diff --git a/apps/server/src/main/kotlin/com/correx/apps/server/Main.kt b/apps/server/src/main/kotlin/com/correx/apps/server/Main.kt
index 871552a5..d61a367a 100644
--- a/apps/server/src/main/kotlin/com/correx/apps/server/Main.kt
+++ b/apps/server/src/main/kotlin/com/correx/apps/server/Main.kt
@@ -385,10 +385,23 @@ fun main() {
     // Plan-compile gate: wrap the ExecutionPlanCompiler as a post-stage check so a compile-invalid
     // architect plan is handed back through the retry-feedback loop instead of parking the session in
     // ACTIVE (post-planning compile dead-end). Returns null on success, the compiler message on failure.
+    // Operator sampling defaults ([sampling] config) sent on every stage inference request. maxTokens=0
+    // is a placeholder; the loader/compiler pin it per-stage to the token budget via copy().
+    val stageSamplingDefaults = with(correxConfig.sampling) {
+        GenerationConfig(
+            temperature = temperature,
+            topP = topP,
+            maxTokens = 0,
+            topK = topK,
+            minP = minP,
+            repeatPenalty = repeatPenalty,
+        )
+    }
     val planCompiler = ExecutionPlanCompiler(
         artifactKindRegistry,
         toolRegistry.all().map { it.name }.toSet(),
         injectRecovery = true,
+        samplingDefaults = stageSamplingDefaults,
     )
     // Startup-load orchestration tuning from the [orchestration] config section into the kernel's
     // OrchestrationTuning. ponytail: read once at boot — a config edit needs a restart to take effect
@@ -449,7 +462,7 @@ fun main() {
         tuning = orchestrationTuning,
     )
     val workflowRegistry = FileSystemWorkflowRegistry(
-        InfrastructureModule.createWorkflowLoader(configArtifactKindsEarly),
+        InfrastructureModule.createWorkflowLoader(configArtifactKindsEarly, stageSamplingDefaults),
     )
     // Builds the router facade from a config snapshot, mapping the [router] block onto the domain
     // TalkieConfig. Reused by ConfigService's rebuild hook so router knob edits apply live (the
diff --git a/core/config/src/main/kotlin/com/correx/core/config/ConfigLoader.kt b/core/config/src/main/kotlin/com/correx/core/config/ConfigLoader.kt
index 28f12577..b76312a2 100644
--- a/core/config/src/main/kotlin/com/correx/core/config/ConfigLoader.kt
+++ b/core/config/src/main/kotlin/com/correx/core/config/ConfigLoader.kt
@@ -662,6 +662,15 @@ object ConfigLoader {
             )
         }
 
+        val samplingSection = sections["sampling"] ?: emptyMap()
+        val sampling = SamplingConfig(
+            temperature = asDouble(samplingSection["temperature"], 0.7),
+            topP = asDouble(samplingSection["top_p"], 1.0),
+            topK = samplingSection["top_k"]?.let { asInt(it) },
+            minP = samplingSection["min_p"]?.let { asDouble(it) },
+            repeatPenalty = samplingSection["repeat_penalty"]?.let { asDouble(it) },
+        )
+
         return CorrexConfig(
             server = server,
             tui = tui,
@@ -675,6 +684,7 @@ object ConfigLoader {
             project = project,
             personalization = personalization,
             orchestration = orchestration,
+            sampling = sampling,
         )
     }
 
diff --git a/core/config/src/main/kotlin/com/correx/core/config/CorrexConfig.kt b/core/config/src/main/kotlin/com/correx/core/config/CorrexConfig.kt
index 565e6b02..a53d85c6 100644
--- a/core/config/src/main/kotlin/com/correx/core/config/CorrexConfig.kt
+++ b/core/config/src/main/kotlin/com/correx/core/config/CorrexConfig.kt
@@ -16,9 +16,26 @@ data class CorrexConfig(
     val project: ProjectConfig = ProjectConfig(),
     val personalization: PersonalizationConfig = PersonalizationConfig(),
     val orchestration: OrchestrationKnobs = OrchestrationKnobs(),
+    val sampling: SamplingConfig = SamplingConfig(),
     val health: HealthConfig = HealthConfig(),
 )
 
