diff --git a/src/features/assistant/useEncounterAdvisor.ts b/src/features/assistant/useEncounterAdvisor.ts
index eae059e..5571e9a 100644
--- a/src/features/assistant/useEncounterAdvisor.ts
+++ b/src/features/assistant/useEncounterAdvisor.ts
@@ -108,6 +108,11 @@ export function useEncounterAdvisor(campaign: Campaign, encounter: Encounter) {
       const existing: ReinforceMonster[] = monsterCombatants.map((c) => ({ name: c.name, cr: c.cr, level: c.level }));
       const current = computeBudget(campaign.system, partyLevels, existing.map((m) => ({ cr: m.cr, level: m.level }))).difficulty;
       const target = nextTarget(campaign.system, current);
+      // When the fight is already at (or above) the hardest tier, "add creatures to
+      // reach <target>" is a contradiction — it makes reasoning models spiral and burn
+      // their whole token budget. Only ask the AI when the target is genuinely harder;
+      // otherwise fall through to the deterministic path (which eases an over-tuned fight).
+      const targetIsHarder = (ORDINAL[target.toLowerCase()] ?? 99) > (ORDINAL[current.toLowerCase()] ?? -1);
 
       // Candidates the AI may choose from: the creatures already in the fight
       // (so it can say "add another goblin") plus level-appropriate bestiary picks.
@@ -125,7 +130,7 @@ export function useEncounterAdvisor(campaign: Campaign, encounter: Encounter) {
       const candidates = [...existingCands, ...poolCands];
       if (candidates.length === 0) { setMessage('No suitable creatures found for this party.'); setState('error'); return; }
 
-      if (canUseLlm) {
+      if (canUseLlm && targetIsHarder) {
         const prompt = buildBalancePrompt(ctx, { difficulty: current, targetDifficulty: target, candidates });
         const res = await complete<BalanceSuggestion>(getLlmConfig(), { system: prompt.system, user: prompt.user, schema: balanceSuggestionSchema });
         if (res.ok && 'data' in res) {
diff --git a/src/lib/llm/client.ts b/src/lib/llm/client.ts
index e8945db..c098bb7 100644
--- a/src/lib/llm/client.ts
+++ b/src/lib/llm/client.ts
@@ -1,7 +1,10 @@
 import type { CompleteOptions, LlmConfig, LlmErrorKind, LlmResult } from './types';
 
 const DEFAULT_TIMEOUT = 30_000;
-const DEFAULT_MAX_TOKENS = 1024;
+// Generous so reasoning models (e.g. DeepSeek, whose chain-of-thought counts
+// against the completion budget) have room to finish thinking AND emit the JSON.
+// A small cap truncates them mid-reasoning, leaving an empty `content`.
+const DEFAULT_MAX_TOKENS = 4096;
 
 function err<T>(error: LlmErrorKind, message: string): LlmResult<T> {
   return { ok: false, error, message };
@@ -176,6 +179,9 @@ export async function complete<T>(cfg: LlmConfig, opts: CompleteOptions<T>): Pro
     const text = extractText(cfg.provider, json);
     if (!wantJson) return { ok: true, text };
 
+    if (text.trim() === '') {
+      return err('parse', 'The model returned an empty response — it likely ran out of output tokens before answering (raise max tokens, or the model spent its whole budget reasoning).');
+    }
     const parsed = tryParseJson(text);
     if (parsed === undefined) return err('parse', 'The model did not return valid JSON.');
     const result = opts.schema!.safeParse(parsed);