Add context compaction, slash commands, and HTTP/SSE MCP transport

Features implemented:
- Context compaction: /compact command to summarize older messages and
  reclaim context space when conversations grow long
- Slash commands: User-defined markdown commands in .yo/commands/<name>.md
  with $ARGUMENTS placeholder and optional YAML frontmatter
- HTTP/SSE MCP transport: Support for HTTP and SSE transports in addition
  to stdio for Model Context Protocol servers
- force_continue: Stop hooks can now request continuation with a new prompt

New files:
- src/compact.rs: Context compaction/summarization module
- src/commands.rs: Slash commands system

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: jgarzik

src/agent.rs

@@ -2,7 +2,7 @@
 
 use crate::{
     cli::Context,
-    llm,
+    llm::{self, LlmClient},
     plan::{self, PlanPhase},
     policy::Decision,
     tools,
@@ -34,6 +34,15 @@ impl CommandStats {
     }
 }
 
+/// Result of a turn, including stats and continuation info
+#[derive(Debug, Default, Clone)]
+pub struct TurnResult {
+    pub stats: CommandStats,
+    /// If true, a Stop hook requested continuation with the given prompt
+    pub force_continue: bool,
+    pub continue_prompt: Option<String>,
+}
+
 const SYSTEM_PROMPT: &str = r#"You are an agentic coding assistant running locally.
 You can only access files via tools. All paths are relative to the project root.
 Use Glob/Grep to find files before Read. Before Edit/Write, explain what you will change.
@@ -53,12 +62,8 @@ fn verbose(ctx: &Context, message: &str) {
     }
 }
 
-pub fn run_turn(
-    ctx: &Context,
-    user_input: &str,
-    messages: &mut Vec<Value>,
-) -> Result<CommandStats> {
-    let mut stats = CommandStats::default();
+pub fn run_turn(ctx: &Context, user_input: &str, messages: &mut Vec<Value>) -> Result<TurnResult> {
+    let mut result = TurnResult::default();
     let _ = ctx.transcript.borrow_mut().user_message(user_input);
 
     messages.push(json!({
@@ -228,8 +233,8 @@ pub fn run_turn(
 
         // Track token usage from this LLM call
         if let Some(usage) = &response.usage {
-            stats.input_tokens += usage.prompt_tokens;
-            stats.output_tokens += usage.completion_tokens;
+            result.stats.input_tokens += usage.prompt_tokens;
+            result.stats.output_tokens += usage.completion_tokens;
 
             // Record cost for this operation
             let turn_number = *ctx.turn_counter.borrow();
@@ -334,7 +339,7 @@ pub fn run_turn(
             let args: Value = serde_json::from_str(&tc.function.arguments).unwrap_or(json!({}));
 
             // Count this tool use
-            stats.tool_uses += 1;
+            result.stats.tool_uses += 1;
 
             trace(
                 ctx,
@@ -434,9 +439,9 @@ pub fn run_turn(
                     }
                 } else if name == "Task" {
                     // Execute Task tool (subagent delegation)
-                    let (result, sub_stats) = tools::task::execute(args.clone(), ctx)?;
-                    stats.merge(&sub_stats);
-                    result
+                    let (task_result, sub_stats) = tools::task::execute(args.clone(), ctx)?;
+                    result.stats.merge(&sub_stats);
+                    task_result
                 } else if name.starts_with("mcp.") {
                     // Execute MCP tool
                     let start = std::time::Instant::now();
@@ -531,8 +536,28 @@ pub fn run_turn(
         }
     }
 
-    // Run Stop hooks (note: force_continue not implemented yet)
-    let _ = ctx.hooks.borrow().on_stop("end_turn", None);
+    // Run Stop hooks - may request continuation
+    let last_assistant_message = messages.iter().rev().find_map(|m| {
+        if m["role"].as_str() == Some("assistant") {
+            m["content"].as_str().map(|s| s.to_string())
+        } else {
+            None
+        }
+    });
+
+    let (force_continue, continue_prompt) = ctx
+        .hooks
+        .borrow()
+        .on_stop("end_turn", last_assistant_message.as_deref());
+
+    // If force_continue is requested, signal to caller to run another turn
+    if force_continue {
+        if let Some(prompt) = continue_prompt {
+            result.force_continue = true;
+            result.continue_prompt = Some(prompt);
+            verbose(ctx, "Stop hook requested continuation");
+        }
+    }
 
-    Ok(stats)
+    Ok(result)
 }