agent402-openclaw 0.4.2 → 0.4.4

package/README.md

@@ -1,5 +1,7 @@
 # agent402-openclaw
 
+Part of [Agent402](https://agent402.tools), the open-source paid door for AI agents.
+
 Agent402 as an [OpenClaw](https://openclaw.ai) model provider: auto-routed and
 explicit frontier models at a **flat per-call price**, paid **by card** (a prepaid
 credits key, no wallet) or in **USDC over x402** from a wallet. The same key and
@@ -46,6 +48,7 @@ client at `http://127.0.0.1:8412/v1` with model `auto`.
 ```bash
 npm i @x402/fetch @x402/evm viem
 export AGENT402_WALLET_KEY=0x...     # an EVM key holding USDC on Base
+export AGENT402_MAX_PER_CALL_USD=2   # per-call ceiling checked before signing (default 2, the gateway's metered cap; 0 disables)
 ```
 
 With no credits key present the proxy signs an x402 payment per call from that
@@ -54,11 +57,11 @@ wallet. The key never leaves the machine.
 Two ways the wallet can pay a metered call:
 
 - **exact** (default): the 402 quotes the call from its body (input + your
-  `max_tokens` at the model's list price, x1.15) and the wallet pays that quote.
+  `max_tokens`) and the wallet pays that quote.
 - **upto, actual usage**: run `agent402-openclaw permit2-approve` once (one USDC
   approval transaction on Base; the wallet needs a little ETH for gas). From
   then on the proxy authorizes the quote as a CEILING and the gateway settles
-  what the call actually cost x1.15, so a short answer costs a fraction of the
+  what the call actually used, so a short answer costs a fraction of the
   quote. `agent402-openclaw doctor` says which mode the wallet is in.
 
 Card / credits buyers already pay actual usage on metered calls.
@@ -68,12 +71,14 @@ Card / credits buyers already pay actual usage on metered calls.
 `auto` (routed per prompt, flat $0.01/call) plus every id from
 `GET https://agent402.tools/v1/models`. Explicit models are **metered by default**:
 the proxy sends them to the gateway's metered route, where each request is quoted
-from its body (exact-BPE input plus your `max_tokens` at the model's list price,
-times 1.15, from $0.001, capped at $2 per call), so a short call costs a fraction
+from its body (input plus your `max_tokens`, from $0.001, capped at $2 per
+call), so a short call costs a fraction
 of a cent and a long one pays for what it asks. `--flat` (or
 `AGENT402_PRICING=flat`) keeps every model on its flat per-call tier instead.
-Either way OpenClaw's per-token cost fields stay zero; the price is per call. A model sent to the wrong tier is answered with a
-400 naming the right one; nothing is charged. A client-supplied `Idempotency-Key`
+Either way OpenClaw's per-token cost fields stay zero; the price is per call. A model sent to another flat
+tier's route is priced at its home tier: the 402 quotes that tier's price, and the paid call is served under
+that tier's caps (the answer says so in `agent402_tier`). The router, grounded and stealth tiers still answer a
+400 naming the right route; nothing is charged. A client-supplied `Idempotency-Key`
 is passed through (an x402 retry with the same key replays the paid answer);
 without one, each call is its own payment.
 

package/index.js

@@ -47,10 +47,44 @@ export const PERMIT2_READY_MIN = 10n ** 12n; // 1,000,000 USDC (6 decimals): a r
 
