fix: align stale provider defaults and branch Claude request shaping on model generation

src/main/providers/ClaudeModelCapabilities.scala

@@ -0,0 +1,111 @@
+// ABOUTME: Per-generation Anthropic API capabilities (thinking mode, sampling-parameter support)
+// ABOUTME: Single place encoding which Claude models take extended thinking vs adaptive thinking + effort
+package org.nlogo.extensions.llm.providers
+
+/**
+ * Which thinking request shape a Claude model accepts.
+ *
+ * Anthropic changed the thinking API across generations, and the two shapes are
+ * mutually exclusive -- sending the wrong one is an HTTP 400:
+ *
+ *  - Extended: `thinking: {type: "enabled", budget_tokens: N}`.
+ *    Claude 4.5 and earlier. Rejected by 4.7 and later.
+ *  - Adaptive: `thinking: {type: "adaptive"}` with depth steered by
+ *    `output_config.effort`. Claude 4.6 and later; the only mode on 4.7+.
+ *
+ * Claude 4.6 accepts both; adaptive is preferred there because extended
+ * thinking is deprecated on that generation.
+ */
+enum ClaudeThinkingMode:
+  case Extended
+  case Adaptive
+
+/**
+ * Capability lookup for Anthropic model identifiers.
+ *
+ * Kept separate from ReasoningModelDetector on purpose: that object answers the
+ * cross-provider question "should thinking be on at all", while these are
+ * Anthropic request-shape details that only ClaudeProvider needs. Folding them
+ * in would grow the shared cross-provider string-matching surface.
+ *
+ * Matching is on model-name substrings because Anthropic model IDs are
+ * versioned strings and the extension accepts user-supplied and override-config
+ * model names that are not in the bundled registry.
+ */
+object ClaudeModelCapabilities {
+
+  /**
+   * Generations that predate adaptive thinking, and so must use the legacy
+   * `{type: "enabled", budget_tokens: N}` shape.
+   *
+   * Claude 3.x is included for completeness: those models are retired, but a
+   * user pinning one via override config should still get the shape their
+   * model expects rather than a guaranteed 400 from the adaptive shape.
+   */
+  private val ExtendedThinkingMarkers = Seq(
+    "claude-3-5", "claude-3-7", "claude-3-opus", "claude-3-haiku", "claude-3-sonnet",
+    "claude-opus-4-0", "claude-opus-4-1", "claude-opus-4-5",
+    "claude-sonnet-4-0", "claude-sonnet-4-5",
+    "claude-haiku-4-5",
+    "claude-opus-4-20", "claude-sonnet-4-20"
+  )
+
+  /**
+   * Generations that still ACCEPT `temperature`/`top_p`/`top_k`.
+   *
+   * Claude 4.7 and later, plus the Fable/Mythos line, reject a non-default
+   * `temperature` with a 400 on EVERY request, thinking or not — which is why
+   * the non-thinking path has to honour this too.
+   *
+   * This is deliberately an allowlist of older generations rather than a denylist
+   * of newer ones. A denylist has to enumerate every future model ID, so
+   * `claude-sonnet-4-7` or `claude-opus-4-9` would match nothing, fall through as
+   * permitted, and be sent a temperature they reject. Listing what is known to
+   * accept sampling params instead makes the unknown case default FORWARD — the
+   * same direction `thinkingMode` already defaults, so the two stay consistent.
+   */
+  private val SamplingParamsMarkers = Seq(
+    "claude-3-5", "claude-3-7", "claude-3-opus", "claude-3-haiku", "claude-3-sonnet",
+    "claude-opus-4-0", "claude-opus-4-1", "claude-opus-4-5", "claude-opus-4-6",
+    "claude-sonnet-4-0", "claude-sonnet-4-5", "claude-sonnet-4-6",
+    "claude-haiku-4-5",
+    "claude-opus-4-20", "claude-sonnet-4-20"
+  )
+
+  private def matches(model: String, markers: Seq[String]): Boolean = {
+    val m = model.toLowerCase
+    markers.exists(m.contains)
+  }
+
+  /**
+   * Thinking request shape for a model.
+   *
+   * Defaults to Adaptive for unrecognized names: new Anthropic models move
+   * forward, not back, so an unknown identifier is far more likely to be a
+   * newer adaptive-only model than a pre-4.6 one.
+   */
+  def thinkingMode(model: String): ClaudeThinkingMode =
+    if (matches(model, ExtendedThinkingMarkers)) ClaudeThinkingMode.Extended
+    else ClaudeThinkingMode.Adaptive
+
+  /**
+   * Whether a `temperature` value may be sent for this model at all.
+   *
+   * False for 4.7+ regardless of thinking state.
+   */
+  def supportsSamplingParams(model: String): Boolean =
+    matches(model, SamplingParamsMarkers)
+
+  /**
+   * Map the extension's reasoning_effort config onto Anthropic's
+   * `output_config.effort` value.
+   *
+   * The extension accepts none|low|medium|high|xhigh. Anthropic accepts
+   * low|medium|high|xhigh|max -- there is no "none", so it is treated as
+   * "no explicit effort" and the API default (high) applies.
+   */
+  def effortValue(reasoningEffort: Option[String]): Option[String] =
+    reasoningEffort.map(_.toLowerCase.trim).collect {
+      case e @ ("low" | "medium" | "high" | "xhigh" | "max") => e
+    }
+}

