feat: Add early trigger, update state logic and README.md

README.md

@@ -0,0 +1,29 @@
+# Pi Auto-Compressor
+
+A lightweight, invisible background extension for the Pi coding agent that automatically manages your context window using a Hermes-style "middle-slice" compression strategy.
+
+## How it works (Hermes-Style)
+Unlike Pi's built-in `/compact` command (which flattens your entire conversation history in the database), the Auto-Compressor runs entirely in the background during the `context` event before the LLM even sees the prompt.
+
+When your context size exceeds the soft threshold (default 50% of max context window):
+1. **Middle Slicing:** It safely carves out the "middle" of your conversation using token math, preserving the System Prompt (the head) and your most recent messages (the tail). It never splits `tool_call` and `tool_result` pairs.
+2. **Background Summarization:** It passes that middle slice to a fast/cheap LLM (like Gemini Flash) to build a structured "Context Checkpoint" (Goals, Progress, Blockers, Key Decisions).
+3. **Seamless Replacement:** It replaces the raw middle slice in the context window with the generated summary, preceded by a `[CONTEXT COMPACTION — REFERENCE ONLY]` tag.
+
+The main agent never gets bogged down by huge logs, and your API calls stay cheap, but your full history remains intact in Pi's database!
+
+## Built-in Sweeping
+The extension also continuously sweeps tool outputs in the background:
+- **Deduplication:** If a tool is called multiple times with the exact same arguments (e.g. `ls` or `cat` on the same file), it replaces older duplicate outputs with a small placeholder tombstone.
+- **Error Purging:** If a tool fails, the error stays in context for a few turns so the agent can fix it, but is then purged to keep the context clean of dead failure traces.
+
+## Commands
+You don't *need* to use any commands—the extension runs automatically. However, if you want to inspect its behavior or trigger it manually, use the `/acp` command:
+
+- `/acp` - Show stats on how many tokens have been saved, tools deduplicated, and whether a summary currently exists.
+- `/acp compress` - Force a context compression on the next turn, regardless of token thresholds.
+
+## Compatibility with `/compact`
+This extension **does not conflict** with Pi's built-in `/compact` command.
+- **`/compact`**: Destructively modifies your actual session branch in the database, squashing history into a single node.
+- **Auto-Compressor**: Ephemeral modification of the context array sent to the API. It saves tokens dynamically without destroying your local branch history.

index.ts

@@ -84,8 +84,15 @@ export default function (pi: ExtensionAPI) {
   })
 
   pi.registerCommand("acp", {
-    description: "Auto-Compressor stats",
+    description: "Auto-Compressor stats and manual trigger",
     async handler(args, ctx) {
+      const argsStr = args.trim().toLowerCase();
+      if (argsStr === "compress") {
+        state.forceCompressNext = true;
+        ctx.ui.notify("Manual compression triggered. It will run in the background on the next agent turn.", "info");
+        return;
+      }
+      
       const usage = ctx.getContextUsage ? ctx.getContextUsage() : null;
       let tokenStr = "unavailable";
       if (usage && usage.tokens !== null) {
@@ -99,7 +106,9 @@ export default function (pi: ExtensionAPI) {
         `  Pruned Tool Outputs (Deduplication/Errors): ${state.prunedToolIds.size}`,
         `  Current User Turn: ${state.currentTurn}`,
         `  Summary Exists (Has Compressed): ${state.previousSummary !== null ? "Yes" : "No"}`,
-        `  Current Context Tokens: ${tokenStr}`
+        `  Current Context Tokens: ${tokenStr}`,
+        "",
+        "Type '/acp compress' to force a compression on the next turn."
       ];
       ctx.ui.notify(lines.join("\n"), "info");
     }

pruner.ts

@@ -358,7 +358,11 @@ export async function applyPruning(
     AUTO_COMPRESS_CONFIG.minimumContextLength
   );
   
-  if (totalTokens >= thresholdTokens && msgs.length > AUTO_COMPRESS_CONFIG.protectFirstN + 4) {
+  if (
+    state.forceCompressNext || 
+    (totalTokens >= thresholdTokens && msgs.length > AUTO_COMPRESS_CONFIG.protectFirstN + 4)
+  ) {
+    state.forceCompressNext = false;
     const tailBudget = Math.floor(thresholdTokens * AUTO_COMPRESS_CONFIG.summaryTargetRatio);
     const compressStart = alignBoundaryForward(msgs, AUTO_COMPRESS_CONFIG.protectFirstN);
     const compressEnd = findTailCutByTokens(msgs, compressStart, tailBudget);

state.ts

@@ -17,6 +17,7 @@ export interface DcpState {
   totalPruneCount: number
   previousSummary: string | null
   compressionCount: number
+  forceCompressNext?: boolean
 }
 
 export function createState(): DcpState {
@@ -28,6 +29,7 @@ export function createState(): DcpState {
     totalPruneCount: 0,
     previousSummary: null,
     compressionCount: 0,
+    forceCompressNext: false,
   }
 }
 
@@ -39,6 +41,7 @@ export function resetState(state: DcpState): void {
   state.totalPruneCount = 0
   state.previousSummary = null
   state.compressionCount = 0
+  state.forceCompressNext = false
 }
 
 function sortObjectKeys(value: unknown): unknown {