cortad 0.3.1-rc.13 → 0.3.1-rc.15

package/lib/contact-text.mjs

@@ -29,7 +29,7 @@ function sessionLine(v) {
   if (!s || !(v.requests > 1)) return null;
   if (!s.held) return "No later request's prompt carried an earlier one.";
   const by = CARRIER[s.carrier] && s.key ? CARRIER[s.carrier](s.key) : "nothing the request sent, so every caller shares what the app keeps";
-  return `A later request's prompt carried an earlier one, tied by ${by}.`;
+  return `A later request's prompt carried an earlier one, held by ${by}.`;
 }
 
 function rulesLine(r) {
@@ -97,8 +97,12 @@ function readyLine(r) {
     ? `about ${plural(r.minutes, "minute")} if your app answers several requests at once, or up to about ${plural(one.minutes, "minute")}${one.trials < r.trials ? ` and ${plural(one.trials, "trial")}` : ""} if it answers one at a time`
     : `about ${plural(r.minutes, "minute")}`);
   const usd = (x) => `$${x.toFixed(2)}`;
-  const cost = has(r.usd) && `about ${usd(r.usd)}${has(r.usdHigh) ? ` to ${usd(r.usdHigh)}` : ""}${soFar} on your ${r.provider ? `${r.provider} ` : ""}key${r.model ? ` for ${r.model}` : ""}${r.estimated ? ", estimated from the length of the text since your provider sent no token counts" : ""}`;
-  const parts = [fit, has(r.replies) && `about ${plural(r.replies, "reply", "replies")}`, time, cost].filter(Boolean);
+  // Some endpoints had no published rate or token counts: the price covers the others and says so.
+  const some = r.pricedOn ? `at least ${usd(r.usd)}, priced on ${num(r.pricedOn.endpoints)} of ${plural(r.pricedOn.of, "endpoint")}` : null;
+  const cost = has(r.usd) && `${some ?? `about ${usd(r.usd)}${has(r.usdHigh) ? ` to ${usd(r.usdHigh)}` : ""}`}${soFar} on your ${r.provider ? `${r.provider} ` : ""}key${r.model ? ` for ${r.model}` : ""}${r.estimated ? ", estimated from the length of the text since your provider sent no token counts" : ""}`;
+  // Follow-ups are played only where something breaks: said apart, never in the count above.
+  const followUps = r.followUps ? `and up to ${plural(r.followUps, "follow-up")} where something breaks` : null;
+  const parts = [fit, followUps, has(r.replies) && `about ${plural(r.replies, "reply", "replies")}`, time, cost].filter(Boolean);
   const rules = !has(r.rules) ? "" : r.rules === 0 ? " No rules of your own were read from the code, so the run checks replies against generic standards." : ` Replies are checked against the ${plural(r.rules, "rule")} read from your code.`;
   return `A run on the proven endpoints: ${parts.join(", ")}.${r.reading ? " The read is still adding trials; the run waits for it and plays them all." : ""}${rules}`;
 }
@@ -114,6 +118,6 @@ export function contactLines(c, data = []) {
   if (missing.length) lines.push(`Endpoints the read found that no request has reached (${num(missing.length)}):`, ...missing.map((m) => `  ${m.method} ${m.path}${m.file ? `  ${at(m.file, m.line)}` : ""}${m.excluded ? `  ${m.excluded}` : ""}`));
   const ready = readyLine(c.ready);
   if (ready) lines.push(ready);
-  for (const b of c.ready?.blocked ?? []) lines.push(`Run waits${sideOf(b.side)}: ${b.door}: ${b.why}${b.question ? ` ${b.question}` : ""}`);
+  for (const b of c.ready?.blocked ?? []) lines.push(`Run is held${sideOf(b.side)}: ${b.door}: ${b.why}${b.question ? ` ${b.question}` : ""}`);
   return [...lines, ...said];
 }

package/lib/dissection-text.mjs

@@ -1,4 +1,4 @@
-import { clip, closeToCall, has, num, plural } from "./words.mjs";
+import { clip, has, num, plural } from "./words.mjs";
 
 // A finished run's dissection, as the server cut it (world-agent/dissection.ts): the run in a line,
 // what only running the app showed, what a code review would also have found, the gates a knock
@@ -167,7 +167,7 @@ function journeyLines(b, said) {
   const lines = [head, "How it went, at each conversation's last reply:"];
   const a = b.all;
   if (a?.got || a?.frustrated || a?.askedForPerson) {
-    const parts = [a.got && gotSaid(a.got, said), a.unsure && closeToCall(a.unsure, said), a.frustrated && `${num(a.frustrated.k)} grew frustrated`, a.askedForPerson && `${num(a.askedForPerson.k)} asked for a person midway`].filter(Boolean);
+    const parts = [a.got && gotSaid(a.got, said), a.unsure && `${num(a.unsure)} could not be told`, a.frustrated && `${num(a.frustrated.k)} grew frustrated`, a.askedForPerson && `${num(a.askedForPerson.k)} asked for a person midway`].filter(Boolean);
     lines.push(`  All ${plural(b.conversations, "conversation")}: ${parts.join("; ")}.`);
   }
   for (const r of j.rows.slice(0, JOURNEYS_SHOWN)) {
@@ -192,7 +192,7 @@ function heldEverywhereLines(b) {
 function notMeasuredLines(d, b) {
   const unplayed = b.journeys?.unplayed ?? [];
   const lines = [d.held, d.notMeasured, unplayed.length && `No conversation reached ${unplayed.length === 1 ? "this journey" : `these ${num(unplayed.length)} journeys`}: ${flat(unplayed.join(", "), 200)}.`].filter(Boolean);
-  return lines.length ? ["What the numbers leave out, and why:", ...lines.map((l) => `  ${l}`)] : [];
+  return lines.length ? ["What was not measured, and why:", ...lines.map((l) => `  ${l}`)] : [];
 }
 
 // `room.lines`: what the one-line findings may take; past it a finding that may lead, an open gate or

package/lib/failed-text.mjs

@@ -7,15 +7,15 @@ import { at, clip, plural } from "./words.mjs";
 
