specmem-hardwicksoftware 3.7.34 → 3.7.36

package/scripts/specmem-init.cjs

@@ -1291,49 +1291,126 @@ fi
     }
   }
 
-  // ── Step 3b: Add ANTHROPIC_BASE_URL proxy routing to bashrc ─────────────
-  // Always routes Claude Code API calls through the compaction proxy.
-  // Proxy runs in passthrough mode when disabled — seamless on/off toggling.
-  // Does NOT touch ANTHROPIC_API_KEY — only sets the base URL.
-  const proxyMarker = '# specmem-proxy-env';
+  // ── Step 3b: Start compaction proxy daemon ─────────────────────────────
+  // CRITICAL: Proxy MUST be confirmed alive BEFORE setting ANTHROPIC_BASE_URL.
+  // If we set the env var pointing at a dead proxy, Claude gets ECONNREFUSED
+  // on every API call and is completely bricked. Daemon first, env var second.
   const defaultProxyPort = process.env.COMPACTION_PROXY_PORT || '4080';
-  let hasProxyFix = false;
-  try {
-    if (fs.existsSync(bashrcPath)) {
-      const currentBashrc = fs.readFileSync(bashrcPath, 'utf8');
-      hasProxyFix = currentBashrc.includes(proxyMarker);
-      // If old conditional version exists, replace it with the unconditional one
-      if (hasProxyFix && currentBashrc.includes('.compaction-proxy-port')) {
-        const cleaned = currentBashrc.replace(/\n?# specmem-proxy-env\n(?:# [^\n]*\n)*if \[ -f "\$HOME\/\.claude\/\.compaction-proxy-port" \];[\s\S]*?fi\n?/g, '');
-        fs.writeFileSync(bashrcPath, cleaned);
-        hasProxyFix = false; // Will re-add the new version below
-        initLog('Replaced old conditional proxy bashrc block with unconditional version');
+  let proxyAlive = false;
+  let proxyPort = defaultProxyPort;
+  {
+    const pidFile = path.join(os.homedir(), '.claude', '.compaction-proxy.pid');
+    const portFile = path.join(os.homedir(), '.claude', '.compaction-proxy-port');
+    // Kill any existing proxy daemon
+    try {
+      if (fs.existsSync(pidFile)) {
+        let pid;
+        try {
+          const pidData = JSON.parse(fs.readFileSync(pidFile, 'utf8'));
+          pid = pidData.pid || parseInt(pidData);
+        } catch { pid = parseInt(fs.readFileSync(pidFile, 'utf8').trim()); }
+        if (pid > 0) {
+          try { process.kill(pid, 0); process.kill(pid, 'SIGTERM'); initLog(`Killed stale proxy daemon PID ${pid}`); } catch {}
+        }
+        try { fs.unlinkSync(pidFile); } catch {}
       }
-    }
-  } catch (e) { /* no .bashrc */ }
-
-  if (!hasProxyFix) {
-    const proxyFix = `
-${proxyMarker}
-# SpecMem: Always route API calls through compaction proxy (passthrough when disabled)
-export ANTHROPIC_BASE_URL="http://127.0.0.1:\${COMPACTION_PROXY_PORT:-${defaultProxyPort}}"
-`;
+      if (fs.existsSync(portFile)) try { fs.unlinkSync(portFile); } catch {}
+    } catch {}
+    // Spawn fresh proxy daemon
     try {
-      fs.appendFileSync(bashrcPath, proxyFix);
-      result.proxyEnvFixed = true;
-      result.fixed = true;
-      initLog('Added unconditional ANTHROPIC_BASE_URL proxy routing to .bashrc');
+      const daemonPath = require.resolve('specmem-hardwicksoftware/dist/mcp/compactionProxyDaemon.js');
+      const { spawn } = require('child_process');
+      const child = spawn(process.execPath, [daemonPath], {
+        detached: true, stdio: 'ignore', env: { ...process.env, SPECMEM_DAEMON: '1' }
+      });
+      child.unref();
+      // Wait for daemon to write port file (up to 5s)
+      const portWaitFile = path.join(os.homedir(), '.claude', '.compaction-proxy-port');
+      for (let i = 0; i < 50; i++) {
+        if (fs.existsSync(portWaitFile)) {
+          const p = fs.readFileSync(portWaitFile, 'utf8').trim();
+          if (p && parseInt(p) > 0) {
+            proxyPort = p;
+            // Verify the daemon process is still alive (signal 0 = existence check)
