squeezr-ai 1.11.0 → 1.11.2

package/README.md

@@ -107,6 +107,67 @@ Your provider's API (Anthropic / OpenAI / Google / Ollama)
 
 Recent content is always preserved untouched — by default the last 3 tool results are never compressed. Your CLI always has full context for what it's currently working on.
 
+### Does compression make the AI "dumber"?
+
+No — it's the opposite. Without Squeezr, long sessions hit the context window limit and the CLI **silently drops old messages entirely**. You lose them with no way to get them back.
+
+With Squeezr, old messages are **summarized, not deleted**. A 3,000-token git diff from message #15 becomes a ~150-token summary like:
+
+```
+[squeezr:a3f2c1] git diff: modified src/auth.ts — validateToken:
+added expiry logging + refreshToken call. 3 files changed.
+```
+
+The AI knows *what* you did, not every exact line. And if it needs the full original, it calls `squeezr_expand(a3f2c1)` and gets it back — losslessly.
+
+| Scenario | Message #15 at turn #100 |
+|---|---|
+| **No compression** | Probably dropped by the CLI (doesn't fit) |
+| **With Squeezr** | Summarized but present, expandable on demand |
+
+The trade-off: less detail, but more memory. Without Squeezr the AI forgets entire messages. With Squeezr it has a one-line note about every decision made — and can retrieve the full context when needed.
+
+---
+
+## Deterministic compression engine
+
+Before any AI model is involved, Squeezr runs a full deterministic compression pipeline on every tool result. This is a zero-cost, zero-latency layer that handles the most common developer outputs with specialized parsers:
+
+| Tool output | What Squeezr does | Typical savings |
+|---|---|---|
+| **git status** | Parses staged/modified/untracked, drops noise lines | 70-85% |
+| **git diff** | Extracts changed function names, strips context lines (adaptive), summarizes large diffs | 65-92% |
+| **git log** | Compacts to `hash msg (author, date)`, caps entries by pressure | 70-90% |
+| **cargo test/build/clippy** | Extracts only failures, errors, warnings | 80-95% |
+| **vitest/jest/playwright** | Extracts failed tests and assertion errors | 80-95% |
+| **tsc** | Groups errors by file, keeps only error lines | 75-90% |
+| **eslint/biome** | Compacts to file + rule + message | 70-85% |
+| **prettier** | Keeps only files-changed summary | 80-90% |
+| **next build** | Extracts errors and route summary | 75-85% |
+| **pytest** | Extracts FAILED lines and short summaries | 80-95% |
+| **npm/pnpm install** | Strips progress bars, keeps final summary | 85-90% |
+| **npm outdated** | Compact table format | 60-75% |
+| **docker ps/images/logs** | Compact output, strips timestamps | 70-80% |
+| **kubectl get/describe/logs** | Strips timestamps, compacts tables | 70-80% |
+| **gh pr/run/issue** | Strips decorations, keeps data | 65-75% |
+| **curl/wget** | Strips progress, keeps response body/headers | 60-80% |
+| **terraform plan/apply** | Extracts changes summary | 70-85% |
+| **prisma** | Compacts migration and schema output | 65-80% |
+| **Grep results** | Groups by file, caps matches per file (adaptive) | 60-80% |
+| **Read (large files)** | >500 lines: imports + signatures only. >200 lines: head + tail | 70-95% |
+| **Glob** | Compacts file listings into directory summaries | 50-70% |
+
+Additionally, on **all** outputs regardless of tool:
+- ANSI escape codes stripped
+- Progress bars and spinners removed
+- Repeated lines collapsed (`"... repeated N more times"`)
+- Duplicate stack traces deduplicated (Node.js and Python)
+- Inline JSON minified (objects >200 chars)
+- Timestamps stripped (ISO 8601, bracketed, bare time formats)
+- Excessive whitespace collapsed
+
+This engine runs in pure Node.js — microseconds per result, no API calls, no cost. It handles the bulk of the compression work. The AI layer (Haiku/GPT-4o-mini) only kicks in afterward on older messages where further summarization is needed.
+
 ---
 
