mobygate 0.8.0 → 0.8.2

package/lib/request-capture.js

@@ -0,0 +1,394 @@
+/**
+ * Request capture — diagnostic dump of inbound /v1/messages and
+ * /v1/chat/completions request bodies (and response usage) to disk,
+ * with a human-readable summary that breaks down system-block sizes,
+ * cache_control markers, tool blocks, message counts, token estimates,
+ * and (when response data is available) actual cache hit rates.
+ *
+ * Off by default. Three ways to turn it on, in order of precedence:
+ *
+ *   1. Env var:    MOBY_CAPTURE=1 mobygate start
+ *   2. Touch file: touch ~/.mobygate/.capture-enabled
+ *   3. (env/file unset → off)
+ *
+ * The touch-file path lets the dashboard toggle capture live without
+ * restarting mobygate. Removing the file disables capture immediately.
+ *
+ * Output: ~/.mobygate/captures/{timestamp}_{path}_{requestId}.{json,summary.txt}
+ *
+ *   .json         — raw request body (pretty-printed)
+ *   .summary.txt  — analysis: system blocks, cache markers, message
+ *                   timeline, tool definitions, token breakdown, and
+ *                   (after response lands) actual usage with cache hits
+ *
+ * Auto-rotation: oldest captures are deleted to keep total count
+ * under MOBY_CAPTURE_KEEP (default 100 captures = 200 files since we
+ * write 2 per request).
+ *
+ * Throws nothing — capture failures log a warning and return. Capture
+ * never blocks request processing.
+ */
+
+import { writeFile, mkdir, appendFile, readdir, unlink, stat } from 'fs/promises';
+import { existsSync } from 'fs';
+import { join } from 'path';
+import { homedir } from 'os';
+
+const CAPTURE_DIR = process.env.MOBYGATE_CAPTURE_DIR
+  || join(process.env.MOBYGATE_HOME || join(homedir(), '.mobygate'), 'captures');
+
+const TOGGLE_FILE = join(process.env.MOBYGATE_HOME || join(homedir(), '.mobygate'), '.capture-enabled');
+
+const KEEP_COUNT = parseInt(process.env.MOBY_CAPTURE_KEEP || '100', 10);
+
+// In-memory map of requestId → summary file path. Populated by
+// captureRequest() and consumed by captureResponse() so we can append
+// response data to the same summary file we wrote on the way in.
+const inFlightSummaries = new Map();
+
+let dirEnsured = false;
+
+async function ensureDir() {
+  if (dirEnsured) return;
+  if (!existsSync(CAPTURE_DIR)) {
+    await mkdir(CAPTURE_DIR, { recursive: true });
+  }
+  dirEnsured = true;
+}
+
+/**
+ * Estimate token count from a string. Rough — 4 chars per token is
+ * the standard back-of-envelope for English+code mixed content.
+ */
+function estimateTokens(s) {
+  if (!s) return 0;
+  return Math.round(String(s).length / 4);
+}
+
+/**
+ * Walk a content array (or string) and sum total characters across all
+ * text blocks. Anthropic's content can be a bare string or an array of
+ * { type: 'text'|'image'|'tool_use'|'tool_result', ... } blocks.
+ */
+function contentBytes(content) {
+  if (typeof content === 'string') return content.length;
+  if (!Array.isArray(content)) return 0;
+  let total = 0;
+  for (const block of content) {
+    if (typeof block === 'string') { total += block.length; continue; }
+    if (!block || typeof block !== 'object') continue;
+    if (typeof block.text === 'string') total += block.text.length;
+    if (typeof block.input === 'object') total += JSON.stringify(block.input).length;
+    if (typeof block.content === 'string') total += block.content.length;
+    if (Array.isArray(block.content)) total += contentBytes(block.content);
+  }
+  return total;
+}
+
+/**
+ * Returns a tool's name from either Anthropic-shape (top-level `name`)
+ * or OpenAI-shape (nested under `function.name`). Used by the summary
+ * tools listing — earlier we showed "(unnamed)" for OpenAI tools because
+ * we only checked the top-level `name` field.
