@chatpanel/gateway 0.6.68 → 0.6.70

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chatpanel/gateway",
-  "version": "0.6.68",
+  "version": "0.6.70",
   "description": "Local privacy gateway \u2014 redacts PII out of OpenAI/Anthropic API traffic before it reaches a model, then restores it in the reply. Point opencode, codex, aider, Claude Code, etc. at it.",
   "type": "module",
   "bin": {

package/src/bridge.js

@@ -17,6 +17,47 @@ import os from 'node:os';
 
 const DEFAULT_TOKEN_PATH = join(os.homedir(), '.chatpanel', 'bridge-token');
 
+export const DEFAULT_BRIDGE_URL = 'http://127.0.0.1:4319';
+const BRIDGE_PROBE_TTL_MS = 10_000;
+let bridgeResolved = { at: 0, cfgUrl: '', url: '', fell: false };
+
+async function bridgeAnswers(url, timeoutMs) {
+  try {
+    const r = await fetch(`${url}/health`, { signal: AbortSignal.timeout(timeoutMs) });
+    return r.ok;
+  } catch { return false; }
+}
+
+/**
+ * Where the bridge is, checked rather than believed.
+ *
+ * The persisted bridge.url was found pointing at an ephemeral port nothing answered on,
+ * while the real bridge sat on 4319: every agent showed unavailable, every agent turn failed,
+ * and the startup line said so in a log nobody was reading. A configured address that does
+ * not answer while the default one does is a stale setting, not a decision, so the default
+ * is used and the fall-back is logged. Probed at most every ten seconds.
+ */
+export async function resolveBridgeUrl(cfg, { fallback, timeoutMs = 1500, now = Date.now() } = {}) {
+  // CHATPANEL_BRIDGE_FALLBACK=off disables the fallback: the test suite sets it, because a
+  // test whose fake bridge has gone away would otherwise find the developer's REAL bridge on
+  // 4319 and send its turns to a live coding agent.
+  if (fallback === undefined) fallback = process.env.CHATPANEL_BRIDGE_FALLBACK === 'off' ? '' : DEFAULT_BRIDGE_URL;
+  const cfgUrl = String(cfg?.bridge?.url || '').replace(/\/$/, '');
+  if (bridgeResolved.url && bridgeResolved.cfgUrl === cfgUrl && now - bridgeResolved.at < BRIDGE_PROBE_TTL_MS) return bridgeResolved.url;
+  let url = cfgUrl || fallback;
+  let fell = false;
+  if (fallback && url !== fallback && !(await bridgeAnswers(url, timeoutMs)) && await bridgeAnswers(fallback, timeoutMs)) {
+    url = fallback;
+    fell = true;
+    if (!bridgeResolved.fell || bridgeResolved.cfgUrl !== cfgUrl) console.log(`[gateway] bridge.url ${cfgUrl} is not answering; the bridge on ${fallback} is — using it (fix the address in Settings to silence this)`);
+  }
+  bridgeResolved = { at: now, cfgUrl, url, fell };
+  return url;
+}
+
+/** Test seam. */
+export function resetBridgeResolution() { bridgeResolved = { at: 0, cfgUrl: '', url: '', fell: false }; }
+
 export function readBridgeToken(cfgToken, tokenPath = DEFAULT_TOKEN_PATH) {
   if (cfgToken) return cfgToken;
   try {

