chain + API param, verbatim) | Capability | Advert max_output_tokens |
+|---|---|---|---|
+| `kimi-code` | `Kimi-K2.7-Code` | `llm.text.generate.v1` | 262144 |
+| `kimi-code-chat` | `Kimi-K2.7-Code` | `llm.chat.v1` | 262144 |
+| `glm` | `GLM-5.2-NVFP4` | `llm.text.generate.v1` | 512000 |
+| `glm-chat` | `GLM-5.2-NVFP4` | `llm.chat.v1` | 512000 |
+| `ds-flash-htz` | `DeepSeek-V4-Flash-0731` | `llm.text.generate.v1` | 512000 |
+| `ds-flash-htz-chat` | `DeepSeek-V4-Flash-0731` | `llm.chat.v1` | 512000 |
+| `qwen35b` | `Qwen/Qwen3.6-35B-A3B-FP8` | `llm.text.generate.v1` | 262144 |
+| `qwen35b-chat` | `Qwen/Qwen3.6-35B-A3B-FP8` | `llm.chat.v1` | 262144 |
+
+Shared advert parameters: price **200000 lovelace (0.2 AP3X)** flat per job,
+bonds **1000000 lovelace (1 AP3X)** both sides, `max_processing_ms` **300000**
+for one-off / **1800000** for chat (the session spans the whole conversation).
+`max_output_tokens` is set to the model's context length: no `max_tokens` is
+forwarded upstream (`OPENAI_MAX_TOKENS` unset), so the advert value only gates
+what buyers may request.
+
+Names follow the brand-by-model convention; `ds-flash-htz` carries a `-htz`
+suffix only because `deepseek-flash-*` was already taken by the OpenRouter
+fleet serving `deepseek/deepseek-v4-flash`.
+
+## 2. Prerequisites
+
+1. A Hetzner account with Inference access and an API key (usage-based billing
+   to you as the node operator — the flat 0.2 AP3X job price is a business
+   choice, not cost-derived). One key is shared by all 8 suppliers; Hetzner
+   rate limits apply to the aggregate.
+2. Eight funded supplier wallets (50 AP3X each recommended; bonds are 1 AP3X
+   per in-flight job and returned on completion).
+3. The usual mainnet supplier deploy prerequisites (`deploy/README.md`).
+4. **DNS**: one A record per supplier — `mp-suppliers-<name>.vector.apexfusion.org`
+   → `91.98.147.172`, **DNS-only/unproxied**. There is NO wildcard for
+   `*.vector.apexfusion.org`; and because traefik issues certs on-box via
+   Let's Encrypt HTTP-01 (port 80), a proxied/CDN record would break issuance.
+   Certs are issued automatically within ~a minute of the record propagating —
+   no restarts needed.
+
+## 3. Verify the backend before spending on-chain
+
+The four model ids come from `GET https://inference.hetzner.com/api/v1/models`
+and are sent **verbatim** as the upstream `model` param (the runtime uses
+`advert.model`). Before posting adverts, smoke-test each model:
+
+```bash
+curl -sS -X POST https://inference.hetzner.com/api/v1/chat/completions \
+  -H "Authorization: Bearer $HETZNER_KEY" -H "Content-Type: application/json" \
+  -d '{"model":"GLM-5.2-NVFP4","messages":[{"role":"user","content":"ping"}]}' \
+  | jq '.choices[0].message.content'
+```
+
+`choices[0].message.content` must be a non-empty string with **no `max_tokens`
+sent** (matches runtime config) — the supplier throws `openai_malformed`
+otherwise and forfeits its 1 AP3X bond per failed job. For chat suppliers also
+verify a `tools`/`tool_choice` round-trip. All four models passed both checks
+on 2026-08-11.
+
+## 4. Per-supplier bring-up (repeat ×8)
+
+```bash
+# 1. wallet — prints privateKeyHex, publicKeyHex, pubKeyHash, address
+pnpm --filter @marketplace/supplier tx:gen-keypair --network 1
+# fund the printed address (50 AP3X recommended), then:
+
+# 2. env file on the mainnet host
+cp supplier/.env.hetzner.example /root/agents-marketplace/supplier/.env.<name>
+chmod 600 /root/agents-marketplace/supplier/.env.<name>
+# fill in: ALL FOUR wallet vars (PRIV_KEY_HEX, ADDRESS, PKH, PUB_KEY_HEX —
+# they are NOT derived at boot; missing derived vars ⇒ 403 wrong_supplier),
+# OPENAI_API_KEY, and the ADVERT_* values from the table above.
+
+# 3. advert (one-off shown; chat: --capability-id llm.chat.v1 --max-processing-ms 1800000)
+pnpm --filter @marketplace/supplier tx:post-advert \
+  --capability-id llm.text.generate.v1 \
+  --model 'GLM-5.2-NVFP4' \
+  --max-output-tokens 512000 \
+  --max-processing-ms 300000 \
+  --price-lovelace 200000 \
+  --endpoint-url https://mp-suppliers-<name>.vector.apexfusion.org
+# paste the printed "<txHash>#0" into ADVERT_REF in the env file
+
+# 4. first start is always manual (CD skips compose projects with no running
+#    containers); subsequent rollouts ride the mainnet CD pipeline
+docker compose -f deploy/mainnet/docker-compose.supplier-<name>.yml up -d
+
+# 5. register the wallet for balance alerts
+#    add {name, address} to wallet-monitor/wallets.json on the host
+```
+
+## 5. Verification
+
+- `curl https://mp-suppliers-<name>.vector.apexfusion.org/healthz` → `{"ok":true}`;
+  `/capability` shows the right model + pkh.
+- Indexer: buyer `GET /v1/indexer/suppliers` lists the new entries with
+  `advert_status: Active`, `status: free`.
+- Gateway `GET /openai/v1/models` includes the four Hetzner model ids.
+- Watch `docker logs -f marketplace-mainnet-supplier-<name>` through the first
+  jobs for `openai_malformed` (reasoning-model content-shape risk).
+
+## 6. Footguns
+
+- `OPENAI_BASE_URL=https://inference.hetzner.com/api` — **no `/v1`**; the
+  client appends `/v1/chat/completions`.
+- Do **not** set `OPENAI_REASONING` — `reasoning:{enabled:false}` is an
+  OpenRouter-only param.
+- The advert `--model` must match the Hetzner id exactly, including the
+  `Qwen/` prefix on `Qwen/Qwen3.6-35B-A3B-FP8` (slash is fine on-chain;
+  precedent `moonshotai/kimi-k2.6`).
+- `Kimi-K2.7-Code` and `DeepSeek-V4-Flash-0731` auto-join the buyer's PDF pool
+  (case-insensitive substring allowlist `["kimi","deepseek","gpt"]` in
+  `buyer/src/pdf/caps.ts`); GLM and Qwen stay out. Deliberate — adjust
+  `PDF_MODEL_ALLOWLIST`/`PDF_MODEL_DENYLIST` in the buyer env to change it.
+- `OPENAI_TIMEOUT_MS` must stay ≤ the advert's `max_processing_ms`.

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
+1. You type a **prompt** and set a **price** in the web UI
-2. The app posts an **escrow transaction** on-chain, locking your payment + bond
+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.
+- The seed phrase controls your wallet. Back it up securely.
-- Escrow protects you: the supplier cannot take your funds without delivering a result
+- 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** |
+| 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-ADA collateral, working}` after settle / on a tick |
+| 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`.
@@ -121,7 +121,7 @@ All on-chain steps run inside the per-key `mutex.run(...)`.
 
 ## 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`.
@@ -145,7 +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 seed phrase controls your wallet. 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 node only responds to buyers who have already locked value on-chain, so you cannot be scammed out of compute.
 
 ## Troubleshooting
 

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/src/account/routes.ts

@@ -19,6 +19,26 @@ import { badRequest, forbidden } from "../openai/errors.js";
 
 const ACTIVE_ESCROW_STATES = new Set(["Open", "Claimed", "Submitted"]);
 
+/** Bound on per-request escrow liveness lookups; rows past the cap count as
+ * locked unverified (a key with this many active escrows has bigger problems). */
+const LIVENESS_CHECK_CAP = 32;
+
+/** True when the escrow UTxO behind an indexer row is still unspent. The
+ * indexer never updates rows on terminal spends (Accept/Reclaim/Release
+ * consume the escrow without a new escrow output), so "active" rows can be
+ * long spent. Unparseable refs and failed lookups count as live so errors
+ * inflate the locked figure rather than hide funds. */
+async function escrowUtxoIsLive(deps: GatewayDeps, utxoRef: string): Promise<boolean> {
+  const m = /^([0-9a-fA-F]{64})#(\d+)$/.exec(utxoRef);
+  if (!m) return true;
+  try {
+    const utxo = await deps.chain.queryUtxo({ txHash: m[1], index: Number(m[2]) });
+    return utxo !== null;
+  } catch {
+    return true;
+  }
+}
+
 function ap3x(lovelace: bigint): string {
   return (Number(lovelace) / 1e6).toFixed(2);
 }
@@ -74,8 +94,15 @@ export function makeAccountHandler(deps: GatewayDeps) {
       if (r.ok) {
         const rows = (await r.json()) as Array<Record<string, unknown>>;
         if (Array.isArray(rows)) {
-          for (const row of rows) {
+          const active = rows.filter((row) => ACTIVE_ESCROW_STATES.has(String(row.state)));
-            if (!ACTIVE_ESCROW_STATES.has(String(row.state))) continue;
+          const liveness = await Promise.all(
+            active.map((row, i) =>
+              i < LIVENESS_CHECK_CAP ? escrowUtxoIsLive(deps, String(row.utxo_ref ?? "")) : Promise.resolve(true),
+            ),
+          );
+          for (let i = 0; i < active.length; i++) {
+            if (!liveness[i]) continue;
+            const row = active[i];
             lockedInEscrow +=
               BigInt(String(row.payment_lovelace ?? "0")) +
               BigInt(String(row.buyer_bond_lovelace ?? "0")) +

gateway/src/config.ts

@@ -53,6 +53,9 @@ export interface GatewayConfig {
   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 {
@@ -138,6 +141,7 @@ export function loadConfig(env: Record<string, string | undefined>): GatewayConf
     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(/\/+$/, ""))

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/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
+ * supplier_bond, raised to the Submitted-state min-ada floor so long
- * that also need a pure-ADA collateral UTxO ≥ 5 ADA (assertCollateralCandidate).
+ * capability ids (e.g. the 33-byte OCR one) don't get bumped mid-flight —
- * So required = price + buyer_bond + supplier_bond + collateral + fee reserve,
+ * see escrowLockFloor (packages/shared/.../minAdaFloor.ts) and
- * AND the wallet must currently hold a pure-ADA UTxO ≥ collateral. We check both
+ * postOcrEscrow.ts:205, the SAME floor a POST actually locks. Accept/Reclaim
- * up front and return a clean 402 instead of failing mid-flight.
+ * 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 =
+  const economicTotal = cost.priceLovelace + cost.buyerBondLovelace + cost.supplierBondLovelace;
-    cost.priceLovelace +
+  const escrowFloor = escrowLockFloor(representativeEscrowDatum(cost), economicTotal);
-    cost.buyerBondLovelace +
+  const required = escrowFloor + COLLATERAL_MIN_LOVELACE + FEE_RESERVE_LOVELACE;
-    cost.supplierBondLovelace +
-    COLLATERAL_MIN_LOVELACE +
-    FEE_RESERVE_LOVELACE;
   return {
     ok: available >= required && collateralOk,
     availableLovelace: available,

gateway/src/openai/chatCompletions.ts

@@ -108,7 +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/sessions.ts

@@ -93,7 +93,7 @@ export async function openSessionCore(
   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(),

gateway/src/routing/selectSupplier.ts

@@ -53,6 +53,9 @@ 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;
@@ -82,7 +85,7 @@ 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" && !opts.ignoreStatusFor?.has(raw.supplier_pkh)) {
       continue;

gateway/src/sdk/registry.ts

@@ -15,19 +15,88 @@ import { open as unseal } from "../crypto/seal.js";
 import { deriveWalletKey } from "../wallet.js";
 import type { ApiKeyRow } from "../db/store.js";
 
