diff --git a/ROADMAP.md b/ROADMAP.md
index 7653649..6def3e4 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -15,7 +15,7 @@ This release focuses on making Cass easier to interrupt, easier to audit, and sa
 
 ### Context Management
 
-- [ ] **Replace message-count trimming with safer context budgeting.** Move away from relying only on `context_message_limit`.
+- [x] **Replace message-count trimming with safer context budgeting.** Move away from relying only on `context_message_limit`.
   - Keep provider-reported token usage when available, but do not rely on returned usage as the only pre-request budgeting mechanism.
   - Preserve valid tool-call structure when compacting context; do not leave orphaned tool result messages.
   - Compact or summarize older tool outputs when needed instead of dropping records blindly.
diff --git a/docs/configuration.md b/docs/configuration.md
index f38cf1b..33815ca 100644
--- a/docs/configuration.md
+++ b/docs/configuration.md
@@ -30,7 +30,7 @@ Fields:
 - `default_provider`: optional provider id from `providers.json`. If omitted, Cass infers the provider from `default_model` when possible.
 - `default_model`: optional model id to use by default.
 - `default_access_mode`: `"read-only"` or `"full-access"`.
-- `context_message_limit`: optional number of recent non-system messages sent to the model.
+- `context_message_limit`: optional legacy upper bound for recent non-system messages. Cass primarily budgets context from model metadata (`context_length` and `max_output_tokens`), compacts older tool outputs when needed, and trims only along valid tool-call boundaries.
 - `model_tool_result_limit`: optional max bytes of tool output sent back to the model.
 - `ui_tool_result_limit`: optional max bytes of tool output shown in the UI unless full output is toggled.
 - `show_reasoning`: optional boolean, defaults to `false`. Shows provider-streamed reasoning in the transcript. Reasoning is persisted and sent back in future model context using the provider's reasoning field, such as `reasoning_content` or `reasoning`.
diff --git a/src/agent.rs b/src/agent.rs
index 47674d5..220e8dc 100644
--- a/src/agent.rs
+++ b/src/agent.rs
@@ -7,6 +7,7 @@ use crate::providers::types::ModelMessage;
 use crate::tools::{self, ToolContext, ToolRuntimeEvent};
 use anyhow::Result;
 use serde_json::Value;
+use std::collections::BTreeSet;
 use std::path::PathBuf;
 use tokio::sync::mpsc;
 
@@ -106,11 +107,7 @@ pub async fn run_turn(
             &settings.config.model,
             &allowed,
         );
-        let mut messages = build_messages(
-            &conversation.records,
-            system,
-            settings.config.context_message_limit,
-        );
+        let mut messages = build_messages(&conversation.records, system, &settings.config);
         if retrying_empty_final {
             messages.push(ModelMessage::User {
                 content: EMPTY_FINAL_RETRY_PROMPT.to_string(),
@@ -240,54 +237,453 @@ fn append_visible_assistant(
     })
 }
 
-fn build_messages(records: &[Record], system: String, limit: usize) -> Vec<ModelMessage> {
-    let mut non_system = Vec::new();
-    for r in records {
-        match r {
-            Record::User { content, .. } => non_system.push(ModelMessage::User {
-                content: content.clone(),
-            }),
-            Record::Assistant {
-                content,
-                reasoning,
-                reasoning_field,
-                tool_calls,
-                ..
-            } => non_system.push(ModelMessage::Assistant {
-                content: content.clone(),
-                reasoning: reasoning.clone(),
-                reasoning_field: reasoning_field.clone(),
-                tool_calls: tool_calls.clone(),
-            }),
-            Record::Tool {
-                tool_call_id,
-                name,
-                content,
-                ..