 // A row the run held back names a door and why, or, when no door could take it, only the reason.
 export const heldBackLine = (h) => h.why === "not testable here"
-  ? `Set aside: ${plural(h.trials, "trial")} cannot be played here: ${clip(h.surface, 200).replace(/\.$/, "")}.`
-  : `Set aside: ${plural(h.trials, "trial")} on ${h.surface}${h.why ? `, ${clip(h.why, 200)}` : ""}.`;
+  ? `Held back: ${plural(h.trials, "trial")} cannot be played here: ${clip(h.surface, 200).replace(/\.$/, "")}.`
+  : `Held back: ${plural(h.trials, "trial")} on ${h.surface}${h.why ? `, ${clip(h.why, 200)}` : ""}.`;
 
 const HELD_SHOWN = 6;
 
 export function heldLines(held) {
   const rows = (held ?? []).filter((h) => h.trials > 0);
   const lines = rows.slice(0, HELD_SHOWN).map(heldBackLine);
-  if (rows.length > HELD_SHOWN) lines.push(`Set aside: ${plural(rows.slice(HELD_SHOWN).reduce((n, h) => n + h.trials, 0), "more trial")} for ${plural(rows.length - HELD_SHOWN, "other reason")}.`);
+  if (rows.length > HELD_SHOWN) lines.push(`Held back: ${plural(rows.slice(HELD_SHOWN).reduce((n, h) => n + h.trials, 0), "more trial")} for ${plural(rows.length - HELD_SHOWN, "other reason")}.`);
   return lines;
 }
 

package/lib/keeping.mjs

@@ -42,7 +42,7 @@ export function makeKeeping({ root, work, ledgerFile, onPath, say, settings, lau
   let seen = 0;
   const told = new Set();
   const tell = (lines) => { for (const line of lines) say(lower(line)); };
-  const held = (blocked) => tell(blocked.map((b) => `Run waits until this is fixed: ${lower(b.why)}`));
+  const held = (blocked) => tell(blocked.map((b) => `Run is held until this is fixed: ${lower(b.why)}`));
 
   async function keepAll(started) {
     const { values, from, appDir, sources } = settings();

package/lib/mcp.mjs

@@ -38,7 +38,7 @@ export const TOOLS = [
   { name: "field_connect", title: "Connect production", inputSchema: none,
     description: "How production replies reach Cortad, so real conversations are read with the same checks. The ingest key is created by the owner in the browser." },
   { name: "field", title: "Production numbers", annotations: readOnly,
-    description: "Production in numbers: conversations read, checks passed, resolved, frustrated, asks for a human, and the rules broken most. Message text stays out of the answer.",
+    description: "Production in numbers: conversations read, checks held, resolved, frustrated, asks for a human, and the rules broken most. Message text stays out of the answer.",
     inputSchema: { type: "object", properties: { days: { type: "integer", minimum: 1, maximum: 180, description: "Window in days. Default 30." } }, additionalProperties: false } },
 ];
 const NAMES = new Set(TOOLS.map((t) => t.name));

package/lib/numbers-text.mjs

@@ -1,4 +1,4 @@
-import { clip, closeToCall, has, num, pack, pageOf, plural } from "./words.mjs";
+import { clip, has, num, pack, pageOf, plural } from "./words.mjs";
 
 // Every number the run measured, whole, one call away (`findings numbers`): each group with its count
 // by class, of how many, the interval on the share that matters, and one exchange from the worst
@@ -14,9 +14,9 @@ const SPLIT_WORDS = { journey: "By journey", segment: "By persona segment", situ
 
 // One group as a row under its header: the counts in the header's order, of how many, the average
 // on a scale, then the share named with its interval.
-function row(t, d, said) {
+function row(t, d) {
   const of = units(d.of, t.unit);
-  const unsure = d.unsure ? `; ${closeToCall(d.unsure, said)}` : "";
+  const unsure = d.unsure ? ` (${num(d.unsure)} unsure)` : "";
   const share = d.bad && d.of ? `, ${interval(d.bad)}` : "";
   if (t.kind === "question") return `${num(d.counts?.broke ?? 0)} of ${of}${share}${unsure}`;
   const counts = t.classes.map((c) => num(d.counts?.[c] ?? 0)).join(", ");
@@ -30,14 +30,13 @@ const headerOf = (t) => (t.kind === "question"
   : `  Counted as: ${isScale(t) ? `levels ${t.classes.join(", ")}` : t.classes.join(", ")} of ${t.unit}${t.badSaid ? `; then how many ${t.badSaid}, with the 95% interval` : ""}.`);
 
 function tableBlock(t) {
-  const said = {};
   const lines = [`${t.kind === "question" ? `${t.id}  ` : ""}${clip(t.title, 400)}${t.from ? ` (${t.from})` : ""}`];
   if (t.legend) lines.push(`  ${clip(t.legend, 500)}`);
-  lines.push(headerOf(t), `  All: ${row(t, t.all, said)}`);
+  lines.push(headerOf(t), `  All: ${row(t, t.all)}`);
   for (const split of ["journey", "segment", "situation", "reply"]) {
     const groups = t.by?.[split] ?? [];
     if (!groups.length || (split !== "reply" && groups.length === 1 && groups[0].of === t.all.of)) continue;
-    lines.push(`  ${SPLIT_WORDS[split]}:`, ...groups.map((d) => `    ${clip(d.value, 90)}${d.kept ? " (kept back)" : ""}: ${row(t, d, said)}`));
+    lines.push(`  ${SPLIT_WORDS[split]}:`, ...groups.map((d) => `    ${clip(d.value, 90)}${d.kept ? " (kept back)" : ""}: ${row(t, d)}`));
   }
   const w = t.worst;
   if (w) {

package/lib/proof.mjs

@@ -547,8 +547,12 @@ const sentence = ({ model, site }) => (model ? " The model said it."
 // The tools the request's last model call asked for, when they ran inside this request and no model
 // call came after them: the app closed the request before the model could answer from their result
 // (an AI SDK route left at one step). A tool the page runs itself and sends back is not this.
+// A last call that wrote words answered: its request carries the tool calls before it in its history
+// (MIT Learn AI's answering call resent search_courses), which is not the request stopping there.
 function endedOnTool(ex) {
-  const asked = [...new Set(arr(generation(ex).at(-1)?.called).map((c) => str(c?.name)).filter(Boolean))];
+  const last = generation(ex).at(-1);
+  if (str(last?.reply).trim()) return null;
+  const asked = [...new Set(arr(last?.called).map((c) => str(c?.name)).filter(Boolean))];
   if (!asked.length) return null;
   const inApp = new Set(ex.deps.flatMap((d) => arr(d.called).map((c) => str(c?.name))));
   const streamed = eventsOf(str(ex.reply)).events.some((e) => e && typeof e === "object" && /tool.*(?:output|result)/i.test(str(e.type)));

package/lib/pyhook/sitecustomize.py

@@ -171,7 +171,14 @@ def _install():
     # What the app's tools answered, as the prompt of the next model call carries them (chat tool
     # messages, responses function outputs, Anthropic tool_result blocks, Gemini functionResponse
     # parts): the material a reply's facts rest on. Bounded per call.
-    tool_text, tools_max = 3000, 12
+    # A tool's answer is kept to the cap the reader takes, and one longer says it was cut (wire.cjs toolText).
+    tool_text, tools_max = 12000, 12
+
+    def tool_text_row(name, full, **extra):
+        out = {"name": str(name or "")[:80], "text": full[:tool_text], **extra}
+        if len(full) > tool_text:
+            out["cut"] = True
+        return out
 
