skalpel 4.0.47 → 4.0.49

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "skalpel",
-  "version": "4.0.47",
+  "version": "4.0.49",
   "description": "Behavioral graph for AI-assisted coding — learns how you work and steers Claude Code + Codex in real time.",
   "type": "module",
   "bin": {

package/skalpel-setup.mjs

@@ -66,6 +66,7 @@ const KNOWN_SUBS = new Set([
   "catches",
   "card",
   "ledger",
+  "eval",
   "analytics",
   "verify",
   "yes",
@@ -87,6 +88,7 @@ usage:
   skalpel catches                               scan your own history for "done" claims over failing tests
   skalpel card [--last|--id <id>] [--md]        share one verified catch: the claim + its own failing proof
   skalpel ledger                                your running tally: N "done"s checked, M verified FALSE
+  skalpel eval [--variant <v>] [--json]         offline replay of your history: lie frequency, measured cost, catch accuracy
   skalpel analytics                             interactive local analytics — caught lies, measured cost, steers, re-asks
   skalpel verify on|off|status                  arm/disarm the live catch (re-runs your own proof on a "done" claim)
   skalpel yes | skalpel no                      was the last catch right? (feeds precision; "no" stops that proof this session)
@@ -125,11 +127,12 @@ function validateArgv() {
       process.exit(0);
     }
     if (BOOL_FLAGS.has(a)) continue; // valueless flag (e.g. `uninstall --purge`)
-    // `skalpel card` owns its own flags (--last / --id <id> / --md / --plain) and forwards them to
-    // card.mjs. They are NOT setup options, so setup must not reject them as unknown — pass any token
-    // after the `card` subcommand straight through (card.mjs validates its own flags). Global -v/-h
-    // above still win for a bare `skalpel -v`. Scoped to `card` so no other subcommand's contract shifts.
-    if (i > 0 && argv[0] === "card") continue;
+    // `skalpel card` (--last / --id <id> / --md / --plain) and `skalpel eval` (--variant / --limit /
+    // --json / --home / --out) own their own flags and forward them to card.mjs / efficacy-eval.mjs.
+    // They are NOT setup options, so setup must not reject them as unknown — pass any token after those
+    // subcommands straight through (each module validates its own flags). Global -v/-h above still win
+    // for a bare `skalpel -v`. Scoped so no other subcommand's contract shifts.
+    if (i > 0 && (argv[0] === "card" || argv[0] === "eval")) continue;
     if (a.startsWith("-")) {
       console.error(`skalpel: unknown option \`${a}\`\n\n${USAGE}`);
       process.exit(2);
@@ -993,6 +996,17 @@ async function main() {
     spawnSync("node", [join(__dir, "ledger.mjs"), ...argv.slice(1)], { stdio: "inherit" });
     return;
   }
+  // `skalpel eval` — THE EFFICACY EVAL: a LOCAL, READ-ONLY, ZERO-NETWORK offline REPLAY of the user's own
+  // Claude Code + Codex history that measures whether the "agent lies about done" fire is real FOR THEM and
+  // how much it costs — lie frequency (confirmed-lie vs true-done vs unverifiable), MEASURED red→green cost
+  // (real timestamps only, never estimated), and catch accuracy (raw-flag → surfaced vs harness-suppressed,
+  // precision + relative recall) with a pluggable catch-config `--variant` so a stricter gate replays in
+  // seconds. Reuses the SAME detectors as the live catch (no divergent claim logic); never re-runs a proof,
+  // never touches the network, never fabricates a number; writes nothing unless the user asks with `--out`.
+  if (sub === "eval") {
+    spawnSync("node", [join(__dir, "efficacy-eval.mjs"), ...argv.slice(1)], { stdio: "inherit" });
+    return;
+  }
   // `skalpel analytics` — INTERACTIVE LOCAL ANALYTICS: a keyboard-navigable, READ-ONLY terminal view
   // derived from the same honest local data as `skalpel ledger` (verify-shadow.log + steers.ndjson +
   // insights.ndjson). Zero network, zero writes, zero estimates. Degrades to a plain static snapshot