+            try { process.kill(child.pid, 0); proxyAlive = true; } catch {}
+            initLog(`Proxy daemon port file appeared (port ${p}, PID ${child.pid}, alive=${proxyAlive})`);
+            break;
+          }
+        }
+        const { execSync: _es } = require('child_process');
+        _es('sleep 0.1', { stdio: 'ignore' });
+      }
+      if (proxyAlive) {
+        console.log(`${c.green}✓ Compaction proxy daemon started (port ${proxyPort})${c.reset}`);
+      } else {
+        console.log(`${c.yellow}⚠ Compaction proxy daemon failed to start — Claude will connect directly to Anthropic${c.reset}`);
+      }
     } catch (e) {
-      initLog('Failed to write .bashrc proxy fix', e);
+      initLog(`Failed to spawn proxy daemon: ${e.message}`);
+      console.log(`${c.yellow}⚠ Could not start proxy daemon: ${e.message}${c.reset}`);
     }
   }
 
-  // Always set for current session — proxy is always in the path
+  // ── Step 3c: Configure ANTHROPIC_BASE_URL routing ─────────────────────
+  // .bashrc: conditional on port file (safe for future shells — no routing to dead proxy)
+  // process.env: only set if proxy confirmed alive (safe for THIS session)
+  // Respects `specmem proxy enable/disable` — disabled proxy still runs in passthrough,
+  // port file still exists, so traffic still routes through it (just uncompressed).
   {
-    const proxyUrl = `http://127.0.0.1:${defaultProxyPort}`;
-    if (process.env.ANTHROPIC_BASE_URL !== proxyUrl) {
+    const proxyMarker = '# specmem-proxy-env';
+    let hasProxyFix = false;
+    let needsUpgrade = false;
+    try {
+      if (fs.existsSync(bashrcPath)) {
+        const bashrc = fs.readFileSync(bashrcPath, 'utf8');
+        hasProxyFix = bashrc.includes(proxyMarker);
+        // Detect old unconditional version (no port file check) — needs upgrade to conditional
+        if (hasProxyFix && !bashrc.includes('.compaction-proxy-port')) {
+          needsUpgrade = true;
+        }
+      }
+    } catch {}
+
+    // Replace unconditional with safe conditional version
+    if (needsUpgrade) {
+      try {
+        const bashrc = fs.readFileSync(bashrcPath, 'utf8');
+        const cleaned = bashrc.replace(/\n?# specmem-proxy-env\n(?:[^\n]+\n)+/, '\n');
+        fs.writeFileSync(bashrcPath, cleaned);
+        hasProxyFix = false;
+        initLog('Removed unconditional proxy block from .bashrc (upgrading to conditional)');
+      } catch {}
+    }
+
+    // Write conditional .bashrc block — only routes when port file exists
+    // This means: proxy running (enabled OR disabled/passthrough) → routes through proxy
+    //             proxy not running (dead/never started) → Claude talks directly to Anthropic
+    if (!hasProxyFix) {
+      const proxyFix = `
+${proxyMarker}
+# SpecMem: Route API calls through compaction proxy when it's running
+if [ -f "\$HOME/.claude/.compaction-proxy-port" ]; then
+  export ANTHROPIC_BASE_URL="http://127.0.0.1:\$(cat "\$HOME/.claude/.compaction-proxy-port")"
+fi
+`;
+      try {
+        fs.appendFileSync(bashrcPath, proxyFix);
+        result.proxyEnvFixed = true;
+        result.fixed = true;
+        initLog('Added conditional ANTHROPIC_BASE_URL proxy routing to .bashrc');
+      } catch (e) {
+        initLog('Failed to write .bashrc proxy fix', e);
+      }
+    }
+
+    // Current session: only set if proxy is confirmed alive RIGHT NOW
+    if (proxyAlive) {
+      const proxyUrl = `http://127.0.0.1:${proxyPort}`;
       process.env.ANTHROPIC_BASE_URL = proxyUrl;
       initLog(`Set ANTHROPIC_BASE_URL=${proxyUrl} for current session`);
+    } else {
+      delete process.env.ANTHROPIC_BASE_URL;
+      initLog('Proxy not alive — ANTHROPIC_BASE_URL not set, Claude connects directly to Anthropic');