-/** Tiny promise-chain mutex: run(fn) executes fns strictly one at a time. */
+/** Default per-run deadline. Chain reads and tx submits settle in seconds;
+ * anything past this is the 2026-08-12 freeze mode (a fetch that never
+ * settles), not a slow success. */
+export const DEFAULT_MUTEX_TIMEOUT_MS = 180_000;
+
+/** Default queue bound per mutex. A healthy key drains its queue in seconds;
+ * a queue this deep means the head is stuck and callers should fail fast. */
+export const DEFAULT_MUTEX_MAX_QUEUE = 64;
+
+/** A run() exceeded its deadline and was abandoned so queued work can proceed. */
+export class MutexTimeoutError extends Error {
+  constructor(label: string, timeoutMs: number) {
+    super(`mutex run '${label}' exceeded ${timeoutMs}ms and was abandoned; queued work proceeds`);
+    this.name = "MutexTimeoutError";
+  }
+}
+
+/** run() was refused because the queue is at its bound (head likely stuck). */
+export class MutexQueueFullError extends Error {
+  constructor(label: string, depth: number) {
+    super(`mutex queue full (${depth} pending) — refusing '${label}'; the head of the queue is likely stuck`);
+    this.name = "MutexQueueFullError";
+  }
+}
+
+function runWithDeadline<T>(fn: () => Promise<T>, timeoutMs: number, label: string): Promise<T> {
+  return new Promise<T>((resolvePromise, rejectPromise) => {
+    const timer = setTimeout(() => rejectPromise(new MutexTimeoutError(label, timeoutMs)), timeoutMs);
+    // Node-only nicety: don't hold the process open for a pending deadline.
+    timer.unref?.();
+    fn().then(
+      (v) => {
+        clearTimeout(timer);
+        resolvePromise(v);
+      },
+      (e) => {
+        clearTimeout(timer);
+        rejectPromise(e);
+      },
+    );
+  });
+}
+
+/**
+ * Tiny promise-chain mutex: run(fn) executes fns strictly one at a time —
+ * now with a per-run deadline and a queue bound (2026-08-12 freeze fix).
+ *
+ * On deadline the stuck fn is NOT cancelled (promises aren't); it is
+ * abandoned: its run() rejects with MutexTimeoutError and the chain advances.
+ * A zombie that settles later can at worst lose a UTxO-contention race with
+ * the run that replaced it — strictly better than the alternative, where one
+ * stuck chain call convoys every subsequent request on the key forever.
+ */
 export class Mutex {
   private tail: Promise<unknown> = Promise.resolve();
+  private depth = 0;
 
-  run<T>(fn: () => Promise<T>): Promise<T> {
+  constructor(private readonly opts: { timeoutMs?: number; maxQueue?: number } = {}) {}
-    const result = this.tail.then(fn, fn);
+
-    // Keep the chain alive even if fn rejects; swallow here so the next
+  run<T>(fn: () => Promise<T>, label = "op"): Promise<T> {
-    // run() isn't poisoned by a prior rejection.
+    const timeoutMs = this.opts.timeoutMs ?? DEFAULT_MUTEX_TIMEOUT_MS;
-    this.tail = result.then(
+    const maxQueue = this.opts.maxQueue ?? DEFAULT_MUTEX_MAX_QUEUE;
+
+    if (this.depth >= maxQueue) {
+      return Promise.reject(new MutexQueueFullError(label, this.depth));
+    }
+    this.depth++;
+
+    // Keep the chain alive even if fn rejects or times out; swallow here so
+    // the next run() isn't poisoned by a prior rejection.
+    const result = this.tail.then(
+      () => runWithDeadline(fn, timeoutMs, label),
+      () => runWithDeadline(fn, timeoutMs, label),
+    );
+    const settled = result.finally(() => {
+      this.depth--;
+    });
+    this.tail = settled.then(
       () => undefined,
       () => undefined,
     );
-    return result;
+    return settled;
   }
 }
 

gateway/src/server.ts

@@ -14,6 +14,7 @@ import { ipRateLimit, keyRateLimit, demoIpRateLimit } from "./middleware/rateLim
 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,
@@ -56,7 +57,11 @@ export function createApp(deps: GatewayDeps): Express {
   app.set("trust proxy", true);
   // CORS before body-parsing so preflight OPTIONS short-circuits cheaply.
   app.use(corsMiddleware(deps.config.corsOrigins));
-  app.use(express.json({ limit: "1mb" }));
+  // /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 });
@@ -80,6 +85,12 @@ export function createApp(deps: GatewayDeps): Express {
   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/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
+ * {≥5 AP3X collateral UTxO, working UTxO} via the shared consolidate builder, so
- * the NEXT request's Accept/Reclaim script-spend always finds a pure-ADA
+ * 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/http.test.ts

@@ -46,6 +46,7 @@ function makeDeps(
     chatSettleMode: overrides?.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"]);

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

@@ -45,6 +45,7 @@ function makeSweepDeps(opts: { chatSettleMode: "full" | "ticket"; escrows: Escro
     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"];

gateway/test/unit.test.ts

@@ -208,8 +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 };
+    const cost = {
-    // required = 2 + 1 + 1 + 5 (collateral) + 2 (fee) = 11 ADA
+      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/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/escrow/release.ts

@@ -17,6 +17,8 @@ import type { WalletKey, BuildResult } from "../types.js";
 import { TxConstructionError } from "../types.js";
 import { encodeTxBody, sha256Hex } from "../internal/testTxBody.js";
 import { ACCEPT_WINDOW_MS, mockSlotToWallclockMs } from "../internal/constants.js";
+import { detectCborBackend } from "../internal/cborBackend.js";
+import type { LiveOgmiosProvider } from "../../chain/LiveOgmiosProvider.js";
 
 export interface ReleaseParams {
   chain: ChainProvider;
@@ -65,8 +67,10 @@ export async function buildReleaseTx(
   }
 
   // 4. Tip must be at or past submitted_at + ACCEPT_WINDOW.
+  // Live backend uses real POSIX ms; mock backend keeps slot*1000.
   const tipSlot = await chain.tip();
-  const tipMs = mockSlotToWallclockMs(tipSlot);
+  const isLive = detectCborBackend(chain) === "live";
+  const tipMs = isLive ? Date.now() : mockSlotToWallclockMs(tipSlot);
   const threshold = datum.submitted_at + ACCEPT_WINDOW_MS;
   if (tipMs < threshold) {
     throw new TxConstructionError(
@@ -75,6 +79,21 @@ export async function buildReleaseTx(
     );
   }
 
+  // Live path: real Cardano CBOR via lucid-evolution.
+  if (isLive) {
+    const liveCborPath = "../internal/liveCbor.js";
+    const { buildLiveTxForRelease } = await import(/* @vite-ignore */ liveCborPath);
+    return buildLiveTxForRelease({
+      chain: chain as LiveOgmiosProvider,
+      supplierKey,
+      escrowRef,
+      escrowUtxo: utxo,
+      datum,
+      tipMs,
+    });
+  }
+
+  // Mock path: synthetic testTxBody (preserves existing test behaviour).
   const supplierDue =
     datum.payment_lovelace + datum.supplier_bond_lovelace + datum.buyer_bond_lovelace;
 

packages/shared/src/tx/index.ts

@@ -7,6 +7,7 @@ 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";
@@ -37,6 +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

@@ -10,8 +10,8 @@
  * for vkh enterprise). Input UTxOs may carry testnet-prefixed addresses — lucid
  * does not validate address network for inputs (only for outputs).
  *
- * Deferred builders (postAdvert, updateAdvert, retireAdvert, reclaim, release)
+ * Deferred builders: only updateAdvert remains on the synthetic testTxBody
- * are NOT touched in M1-F-4. Those continue using the synthetic testTxBody path.
+ * path. postAdvert, retireAdvert, reclaim, and release all have live builders.
  *
  * Catherine M1-F-4-green.
  */
@@ -61,6 +61,7 @@ const REDEEMER_CLAIM = 121;
 const REDEEMER_SUBMIT = 122;
 const REDEEMER_ACCEPT = 123;
 const REDEEMER_RECLAIM = 124;
+const REDEEMER_RELEASE = 125;
 
 // AdvertRedeemer Constr indices per lib/marketplace/types.ak:
 //   PostAdvert    → Constr0 (tag 121, creation event — never seen by spend)
@@ -121,6 +122,15 @@ export interface LiveReclaimParams {
   tipMs: number;
 }
 
+export interface LiveReleaseParams {
+  chain: LiveOgmiosProvider;
+  supplierKey: WalletKey;
+  escrowRef: OutputReference;
+  escrowUtxo: Utxo;
+  datum: EscrowDatum;
+  tipMs: number;
+}
+
 export interface LivePostAdvertParams {
   chain: LiveOgmiosProvider;
   walletKey: WalletKey;
@@ -778,6 +788,77 @@ export async function buildLiveTxForReclaim(params: LiveReclaimParams): Promise<
   return { txCborHex, expectedTxHash };
 }
 
+// ─── Release ─────────────────────────────────────────────────────────
+
+export async function buildLiveTxForRelease(params: LiveReleaseParams): Promise<BuildResult> {
+  const provider = buildLucidProvider(params.chain);
+  const { lucid } = await createLucidContext(provider, params.supplierKey, {
+    networkId: 1,
+    systemStartUnix: 0,
+    slotLengthMs: 1000,
+  }, { usePresetProtocolParameters: true });
+
+  const { script, address: escrowAddress } = loadEscrowScript();
+  const spendCtx = await resolveSpendCtx(lucid, "ESCROW_REF_UTXO", script);
+  const lucidInput = chainUtxoToLucidScriptInput(params.escrowUtxo, escrowAddress);
+  const releaseRedeemerHex = encodeNullaryConstrHex(REDEEMER_RELEASE);
+
+  if (params.datum.submitted_at === null) {
+    throw new TxConstructionError(
+      "submitted_at missing",
+      "Submitted-state escrow must carry a submitted_at timestamp",
+    );
+  }
+
+  // Release pays everything to the supplier: payment + supplier_bond + buyer_bond.
+  const supplierDue =
+    params.datum.payment_lovelace +
+    params.datum.supplier_bond_lovelace +
+    params.datum.buyer_bond_lovelace;
+  const supplierLovelace = supplierDue < MIN_UTXO_LOVELACE ? MIN_UTXO_LOVELACE : supplierDue;
+  const supplierAddress = pkhToEnterpriseAddress(params.datum.supplier_pkh, MAINNET_ID);
+
+  const realWalletUtxos = await lucid.wallet().getUtxos();
+  assertCollateralCandidate(realWalletUtxos);
+  const presetWalletInputs = realWalletUtxos;
+
+  // escrow.ak:handle_release checks `lower_bound >= submitted_at + ACCEPT_WINDOW`.
+  // Clamp validFrom to that threshold: the Reclaim-style tip-60s back-pad alone
+  // would violate the validator whenever the tip is less than 60s past the window.
+  const threshold = params.datum.submitted_at + ACCEPT_WINDOW_MS;
+  const validFromMs = Math.max(threshold, params.tipMs - 60_000);
+
+  let signed;
+  try {
+    const txBuilder = (
+      spendCtx.kind === "ref"
+        ? lucid.newTx().readFrom([spendCtx.refUtxo])
+        : lucid.newTx().attach.SpendingValidator(spendCtx.script)
+    )
+      .collectFrom([lucidInput], releaseRedeemerHex)
+      .pay.ToAddress(supplierAddress, { lovelace: supplierLovelace })
+      .addSignerKey(params.supplierKey.pubKeyHash)
+      .validFrom(validFromMs)
+      .validTo(params.tipMs + 60_000)
+      .setMinFee(500_000n);
+
+    // Same UPLC eval delegation as Accept/Reclaim — lucid's CML evaluator
+    // falsely rejects valid Plutus V3 spends; route to Ogmios.
+    const completed = await txBuilder.complete({
+      presetWalletInputs,
+      localUPLCEval: false,
+    });
+    signed = await completed.sign.withWallet().complete();
+  } catch (err) {
+    rethrowAsTxError(err, "release build failed");
+  }
+
+  const txCborHex = signed.toCBOR();
+  const expectedTxHash = signed.toHash();
+  await params.chain.submitTx(txCborHex);
+  return { txCborHex, expectedTxHash };
+}
+
 // ─── PostAdvert ──────────────────────────────────────────────────────
 
 export async function buildLiveTxForAdvert(

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/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
+ * whether the wallet needs reshaping or is already usable for script spends.
- * {≥collateral, working}). The same downstream tx-builder handles
+ *
- * consolidate and split — only the "reason" tag differs for logging.
+ * 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.hetzner.example