 ## Supported CLIs and providers
@@ -116,12 +177,12 @@ Squeezr auto-detects which provider each request targets from the auth headers.
 | CLI | Set this env var | Compresses with | Extra keys needed |
 |---|---|---|---|
 | **Claude Code** | `ANTHROPIC_BASE_URL=http://localhost:8080` | Claude Haiku | None |
-| **Codex CLI** | `OPENAI_BASE_URL=http://localhost:8080` | GPT-4o-mini | None |
-| **Aider** (OpenAI backend) | `OPENAI_BASE_URL=http://localhost:8080` | GPT-4o-mini | None |
+| **Codex CLI** | `openai_base_url=http://localhost:8080` | GPT-4o-mini | None |
+| **Aider** (OpenAI backend) | `openai_base_url=http://localhost:8080` | GPT-4o-mini | None |
 | **Aider** (Anthropic backend) | `ANTHROPIC_BASE_URL=http://localhost:8080` | Claude Haiku | None |
-| **OpenCode** | `OPENAI_BASE_URL=http://localhost:8080` | GPT-4o-mini | None |
+| **OpenCode** | `openai_base_url=http://localhost:8080` | GPT-4o-mini | None |
 | **Gemini CLI** | `GEMINI_API_BASE_URL=http://localhost:8080` | Gemini Flash 8B | None |
-| **Ollama** (any CLI) | `OPENAI_BASE_URL=http://localhost:8080` | Local model (configurable) | None |
+| **Ollama** (any CLI) | `openai_base_url=http://localhost:8080` | Local model (configurable) | None |
 
 Squeezr extracts the API key from the request itself and reuses it for compression. Zero extra setup.
 
@@ -142,13 +203,13 @@ export ANTHROPIC_BASE_URL=http://localhost:8080        # macOS / Linux
 $env:ANTHROPIC_BASE_URL="http://localhost:8080"        # Windows PowerShell
 