     }
   }
 
@@ -7190,6 +7267,11 @@ async function launchScreenSessions(projectPath, ui) {
     return result;
   }
 
+  // Ensure xclip is available for TUI clipboard support
+  try { execSync('which xclip', { stdio: 'ignore' }); } catch {
+    try { execSync('apt-get install -y xclip 2>/dev/null', { stdio: 'ignore', timeout: 30000 }); } catch {}
+  }
+
   // Get current running screens (wipe dead ones first)
   ui.setStatus('Detecting running sessions...');
   try { execSync('screen -wipe 2>/dev/null || true', { stdio: 'ignore' }); } catch {}
@@ -7263,9 +7345,21 @@ async function launchScreenSessions(projectPath, ui) {
       // Uses screen hardcopy to tmpfs on-demand instead of continuous logging
       // -h 5000 sets scrollback buffer to 5000 lines for hardcopy capture
       const claudeBin = getClaudeBinary();
-      // Always set ANTHROPIC_BASE_URL — proxy runs in passthrough mode when disabled
-      const proxyPort = process.env.COMPACTION_PROXY_PORT || '4080';
-      const proxyEnv = `ANTHROPIC_BASE_URL="http://127.0.0.1:${proxyPort}" `;
+      // Only route through proxy if it's confirmed alive (port file exists)
+      // Prevents launching Claude pointed at a dead proxy (ECONNREFUSED on all API calls)
+      const proxyPortFile = path.join(os.homedir(), '.claude', '.compaction-proxy-port');
+      let proxyEnv = '';
+      try {
+        if (fs.existsSync(proxyPortFile)) {
+          const pPort = fs.readFileSync(proxyPortFile, 'utf8').trim();
+          if (pPort && parseInt(pPort) > 0) {
+            proxyEnv = `ANTHROPIC_BASE_URL="http://127.0.0.1:${pPort}" `;
+          }
+        }
+      } catch {}
+      if (!proxyEnv) {
+        initLog('Proxy port file not found at screen launch — Claude will connect directly to Anthropic');
+      }
       execSync(`screen -h 5000 -dmS ${claudeSession} bash -c 'cd "${projectPath}" && ${proxyEnv}SPECMEM_DASHBOARD=1 "${claudeBin}" 2>&1; exec bash'`, { stdio: 'ignore' });
       await sleep(300);
 

package/specmem/model-config.json

@@ -35,12 +35,12 @@
   },
   "resources": {
     "cpuMin": 10,
-    "cpuMax": 30,
+    "cpuMax": 50,
     "cpuCoreMin": 1,
-    "cpuCoreMax": 4,
-    "ramMinMb": 4000,
-    "ramMaxMb": 15000,
-    "updatedAt": "2026-02-24T20:01:48.721Z"
+    "cpuCoreMax": 5,
+    "ramMinMb": 6000,
+    "ramMaxMb": 19500,
+    "updatedAt": "2026-03-01T19:38:15.430Z"
   },
   "resourcePool": {
     "embedding": {
@@ -95,7 +95,7 @@
   },
   "heavyOps": {
     "enabled": true,
-    "enabledAt": "2026-02-26T14:18:50.747Z",
+    "enabledAt": "2026-03-01T13:41:50.330Z",
     "originalBatchSize": 32,
     "batchSizeMultiplier": 2,
     "throttleReduction": 0.2

package/specmem/supervisord.conf

@@ -1,6 +1,6 @@
 ; ============================================
 ; SPECMEM BRAIN CONTAINER - DYNAMIC SUPERVISORD CONFIG
-; Generated by specmem-init at 2026-02-26T20:07:11.788Z
+; Generated by specmem-init at 2026-03-02T00:36:22.804Z
 ; Thread counts from model-config.json resourcePool
 ; ============================================
 

package/svg-sections/readme-token-compaction.svg

@@ -0,0 +1,246 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 800 1020" width="800" height="1020">
+  <defs>
+    <linearGradient id="bgGrad" x1="0%" y1="0%" x2="100%" y2="100%">
+      <stop offset="0%" style="stop-color:#0d1117"/>
+      <stop offset="100%" style="stop-color:#161b22"/>
+    </linearGradient>
+    <linearGradient id="cardGrad" x1="0%" y1="0%" x2="100%" y2="100%">
+      <stop offset="0%" style="stop-color:#1c2128;stop-opacity:0.9"/>
+      <stop offset="100%" style="stop-color:#21262d;stop-opacity:0.85"/>
+    </linearGradient>
+    <linearGradient id="titleGrad" x1="0%" y1="0%" x2="100%" y2="0%">
