ambitry 0.1.1 → 0.1.3

package/README.md

@@ -150,6 +150,8 @@ Point a provider at a different upstream — Azure, a gateway, a local model —
 
 Cost is computed from published per-token rates. Models with no known rate report `null` rather than a guess; a cost dashboard that quietly invents numbers is worse than one that admits ignorance.
 
+**One request modification, and the only one.** OpenAI-compatible providers omit token usage from streamed responses unless `stream_options.include_usage` is set, and almost no agent code sets it — so without this, cost and spend caps would be blank for most streaming traffic. Ambitry adds that flag to outgoing OpenAI streaming requests when you haven't set it yourself. The visible effect is one extra terminal chunk carrying usage, which SDKs treat as a normal chunk with no choices. If you set `stream_options` yourself, it's left alone, and Anthropic requests are never modified.
+
 ## Development
 
 ```bash

package/dist/cli.js

@@ -1,5 +1,6 @@
 #!/usr/bin/env node
 import { resolve } from 'node:path';
+import { rmSync } from 'node:fs';
 import { Store } from "./db.js";
 import { Controls, loadPolicy } from "./policy.js";
 import { createServer } from "./server.js";
@@ -67,6 +68,13 @@ function main(argv) {
     const port = Number(flag(argv, 'port') ?? 8787);
     // Demo traffic goes to its own file so it never mixes with real traces.
     const dbPath = resolve(flag(argv, 'db') ?? (demo ? 'ambitry-demo.db' : 'ambitry.db'));
+    // Each demo run starts clean. Otherwise a second run doubles every number
+    // on the dashboard, which is confusing on its own and useless to record.
+    if (demo) {
+        for (const suffix of ['', '-journal', '-wal', '-shm']) {
+            rmSync(`${dbPath}${suffix}`, { force: true });
+        }
+    }
     const policyPath = resolve(flag(argv, 'policy') ?? 'ambitry.json');
     let controls;
     try {

package/dist/server.js

@@ -33,9 +33,22 @@ export function createServer({ store, controls }) {
                 examples: ['http://localhost:8787/anthropic', 'http://localhost:8787/openai/v1'],
             });
         }
-        const body = await readBody(req);
+        let body = await readBody(req);
         const parsed = parseRequest(route.provider.id, body);
         const requestJson = safeJson(body);
+        // OpenAI-compatible providers omit usage from streamed responses unless
+        // stream_options.include_usage is set, and almost no agent code sets it.
+        // Without this the cost column is empty for most streaming traffic, which
+        // guts the spend caps. Anthropic always reports usage, so this is only
+        // needed here. The cost is one extra terminal chunk carrying usage, which
+        // SDKs treat as a normal chunk with no choices.
+        if (route.provider.id === 'openai' && parsed.stream && requestJson && typeof requestJson === 'object') {
+            const r = requestJson;
+            if (r.stream_options === undefined) {
+                r.stream_options = { include_usage: true };
+                body = Buffer.from(JSON.stringify(r));
+            }
+        }
         // Scanned on the way out, so a leak is caught on the request that carries
         // it rather than after the provider has already received it.
         const findings = scanRequest(requestJson);

package/dist/viewer.js

@@ -125,9 +125,14 @@ async function refresh() {
   // The headline. What people came for is "am I leaking anything", and the
   // answer belongs above the fold, not inside a trace nobody opens.
   const f = status.findings || [];
+  const total = f.reduce((n, x) => n + x.occurrences, 0);
+  // "Credentials" only when they all are. An email address is personal data,
+  // not a credential, and this line is the first thing anyone reads — the
+  // whole positioning rests on not overstating.
+  const allCreds = f.every(x => x.severity === 'critical');
+  const noun = allCreds ? 'credential' : 'sensitive value';
   $('alert').innerHTML = !f.length ? '' :
-    '<div class="alert"><h3>' + f.reduce((n, x) => n + x.occurrences, 0) +
-    ' credential' + (f.reduce((n, x) => n + x.occurrences, 0) === 1 ? '' : 's') +
+    '<div class="alert"><h3>' + total + ' ' + noun + (total === 1 ? '' : 's') +
     ' sent to your model provider</h3>' +
     '<p>Found in outgoing prompts over the last 7 days. The finding keeps only a masked preview — ' +
     'but the prompt itself is in your local trace file, so treat that file as sensitive.</p>' +

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ambitry",
-  "version": "0.1.1",
+  "version": "0.1.3",
   "description": "Audit log and permission layer for AI agents. One-line integration, zero runtime dependencies.",
   "type": "module",
   "bin": {