src/main/providers/ClaudeProvider.scala

@@ -39,14 +39,13 @@ class ClaudeProvider(implicit ec: ExecutionContext) extends BaseHttpProvider {
   }
 
   override protected def buildHeaders(apiKey: Option[String]): Map[String, String] = {
-    // Note: This reads ENABLE_THINKING from the provider's configStore, which stays in sync
-    // because LLMExtension invalidates the provider (currentProvider = None) on thinking config changes.
-    val thinkingEnabled = configStore.get(ConfigStore.ENABLE_THINKING).exists(_.toLowerCase == "true")
-    val version = if (thinkingEnabled) "2025-04-15" else "2023-06-01"
+    // 2023-06-01 is the only current Anthropic API version; thinking does not
+    // require a different one. (A previous version bumped this to "2025-04-15"
+    // when thinking was on, which is not a version Anthropic publishes.)
     Map(
       "x-api-key" -> apiKey.getOrElse(throw new IllegalStateException("API key required for Claude")),
       "content-type" -> "application/json",
-      "anthropic-version" -> version
+      "anthropic-version" -> ClaudeProvider.ApiVersion
     )
   }
 
@@ -78,28 +77,48 @@ class ClaudeProvider(implicit ec: ExecutionContext) extends BaseHttpProvider {
       baseRequest("system") = sysMsg.content
     }
 
-    if (isThinking) {
-      // Anthropic requires budget >= 1024 AND budget < max_tokens, so max_tokens must be > 1024
-      if (maxTokens <= 1024) {
-        throw new RuntimeException(
-          s"Claude thinking requires max_tokens > 1024 (current: $maxTokens). " +
-          "The thinking budget must be at least 1024 and less than max_tokens."
-        )
-      }
+    // Newer Claude generations (4.7+) reject non-default temperature/top_p/top_k
+    // with a 400 on EVERY request, thinking or not -- so this gate also applies
+    // to the non-thinking path below.
+    val allowsSampling = ClaudeModelCapabilities.supportsSamplingParams(request.model)
 
-      // Anthropic requires temperature=1.0 when thinking is enabled
-      baseRequest("temperature") = 1.0
+    if (isThinking) {
+      ClaudeModelCapabilities.thinkingMode(request.model) match {
+        case ClaudeThinkingMode.Extended =>
+          // Legacy shape: budget >= 1024 AND budget < max_tokens, so max_tokens must be > 1024
+          if (maxTokens <= 1024) {
+            throw new RuntimeException(
+              s"Claude thinking requires max_tokens > 1024 (current: $maxTokens). " +
+              "The thinking budget must be at least 1024 and less than max_tokens."
+            )
+          }
 