+      <stop offset="0%" style="stop-color:#00bfff"/>
+      <stop offset="50%" style="stop-color:#a855f7"/>
+      <stop offset="100%" style="stop-color:#00bfff"/>
+    </linearGradient>
+    <linearGradient id="cyanPurple" x1="0%" y1="0%" x2="100%" y2="0%">
+      <stop offset="0%" style="stop-color:#00bfff"/>
+      <stop offset="100%" style="stop-color:#a855f7"/>
+    </linearGradient>
+    <linearGradient id="greenGrad" x1="0%" y1="0%" x2="100%" y2="0%">
+      <stop offset="0%" style="stop-color:#22c55e"/>
+      <stop offset="100%" style="stop-color:#4ade80"/>
+    </linearGradient>
+    <linearGradient id="orangeGrad" x1="0%" y1="0%" x2="100%" y2="0%">
+      <stop offset="0%" style="stop-color:#f59e0b"/>
+      <stop offset="100%" style="stop-color:#f78534"/>
+    </linearGradient>
+    <filter id="glow" x="-20%" y="-20%" width="140%" height="140%">
+      <feGaussianBlur in="SourceAlpha" stdDeviation="3" result="blur"/>
+      <feFlood flood-color="#00bfff" flood-opacity="0.3"/>
+      <feComposite in2="blur" operator="in"/>
+      <feMerge><feMergeNode/><feMergeNode in="SourceGraphic"/></feMerge>
+    </filter>
+    <filter id="glowGreen" x="-20%" y="-20%" width="140%" height="140%">
+      <feGaussianBlur in="SourceAlpha" stdDeviation="4" result="blur"/>
+      <feFlood flood-color="#22c55e" flood-opacity="0.4"/>
+      <feComposite in2="blur" operator="in"/>
+      <feMerge><feMergeNode/><feMergeNode in="SourceGraphic"/></feMerge>
+    </filter>
+  </defs>
+
+  <!-- Background -->
+  <rect width="800" height="1020" rx="16" ry="16" fill="url(#bgGrad)"/>
+  <rect width="800" height="1020" rx="16" ry="16" fill="none" stroke="url(#cyanPurple)" stroke-width="1" stroke-opacity="0.2"/>
+
+  <!-- Title Section -->
+  <text x="400" y="36" text-anchor="middle" font-family="system-ui, -apple-system, 'Segoe UI', Roboto, sans-serif" font-size="20" font-weight="700" fill="url(#titleGrad)">Token Compaction Pipeline</text>
+  <text x="400" y="56" text-anchor="middle" font-family="system-ui, sans-serif" font-size="11" fill="#6e7681">5-layer MITM proxy between Claude Code and Anthropic API. Every request gets optimized.</text>
+
+  <!-- Big stat banner -->
+  <rect x="40" y="72" width="720" height="70" rx="12" fill="url(#cardGrad)" stroke="#22c55e" stroke-width="1.5" filter="url(#glowGreen)"/>
+  <text x="200" y="103" text-anchor="middle" font-family="system-ui, sans-serif" font-size="36" font-weight="800" fill="#22c55e">~60%</text>
+  <text x="200" y="125" text-anchor="middle" font-family="system-ui, sans-serif" font-size="11" fill="#6e7681">input token reduction</text>
+  <text x="400" y="103" text-anchor="middle" font-family="system-ui, sans-serif" font-size="36" font-weight="800" fill="#00bfff">5</text>
+  <text x="400" y="125" text-anchor="middle" font-family="system-ui, sans-serif" font-size="11" fill="#6e7681">compression layers</text>
+  <text x="600" y="103" text-anchor="middle" font-family="system-ui, sans-serif" font-size="36" font-weight="800" fill="#a855f7">0ms</text>
+  <text x="600" y="125" text-anchor="middle" font-family="system-ui, sans-serif" font-size="11" fill="#6e7681">added latency (async)</text>
+
+  <!-- Pipeline flow label -->
+  <text x="400" y="170" text-anchor="middle" font-family="system-ui, sans-serif" font-size="14" font-weight="600" fill="#e6edf3">Request Flow: Claude Code  &#x2192;  Compaction Proxy  &#x2192;  Anthropic API</text>
+
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <!-- LAYER 1: Context Vault  (y=188, h=126)                -->
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <g transform="translate(30, 188)">
+    <rect width="740" height="126" rx="12" fill="url(#cardGrad)" stroke="#22c55e" stroke-width="1.5"/>
+    <rect width="740" height="3" rx="1.5" fill="url(#greenGrad)"/>
