@elisym/cli 0.17.0 → 0.17.2

package/dist/index.js

@@ -406,7 +406,16 @@ function createOpenAICompatibleProvider(config) {
           },
           body: JSON.stringify({
             model,
-            max_tokens: 1,
+            // Some xAI / DeepSeek model variants are reasoning models that
+            // burn output tokens on internal chain-of-thought before any
+            // visible reply; with `max_tokens: 1` the API returns HTTP 400
+            // and we falsely flip the (provider, model) pair to unhealthy
+            // on a valid key. 256 leaves room for the minimum reasoning
+            // budget on every current variant; for non-reasoning models
+            // the probe still stops at the first natural token, so the
+            // extra ceiling does not change the per-probe cost in
+            // practice.
+            max_tokens: 256,
             messages: [{ role: "user", content: "." }]
           })
         },
@@ -693,7 +702,17 @@ async function verifyKeyDeep2(apiKey, model, signal) {
         },
         body: JSON.stringify({
           model,
-          ...reasoning ? { max_completion_tokens: 1 } : { max_tokens: 1 },
+          // GPT-5 reasoning models count internal chain-of-thought tokens
+          // against `max_completion_tokens`. With a budget of 1 the model
+          // exhausts it on reasoning before producing any visible content
+          // and the API responds HTTP 400 ("max_tokens reached") - which
+          // would falsely flip the (provider, model) pair to unhealthy on
+          // an otherwise valid key. 256 covers a "." prompt's reasoning
+          // budget on every current GPT-5 variant we tested with margin;
+          // non-reasoning models still take max_tokens: 1 since they emit
+          // visible content directly. Probe cost stays sub-cent per
+          // startup probe (reasoning tokens are billed at output rates).
+          ...reasoning ? { max_completion_tokens: 256 } : { max_tokens: 1 },
           messages: [{ role: "user", content: "." }]
         })
       },
@@ -2016,8 +2035,10 @@ function resolveHealthPair(skill) {
   return null;
 }
 var RATE_LIMIT_WINDOW_MS = 10 * 60 * 1e3;
-var MAX_JOBS_PER_CUSTOMER = 20;
-var GLOBAL_MAX_JOBS_PER_WINDOW = 200;
+var FREE_MAX_JOBS_PER_CUSTOMER = 20;
+var FREE_GLOBAL_MAX_JOBS_PER_WINDOW = 200;
+var PAID_MAX_JOBS_PER_CUSTOMER = 200;
+var PAID_GLOBAL_MAX_JOBS_PER_WINDOW = 2e3;
 var MAX_TRACKED_CUSTOMERS = 1e3;
 var GLOBAL_LIMITER_KEY = "__global__";
 var AgentRuntime = class {
@@ -2041,16 +2062,32 @@ var AgentRuntime = class {
   recoveryInterval = null;
   gcInterval = null;
   stopped = false;
-  /** Per-customer sliding-window rate limiter (keyed on customer pubkey). */
-  customerLimiter = createSlidingWindowLimiter({
+  /** Per-customer sliding-window rate limiter for free skills. */
+  freeCustomerLimiter = createSlidingWindowLimiter({
     windowMs: RATE_LIMIT_WINDOW_MS,
-    maxPerWindow: MAX_JOBS_PER_CUSTOMER,
+    maxPerWindow: FREE_MAX_JOBS_PER_CUSTOMER,
     maxKeys: MAX_TRACKED_CUSTOMERS
   });
-  /** Global sliding-window rate limiter (Sybil protection). */
-  globalLimiter = createSlidingWindowLimiter({
+  /** Global sliding-window rate limiter for free skills (Sybil protection). */
+  freeGlobalLimiter = createSlidingWindowLimiter({
     windowMs: RATE_LIMIT_WINDOW_MS,
-    maxPerWindow: GLOBAL_MAX_JOBS_PER_WINDOW,
+    maxPerWindow: FREE_GLOBAL_MAX_JOBS_PER_WINDOW,
+    maxKeys: 1
+  });
+  /**
+   * Per-customer sliding-window limiter for paid skills (10x looser than free).
+   * Payment is the primary economic deterrent; this cap exists to bound the
+   * "claim paid skill but never pay" queue-spam vector.
+   */
+  paidCustomerLimiter = createSlidingWindowLimiter({
+    windowMs: RATE_LIMIT_WINDOW_MS,
+    maxPerWindow: PAID_MAX_JOBS_PER_CUSTOMER,
+    maxKeys: MAX_TRACKED_CUSTOMERS
+  });
+  /** Global sliding-window limiter for paid skills (Sybil protection, 10x free). */
+  paidGlobalLimiter = createSlidingWindowLimiter({
+    windowMs: RATE_LIMIT_WINDOW_MS,
+    maxPerWindow: PAID_GLOBAL_MAX_JOBS_PER_WINDOW,
     maxKeys: 1
   });
   /**
@@ -2228,17 +2265,20 @@ var AgentRuntime = class {
         });
         return;
       }
-      if (!this.customerLimiter.peek(job.customerId).allowed) {
+      const matched = this.skills.route(job.tags);
+      const isPaid = matched ? matched.priceSubunits > 0 : false;
+      const customerLimiter = isPaid ? this.paidCustomerLimiter : this.freeCustomerLimiter;
+      const globalLimiter = isPaid ? this.paidGlobalLimiter : this.freeGlobalLimiter;
+      if (!customerLimiter.peek(job.customerId).allowed) {
         this.transport.sendFeedback(job, { type: "error", message: "Rate limited, try again later" }).catch(() => {
         });
         return;
       }
-      if (!this.globalLimiter.peek(GLOBAL_LIMITER_KEY).allowed) {
+      if (!globalLimiter.peek(GLOBAL_LIMITER_KEY).allowed) {
         this.transport.sendFeedback(job, { type: "error", message: "Server busy, try again later" }).catch(() => {
         });
         return;
       }
-      const matched = this.skills.route(job.tags);
       const isFreeLlm = matched?.mode === "llm" && matched.priceSubunits === 0;
       let perCustomerLimiter;
       let perSkillKey;
@@ -2259,8 +2299,8 @@ var AgentRuntime = class {
           return;
         }
       }
-      this.customerLimiter.check(job.customerId);
-      this.globalLimiter.check(GLOBAL_LIMITER_KEY);
+      customerLimiter.check(job.customerId);
+      globalLimiter.check(GLOBAL_LIMITER_KEY);
       if (isFreeLlm && perCustomerLimiter && perSkillKey) {
         this.freeLlmLimiters.globalLimiter.check(FREE_LLM_GLOBAL_KEY);
         perCustomerLimiter.check(perSkillKey);
@@ -2294,8 +2334,10 @@ var AgentRuntime = class {
   }
   /** Drop expired hits from every sliding-window limiter. */
   cleanupRateLimits() {
-    this.customerLimiter.prune();
-    this.globalLimiter.prune();
+    this.freeCustomerLimiter.prune();
+    this.freeGlobalLimiter.prune();
+    this.paidCustomerLimiter.prune();
+    this.paidGlobalLimiter.prune();
     this.freeLlmLimiters.globalLimiter.prune();
     this.freeLlmLimiters.prunePerCustomer();
   }