Block oversized requests before providers hard-fail

The runtime already tracked rough token estimates for compaction, but provider-bound
requests still relied on naive model output limits and could be sent upstream even
when the selected model could not fit the estimated prompt plus requested output.

This adds a small model token/context registry in the API layer, estimates request
size from the serialized prompt payload, and fails locally with a dedicated
context-window error before Anthropic or xAI calls are made. Focused integration
coverage asserts the preflight fires before any HTTP request leaves the process.

Constraint: Keep the first pass minimal and reusable across both Anthropic and OpenAI-compatible providers
Rejected: Auto-compact-and-retry in the same patch | broader control-flow change than the requested minimal preflight
Confidence: medium
Scope-risk: narrow
Reversibility: clean
Directive: Expand the model registry before enabling preflight for additional providers or aliases
Tested: cargo build -p api -p tools -p rusty-claude-cli; cargo test -p api
Not-tested: End-to-end CLI auto-compaction or retry behavior after a local context_window_blocked failure

rust/crates/api/src/providers/mod.rs

@@ -1,6 +1,8 @@
 use std::future::Future;
 use std::pin::Pin;
 
+use serde::Serialize;
+
 use crate::error::ApiError;
 use crate::types::{MessageRequest, MessageResponse};
 
@@ -40,6 +42,12 @@ pub struct ProviderMetadata {
     pub default_base_url: &'static str,
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct ModelTokenLimit {
+    pub max_output_tokens: u32,
+    pub context_window_tokens: u32,
+}
+
 const MODEL_REGISTRY: &[(&str, ProviderMetadata)] = &[
     (
         "opus",
@@ -182,17 +190,86 @@ pub fn detect_provider_kind(model: &str) -> ProviderKind {
 
 #[must_use]
 pub fn max_tokens_for_model(model: &str) -> u32 {
+    model_token_limit(model).map_or_else(
+        || {
+            let canonical = resolve_model_alias(model);
+            if canonical.contains("opus") {
+                32_000
+            } else {
+                64_000
+            }
+        },
+        |limit| limit.max_output_tokens,
+    )
+}
+
+#[must_use]
+pub fn model_token_limit(model: &str) -> Option<ModelTokenLimit> {
     let canonical = resolve_model_alias(model);
-    if canonical.contains("opus") {
-        32_000
-    } else {
-        64_000
+    match canonical.as_str() {
+        "claude-opus-4-6" => Some(ModelTokenLimit {
+            max_output_tokens: 32_000,
+            context_window_tokens: 200_000,
+        }),
+        "claude-sonnet-4-6" | "claude-haiku-4-5-20251213" => Some(ModelTokenLimit {
+            max_output_tokens: 64_000,
+            context_window_tokens: 200_000,
+        }),
+        "grok-3" | "grok-3-mini" => Some(ModelTokenLimit {
+            max_output_tokens: 64_000,
+            context_window_tokens: 131_072,
+        }),
+        _ => None,
     }
 }
 