@@ -0,0 +1,105 @@
+# supplier/.env.hetzner.example
+#
+# Preset for a supplier that uses Hetzner Inference
+# (https://inference.hetzner.com) as its COMPUTE BACKEND to fulfil bonded
+# inference work on Vector. Hetzner exposes an OpenAI-compatible API, so this
+# reuses the existing LLM_BACKEND=openai path with NO code changes — config only.
+#
+# One template for all 8 Hetzner-backed suppliers (4 models x {one-off, chat}):
+#   kimi-code[-chat]     → Kimi-K2.7-Code            (ctx 262144)
+#   glm[-chat]           → GLM-5.2-NVFP4             (ctx 512000)
+#   ds-flash-htz[-chat]  → DeepSeek-V4-Flash-0731    (ctx 512000)
+#   qwen35b[-chat]       → Qwen/Qwen3.6-35B-A3B-FP8  (ctx 262144)
+#
+# Copy to supplier/.env.<name> on the host (chmod 600) and fill in real values.
+# .env.* is gitignored — NEVER commit a populated copy.
+#
+# Runbook: docs/HETZNER_SETUP.md
+#
+# Lines marked SECRET must never be shared, logged, or committed.
+
+# ─── Wallet identity ──────────────────────────────────────────────────────
+# Generate all four values in one shot (from repo root):
+#   pnpm --filter @marketplace/supplier tx:gen-keypair --network 1
+#
+# ALL FOUR are required. The supplier does NOT re-derive address/pkh/pubkey
+# from the private key at boot — leaving them unset silently yields an empty
+# wallet identity and every escrow is rejected 403 wrong_supplier (this took
+# down the TTS supplier once; don't repeat it).
+
+# 64-char hex Ed25519 private key seed for this supplier's wallet. SECRET.
+SUPPLIER_PRIV_KEY_HEX=PASTE_64_HEX_CHARS_HERE
+SUPPLIER_ADDRESS=PASTE_ADDR1_HERE
+SUPPLIER_PKH=PASTE_28_BYTE_HEX_HERE
+SUPPLIER_PUB_KEY_HEX=PASTE_32_BYTE_HEX_HERE
+
+# ─── Chain plumbing (mainnet) ─────────────────────────────────────────────
+OGMIOS_URL=https://ogmios.vector.mainnet.apexfusion.org
+NETWORK_ID=1
+VECTOR_ZERO_TIME_MS=1756485600000
+
+# CIP-33 reference scripts (shared mainnet deployment).
+ESCROW_REF_UTXO=c8d84c6d67ec67a1efe5e9c6c06d53020e05d1bb96d1c55ecb1eb7d5010c4d54#0
+ADVERT_REF_UTXO=c8d84c6d67ec67a1efe5e9c6c06d53020e05d1bb96d1c55ecb1eb7d5010c4d54#1
+
+# OutputReference of this supplier's on-chain advertisement UTxO.
+# Post it first (see tx:post-advert below), then paste the printed "<txHash>#0".
+ADVERT_REF=PASTE_TXHASH_HERE#0
+
+# LIVE_CHAIN: only the literal "1" opts in to real Claim/Submit txs.
+LIVE_CHAIN=1
+
+# ─── Backend: Hetzner Inference (OpenAI-compatible) ───────────────────────
+LLM_BACKEND=openai
+
+# Hetzner base URL. The supplier appends /v1/chat/completions itself, so do
+# NOT include the /api/v1 → only up to /api.
+OPENAI_BASE_URL=https://inference.hetzner.com/api
+
+# Hetzner Inference API key. Shared across all 8 Hetzner-backed suppliers —
+# rate limits apply to the aggregate. SECRET.
+OPENAI_API_KEY=PASTE_HETZNER_KEY_HERE
+
+# All four models are reasoning-style; generations can be slow.
+OPENAI_TIMEOUT_MS=300000
+
+# Do NOT set OPENAI_REASONING — `reasoning:{enabled:false}` is an
+# OpenRouter-only param.
+# OPENAI_REASONING=
+
+# OPENAI_MAX_TOKENS intentionally unset: no max_tokens is forwarded upstream,
+# Hetzner grants context − prompt. The advert's max_output_tokens (= model
+# context length) only caps what buyers may request.
+# OPENAI_MAX_TOKENS=
+
+PORT=8080
+
+# ─── tx:post-advert — run once before first deploy ────────────────────────
+# One-off (llm.text.generate.v1) suppliers, from repo root:
+#   pnpm --filter @marketplace/supplier tx:post-advert \
+#     --capability-id llm.text.generate.v1 \
+#     --model <EXACT Hetzner model id, see table above> \
+#     --max-output-tokens <262144 or 512000 per table> \
+#     --max-processing-ms 300000 \
+#     --price-lovelace 200000 \
+#     --endpoint-url https://mp-suppliers-<name>.vector.apexfusion.org
+#
+# Chat (llm.chat.v1) suppliers: same, but
+#     --capability-id llm.chat.v1 --max-processing-ms 1800000
+# (the session spans the whole conversation; Claim+Submit must land before
+# deliver_by — mirrors the kimi-chat fleet).
+#
+# CRITICAL: --model goes on-chain AND is sent verbatim to Hetzner as the API
+# `model` param (jobRunner uses advert.model). It must be the EXACT id from
+# GET https://inference.hetzner.com/api/v1/models — including the "Qwen/"
+# prefix on Qwen/Qwen3.6-35B-A3B-FP8.
+ADVERT_CAPABILITY_ID=llm.text.generate.v1
+ADVERT_MODEL=GLM-5.2-NVFP4
+ADVERT_MAX_OUTPUT_TOKENS=512000
+ADVERT_MAX_PROCESSING_MS=300000
+ADVERT_PRICE_LOVELACE=200000
+ADVERT_SUPPLIER_BOND_LOVELACE=1000000
+ADVERT_BUYER_BOND_LOVELACE=1000000
+ADVERT_ENDPOINT_URL=https://mp-suppliers-glm.vector.apexfusion.org
+ADVERT_DETAIL_URI=
+ADVERT_DETAIL_HASH=

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)"
+}

supplier/src/config.ts

@@ -36,6 +36,15 @@
  *   OLLAMA_TIMEOUT_MS      — ollama request timeout in ms (default 120_000)
  *   OPENAI_TIMEOUT_MS      — openai request timeout in ms (default 180_000 — GPT
  *                            first-token latency is longer than local Ollama)
+ *   OPENAI_SESSION_PASSTHROUGH
+ *                          — "1" switches chat-session turns to the stateful-
+ *                            upstream contract (OpenClaw): escrow ref as the
+ *                            OpenAI `user` field + delta-only messages per turn.
+ *                            Strict "1" only. Default off. Requires
+ *                            CAPABILITY_KIND="chat-session".
+ *   OPENAI_MODEL_OVERRIDE  — fixed upstream model id sent instead of the
+ *                            advert's buyer-facing model (OpenClaw accepts only
+ *                            "openclaw" / "openclaw/<agentId>"). Default unset.
  *   PIPER_TIMEOUT_MS       — piper request timeout in ms (default 120_000)
  *   LIVE_CHAIN             — "1" opts in to live-chain submit/await (default off).
  *                            Any other value (including "true", "yes", "TRUE")
@@ -72,8 +81,26 @@ export interface AdvertRef {
  *   "chat"         — one-off `llm.text.generate.v1` (single prompt → single completion)
  *   "tts"          — `audio.synthesize.piper.v1`
  *   "chat-session" — multi-turn `llm.chat.v1` (escrow bookends, off-chain turns,
- *                    streaming). Requires LLM_BACKEND="openai". */
+ *                    streaming). Requires LLM_BACKEND="openai".
-export type CapabilityKind = "chat" | "tts" | "chat-session";
+ *   "ocr"          — model-scoped `ocr.page.extract.<model-slug>.v1` (one page
+ *                    image → one extraction). Upstream is an OpenAI-compatible
+ *                    VISION endpoint (OPENAI_BASE_URL), e.g. vLLM serving
+ *                    datalab-to/chandra-ocr-2. */
+export type CapabilityKind = "chat" | "tts" | "chat-session" | "ocr";
+
+/** Served-checkpoint probe cadence for the ocr capability.
+ *   "per_job" — probe the upstream before every Claim (default when
+ *               REVISION_PROBE_URL is set, and always the default for the
+ *               datalab upstream, whose health route is derived).
+ *   "off"     — no probing. */
+export type RevisionProbeMode = "per_job" | "off";
+
+/** Upstream flavor for the ocr capability.
+ *   "openai-vision" — OpenAI-compatible vision endpoint (vLLM self-hosting;
+ *                     supports the revision-pin attestation probe).
+ *   "datalab"       — Datalab hosted document API (submit + poll; liveness
+ *                     probe only — hosted supply has no checkpoint to pin). */
+export type OcrUpstream = "openai-vision" | "datalab";
 
 /** LLM backend selected for the chat capability. */
 export type LlmBackend = "ollama" | "openai";
@@ -111,6 +138,28 @@ export interface SupplierConfig {
    * the model context or the provider 400s (e.g. kimi context = 262144).
    */
   openaiMaxTokens: number;
+  /**
+   * When true, chat-session turns speak the STATEFUL-upstream contract
+   * (e.g. an OpenClaw gateway): each turn sends the escrow ref as the OpenAI
+   * `user` field (OpenClaw keys one persistent agent session per `user`, so
+   * browser/tool state survives across turns) and sends ONLY the current
+   * turn's delta messages — the upstream session carries the history, and
+   * re-sending the full transcript would duplicate it into the upstream
+   * context every turn. The supplier still accumulates the full transcript
+   * locally for the receipt. Env: OPENAI_SESSION_PASSTHROUGH="1". Default
+   * off (stateless upstreams get the full transcript each turn). Requires
+   * capabilityKind="chat-session".
+   */
+  openaiSessionPassthrough: boolean;
+  /**
+   * When non-empty, sent as the upstream `model` instead of the advert's
+   * model id. For upstreams that reject arbitrary model strings — OpenClaw's
+   * chat-completions endpoint 400s on anything but "openclaw" or
+   * "openclaw/<agentId>" — while the advert carries the buyer-facing id the
+   * gateway routes on. Env: OPENAI_MODEL_OVERRIDE. Default "" (advert model
+   * sent verbatim).
+   */
+  openaiModelOverride: string;
   /** Empty string when capabilityKind="chat". */
   piperUrl: string;
   advertRef: AdvertRef;
@@ -135,6 +184,37 @@ export interface SupplierConfig {
   maxChatSessions: number;
   /** Chain settlement mode for chat sessions. Default "full". */
   chatSettleMode: ChatSettleMode;
+  /**
+   * Upstream flavor for the ocr capability. Env: OCR_UPSTREAM
+   * ("openai-vision" default | "datalab").
+   */
+  ocrUpstream: OcrUpstream;
+  /** Datalab API key (X-API-Key). Required when ocrUpstream="datalab". */
+  datalabApiKey: string;
+  /** Datalab API origin. Env: DATALAB_BASE_URL, default https://www.datalab.to */
+  datalabBaseUrl: string;
+  /** Datalab quality mode: fast | balanced | accurate. Env: DATALAB_MODE. */
+  datalabMode: string;
+  /** Whole-call budget for submit+poll, ms. Env: DATALAB_TIMEOUT_MS. */
+  datalabTimeoutMs: number;
+  /**
+   * Overrides the built-in per-format OCR instruction prompt when non-empty
+   * (capabilityKind="ocr" + openai-vision upstream only). Env: OCR_PROMPT_OVERRIDE.
+   */
+  ocrPromptOverride: string;
+  /**
+   * Expected served-checkpoint SHA (40-hex) for the ocr upstream. When set
+   * together with revisionProbeUrl, a probe mismatch refuses the Claim.
+   * Env: HF_MODEL_REVISION. "" = no pin.
+   */
+  hfModelRevision: string;
+  /**
+   * Serving-runtime info route reporting the loaded checkpoint (TGI-style
+   * /info or vLLM equivalent). Env: REVISION_PROBE_URL. "" = probing off.
+   */
+  revisionProbeUrl: string;
+  /** Probe cadence; see RevisionProbeMode. Env: REVISION_PROBE_MODE. */
+  revisionProbeMode: RevisionProbeMode;
   /**
    * When true, the supplier boots a LiveOgmiosProvider (real submitTx/awaitTx).
    * Default false → ReadOnlyOgmiosProvider (safe; no chain writes).
@@ -186,8 +266,13 @@ export function loadConfig(env: Record<string, string | undefined>): SupplierCon
   // CAPABILITY_KIND drives whether OLLAMA_URL or PIPER_URL is required.
   // Default to "chat" so existing chat suppliers keep booting unchanged.
   const capKindStr = env.CAPABILITY_KIND ?? "chat";
-  if (capKindStr !== "chat" && capKindStr !== "tts" && capKindStr !== "chat-session") {
+  if (
-    throw new Error('loadConfig: CAPABILITY_KIND must be "chat", "tts", or "chat-session"');
+    capKindStr !== "chat" &&
+    capKindStr !== "tts" &&
+    capKindStr !== "chat-session" &&
+    capKindStr !== "ocr"
+  ) {
+    throw new Error('loadConfig: CAPABILITY_KIND must be "chat", "tts", "chat-session", or "ocr"');
   }
   const capabilityKind: CapabilityKind = capKindStr;
 