-            } => non_system.push(ModelMessage::Tool {
-                tool_call_id: tool_call_id.clone(),
-                name: name.clone(),
-                content: content.clone(),
-            }),
-            _ => {}
-        }
-    }
-    if non_system.len() > limit {
-        non_system = non_system.split_off(non_system.len() - limit);
-    }
+const FALLBACK_CONTEXT_TOKENS: usize = 32_000;
+const FALLBACK_OUTPUT_RESERVE_TOKENS: usize = 4_096;
+const MIN_INPUT_BUDGET_TOKENS: usize = 1_024;
+const TOOL_OUTPUT_COMPACT_CHARS: usize = 1_200;
+const TOOL_OUTPUT_TINY_CHARS: usize = 320;
+
+fn build_messages(records: &[Record], system: String, config: &Config) -> Vec<ModelMessage> {
     let mut messages = vec![ModelMessage::System { content: system }];
-    messages.extend(non_system);
-    repair_tool_message_prefix(messages)
+    messages.extend(records.iter().filter_map(record_to_model_message));
+    messages = sanitize_tool_message_structure(messages);
+
+    let budget = context_budget_tokens(config);
+    if estimate_messages_tokens(&messages) > budget {
+        compact_tool_outputs(&mut messages, budget);
+        messages = trim_to_context_budget(messages, budget);
+    }
+    messages = trim_to_message_limit(messages, config.context_message_limit.max(1));
+    sanitize_tool_message_structure(messages)
 }
 
-fn repair_tool_message_prefix(mut messages: Vec<ModelMessage>) -> Vec<ModelMessage> {
-    // If context trimming begins with tool results, remove them because OpenAI-compatible APIs
-    // require tool messages to follow an assistant tool call.
-    while matches!(messages.get(1), Some(ModelMessage::Tool { .. })) {
-        messages.remove(1);
+fn record_to_model_message(record: &Record) -> Option<ModelMessage> {
+    match record {
+        Record::User { content, .. } => Some(ModelMessage::User {
+            content: content.clone(),
+        }),
+        Record::Assistant {
+            content,
+            reasoning,
+            reasoning_field,
+            tool_calls,
+            ..
+        } => Some(ModelMessage::Assistant {
+            content: content.clone(),
+            reasoning: reasoning.clone(),
+            reasoning_field: reasoning_field.clone(),
+            tool_calls: tool_calls.clone(),
+        }),
+        Record::Tool {
+            tool_call_id,
+            name,
+            content,
+            ..
+        } => Some(ModelMessage::Tool {
+            tool_call_id: tool_call_id.clone(),
+            name: name.clone(),
+            content: content.clone(),
+        }),
+        _ => None,
+    }
+}
+
+fn context_budget_tokens(config: &Config) -> usize {
+    let context_length = config
+        .model_metadata
+        .as_ref()
+        .and_then(|m| m.context_length)
+        .and_then(|v| usize::try_from(v).ok())
+        .unwrap_or(FALLBACK_CONTEXT_TOKENS)
+        .max(MIN_INPUT_BUDGET_TOKENS * 2);
+    let output_reserve = config
+        .model_metadata
+        .as_ref()
+        .and_then(|m| m.max_output_tokens)
+        .and_then(|v| usize::try_from(v).ok())
+        .unwrap_or(FALLBACK_OUTPUT_RESERVE_TOKENS)
+        .min(context_length / 2);
+    let safety_reserve = (context_length / 50).clamp(512, 8_192);
+    context_length
+        .saturating_sub(output_reserve + safety_reserve)
+        .max(MIN_INPUT_BUDGET_TOKENS)
+}
+
+fn compact_tool_outputs(messages: &mut [ModelMessage], budget: usize) {
+    let newest_tool_idx = messages
+        .iter()
+        .rposition(|message| matches!(message, ModelMessage::Tool { .. }));
+
+    for idx in 1..messages.len() {
+        if Some(idx) == newest_tool_idx {
+            continue;
+        }
+        compact_tool_output_at(messages, idx, TOOL_OUTPUT_COMPACT_CHARS);
+        if estimate_messages_tokens(messages) <= budget {
+            return;
+        }
+    }
+
+    for idx in 1..messages.len() {
+        compact_tool_output_at(messages, idx, TOOL_OUTPUT_TINY_CHARS);
+        if estimate_messages_tokens(messages) <= budget {
+            return;
+        }
+    }
+}
+
+fn compact_tool_output_at(messages: &mut [ModelMessage], idx: usize, target_chars: usize) {
+    let Some(ModelMessage::Tool { content, .. }) = messages.get_mut(idx) else {
+        return;
+    };
+    if content.chars().count() <= target_chars {
+        return;
+    }
+    *content = compact_text(content, target_chars);
+}
+
+fn compact_text(content: &str, target_chars: usize) -> String {
+    let original_chars = content.chars().count();