+pub fn preflight_message_request(request: &MessageRequest) -> Result<(), ApiError> {
+    let Some(limit) = model_token_limit(&request.model) else {
+        return Ok(());
+    };
+
+    let estimated_input_tokens = estimate_message_request_input_tokens(request);
+    let estimated_total_tokens = estimated_input_tokens.saturating_add(request.max_tokens);
+    if estimated_total_tokens > limit.context_window_tokens {
+        return Err(ApiError::ContextWindowExceeded {
+            model: resolve_model_alias(&request.model),
+            estimated_input_tokens,
+            requested_output_tokens: request.max_tokens,
+            estimated_total_tokens,
+            context_window_tokens: limit.context_window_tokens,
+        });
+    }
+
+    Ok(())
+}
+
+fn estimate_message_request_input_tokens(request: &MessageRequest) -> u32 {
+    let mut estimate = estimate_serialized_tokens(&request.messages);
+    estimate = estimate.saturating_add(estimate_serialized_tokens(&request.system));
+    estimate = estimate.saturating_add(estimate_serialized_tokens(&request.tools));
+    estimate = estimate.saturating_add(estimate_serialized_tokens(&request.tool_choice));
+    estimate
+}
+
+fn estimate_serialized_tokens<T: Serialize>(value: &T) -> u32 {
+    serde_json::to_vec(value)
+        .ok()
+        .map_or(0, |bytes| (bytes.len() / 4 + 1) as u32)
+}
+
 #[cfg(test)]
 mod tests {
-    use super::{detect_provider_kind, max_tokens_for_model, resolve_model_alias, ProviderKind};
+    use serde_json::json;
+
+    use crate::error::ApiError;
+    use crate::types::{
+        InputContentBlock, InputMessage, MessageRequest, ToolChoice, ToolDefinition,
+    };
+
+    use super::{
+        detect_provider_kind, max_tokens_for_model, model_token_limit, preflight_message_request,
+        resolve_model_alias, ProviderKind,
+    };
 
     #[test]
     fn resolves_grok_aliases() {
@@ -215,4 +292,86 @@ mod tests {
         assert_eq!(max_tokens_for_model("opus"), 32_000);
         assert_eq!(max_tokens_for_model("grok-3"), 64_000);
     }
+
+    #[test]
+    fn returns_context_window_metadata_for_supported_models() {
+        assert_eq!(
+            model_token_limit("claude-sonnet-4-6")
+                .expect("claude-sonnet-4-6 should be registered")
+                .context_window_tokens,
+            200_000
+        );
+        assert_eq!(
+            model_token_limit("grok-mini")
+                .expect("grok-mini should resolve to a registered model")
+                .context_window_tokens,
+            131_072
+        );
+    }
+
+    #[test]
+    fn preflight_blocks_requests_that_exceed_the_model_context_window() {
+        let request = MessageRequest {
+            model: "claude-sonnet-4-6".to_string(),
+            max_tokens: 64_000,
+            messages: vec![InputMessage {
+                role: "user".to_string(),
+                content: vec![InputContentBlock::Text {
+                    text: "x".repeat(600_000),
+                }],
+            }],
+            system: Some("Keep the answer short.".to_string()),
+            tools: Some(vec![ToolDefinition {
+                name: "weather".to_string(),
+                description: Some("Fetches weather".to_string()),
+                input_schema: json!({
+                    "type": "object",
+                    "properties": { "city": { "type": "string" } },
+                }),
+            }]),
+            tool_choice: Some(ToolChoice::Auto),
+            stream: true,
+        };
+
+        let error = preflight_message_request(&request)
+            .expect_err("oversized request should be rejected before the provider call");
+
+        match error {
+            ApiError::ContextWindowExceeded {
+                model,
+                estimated_input_tokens,
+                requested_output_tokens,
+                estimated_total_tokens,
+                context_window_tokens,
+            } => {
+                assert_eq!(model, "claude-sonnet-4-6");
+                assert!(estimated_input_tokens > 136_000);
+                assert_eq!(requested_output_tokens, 64_000);
+                assert!(estimated_total_tokens > context_window_tokens);
+                assert_eq!(context_window_tokens, 200_000);
+            }
+            other => panic!("expected context-window preflight failure, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn preflight_skips_unknown_models() {
+        let request = MessageRequest {
+            model: "unknown-model".to_string(),
+            max_tokens: 64_000,
+            messages: vec![InputMessage {
+                role: "user".to_string(),
+                content: vec![InputContentBlock::Text {
+                    text: "x".repeat(600_000),
+                }],
+            }],
+            system: None,
+            tools: None,
+            tool_choice: None,
+            stream: false,
+        };
+
+        preflight_message_request(&request)
+            .expect("models without context metadata should skip the guarded preflight");
+    }
 }