Bump Anthropic max_tokens default and add per-node max_tokens override

Raise the Anthropic adapter fallback from 4096 to 16384 to prevent
truncation of large tool call JSON when the model isn't in the catalog.

Add max_tokens as a configurable DOT node attribute that flows through
SessionConfig to LLM requests, following the same pattern as
reasoning_effort. Priority: node attribute > catalog > provider default.

Ported from kilroy (danshapiro/kilroy) commits 99a5cd7 and 78fadad.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

crates/arc-agent/src/config.rs

@@ -17,6 +17,9 @@ pub struct SessionConfig {
     pub reasoning_effort: Option<String>,
     pub tool_output_limits: HashMap<String, usize>,
     pub tool_line_limits: HashMap<String, usize>,
+    /// Override the provider's default max_tokens when set.
+    /// Node-level attribute takes priority over the model catalog default.
+    pub max_tokens: Option<i64>,
     pub enable_loop_detection: bool,
     pub loop_detection_window: usize,
     pub max_subagent_depth: usize,
@@ -45,6 +48,7 @@ impl std::fmt::Debug for SessionConfig {
                 &self.default_command_timeout_ms,
             )
             .field("max_command_timeout_ms", &self.max_command_timeout_ms)
+            .field("max_tokens", &self.max_tokens)
             .field("reasoning_effort", &self.reasoning_effort)
             .field("tool_output_limits", &self.tool_output_limits)
             .field("tool_line_limits", &self.tool_line_limits)
@@ -73,6 +77,7 @@ impl Default for SessionConfig {
             max_tool_rounds_per_input: 200,
             default_command_timeout_ms: 10_000,
             max_command_timeout_ms: 600_000,
+            max_tokens: None,
             reasoning_effort: None,
             tool_output_limits: HashMap::new(),
             tool_line_limits: HashMap::new(),

crates/arc-agent/src/session.rs

@@ -622,8 +622,9 @@ impl Session {
             response_format: None,
             temperature: None,
             top_p: None,
-            max_tokens: arc_llm::catalog::get_model_info(self.provider_profile.model())
-                .and_then(|m| m.max_output),
+            max_tokens: self.config.max_tokens.or_else(||
+                arc_llm::catalog::get_model_info(self.provider_profile.model())
+                    .and_then(|m| m.max_output)),
             stop_sequences: None,
             reasoning_effort: self.config.reasoning_effort.clone(),
             metadata: None,

crates/arc-llm/src/providers/anthropic.rs

@@ -1066,7 +1066,7 @@ fn build_api_request(
     let api_request = ApiRequest {
         model: request.model.clone(),
         messages: api_messages,
-        max_tokens: request.max_tokens.unwrap_or(4096),
+        max_tokens: request.max_tokens.unwrap_or(16384),
         system: system_value,
         temperature: request.temperature,
         top_p: request.top_p,

crates/arc-workflows/src/cli/backend.rs

@@ -59,6 +59,7 @@ impl AgentBackend {
         let mut profile = self.build_profile();
 
         let config = SessionConfig {
+            max_tokens: node.max_tokens(),
             reasoning_effort: Some(node.reasoning_effort().to_string()),
             ..SessionConfig::default()
         };
@@ -131,7 +132,8 @@ impl CodergenBackend for AgentBackend {
             .map(String::from)
             .or_else(|| Some(self.provider.as_str().to_string()));
 
-        let max_tokens = arc_llm::catalog::get_model_info(model).and_then(|m| m.max_output);
+        let max_tokens = node.max_tokens().or_else(||
+            arc_llm::catalog::get_model_info(model).and_then(|m| m.max_output));
 
         let request = arc_llm::types::Request {
             model: model.to_string(),

crates/arc-workflows/src/graph/types.rs

@@ -205,7 +205,12 @@ impl Node {
         self.str_attr("llm_provider")
     }
 
-    #[must_use] 
+    #[must_use]
+    pub fn max_tokens(&self) -> Option<i64> {
+        self.int_attr("max_tokens").filter(|&v| v > 0)
+    }
+
+    #[must_use]
     pub fn reasoning_effort(&self) -> &str {
         self.str_attr("reasoning_effort").unwrap_or("high")
     }