 # Codex / Aider / OpenCode
-export OPENAI_BASE_URL=http://localhost:8080
+export openai_base_url=http://localhost:8080
 
 # Gemini CLI
 export GEMINI_API_BASE_URL=http://localhost:8080
 
 # Ollama
-export OPENAI_BASE_URL=http://localhost:8080
+export openai_base_url=http://localhost:8080
 ```
 
 Or use the shell installer to set up the env var permanently and register Squeezr as a login service:

package/bin/squeezr.js

@@ -47,6 +47,46 @@ function runNode(script, extraArgs = []) {
   child.on('exit', code => process.exit(code ?? 0))
 }
 
+async function startDaemon() {
+  const distIndex = path.join(ROOT, 'dist', 'index.js')
+  if (!fs.existsSync(distIndex)) {
+    console.error(`Error: ${distIndex} not found. Run 'npm run build' first.`)
+    process.exit(1)
+  }
+
+  // Check if already running
+  const port = process.env.SQUEEZR_PORT || 8080
+  const running = await new Promise(resolve => {
+    const req = http.get(`http://localhost:${port}/squeezr/health`, res => {
+      resolve(res.statusCode === 200)
+      res.destroy()
+    })
+    req.on('error', () => resolve(false))
+    req.setTimeout(2000, () => { req.destroy(); resolve(false) })
+  })
+  if (running) {
+    console.log(`Squeezr is already running on port ${port}`)
+    return
+  }
+
+  // Launch detached background process
+  const logDir = path.join(os.homedir(), '.squeezr')
+  const logFile = path.join(logDir, 'squeezr.log')
+  fs.mkdirSync(logDir, { recursive: true })
+  const logFd = fs.openSync(logFile, 'a')
+  const child = spawn(process.execPath, [distIndex], {
+    detached: true,
+    stdio: ['ignore', logFd, logFd],
+    windowsHide: true,
+    cwd: ROOT,
+    env: { ...process.env, SQUEEZR_DAEMON: '1' },
+  })
+  child.unref()
+  fs.closeSync(logFd)
+  console.log(`Squeezr started in background (pid ${child.pid})`)
+  console.log(`Logs → ${logFile}`)
+}
+
 function showLogs() {
   const logFile = path.join(os.homedir(), '.squeezr', 'squeezr.log')
   if (!fs.existsSync(logFile)) {
@@ -141,7 +181,7 @@ function setupWindows() {
   // 1. Set env vars permanently via setx (user scope, no admin needed)
   const vars = {
     ANTHROPIC_BASE_URL: 'http://localhost:8080',
-    OPENAI_BASE_URL: 'http://localhost:8080',
+    openai_base_url: 'http://localhost:8080',
     GEMINI_API_BASE_URL: 'http://localhost:8080',
   }
   for (const [key, value] of Object.entries(vars)) {
@@ -217,7 +257,7 @@ function setupUnix() {
   const shellBlock = [
     `# squeezr env vars`,
     `export ANTHROPIC_BASE_URL=http://localhost:${port}`,
-    `export OPENAI_BASE_URL=http://localhost:${port}`,
+    `export openai_base_url=http://localhost:${port}`,
     `export GEMINI_API_BASE_URL=http://localhost:${port}`,
     `# squeezr auto-heal: start proxy if not running`,
     `if ! curl -sf http://localhost:${port}/squeezr/health >/dev/null 2>&1; then`,
@@ -344,7 +384,7 @@ function setupWSL() {
   const shellBlock = [
     `# squeezr env vars`,
     `export ANTHROPIC_BASE_URL=http://localhost:${port}`,
-    `export OPENAI_BASE_URL=http://localhost:${port}`,
+    `export openai_base_url=http://localhost:${port}`,
     `export GEMINI_API_BASE_URL=http://localhost:${port}`,
     `# squeezr auto-heal: start proxy if not running`,
     `if ! curl -sf http://localhost:${port}/squeezr/health >/dev/null 2>&1; then`,
@@ -381,7 +421,7 @@ function setupWSL() {
   const setxExe = '/mnt/c/Windows/System32/setx.exe'
   const winVars = {
     ANTHROPIC_BASE_URL: 'http://localhost:8080',
-    OPENAI_BASE_URL: 'http://localhost:8080',
+    openai_base_url: 'http://localhost:8080',
     GEMINI_API_BASE_URL: 'http://localhost:8080',
   }
   if (fs.existsSync(setxExe)) {
@@ -489,7 +529,7 @@ Done!
 switch (command) {
   case undefined:
   case 'start':
-    runNode('index.js')
+    startDaemon()
     break
 
   case 'setup':

package/dist/index.js

@@ -11,9 +11,19 @@ serve({ fetch: app.fetch, port: PORT }, () => {
     console.log(`Backends: Anthropic → Haiku | OpenAI → GPT-4o-mini | Gemini → Flash-8B | Local → ${config.localCompressionModel}`);
     console.log(`Stats: http://localhost:${PORT}/squeezr/stats`);
 });
-process.on('SIGINT', () => {
-    const s = stats.summary();
-    console.log(`\n[squeezr] Session summary: ${s.requests} requests | -${s.total_saved_chars.toLocaleString()} chars (~${s.total_saved_tokens.toLocaleString()} tokens, ${s.savings_pct}% saved)`);
-    process.exit(0);
-});
+const isDaemon = !!process.env.SQUEEZR_DAEMON;
+if (isDaemon) {
+    // Daemon mode: ignore SIGINT (Ctrl+C) and SIGHUP (terminal close)
+    // Only stop via `squeezr stop` which sends SIGTERM
+    process.on('SIGINT', () => { });
+    process.on('SIGHUP', () => { });
+}
+else {
+    // Dev mode (npm run dev): allow Ctrl+C to stop
+    process.on('SIGINT', () => {
+        const s = stats.summary();
+        console.log(`\n[squeezr] Session summary: ${s.requests} requests | -${s.total_saved_chars.toLocaleString()} chars (~${s.total_saved_tokens.toLocaleString()} tokens, ${s.savings_pct}% saved)`);
+        process.exit(0);
+    });
+}
 process.on('SIGTERM', () => process.exit(0));

package/dist/server.js

@@ -236,7 +236,8 @@ app.get('/squeezr/expand/:id', (c) => {
 app.all('*', async (c) => {
     const upstream = detectUpstream(c.req.raw.headers);
     const url = new URL(c.req.url);
-    const targetUrl = `${upstream}${url.pathname}${url.search}`;
+    const targetPath = url.pathname === '/responses' ? '/v1/responses' : url.pathname;
+    const targetUrl = `${upstream}${targetPath}${url.search}`;
     const body = await c.req.arrayBuffer();
     const fwdHeaders = forwardHeaders(c.req.raw.headers);
     const resp = await fetch(targetUrl, {

package/dist/version.d.ts

@@ -1 +1 @@
-export declare const VERSION = "1.11.0";
+export declare const VERSION = "1.11.1";

package/dist/version.js

@@ -1 +1 @@
-export const VERSION = '1.11.0';
+export const VERSION = '1.11.1';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "squeezr-ai",
-  "version": "1.11.0",
+  "version": "1.11.2",
   "description": "AI proxy that compresses Claude Code, Codex, Aider, Gemini CLI and Ollama context windows to save thousands of tokens per session",
   "keywords": [
     "claude",