+ */
+function toolName(t) {
+  if (!t || typeof t !== 'object') return '(unnamed)';
+  if (typeof t.name === 'string') return t.name;
+  if (t.function && typeof t.function.name === 'string') return t.function.name;
+  return '(unnamed)';
+}
+
+/**
+ * Build a human-readable analysis of an Anthropic-shape body. Works for
+ * both /v1/messages (native) and translated /v1/chat/completions bodies
+ * where messages have content arrays.
+ */
+function analyzeAnthropic(body) {
+  const lines = [];
+  lines.push(`model:           ${body.model || '(none)'}`);
+  lines.push(`stream:          ${!!body.stream}`);
+  lines.push(`max_tokens:      ${body.max_tokens ?? body.max_completion_tokens ?? '(none)'}`);
+  lines.push(`temperature:     ${body.temperature ?? '(default)'}`);
+  lines.push(`session_id:      ${body.session_id ?? '(none)'}`);
+  lines.push('');
+
+  // System block(s) — Anthropic accepts string or array of {type, text, cache_control?}
+  const sys = body.system;
+  if (typeof sys === 'string') {
+    lines.push(`system:          1 block (string), ${sys.length} bytes, ~${estimateTokens(sys)} tokens`);
+    lines.push(`  cache_control: NONE (system is bare string — markers only work on array form)`);
+  } else if (Array.isArray(sys)) {
+    const totalBytes = sys.reduce((acc, b) => acc + (b?.text?.length || 0), 0);
+    lines.push(`system:          ${sys.length} blocks (array), ${totalBytes} bytes, ~${estimateTokens(' '.repeat(totalBytes))} tokens`);
+    sys.forEach((block, i) => {
+      const bytes = block?.text?.length || 0;
+      const marker = block?.cache_control ? ` [cache_control: ${JSON.stringify(block.cache_control)}]` : '';
+      lines.push(`  [${i}] ${block?.type || '?'} ${bytes} bytes${marker}`);
+    });
+    const cached = sys.filter(b => b?.cache_control).length;
+    lines.push(`  cache_control: ${cached}/${sys.length} system blocks marked`);
+  } else {
+    lines.push(`system:          (none)`);
+  }
+  lines.push('');
+
+  // Messages breakdown
+  const msgs = body.messages || [];
+  lines.push(`messages:        ${msgs.length}`);
+  let totalContentBytes = 0;
+  let imageCount = 0;
+  let toolUseCount = 0;
+  let toolResultCount = 0;
+  let cacheControlInMessages = 0;
+  msgs.forEach((m, i) => {
+    const bytes = contentBytes(m.content);
+    totalContentBytes += bytes;
+    if (Array.isArray(m.content)) {
+      for (const b of m.content) {
+        if (b?.type === 'image') imageCount += 1;
+        if (b?.type === 'tool_use') toolUseCount += 1;
+        if (b?.type === 'tool_result') toolResultCount += 1;
+        if (b?.cache_control) cacheControlInMessages += 1;
+      }
+    }
+    if (i < 3 || i >= msgs.length - 2) {
+      const role = m.role || '?';
+      const preview = (typeof m.content === 'string' ? m.content : JSON.stringify(m.content)).slice(0, 80).replace(/\s+/g, ' ');
+      lines.push(`  [${i}] ${role.padEnd(10)} ${bytes.toString().padStart(7)} b   ${preview}${preview.length >= 80 ? '…' : ''}`);
+    } else if (i === 3 && msgs.length > 5) {
+      lines.push(`  ... ${msgs.length - 5} more messages omitted ...`);
+    }
+  });
+  lines.push('');
+  lines.push(`messages bytes:  ${totalContentBytes} (~${estimateTokens(' '.repeat(totalContentBytes))} tokens)`);
+  lines.push(`images:          ${imageCount}`);
+  lines.push(`tool_use:        ${toolUseCount}`);
+  lines.push(`tool_result:     ${toolResultCount}`);
+  lines.push(`cache_control in messages: ${cacheControlInMessages}`);
+  lines.push('');
+
+  // Tools (declared client tools) — handle both Anthropic and OpenAI shapes
+  if (Array.isArray(body.tools)) {
+    const toolBytes = JSON.stringify(body.tools).length;
