diff --git a/core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestrator.kt b/core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestrator.kt
index 8283d69c..96090258 100644
--- a/core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestrator.kt
+++ b/core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestrator.kt
@@ -1542,7 +1542,13 @@ abstract class SessionOrchestrator(
     protected open suspend fun estimateTokens(content: String): Int {
         val t = tokenizer
         if (t != null) {
-            return runCatching { t.countTokens(content) }.getOrElse { fallbackTokenEstimate(content) }
+            // A zero count for non-blank content means the tokenizer endpoint is lying
+            // (e.g. a malformed request silently answered with an empty token list) —
+            // trust the heuristic instead, or every budget check goes blind.
+            return runCatching { t.countTokens(content) }
+                .getOrElse { fallbackTokenEstimate(content) }
+                .takeIf { it > 0 || content.isBlank() }
+                ?: fallbackTokenEstimate(content)
         }
         return fallbackTokenEstimate(content)
     }
diff --git a/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppApiModels.kt b/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppApiModels.kt
index b970ee38..12821e0c 100644
--- a/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppApiModels.kt
+++ b/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppApiModels.kt
@@ -49,7 +49,7 @@ data class Usage(
 )
 
 @Serializable
-data class TokenizeRequest(val tokens: List<String>)
+data class TokenizeRequest(val content: String)
 
 @Serializable
 data class TokenizeResponse(val tokens: List<Int>)
diff --git a/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppTokenizer.kt b/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppTokenizer.kt
index c1ff9d1d..11daabee 100644
--- a/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppTokenizer.kt
+++ b/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppTokenizer.kt
@@ -19,7 +19,7 @@ class LlamaCppTokenizer(
         httpClient.post("$baseUrl/tokenize") {
             contentType(ContentType.Application.Json)
             accept(ContentType.Application.Json)
-            setBody(TokenizeRequest(tokens = listOf(text)))
+            setBody(TokenizeRequest(content = text))
         }.body<TokenizeResponse>().tokens.map { Token(it) }
 
     override suspend fun countTokens(text: String): Int = tokenize(text).size
diff --git a/testing/deterministic/src/test/kotlin/DefaultContextPackBuilderTest.kt b/testing/deterministic/src/test/kotlin/DefaultContextPackBuilderTest.kt
index 4f867479..5d7aac25 100644
--- a/testing/deterministic/src/test/kotlin/DefaultContextPackBuilderTest.kt
+++ b/testing/deterministic/src/test/kotlin/DefaultContextPackBuilderTest.kt
@@ -31,6 +31,36 @@ class DefaultContextPackBuilderTest {
         tokenEstimate = tokens
     )
 
+    private fun typedEntry(id: String, layer: ContextLayer, tokens: Int, sourceType: String) = ContextEntry(
+        id = ContextEntryId(id),
+        layer = layer,
+        content = "content-$id",
+        sourceType = sourceType,
+        sourceId = id,
+        tokenEstimate = tokens
+    )
+
+    @Test
+    fun `oversized tool results from a stage tool loop are trimmed to budget`() {
+        // Live repro (2026-06-11): analyst stage with three file_read results of ~5.6k/11.9k/10.7k
+        // tokens sailed through a 16384 budget untrimmed and produced a 34k-token prompt.
+        val entries = listOf(
+            typedEntry("sys", ContextLayer.L0, 3500, "systemPrompt"),
+            typedEntry("task", ContextLayer.L1, 150, "agentPrompt"),
+            typedEntry("call1", ContextLayer.L2, 40, "assistantToolCall"),
+            typedEntry("res1", ContextLayer.L2, 5600, "toolResult"),
+            typedEntry("call2", ContextLayer.L2, 40, "assistantToolCall"),
+            typedEntry("res2", ContextLayer.L2, 11900, "toolResult"),
+            typedEntry("call3", ContextLayer.L2, 40, "assistantToolCall"),
+            typedEntry("res3", ContextLayer.L2, 10700, "toolResult"),
+        )
+        val pack = builder.build(packId, sessionId, stageId, entries, TokenBudget(limit = 16384))
+        assertTrue(
+            pack.budgetUsed <= 16384,
+            "budgetUsed ${pack.budgetUsed} must not exceed the 16384 stage budget",
+        )
+    }
+
     @Test
     fun `build groups entries by layer`() {
         val entries = listOf(