diff --git a/src/autonomy/llm.js b/src/autonomy/llm.js
index 50af7d8..d532629 100644
--- a/src/autonomy/llm.js
+++ b/src/autonomy/llm.js
@@ -8,7 +8,19 @@ function extractJson(text) {
   }
 }
 
-async function callCoordinator(config, prompt) {
+// OpenRouter reserves max_tokens against the key's remaining budget up front, and
+// that affordable ceiling shrinks as the balance depletes. An unbounded request is
+// refused outright, so "no cap" is not an option on a limited key -- it produces no
+// output at all rather than truncated output. This pulls the real ceiling out of the
+// refusal so we can retry just under it instead of guessing a fixed number.
+function affordableTokens(message) {
+  const match = /can only afford (\d+)/i.exec(String(message || ''));
+  if (!match) return null;
+  const affordable = Number(match[1]);
+  return Number.isFinite(affordable) && affordable > 256 ? affordable : null;
+}
+
+async function callCoordinator(config, prompt, options = {}) {
   const apiKey = String(config?.openRouter?.apiKey || '').trim();
   if (!apiKey) throw new Error('OpenRouter API key is not configured');
   const timeoutMs = Math.max(1000, Number(config?.openRouter?.timeoutMs || process.env.OPEN_ROUTER_TIMEOUT_MS) || 60000);
@@ -23,11 +35,12 @@ async function callCoordinator(config, prompt) {
       body: JSON.stringify({
         model: config.openRouter.llmModel,
         temperature: 0,
-        // Without this OpenRouter reserves the model's entire output window against
-        // the key's remaining budget and rejects the call with a 402 before it ever
-        // runs -- 131k tokens reserved to produce a few hundred. Reasoning models
-        // also spend completion tokens on reasoning, so leave real headroom.
-        max_tokens: Math.max(512, Number(config?.openRouter?.maxTokens || process.env.OPEN_ROUTER_MAX_TOKENS) || 8000),
+        // Sized against reality rather than guessed: the largest proposal this has
+        // ever produced was ~1,258 tokens carrying 12 predictions, the average is
+        // ~50. Reasoning models spend completion tokens thinking first, so this is
+        // still several times the worst case actually observed.
+        max_tokens: options.maxTokens
+          || Math.max(512, Number(config?.openRouter?.maxTokens || process.env.OPEN_ROUTER_MAX_TOKENS) || 6000),
         response_format: { type: 'json_object' },
         messages: [
           { role: 'system', content: 'You are a coordinator. Extract only evidence-backed categorical hypotheses. Never output probabilities, expected returns, confidence scores, position sizes, or trade actions.' },
@@ -43,10 +56,28 @@ async function callCoordinator(config, prompt) {
   }
   if (!response.ok) {
     const body = await response.text().catch(() => '');
+    // A 402 names the ceiling the key can currently afford. Retry once just under
+    // it rather than failing the event, but only downwards, so a shrinking budget
+    // degrades output length instead of stopping the pipeline dead.
+    const affordable = response.status === 402 ? affordableTokens(body) : null;
+    if (affordable && !options.retriedForBudget) {
+      const retryTokens = Math.floor(affordable * 0.9);
+      console.warn(`[llm] budget only affords ${affordable} tokens, retrying with max_tokens=${retryTokens}`);
+      return callCoordinator(config, prompt, { maxTokens: retryTokens, retriedForBudget: true });
+    }
     throw new Error(`coordinator request failed with ${response.status}: ${body.slice(0, 300)}`);
   }
+
   const body = await response.json();
-  return extractJson(body?.choices?.[0]?.message?.content);
+  const choice = body?.choices?.[0];
+
+  // Truncated json is worse than no json, because a partial object can occasionally
+  // still parse and quietly lose predictions. Fail loudly on the reason field rather
+  // than letting extractJson guess at a half-written response.
+  if (choice?.finish_reason === 'length') {
+    throw new Error('coordinator response was truncated by max_tokens, raise OPEN_ROUTER_MAX_TOKENS');
+  }
+  return extractJson(choice?.message?.content);
 }
 
 module.exports = { extractJson, callCoordinator };