@@ -216,6 +301,19 @@ export function loadConfig(env: Record<string, string | undefined>): SupplierCon
     openaiBaseUrl = requireField(env, "OPENAI_BASE_URL");
     ollamaUrl = env.OLLAMA_URL ?? "";
     piperUrl = env.PIPER_URL ?? "";
+  } else if (capabilityKind === "ocr") {
+    // Upstream is either an OpenAI-compatible vision endpoint (self-hosted
+    // vLLM; OPENAI_BASE_URL required) or the Datalab hosted API
+    // (DATALAB_API_KEY required, validated below). Ollama and piper are
+    // unused regardless of LLM_BACKEND.
+    const upstream = env.OCR_UPSTREAM ?? "openai-vision";
+    if (upstream === "openai-vision") {
+      openaiBaseUrl = requireField(env, "OPENAI_BASE_URL");
+    } else {
+      openaiBaseUrl = env.OPENAI_BASE_URL ?? "";
+    }
+    ollamaUrl = env.OLLAMA_URL ?? "";
+    piperUrl = env.PIPER_URL ?? "";
   } else if (capabilityKind === "chat") {
     if (llmBackend === "openai") {
       openaiBaseUrl = requireField(env, "OPENAI_BASE_URL");
@@ -312,6 +410,22 @@ export function loadConfig(env: Record<string, string | undefined>): SupplierCon
     openaiMaxTokens = Number(maxTokStr);
   }
 
+  // OPENAI_SESSION_PASSTHROUGH — "1" switches chat-session turns to the
+  // stateful-upstream contract (OpenClaw): escrow ref as the OpenAI `user`
+  // field + delta-only messages (the upstream session carries history).
+  // Strict "1" only, mirroring LIVE_CHAIN.
+  const openaiSessionPassthrough = env.OPENAI_SESSION_PASSTHROUGH === "1";
+  if (openaiSessionPassthrough && capabilityKind !== "chat-session") {
+    throw new Error(
+      'loadConfig: OPENAI_SESSION_PASSTHROUGH requires CAPABILITY_KIND="chat-session"',
+    );
+  }
+
+  // OPENAI_MODEL_OVERRIDE — fixed upstream model id sent instead of the
+  // advert's (buyer-facing) model. Needed when the upstream validates the
+  // model field (OpenClaw accepts only "openclaw" / "openclaw/<agentId>").
+  const openaiModelOverride = env.OPENAI_MODEL_OVERRIDE ?? "";
+
   // CHAT_IDLE_TIMEOUT_MS — idle auto-end window for chat sessions (default 5 min).
   const chatIdleStr = env.CHAT_IDLE_TIMEOUT_MS;
   let chatIdleTimeoutMs = 300_000;
@@ -348,6 +462,52 @@ export function loadConfig(env: Record<string, string | undefined>): SupplierCon
   }
   const chatSettleMode: ChatSettleMode = settleModeStr;
 
+  // OCR-capability extras. All optional; harmless when set on other kinds
+  // (ignored by the chat/tts/chat-session paths).
+  const ocrUpstreamStr = env.OCR_UPSTREAM ?? "openai-vision";
+  if (ocrUpstreamStr !== "openai-vision" && ocrUpstreamStr !== "datalab") {
+    throw new Error('loadConfig: OCR_UPSTREAM must be "openai-vision" or "datalab"');
+  }
+  const ocrUpstream: OcrUpstream = ocrUpstreamStr;
+
+  const datalabApiKey = env.DATALAB_API_KEY ?? "";
+  if (capabilityKind === "ocr" && ocrUpstream === "datalab" && datalabApiKey === "") {
+    throw new Error('loadConfig: OCR_UPSTREAM="datalab" requires DATALAB_API_KEY');
+  }
+  const datalabBaseUrl = env.DATALAB_BASE_URL ?? "https://www.datalab.to";
+  const datalabMode = env.DATALAB_MODE ?? "accurate";
+  if (datalabMode !== "fast" && datalabMode !== "balanced" && datalabMode !== "accurate") {
+    throw new Error('loadConfig: DATALAB_MODE must be "fast", "balanced", or "accurate"');
+  }
+  const datalabTimeoutStr = env.DATALAB_TIMEOUT_MS;
+  let datalabTimeoutMs = 120_000;
+  if (datalabTimeoutStr !== undefined && datalabTimeoutStr !== "") {
+    if (!POS_INT_RE.test(datalabTimeoutStr)) {
+      throw new Error("loadConfig: DATALAB_TIMEOUT_MS must be a positive integer");
+    }
+    datalabTimeoutMs = Number(datalabTimeoutStr);
+  }
+
+  const ocrPromptOverride = env.OCR_PROMPT_OVERRIDE ?? "";
+  const hfModelRevision = env.HF_MODEL_REVISION ?? "";
+  if (hfModelRevision !== "" && !/^[0-9a-fA-F]{40}$/.test(hfModelRevision)) {
+    throw new Error("loadConfig: HF_MODEL_REVISION must be a 40-hex commit SHA when set");
+  }
+  const revisionProbeUrl = env.REVISION_PROBE_URL ?? "";
+  // Datalab upstream: the health route is derived from DATALAB_BASE_URL, so
+  // the liveness probe defaults ON without a REVISION_PROBE_URL. The
+  // openai-vision upstream keeps the explicit-URL requirement.
+  const probeDefault =
+    ocrUpstream === "datalab" || revisionProbeUrl !== "" ? "per_job" : "off";
+  const probeModeStr = env.REVISION_PROBE_MODE ?? probeDefault;
+  if (probeModeStr !== "per_job" && probeModeStr !== "off") {
+    throw new Error('loadConfig: REVISION_PROBE_MODE must be "per_job" or "off"');
+  }
+  if (probeModeStr === "per_job" && ocrUpstream === "openai-vision" && revisionProbeUrl === "" && capabilityKind === "ocr") {
+    throw new Error('loadConfig: REVISION_PROBE_MODE="per_job" requires REVISION_PROBE_URL for the openai-vision upstream');
+  }
+  const revisionProbeMode: RevisionProbeMode = probeModeStr;
+
   const piperTimeoutStr = env.PIPER_TIMEOUT_MS;
   let piperTimeoutMs = 120_000;
   if (piperTimeoutStr !== undefined && piperTimeoutStr !== "") {
@@ -380,6 +540,8 @@ export function loadConfig(env: Record<string, string | undefined>): SupplierCon
     openaiApiKey,
     openaiReasoningDisabled,
     openaiMaxTokens,
+    openaiSessionPassthrough,
+    openaiModelOverride,
     piperUrl,
     advertRef,
     networkId,
@@ -392,6 +554,15 @@ export function loadConfig(env: Record<string, string | undefined>): SupplierCon
     chatIdleTimeoutMs,
     maxChatSessions,
     chatSettleMode,
+    ocrUpstream,
+    datalabApiKey,
+    datalabBaseUrl,
+    datalabMode,
+    datalabTimeoutMs,
+    ocrPromptOverride,
+    hfModelRevision,
+    revisionProbeUrl,
+    revisionProbeMode,
     liveChain,
     walletHealthIntervalMs,
   };

supplier/src/datalabOcr.ts

@@ -0,0 +1,249 @@
+/**
+ * supplier/src/datalabOcr.ts — HTTP client for the Datalab hosted document
+ * API (https://www.datalab.to), used when OCR_UPSTREAM="datalab".
+ *
+ * Contract (verified against documentation.datalab.to, 2026-08-06):
+ *   POST {base}/api/v1/convert            multipart/form-data, X-API-Key
+ *     fields: file (binary), output_format (markdown|html|json), mode
+ *     → { success, error, request_id, request_check_url }
+ *   GET  {base}/api/v1/convert/{id}       X-API-Key
+ *     → { status: "complete" | ..., markdown|html|json, page_count, error }
+ *
+ * callDatalabOcr() submits ONE page image and polls until complete, then
+ * returns the SAME result shape as ocrVision.callOcrVision so runOcrJob's
+ * receipt construction is shared. Token counts are pinned to 0 — the
+ * Datalab API has no token semantics (mirrors the Piper/TTS convention);
+ * wallclock_ms brackets submit→complete locally.
+ *
+ * JSON output: when output_format="json" the result field can be an object;
+ * it is serialized with JSON.stringify to keep `content` a string. The
+ * receipt commits sha256(canonical({content, output_format})) over exactly
+ * the string the buyer receives, so determinism holds regardless.
+ *
+ * Error reasons (mirror the taxonomy of the other adapters):
+ *   "datalab_failure"   — non-2xx, network error, success:false, job error
+ *   "datalab_timeout"   — submit+poll exceeded timeoutMs
+ *   "datalab_malformed" — response JSON missing the expected fields
+ *
+ * probeDatalabHealth() — pre-Claim liveness gate for the route (the
+ * revision-probe slot repurposed for hosted supply): GETs {base}/api/v1/health
+ * with the API key; a non-OK answer refuses the Claim so escrows are not
+ * stranded behind a dead upstream.
+ */
+
+import type { OcrRequest } from "@marketplace/shared/tx";
+import type { OcrVisionResult } from "./ocrVision.js";
+
+export interface CallDatalabOcrParams {
+  /** API origin, no trailing slash. Default deployment: https://www.datalab.to */
+  baseUrl: string;
+  apiKey: string;
+  request: OcrRequest;
+  /** Whole-call budget: submit + polling until complete. */
+  timeoutMs: number;
+  /** Datalab quality mode: fast | balanced | accurate. */
+  mode: string;
+  /** Poll cadence; overridable in tests. Default 2000. */
+  pollIntervalMs?: number;
+}
+
+export type DatalabErrorReason = "datalab_failure" | "datalab_timeout" | "datalab_malformed";
+
+export class DatalabError extends Error {
+  public readonly reason: DatalabErrorReason;
+  constructor(reason: DatalabErrorReason, message?: string) {
+    super(message ?? reason);
+    this.name = "DatalabError";
+    this.reason = reason;
+  }
+}
+
+function isAbortError(err: unknown): boolean {
+  if (err instanceof Error) {
+    if (err.name === "AbortError") return true;
+    if ((err as { code?: string }).code === "ABORT_ERR") return true;
+  }
+  return false;
+}
+
+const EXT_BY_MIME: Record<string, string> = {
+  "image/png": "png",
+  "image/jpeg": "jpg",
+  "image/webp": "webp",
+};
+
+function sleep(ms: number, signal: AbortSignal): Promise<void> {
+  return new Promise((resolve, reject) => {
+    const t = setTimeout(resolve, ms);
+    signal.addEventListener("abort", () => {
+      clearTimeout(t);
+      const err = new Error("aborted");
+      err.name = "AbortError";
+      reject(err);
+    }, { once: true });
+  });
+}
+
+export async function callDatalabOcr(params: CallDatalabOcrParams): Promise<OcrVisionResult> {
+  const { baseUrl, apiKey, request, timeoutMs, mode } = params;
+  const pollIntervalMs = params.pollIntervalMs ?? 2_000;
+  const base = baseUrl.replace(/\/+$/, "");
+
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(), timeoutMs);
+  const startedAt = Date.now();
+
+  try {
+    // ── 1. Submit ─────────────────────────────────────────────────────
+    const form = new FormData();
+    const bytes = Buffer.from(request.image_b64, "base64");
+    const ext = EXT_BY_MIME[request.mime] ?? "png";
+    form.append("file", new Blob([bytes], { type: request.mime }), `page.${ext}`);
+    form.append("output_format", request.output_format);
+    form.append("mode", mode);
+
+    let submitRes: Response;
+    try {
+      submitRes = await fetch(`${base}/api/v1/convert`, {
+        method: "POST",
+        headers: { "X-API-Key": apiKey },
+        body: form,
+        signal: controller.signal,
+      });
+    } catch (err) {
+      if (isAbortError(err)) {
+        throw new DatalabError("datalab_timeout", `Datalab submit exceeded ${timeoutMs}ms`);
+      }
+      throw new DatalabError("datalab_failure",
+        `Datalab submit failed: ${(err as Error)?.message ?? String(err)}`);
+    }
+    if (!submitRes.ok) {
+      let detail = "";
+      try { detail = await submitRes.text(); } catch { /* body unavailable */ }
+      throw new DatalabError("datalab_failure",
+        `Datalab submit returned HTTP ${submitRes.status}${detail ? `: ${detail.slice(0, 200)}` : ""}`);
+    }
+
+    let submitBody: Record<string, unknown>;
+    try {
+      submitBody = await submitRes.json() as Record<string, unknown>;
+    } catch {
+      throw new DatalabError("datalab_malformed", "Datalab submit response was not JSON");
+    }
+    if (submitBody.success !== true) {
+      throw new DatalabError("datalab_failure",
+        `Datalab submit success=false: ${String(submitBody.error ?? "no error detail")}`);
+    }
+    const requestId = submitBody.request_id;
+    if (typeof requestId !== "string" || requestId.length === 0) {
+      throw new DatalabError("datalab_malformed", "Datalab submit response missing request_id");
+    }
+
+    // ── 2. Poll until complete ────────────────────────────────────────
+    const checkUrl = typeof submitBody.request_check_url === "string" && submitBody.request_check_url.length > 0
+      ? submitBody.request_check_url
+      : `${base}/api/v1/convert/${requestId}`;
+
+    for (;;) {
+      let pollRes: Response;
+      try {
+        pollRes = await fetch(checkUrl, {
+          headers: { "X-API-Key": apiKey },
+          signal: controller.signal,
+        });
+      } catch (err) {
+        if (isAbortError(err)) {
+          throw new DatalabError("datalab_timeout", `Datalab poll exceeded ${timeoutMs}ms budget`);
+        }
+        throw new DatalabError("datalab_failure",
+          `Datalab poll failed: ${(err as Error)?.message ?? String(err)}`);
+      }
+      if (!pollRes.ok) {
+        throw new DatalabError("datalab_failure", `Datalab poll returned HTTP ${pollRes.status}`);
+      }
+      let poll: Record<string, unknown>;
+      try {
+        poll = await pollRes.json() as Record<string, unknown>;
+      } catch {
+        throw new DatalabError("datalab_malformed", "Datalab poll response was not JSON");
+      }
+
+      const status = typeof poll.status === "string" ? poll.status.toLowerCase() : "";
+      if (status === "complete") {
+        if (poll.error && typeof poll.error === "string" && poll.error.length > 0) {
+          throw new DatalabError("datalab_failure", `Datalab conversion error: ${poll.error}`);
+        }
+        const raw = poll[request.output_format];
+        let content: string;
+        if (typeof raw === "string") {
+          content = raw;
+        } else if (raw !== null && raw !== undefined && typeof raw === "object") {
+          content = JSON.stringify(raw);
+        } else {
+          throw new DatalabError("datalab_malformed",
+            `Datalab result missing "${request.output_format}" field`);
+        }
+        if (content.length === 0) {
+          throw new DatalabError("datalab_malformed", "Datalab result content is empty");
+        }
+        return {
+          content,
+          prompt_tokens: 0,
+          completion_tokens: 0,
+          wallclock_ms: Date.now() - startedAt,
+        };
+      }
+      if (status === "failed" || status === "error") {
+        throw new DatalabError("datalab_failure",
+          `Datalab conversion status=${status}: ${String(poll.error ?? "no detail")}`);
+      }
+
+      try {
+        await sleep(pollIntervalMs, controller.signal);
+      } catch (err) {
+        if (isAbortError(err)) {
+          throw new DatalabError("datalab_timeout", `Datalab poll exceeded ${timeoutMs}ms budget`);
+        }
+        throw err;
+      }
+    }
+  } finally {
+    clearTimeout(timer);
+  }
+}
+
+// ─── Liveness probe (pre-Claim gate for hosted supply) ─────────────────────
+
+export interface ProbeDatalabParams {
+  baseUrl: string;
+  apiKey: string;
+  timeoutMs: number;
+}
+
+export interface ProbeDatalabResult {
+  ok: boolean;
+  detail: string;
+}
+
+export async function probeDatalabHealth(params: ProbeDatalabParams): Promise<ProbeDatalabResult> {
+  const base = params.baseUrl.replace(/\/+$/, "");
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(), params.timeoutMs);
+  try {
+    const res = await fetch(`${base}/api/v1/health`, {
+      headers: { "X-API-Key": params.apiKey },
+      signal: controller.signal,
+    });
+    if (!res.ok) {
+      return { ok: false, detail: `health HTTP ${res.status}` };
+    }
+    return { ok: true, detail: "health ok" };
+  } catch (err) {
+    const msg = isAbortError(err)
+      ? `health probe exceeded ${params.timeoutMs}ms`
+      : `health probe failed: ${(err as Error)?.message ?? String(err)}`;
+    return { ok: false, detail: msg };
+  } finally {
+    clearTimeout(timer);
+  }
+}