+
+    <circle cx="36" cy="30" r="14" fill="none" stroke="#22c55e" stroke-width="1.5" stroke-opacity="0.7"/>
+    <text x="36" y="35" text-anchor="middle" font-family="system-ui, sans-serif" font-size="13" font-weight="700" fill="#22c55e">1</text>
+
+    <text x="60" y="34" font-family="system-ui, sans-serif" font-size="14" font-weight="700" fill="#4ade80">&#x1F3E6; Context Vault</text>
+    <rect x="210" y="20" width="90" height="20" rx="10" fill="#22c55e" fill-opacity="0.15" stroke="#22c55e" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="255" y="34" text-anchor="middle" font-family="'SF Mono', 'Fira Code', Consolas, monospace" font-size="10" font-weight="600" fill="#4ade80">99% savings</text>
+
+    <text font-family="system-ui, sans-serif" font-size="11" fill="#8b949e">
+      <tspan x="24" y="58">Tool results larger than 5KB are auto-stashed to PostgreSQL with full-text indexing (tsvector).</tspan>
+      <tspan x="24" dy="15">Claude receives a compact receipt (~800 bytes) with preview, vocabulary, and retrieval commands.</tspan>
+      <tspan x="24" dy="15">On-demand retrieval via BM25-ranked search. 24h TTL with automatic expiry.</tspan>
+    </text>
+
+    <rect x="24" y="106" width="80" height="14" rx="3" fill="#ef4444" fill-opacity="0.2" stroke="#ef4444" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="64" y="117" text-anchor="middle" font-family="'SF Mono', monospace" font-size="8" fill="#f87171">315KB result</text>
+    <text x="118" y="117" font-family="system-ui, sans-serif" font-size="11" fill="#484f58">&#x2192;</text>
+    <rect x="138" y="106" width="68" height="14" rx="3" fill="#22c55e" fill-opacity="0.2" stroke="#22c55e" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="172" y="117" text-anchor="middle" font-family="'SF Mono', monospace" font-size="8" fill="#4ade80">800B receipt</text>
+    <text x="225" y="117" font-family="'SF Mono', monospace" font-size="8" fill="#484f58">stash_the_goods() / dig_in_the_vault()</text>
+  </g>
+
+  <!-- Connector 1→2 -->
+  <line x1="66" y1="314" x2="66" y2="330" stroke="#30363d" stroke-width="2" stroke-dasharray="3,3"/>
+  <polygon points="62,328 66,336 70,328" fill="#30363d"/>
+
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <!-- LAYER 2: Steno + YCC  (y=336, h=180)                  -->
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <g transform="translate(30, 336)">
+    <rect width="740" height="180" rx="12" fill="url(#cardGrad)" stroke="#00bfff" stroke-width="1.5"/>
+    <rect width="740" height="3" rx="1.5" fill="#00bfff"/>
+
+    <circle cx="36" cy="30" r="14" fill="none" stroke="#00bfff" stroke-width="1.5" stroke-opacity="0.7"/>
+    <text x="36" y="35" text-anchor="middle" font-family="system-ui, sans-serif" font-size="13" font-weight="700" fill="#00bfff">2</text>
+
+    <text x="60" y="34" font-family="system-ui, sans-serif" font-size="14" font-weight="700" fill="#00bfff">&#x1F916; Steno Compression</text>
+    <rect x="240" y="20" width="90" height="20" rx="10" fill="#00bfff" fill-opacity="0.12" stroke="#00bfff" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="285" y="34" text-anchor="middle" font-family="'SF Mono', 'Fira Code', Consolas, monospace" font-size="10" font-weight="600" fill="#00bfff">~40% savings</text>
+
+    <text font-family="system-ui, sans-serif" font-size="11" fill="#8b949e">
+      <tspan x="24" y="58">Strips ~80 filler words (the, a, is, are, was, were, of, to, in, for ...) that carry minimal semantic</tspan>
+      <tspan x="24" dy="15">meaning. Applies ~100 programming abbreviations: function&#x2192;fn, configuration&#x2192;cfg, database&#x2192;db,</tspan>
+      <tspan x="24" dy="15">implementation&#x2192;impl, authentication&#x2192;auth, environment&#x2192;env, repository&#x2192;repo, and more.</tspan>