+/**
+ * Sampling knobs sent to the llama-server on every *stage* inference request (the main agentic
+ * loop). [temperature] and [topP] default to the former hardcoded stage values. [topK], [minP] and
+ * [repeatPenalty] are null by default, meaning the request omits them and the model keeps its own
+ * default — set them to tighten a rambling local model. maxTokens is not here: it is pinned per-stage
+ * to the token budget. Talkie chat/narration have their own [GenerationSettings]/[NarrationSettings].
+ */
+@Serializable
+data class SamplingConfig(
+    val temperature: Double = 0.7,
+    val topP: Double = 1.0,
+    val topK: Int? = null,
+    val minP: Double? = null,
+    val repeatPenalty: Double? = null,
+)
+
 /**
  * Continuous health watch (observability-spec §4). When [enabled], a background monitor polls
  * the on-disk footprint every [intervalMs] and records a degraded/restored event on the edge.
diff --git a/core/config/src/main/kotlin/com/correx/core/config/CorrexConfigWriter.kt b/core/config/src/main/kotlin/com/correx/core/config/CorrexConfigWriter.kt
index 5edf05f3..e24b676e 100644
--- a/core/config/src/main/kotlin/com/correx/core/config/CorrexConfigWriter.kt
+++ b/core/config/src/main/kotlin/com/correx/core/config/CorrexConfigWriter.kt
@@ -102,6 +102,13 @@ object CorrexConfigWriter {
         b.kv("recovery_route_budget", cfg.orchestration.recoveryRouteBudget)
         b.kv("intent_route_budget", cfg.orchestration.intentRouteBudget)
 
+        b.section("sampling")
+        b.kv("temperature", cfg.sampling.temperature)
+        b.kv("top_p", cfg.sampling.topP)
+        cfg.sampling.topK?.let { b.kv("top_k", it) }
+        cfg.sampling.minP?.let { b.kv("min_p", it) }
+        cfg.sampling.repeatPenalty?.let { b.kv("repeat_penalty", it) }
+
         b.section("personalization")
         b.kv("enabled", cfg.personalization.enabled)
         b.kv("learn", cfg.personalization.learn)
diff --git a/core/inference/src/main/kotlin/com/correx/core/inference/GenerationConfig.kt b/core/inference/src/main/kotlin/com/correx/core/inference/GenerationConfig.kt
index f862be6a..f9fe6c78 100644
--- a/core/inference/src/main/kotlin/com/correx/core/inference/GenerationConfig.kt
+++ b/core/inference/src/main/kotlin/com/correx/core/inference/GenerationConfig.kt
@@ -13,4 +13,9 @@ data class GenerationConfig(
     val maxTokens: Int,
     val stopSequences: List<String> = emptyList(),
     val seed: Long? = null,          // null = non-deterministic; set for replay
+    // Sampling knobs. null = omit from the request so the provider/model keeps its own default,
+    // preserving prior behavior. Serialized only when set (top_k / min_p / repeat_penalty).
+    val topK: Int? = null,
+    val minP: Double? = null,
+    val repeatPenalty: Double? = null,
 )
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 e71a579e..48d6c846 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
@@ -3,6 +3,7 @@ package com.correx.infrastructure.inference.llama.cpp
 import com.correx.core.inference.ChatMessage
 import com.correx.core.inference.ToolCallRequest
 import com.correx.core.inference.ToolDefinition
+import kotlinx.serialization.EncodeDefault
 import kotlinx.serialization.SerialName
 import kotlinx.serialization.Serializable
 
