feat: integrate Headroom compression - compress large analysis/code/devops contexts (≥5K tokens) via context event

extensions/smart-router/index.ts

@@ -5,11 +5,30 @@ import type {
   BeforeAgentStartEventResult,
   ExtensionCommandContext,
 } from "@earendil-works/pi-coding-agent";
+import { compress } from "headroom-ai";
 
 // Global state for manual model lock
 let isLocked = false;
 let lockedModel: any = null;
 
+// Tags that trigger Headroom context compression.
+// read / discuss / search skip compression entirely.
+const COMPRESS_TAGS = new Set([
+  "devops-low",
+  "devops-high",
+  "code-analysis-low",
+  "code-analysis-high",
+  "codewrite-low",
+  "codewrite-high",
+]);
+
+// Minimum message size (in chars) before compression activates.
+// ~5K tokens ≈ 20K characters (rough 4:1 ratio).
+const COMPRESS_MIN_CHARS = 20_000;
+
+// Current routing tag for the active turn (used by context handler)
+let currentTag: string | null = null;
+
 // Model ID mappings for routing
 const MODELS: Record<string, { provider: string; id: string }> = {
   "free-core":          { provider: "openrouter", id: "free" },
@@ -193,6 +212,9 @@ export default function (pi: ExtensionAPI) {
         pi.setThinkingLevel(thinkingLevel as any);
       }
 
+      // Store tag for compression check in context event
+      currentTag = tag;
+
       // Show routing decision in status bar
       ctx.ui.setStatus("router", `🎯 ${tag} → ${modelLabel(modelKey)}`);
 
@@ -203,4 +225,30 @@ export default function (pi: ExtensionAPI) {
       ctx.ui.setStatus("router", `⚠️ fallback ${modelLabel("free-core")}`);
     }
   });
+
+  // 4. Compress large contexts before LLM turns (Headroom)
+  pi.on("context", async (event, ctx) => {
+    // Only compress for analysis/coding/devops tags
+    if (!currentTag || !COMPRESS_TAGS.has(currentTag)) return;
+
+    // Quick size check before calling the proxy
+    const totalChars = JSON.stringify(event.messages).length;
+    if (totalChars < COMPRESS_MIN_CHARS) return;
+
+    try {
+      const result = await compress(event.messages, {
+        baseUrl: "http://localhost:8787",
+        fallback: true,
+        timeout: 15_000,
+      });
+      if (result.messages && result.messages.length > 0) {
+        const saved = ((result.tokensBefore - result.tokensAfter) / result.tokensBefore * 100).toFixed(0);
+        ctx.ui.setStatus("compression", `📦 ${saved}%`);
+        return { messages: result.messages };
+      }
+    } catch {
+      // Proxy down — pass through (fallback: true already handles transport errors)
+      ctx.ui.setStatus("compression", "⚠️ offline");
+    }
+  });
 }