supplier/src/jobRunner.ts

@@ -38,11 +38,15 @@ import type { SupplierConfig } from "./config.js";
 import * as ollama from "./ollama.js";
 import * as openai from "./openai.js";
 import * as piper from "./piper.js";
+import * as ocrVision from "./ocrVision.js";
+import * as datalabOcr from "./datalabOcr.js";
+import type { OcrRequest } from "@marketplace/shared/tx";
 import type {
   JobStore,
   JobResponsePayload,
   ChatJobResponsePayload,
   TtsJobResponsePayload,
+  OcrJobResponsePayload,
 } from "./jobs.js";
 
 export interface RunChatJobDeps {
@@ -88,7 +92,7 @@ export async function runChatJob(params: RunChatJobParams): Promise<void> {
       if (deps.config.llmBackend === "openai") {
         inference = await openai.callOpenAi({
           baseUrl: deps.config.openaiBaseUrl,
-          model: advert.model,
+          model: deps.config.openaiModelOverride || advert.model,
           messages: requestBody.messages,
           timeoutMs: deps.config.openaiTimeoutMs,
           apiKey: deps.config.openaiApiKey,
@@ -364,3 +368,173 @@ export async function runTtsJob(params: RunTtsJobParams): Promise<void> {
     }
   }
 }
+
+// ─── runOcrJob ────────────────────────────────────────────────────────────
+//
+// Mirror of runTtsJob for model-scoped `ocr.page.extract.<model-slug>.v1`
+// capabilities. Same 7-step shape (setRunning → upstream → receipt →
+// buildSubmit → awaitTx → complete → finally release). Differences from the
+// TTS path:
+//
+//   - Upstream is `ocrVision.callOcrVision` against an OpenAI-compatible
+//     VISION endpoint (vLLM serving e.g. datalab-to/chandra-ocr-2); the
+//     one-page image rides as a data-URL content part.
+//   - response_hash = sha256(canonical({content, output_format})) — commits
+//     the extraction text AND the shape it was requested in.
+//   - prompt_tokens / completion_tokens come from upstream usage (vision
+//     tokens are real tokens, unlike Piper).
+
+export interface RunOcrJobParams {
+  deps: RunChatJobDeps;
+  jobId: string;
+  escrowRef: string;
+  claimedRef: OutputReference;
+  advert: AdvertDatum;
+  escrowDatum: EscrowDatum;
+  /** The OCR request envelope. The PROMPT_HASH committed by the buyer in
+   * the escrow datum is sha256(canonicalize(this same object)), validated
+   * by the route before Claim. */
+  requestBody: OcrRequest;
+}
+
+export async function runOcrJob(params: RunOcrJobParams): Promise<void> {
+  const { deps, jobId, escrowRef, claimedRef, advert, escrowDatum, requestBody } = params;
+
+  deps.jobs.setRunning(jobId);
+
+  try {
+    // ── 2. Call the OCR upstream (openai-vision or datalab) ───────────
+    let inference: ocrVision.OcrVisionResult;
+    try {
+      if (deps.config.ocrUpstream === "datalab") {
+        inference = await datalabOcr.callDatalabOcr({
+          baseUrl: deps.config.datalabBaseUrl,
+          apiKey: deps.config.datalabApiKey,
+          request: requestBody,
+          timeoutMs: deps.config.datalabTimeoutMs,
+          mode: deps.config.datalabMode,
+        });
+      } else {
+        const configuredMax = deps.config.openaiMaxTokens;
+        const maxTokens =
+          configuredMax > 0
+            ? Math.min(configuredMax, advert.max_output_tokens)
+            : advert.max_output_tokens;
+        inference = await ocrVision.callOcrVision({
+          baseUrl: deps.config.openaiBaseUrl,
+          model: advert.model,
+          request: requestBody,
+          timeoutMs: deps.config.openaiTimeoutMs,
+          apiKey: deps.config.openaiApiKey,
+          maxTokens,
+          promptOverride: deps.config.ocrPromptOverride,
+        });
+      }
+    } catch (err) {
+      const rawReason =
+        (err as ocrVision.OcrVisionError | datalabOcr.DatalabError)?.reason ??
+        (deps.config.ocrUpstream === "datalab" ? "datalab_failure" : "ocr_failure");
+      const message = err instanceof Error ? err.message : String(err);
+      // *_timeout collapses to the failure reason for jobs.fail, matching
+      // the chat/tts runner failure-code table (502 only).
+      const collapsedReason =
+        rawReason === "ocr_timeout" ? "ocr_failure"
+          : rawReason === "datalab_timeout" ? "datalab_failure"
+            : rawReason;
+      console.warn(
+        `[job_failed] jobId=${jobId} reason=${rawReason} httpStatus=502 msg=${message}`,
+      );
+      deps.jobs.fail(jobId, {
+        httpStatus: 502,
+        reason: collapsedReason,
+        message,
+      });
+      return;
+    }
+
+    // ── 3. Build + sign receipt ───────────────────────────────────────
+    const responseHash = sha256Hex(canonicalize({
+      content: inference.content,
+      output_format: requestBody.output_format,
+    }));
+
+    const receipt = buildReceipt({
+      prompt_hash: escrowDatum.prompt_hash,
+      response_hash: responseHash,
+      model: advert.model,
+      prompt_tokens: inference.prompt_tokens,
+      completion_tokens: inference.completion_tokens,
+      wallclock_ms: inference.wallclock_ms,
+      supplier_pkh: deps.supplierKey.pubKeyHash,
+      escrow_ref: escrowRef,
+    });
+    const signed = signReceipt(receipt, deps.supplierKey.privateKeyHex);
+    const resultHash = receiptResultHash(signed);
+
+    // ── 4+5. Submit tx: build + confirm as ONE wallet critical section ──
+    const submitOutcome = await deps.state.walletMutex.run(async () => {
+      let built: { txCborHex: string; expectedTxHash: string };
+      try {
+        built = await buildSubmitTx({
+          chain: deps.chain,
+          supplierKey: deps.supplierKey,
+          escrowRef: claimedRef,
+          receiptHash: resultHash,
+        });
+      } catch (err) {
+        return { kind: "build_failed" as const, err };
+      }
+      try {
+        await deps.chain.awaitTx(built.expectedTxHash, 60_000);
+      } catch (err) {
+        return { kind: "await_failed" as const, err };
+      }
+      return { kind: "ok" as const };
+    });
+    if (submitOutcome.kind === "build_failed") {
+      const message = submitOutcome.err instanceof Error ? submitOutcome.err.message : String(submitOutcome.err);
+      console.warn(
+        `[job_failed] jobId=${jobId} reason=submit_failed httpStatus=502 msg=${message}`,
+      );
+      deps.jobs.fail(jobId, {
+        httpStatus: 502,
+        reason: "submit_failed",
+        message: `Submit tx failed: ${message}`,
+      });
+      return;
+    }
+    if (submitOutcome.kind === "await_failed") {
+      const msg = submitOutcome.err instanceof Error ? submitOutcome.err.message : String(submitOutcome.err);
+      console.warn(
+        `[job_failed] jobId=${jobId} reason=submit_timeout httpStatus=502 msg=${msg}`,
+      );
+      deps.jobs.fail(jobId, {
+        httpStatus: 502,
+        reason: "submit_timeout",
+        message: `Submit awaitTx failed: ${msg}`,
+      });
+      return;
+    }
+
+    // ── 6. Mark complete ──────────────────────────────────────────────
+    const payload: OcrJobResponsePayload = {
+      kind: "ocr",
+      output_format: requestBody.output_format,
+      content: inference.content,
+      usage: {
+        prompt_tokens: inference.prompt_tokens,
+        completion_tokens: inference.completion_tokens,
+        total_tokens: inference.prompt_tokens + inference.completion_tokens,
+      },
+      receipt: signed.receipt as unknown as Record<string, unknown>,
+      receipt_signature: signed.signature,
+    };
+    deps.jobs.complete(jobId, payload);
+  } finally {
+    try {
+      deps.state.release(escrowRef);
+    } catch {
+      /* never let lock-release escape */
+    }
+  }
+}

supplier/src/jobs.ts

@@ -63,13 +63,35 @@ export interface TtsJobResponsePayload {
   receipt_signature: string;
 }
 