@@ -15,6 +16,12 @@ data class ChatCompletionRequest(
     @SerialName("max_tokens") val maxTokens: Int,
     @SerialName("stop") val stopSequences: List<String> = emptyList(),
     val seed: Long? = null,
+    // EncodeDefault.NEVER overrides the class-level encodeDefaults=true so an unset (null) sampling
+    // knob is omitted from the JSON entirely, letting the model keep its own default (rather than
+    // sending "top_k": null, which llama.cpp may reject or misread).
+    @EncodeDefault(EncodeDefault.Mode.NEVER) @SerialName("top_k") val topK: Int? = null,
+    @EncodeDefault(EncodeDefault.Mode.NEVER) @SerialName("min_p") val minP: Double? = null,
+    @EncodeDefault(EncodeDefault.Mode.NEVER) @SerialName("repeat_penalty") val repeatPenalty: Double? = null,
     val stream: Boolean = false,
     val grammar: String? = null,
     val tools: List<ToolDefinition>? = null,
diff --git a/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppInferenceProvider.kt b/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppInferenceProvider.kt
index 54ba0906..1dbeb9ae 100644
--- a/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppInferenceProvider.kt
+++ b/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppInferenceProvider.kt
@@ -170,6 +170,9 @@ class LlamaCppInferenceProvider(
             maxTokens = request.generationConfig.maxTokens,
             stopSequences = request.generationConfig.stopSequences,
             seed = request.generationConfig.seed,
+            topK = request.generationConfig.topK,
+            minP = request.generationConfig.minP,
+            repeatPenalty = request.generationConfig.repeatPenalty,
             stream = false,
             grammar = grammar,
             tools = tools,
diff --git a/infrastructure/inference/llama_cpp/src/test/kotlin/com/correx/infrastructure/inference/llama/cpp/SamplingRequestSerializationTest.kt b/infrastructure/inference/llama_cpp/src/test/kotlin/com/correx/infrastructure/inference/llama/cpp/SamplingRequestSerializationTest.kt
new file mode 100644
index 00000000..2a5aead1
--- /dev/null
+++ b/infrastructure/inference/llama_cpp/src/test/kotlin/com/correx/infrastructure/inference/llama/cpp/SamplingRequestSerializationTest.kt
@@ -0,0 +1,37 @@
+package com.correx.infrastructure.inference.llama.cpp
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import kotlin.test.Test
+import kotlin.test.assertFalse
+import kotlin.test.assertTrue
+
+// Guards the EncodeDefault.NEVER behavior on the sampling knobs: with encodeDefaults=true (matching
+// the provider's Json), an unset (null) top_k/min_p/repeat_penalty must be OMITTED from the body so
+// the model keeps its own default; a set value must appear.
+class SamplingRequestSerializationTest {
+    private val json = Json { encodeDefaults = true }
+
+    @Test
+    fun `unset sampling knobs are omitted from the request body`() {
+        val body = ChatCompletionRequest(
+            model = "m", messages = emptyList(), temperature = 0.7, topP = 1.0, maxTokens = 16,
+        )
+        val out = json.encodeToString(body)
+        assertFalse("top_k" in out, out)
+        assertFalse("min_p" in out, out)
+        assertFalse("repeat_penalty" in out, out)
+    }
+
+    @Test
+    fun `set sampling knobs are serialized`() {
+        val body = ChatCompletionRequest(
+            model = "m", messages = emptyList(), temperature = 0.7, topP = 1.0, maxTokens = 16,
+            topK = 40, minP = 0.05, repeatPenalty = 1.1,
+        )
+        val out = json.encodeToString(body)
+        assertTrue("\"top_k\":40" in out, out)
+        assertTrue("\"min_p\":0.05" in out, out)
+        assertTrue("\"repeat_penalty\":1.1" in out, out)
+    }
+}
diff --git a/infrastructure/inference/openai_compat/src/main/kotlin/com/correx/infrastructure/inference/openai/OpenAiApiModels.kt b/infrastructure/inference/openai_compat/src/main/kotlin/com/correx/infrastructure/inference/openai/OpenAiApiModels.kt
index fb6cf427..9509403a 100644
--- a/infrastructure/inference/openai_compat/src/main/kotlin/com/correx/infrastructure/inference/openai/OpenAiApiModels.kt
+++ b/infrastructure/inference/openai_compat/src/main/kotlin/com/correx/infrastructure/inference/openai/OpenAiApiModels.kt
@@ -2,6 +2,7 @@ package com.correx.infrastructure.inference.openai
 
 import com.correx.core.inference.ToolCallRequest
 import com.correx.core.inference.ToolDefinition
+import kotlinx.serialization.EncodeDefault
 import kotlinx.serialization.SerialName
 import kotlinx.serialization.Serializable
 