+    let head_chars = (target_chars * 2 / 3).max(1);
+    let tail_chars = target_chars.saturating_sub(head_chars).max(1);
+    let head: String = content.chars().take(head_chars).collect();
+    let tail: String = content
+        .chars()
+        .rev()
+        .take(tail_chars)
+        .collect::<Vec<_>>()
+        .into_iter()
+        .rev()
+        .collect();
+    format!(
+        "[Cass compacted this tool output from {original_chars} chars to fit the model context. Head/tail excerpt follows.]\n{head}\n… omitted …\n{tail}"
+    )
+}
+
+fn trim_to_context_budget(mut messages: Vec<ModelMessage>, budget: usize) -> Vec<ModelMessage> {
+    let mut omitted = false;
+    while estimate_messages_tokens(&messages) > budget && messages.len() > 1 {
+        let prefix_end = safe_prefix_end(&messages);
+        if prefix_end <= 1 {
+            break;
+        }
+        messages.drain(1..prefix_end);
+        omitted = true;
+    }
+
+    if omitted {
+        let note = ModelMessage::System {
+            content: "Cass omitted earlier conversation messages to fit the model context budget. Included tool results still follow their matching assistant tool calls; older large tool outputs may be compacted.".to_string(),
+        };
+        messages.insert(1, note);
+        while estimate_messages_tokens(&messages) > budget && messages.len() > 2 {
+            let prefix_end = safe_prefix_end(&messages);
+            if prefix_end <= 1 {
+                break;
+            }
+            messages.drain(1..prefix_end);
+        }
+    }
+
+    messages
+}
+
+fn trim_to_message_limit(mut messages: Vec<ModelMessage>, limit: usize) -> Vec<ModelMessage> {
+    while messages.len().saturating_sub(1) > limit {
+        let prefix_end = safe_prefix_end(&messages);
+        if prefix_end <= 1 {
+            break;
+        }
+        messages.drain(1..prefix_end);
     }
     messages
 }
 
+fn safe_prefix_end(messages: &[ModelMessage]) -> usize {
+    let Some(first) = messages.get(1) else {
+        return 1;
+    };
+    match first {
+        ModelMessage::Tool { .. } => {
+            let mut end = 2;
+            while matches!(messages.get(end), Some(ModelMessage::Tool { .. })) {
+                end += 1;
+            }
+            end
+        }
+        ModelMessage::Assistant { tool_calls, .. } if !tool_calls.is_empty() => {
+            let expected: BTreeSet<&str> = tool_calls.iter().map(|call| call.id.as_str()).collect();
+            let mut seen = BTreeSet::new();
+            let mut end = 2;
+            while let Some(ModelMessage::Tool { tool_call_id, .. }) = messages.get(end) {
+                if !expected.contains(tool_call_id.as_str()) {
+                    break;
+                }
+                seen.insert(tool_call_id.as_str());
+                end += 1;
+                if seen == expected {
+                    break;
+                }
+            }
+            end
+        }
+        _ => 2,
+    }
+}
+
+fn sanitize_tool_message_structure(messages: Vec<ModelMessage>) -> Vec<ModelMessage> {
+    let mut out = Vec::with_capacity(messages.len());
+    let mut iter = messages.into_iter();
+    if let Some(system) = iter.next() {
+        out.push(system);
+    }
+    let rest: Vec<_> = iter.collect();
+    let mut idx = 0;
+    while idx < rest.len() {
+        match &rest[idx] {
+            ModelMessage::Tool { .. } => {
+                idx += 1;
+            }
+            ModelMessage::Assistant { tool_calls, .. } if !tool_calls.is_empty() => {
+                let expected: BTreeSet<&str> =
+                    tool_calls.iter().map(|call| call.id.as_str()).collect();
+                let mut seen = BTreeSet::new();
+                let mut end = idx + 1;
+                while let Some(ModelMessage::Tool { tool_call_id, .. }) = rest.get(end) {
+                    if !expected.contains(tool_call_id.as_str()) {
+                        break;
+                    }
+                    seen.insert(tool_call_id.as_str());
+                    end += 1;
+                    if seen == expected {
+                        break;
+                    }
+                }
+
+                if seen == expected {
+                    out.extend(rest[idx..end].iter().cloned());
+                } else if let ModelMessage::Assistant {
+                    content,
+                    reasoning,
+                    reasoning_field,
+                    ..