+    </text>
+
+    <!-- Before/after example -->
+    <rect x="24" y="104" width="340" height="14" rx="3" fill="#1a1f26" stroke="#21262d" stroke-width="0.8"/>
+    <text x="30" y="115" font-family="'SF Mono', monospace" font-size="8" fill="#6e7681">"the function configuration is required"</text>
+    <text x="372" y="115" font-family="system-ui, sans-serif" font-size="11" fill="#484f58">&#x2192;</text>
+    <rect x="390" y="104" width="130" height="14" rx="3" fill="#00bfff" fill-opacity="0.1" stroke="#00bfff" stroke-width="0.8" stroke-opacity="0.3"/>
+    <text x="396" y="115" font-family="'SF Mono', monospace" font-size="8" fill="#58a6ff">"fn cfg reqd"</text>
+    <text x="540" y="115" font-family="system-ui, sans-serif" font-size="8" fill="#484f58">KEEP_WORDS: not, no, never, must, all</text>
+
+    <!-- ── YCC Sub-layer ── -->
+    <line x1="24" y1="128" x2="716" y2="128" stroke="#30363d" stroke-width="0.8" stroke-dasharray="4,3"/>
+
+    <!-- YC Logo -->
+    <rect x="24" y="137" width="36" height="36" rx="8" fill="#a855f7" fill-opacity="0.2" stroke="#a855f7" stroke-width="1.5" stroke-opacity="0.6"/>
+    <text x="42" y="161" text-anchor="middle" font-family="system-ui, sans-serif" font-size="16" font-weight="900" fill="#c084fc">YC</text>
+
+    <!-- Yung Cracka Compress title + badge -->
+    <text x="72" y="152" font-family="system-ui, sans-serif" font-size="13" font-weight="700" fill="#c084fc">&#x1F977;&#x1F3FB; Yung Cracka Compress</text>
+    <rect x="260" y="140" width="68" height="18" rx="9" fill="#a855f7" fill-opacity="0.12" stroke="#a855f7" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="294" y="153" text-anchor="middle" font-family="'SF Mono', monospace" font-size="9" font-weight="600" fill="#c084fc">opt-in</text>
+    <text x="340" y="153" font-family="'SF Mono', monospace" font-size="8" fill="#484f58">SLANG_ENABLED: false</text>
+
+    <!-- YCC description -->
+    <text x="72" y="172" font-family="system-ui, sans-serif" font-size="10" fill="#6e7681">DMV x Florida slang layer. 90+ entries LLMs already understand. definitely&#x2192;def, probably&#x2192;prolly, through&#x2192;thru</text>
+  </g>
+
+  <!-- Connector 2→3 -->
+  <line x1="66" y1="516" x2="66" y2="532" stroke="#30363d" stroke-width="2" stroke-dasharray="3,3"/>
+  <polygon points="62,530 66,538 70,530" fill="#30363d"/>
+
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <!-- LAYER 3: Neural MT  (y=538, h=142)                    -->
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <g transform="translate(30, 538)">
+    <rect width="740" height="142" rx="12" fill="url(#cardGrad)" stroke="#a855f7" stroke-width="1.5"/>
+    <rect width="740" height="3" rx="1.5" fill="#a855f7"/>
+
+    <circle cx="36" cy="30" r="14" fill="none" stroke="#a855f7" stroke-width="1.5" stroke-opacity="0.7"/>
+    <text x="36" y="35" text-anchor="middle" font-family="system-ui, sans-serif" font-size="13" font-weight="700" fill="#a855f7">3</text>
+
+    <text x="60" y="34" font-family="system-ui, sans-serif" font-size="14" font-weight="700" fill="#a855f7">&#x1F9D1;&#x200D;&#x1F4BB; Neural MT (EN&#x2192;ZH)</text>
+    <rect x="270" y="20" width="105" height="20" rx="10" fill="#a855f7" fill-opacity="0.12" stroke="#a855f7" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="322" y="34" text-anchor="middle" font-family="'SF Mono', 'Fira Code', Consolas, monospace" font-size="10" font-weight="600" fill="#a855f7">+30% on top</text>
+
+    <text font-family="system-ui, sans-serif" font-size="11" fill="#8b949e">
+      <tspan x="24" y="58">Steno'd tool_result text is translated EN&#x2192;ZH via local Argos Translate (air-gapped, no API calls).</tspan>
