mcp-context-cost 0.7.0 → 0.9.0

package/dist/sweep/sweep-all.js

@@ -14,9 +14,10 @@
 import { readFileSync } from 'node:fs';
 import { parse } from 'yaml';
 import { measureServer } from './run.js';
+import { DockerHarnessFault } from './docker.js';
 import { writeLeaderboard } from './report.js';
 import { appendHistory } from './history.js';
-import { snapshot, verdict, restore } from './harness-guard.js';
+import { FAULT_RATIO, MIN_REGRESSIONS, snapshot, verdict, restore } from './harness-guard.js';
 import { selectShard, shardIndexForDate } from './shard.js';
 function arg(name) {
     const i = process.argv.indexOf(`--${name}`);
@@ -64,17 +65,31 @@ const prior = snapshot(entries.map((e) => e.name));
 const queue = [...entries];
 const summary = {};
 const statuses = new Map();
+// Servers docker itself failed to run — never a measurement (measureServer
+// throws before persisting), so each one's previous record simply stands.
+const dockerFaults = new Map();
 async function worker() {
     for (let e = queue.shift(); e; e = queue.shift()) {
         const started = Date.now();
-        const m = await measureServer(e.name, e.command, {
-            timeoutMs: (e.timeoutSeconds ?? defaultTimeout) * 1000,
-            docker,
-            dockerImage: e.dockerImage,
-            dummyEnv: e.env ?? [],
-            dummyEnvValues: e.envValues,
-            needsGit: e.needsGit,
-        });
+        let m;
+        try {
+            m = await measureServer(e.name, e.command, {
+                timeoutMs: (e.timeoutSeconds ?? defaultTimeout) * 1000,
+                docker,
+                dockerImage: e.dockerImage,
+                dummyEnv: e.env ?? [],
+                dummyEnvValues: e.envValues,
+                needsGit: e.needsGit,
+            });
+        }
+        catch (err) {
+            if (!(err instanceof DockerHarnessFault))
+                throw err;
+            dockerFaults.set(e.name, err.message);
+            summary[e.name] = 'docker harness fault — not measured; previous record untouched';
+            console.log(`  ${e.name}: ${summary[e.name]}`);
+            continue;
+        }
         const secs = ((Date.now() - started) / 1000).toFixed(0);
         statuses.set(e.name, m.status);
         summary[e.name] =
@@ -85,6 +100,19 @@ async function worker() {
     }
 }
 await Promise.all(Array.from({ length: Math.max(1, concurrency) }, () => worker()));
+// A docker fault on most of the slice is the harness-fault story with an
+// earlier symptom — the guard below can't see it (nothing regressed on disk;
+// the throws happened before anything was written), so it is judged here by
+// the guard's own thresholds. Below them, the sweep publishes what it did
+// measure and the faulted servers wait for the next cycle.
+if (dockerFaults.size >= MIN_REGRESSIONS && dockerFaults.size / entries.length >= FAULT_RATIO) {
+    console.error(`\nHARNESS FAULT — docker could not run for ${dockerFaults.size} of ${entries.length} servers; ` +
+        `refusing to publish this sweep.\n` +
+        [...dockerFaults].map(([n, msg]) => `  ${n}: ${msg}`).join('\n') +
+        `\n  Nothing was overwritten — every previous record stands. ` +
+        `Check the Docker daemon and registry path, then re-run.`);
+    process.exit(1);
+}
 // Before publishing anything: is this sweep a statement about the servers, or
 // about the machine that measured them?
 const v = verdict(prior, statuses);
@@ -105,4 +133,8 @@ if (v.fault) {
 writeLeaderboard(doc.servers);
 const h = appendHistory();
 const measured = Object.values(summary).filter((s) => s.includes('tokens')).length;
+if (dockerFaults.size > 0) {
+    console.warn(`\nDOCKER FAULT on ${[...dockerFaults.keys()].join(', ')} — not measured this sweep; ` +
+        `previous records untouched. The next cycle re-attempts ${dockerFaults.size === 1 ? 'it' : 'them'}.`);
+}
 console.log(`done: ${measured}/${entries.length} measured; leaderboard + history (${h.rows} rows) regenerated`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mcp-context-cost",
-  "version": "0.7.0",
+  "version": "0.9.0",
   "description": "Measure what your MCP servers cost in context tokens — audit your own config, or badge the server you publish",
   "type": "module",
   "license": "MIT",
@@ -38,6 +38,7 @@
     "test:badge": "./upstream/tests/badge-test.sh",
     "typecheck": "tsc --noEmit && tsc -p tsconfig.tools.json",
     "adoption": "tsx tools/measure-adoption.ts",
+    "cross-check": "tsx src/sweep/cross-check.ts",
     "divergence": "tsx tools/measure-divergence.ts",
     "session-start": "tsx src/sweep/session-start.ts",
     "sweep": "tsx src/sweep/run.ts",