package/src/configstore.js

@@ -83,6 +83,21 @@ export function applyNerModelSelection(cfg, id) {
 }
 
 // Merge an editable patch into the live cfg. Only known fields; ignores the rest.
+/**
+ * Is this API destination the gateway itself? A loop: a turn routed there comes straight
+ * back in. It arrived once from a backup import — the extension keeps the gateway in its own
+ * endpoint list, and the desktop copied that list into the gateway's destinations — and its
+ * model ids (the agents') then shadowed the real agents. Refused at the door, with a log line.
+ */
+export function isSelfDestination(d, cfg = {}) {
+  if (!d || d.type !== 'api') return false;
+  let u;
+  try { u = new URL(String(d.baseUrl || '')); } catch { return false; }
+  const loop = /^(127\.0\.0\.1|localhost|\[::1\]|0\.0\.0\.0)$/i.test(u.hostname);
+  const port = Number(u.port) || (u.protocol === 'https:' ? 443 : 80);
+  return loop && port === (Number(cfg.port) || 4320);
+}
+
 export function applyConfigPatch(cfg, patch = {}) {
   if (patch.backend === 'bridge' || patch.backend === 'api') cfg.backend = patch.backend;
   if (Array.isArray(patch.destinations)) {
@@ -91,6 +106,7 @@ export function applyConfigPatch(cfg, patch = {}) {
     const prev = new Map((Array.isArray(cfg.destinations) ? cfg.destinations : []).map((d) => [d.id, d]));
     cfg.destinations = patch.destinations
       .filter((d) => d && typeof d.id === 'string' && (d.type === 'agent' || d.type === 'api'))
+      .filter((d) => { const self = isSelfDestination(d, cfg); if (self) console.log(`[gateway] refusing destination "${d.id}": ${d.baseUrl} is this gateway — a loop`); return !self; })
       .map((d) => {
         const out = { id: d.id, type: d.type, models: Array.isArray(d.models) ? d.models.filter((m) => typeof m === 'string' && m) : [] };
         if (d.type === 'agent') out.agent = d.agent || d.id;

package/src/router.js

@@ -12,7 +12,7 @@
 //   }
 
 import { secureFetch } from './secure-fetch.js';
-import { readBridgeToken } from './bridge.js';
+import { readBridgeToken, resolveBridgeUrl } from './bridge.js';
 //
 // /v1/models aggregates every destination's models so clients can discover them.
 
@@ -29,10 +29,15 @@ const KNOWN_AGENTS = ['codex', 'claude', 'opencode', 'pi', 'kiro', 'antigravity'
 export function listDestinations(cfg) {
   const configured = (Array.isArray(cfg.destinations) ? cfg.destinations : []).filter(Boolean);
   const haveId = new Set(configured.map((d) => d.id));
-  const out = [...configured];
-  for (const a of KNOWN_AGENTS) {
-    if (!haveId.has(a)) out.push({ id: a, type: 'agent', agent: a, models: [a] });
-  }
+  // AGENTS FIRST. The model list de-duplicates by id in this order, and resolveDestination
+  // takes the first destination that serves a model — so an API destination that happened to
+  // list "codex" or "claude" (a gateway pointed at itself did exactly that) used to swallow
+  // the agents: they vanished from the picker and a turn for one was proxied to the loop.
+  const out = [
+    ...configured.filter((d) => d.type === 'agent'),
+    ...KNOWN_AGENTS.filter((a) => !haveId.has(a)).map((a) => ({ id: a, type: 'agent', agent: a, models: [a] })),
+    ...configured.filter((d) => d.type !== 'agent'),
+  ];
   if (cfg.backend === 'api' && !configured.some((d) => d.type === 'api')) {
     out.push({ id: 'openai', type: 'api', protocol: 'openai', baseUrl: cfg.upstreams?.openai?.baseUrl, models: [] });
     out.push({ id: 'anthropic', type: 'api', protocol: 'anthropic', baseUrl: cfg.upstreams?.anthropic?.baseUrl, models: [] });
@@ -156,7 +161,7 @@ export function aggregateModels(cfg) {
  * spends a subprocess per agent to describe something unusable.
  */
 async function bridgeAgentModels(cfg, installed, timeoutMs) {
-  const base = String(cfg?.bridge?.url || '').replace(/\/$/, '');
+  const base = await resolveBridgeUrl(cfg);
   if (!base || !installed) return new Map();
   const token = readBridgeToken(cfg.bridge?.token);
   const ids = [...installed.entries()].filter(([, ok]) => ok).map(([id]) => id);
@@ -183,7 +188,7 @@ async function bridgeAgentModels(cfg, installed, timeoutMs) {
 