+      <tspan x="24" dy="15">Loop-back verification: ZH&#x2192;EN back-translation compared via LCS Dice + Jaccard (threshold 0.65).</tspan>
+      <tspan x="24" dy="15">Failed verifications fall back to steno-only. Passthrough vocab for untranslatable tech terms (webpack,</tspan>
+      <tspan x="24" dy="15">nginx, docker, etc). Word-level TM cache grows over time &#x2014; repeated words skip the socket entirely.</tspan>
+    </text>
+
+    <!-- Verification flow diagram -->
+    <rect x="24" y="120" width="58" height="14" rx="3" fill="#00bfff" fill-opacity="0.1" stroke="#00bfff" stroke-width="0.8" stroke-opacity="0.3"/>
+    <text x="53" y="131" text-anchor="middle" font-family="'SF Mono', monospace" font-size="8" fill="#58a6ff">steno'd</text>
+    <text x="90" y="131" font-family="system-ui, sans-serif" font-size="9" fill="#484f58">&#x2192;</text>
+    <rect x="102" y="120" width="34" height="14" rx="3" fill="#a855f7" fill-opacity="0.15" stroke="#a855f7" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="119" y="131" text-anchor="middle" font-family="'SF Mono', monospace" font-size="8" fill="#c084fc">ZH</text>
+    <text x="144" y="131" font-family="system-ui, sans-serif" font-size="9" fill="#484f58">&#x2192;</text>
+    <rect x="156" y="120" width="64" height="14" rx="3" fill="#f59e0b" fill-opacity="0.12" stroke="#f59e0b" stroke-width="0.8" stroke-opacity="0.3"/>
+    <text x="188" y="131" text-anchor="middle" font-family="'SF Mono', monospace" font-size="8" fill="#fbbf24">back-EN</text>
+    <text x="228" y="131" font-family="system-ui, sans-serif" font-size="9" fill="#484f58">&#x2192;</text>
+    <rect x="242" y="120" width="48" height="14" rx="3" fill="#22c55e" fill-opacity="0.15" stroke="#22c55e" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="266" y="131" text-anchor="middle" font-family="'SF Mono', monospace" font-size="8" fill="#4ade80">&gt;=0.65</text>
+    <text x="298" y="131" font-family="system-ui, sans-serif" font-size="9" fill="#4ade80">use ZH</text>
+    <text x="350" y="131" font-family="system-ui, sans-serif" font-size="9" fill="#484f58">|</text>
+    <rect x="364" y="120" width="48" height="14" rx="3" fill="#ef4444" fill-opacity="0.15" stroke="#ef4444" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="388" y="131" text-anchor="middle" font-family="'SF Mono', monospace" font-size="8" fill="#f87171">&lt;0.65</text>
+    <text x="420" y="131" font-family="system-ui, sans-serif" font-size="9" fill="#f87171">keep steno</text>
+    <text x="490" y="131" font-family="system-ui, sans-serif" font-size="8" fill="#484f58">TM cache + synonym learning from failures</text>
+  </g>
+
+  <!-- Connector 3→4 -->
+  <line x1="66" y1="680" x2="66" y2="696" stroke="#30363d" stroke-width="2" stroke-dasharray="3,3"/>
+  <polygon points="62,694 66,702 70,694" fill="#30363d"/>
+
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <!-- LAYER 4: Old-Message Stripping  (y=702, h=114)        -->
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <g transform="translate(30, 702)">
+    <rect width="740" height="114" rx="12" fill="url(#cardGrad)" stroke="#f59e0b" stroke-width="1.5"/>
+    <rect width="740" height="3" rx="1.5" fill="url(#orangeGrad)"/>
+
+    <circle cx="36" cy="30" r="14" fill="none" stroke="#f59e0b" stroke-width="1.5" stroke-opacity="0.7"/>
+    <text x="36" y="35" text-anchor="middle" font-family="system-ui, sans-serif" font-size="13" font-weight="700" fill="#f59e0b">4</text>
+
+    <text x="60" y="34" font-family="system-ui, sans-serif" font-size="14" font-weight="700" fill="#fbbf24">&#x1F5D1;&#xFE0F; Old-Message Stripping</text>