     def text_of(v):
         if isinstance(v, str):
@@ -214,9 +221,10 @@ def _install():
         out, names = [], {}
 
         def add(name, text):
-            t = text_of(text)[:tool_text]
+            full = text_of(text)
+            t = full[:tool_text]
             if t.strip() and len(out) < tools_max and all(o["text"] != t for o in out):
-                out.append({"name": str(name or "")[:80], "text": t})
+                out.append(tool_text_row(name, full))
 
         messages = items(body.get("messages"))
         for m in messages:
@@ -645,9 +653,10 @@ def _install():
         calls, names, answers = {}, {}, []
 
         def answer(name, said):
-            t = text_of(said)[:tool_text]
+            full = text_of(said)
+            t = full[:tool_text]
             if name and t.strip() and len(answers) < tools_max and not any(a["name"] == name and a["text"] == t for a in answers):
-                answers.append({"name": str(name)[:80], "text": t, "provider": True})
+                answers.append(tool_text_row(name, full, provider=True))
 
         for e in events:
             if not isinstance(e, dict):
@@ -1731,7 +1740,7 @@ def _install():
             text = "raised %s: %s" % (type(error).__name__, error) if error is not None else returned(value)
             row = {"at": int(time.time() * 1000), "host": "in-app", "status": 200,
                    "called": [{"name": name, "arguments": args_text(plain_of(named))}],
-                   "tools": [{"name": name, "text": str(text)[:tool_text]}]}
+                   "tools": [tool_text_row(name, str(text))]}
             row.update(inside_of(req))
             write({"dep": row})
         except Exception:
@@ -2448,7 +2457,7 @@ def _install():
             if held_here(url, method, lambda: auth_of(self, k)) and not reads_only(text(body, reply_max)):
                 port = held_port()
                 if port is None:
-                    raise module.ClientConnectionError("this write was kept from leaving and could not be answered")
+                    raise module.ClientConnectionError("this write was held and could not be answered")
                 parts = urlsplit(url)
                 n = str(next(held_sent))
                 held_waiting[n] = (url, method, ctx.get())

package/lib/read-text.mjs

@@ -13,7 +13,7 @@ const rulesHead = (rules) => {
   return `Rules in the code: ${num(rules.count)}${has(files) ? `, in ${plural(files, "file")}` : ""}.`;
 };
 // What the last run found about the rule beside it; a rule it never asked is not a rule that held.
-const measuredOf = (m) => (m === null ? "; asked of no reply yet" : m ? `; asked of ${plural(m.asked, "reply", "replies")}, passed in ${num(m.held)}${m.asked > m.held ? `, broke in ${num(m.asked - m.held)}` : ""}` : "");
+const measuredOf = (m) => (m === null ? "; not measured yet" : m ? `; asked of ${plural(m.asked, "reply", "replies")}, held in ${num(m.held)}${m.asked > m.held ? `, broke in ${num(m.asked - m.held)}` : ""}` : "");
 const ruleLine = (e) => `  ${e.path ? `${at(e.path, e.line)}  ` : ""}"${clip(e.text, 300)}"${"measured" in e ? measuredOf(e.measured) : ""}`;
 const headOf = (label, parts) => (parts.some(Boolean) ? `${label}: ${parts.filter(Boolean).join(", ")}.` : `${label}:`);
 const standardsHead = (m) => headOf("Engineering standards", [has(m.decided) && `${num(m.decided)} decided`, has(m.met) && `${num(m.met)} met`, has(m.missed) && `${num(m.missed)} missed${has(m.problems) ? `, from ${plural(m.problems, "problem")}` : ""}`]);
@@ -65,7 +65,7 @@ export function readLines(r) {
   const q = r.questions;
   if (q) lines.push(`Questions: ${num(q.total)}${has(q.everywhere) ? `; ${num(q.everywhere)} asked in every conversation` : ""}${has(q.placed) ? `, ${num(q.placed)} placed in the situations they fit` : ""}.`);
   const t = r.trials;
-  if (t) lines.push(trialsHead(t), ...(t.held ?? []).map((h) => `  ${h.surface}: ${plural(h.n, "trial")} set aside${heldWhy(h)}`));
+  if (t) lines.push(trialsHead(t), ...(t.held ?? []).map((h) => `  ${h.surface}: ${plural(h.n, "trial")} held back${heldWhy(h)}`));
   return lines;
 }
 
@@ -104,7 +104,7 @@ function sectionOf(r, show) {
     default: {
       const t = r.trials ?? {};
       const all = t.all ?? (t.held ?? []).map((h) => ({ surface: h.surface, held: h.n, why: h.why, side: h.side }));
-      const line = (x) => `  ${x.surface}: ${[has(x.written) && `${num(x.written)} written`, has(x.playable) && `${num(x.playable)} playable`, x.held && `${num(x.held)} set aside`].filter(Boolean).join(", ")}${x.held ? heldWhy(x) : ""}`;
+      const line = (x) => `  ${x.surface}: ${[has(x.written) && `${num(x.written)} written`, has(x.playable) && `${num(x.playable)} playable`, x.held && `${num(x.held)} held back`].filter(Boolean).join(", ")}${x.held ? heldWhy(x) : ""}`;
       return { head: trialsHead(t), lines: all.map(line), continued: "Trials, continued." };
     }
   }

package/lib/stick.mjs

@@ -74,7 +74,7 @@ function write(root, rel, change, done) {
   if (next.trim() === "") {
     rmSync(file);
     for (let dir = dirname(file); dir.startsWith(`${resolve(root)}${sep}`) && !readdirSync(dir).length; dir = dirname(dir)) rmdirSync(dir);
-    return `${rel}: removed, it had only what stick wrote`;
+    return `${rel}: removed, it held only what stick wrote`;
   }
   mkdirSync(dirname(file), { recursive: true });
   writeFileSync(file, next);

package/lib/stores.mjs

@@ -126,6 +126,8 @@ function fromParts(values) {
   for (const [p, hostName] of hosts) {
     // A host setting left blank is a store the app is set up not to use (QDRANT_HOST= beside QDRANT_PATH).
     if (hostName && !String(values[hostName] ?? "").trim()) continue;
+    // A host setting that holds a whole address (REDIS_HOST=redis://...) is fromUrl's, not a host.
+    if (hostName && String(values[hostName]).includes("://")) continue;
     const parts = {};
     for (const name of Object.keys(values)) {
       if (!name.startsWith(p) || name === hostName) continue;
@@ -235,9 +237,13 @@ const HELD = {
   bypassed: () => ["your app reached it anyway, since it loads its settings file over the settings it was started with", "Load that file without overriding what is already set"],
   other: (s) => ["the copy did not go through", `Point ${s.names[0]} at a database on this machine`],
 };
+// The other way out is the owner's yes on the card in the browser, which lets trials write into it as
+// it is: an agent told only to move the store spent ten minutes on a local Qdrant the card had offered
+// to do without (ulaim, 2026-10-01).
 const heldWhy = (s, why, tool) => {
   const [reason, fix] = (HELD[why] ?? HELD.other)(s, tool);
-  return `Trials would write into the ${whatOf(s)} that ${listed(s.names)} ${s.names.length === 1 ? "names" : "name"}, on a server off this machine, and ${reason}. ${fix}, then run the command again.`;
+  const yes = why === "started" ? "" : " Or the owner can say yes to it on the card in the browser, and trials write into it as it is.";
+  return `Trials would write into the ${whatOf(s)} that ${listed(s.names)} ${s.names.length === 1 ? "names" : "name"}, on a server off this machine, and ${reason}. ${fix}, then run the command again.${yes}`;
 };
 const hostedKept = (s, got) => (got.how === "neon"
   ? `Your app runs against a Neon branch of its ${whatOf(s)}, made for this session through ${listed(s.names)}; the original is not touched, and the branch is deleted when this command ends.`

