tryharness 0.7.22 → 0.7.25

package/lib/provenance.mjs

@@ -6,12 +6,27 @@
 
 const USER_WRITERS = new Set(["user", "harness"]);
 
-export function provenanceFor(originAgent, parent = null, claimed = null, { personSession = false } = {}) {
+export function provenanceFor(
+  originAgent,
+  parent = null,
+  claimed = null,
+  { personSession = false } = {}
+) {
   // A claim counts only when it tightens: an 'agent' label and its chain. A claim of 'user' adds nothing.
   const taint = claimed?.trust === "agent" ? claimed : null;
-  const chain = [...(parent?.chain ?? (Array.isArray(taint?.chain) ? taint.chain.map(String) : [])), originAgent].slice(-16);
-  const tainted = parent?.trust === "agent" || Boolean(taint) || (!USER_WRITERS.has(originAgent) && !personSession);
-  return { trust: originAgent === "user" || !tainted ? "user" : "agent", chain };
+  const chain = [
+    ...(parent?.chain ??
+      (Array.isArray(taint?.chain) ? taint.chain.map(String) : [])),
+    originAgent,
+  ].slice(-16);
+  const tainted =
+    parent?.trust === "agent" ||
+    Boolean(taint) ||
+    (!USER_WRITERS.has(originAgent) && !personSession);
+  return {
+    trust: originAgent === "user" || !tainted ? "user" : "agent",
+    chain,
+  };
 }
 
 // What an agent reads before an agent-written brief: who wrote it, mechanically, every time.

package/lib/receipts.mjs

@@ -0,0 +1,82 @@
+import { insideRoot, isTestCommand } from "./trails.mjs";
+
+// Receipts for agents' work: what a run's own session file shows it did after its last change, never what it says it
+// did. A run that changed files and then ran its tests with a zero exit is "tested"; one whose last test after its last
+// change failed is "tests failing"; one that changed files and ran no test after the change is "not tested after its
+// last change". The words come from code over the transcript, so an agent cannot talk its way to "tested". Changes are
+// code changes inside the run's folder: notes, docs and files elsewhere (an agent's own memory) do not need tests.
+
+const DOC = /\.(?:md|mdx|txt|rst)$/i;
+
+/** { verdict, changed, lastChangeAt, tests: [{ command, ok, at, afterChange }] } for a trail (trails.mjs). */
+export function receipt(trail, { root = trail?.cwd } = {}) {
+  const calls = trail?.calls ?? [];
+  const code = (call) =>
+    call.edited
+      .map((token) => (root ? insideRoot(token, root, trail.cwd) : token))
+      .filter((file) => file && !DOC.test(file));
+  let lastChange = -1;
+  calls.forEach((call, i) => {
+    if (code(call).length) lastChange = i;
+  });
+  const changed = [...new Set(calls.flatMap(code))];
+  const tests = calls
+    .map((call, i) => ({ call, i }))
+    .filter(({ call }) => call.command && isTestCommand(call.command))
+    .map(({ call, i }) => ({
+      command: call.command.split("\n")[0].slice(0, 200),
+      ok: !call.error,
+      at: call.at,
+      afterChange: i > lastChange,
+    }));
+  const after = tests.filter((test) => test.afterChange);
+  const verdict = !changed.length
+    ? "no changes"
+    : !after.length
+      ? "not tested after its last change"
+      : after.at(-1).ok
+        ? "tested"
+        : "tests failing";
+  // The same test failing again and again is the clearest sign of a run going round in circles.
+  let repeated = 0;
+  let streak = 0;
+  tests.forEach((test, i) => {
+    streak =
+      !test.ok &&
+      i > 0 &&
+      !tests[i - 1].ok &&
+      tests[i - 1].command === test.command
+        ? streak + 1
+        : test.ok
+          ? 0
+          : 1;
+    repeated = Math.max(repeated, streak);
+  });
+  return {
+    verdict,
+    changed,
+    lastChangeAt: lastChange >= 0 ? calls[lastChange].at : null,
+    tests: after.length ? after : tests.slice(-3),
+    repeatedFailures: repeated,
+  };
+}
+
+/** The line the thread shows for a receipt. */
+export function receiptLine(agentName, result) {
+  const line = verdictLine(agentName, result);
+  return result.repeatedFailures >= 3
+    ? `${line} The same test failed ${result.repeatedFailures} times in a row during the run.`
+    : line;
+}
+
+function verdictLine(agentName, result) {
+  const files = `${result.changed.length} file${result.changed.length === 1 ? "" : "s"}`;
+  if (result.verdict === "no changes")
+    return `Receipt: ${agentName} changed no files in this run.`;
+  const last = result.tests.filter((test) => test.afterChange).at(-1);
+  if (result.verdict === "tested")
+    return `Receipt: ${agentName} changed ${files}, then ran \`${last.command}\` and it passed.`;
+  if (result.verdict === "tests failing")
+    return `Receipt: ${agentName} changed ${files}; its last test after the change, \`${last.command}\`, failed.`;
+  return `Receipt: ${agentName} changed ${files} and ran no tests after its last change.`;
+}