 /** id → installed, from the bridge's own /health. `null` when it could not be asked. */
 async function bridgeAgentAvailability(cfg, timeoutMs) {
-  const base = String(cfg?.bridge?.url || '').replace(/\/$/, '');
+  const base = await resolveBridgeUrl(cfg);
   if (!base) return null;
   try {
     const token = readBridgeToken(cfg.bridge?.token);

package/src/server.js

@@ -24,6 +24,7 @@ import { redactSegments, segment } from './redact.js';
 import { pipeRestoredStream, pipeRestoredOpenAIStream, makeTokenRestorer, restoreDeep } from './stream.js';
 import { restoreText, gatedDictionary, narrowSpecs, makeToolHarness, placeholderToolNote, assertEndpointUrl } from '@chatpanel/pii';
 import { ensureGatewayToken, isAdminAuthorized } from './gateway-token.js';
+import { resolveBridgeUrl } from './bridge.js';
 import { secureFetch } from './secure-fetch.js';
 import { streamBridgeChat, readBridgeToken, openBridgeChat } from './bridge.js';
 import { createRelaySession, getRelaySession, endRelaySession, pumpBridgeStream, deliverToolResult, toolsToSpecs, parseToolCallId } from './toolrelay.js';
@@ -55,7 +56,7 @@ import * as openai from './openai.js';
 import * as responses from './responses.js';
 import * as anthropic from './anthropic.js';
 
-export const VERSION = '0.6.68';
+export const VERSION = '0.6.70';
 
 // WARM search tier — SQLite + FTS5 record store (falls back to an encrypted-JSON
 // store if SQLite can't load), fed by the extension's ingest sync + backup-ingest.
@@ -170,7 +171,7 @@ function mkTrace(sink) {
       const entry = /** @type {any} */ ({ ...this.meta, timings });
       setImmediate(() => {
         sink(entry);
-        console.log(`[gateway] model=${entry.model || '-'} → ${entry.dest ? `${entry.dest}(${entry.type})` : 'none'} · redacted ${entry.redacted || 0}${entry.sanitized ? ` · scrubbed ${entry.sanitized} hidden` : ''}${entry.narrowed ? ` · narrowed -${entry.narrowed}` : ''} · ${fmtTimings(timings)}`);
+        console.log(`[gateway] model=${entry.model || '-'} → ${entry.dest ? `${entry.dest}(${entry.type})` : 'none'} · ${entry.redaction === 'off' ? 'redaction OFF (client request)' : `redacted ${entry.redacted || 0}`}${entry.sanitized ? ` · scrubbed ${entry.sanitized} hidden` : ''}${entry.narrowed ? ` · narrowed -${entry.narrowed}` : ''} · ${fmtTimings(timings)}`);
       });
     },
   };
@@ -356,13 +357,14 @@ async function startRelay(req, res, { kind, adapter, agent }, body, vault, cfg,
   // Full tier for everyone here (the free allowance is enforced in the main
   // handler), but the custom dictionary stays capped for free.
   const redactOpts = { tier: cfg.redaction.tier === 'full' ? 'full' : 'basic', dictionary: gatedDictionary(cfg.redaction, isPro), entities: [] };
-  const s = createRelaySession({ vault, redactOpts, bridgeUrl: cfg.bridge.url, token, harness });
+  const bridgeUrl = await resolveBridgeUrl(cfg);
+  const s = createRelaySession({ vault, redactOpts, bridgeUrl, token, harness });
   const ttl = setTimeout(() => endRelaySession(s.id), 135_000); // bridge tool-call timeout is 120s
   // The placeholder note is already in `system` (injected into the body after
   // redaction in the main handler), so toTurn() carried it here — nothing to add.
   let resp;
   try {
-    resp = await openBridgeChat({ bridgeUrl: cfg.bridge.url, agent, token, messages, system, specs: toolsToSpecs(tools), options: {}, signal: undefined });
+    resp = await openBridgeChat({ bridgeUrl, agent, token, messages, system, specs: toolsToSpecs(tools), options: {}, signal: undefined });
   } catch (e) { clearTimeout(ttl); endRelaySession(s.id); trace?.commit(); return sendJson(res, 502, { error: { message: `bridge: ${e.message}`, type: 'bridge_error' } }); }
   s.reader = resp.body.getReader();
   res.writeHead(200, { 'content-type': 'text/event-stream', 'cache-control': 'no-cache', connection: 'keep-alive' });
@@ -420,7 +422,7 @@ async function handleBridge(req, res, { kind, adapter, redactable, pathname, age
   // agent id, which leaves the agent on its own default exactly as before.
   const { agentModel } = parseAgentModel(body?.model, cfg);
   const turn = {
-    bridgeUrl: cfg.bridge.url, agent, token, messages, system, signal: ac.signal,
+    bridgeUrl: await resolveBridgeUrl(cfg), agent, token, messages, system, signal: ac.signal,
     ...(agentModel ? { options: { model: agentModel } } : {}),
   };
 