+/**
+ * OCR-flavoured terminal payload. Stored when a `runOcrJob` succeeds.
+ * `content` is the extraction in the requested output_format (markdown/
+ * html/json as text). The receipt's response_hash commits
+ * sha256(canonical({content, output_format})).
+ */
+export interface OcrJobResponsePayload {
+  kind: "ocr";
+  output_format: string;     // "markdown" | "html" | "json"
+  content: string;
+  usage: {
+    prompt_tokens: number;
+    completion_tokens: number;
+    total_tokens: number;
+  };
+  receipt: Record<string, unknown>;
+  receipt_signature: string;
+}
+
 /**
  * Discriminated union of every job-payload shape the store can hold.
  * `kind` defaults to "chat" on the chat shape (older code paths don't
  * set it; new `runChatJob` writes can set kind: "chat" explicitly to make
  * the discriminator narrowing reliable).
  */
-export type JobResponsePayload = ChatJobResponsePayload | TtsJobResponsePayload;
+export type JobResponsePayload =
+  | ChatJobResponsePayload
+  | TtsJobResponsePayload
+  | OcrJobResponsePayload;
 
 export interface JobFailure {
   httpStatus: number;

supplier/src/ocrVision.ts

@@ -0,0 +1,271 @@
+/**
+ * supplier/src/ocrVision.ts — HTTP client for an OpenAI-compatible VISION
+ * /v1/chat/completions endpoint (vLLM serving an OCR VLM, e.g.
+ * datalab-to/chandra-ocr-2). Used by the `ocr` capability kind.
+ *
+ * callOcrVision({ baseUrl, model, request, timeoutMs, ... })
+ *   POST ${baseUrl}/v1/chat/completions with a single user message whose
+ *   content is OpenAI vision content parts:
+ *     [{type:"image_url", image_url:{url:"data:<mime>;base64,<b64>"}},
+ *      {type:"text", text:<instruction for the output format>}]
+ *   Returns { content, prompt_tokens, completion_tokens, wallclock_ms }
+ *   — the same shape as ollama.ts / openai.ts so receipt construction is
+ *   identical downstream.
+ *
+ * The instruction text per output_format is a serving-side implementation
+ * detail (like temperature); the buyer's on-chain commitment covers the
+ * OcrRequest envelope, not the serving prompt. The exact Chandra prompt
+ * convention gets pinned against a live endpoint during integration —
+ * override per deployment via OCR_PROMPT_OVERRIDE without a code change.
+ *
+ * Error reasons:
+ *   "ocr_failure"   — non-2xx HTTP, network error, unexpected fetch failure
+ *   "ocr_timeout"   — request exceeded timeoutMs
+ *   "ocr_malformed" — response missing choices[0].message.content (or empty)
+ *
+ * probeRevision() — best-effort served-checkpoint probe (light v0). GETs a
+ * runtime info route and extracts a revision/SHA field if one is present.
+ * The route handler refuses to Claim on a mismatch when a pin is
+ * configured. When the receipt schema is extended to carry it, the probed
+ * value feeds `model_revision`; until then it is a claim gate and
+ * a log line.
+ */
+
+import type { OcrRequest } from "@marketplace/shared/tx";
+
+export interface CallOcrVisionParams {
+  baseUrl: string;
+  model: string;
+  request: OcrRequest;
+  timeoutMs: number;
+  /** Optional bearer token (hosted endpoints). "" sends no header. */
+  apiKey?: string;
+  /** Hard output-token ceiling forwarded as max_tokens when > 0. */
+  maxTokens?: number;
+  /** Overrides the built-in per-format instruction when non-empty. */
+  promptOverride?: string;
+}
+
+export interface OcrVisionResult {
+  content: string;
+  prompt_tokens: number;
+  completion_tokens: number;
+  wallclock_ms: number;
+}
+
+export type OcrVisionErrorReason = "ocr_failure" | "ocr_timeout" | "ocr_malformed";
+
+export class OcrVisionError extends Error {
+  public readonly reason: OcrVisionErrorReason;
+  constructor(reason: OcrVisionErrorReason, message?: string) {
+    super(message ?? reason);
+    this.name = "OcrVisionError";
+    this.reason = reason;
+  }
+}
+
+function isAbortError(err: unknown): boolean {
+  if (err instanceof Error) {
+    if (err.name === "AbortError") return true;
+    if ((err as { code?: string }).code === "ABORT_ERR") return true;
+  }
+  return false;
+}
+
+/** Default extraction instructions per output_format. Placeholder wording
+ * pending the Chandra-convention pin (integration item); serving deployments
+ * override via OCR_PROMPT_OVERRIDE. */
+export function defaultOcrInstruction(outputFormat: string): string {
+  switch (outputFormat) {
+    case "html":
+      return "Convert this document page to clean, semantic HTML. Preserve reading order, tables, form fields and layout structure. Output only the HTML.";
+    case "json":
+      return "Extract this document page as JSON with layout blocks (type, bbox, text) preserving reading order, tables and form fields. Output only the JSON.";
+    case "markdown":
+    default:
+      return "Convert this document page to clean markdown. Preserve reading order, tables and structure. Transcribe handwriting faithfully. Output only the markdown.";
+  }
+}
+
+export async function callOcrVision(params: CallOcrVisionParams): Promise<OcrVisionResult> {
+  const { baseUrl, model, request, timeoutMs, apiKey, maxTokens, promptOverride } = params;
+  const url = `${baseUrl.replace(/\/+$/, "")}/v1/chat/completions`;
+
+  const instruction =
+    promptOverride && promptOverride.length > 0
+      ? promptOverride
+      : defaultOcrInstruction(request.output_format);
+
+  const payload: Record<string, unknown> = {
+    model,
+    messages: [
+      {
+        role: "user",
+        content: [
+          {
+            type: "image_url",
+            image_url: { url: `data:${request.mime};base64,${request.image_b64}` },
+          },
+          { type: "text", text: instruction },
+        ],
+      },
+    ],
+    stream: false,
+  };
+  if (typeof maxTokens === "number" && maxTokens > 0) payload.max_tokens = maxTokens;
+  const body = JSON.stringify(payload);
+
+  const headers: Record<string, string> = { "content-type": "application/json" };
+  if (apiKey) {
+    headers.authorization = `Bearer ${apiKey}`;
+  }
+
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(), timeoutMs);
+
+  const startedAt = Date.now();
+  let response: Response;
+  try {
+    response = await fetch(url, {
+      method: "POST",
+      headers,
+      body,
+      signal: controller.signal,
+    });
+  } catch (err) {
+    if (isAbortError(err)) {
+      throw new OcrVisionError("ocr_timeout", `OCR request exceeded ${timeoutMs}ms`);
+    }
+    throw new OcrVisionError(
+      "ocr_failure",
+      `OCR fetch failed: ${(err as Error)?.message ?? String(err)}`,
+    );
+  } finally {
+    clearTimeout(timer);
+  }
+
+  if (!response.ok) {
+    let detail = "";
+    try {
+      detail = await response.text();
+    } catch {
+      // ignore — body unavailable
+    }
+    throw new OcrVisionError(
+      "ocr_failure",
+      `OCR upstream returned HTTP ${response.status}${detail ? `: ${detail.slice(0, 200)}` : ""}`,
+    );
+  }
+
+  let parsed: unknown;
+  try {
+    parsed = await response.json();
+  } catch (err) {
+    throw new OcrVisionError(
+      "ocr_malformed",
+      `OCR response was not valid JSON: ${(err as Error)?.message ?? String(err)}`,
+    );
+  }
+
+  if (!parsed || typeof parsed !== "object") {
+    throw new OcrVisionError("ocr_malformed", "OCR response was not an object");
+  }
+
+  const obj = parsed as Record<string, unknown>;
+  const choicesRaw = obj.choices;
+  if (!Array.isArray(choicesRaw) || choicesRaw.length === 0) {
+    throw new OcrVisionError("ocr_malformed", "OCR response missing 'choices' array");
+  }
+  const firstChoice = choicesRaw[0];
+  if (!firstChoice || typeof firstChoice !== "object") {
+    throw new OcrVisionError("ocr_malformed", "OCR response choices[0] not an object");
+  }
+  const messageRaw = (firstChoice as Record<string, unknown>).message;
+  if (!messageRaw || typeof messageRaw !== "object") {
+    throw new OcrVisionError("ocr_malformed", "OCR response missing 'choices[0].message'");
+  }
+  const content = (messageRaw as Record<string, unknown>).content;
+  if (typeof content !== "string" || content.length === 0) {
+    throw new OcrVisionError(
+      "ocr_malformed",
+      "OCR response missing/empty choices[0].message.content",
+    );
+  }
+
+  const usageRaw = obj.usage;
+  let promptTokens = 0;
+  let completionTokens = 0;
+  if (usageRaw && typeof usageRaw === "object") {
+    const usage = usageRaw as Record<string, unknown>;
+    if (typeof usage.prompt_tokens === "number") promptTokens = usage.prompt_tokens;
+    if (typeof usage.completion_tokens === "number") completionTokens = usage.completion_tokens;
+  }
+
+  return {
+    content,
+    prompt_tokens: promptTokens,
+    completion_tokens: completionTokens,
+    wallclock_ms: Date.now() - startedAt,
+  };
+}
+
+// ─── Revision probe (light v0) ──────────────────────────────────────────────
+
+export interface ProbeRevisionParams {
+  probeUrl: string;
+  timeoutMs: number;
+  apiKey?: string;
+}
+
+export interface ProbeRevisionResult {
+  ok: boolean;
+  /** 40-hex checkpoint SHA when the runtime reports one; "" otherwise. */
+  servedSha: string;
+  /** Human-readable detail for logs/alerts. */
+  detail: string;
+}
+
+const SHA40_RE = /^[0-9a-f]{40}$/i;
+
+/** Fields observed across serving runtimes that carry a checkpoint id
+ * (TGI /info: model_sha; vLLM variants: sha / revision / model_revision). */
+const SHA_FIELDS = ["model_sha", "sha", "revision", "model_revision"];
+
+export async function probeRevision(params: ProbeRevisionParams): Promise<ProbeRevisionResult> {
+  const { probeUrl, timeoutMs, apiKey } = params;
+  const headers: Record<string, string> = {};
+  if (apiKey) headers.authorization = `Bearer ${apiKey}`;
+
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(), timeoutMs);
+  try {
+    const response = await fetch(probeUrl, { headers, signal: controller.signal });
+    if (!response.ok) {
+      return { ok: false, servedSha: "", detail: `probe HTTP ${response.status}` };
+    }
+    let parsed: unknown;
+    try {
+      parsed = await response.json();
+    } catch {
+      return { ok: false, servedSha: "", detail: "probe response not JSON" };
+    }
+    if (!parsed || typeof parsed !== "object") {
+      return { ok: false, servedSha: "", detail: "probe response not an object" };
+    }
+    const obj = parsed as Record<string, unknown>;
+    for (const field of SHA_FIELDS) {
+      const v = obj[field];
+      if (typeof v === "string" && SHA40_RE.test(v)) {
+        return { ok: true, servedSha: v.toLowerCase(), detail: `probe ${field}=${v}` };
+      }
+    }
+    return { ok: true, servedSha: "", detail: "probe answered without a revision field" };
+  } catch (err) {
+    const msg = isAbortError(err)
+      ? `probe exceeded ${timeoutMs}ms`
+      : `probe failed: ${(err as Error)?.message ?? String(err)}`;
+    return { ok: false, servedSha: "", detail: msg };
+  } finally {
+    clearTimeout(timer);
+  }
+}

supplier/src/openai.ts

@@ -52,6 +52,13 @@ export interface CallOpenAiParams {
    * context window (some providers 400 when it exceeds context).
    */
   maxTokens?: number;
+  /**
+   * Forwarded as the OpenAI `user` field when non-empty. Stateful upstreams
+   * (e.g. an OpenClaw gateway) key their agent session off it, so passing the
+   * escrow ref pins one upstream session per marketplace chat session.
+   * Stateless providers ignore it.
+   */
+  user?: string;
   /**
    * When true, sends `reasoning: { enabled: false }` to disable OpenRouter
    * "thinking" tokens — faster/cheaper straight answers, and avoids reasoning
@@ -93,11 +100,12 @@ function isAbortError(err: unknown): boolean {
 }
 