package/lib/text.mjs

@@ -30,6 +30,8 @@ export const finished = (run) => Boolean(run) && (run.finished ?? DONE.has(run.s
 export function statusText(d) {
   const lines = [];
   if (d.repository?.name) lines.push(`Cortad · ${d.repository.name}`);
+  // What changed since the last status call, first: the read landing, endpoints proven, a run.
+  if (d.news?.length) lines.push("New since your last call:", ...d.news.map((n) => `  ${n}`));
   const plan = planLine(d.plan);
   if (plan) lines.push(plan);
   lines.push(...appLines(d));
@@ -86,7 +88,7 @@ function measuredLines(d) {
   if (of) lines.push(`${num(d.questionsAsked)} of ${plural(d.questionsOf, "check")} measured.`);
   // The dissection says what was not measured in its own line.
   const groups = d.dissection ? [] : (d.notMeasured?.groups ?? []).filter((g) => g.count > 0 && g.said);
-  if (groups.length) lines.push(`${plural(d.notMeasured.total, "check")} with no reply to read: ${groups.map((g) => clip(g.said, 240)).join("; ")}.`);
+  if (groups.length) lines.push(`${num(d.notMeasured.total)} not measured: ${groups.map((g) => clip(g.said, 240)).join("; ")}.`);
   // An answer from before failedReplies says its failures here, never folded into a number.
   if (!d.failedReplies && d.replies?.crashed > 0) lines.push(`${num(d.replies.crashed)} of ${plural(d.replies.total, "reply", "replies")} were failures and are not counted.`);
   // The server's own sentence already says the count.
@@ -171,7 +173,7 @@ function trialsLine(d) {
   const t = d.trials;
   // A verify replays the trials it names, so it has no written count of its own to add up.
   if (!t || !has(t.written) || d.kind === "verify") return null;
-  const parts = [`${num(t.chosen)} chosen for this run`, t.heldBack && `${num(t.heldBack)} set aside`, t.notPlanned && `${num(t.notPlanned)} not planned`].filter(Boolean);
+  const parts = [`${num(t.chosen)} chosen for this run`, t.heldBack && `${num(t.heldBack)} held back`, t.notPlanned && `${num(t.notPlanned)} not planned`].filter(Boolean);
   const rest = t.notPlayed ? `; ${num(t.played)} played, ${num(t.notPlayed)} ${finished(d) ? "not played" : "still to play"}` : "";
   return `Trials: ${num(t.written)} written: ${parts.join(", ")}${rest}.`;
 }
@@ -299,7 +301,7 @@ function ledgerLines(l) {
     for (const f of l.fixed) lines.push(`  ${f.path ? `${at(f.path, f.line)} (${f.findingId})` : f.findingId}: ${f.verdict}; ${num(f.after.failed)} of ${plural(f.after.replays, "replay")} failed, against ${num(f.before.failed)} of ${num(f.before.trials)} before${keptSaid(f.heldOut)}.`);
   }
   const n = l.nextRun;
-  if (n) lines.push(`The next run would play ${[plural(n.trials, "trial"), has(n.heldOut) && `${num(n.heldOut)} kept back`, has(n.newFromChanges) && `${num(n.newFromChanges)} new from the changes`].filter(Boolean).join(", ")}.`);
+  if (n) lines.push(`The next run would play ${[plural(n.trials, "trial"), has(n.heldOut) && `${num(n.heldOut)} held out`, has(n.newFromChanges) && `${num(n.newFromChanges)} new from the changes`].filter(Boolean).join(", ")}.`);
   const p = l.plan;
   if (p) lines.push(`The ${p.name} plan, $${p.monthlyUsd} a month${has(p.allowed) ? `, includes ${units(p.allowed, p.unit)}` : ""}.`);
   if (l.field === null) lines.push("Production: not connected.");
@@ -342,7 +344,7 @@ function findingsTail(d) {
   for (const s of [d.read, d.heldBack, d.findings?.length ? null : d.why]) if (s) lines.push(s);
   // A finding in a situation the run keeps back: what failed and where, so the rule can be fixed;
   // its inputs and trials stay unseen, and verify replays it all the same.
-  for (const k of d.keptBack ?? []) lines.push(`Kept back ${k.id}: "${clip(k.asks, 160)}"${k.file ? ` at ${at(k.file, k.line)}` : ""}, ${k.trials ? `failed in ${num(k.trials.failed)} of ${plural(k.trials.of, "trial")}` : `passed in ${kn(k.rate)}${has(k.reply) ? ` at reply ${num(k.reply)}` : ""}`}${k.quote ? `; the reply said "${clip(k.quote, 120)}"` : ""}. Its inputs are not shown; verify ${k.id} replays them.`);
+  for (const k of d.keptBack ?? []) lines.push(`Kept back ${k.id}: "${clip(k.asks, 160)}"${k.file ? ` at ${at(k.file, k.line)}` : ""}, ${k.trials ? `failed in ${num(k.trials.failed)} of ${plural(k.trials.of, "trial")}` : `held ${kn(k.rate)}${has(k.reply) ? ` at reply ${num(k.reply)}` : ""}`}${k.quote ? `; the reply said "${clip(k.quote, 120)}"` : ""}. Its inputs are not shown; verify ${k.id} replays them.`);
   return lines.join("\n");
 }
 