@@ -20,6 +21,11 @@ data class OpenAiChatCompletionRequest(
     @SerialName("max_tokens") val maxTokens: Int,
     @SerialName("stop") val stopSequences: List<String>? = null,
     val seed: Long? = null,
+    // Non-standard OpenAI params accepted by local backends (vLLM, llama.cpp --api). EncodeDefault.NEVER
+    // omits them when unset so a strict endpoint never sees an unknown key unless the operator opts in.
+    @EncodeDefault(EncodeDefault.Mode.NEVER) @SerialName("top_k") val topK: Int? = null,
+    @EncodeDefault(EncodeDefault.Mode.NEVER) @SerialName("min_p") val minP: Double? = null,
+    @EncodeDefault(EncodeDefault.Mode.NEVER) @SerialName("repeat_penalty") val repeatPenalty: Double? = null,
     val stream: Boolean = false,
     val tools: List<ToolDefinition>? = null,
 )
diff --git a/infrastructure/inference/openai_compat/src/main/kotlin/com/correx/infrastructure/inference/openai/OpenAiCompatInferenceProvider.kt b/infrastructure/inference/openai_compat/src/main/kotlin/com/correx/infrastructure/inference/openai/OpenAiCompatInferenceProvider.kt
index 34c26d4a..3027392c 100644
--- a/infrastructure/inference/openai_compat/src/main/kotlin/com/correx/infrastructure/inference/openai/OpenAiCompatInferenceProvider.kt
+++ b/infrastructure/inference/openai_compat/src/main/kotlin/com/correx/infrastructure/inference/openai/OpenAiCompatInferenceProvider.kt
@@ -104,6 +104,9 @@ class OpenAiCompatInferenceProvider(
             maxTokens = request.generationConfig.maxTokens,
             stopSequences = request.generationConfig.stopSequences.ifEmpty { null },
             seed = request.generationConfig.seed,
+            topK = request.generationConfig.topK,
+            minP = request.generationConfig.minP,
+            repeatPenalty = request.generationConfig.repeatPenalty,
             stream = false,
             tools = tools,
         )
diff --git a/infrastructure/src/main/kotlin/com/correx/infrastructure/InfrastructureModule.kt b/infrastructure/src/main/kotlin/com/correx/infrastructure/InfrastructureModule.kt
index 98ea2fb8..309aac21 100644
--- a/infrastructure/src/main/kotlin/com/correx/infrastructure/InfrastructureModule.kt
+++ b/infrastructure/src/main/kotlin/com/correx/infrastructure/InfrastructureModule.kt
@@ -17,6 +17,7 @@ import com.correx.core.events.EventDispatcher
 import com.correx.core.events.stores.EventStore
 import com.correx.core.inference.CapabilityScore
 import com.correx.core.inference.Embedder
+import com.correx.core.inference.GenerationConfig
 import com.correx.core.inference.InferenceProvider
 import com.correx.core.inference.InferenceRouter
 import com.correx.core.inference.ModelCapability
@@ -210,10 +211,13 @@ object InfrastructureModule {
             ),
         )
 
-    fun createWorkflowLoader(extraKinds: List<ArtifactKind> = emptyList()): WorkflowLoader {
+    fun createWorkflowLoader(
+        extraKinds: List<ArtifactKind> = emptyList(),
+        samplingDefaults: GenerationConfig = GenerationConfig(temperature = 0.7, topP = 1.0, maxTokens = 0),
+    ): WorkflowLoader {
         val registry = DefaultArtifactKindRegistry()
         extraKinds.forEach { registry.register(it) }
-        return TomlWorkflowLoader(registry)
+        return TomlWorkflowLoader(registry, samplingDefaults)
     }
 