 export async function callOpenAi(params: CallOpenAiParams): Promise<OpenAiResult> {
-  const { baseUrl, model, messages, timeoutMs, apiKey, maxTokens, disableReasoning } = params;
+  const { baseUrl, model, messages, timeoutMs, apiKey, maxTokens, disableReasoning, user } = params;
   const url = `${baseUrl}/v1/chat/completions`;
   const payload: Record<string, unknown> = { model, messages, stream: false };
   if (typeof maxTokens === "number" && maxTokens > 0) payload.max_tokens = maxTokens;
   if (disableReasoning) payload.reasoning = { enabled: false };
+  if (user) payload.user = user;
   const body = JSON.stringify(payload);
 
   const headers: Record<string, string> = { "content-type": "application/json" };
@@ -215,7 +223,7 @@ export async function callOpenAiStream(
   params: CallOpenAiParams,
   onToken: (delta: string) => void,
 ): Promise<OpenAiResult> {
-  const { baseUrl, model, messages, timeoutMs, apiKey, maxTokens, disableReasoning, tools, toolChoice } = params;
+  const { baseUrl, model, messages, timeoutMs, apiKey, maxTokens, disableReasoning, tools, toolChoice, user } = params;
   const url = `${baseUrl}/v1/chat/completions`;
   const payload: Record<string, unknown> = {
     model,
@@ -225,6 +233,7 @@ export async function callOpenAiStream(
   };
   if (typeof maxTokens === "number" && maxTokens > 0) payload.max_tokens = maxTokens;
   if (disableReasoning) payload.reasoning = { enabled: false };
+  if (user) payload.user = user;
   if (Array.isArray(tools) && tools.length > 0) {
     payload.tools = tools;
     if (toolChoice !== undefined) payload.tool_choice = toolChoice;

supplier/src/server.ts

@@ -44,12 +44,19 @@ import {
   mockSlotToWallclockMs,
   detectCborBackend,
   normalizeChatMessage,
+  ocrPromptHash,
+  ALLOWED_OCR_MIMES,
+  ALLOWED_OCR_OUTPUT_FORMATS,
+  MAX_OCR_IMAGE_B64_CHARS,
 } from "@marketplace/shared/tx";
+import type { OcrRequest } from "@marketplace/shared/tx";
 
 import type { SupplierState } from "./state.js";
 import type { SupplierConfig } from "./config.js";
 import { JobStore } from "./jobs.js";
-import { runChatJob, runTtsJob } from "./jobRunner.js";
+import { runChatJob, runTtsJob, runOcrJob } from "./jobRunner.js";
+import { probeRevision } from "./ocrVision.js";
+import { probeDatalabHealth } from "./datalabOcr.js";
 import { ChatSessionStore, type ChatSessionRecord } from "./chatSession.js";
 import { endChatSession, type EndChatSessionDeps } from "./chatSessionRunner.js";
 import { callOpenAiStream } from "./openai.js";
@@ -427,9 +434,9 @@ function makeGetJobHandler(deps: ResolvedDeps) {
       // Defensive narrowing: this poll route only knows how to render the
       // chat shape. A future shared JobStore could mix capabilities; we'd
       // want each kind's poll to be sure it's reading its own.
-      if ("audio_b64" in payload) {
+      if (payload.kind === "ocr" || "audio_b64" in payload) {
         return jsonError(res, 500, "wrong_payload_kind",
-          "tts payload returned to chat poll route");
+          "non-chat payload returned to chat poll route");
       }
       return res.status(200).json({
         choices: payload.choices,
@@ -696,6 +703,303 @@ function makeGetTtsJobHandler(deps: ResolvedDeps) {
   };
 }
 
+// ─── POST /v1/ocr/extract ───────────────────────────────────────────────────
+//
+// Mirror of makeTtsHandler for model-scoped `ocr.page.extract.<model-slug>.v1`
+// capabilities. Same lifecycle (validate → resolve advert → resolve escrow →
+// hash check → claim → fire-and-forget runner → 202). Differences:
+//   - body shape: { image_b64, mime, output_format } (one page per job)
+//   - prompt_hash is ocrPromptHash(envelope) — shared with the buyer-side
+//     builder in packages/shared/src/tx/escrow/postOcrEscrow.ts
+//   - optional served-revision probe BEFORE Claim: a mispinned or
+//     unreachable upstream refuses the job instead of stranding the escrow
+//     in Claimed.
+
+/** Plain base64 (standard alphabet, optional padding). Shared shape with the
+ * buyer-side validator; kept local so route validation stays dependency-light. */
+const OCR_BASE64_RE = /^[A-Za-z0-9+/]+={0,2}$/;
+
+function makeOcrHandler(deps: ResolvedDeps) {
+  return async (req: Request, res: Response, next: NextFunction) => {
+    // Slot held by THIS request (null until acquired / after hand-off).
+    let acquiredRef: string | null = null;
+    try {
+      // ── 1. Header ───────────────────────────────────────────────────
+      const headerVal = req.header("X-Escrow-Ref");
+      if (!headerVal) {
+        return jsonError(res, 400, "escrow_ref_required", "X-Escrow-Ref header is required");
+      }
+      const escrowRef = parseEscrowRef(headerVal);
+      if (escrowRef === null) {
+        return jsonError(res, 400, "escrow_ref_malformed",
+          'X-Escrow-Ref must match "<64-hex>#<int>"');
+      }
+      const escrowRefStr = `${escrowRef.txHash}#${escrowRef.index}`;
+
+      // ── 2. Body ─────────────────────────────────────────────────────
+      const body = (req.body ?? {}) as Record<string, unknown>;
+      const imageB64 = typeof body.image_b64 === "string" ? body.image_b64 : "";
+      if (imageB64.length === 0) {
+        return jsonError(res, 400, "image_required",
+          "body.image_b64 must be a non-empty base64 string");
+      }
+      if (imageB64.length > MAX_OCR_IMAGE_B64_CHARS) {
+        return jsonError(res, 413, "image_too_large",
+          `body.image_b64 exceeds ${MAX_OCR_IMAGE_B64_CHARS} chars`);
+      }
+      if (!OCR_BASE64_RE.test(imageB64)) {
+        return jsonError(res, 400, "image_not_base64",
+          "body.image_b64 must be plain base64 (no data: prefix, no whitespace)");
+      }
+      const mime = typeof body.mime === "string" ? body.mime : "";
+      if (!ALLOWED_OCR_MIMES.has(mime)) {
+        return jsonError(res, 400, "mime_invalid",
+          `mime must be one of: ${[...ALLOWED_OCR_MIMES].join(", ")}`);
+      }
+      const outputFormat = typeof body.output_format === "string" ? body.output_format : "";
+      if (!ALLOWED_OCR_OUTPUT_FORMATS.has(outputFormat)) {
+        return jsonError(res, 400, "output_format_invalid",
+          `output_format must be one of: ${[...ALLOWED_OCR_OUTPUT_FORMATS].join(", ")}`);
+      }
+      const ocrRequest: OcrRequest = { image_b64: imageB64, mime, output_format: outputFormat };
+
+      // ── 3. Advert ───────────────────────────────────────────────────
+      const advertResult = await fetchActiveAdvert(deps);
+      if ("error" in advertResult) {
+        const e = advertResult.error;
+        return jsonError(res, e.status, e.reason, e.message);
+      }
+      const advert = advertResult.datum;
+
+      // ── 4. Escrow ───────────────────────────────────────────────────
+      const escrowUtxo = await deps.chain.queryUtxo(escrowRef);
+      if (escrowUtxo === null || !escrowUtxo.datumHex) {
+        return jsonError(res, 404, "escrow_not_found",
+          `escrow UTxO ${escrowRefStr} not found on chain`);
+      }
+      let escrowDatum: EscrowDatum;
+      try {
+        escrowDatum = decodeEscrowDatum(escrowUtxo.datumHex);
+      } catch (err) {
+        return jsonError(res, 404, "escrow_decode_failed", (err as Error).message);
+      }
+
+      // ── 5. State / identity / capability ────────────────────────────
+      if (escrowDatum.state !== "Open") {
+        return jsonError(res, 409, "escrow_not_claimable",
+          `escrow state is ${escrowDatum.state}, expected Open`);
+      }
+      if (escrowDatum.supplier_pkh !== deps.supplierKey.pubKeyHash) {
+        return jsonError(res, 403, "wrong_supplier",
+          "escrow supplier_pkh does not match this node");
+      }
+      if (escrowDatum.capability_id !== advert.capability_id) {
+        return jsonError(res, 409, "capability_mismatch",
+          `escrow capability ${escrowDatum.capability_id} != advert ${advert.capability_id}`);
+      }
+
+      // ── 6. Hash checks ──────────────────────────────────────────────
+      const expectedRequestSpecHash = sha256Hex(canonicalize({
+        capability_id: advert.capability_id,
+        max_output_tokens: advert.max_output_tokens,
+        model: advert.model,
+      }));
+      if (escrowDatum.request_spec_hash !== expectedRequestSpecHash) {
+        return jsonError(res, 409, "request_spec_mismatch",
+          "request_spec_hash in escrow does not match advert spec");
+      }
+      // The OCR prompt commitment is the canonicalisation of the full
+      // request envelope. The buyer SDK computes the same shape (shared
+      // ocrPromptHash) when building the escrow datum.
+      const expectedPromptHash = ocrPromptHash(ocrRequest);
+      if (escrowDatum.prompt_hash !== expectedPromptHash) {
+        return jsonError(res, 409, "prompt_mismatch",
+          "prompt_hash in escrow does not match request body");
+      }
+
+      // ── 7. Deadline ─────────────────────────────────────────────────
+      const tipSlot = await deps.chain.tip();
+      const isLive = detectCborBackend(deps.chain) === "live";
+      const nowMs = isLive
+        ? Date.now()
+        : Math.max(mockSlotToWallclockMs(tipSlot), escrowDatum.posted_at);
+      if (nowMs >= escrowDatum.deliver_by) {
+        return jsonError(res, 408, "past_deliver_by",
+          `now ${nowMs} >= deliver_by ${escrowDatum.deliver_by}`);
+      }
+
+      // ── 7.5. Upstream probe (before Claim, when configured) ─────────
+      // A failure REFUSES the job here — claiming with a dead or mispinned
+      // upstream would strand the escrow (and, for self-hosted supply,
+      // corrupt the attestation story). 503 tells the buyer to retry/route
+      // elsewhere. Datalab upstream: liveness only (hosted supply has no
+      // checkpoint to pin). openai-vision upstream: liveness + optional
+      // served-revision match against HF_MODEL_REVISION.
+      if (deps.config.revisionProbeMode === "per_job") {
+        if (deps.config.ocrUpstream === "datalab") {
+          const probe = await probeDatalabHealth({
+            baseUrl: deps.config.datalabBaseUrl,
+            apiKey: deps.config.datalabApiKey,
+            timeoutMs: 10_000,
+          });
+          if (!probe.ok) {
+            console.warn(`[upstream_probe_failed] ${probe.detail}`);
+            return jsonError(res, 503, "upstream_probe_failed",
+              `upstream health probe failed: ${probe.detail}`);
+          }
+        } else if (deps.config.revisionProbeUrl !== "") {
+          const probe = await probeRevision({
+            probeUrl: deps.config.revisionProbeUrl,
+            timeoutMs: 10_000,
+            apiKey: deps.config.openaiApiKey,
+          });
+          if (!probe.ok) {
+            console.warn(`[revision_probe_failed] ${probe.detail}`);
+            return jsonError(res, 503, "revision_probe_failed",
+              `served-revision probe failed: ${probe.detail}`);
+          }
+          if (
+            deps.config.hfModelRevision !== "" &&
+            probe.servedSha !== "" &&
+            probe.servedSha !== deps.config.hfModelRevision.toLowerCase()
+          ) {
+            console.warn(
+              `[revision_mismatch] expected=${deps.config.hfModelRevision} served=${probe.servedSha}`,
+            );
+            return jsonError(res, 503, "revision_mismatch",
+              "served checkpoint does not match the pinned revision");
+          }
+        }
+      }
+
+      // ── 8. Acquire session slot ─────────────────────────────────────
+      if (!deps.state.tryAcquire(escrowRefStr)) {
+        return jsonError(res, 409, "supplier_busy", "supplier is already working another job");
+      }
+      acquiredRef = escrowRefStr;
+
+      // ── 9. Claim tx: build + confirm as ONE wallet critical section ──
+      const claimOutcome = await deps.state.walletMutex.run(async () => {
+        let built;
+        try {
+          built = await buildClaimTx({
+            chain: deps.chain,
+            supplierKey: deps.supplierKey,
+            escrowRef,
+          });
+        } catch (err) {
+          return { kind: "build_failed" as const, err };
+        }
+        try {
+          await deps.chain.awaitTx(built.expectedTxHash, 60_000);
+        } catch (err) {
+          return { kind: "await_failed" as const, err };
+        }
+        return { kind: "ok" as const, built };
+      });
+      if (claimOutcome.kind === "build_failed") {
+        deps.state.release(escrowRefStr);
+        acquiredRef = null;
+        triggerOnFailureConsolidate({
+          chain: deps.chain,
+          state: deps.state,
+          supplierKey: deps.supplierKey,
+        });
+        return jsonError(res, 503, "chain_submit_failed",
+          `Claim tx submit failed: ${(claimOutcome.err as Error).message}`);
+      }
+      if (claimOutcome.kind === "await_failed") {
+        deps.state.release(escrowRefStr);
+        acquiredRef = null;
+        return jsonError(res, 504, "claim_timeout",
+          `Claim awaitTx failed: ${(claimOutcome.err as Error).message}`);
+      }
+
+      // ── 10. Spawn runner + 202 ──────────────────────────────────────
+      const claimedRef: OutputReference = {
+        txHash: claimOutcome.built.expectedTxHash,
+        index: 0,
+      };
+      const jobId = deps.jobs.create(escrowRefStr);
+
+      // Slot release now happens inside runOcrJob's finally.
+      acquiredRef = null;
+      void runOcrJob({
+        deps: {
+          chain: deps.chain,
+          state: deps.state,
+          config: deps.config,
+          supplierKey: deps.supplierKey,
+          jobs: deps.jobs,
+        },
+        jobId,
+        escrowRef: escrowRefStr,
+        claimedRef,
+        advert,
+        escrowDatum,
+        requestBody: ocrRequest,
+      });
+
+      return res.status(202).json({
+        job_id: jobId,
+        status: "accepted",
+        escrow_ref: escrowRefStr,
+      });
+    } catch (err) {
+      if (acquiredRef) {
+        try { deps.state.release(acquiredRef); } catch { /* ignore */ }
+      }
+      next(err);
+      return;
+    }
+  };
+}
+
+// ─── GET /v1/ocr/extract/:jobId ─────────────────────────────────────────────
+function makeGetOcrJobHandler(deps: ResolvedDeps) {
+  return (req: Request, res: Response) => {
+    const rawJobId = req.params.jobId;
+    const jobId = typeof rawJobId === "string" ? rawJobId : "";
+    if (!UUID_V4_RE.test(jobId)) {
+      return jsonError(res, 400, "invalid_job_id", "jobId must be a UUIDv4 string");
+    }
+    const record = deps.jobs.get(jobId);
+    if (!record) {
+      return jsonError(res, 404, "job_not_found", `no job found with id ${jobId}`);
+    }
+    res.setHeader("Content-Type", "application/json");
+    if (record.status === "accepted" || record.status === "running") {
+      return res.status(202).json({
+        status: record.status,
+        escrow_ref: record.escrowRef,
+      });
+    }
+    if (record.status === "done") {
+      const payload = record.responsePayload!;
+      if (!("kind" in payload) || payload.kind !== "ocr") {
+        return jsonError(res, 500, "wrong_payload_kind",
+          "non-ocr payload returned to ocr poll route");
+      }
+      return res.status(200).json({
+        output_format: payload.output_format,
+        content: payload.content,
+        usage: payload.usage,
+        receipt: payload.receipt,
+        receipt_signature: payload.receipt_signature,
+        escrow_ref: record.escrowRef,
+      });
+    }
+    // failed
+    const f = record.failure!;
+    return res.status(f.httpStatus).json({
+      status: "failed",
+      reason: f.reason,
+      message: f.message,
+      escrow_ref: record.escrowRef,
+    });
+  };
+}
+
 // ─── /v1/chat/{start,message,end} — multi-turn chat session (llm.chat.v1) ───
 //
 // Escrow bookends with off-chain turns:
@@ -968,17 +1272,23 @@ function makeChatSessionHandlers(deps: ResolvedDeps) {
       };
 