@@ -414,10 +416,10 @@ const pct = (x) => `${Math.round(x * 100)}%`;
 // whether the passages MISS the point left an agent unable to tell pass from fail.
 function shareLine(f) {
   const t = f.trials;
-  if (t) return `  Failed in ${num(t.failed)} of ${plural(t.of, "trial")}, between ${pct(t.lo)} and ${pct(t.hi)} of trials${f.unsettled ? "; not settled yet: its interval is still wider than one visit in five" : ""}.`;
+  if (t) return `  Failed in ${num(t.failed)} of ${plural(t.of, "trial")}, between ${pct(t.lo)} and ${pct(t.hi)} of trials${f.unsettled ? "; unsettled: too few trials yet to say it fails in one visit in five" : ""}.`;
   const r = f.rate;
   if (!r) return "";
-  return `  Broke in ${num(r.n - r.k)} of ${plural(r.n, "reply", "replies")}${r.n ? `, ${Math.round((100 * (r.n - r.k)) / r.n)}%` : ""}${has(r.lo) ? `, interval ${Math.round((1 - r.hi) * 100)}% to ${Math.round((1 - r.lo) * 100)}%` : ""}.${f.unsettled ? " Not settled yet: its interval is still wider than one visit in five." : ""}`;
+  return `  Broke in ${num(r.n - r.k)} of ${plural(r.n, "reply", "replies")}${r.n ? `, ${Math.round((100 * (r.n - r.k)) / r.n)}%` : ""}${has(r.lo) ? `, interval ${Math.round((1 - r.hi) * 100)}% to ${Math.round((1 - r.lo) * 100)}%` : ""}.${f.unsettled ? " Unsettled." : ""}`;
 }
 
 // One quoted break: what the trial sent, the reply, the sentence it broke on, what the reply was
@@ -429,8 +431,6 @@ function quoteLines(q, f, logs) {
   if (q.asked) lines.push(`  Sent: "${clip(q.asked, 300)}"`);
   lines.push(`  Reply ${q.reply}: "${clip(q.answer ?? q.quote, 400)}"${about ? ` (${about})` : ""}`);
   if (q.answer && q.quote && q.quote !== q.answer) lines.push(`    It broke on: "${clip(q.quote, 300)}"`);
-  // How sure the reading was that this reply broke the check, and who decided it.
-  if (has(q.p)) lines.push(`    ${num(Math.round(q.p * 100))} in 100 sure${q.decidedBy === "code" ? ", decided by code" : q.decidedBy === "text-judge" ? ", decided by the second judge" : ""}.`);
   if (q.after) lines.push(`    Then the customer: "${clip(q.after, 300)}"`);
   if (q.checked) lines.push(`    ${q.checked}`);
   for (const p of (q.passages ?? []).slice(0, 3)) lines.push(`    Passage${p.source ? ` (${clip(p.source, 60)})` : ""}: "${clip(p.text, 240)}"`);
@@ -472,11 +472,11 @@ export function fieldText(d) {
   if (!t || !t.n) return [`No production conversations read in the last ${plural(d.days, "day")}.`, ...link].join("\n");
   const lines = [
     `Production, last ${plural(d.days, "day")}: ${plural(t.n, "conversation")}, ${num(t.read)} read.`,
-    `Rule checks passed: ${pct(t.rulingsHeld, t.rulings)} of ${num(t.rulings)}${t.rulingsUnsure ? ` (${num(t.rulingsUnsure)} too close to call)` : ""}. Resolved ${pct(t.resolved, t.read)}, frustrated ${pct(t.frustrated, t.read)}, asked for a human ${pct(t.wantsHuman, t.read)}, unanswered ${pct(t.unanswered, t.read)}.`,
+    `Rule checks held: ${pct(t.rulingsHeld, t.rulings)} of ${num(t.rulings)}${t.rulingsUnsure ? ` (${num(t.rulingsUnsure)} unsure)` : ""}. Resolved ${pct(t.resolved, t.read)}, frustrated ${pct(t.frustrated, t.read)}, asked for a human ${pct(t.wantsHuman, t.read)}, unanswered ${pct(t.unanswered, t.read)}.`,
   ];
   const broke = (d.rules ?? []).filter((r) => r.broke > 0).sort((a, b) => b.broke - a.broke).slice(0, 5);
   if (broke.length) lines.push(`Rules broken most: ${broke.map((r) => `${r.id} (${num(r.broke)})`).join(", ")}.`);
-  if (d.journeys?.length) lines.push(`By journey: ${d.journeys.slice(0, 6).map((j) => `${j.value} ${plural(j.n, "conversation")}, ${pct(j.rulingsHeld, j.rulings)} passed`).join("; ")}.`);
+  if (d.journeys?.length) lines.push(`By journey: ${d.journeys.slice(0, 6).map((j) => `${j.value} ${plural(j.n, "conversation")}, ${pct(j.rulingsHeld, j.rulings)} held`).join("; ")}.`);
   return [...lines, ...link].join("\n");
 }
 

package/lib/trace.cjs

@@ -17,7 +17,7 @@ if (FILE) {
     const http = require("node:http");
     const https = require("node:https");
     const als = new AsyncLocalStorage();
-    const { MAX, REPLY_MAX, EMBEDDING, TOOL_TEXT, scrubbed, list, clipped, argsText, turnOf, isModelCall, isRetrieval, passagesFrom, decoded, cookieNames, makeRecorder } = require("./wire.cjs");
+    const { MAX, REPLY_MAX, EMBEDDING, TOOL_TEXT, toolText, scrubbed, list, clipped, argsText, turnOf, isModelCall, isRetrieval, passagesFrom, decoded, cookieNames, makeRecorder } = require("./wire.cjs");
     const row = (value) => { try { fs.appendFileSync(FILE, JSON.stringify(scrubbed(value)) + "\n", { mode: 0o600 }); } catch { /* the command is gone */ } };
     // What the app writes into its own folder while a run is on: each file copied once before its
     // first write, so the command puts every one back when the run ends (lib/writes.mjs). Trials
@@ -611,7 +611,7 @@ if (FILE) {
       args.forEach((a, i) => { named[IDENT.test(names[i] || "") ? names[i] : String(i)] = plainOf(a); });
       let text;
       try { text = error ? `raised ${(error && error.name) || "an error"}: ${(error && error.message) || String(error)}` : returned(value); } catch { text = "returned a value this hook cannot read"; }
-      row({ dep: { at: Date.now(), host: "in-app", status: 200, ...(ctx ? { ex: ctx.id } : {}), ...(turn ? { turn } : {}), called: [{ name, arguments: argsText(named) }], tools: [{ name, text: String(text).slice(0, TOOL_TEXT) }] } });
+      row({ dep: { at: Date.now(), host: "in-app", status: 200, ...(ctx ? { ex: ctx.id } : {}), ...(turn ? { turn } : {}), called: [{ name, arguments: argsText(named) }], tools: [toolText(name, String(text))] } });
     };
     const TOOL = Symbol.for("cortad.wrapped");
     const wrappers = new WeakMap();

