skalpel 4.0.43 → 4.0.45

package/card.mjs

@@ -69,7 +69,20 @@ export function isGenuineCatch(row) {
   // Ship-status catches (category:"ship", outcome SHIP_FAIL) record the check under probe_command, not
   // proof_command — accept either so a genuine "you said deployed, it's 404" catch is not invisible.
   const command = row.proof_command || row.probe_command;
-  return failed && !!command && !!row.claim_text;
+  if (!failed || !command || !row.claim_text) return false;
+  // RETRO catches (category:"retro") are RECORDED at claim time — never re-run. The whole indictment rests
+  // on the exit code the transcript already recorded, so a retro row is a genuine catch ONLY when it carries
+  // that recorded non-zero exit code; without it we can't render the "already failed" provenance truthfully.
+  if (row.category === "retro") {
+    return Number.isInteger(row.recorded_exit_code) && row.recorded_exit_code !== 0;
+  }
+  return true;
+}
+
+// A retro row is RECORDED at claim time (from the user's own transcript), never re-run — so its wording
+// must never borrow the live card's "skalpel re-ran its own proof" language.
+function isRetro(row) {
+  return row && row.category === "retro";
 }
 
 // readCatches() → all genuine-catch rows in the shadow log, oldest→newest as written. Fail-open: any
@@ -133,9 +146,17 @@ export function extractCode(evidence) {
 
 // verdictPhrase(row) → the honest one-line verdict. Always truthful for a catch row.
 function verdictPhrase(row) {
-  const code = extractCode(row.evidence);
   const failN = extractFailCount(row.evidence);
   const count = failN != null ? ` · ${failN} failing` : "";
+  // RETRO: the exit code was RECORDED in the transcript at claim time — we never re-ran it, so the verdict
+  // must say "had already failed", never "re-ran". recorded_exit_code is a required field for a retro catch.
+  if (isRetro(row)) {
+    const n = Number.isInteger(row.recorded_exit_code)
+      ? ` (recorded exit ${row.recorded_exit_code})`
+      : "";
+    return `FAIL — its own test had already failed when it said this${n}${count}`;
+  }
+  const code = extractCode(row.evidence);
   if (code && code.kind === "http") return `FAIL — its own proof got HTTP ${code.value}${count}`;
   if (code && code.kind === "exit") return `FAIL — its own proof exited ${code.value}${count}`;
   return `FAIL — its own proof re-ran and exited non-zero${count}`;
@@ -152,13 +173,18 @@ export function renderPlainCard(row) {
   const claim = clean(row.claim_text, 200) || "it's done";
   const proof = clean(row.proof_command || row.probe_command, 200) || "its own proof";
   const evidence = clean(row.evidence, 200);
+  const retro = isRetro(row);
   const L = [];
-  L.push("🔬 skalpel — verified catch");
+  L.push(retro ? "🔬 skalpel — recorded catch" : "🔬 skalpel — verified catch");
   L.push("");
   L.push("The agent claimed:");
   L.push(`  "${claim}"`);
   L.push("");
-  L.push("skalpel re-ran the agent's OWN proof, out-of-band:");
+  L.push(
+    retro
+      ? "Its own test had already failed when it said this — recorded in your transcript at claim time:"
+      : "skalpel re-ran the agent's OWN proof, out-of-band:",
+  );
   L.push(`  $ ${proof}`);
   L.push(`  → ${verdictPhrase(row)}`);
   if (evidence) {
@@ -167,8 +193,16 @@ export function renderPlainCard(row) {
     L.push(`  ${evidence}`);
   }
   L.push("");
-  L.push(`caught ${stamp(row.ts)}`);
-  L.push("The agent said done. Its own proof, re-run, disagreed.");
+  L.push(
+    retro
+      ? `recorded ${stamp(row.ts)} (your transcript, at claim time)`
+      : `caught ${stamp(row.ts)}`,
+  );
+  L.push(
+    retro
+      ? "The agent said done. Its own test, recorded at that moment, had already failed."
+      : "The agent said done. Its own proof, re-run, disagreed.",
+  );
   return L.join("\n");
 }
 
@@ -177,13 +211,18 @@ export function renderMarkdownCard(row) {
   const claim = clean(row.claim_text, 200) || "it's done";
   const proof = clean(row.proof_command || row.probe_command, 200) || "its own proof";
   const evidence = clean(row.evidence, 200);
+  const retro = isRetro(row);
   const FENCE = "```";
   const L = [];
-  L.push("**🔬 skalpel — verified catch**");
+  L.push(retro ? "**🔬 skalpel — recorded catch**" : "**🔬 skalpel — verified catch**");
   L.push("");
   L.push(`> The agent claimed: _"${claim}"_`);
   L.push("");
-  L.push("skalpel re-ran the agent's own proof, out-of-band:");
+  L.push(
+    retro
+      ? "Its own test had already failed when it said this — recorded in your transcript at claim time:"
+      : "skalpel re-ran the agent's own proof, out-of-band:",
+  );
   L.push(FENCE);
   L.push(`$ ${proof}`);
   L.push(`→ ${verdictPhrase(row)}`);
@@ -194,7 +233,11 @@ export function renderMarkdownCard(row) {
     L.push(evidence);
     L.push(FENCE);
   }
-  L.push(`_Caught ${stamp(row.ts)}. The agent said done; its own proof, re-run, disagreed._`);
+  L.push(
+    retro
+      ? `_Recorded ${stamp(row.ts)} in your transcript, at claim time. The agent said done; its own test had already failed._`
+      : `_Caught ${stamp(row.ts)}. The agent said done; its own proof, re-run, disagreed._`,
+  );
   return L.join("\n");
 }
 
@@ -203,14 +246,21 @@ function renderTerminalCard(row) {
   const claim = clean(row.claim_text, 200) || "it's done";
   const proof = clean(row.proof_command || row.probe_command, 200) || "its own proof";
   const evidence = clean(row.evidence, 180);
+  const retro = isRetro(row);
   const L = [];
   L.push("");
-  L.push(`  ${B}🔬 skalpel${X} ${D}·${X} ${B}${R}verified catch${X}`);
+  L.push(
+    `  ${B}🔬 skalpel${X} ${D}·${X} ${B}${R}${retro ? "recorded catch" : "verified catch"}${X}`,
+  );
   L.push(`  ${D}${"─".repeat(70)}${X}`);
   L.push(`  ${D}The agent claimed:${X}`);
   L.push(`    ${B}"${claim}"${X}`);
   L.push("");
-  L.push(`  ${D}skalpel re-ran the agent's OWN proof, out-of-band:${X}`);
+  L.push(
+    retro
+      ? `  ${D}Its own test had already failed when it said this — recorded at claim time:${X}`
+      : `  ${D}skalpel re-ran the agent's OWN proof, out-of-band:${X}`,
+  );
   L.push(`    ${D}$${X} ${proof}`);
   L.push(`    ${R}${B}→ ${verdictPhrase(row)}${X}`);
   if (evidence) {
@@ -219,7 +269,9 @@ function renderTerminalCard(row) {
   }
   L.push(`  ${D}${"─".repeat(70)}${X}`);
   L.push(
-    `  ${D}caught ${stamp(row.ts)} — the agent said done; its own proof, re-run, disagreed.${X}`,
+    retro
+      ? `  ${D}recorded ${stamp(row.ts)} — the agent said done; its own test had already failed.${X}`
+      : `  ${D}caught ${stamp(row.ts)} — the agent said done; its own proof, re-run, disagreed.${X}`,
   );
   L.push("");
   return L.join("\n");