diff --git a/README.md b/README.md
index 6bab1f6..bb7ddaf 100644
--- a/README.md
+++ b/README.md
@@ -120,7 +120,7 @@ Restart OpenCode. The plugin will automatically load.
 │                                                      │
 │  Hooks:                                              │
 │  ┌─────────────────┐      ┌────────────────────┐     │
-│  │  chat.params    │─────▶│ Context Injection  │     │
+│  │  chat.message   │─────▶│ Context Injection  │     │
 │  │  (pre-LLM)      │      │ - Golden Rules     │     │
 │  │                 │      │ - Past Learnings   │     │
 │  │                 │      │ - Heuristics       │     │
diff --git a/package-lock.json b/package-lock.json
index 0e37efe..e049e66 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -1,12 +1,12 @@
 {
   "name": "opencode-elf",
-  "version": "0.5.3",
+  "version": "0.6.0",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "opencode-elf",
-      "version": "0.5.3",
+      "version": "0.6.0",
       "license": "MIT",
       "dependencies": {
         "@libsql/client": "^0.14.0",
diff --git a/package.json b/package.json
index cad4f94..f9fc39b 100644
--- a/package.json
+++ b/package.json
@@ -51,7 +51,7 @@
   "opencode": {
     "type": "plugin",
     "hooks": [
-      "chat.params",
+      "chat.message",
       "event"
     ]
   },
diff --git a/src/index.ts b/src/index.ts
index 6896333..34cacf7 100644
--- a/src/index.ts
+++ b/src/index.ts
@@ -75,18 +75,21 @@ export const ELFPlugin: Plugin = async ({ directory }: PluginInput) => {
   // 3. Return hooks immediately
   return {
     /**
-     * Chat params hook - Inject ELF context before the LLM processes the message
+     * Chat message hook - Inject ELF context into system prompt before LLM processing
      */
-    "chat.params": async (params: Record<string, unknown>) => {
+    "chat.message": async (input, output) => {
       const start = Date.now();
 
       try {
         // Wait for init to finish (only affects the very first message)
         await ensureReady();
 
-        const input = params.input as Record<string, unknown> | undefined;
-        const message = input?.message as Record<string, unknown> | undefined;
-        const userMessage = message?.text as string | undefined;
+        // Extract user text from message parts (TextParts have type "text" and a text field)
+        const userMessage = output.parts
+          .filter((p): p is Extract<typeof p, { type: "text" }> => p.type === "text")
+          .map(p => p.text)
+          .join("\n")
+          .trim();
 
         if (!userMessage) return;
 
@@ -110,13 +113,11 @@ export const ELFPlugin: Plugin = async ({ directory }: PluginInput) => {
           context.relevantLearnings.length > 0 ||
           context.heuristics.length > 0) {
 
-          // Inject into system prompt or prepend to user message
-          const systemPrompt = input?.systemPrompt as string | undefined;
-          if (systemPrompt && input) {
-            input.systemPrompt = `${systemPrompt}\n\n${elfMemory}`;
-          } else if (message) {
-            message.text = `${elfMemory}\n\n${userMessage}`;
-          }
+          // Inject into the system prompt via output.message.system
+          const currentSystem = output.message.system || "";
+          output.message.system = currentSystem
+            ? `${currentSystem}\n\n${elfMemory}`
+            : elfMemory;
 
           // Record metrics
           const duration = Date.now() - start;
@@ -129,7 +130,7 @@ export const ELFPlugin: Plugin = async ({ directory }: PluginInput) => {
         }
       } catch (error) {
         // Fail open: If ELF fails, log it but don't break the user's chat
-        console.error("ELF: Error in chat.params hook", error);
+        console.error("ELF: Error in chat.message hook", error);
       }
     },