package/lib/wire.cjs

@@ -55,7 +55,11 @@ const partsOf = (text) => text.split(/\{[^}]*\}|\$\{[^}]*\}|%[sd]|<[^>]{1,40}>/)
 // messages of a chat body, the function outputs of a responses body, the tool_result blocks of
 // an Anthropic one, the functionResponse parts of a Gemini one. They are the material a reply's
 // facts rest on, and without them a real ticket id read as invented. Bounded per call.
-const TOOL_TEXT = 3000, TOOLS_MAX = 12;
+// A tool's answer is kept to the cap the reader takes (12,000 characters), and one longer says it was
+// cut: MIT Learn AI's 8,000-character search results were kept at 3,000, and a podcast episode in
+// the part cut off was graded as made up. A cut answer stands the passage questions down.
+const TOOL_TEXT = 12000, TOOLS_MAX = 12;
+const toolText = (name, full, extra = {}) => ({ name: String(name || "").slice(0, 80), text: full.slice(0, TOOL_TEXT), ...extra, ...(full.length > TOOL_TEXT ? { cut: true } : {}) });
 const textOf = (v) => (typeof v === "string" ? v : Array.isArray(v) ? v.map((p) => (p && typeof p === "object" ? (p.text || p.content || "") : String(p || ""))).filter(Boolean).join("\n") : v && typeof v === "object" ? JSON.stringify(v) : "");
 // Only what came after the person's latest message belongs to this turn: a thread the app
 // resends whole carries every earlier turn's tool answers too.
@@ -71,7 +75,7 @@ const toolsIn = (sent) => {
   if (!body || typeof body !== "object") return undefined;
   const out = [];
   const names = new Map();
-  const add = (name, text) => { const t = textOf(text).slice(0, TOOL_TEXT); if (t.trim() && out.length < TOOLS_MAX && !out.some((o) => o.text === t)) out.push({ name: String(name || "").slice(0, 80), text: t }); };
+  const add = (name, text) => { const full = textOf(text); const t = full.slice(0, TOOL_TEXT); if (t.trim() && out.length < TOOLS_MAX && !out.some((o) => o.text === t)) out.push(toolText(name, full)); };
   for (const m of list(body.messages)) {
     for (const c of list(m.tool_calls)) if (c.id && c.function) names.set(c.id, c.function.name);
     for (const c of list(m.content)) if (c.type === "tool_use" && c.id) names.set(c.id, c.name);
@@ -323,7 +327,7 @@ const providerIn = (sent, events) => {
   const offered = list(body && body.tools).map((t) => t.type).filter((t) => typeof t === "string" && t !== "function" && t !== "custom");
   const named = (kind) => offered.find((t) => t.startsWith(kind)) || kind;
   const calls = new Map(), names = new Map(), answers = [];
-  const answer = (name, text) => { const t = textOf(text).slice(0, TOOL_TEXT); if (name && t.trim() && answers.length < TOOLS_MAX && !answers.some((a) => a.name === name && a.text === t)) answers.push({ name: String(name).slice(0, 80), text: t, provider: true }); };
+  const answer = (name, text) => { const full = textOf(text); const t = full.slice(0, TOOL_TEXT); if (name && t.trim() && answers.length < TOOLS_MAX && !answers.some((a) => a.name === name && a.text === t)) answers.push(toolText(name, full, { provider: true })); };
   for (const e of events) {
     if (!e || typeof e !== "object") continue;
     const order = [];
@@ -813,6 +817,6 @@ function makeRecorder({ row, rulesFile, onTurn = () => {}, onEnded = () => {} })
 
 module.exports = {
   MAX, REPLY_MAX, EMBEDDING, TOOL_TEXT, MODEL_HOST,
-  scrubbed, list, clipped, argsText, turnOf, isModelCall, isRetrieval, passagesFrom, decoded, cookieNames,
+  scrubbed, list, clipped, argsText, turnOf, toolText, isModelCall, isRetrieval, passagesFrom, decoded, cookieNames,
   makeRecorder,
 };

package/lib/words.mjs

@@ -12,8 +12,6 @@ export const has = (v) => v !== undefined && v !== null;
 export const num = (x) => (typeof x === "number" ? Math.round(x).toLocaleString("en-US") : String(x));
 export const plural = (n, one, many = `${one}s`) => `${num(n)} ${n === 1 ? one : many}`;
 export const upper = (s) => `${s.charAt(0).toUpperCase()}${s.slice(1)}`;
-// A reading between 35 and 85 in 100 is too close to call; a block says that band the first time.
-export const closeToCall = (n, said = {}) => { const band = said.close ? "" : " (between 35 and 85 in 100)"; said.close = true; return `${num(n)} too close to call${band}`; };
 // A line past the end of the file as it stands on this machine is not an address: the file moved
 // on since it was read, and an agent was sent to line 204 of a 172-line file. The file stays.
 // Read each time, never kept: the agent edits these files while this server runs. Only a path

package/local.mjs

@@ -12,6 +12,7 @@
 // on with the key this connect leaves in ~/.cortad. Nothing here touches git. Your environment
 // never leaves this machine. Ctrl-C ends everything.
 