+                } = &rest[idx]
+                {
+                    if !content.trim().is_empty() || !reasoning.trim().is_empty() {
+                        out.push(ModelMessage::Assistant {
+                            content: content.clone(),
+                            reasoning: reasoning.clone(),
+                            reasoning_field: reasoning_field.clone(),
+                            tool_calls: Vec::new(),
+                        });
+                    }
+                    while end < rest.len()
+                        && matches!(rest.get(end), Some(ModelMessage::Tool { .. }))
+                    {
+                        end += 1;
+                    }
+                }
+                idx = end;
+            }
+            message => {
+                out.push(message.clone());
+                idx += 1;
+            }
+        }
+    }
+    out
+}
+
+fn estimate_messages_tokens(messages: &[ModelMessage]) -> usize {
+    messages.iter().map(estimate_message_tokens).sum()
+}
+
+fn estimate_message_tokens(message: &ModelMessage) -> usize {
+    const MESSAGE_OVERHEAD: usize = 8;
+    match message {
+        ModelMessage::System { content } | ModelMessage::User { content } => {
+            MESSAGE_OVERHEAD + estimate_text_tokens(content)
+        }
+        ModelMessage::Assistant {
+            content,
+            reasoning,
+            reasoning_field,
+            tool_calls,
+        } => {
+            MESSAGE_OVERHEAD
+                + estimate_text_tokens(content)
+                + estimate_text_tokens(reasoning)
+                + reasoning_field
+                    .as_deref()
+                    .map(estimate_text_tokens)
+                    .unwrap_or(0)
+                + tool_calls
+                    .iter()
+                    .map(|call| {
+                        estimate_text_tokens(&call.id)
+                            + estimate_text_tokens(&call.name)
+                            + estimate_text_tokens(&call.arguments.to_string())
+                            + 12
+                    })
+                    .sum::<usize>()
+        }
+        ModelMessage::Tool {
+            tool_call_id,
+            name,
+            content,
+        } => {
+            MESSAGE_OVERHEAD
+                + estimate_text_tokens(tool_call_id)
+                + estimate_text_tokens(name)
+                + estimate_text_tokens(content)
+        }
+    }
+}
+
+fn estimate_text_tokens(text: &str) -> usize {
+    text.chars().count().div_ceil(4).max(1)
+}
+
 #[allow(dead_code)]
 fn _calls(_calls: Vec<StoredToolCall>) {}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::config::default_model_definition;
+    use serde_json::json;
+
+    fn small_config(context_length: u64, max_output_tokens: u64) -> Config {
+        let mut config = Config::default();
+        let mut model = default_model_definition();
+        model.context_length = Some(context_length);
+        model.max_output_tokens = Some(max_output_tokens);
+        config.model_metadata = Some(model);
+        config
+    }
+
+    fn call(id: &str) -> StoredToolCall {
+        StoredToolCall {
+            id: id.to_string(),
+            name: "read".to_string(),
+            arguments: json!({"path":"src/main.rs"}),
+        }
+    }
+
+    fn assert_valid_tool_structure(messages: &[ModelMessage]) {
+        let mut idx = 1;
+        while idx < messages.len() {
+            match &messages[idx] {
+                ModelMessage::Tool { .. } => panic!("orphaned tool message at {idx}"),
+                ModelMessage::Assistant { tool_calls, .. } if !tool_calls.is_empty() => {
+                    let expected: BTreeSet<&str> =
+                        tool_calls.iter().map(|call| call.id.as_str()).collect();
