@elisym/cli 0.16.0 → 0.17.1

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: "." }]
         })
       },
@@ -2158,23 +2177,26 @@ var AgentRuntime = class {
     }
     if (skill.mode !== "llm") {
       const message = err instanceof Error ? err.message : String(err);
-      if (!scriptMessageLooksLikeBillingOrInvalid(message)) {
-        return false;
-      }
       const provider = skill.llmOverride?.provider;
       const model = skill.llmOverride?.model;
       if (!provider || !model) {
         log(
-          `${tag} Script failure looks like billing/invalid ("${message.slice(0, 120)}") but skill "${skill.name}" did not declare provider/model in SKILL.md - cannot gate future jobs.`
+          `${tag} Script "${skill.name}" failed ("${message.slice(0, 120)}") but did not declare provider/model in SKILL.md - cannot gate future jobs.`
         );
         return false;
       }
       const lower = message.toLowerCase();
-      const reason = lower.includes("credit balance") || lower.includes("billing") || lower.includes("insufficient") ? "billing" : "invalid";
+      const looksBillingOrInvalid = scriptMessageLooksLikeBillingOrInvalid(message);
+      const reason = looksBillingOrInvalid && (lower.includes("credit balance") || lower.includes("billing") || lower.includes("insufficient")) ? "billing" : "invalid";
+      const cascade = looksBillingOrInvalid;
+      const cascadeNote = cascade ? this.cascadeSuffix(provider, model) : " (no cascade)";
+      const signalNote = looksBillingOrInvalid ? `${reason} signal in stderr` : `generic exit (no billing/invalid markers, classified as ${reason}, skill-local)`;
       log(
-        `${tag} Script failure carries ${reason} signal in stderr. Marking ${provider}/${model} unhealthy${this.cascadeSuffix(provider, model)}; future jobs against this pair will be refused until recovery probe succeeds.`
+        `${tag} Script failure (${signalNote}). Marking ${provider}/${model} unhealthy${cascadeNote}; future jobs against this pair will be refused until recovery probe succeeds.`
       );
-      this.healthMonitor.markUnhealthyFromJob(provider, model, reason, message.slice(0, 200));
+      this.healthMonitor.markUnhealthyFromJob(provider, model, reason, message.slice(0, 200), {
+        cascade
+      });
       return true;
     }
     return false;