-      // Budget must be >= 1024 and < max_tokens
-      val budget = request.thinkingConfig.flatMap(_.budgetTokens)
-        .map(b => math.max(1024, math.min(b, maxTokens - 1)))
-        .getOrElse(math.max(1024, math.min(4096, maxTokens - 1)))
+          // These models require temperature=1.0 when thinking is enabled
+          if (allowsSampling) {
+            baseRequest("temperature") = 1.0
+          }
 
-      baseRequest("thinking") = ujson.Obj(
-        "type" -> "enabled",
-        "budget_tokens" -> budget
-      )
-    } else {
+          val budget = request.thinkingConfig.flatMap(_.budgetTokens)
+            .map(b => math.max(1024, math.min(b, maxTokens - 1)))
+            .getOrElse(math.max(1024, math.min(4096, maxTokens - 1)))
+
+          baseRequest("thinking") = ujson.Obj(
+            "type" -> "enabled",
+            "budget_tokens" -> budget
+          )
+
+        case ClaudeThinkingMode.Adaptive =>
+          // Adaptive models take no budget_tokens and no temperature; depth is
+          // steered by output_config.effort instead.
+          baseRequest("thinking") = ujson.Obj("type" -> "adaptive")
+
+          ClaudeModelCapabilities
+            .effortValue(request.thinkingConfig.flatMap(_.reasoningEffort))
+            .foreach { effort =>
+              baseRequest("output_config") = ujson.Obj("effort" -> effort)
+            }
+      }
+    } else if (allowsSampling) {
       request.temperature.foreach { temp =>
         baseRequest("temperature") = temp
       }
@@ -166,3 +185,8 @@ class ClaudeProvider(implicit ec: ExecutionContext) extends BaseHttpProvider {
     }
   }
 }
+
+object ClaudeProvider {
+  /** The only Anthropic API version currently published. */
+  val ApiVersion: String = "2023-06-01"
+}

src/main/providers/ModelRegistry.scala

@@ -27,15 +27,20 @@ object ModelRegistry {
   private var modelDirLoaded: Option[String] = None
   private var overrideLoadMessage: Option[String] = None
 
-  // Fallback config in case YAML loading fails (minimal set for stability)
-  private val FALLBACK_CONFIG: Map[String, ProviderModels] = Map(
+  // Fallback config in case YAML loading fails (minimal set for stability).
+  //
+  // Every provider's defaultModel must appear here as well as in models.yaml.
+  // The drift guard in ProviderDefaultsSpec checks descriptors against the
+  // LOADED registry, so a stale entry here survives it — which is how the
+  // retired models #62 removed elsewhere lingered in this map.
+  private[llm] val FALLBACK_CONFIG: Map[String, ProviderModels] = Map(
     "openai" -> ProviderModels(Set("gpt-4o", "gpt-4o-mini", "gpt-4", "gpt-3.5-turbo"), isCustom = false),
     "anthropic" -> ProviderModels(Set(
-      "claude-3-5-sonnet-20241022", "claude-3-5-sonnet-latest",
-      "claude-3-5-haiku-20241022", "claude-3-5-haiku-latest"
+      "claude-opus-5", "claude-sonnet-5",
+      "claude-haiku-4-5-20251001", "claude-opus-4-5-20251101"
     ), isCustom = false),
-    "gemini" -> ProviderModels(Set("gemini-1.5-pro", "gemini-1.5-flash", "gemini-2.0-flash-exp"), isCustom = false),
-    "ollama" -> ProviderModels(Set("llama3.2", "llama3.1", "mistral", "phi4"), isCustom = false),
+    "gemini" -> ProviderModels(Set("gemini-2.5-pro", "gemini-2.5-flash", "gemini-3-pro-preview"), isCustom = false),
+    "ollama" -> ProviderModels(Set("llama3.2:3b", "llama3.2:1b", "mistral", "phi4"), isCustom = false),
     "openrouter" -> ProviderModels(Set("openai/gpt-4o", "openai/gpt-4o-mini", "anthropic/claude-3.5-sonnet", "deepseek/deepseek-r1"), isCustom = false),
     "together" -> ProviderModels(Set("meta-llama/Llama-3.3-70B-Instruct-Turbo", "deepseek-ai/DeepSeek-R1", "Qwen/Qwen2.5-72B-Instruct-Turbo"), isCustom = false)
   )