+    <rect x="272" y="20" width="100" height="20" rx="10" fill="#f59e0b" fill-opacity="0.12" stroke="#f59e0b" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="322" y="34" text-anchor="middle" font-family="'SF Mono', 'Fira Code', Consolas, monospace" font-size="10" font-weight="600" fill="#fbbf24">85%+ savings</text>
+
+    <text font-family="system-ui, sans-serif" font-size="11" fill="#8b949e">
+      <tspan x="24" y="58">tool_result content from messages outside the recent window (last 3 messages preserved) gets</tspan>
+      <tspan x="24" dy="15">replaced with a 200-char preview summary. Stale file reads, grep outputs, and old assistant text</tspan>
+      <tspan x="24" dy="15">blocks are trimmed to first-line anchors. SpecMem hook injections (SM-* markers) also stripped.</tspan>
+    </text>
+
+    <text x="24" y="106" font-family="'SF Mono', monospace" font-size="8" fill="#484f58">threshold: 100 chars | preview: 200 chars | preserves: tool_use_id, is_error, cache_control</text>
+  </g>
+
+  <!-- Connector 4→5 -->
+  <line x1="66" y1="816" x2="66" y2="832" stroke="#30363d" stroke-width="2" stroke-dasharray="3,3"/>
+  <polygon points="62,830 66,838 70,830" fill="#30363d"/>
+
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <!-- LAYER 5: System-Reminder Stripping  (y=838, h=114)    -->
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <g transform="translate(30, 838)">
+    <rect width="740" height="114" rx="12" fill="url(#cardGrad)" stroke="#ef4444" stroke-width="1.5"/>
+    <rect width="740" height="3" rx="1.5" fill="#ef4444"/>
+
+    <circle cx="36" cy="30" r="14" fill="none" stroke="#ef4444" stroke-width="1.5" stroke-opacity="0.7"/>
+    <text x="36" y="35" text-anchor="middle" font-family="system-ui, sans-serif" font-size="13" font-weight="700" fill="#ef4444">5</text>
+
+    <text x="60" y="34" font-family="system-ui, sans-serif" font-size="14" font-weight="700" fill="#f87171">&#x1F9F9; System-Reminder Stripping</text>
+    <rect x="298" y="20" width="100" height="20" rx="10" fill="#ef4444" fill-opacity="0.12" stroke="#ef4444" stroke-width="0.8" stroke-opacity="0.4"/>
+    <text x="348" y="34" text-anchor="middle" font-family="'SF Mono', 'Fira Code', Consolas, monospace" font-size="10" font-weight="600" fill="#f87171">variable</text>
+
+    <text font-family="system-ui, sans-serif" font-size="11" fill="#8b949e">
+      <tspan x="24" y="58">Duplicate &lt;system-reminder&gt; tags stripped from messages. First instance is preserved (sets up</tspan>
+      <tspan x="24" dy="15">context/style), all subsequent identical reminders removed. System prompt dedup tracks content</tspan>
+      <tspan x="24" dy="15">hash &#x2014; same prompt across requests is sent once, then stripped on repeat.</tspan>
+    </text>
+
+    <text x="24" y="106" font-family="'SF Mono', monospace" font-size="8" fill="#484f58">regex: /&lt;system-reminder&gt;[\s\S]*?&lt;\/system-reminder&gt;/g | keeps first, strips rest</text>
+  </g>
+
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <!-- Result bar  (y=970, h=44)                              -->
+  <!-- ═══════════════════════════════════════════════════════ -->
+  <rect x="30" y="968" width="740" height="44" rx="12" fill="#22c55e" fill-opacity="0.08" stroke="#22c55e" stroke-width="1.5" stroke-opacity="0.4"/>
+  <circle cx="66" cy="990" r="12" fill="none" stroke="#22c55e" stroke-width="2"/>
+  <path d="M60 990 L64 995 L74 984" fill="none" stroke="#4ade80" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round"/>
+  <text x="88" y="986" font-family="system-ui, sans-serif" font-size="13" font-weight="600" fill="#4ade80">Combined: ~57-60% total input token reduction.</text>
+  <text x="88" y="1003" font-family="system-ui, sans-serif" font-size="10" fill="#6e7681">All layers run on every request. Falls back gracefully &#x2014; proxy never breaks the API call. Toggle via /config.</text>
+</svg>