+    lines.push(`tools declared:  ${body.tools.length} (${toolBytes} bytes of schema)`);
+    body.tools.slice(0, 10).forEach(t => {
+      lines.push(`  - ${toolName(t)}`);
+    });
+    if (body.tools.length > 10) lines.push(`  ... and ${body.tools.length - 10} more`);
+  } else {
+    lines.push('tools declared:  (none)');
+  }
+  lines.push('');
+
+  // Grand total estimate
+  const sysBytes = typeof sys === 'string' ? sys.length
+    : Array.isArray(sys) ? sys.reduce((a, b) => a + (b?.text?.length || 0), 0)
+    : 0;
+  const toolBytes = Array.isArray(body.tools) ? JSON.stringify(body.tools).length : 0;
+  const grand = sysBytes + totalContentBytes + toolBytes;
+  lines.push(`────`);
+  lines.push(`grand total:     ${grand} bytes ≈ ${estimateTokens(' '.repeat(grand))} input tokens`);
+  lines.push(`  system:        ${sysBytes} (${pct(sysBytes, grand)}%)`);
+  lines.push(`  messages:      ${totalContentBytes} (${pct(totalContentBytes, grand)}%)`);
+  lines.push(`  tool schemas:  ${toolBytes} (${pct(toolBytes, grand)}%)`);
+
+  return lines.join('\n');
+}
+
+function pct(part, total) {
+  if (!total) return '0';
+  return ((part / total) * 100).toFixed(1);
+}
+
+/**
+ * Capture a request to disk. `path` is the route ('/v1/messages' or
+ * '/v1/chat/completions'), used for filename only. `body` is the parsed
+ * request body. `meta` carries session-key resolution info.
+ *
+ * Returns nothing. Errors logged to console.warn and swallowed — capture
+ * is best-effort and must not block requests.
+ */
+export async function captureRequest({ path, body, requestId, sessionKey, sessionKeySource }) {
+  if (!isCaptureEnabled()) return;
+
+  try {
+    await ensureDir();
+
+    const ts = new Date().toISOString().replace(/[:.]/g, '-').replace('T', '_').slice(0, 19);
+    const slug = path.replace(/[\/]/g, '-').replace(/^-/, '');
+    const baseName = `${ts}_${slug}_${requestId}`;
+    const jsonPath = join(CAPTURE_DIR, `${baseName}.json`);
+    const summaryPath = join(CAPTURE_DIR, `${baseName}.summary.txt`);
+
+    const header = [
+      `mobygate request capture`,
+      `─────────────────────────`,
+      `timestamp:       ${new Date().toISOString()}`,
+      `path:            ${path}`,
+      `request_id:      ${requestId}`,
+      `session_key:     ${sessionKey || '(none)'}`,
+      `session_source:  ${sessionKeySource || '(unknown)'}`,
+      ``,
+    ].join('\n');
+
+    const analysis = analyzeAnthropic(body);
+
+    await Promise.all([
+      writeFile(jsonPath, JSON.stringify(body, null, 2), 'utf8'),
+      writeFile(summaryPath, header + analysis + '\n', 'utf8'),
+    ]);
+
+    // Remember the summary path so captureResponse() can append to it.
+    inFlightSummaries.set(requestId, summaryPath);
+
+    // Best-effort prune to stay under the cap. Don't await — let it run
+    // alongside the next request.
+    pruneOldCaptures().catch(() => {});
+
+    console.log(`[capture] ${baseName} (${jsonPath.replace(homedir(), '~')})`);
+  } catch (e) {
+    console.warn(`[capture] failed for ${requestId}: ${e.message}`);
+  }
+}
+
+/**
+ * Append response usage data to the summary file we wrote on the request
+ * side. If the request's summary file isn't found in our in-flight map,
+ * we silently no-op — that means capture wasn't enabled when the request
+ * came in, or this requestId was never captured. Calling captureResponse
+ * is always safe.
+ *
+ * `usage` should be the SDK's NonNullableUsage shape:
+ *   { input_tokens, output_tokens, cache_read_input_tokens,
+ *     cache_creation_input_tokens, ... }
+ *
+ * `meta` carries: durationMs, status, stopReason, model.