@@ -1541,7 +1543,7 @@ export function createGateway(cfg = loadConfig()) {
     // modes for one screen; and the direct-to-bridge path is the one with no policy in front
     // of it, so making it necessary for a feature is how it becomes the habit.
     if (req.method === 'GET' && pathname === '/skills') {
-      const base = String(cfg.bridge?.url || '').replace(/\/$/, '');
+      const base = await resolveBridgeUrl(cfg);
       if (!base) return sendJson(res, 503, { error: { message: 'no bridge is configured', type: 'no_bridge' } });
       const token = readBridgeToken(cfg.bridge?.token);
       try {
@@ -1572,7 +1574,7 @@ export function createGateway(cfg = loadConfig()) {
       if (!id || id === '.' || id === '..' || id.includes('/') || id.includes('\\')) {
         return sendJson(res, 400, { error: { message: 'not a skill id', type: 'invalid_request' } });
       }
-      const base = String(cfg.bridge?.url || '').replace(/\/$/, '');
+      const base = await resolveBridgeUrl(cfg);
       if (!base) return sendJson(res, 503, { error: { message: 'no bridge is configured', type: 'no_bridge' } });
       const token = readBridgeToken(cfg.bridge?.token);
       try {
@@ -1624,6 +1626,7 @@ export function createGateway(cfg = loadConfig()) {
     let redactedCount = 0;
     let sanitizedCount = 0;
     let narrowedTools = 0;
+    let redactionOff = false;
     let isPro = true;
     // Off the hot path: only build a trace when logging is on, so it adds nothing
     // when off (no clock reads, no record, no console line).
@@ -1659,14 +1662,24 @@ export function createGateway(cfg = loadConfig()) {
             type: 'free_limit_reached',
           } });
         }
-        const segs = r.adapter.collectSegments(body, cfg.redaction);
+        // REDACTION OFF, FOR THIS REQUEST, BECAUSE THE USER SAID SO. A note task that reads
+        // "write about NVIDIA GPUs" had NVIDIA replaced with an organisation placeholder, and
+        // the model — seeing only [[ORG_1]] — guessed a different company and wrote about
+        // that. The policy is right for the corpus and wrong for the user's own instruction,
+        // and only the user can tell which a given turn is. So an AUTHENTICATED local client
+        // (the desktop, the extension — both hold the gateway token) may send
+        // `X-ChatPanel-Redaction: off`; the trace records it so the ledger says "redaction
+        // was off for this turn" rather than "0 redactions", which would read as "nothing to
+        // redact". Anonymous callers cannot switch it off: that is what the token is for.
+        redactionOff = String(req.headers['x-chatpanel-redaction'] || '').trim().toLowerCase() === 'off' && isAdminAuthorized(req);
+        const segs = redactionOff ? [] : r.adapter.collectSegments(body, cfg.redaction);
         const ac = new AbortController();
         req.on('close', () => ac.abort());
         const rd0 = trace ? trace.clock() : 0;
         // Redact at the configured tier for everyone (free users get name/org
         // redaction within their allowance); the custom dictionary stays capped for
         // free (isPro decides that inside).
-        const { vault: v, count, sanitized } = await redactSegments(segs, cfg.redaction, {
+        const { vault: v, count, sanitized } = redactionOff ? { vault: null, count: 0, sanitized: 0 } : await redactSegments(segs, cfg.redaction, {
           signal: ac.signal,
           isPro,
           // A detector is the only hop that sees the request BEFORE redaction. It is guarded
@@ -1695,7 +1708,7 @@ export function createGateway(cfg = loadConfig()) {
         // tools (so privacy-aware models USE them instead of refusing). Injected
         // AFTER redaction so the note isn't itself redacted. Covers BOTH the API
         // forward and the relay (which reads system from this same body).
-        if (Array.isArray(body.tools) && body.tools.length && typeof r.adapter.injectSystemNote === 'function') {
+        if (!redactionOff && Array.isArray(body.tools) && body.tools.length && typeof r.adapter.injectSystemNote === 'function') {
           r.adapter.injectSystemNote(body, placeholderToolNote({ toolData: cfg.tools?.toolData }));
         }
         outBody = Buffer.from(JSON.stringify(body), 'utf8');
@@ -1749,7 +1762,7 @@ export function createGateway(cfg = loadConfig()) {
       });
     }
     if (trace) {
-      trace.meta = { t: Date.now(), model: body?.model || null, dest: dest ? dest.id : null, type: dest ? dest.type : null, redacted: redactedCount, sanitized: sanitizedCount, narrowed: narrowedTools, detail: redactionDetail(vault, cfg.logDetail) };
+      trace.meta = { t: Date.now(), model: body?.model || null, dest: dest ? dest.id : null, type: dest ? dest.type : null, redacted: redactedCount, sanitized: sanitizedCount, narrowed: narrowedTools, detail: redactionDetail(vault, cfg.logDetail), redaction: redactionOff ? 'off' : (cfg.redaction.tier || 'basic') };
     }
     if (dest && dest.type === 'api') {
       if (!dest.baseUrl) { trace?.commit(); return sendJson(res, 502, { error: `destination "${dest.id}" has no baseUrl` }); }