+                    let mut seen = BTreeSet::new();
+                    idx += 1;
+                    while let Some(ModelMessage::Tool { tool_call_id, .. }) = messages.get(idx) {
+                        assert!(expected.contains(tool_call_id.as_str()));
+                        seen.insert(tool_call_id.as_str());
+                        idx += 1;
+                        if seen == expected {
+                            break;
+                        }
+                    }
+                    assert_eq!(seen, expected);
+                }
+                _ => idx += 1,
+            }
+        }
+    }
+
+    #[test]
+    fn context_budget_compacts_large_tool_outputs_before_dropping_records() {
+        let records = vec![
+            Record::User {
+                content: "please inspect this file".into(),
+                ts: now_ts(),
+            },
+            Record::Assistant {
+                content: String::new(),
+                reasoning: String::new(),
+                reasoning_field: None,
+                tool_calls: vec![call("call_1")],
+                ts: now_ts(),
+            },
+            Record::Tool {
+                tool_call_id: "call_1".into(),
+                name: "read".into(),
+                ok: true,
+                content: "x".repeat(32_000),
+                ts: now_ts(),
+            },
+            Record::Assistant {
+                content: "I inspected it.".into(),
+                reasoning: String::new(),
+                reasoning_field: None,
+                tool_calls: Vec::new(),
+                ts: now_ts(),
+            },
+        ];
+        let messages = build_messages(&records, "system".into(), &small_config(8_000, 512));
+
+        assert_valid_tool_structure(&messages);
+        assert!(messages.iter().any(|message| matches!(
+            message,
+            ModelMessage::Tool { content, .. } if content.contains("Cass compacted this tool output")
+        )));
+    }
+
+    #[test]
+    fn context_budget_trimming_does_not_leave_orphaned_tool_results() {
+        let records = vec![
+            Record::User {
+                content: "old".repeat(8_000),
+                ts: now_ts(),
+            },
+            Record::Assistant {
+                content: String::new(),
+                reasoning: String::new(),
+                reasoning_field: None,
+                tool_calls: vec![call("call_1")],
+                ts: now_ts(),
+            },
+            Record::Tool {
+                tool_call_id: "call_1".into(),
+                name: "read".into(),
+                ok: true,
+                content: "result".repeat(2_000),
+                ts: now_ts(),
+            },
+            Record::User {
+                content: "new question".into(),
+                ts: now_ts(),
+            },
+        ];
+        let messages = build_messages(&records, "system".into(), &small_config(2_000, 512));
+
+        assert_valid_tool_structure(&messages);
+        assert!(
+            matches!(messages.last(), Some(ModelMessage::User { content }) if content == "new question")
+        );
+    }
+
+    #[test]
+    fn invalid_leading_tool_results_are_removed() {
+        let messages = sanitize_tool_message_structure(vec![
+            ModelMessage::System {
+                content: "system".into(),
+            },
+            ModelMessage::Tool {
+                tool_call_id: "missing".into(),
+                name: "read".into(),
+                content: "orphan".into(),
+            },
+            ModelMessage::User {
+                content: "hello".into(),
+            },
+        ]);
+
+        assert_eq!(messages.len(), 2);
+        assert!(matches!(messages[1], ModelMessage::User { .. }));
+    }
+}