+import { storesOf } from "./lib/stores.mjs";
 import { notAnApp, withoutStores } from "./lib/app-port.mjs";
 import { holdsKeys, secretEnvValues } from "./lib/keys.mjs";
 import { spawn, execFile, execFileSync } from "node:child_process";
@@ -508,6 +509,8 @@ async function verb(job) {
     }
     // The person changed what they said yes to while this command runs.
     case "consent": await allowOutbound(b); return { ok: true };
+    // A line for the agent reading this terminal: the read of the code landing, said the moment it does.
+    case "say": { const line = String(b.line ?? "").replace(/\s+/g, " ").trim().slice(0, 400); if (line) say(line); return { ok: true }; }
     case "restart": return restartApp();
     case "inventory": return inventoryOf(b.probe && typeof b.probe === "object" ? b.probe : {});
     // Your own pages, read here rather than in a world's shell: that shell is sealed away from
@@ -661,7 +664,7 @@ const ownFolder = (dir) => { try { const at = realpathSync(dir); return at === r
 // carries it to the verbs, and the screen gets the app's own words beside it.
 function whyNot(up) {
   const h = up.held;
-  if (h) return `port ${h.port} is in use by another program (pid ${h.pid}${h.command ? `, ${h.command}` : ""}), so your app cannot listen there${up.fixed ? "; your app sets that port in its own code, so it cannot be moved" : ""}. Stop that program, then save a file here: your app is started again by itself.`;
+  if (h) return `port ${h.port} is held by another program (pid ${h.pid}${h.command ? `, ${h.command}` : ""}), so your app cannot listen there${up.fixed ? "; your app sets that port in its own code, so it cannot be moved" : ""}. Stop that program, then save a file here: your app is started again by itself.`;
   const wrongNode = pinned.major && !pinned.bin ? ` This project pins Node ${pinned.major} and this shell runs Node ${shellNode}: switch to ${pinned.major}.` : "";
   return `${up.exited !== null ? "your app stopped before it answered" : "your app did not answer within three minutes"}.${wrongNode} Fix it and save a file here: your app is started again by itself.`;
 }
@@ -861,7 +864,10 @@ async function start(waitMs, tries = 3) {
   try { markRun(writesDirOf(homeOf(projectOf(root))), "session", { keep: true }); } catch { /* the app's writes go unrecorded */ }
   // Taken before the app reads a file: a save after this moment is a change it has not loaded.
   const note = { own: true, at: Date.now(), reloads: RELOADER.test(commandText(cmd, appDir)), answered: false };
-  note.files = sourceOf(root, files);
+  // The settings files count as what it loaded: an env file is rarely tracked, and an agent's edit to
+  // one left the app on the old settings until the command was started again (ulaim, 2026-10-01).
+  note.files = { ...sourceOf(root, files), ...sourceOf(root, envFiles.filter((f) => !/\.(example|sample)$/.test(f)).map((f) => relative(root, f))) };
+  storesAt ??= storesNamed();
   loaded = note;
   const copies = await keeping.env(true);
   await allowOutbound();
@@ -957,10 +963,21 @@ const savedSinceStart = () => Boolean(loaded?.files) && changesOf({
 // Every save that changes what your app runs starts it again, when it does not reload by itself: an
 // agent that saved a fix and asked the app eight seconds later was answered by the old code. Looked
 // at again after each restart, for a save made while it was under way.
+// The stores the settings name, as copies are made of them: a change here is a change of what the app
+// writes into, so its copies and holds are made again for the new settings, and said again.
+const storesNamed = () => { try { return JSON.stringify(storesOf({ ...envExports(envFiles), ...process.env }).map((s) => [s.engine, s.host, s.port, s.db, s.names])); } catch { return ""; } };
+let storesAt = null;
 async function restartOnSave() {
   while (!closing) {
     await sourceChanged();
-    while (!closing && launched && child && !appGone && !loaded?.reloads && savedSinceStart()) await restartApp();
+    while (!closing && launched && child && !appGone && !loaded?.reloads && savedSinceStart()) {
+      const named = storesNamed();
+      const moved = storesAt !== null && named !== storesAt;
+      storesAt = named;
+      if (moved) await keeping.forget().catch(() => {});
+      await restartApp(moved ? "starting your app again on the settings you changed, with its data copied again for them" : undefined);
+      if (moved) await announce().catch(() => null);
+    }
   }
 }
 async function restartNow() {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cortad",
-  "version": "0.3.1-rc.13",
+  "version": "0.3.1-rc.15",
   "description": "Connects the AI app on your machine to Cortad for test conversations, and gives your coding agent the MCP and skill to run them. No dependencies.",
   "bin": {
     "cortad": "local.mjs"

package/skill/SKILL.md

@@ -25,7 +25,7 @@ The connect command ends with "Go back to the browser".
 
 1. The connect command says whether it started the app and the port to send requests to. If it says the app did not start, fix what it names and save; it starts the app again by itself. Requests are seen in an app the command started, or at the port it says to send them to.
 2. Send one real request to each endpoint that reaches the model, the way the app's own client does: its route, its body, its sign-in. On a chat, send a second message in the same conversation. When `status` asks for a request signed in as another customer, send it: each customer you sign in as becomes an account of its own in the run.
-3. `status` then shows what Cortad read (rules, journeys, simulated users, endpoints, standards) and each endpoint your requests reached with what it did inside the app: the model and how many calls a request made, the seconds a reply took, which of the rules read from the code the prompts carried, the tools that ran, the passages handed to the model, and problems a line of code decides, at their file and line. Endpoints the read found that no request reached are listed apart. The line `A run on the proven endpoints:` says how many conversations, about how long and about what it costs on the app's own key; any `Run waits` line says whose side it is on.
+3. `status` then shows what Cortad read (rules, journeys, simulated users, endpoints, standards) and each endpoint your requests reached with what it did inside the app: the model and how many calls a request made, the seconds a reply took, which of the rules read from the code the prompts carried, the tools that ran, the passages handed to the model, and problems a line of code decides, at their file and line. Endpoints the read found that no request reached are listed apart. The line `A run on the proven endpoints:` says how many conversations, about how long and about what it costs on the app's own key; any `Run is held` line says whose side it is on.
 4. The first run is free. Before it, read `status` for what would make the run measure the setup instead of the app:
    - a service down: a `Problem` line where a retrieval got no answer, or a tool came back as an error;
    - a request that ended on a tool call: `The request ended right after ... ran: no model call came after it`, which every conversation on that endpoint would repeat;
@@ -41,7 +41,7 @@ The connect command ends with "Go back to the browser".
 
 ## The first run: the baseline
 
-The first finished run on a repository prints its baseline, in this order: one line on what ran; who the users are, as the read found them in the code, with how many of each one's conversations got what they came for; the journeys and how each went, with the checks that failed most there; the checks no conversation broke; what broke, found only by running the app, each with its file and line; what a code review would also find; what the numbers leave out and why; one line on production.
+The first finished run on a repository prints its baseline, in this order: one line on what ran; who the users are, as the read found them in the code, with how many of each one's conversations got what they came for; the journeys and how each went, with the checks that failed most there; the checks no conversation broke; what broke, found only by running the app, each with its file and line; what a code review would also find; what was not measured and why; one line on production.
 
 How to read it:
 
@@ -49,8 +49,8 @@ How to read it:
 - "Got what they came for" is read at each conversation's last reply, in full or in part.
 - What broke was found by simulated customers who write the way the read says that user writes and answer each reply as it comes. Each finding carries its reach, the exchange (`Sent`, `Broke on`, `Then`) and where the fix belongs.
 - A hand-written script sends the requests someone wrote and checks the strings they expected. The exchange shows whether a finding took a customer who pressed, came back or wrote their own way; say which findings a script would also have caught.
-- What passed is measured too: a check no conversation broke, asked often enough that its interval tops out under one in five, leaving out any that names a tool no conversation ran.
-- "What the numbers leave out, and why" is what this run says nothing about.
+- What held is measured too: a check no conversation broke, asked often enough that its interval tops out under one in five, leaving out any that names a tool no conversation ran.
+- "What was not measured, and why" is what this run says nothing about.
 - Some conversations are kept back from you; a verify replays them too, so a fix is graded on conversations you have not read.
 
 Tell the person what the run found and what it did not, with the numbers as printed.
@@ -78,7 +78,7 @@ The trials, checks, seeds and the conversations kept back belong to Cortad, and
 After the fixes, `run` again. It plays what your changes since the last full run reach: those conversations again, each paired with its earlier self, and 8 to 12 new ones written for the change; it says what it left out. Its text leads with what moved: each behavior's failing count before and now, whether that is beyond chance, and the conversation that turned. Worse comes first.
 
 
-Production comes up after the first run. A verify whose failure is gone, and gone or less often on the conversations kept back from you, while production is not connected, ends with a "For the person" line that offers it; the plan line after a run names the production replies the plan above reads, and a spent plan's answer says whether production is connected. `field_connect` gives the steps; the owner creates the key in the browser. `field` gives the numbers: conversations read, checks passed, the rules broken most. Message text stays out.
+Production comes up after the first run. A verify whose failure is gone, and gone or less often on the conversations kept back from you, while production is not connected, ends with a "For the person" line that offers it; the plan line after a run names the production replies the plan above reads, and a spent plan's answer says whether production is connected. `field_connect` gives the steps; the owner creates the key in the browser. `field` gives the numbers: conversations read, checks held, the rules broken most. Message text stays out.
 