+ */
+export async function captureResponse({ requestId, usage, durationMs, status, stopReason, model, error }) {
+  const summaryPath = inFlightSummaries.get(requestId);
+  if (!summaryPath) return;
+  inFlightSummaries.delete(requestId);
+
+  try {
+    const u = usage || {};
+    const totalInput = (u.input_tokens || 0) + (u.cache_read_input_tokens || 0) + (u.cache_creation_input_tokens || 0);
+    const cacheHitPct = totalInput > 0 ? (((u.cache_read_input_tokens || 0) / totalInput) * 100).toFixed(1) : '0';
+
+    const lines = [
+      ``,
+      `═══ RESPONSE ═══`,
+      `status:                       ${status || '(unknown)'}`,
+      `duration:                     ${durationMs ? durationMs + ' ms' : '(unknown)'}`,
+      `model:                        ${model || '(unknown)'}`,
+      `stop_reason:                  ${stopReason || '(none)'}`,
+    ];
+
+    if (error) {
+      lines.push(`error:                        ${error}`);
+    }
+
+    if (usage) {
+      lines.push(``);
+      lines.push(`usage:`);
+      lines.push(`  input_tokens (uncached):    ${u.input_tokens ?? 0}`);
+      lines.push(`  cache_read_input_tokens:    ${u.cache_read_input_tokens ?? 0}   (charged 0.1x)`);
+      lines.push(`  cache_creation_input_tokens: ${u.cache_creation_input_tokens ?? 0}   (charged 1.25x)`);
+      lines.push(`  output_tokens:              ${u.output_tokens ?? 0}`);
+      lines.push(``);
+      lines.push(`cache hit rate:               ${cacheHitPct}%   (${u.cache_read_input_tokens ?? 0} of ${totalInput} input tokens)`);
+
+      // Effective cost (in equivalent uncached tokens):
+      //   uncached input × 1.0 + cache_read × 0.1 + cache_create × 1.25 + output × 5.0 (per Anthropic Opus pricing)
+      // For reference only — actual billing depends on model.
+      const effectiveInput =
+        (u.input_tokens ?? 0) * 1.0 +
+        (u.cache_read_input_tokens ?? 0) * 0.1 +
+        (u.cache_creation_input_tokens ?? 0) * 1.25;
+      lines.push(`effective input cost:         ${effectiveInput.toFixed(0)} input-tokens-equiv (vs ${totalInput} wire-level)`);
+      const savings = totalInput > 0 ? (((totalInput - effectiveInput) / totalInput) * 100).toFixed(1) : '0';
+      lines.push(`savings from cache:           ${savings}%`);
+    }
+
+    await appendFile(summaryPath, lines.join('\n') + '\n', 'utf8');
+  } catch (e) {
+    console.warn(`[capture] response append failed for ${requestId}: ${e.message}`);
+  }
+}
+
+/**
+ * Prune old capture files to stay under KEEP_COUNT. Sorts by mtime,
+ * keeps the newest 2*KEEP_COUNT files (since each request writes 2
+ * files: .json and .summary.txt). Best-effort — failures swallowed.
+ */
+async function pruneOldCaptures() {
+  if (!existsSync(CAPTURE_DIR)) return;
+  let entries;
+  try {
+    entries = await readdir(CAPTURE_DIR);
+  } catch (e) {
+    return;
+  }
+  if (entries.length <= KEEP_COUNT * 2) return;
+
+  // Stat all files for mtime, sort newest-first, drop the tail.
+  const stats = [];
+  for (const name of entries) {
+    const full = join(CAPTURE_DIR, name);
+    try {
+      const st = await stat(full);
+      if (st.isFile()) stats.push({ name, full, mtime: st.mtimeMs });
+    } catch {}
+  }
+  stats.sort((a, b) => b.mtime - a.mtime);
+  const toDelete = stats.slice(KEEP_COUNT * 2);
+  for (const f of toDelete) {
+    try { await unlink(f.full); } catch {}
+  }
+}
+
+let cachedFlag;
+
+/**
+ * Returns true if request capture is enabled. Three sources, in order:
+ *   1. MOBY_CAPTURE env var (set/unset)
+ *   2. Touch file at ~/.mobygate/.capture-enabled
+ *   3. Default: false
+ *
+ * Cached for 1s to avoid spamming process.env / fs.exists on every
+ * request. The 1s cache is short enough to feel "live" when toggled
+ * from the dashboard, fast enough to not bottleneck request handling.