 /** Which 402 accept to pay: upto on Base when the client can (settle-actual),
  *  else exact on Base, else the first the client supports. Pure. */
-export function selectAccept(accepts, { preferUpto = true } = {}) {
+/** Per-call ceiling in USD, enforced in the accept selector so an over-cap quote
+ *  is refused BEFORE anything is signed. @x402/core 2.23+ ships a $1
+ *  pegged-assets default on every client; this package turns that off (it would
+ *  refuse USDG and gave an opaque vendor error on any metered turn over $1) and
+ *  carries its own ceiling under its own name. The default is $2, the gateway's
+ *  own metered quote cap (GATEWAY_METERED_MAX_QUOTE_USD), so no call the gateway
+ *  would serve is refused here while a runaway quote still stops at the wallet.
+ *  `maxPerCallUsd` in the plugin config or AGENT402_MAX_PER_CALL_USD; `0`/`off`
+ *  disables; a malformed value reads as the default, never as off. */
+export const DEFAULT_MAX_PER_CALL_USD = 2;
+export function maxPerCallUsd(pluginConfig = {}) {
+  const raw = String(pluginConfig.maxPerCallUsd ?? process.env.AGENT402_MAX_PER_CALL_USD ?? "").trim().toLowerCase();
+  if (raw === "0" || raw === "off") return Infinity;
+  const n = Number(raw);
+  return Number.isFinite(n) && n > 0 ? n : DEFAULT_MAX_PER_CALL_USD;
+}
+/** Dollar value of an accept's quote: `amount` (v2) or `maxAmountRequired` (v1)
+ *  in base units over the asset's `extra.decimals` (USDC on Base: 6). Unknown
+ *  shape reads as null, and null is REFUSED under a finite ceiling: a quote we
+ *  cannot read is not one we sign. */
+export function acceptUsd(accept) {
+  const raw = accept?.amount ?? accept?.maxAmountRequired;
+  if (raw == null || !/^\d+$/.test(String(raw))) return null;
+  const dec = Number(accept?.extra?.decimals ?? 6);
+  if (!Number.isInteger(dec) || dec < 0 || dec > 36) return null;
+  return Number(raw) / 10 ** dec;
+}
+export function selectAccept(accepts, { preferUpto = true, maxUsd = Infinity } = {}) {
   const list = Array.isArray(accepts) ? accepts : [];
-  if (preferUpto) { const u = list.find((a) => a?.scheme === "upto" && a?.network === BASE_CAIP2); if (u) return u; }
-  return list.find((a) => a?.scheme === "exact" && a?.network === BASE_CAIP2) || list[0];
+  const pick = (preferUpto && list.find((a) => a?.scheme === "upto" && a?.network === BASE_CAIP2))
+    || list.find((a) => a?.scheme === "exact" && a?.network === BASE_CAIP2) || list[0];
+  if (pick && Number.isFinite(maxUsd)) {
+    const usd = acceptUsd(pick);
+    if (usd == null || usd > maxUsd) {
+      throw new Error(`agent402-openclaw: quote ${usd == null ? "unreadable" : `$${usd}`} exceeds the per-call ceiling $${maxUsd} (maxPerCallUsd / AGENT402_MAX_PER_CALL_USD; 0 disables)`);
+    }
+  }
+  return pick;
 }
 export const uptoReady = (allowance) => { try { return BigInt(allowance ?? 0) >= PERMIT2_READY_MIN; } catch { return false; } };
 
@@ -70,7 +104,9 @@ export async function resolvePayFetch(pluginConfig = {}, log = () => {}) {
       const pub = createPublicClient({ chain: base, transport: http((pluginConfig.baseRpc || process.env.AGENT402_BASE_RPC || DEFAULT_BASE_RPC).trim()) });
       const allowance = await pub.readContract(getPermit2AllowanceReadParams({ tokenAddress: USDC_BASE, ownerAddress: account.address }));
       upto = uptoReady(allowance);
-      const client = new x402Client((v, accepts) => selectAccept(accepts, { preferUpto: upto }));
+      const maxUsd = maxPerCallUsd(pluginConfig);
+      const client = new x402Client((v, accepts) => selectAccept(accepts, { preferUpto: upto, maxUsd }));
+      client.setSpendControls?.(false); // the vendor default ($1, pegged assets only) would refuse USDG and any quote over $1; selectAccept enforces our own maxPerCallUsd (same $1 default) instead
       registerExactEvmScheme(client, { signer });
       if (upto) client.register(BASE_CAIP2, new UptoEvmScheme(signer));
       log(upto
@@ -80,7 +116,8 @@ export async function resolvePayFetch(pluginConfig = {}, log = () => {}) {
     } catch (e) {
       // The allowance read or the upto scheme failed: exact still works.
       log(`[agent402-openclaw] upto unavailable (${String(e?.message || e).slice(0, 120)}) - paying exact`);
-      const client = new x402Client((v, accepts) => selectAccept(accepts, { preferUpto: false }));
+      const client = new x402Client((v, accepts) => selectAccept(accepts, { preferUpto: false, maxUsd: maxPerCallUsd(pluginConfig) }));
+      client.setSpendControls?.(false); // the vendor default ($1, pegged assets only) would refuse USDG and any quote over $1; selectAccept enforces our own maxPerCallUsd (same $1 default) instead
       registerExactEvmScheme(client, { signer });
       return wrapFetchWithPayment(fetch, client);
     }

package/models.js

@@ -77,7 +77,13 @@ export function routesFromCatalog(catalog, { pricing = "metered" } = {}) {
 
 /** Preferred primaries, in order: capable tool-callers that are cheap when
  *  metered. Only consulted among routes that can hold OpenClaw's prompt. */
-const PRIMARY_PREFERENCE = ["anthropic/claude-haiku-4.5", "openai/gpt-4.1-mini", "google/gemini-2.5-flash", "openai/gpt-4o-mini"];
+// Exported so the gateway's live-catalog guard (scripts/test-gateway-model-ids.js)
+// fails the day any of these ids carries an expiration date upstream: the first
+// pick is haiku-4.5, whose retirement floor is 2026-10-15 with no notice yet.
+// google/gemini-2.5-flash left this list 2026-09-18: the live catalog stamps it
+// expiration_date 2026-10-20, which is exactly what that guard was written to
+// catch on its first run.
+export const PRIMARY_PREFERENCE = ["anthropic/claude-haiku-4.5", "openai/gpt-4.1-mini", "openai/gpt-4o-mini"];
 