package/src/stream.js

@@ -17,23 +17,30 @@ import { restoreText, restoreWithAliases } from '@chatpanel/pii';
 // Returns a TransformStream-free chunk transformer: feed it decoded string chunks,
 // it returns the prefix that's safe to forward now and buffers a possibly-partial
 // trailing token. Call flush() at end-of-stream.
+/**
+ * Split `buf` into what is safe to forward now and what may still be the start of a token.
+ *
+ * Two shapes are held back: an unterminated "[[" (a token mid-way), and a trailing SINGLE
+ * "[" — a model tokenizes "[[ORG_1]]" as "[" + "[ORG_1]]" often enough that forwarding the
+ * lone bracket left every restored name wearing one: "[NVIDIA designs GPUs". "[[" cannot
+ * legitimately appear except as a token open, and a lone "[" at a chunk edge costs nothing to
+ * wait one chunk for.
+ */
+function splitSafe(buf) {
+  const open = buf.lastIndexOf('[[');
+  if (open !== -1 && !buf.slice(open).includes(']]')) return [buf.slice(0, open), buf.slice(open)];
+  if (buf.endsWith('[')) return [buf.slice(0, -1), '['];
+  return [buf, ''];
+}
+
 export function makeTokenRestorer(vault) {
   let buf = '';
   return {
     push(chunk) {
       if (!vault) return chunk || '';
       buf += chunk || '';
-      // If an unterminated "[[" sits in the tail, a token may still be forming —
-      // hold from there. "[[" can't legitimately appear except as a token open.
-      const open = buf.lastIndexOf('[[');
       let safe;
-      if (open !== -1 && !buf.slice(open).includes(']]')) {
-        safe = buf.slice(0, open);
-        buf = buf.slice(open);
-      } else {
-        safe = buf;
-        buf = '';
-      }
+      [safe, buf] = splitSafe(buf);
       return restoreText(safe, vault);
     },
     flush() {
@@ -83,10 +90,8 @@ function makeFieldRestorer(vault, restoreFn) {
   return {
     push(chunk) {
       buf += chunk || '';
-      const open = buf.lastIndexOf('[[');
       let safe;
-      if (open !== -1 && !buf.slice(open).includes(']]')) { safe = buf.slice(0, open); buf = buf.slice(open); }
-      else { safe = buf; buf = ''; }
+      [safe, buf] = splitSafe(buf);
       return restoreFn(safe, vault);
     },
     flush() { const out = restoreFn(buf, vault); buf = ''; return out; },