+ */
+export function isCaptureEnabled() {
+  const now = Date.now();
+  if (cachedFlag && cachedFlag.expires > now) return cachedFlag.value;
+
+  let value = false;
+  if (process.env.MOBY_CAPTURE === '1' || process.env.MOBY_CAPTURE === 'true') {
+    value = true;
+  } else if (existsSync(TOGGLE_FILE)) {
+    value = true;
+  }
+
+  cachedFlag = { value, expires: now + 1000 };
+  return value;
+}
+
+/**
+ * Programmatic toggle — flips the touch file. Returns the new state.
+ * Used by the dashboard toggle button.
+ */
+export async function setCaptureEnabled(enabled) {
+  await ensureDir();
+  const dir = join(process.env.MOBYGATE_HOME || join(homedir(), '.mobygate'));
+  if (!existsSync(dir)) await mkdir(dir, { recursive: true });
+  if (enabled) {
+    await writeFile(TOGGLE_FILE, `enabled at ${new Date().toISOString()}\n`, 'utf8');
+  } else {
+    try { await unlink(TOGGLE_FILE); } catch {}
+  }
+  cachedFlag = null; // invalidate so next isCaptureEnabled() reads fresh
+  return isCaptureEnabled();
+}
+
+export const CAPTURE_DIR_PATH = CAPTURE_DIR;
+export const CAPTURE_TOGGLE_FILE = TOGGLE_FILE;

package/lib/session-derive.js

@@ -40,6 +40,12 @@
  *     user message from history mid-conversation, the auto-key changes
  *     and the SDK starts a new session. One turn of double-billing,
  *     then we're back on the new key. Acceptable.
+ *   - **Multi-agent collisions** (fixed in v0.8.2): two agents that
+ *     share boilerplate at the start of their system prompt previously
+ *     collided onto one session key when the trim window only covered
+ *     the boilerplate. SYSTEM_TRIM was raised from 500 to 20000 chars
+ *     to capture the per-agent personality content that follows the
+ *     shared preamble. See note on the constant below for details.
  *
  * Opt-out: `X-Session-Id: none` tells us the client explicitly wants
  * stateless behavior — we return null and the request flows through
@@ -51,7 +57,20 @@
 import { createHash } from 'crypto';
 
 const HASH_LEN = 16;
-const SYSTEM_TRIM = 500;
+// SYSTEM_TRIM was 500 in v0.7.1 — large enough for casual single-agent
+// scenarios (Hermes, single-bot OpenClaw) but caused collisions when
+// multiple agents shared a common boilerplate prefix. Observed in v0.8.1
+// production: Lux + Mercury (two OpenClaw agents) both started their
+// system prompt with the OpenClaw "You are a personal assistant…"
+// boilerplate that filled the first ~500 chars, so their personality
+// markers (loaded from per-agent SOUL.md / IDENTITY.md / etc.) didn't
+// reach the hash and they collided onto the same session key.
+//
+// Bumping to 20kB covers realistic agent system prompts including
+// rich workspace bootstrap (Lux: ~42kB, Mercury: ~80kB total — but
+// the first 20kB has more than enough divergence to fingerprint each).
+// SHA-256 cost on 20kB is ~10-20µs, irrelevant per request.
+const SYSTEM_TRIM = 20000;
 const USER_TRIM = 500;
 
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mobygate",
-  "version": "0.8.0",
+  "version": "0.8.2",
   "description": "OpenAI-compatible local proxy for Claude Max. The Möbius-strip gateway: OpenAI shape in, Claude Max out.",
   "type": "module",
   "main": "server.js",
@@ -58,6 +58,7 @@
     "launchd",
     "server.js",
     "index.html",
+    "inspector.html",
     "mcp-inspect.mjs",
     "README.md",
     "CHANGELOG.md",