src/main/providers/ProviderRegistrations.scala

@@ -48,7 +48,7 @@ object ProviderRegistrations {
       apiKeyConfigKey = "anthropic_api_key",
       baseUrlConfigKey = "anthropic_base_url",
       defaultBaseUrl = "https://api.anthropic.com/v1",
-      defaultModel = "claude-3-5-haiku-latest",
+      defaultModel = "claude-haiku-4-5-20251001",
       defaultMaxTokens = "4000",
       requiresApiKey = true,
       apiKeyPrefix = None,
@@ -76,7 +76,7 @@ object ProviderRegistrations {
       apiKeyConfigKey = "gemini_api_key",
       baseUrlConfigKey = "gemini_base_url",
       defaultBaseUrl = "https://generativelanguage.googleapis.com/v1beta",
-      defaultModel = "gemini-1.5-flash",
+      defaultModel = "gemini-2.5-flash",
       defaultMaxTokens = "2048",
       requiresApiKey = true,
       apiKeyPrefix = None,
@@ -104,7 +104,7 @@ object ProviderRegistrations {
       apiKeyConfigKey = "ollama_api_key",
       baseUrlConfigKey = "ollama_base_url",
       defaultBaseUrl = "http://localhost:11434",
-      defaultModel = "llama3.2",
+      defaultModel = "llama3.2:3b",
       defaultMaxTokens = "2048",
       requiresApiKey = false,
       apiKeyPrefix = None,
@@ -122,7 +122,7 @@ object ProviderRegistrations {
           |   - Or start Ollama app (it runs in background)
           |
           |3. Pull a model:
-          |   - Run: ollama pull llama3.2
+          |   - Run: ollama pull llama3.2:3b
           |   - Or try: ollama pull deepseek-r1:1.5b (smaller)
           |
           |4. Verify installation:

src/main/resources/config/models.yaml

@@ -37,20 +37,29 @@ openai:
   - gpt-3.5-turbo
 
 anthropic:
-  # Claude 4.5 (latest production)
+  # Adaptive thinking (thinking.type "adaptive" + output_config.effort).
+  # These reject thinking.type "enabled" and any non-default temperature.
+  # IDs are dateless from the 4.6 generation on, and still pinned snapshots.
+  - claude-fable-5
+  - claude-opus-5
+  - claude-sonnet-5
+  - claude-opus-4-8
+  - claude-opus-4-7
+
+  # Adaptive thinking, extended thinking deprecated but still accepted
+  - claude-opus-4-6
+  - claude-sonnet-4-6
+
+  # Extended thinking only (thinking.type "enabled" + budget_tokens).
+  # These reject "adaptive".
   - claude-opus-4-5-20251101
   - claude-sonnet-4-5-20250929
   - claude-haiku-4-5-20251001
-  
-  # Claude 4.1 (current)
+
+  # Claude 4.1 / 4 (extended thinking)
   - claude-opus-4-1-20250805
-
-  # Claude 4 (stable)
   - claude-sonnet-4-20250514
   - claude-opus-4-20250514
-
-  # Claude 3.7 (deprecated, retiring Feb 19, 2026)
-  - claude-3-7-sonnet-20250219  # deprecated
 
 gemini:
   # Gemini 3 (latest preview)