feat: massive overhaul to manage costs (#103)

### Explanation

_No response_

### Issue

Closes #102

### Attestations

- [ ] I have read and agree to the [Code of Conduct](https://docs.nhcarrigan.com/community/coc/)
- [ ] I have read and agree to the [Community Guidelines](https://docs.nhcarrigan.com/community/guide/).
- [ ] My contribution complies with the [Contributor Covenant](https://docs.nhcarrigan.com/dev/covenant/).

### Dependencies

- [ ] I have pinned the dependencies to a specific patch version.

### Style

- [ ] I have run the linter and resolved any errors.
- [ ] My pull request uses an appropriate title, matching the conventional commit standards.
- [ ] My scope of feat/fix/chore/etc. correctly matches the nature of changes in my pull request.

### Tests

- [ ] My contribution adds new code, and I have added tests to cover it.
- [ ] My contribution modifies existing code, and I have updated the tests to reflect these changes.
- [ ] All new and existing tests pass locally with my changes.
- [ ] Code coverage remains at or above the configured threshold.

### Documentation

_No response_

### Versioning

_No response_

Reviewed-on: #103
Co-authored-by: Naomi Carrigan <commits@nhcarrigan.com>
Co-committed-by: Naomi Carrigan <commits@nhcarrigan.com>

src/lib/components/StatusBar.svelte

@@ -24,6 +24,12 @@
   import SessionHistoryPanel from "./SessionHistoryPanel.svelte";
   import GitPanel from "./GitPanel.svelte";
   import ProfilePanel from "./ProfilePanel.svelte";
+  import { conversationsStore } from "$lib/stores/conversations";
+  import {
+    generateContextInjection,
+    createSummary,
+    sanitizeForJson,
+  } from "$lib/utils/conversationUtils";
 
   const DISCORD_URL = "https://chat.nhcarrigan.com";
   const DONATE_URL = "https://donate.nhcarrigan.com";
@@ -41,6 +47,7 @@
   let showSessionHistory = $state(false);
   let showGitPanel = $state(false);
   let showProfile = $state(false);
+  let isSummarising = $state(false);
   const progress = $derived($achievementProgress);
   let currentConfig: HikariConfig = $state({
     model: null,
@@ -74,6 +81,11 @@
       text_secondary: null,
       border_color: null,
     },
+    budget_enabled: false,
+    session_token_budget: null,
+    session_cost_budget: null,
+    budget_action: "warn",
+    budget_warning_threshold: 0.8,
   });
 
   let streamerModeActive = $state(false);
@@ -200,6 +212,106 @@
   function toggleAchievements() {
     onToggleAchievements();
   }
+
+  async function handleCompactConversation() {
+    const activeId = get(conversationsStore.activeConversationId);
+    if (!activeId) return;
+
+    isSummarising = true;
+
+    try {
+      const conversationContent = conversationsStore.getConversationForSummary(activeId);
+      const messageCount =
+        get(conversationsStore.activeConversation)?.terminalLines.filter(
+          (l) => l.type === "user" || l.type === "assistant"
+        ).length || 0;
+      const tokenEstimate = conversationsStore.estimateTokenCount(activeId);
+
+      // Create a summary from the conversation content (truncate if too long)
+      // Apply sanitization early to handle any problematic escape sequences
+      const sanitizedContent = sanitizeForJson(conversationContent);
+      const summaryContent =
+        sanitizedContent.length > 4000
+          ? `${sanitizedContent.slice(0, 4000)}\n\n[Truncated for length - original had ${messageCount} messages]`
+          : sanitizedContent;
+
+      // Step 1: Disconnect from Claude to reset context
+      if (connectionStatus === "connected") {
+        await invoke("stop_claude", { conversationId: activeId });
+      }
+
+      // Step 2: Clear messages and store summary
+      conversationsStore.compactWithSummary(activeId, summaryContent, messageCount, tokenEstimate);
+
+      // Step 3: Reconnect to Claude with fresh context
+      const allAllowedTools = [
+        ...(currentConfig.auto_granted_tools || []),
+        ...Array.from(get(claudeStore.grantedTools)),
+      ];
+
+      await invoke("start_claude", {
+        conversationId: activeId,
+        options: {
+          working_dir: workingDirectory || selectedDirectory,
+          model: currentConfig.model || null,
+          api_key: currentConfig.api_key || null,
+          custom_instructions: currentConfig.custom_instructions || null,
+          mcp_servers_json: currentConfig.mcp_servers_json || null,
+          allowed_tools: allAllowedTools,
+        },
+      });
+
+      // Step 4: Send the context summary to Claude as the first message
+      const contextPrompt = generateContextInjection(
+        createSummary(summaryContent, messageCount, tokenEstimate)
+      );
+
+      await invoke("send_prompt", {
+        conversationId: activeId,
+        message: contextPrompt,
+      });
+
+      claudeStore.addLine(
+        "system",
+        "Conversation compacted. Context from previous session has been provided to Claude."
+      );
+    } catch (error) {
+      console.error("Failed to compact conversation:", error);
+      claudeStore.addLine("error", `Failed to compact conversation: ${error}`);
+    } finally {
+      isSummarising = false;
+    }
+  }
+
+  async function handleStartFreshWithContext() {
+    const activeId = get(conversationsStore.activeConversationId);
+    if (!activeId) return;
+
+    const conversationContent = conversationsStore.getConversationForSummary(activeId);
+    const messageCount =
+      get(conversationsStore.activeConversation)?.terminalLines.filter(
+        (l) => l.type === "user" || l.type === "assistant"
+      ).length || 0;
+    const tokenEstimate = conversationsStore.estimateTokenCount(activeId);
+
+    const summary = createSummary(
+      `This is a continuation of a previous conversation. Here's what was discussed:\n\n${conversationContent.slice(0, 4000)}${conversationContent.length > 4000 ? "\n\n[Truncated for length...]" : ""}`,
+      messageCount,
+      tokenEstimate
+    );
+
+    const newConvId = conversationsStore.createConversation("Fresh Start");
+
+    conversationsStore.setSummary(newConvId, summary);
+
+    // Context injection is generated but the actual injection happens via the summary
+    generateContextInjection(summary);
+    claudeStore.addLine("system", "Started fresh conversation with context from previous session.");
+    claudeStore.addLine(
+      "system",
+      `Previous session had ${messageCount} messages (~${tokenEstimate.toLocaleString()} tokens).`
+    );
+  }
 </script>
 
 <div
@@ -446,7 +558,11 @@
 
     {#if showStats}
       <div class="absolute top-full right-0 mt-2 mr-4 z-50">
-        <StatsDisplay />
+        <StatsDisplay
+          onRequestSummary={handleCompactConversation}
+          onStartFreshWithContext={handleStartFreshWithContext}
+          {isSummarising}
+        />
       </div>
     {/if}
     {#if connectionStatus === "connected"}
@@ -473,7 +589,11 @@
   <!-- svelte-ignore a11y_no_static_element_interactions -->
   <div class="fixed inset-0 z-40" onclick={() => (showStats = false)}></div>
   <div class="fixed top-14 right-4 z-50">
-    <StatsDisplay />
+    <StatsDisplay
+      onRequestSummary={handleCompactConversation}
+      onStartFreshWithContext={handleStartFreshWithContext}
+      {isSummarising}
+    />
   </div>
 {/if}