     fun createPromptLoader(): PromptLoader = FileSystemPromptLoader()
diff --git a/infrastructure/workflow/src/main/kotlin/com/correx/infrastructure/workflow/ExecutionPlanCompiler.kt b/infrastructure/workflow/src/main/kotlin/com/correx/infrastructure/workflow/ExecutionPlanCompiler.kt
index 68fa7198..eb60f2f3 100644
--- a/infrastructure/workflow/src/main/kotlin/com/correx/infrastructure/workflow/ExecutionPlanCompiler.kt
+++ b/infrastructure/workflow/src/main/kotlin/com/correx/infrastructure/workflow/ExecutionPlanCompiler.kt
@@ -59,12 +59,6 @@ private const val DEFAULT_STAGE_TOKEN_BUDGET = 16384
 // default, or the model is truncated (finishReason=length) mid-artifact — and a degenerating local
 // model burns the whole 2048 on garbage (e.g. a `<|channel>thought` repetition loop) before it can
 // stop. Mirror the static TomlWorkflowLoader path: pin the completion cap to the stage token budget.
-private val DEFAULT_STAGE_GENERATION = GenerationConfig(
-    temperature = 0.7,
-    topP = 1.0,
-    maxTokens = DEFAULT_STAGE_TOKEN_BUDGET,
-)
-
 class ExecutionPlanCompiler(
     private val registry: ArtifactKindRegistry,
     // Names of every registered tool. A stage that references a tool the runtime can't resolve
@@ -79,7 +73,11 @@ class ExecutionPlanCompiler(
     // retries (Vikunja #41). Off by default so the compiler's own unit tests see only plan stages;
     // the server's freestyle path (Main.kt) turns it on.
     private val injectRecovery: Boolean = false,
+    // Operator sampling defaults for freestyle-compiled stages; maxTokens pinned to the stage budget.
+    // Default reproduces the former hardcoded temperature=0.7/topP=1.0.
+    private val samplingDefaults: GenerationConfig = GenerationConfig(temperature = 0.7, topP = 1.0, maxTokens = 0),
 ) {
+    private val defaultStageGeneration = samplingDefaults.copy(maxTokens = DEFAULT_STAGE_TOKEN_BUDGET)
     private val mapper = JsonMapper.builder()
         .addModule(kotlinModule())
         .disable(DeserializationFeature.FAIL_ON_UNKNOWN_PROPERTIES)
@@ -150,7 +148,7 @@ class ExecutionPlanCompiler(
                 autoBuildGate = s.id == autoGateStageId,
                 semanticReview = s.semanticReview,
                 tokenBudget = DEFAULT_STAGE_TOKEN_BUDGET,
-                generationConfig = DEFAULT_STAGE_GENERATION,
+                generationConfig = defaultStageGeneration,
                 metadata = mapOf("promptInline" to s.prompt),
             )
         }
@@ -168,7 +166,7 @@ class ExecutionPlanCompiler(
                 StageId(RECOVERY_STAGE) to StageConfig(
                     allowedTools = knownTools.ifEmpty { setOf("file_write", "file_edit", "shell") },
                     tokenBudget = DEFAULT_STAGE_TOKEN_BUDGET,
-                    generationConfig = DEFAULT_STAGE_GENERATION,
+                    generationConfig = defaultStageGeneration,
                     metadata = mapOf("role" to "recovery", "promptInline" to RECOVERY_PROMPT),
                 )
             }
diff --git a/infrastructure/workflow/src/main/kotlin/com/correx/infrastructure/workflow/TomlWorkflowLoader.kt b/infrastructure/workflow/src/main/kotlin/com/correx/infrastructure/workflow/TomlWorkflowLoader.kt
index 4e848825..5791b129 100644
--- a/infrastructure/workflow/src/main/kotlin/com/correx/infrastructure/workflow/TomlWorkflowLoader.kt
+++ b/infrastructure/workflow/src/main/kotlin/com/correx/infrastructure/workflow/TomlWorkflowLoader.kt
@@ -76,6 +76,9 @@ private val mapper = TomlMapper.builder()
 
 class TomlWorkflowLoader(
     private val registry: ArtifactKindRegistry = DefaultArtifactKindRegistry(),
+    // Operator sampling defaults applied to every stage's inference request (maxTokens is still pinned
+    // per-stage to the token budget). Defaults reproduce the former hardcoded temperature=0.7/topP=1.0.
+    private val samplingDefaults: GenerationConfig = GenerationConfig(temperature = 0.7, topP = 1.0, maxTokens = 0),
 ) : WorkflowLoader {
     override fun load(path: Path): WorkflowGraph {
         val raw = path.readText()
@@ -116,11 +119,7 @@ class TomlWorkflowLoader(
                 // Propagate the declared token budget to the inference completion cap.
                 // Without this the StageConfig default (maxTokens=2048) is used, truncating
                 // larger artifacts (finishReason=length) → invalid JSON → validation failure.
-                generationConfig = GenerationConfig(
-                    temperature = 0.7,
-                    topP = 1.0,
-                    maxTokens = s.tokenBudget,
-                ),
+                generationConfig = samplingDefaults.copy(maxTokens = s.tokenBudget),
                 maxRetries = s.maxRetries,
                 metadata = s.toMetadata(workflowDir),
             )