 /** The route OpenClaw should make its primary model: the first preferred id
  *  that fits OPENCLAW_MIN_INPUT_CHARS, else the cheapest fitting one, else

package/openclaw.plugin.json

@@ -36,6 +36,13 @@
           "flat"
         ],
         "description": "metered (default): explicit models pay a per-request quote; flat: every model on its flat tier."
+      },
+      "maxPerCallUsd": {
+        "type": [
+          "number",
+          "string"
+        ],
+        "description": "Per-call ceiling in USD for wallet (x402) payments, checked before signing (default 2, the gateway's metered quote cap; 0 or \"off\" disables). Also AGENT402_MAX_PER_CALL_USD."
       }
     }
   },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent402-openclaw",
-  "version": "0.4.2",
+  "version": "0.4.4",
   "description": "OpenClaw model provider for Agent402: auto-routed and explicit LLM models at a flat per-call price, paid by card (prepaid credits key) or in USDC over x402 through a local proxy. Zero dependencies; x402 wallet payment is an optional peer.",
   "type": "module",
   "main": "index.js",

package/proxy.js

@@ -20,6 +20,15 @@
 // own 402->pay retry), which does NOT make two client calls one payment.
 // Streams pass through byte for byte.
 //
+// Refusals: a refused paid call's 402 from an x402 v2 seller carries the whole
+// payment offer in its JSON body (Agent402 mirrors its PAYMENT-REQUIRED header
+// there) beside the refusal's own reason, hint and retry. OpenClaw's model
+// client builds its error message from `error.message`, else from the whole
+// body, so relaying that body put kilobytes of accepts and schemas into the
+// agent's error text in place of the one sentence that says what to fix. Such
+// a body is answered in the OpenAI error shape instead (refusalAsOpenAIError);
+// every other upstream body passes through byte for byte.
+//
 // Loopback only, and browser-hostile on purpose: any web page can POST to
 // 127.0.0.1 with a "simple" no-cors request, and this proxy spends the
 // user's key, so a request carrying an Origin header (browsers always send
@@ -49,6 +58,27 @@ function readBody(req, limit = MAX_BODY) {
   });
 }
 
+// The x402 offer keys a refusal body may carry beside its own fields.
+const OFFER_KEYS = ["x402Version", "resource", "accepts", "extensions"];
+const text = (v) => (typeof v === "string" && v.trim() ? v : null);
+
+/** An upstream error body that carries an x402 offer (a numeric x402Version),
+ *  as an OpenAI-shaped error: the explanation (a problem's detail, else the
+ *  hint, else the error sentence) as error.message, the retry class or reason as
+ *  error.code, the refusal's own fields beside it, and no offer. Null for any
+ *  other body, which the caller relays unchanged. */
+export function refusalAsOpenAIError(status, doc) {
+  if (!doc || typeof doc !== "object" || Array.isArray(doc) || typeof doc.x402Version !== "number") return null;
+  const ours = { ...doc };
+  for (const k of OFFER_KEYS) delete ours[k];
+  const { error, ...fields } = ours;
+  const message = text(fields.detail) || text(fields.hint) || text(error) || text(fields.title) || `Agent402 answered HTTP ${status}`;
+  return {
+    error: { message, type: status === 402 ? "payment_required" : "upstream_error", code: text(fields.retry) || text(fields.reason) || null },
+    ...fields,
+  };
+}
+
 const json = (res, status, obj, headers = {}) => {
   res.writeHead(status, { "content-type": "application/json; charset=utf-8", ...headers });
   res.end(JSON.stringify(obj));
@@ -113,6 +143,18 @@ export async function startProxy({ upstream = DEFAULT_UPSTREAM, creditsKey = nul
         stats.forwarded++;
         const passthrough = {};
         for (const h of ["content-type", "x-credits-balance", "payment-receipt", "x-cache", "cache-control"]) { const v = up.headers.get(h); if (v) passthrough[h] = v; }
+        if (!up.ok && /json/i.test(passthrough["content-type"] || "")) {
+          const raw = await up.text();
+          let doc = null;
+          try { doc = JSON.parse(raw); } catch { /* not JSON after all: relayed as sent */ }
+          const mapped = refusalAsOpenAIError(up.status, doc);
+          if (mapped) {
+            const { "content-type": _ct, ...rest } = passthrough;
+            return json(res, up.status, mapped, rest);
+          }
+          res.writeHead(up.status, passthrough);
+          return res.end(raw);
+        }
         res.writeHead(up.status, passthrough);
         if (!up.body) return res.end();
         const reader = up.body.getReader();