       try {
+        // Stateful upstreams (OpenClaw) key a persistent agent session off the
+        // OpenAI `user` field and carry the history themselves — send only this
+        // turn's delta or the transcript duplicates into the upstream context
+        // every turn. Stateless upstreams get the full transcript as before.
+        const stateful = deps.config.openaiSessionPassthrough;
         const result = await callOpenAiStream(
           {
             baseUrl: deps.config.openaiBaseUrl,
-            model: record.advert.model,
+            model: deps.config.openaiModelOverride || record.advert.model,
-            messages: record.transcript,
+            messages: stateful ? delta : record.transcript,
             timeoutMs: deps.config.openaiTimeoutMs,
             apiKey: deps.config.openaiApiKey,
             maxTokens: deps.config.openaiMaxTokens,
             disableReasoning: deps.config.openaiReasoningDisabled,
             tools,
             toolChoice,
+            user: stateful ? escrowRefStr : undefined,
           },
           (tok) => sse({ type: "token", value: tok }),
         );
@@ -1077,7 +1387,11 @@ export function createApp(deps: SupplierDeps): Application {
 
   const app = express();
   app.disable("x-powered-by");
-  app.use(express.json({ limit: "1mb" }));
+  // /v1/ocr carries a base64 page image and mounts its own larger parser
+  // below; every other route 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));
 
   // /healthz is mounted FIRST and uses no deps — it must succeed regardless
   // of chain/state/config (independent of /status free/working/offline).
@@ -1095,6 +1409,11 @@ export function createApp(deps: SupplierDeps): Application {
   if (resolved.config.capabilityKind === "tts") {
     app.post("/v1/audio/synthesize", makeTtsHandler(resolved));
     app.get("/v1/audio/synthesize/:jobId", makeGetTtsJobHandler(resolved));
+  } else if (resolved.config.capabilityKind === "ocr") {
+    // MAX_OCR_IMAGE_B64_CHARS (~12M chars ≈ 9 MB binary) + JSON envelope
+    // fits comfortably under 16mb.
+    app.post("/v1/ocr/extract", express.json({ limit: "16mb" }), makeOcrHandler(resolved));
+    app.get("/v1/ocr/extract/:jobId", makeGetOcrJobHandler(resolved));
   } else if (resolved.config.capabilityKind === "chat-session") {
     const chat = makeChatSessionHandlers(resolved);
     app.post("/v1/chat/start", chat.start);

supplier/src/walletHealth.ts

@@ -7,15 +7,26 @@
  * and the lane stalls until an operator runs `tx:consolidate-wallet`.
  *
  * Two layers of self-healing:
- *   startWalletHealthTicker — periodic background tick (every N ms). When
+ *   startWalletHealthTicker — periodic background tick (every N ms). Runs
- *     the supplier is idle, runs runConsolidateWallet. planConsolidate
+ *     runConsolidateWallet; planConsolidate's structural health predicate
- *     returns `already-healthy` cheaply for clean 2-UTxO wallets, so this
+ *     short-circuits without a tx when a collateral candidate exists and
- *     is a no-op in steady state — no tx, no fee.
+ *     fragmentation is bounded. A tick on a healthy wallet costs one UTxO
+ *     query and nothing else.
  *   triggerOnFailureConsolidate — fire-and-forget consolidate from the
  *     Claim-failure path. Debounced so back-to-back failures don't queue
  *     multiple consolidates. Returns 503 to the buyer for THIS request; the
  *     next buyer retry (~30-60s later) finds the wallet healthy.
  *
+ * F7 guards (2026-08-08 incident: a non-idempotent health check made every
+ * tick consolidate, burning 0.5 AP3X per 45s until the wallet hit the
+ * floor — 116 AP3X in one soak). Two independent circuit breakers protect
+ * against ANY future non-convergence:
+ *   - consecutive guard: 3 consolidations in a row without ever reaching
+ *     `already-healthy` halts wallet-health for this process.
+ *   - budget guard: more than 6 consolidations inside 24h halts it too.
+ * A halt logs loudly once; consolidate manually and restart the supplier
+ * to re-arm.
+ *
  * Wallet-lock discipline: consolidation spends EVERY wallet UTxO, so it runs
  * inside state.walletMutex — the same mutex that serializes Claims/Submits —
  * queueing behind any in-flight chain op instead of skipping. Session slots
@@ -63,6 +74,29 @@ export interface WalletHealthTicker {
 const inFlight = new WeakSet<SupplierState>();
 const lastOnFailureAt = new WeakMap<SupplierState, number>();
 
+// ── F7 circuit breakers ──────────────────────────────────────────────
+const MAX_CONSECUTIVE_CONSOLIDATIONS = 3;
+const MAX_CONSOLIDATIONS_PER_WINDOW = 6;
+const BUDGET_WINDOW_MS = 24 * 60 * 60 * 1000;
+
+interface GuardCounters {
+  consecutive: number;
+  windowStart: number;
+  windowCount: number;
+  halted: boolean;
+}
+
+const guardCounters = new WeakMap<SupplierState, GuardCounters>();
+
+function guardsFor(state: SupplierState): GuardCounters {
+  let g = guardCounters.get(state);
+  if (!g) {
+    g = { consecutive: 0, windowStart: Date.now(), windowCount: 0, halted: false };
+    guardCounters.set(state, g);
+  }
+  return g;
+}
+
 function defaultLog(line: string): void {
   // eslint-disable-next-line no-console
   console.log(`[wallet-health] ${line}`);
@@ -73,6 +107,8 @@ async function consolidateOnce(
   opts: { collateralLovelace: bigint; awaitTimeoutMs: number; log: (line: string) => void },
 ): Promise<void> {
   const { state } = deps;
+  const guards = guardsFor(state);
+  if (guards.halted) return; // the halt already logged loudly, once
   if (inFlight.has(state)) {
     opts.log("skip: another consolidate already in flight");
     return;
@@ -92,11 +128,36 @@ async function consolidateOnce(
       }),
     );
     if (result.reason === "already-healthy") {
+      guards.consecutive = 0;
       opts.log("already-healthy");
     } else {
+      const now = Date.now();
+      if (now - guards.windowStart > BUDGET_WINDOW_MS) {
+        guards.windowStart = now;
+        guards.windowCount = 0;
+      }
+      guards.consecutive += 1;
+      guards.windowCount += 1;
       opts.log(`consolidated: ${result.reason} txHash=${result.txHash}`);
+      if (guards.consecutive >= MAX_CONSECUTIVE_CONSOLIDATIONS) {
+        guards.halted = true;
+        opts.log(
+          `HALT: ${guards.consecutive} consecutive consolidations without reaching a healthy shape — ` +
+            `wallet-health stopped to protect funds (F7 guard). Investigate planConsolidate vs the ` +
+            `builder's output shape, consolidate manually, restart the supplier to re-arm.`,
+        );
+      } else if (guards.windowCount >= MAX_CONSOLIDATIONS_PER_WINDOW) {
+        guards.halted = true;
+        opts.log(
+          `HALT: consolidation budget exhausted (${guards.windowCount} in 24h) — wallet-health ` +
+            `stopped to protect funds (F7 guard). Restart the supplier to re-arm.`,
+        );
+      }
     }
   } catch (err) {
+    // Neither healthy nor consolidated (e.g. balance too low): no fee was
+    // spent, so the guards don't move — but nothing converged either, so
+    // the consecutive counter is NOT reset.
     opts.log(`consolidate failed: ${(err as Error).message}`);
   } finally {
     inFlight.delete(state);
@@ -143,8 +204,9 @@ export function triggerOnFailureConsolidate(
   });
 }
 
-/** Test-only: reset the in-flight + debounce maps between tests. */
+/** Test-only: reset the in-flight + debounce + guard maps between tests. */
 export function _resetWalletHealthForTests(state: SupplierState): void {
   inFlight.delete(state);
   lastOnFailureAt.delete(state);
+  guardCounters.delete(state);
 }

tests/fixtures/buyer-side/advert-datum-builders.ts

@@ -118,15 +118,17 @@ export function encodeSampleAdvertDatumHex(): string {
     hexToBytes(datum.supplier_pkh),
     utf8Bytes(datum.capability_id),
     utf8Bytes(datum.model),
-    datum.max_output_tokens,
+    // BigInt so cbor-x emits uint64 (0x1b) like the production codec — plain
-    datum.max_processing_ms,
+    // numbers encode minimally (or as float64 past 2^32) and break golden parity.
+    BigInt(datum.max_output_tokens),
+    BigInt(datum.max_processing_ms),
     datum.price_lovelace,
     datum.supplier_bond_lovelace,
     datum.buyer_bond_lovelace,
     utf8Bytes(datum.endpoint_url),
     utf8Bytes(datum.detail_uri),
     hexToBytes(datum.detail_hash),
-    datum.advertised_at,
+    BigInt(datum.advertised_at),
     statusTag,
   ]);