 ## Making it stick
 

package/skill/references/results.md

@@ -57,7 +57,7 @@ Engineering standards: 38 decided, 35 met, 3 missed.
   apps/api/src/tools/search.ts:51  Tool errors reach the model as errors: search returns an empty list when the index is down (decided by a model)
 Questions: 112; 40 asked in every conversation, 72 placed in the situations they fit.
 Trials: 58 written.
-  POST /api/homework/upload: 7 trials set aside, on the app's side: the route needs a signed file URL, and no test account can make one
+  POST /api/homework/upload: 7 trials held back, on the app's side: the route needs a signed file URL, and no test account can make one
 Latest run 8f2a1c4e-5b6d-4e7f-9a0b-1c2d3e4f5a6b: running, 14 of 51 trials played.
 Production: not connected.
 For the person: https://cortad.com/lab shows this in the browser.
@@ -109,7 +109,7 @@ Last run 8f2a1c4e-5b6d-4e7f-9a0b-1c2d3e4f5a6b: 3 findings, 51 of 51 trials playe
 2 fixes verified since that run:
   apps/api/src/agent/prompt.ts:41 (finding:1): gone; 0 of 3 replays failed, against 3 of 3 before; on the conversations kept back from you, 1 of 2 failed before, 0 of 2 after, not beyond chance.
   apps/api/src/agent/system.ts:12 (finding:2): gone; 0 of 7 replays failed, against 4 of 4 before; the fix made the failure worse on the conversations kept back from you (0 of 8 failed before, 6 of 8 after).
-The next run would play 58 trials, 7 kept back, 5 new from the changes.
+The next run would play 58 trials, 7 held out, 5 new from the changes.
 The Hobby plan, $99 a month, includes 10 runs.
 Production: not connected.
 Nothing ran.
@@ -118,14 +118,14 @@ For the person: plans and checkout at https://cortad.com/pricing?checkout=ship
 
 ## run_status
 
-Playing. The call waited up to 45 seconds:
+Playing. The call held up to 45 seconds:
 
 ```
 Run 8f2a1c4e-5b6d-4e7f-9a0b-1c2d3e4f5a6b: running, 14 of 51 trials played.
 next: run_status 8f2a1c4e-5b6d-4e7f-9a0b-1c2d3e4f5a6b
 ```
 
-Finished, the first run on a repository: its baseline. One line on what ran, then who the users are as the read found them in the code, with how many of each one's conversations got what they came for, in full or in part, read at the last reply; the journeys the read named and how each went, with the one or two checks that failed most there; the checks no conversation broke, asked often enough that each fails less than one time in five; then what broke, found only by running the app, each finding with its reach in conversations, the line of the app that wrote the breaking reply, the tools behind it, the layer a fix belongs in (prompt, flow around the model, tool or routing) and the exchange it broke on; the gates a knock found open; what a code review would also have found; what the numbers leave out; the run's model, time and spend; and one line on production. Every rate carries its 95% interval, said in words the first time. The findings take the room the picture leaves, the rest one line each, and findings shows them whole.
+Finished, the first run on a repository: its baseline. One line on what ran, then who the users are as the read found them in the code, with how many of each one's conversations got what they came for, in full or in part, read at the last reply; the journeys the read named and how each went, with the one or two checks that failed most there; the checks no conversation broke, asked often enough that each fails less than one time in five; then what broke, found only by running the app, each finding with its reach in conversations, the line of the app that wrote the breaking reply, the tools behind it, the layer a fix belongs in (prompt, flow around the model, tool or routing) and the exchange it broke on; the gates a knock found open; what a code review would also have found; what was held back and not measured; the run's model, time and spend; and one line on production. Every rate carries its 95% interval, said in words the first time. The findings take the room the picture leaves, the rest one line each, and findings shows them whole.
 
 A finding that sends customers away counts, for one journey or for everything else customers asked at one endpoint, the conversations whose last reply declined them or pointed them somewhere else, or where they asked for a person or said the app could not help and were not handed on; a conversation that ended with the customer served is never counted. It names the line of the prompt, or the tool those asks needed that the call writing the reply was given and never ran, that at least two and at least half of the replies read point at, reading at most 8 of them, and says where the fix belongs is not known when neither holds; it shows what the customer said next, and stays unconfirmed until its conversations were checked a second time. The same cut comes as structured data beside the text:
 
@@ -175,9 +175,9 @@ access:1  GET /api/admin/students answered 200 to a request with no sign-in; it
   Harm: data exposure.
   At apps/api/src/routes/admin.ts:14
   verify access:1 knocks the routes again.
-What the numbers leave out, and why:
-  Set aside: 7 conversations (POST /api/homework/upload: the route needs a signed file URL, and no test account can make one).
-  28 checks with no reply to read: 16 need a conversation past the first reply; 8 never met their condition (the parent asks what the plan costs); 4 have no conversation yet.
+What was not measured, and why:
+  Held: 7 conversations held back (POST /api/homework/upload: the route needs a signed file URL, and no test account can make one).
+  28 checks not measured: 16 need a conversation past the first reply; 8 never met their condition (the parent asks what the plan costs); 4 have no conversation yet.
   No conversation reached this journey: account recovery.
 Your app answered on gpt-4o-mini. The run took 7 min 32 s and spent $0.41 on your key over 214 model calls.
 Replies took 3.1 s at the median and 11.4 s at the slowest, over 138 replies.
@@ -343,7 +343,7 @@ finding:2  Is the reply written in a language other than the one the student wro
   A verify replays its 4 failing conversations. 2 clean replays show it gone.
 
 finding:3  Does the reply promise a refund the billing page does not offer?
-  Failed in 2 of 7 trials, between 8% and 64% of trials; not settled yet: its interval is still wider than one visit in five.
+  Failed in 2 of 7 trials, between 8% and 64% of trials; unsettled: too few trials yet to say it fails in one visit in five.
   Sent: "We are on the paid plan. What happens if we cancel?"
   Reply 1: "You can get a full refund any time in the first 60 days." (trial 77a0c5d1)
   At apps/api/src/agent/prompt.ts:44
@@ -357,7 +357,7 @@ finding:3  Does the reply promise a refund the billing page does not offer?
   "Never state a refund policy the product does not publish." (apps/api/src/agent/prompt.ts:41): broke in 9 of 12 trials.
   "Answer in the language the student writes in." (apps/api/src/agent/system.ts:12): broke in 2 of 26 trials.
 112 of 140 checks measured.
-28 checks with no reply to read: 16 need a conversation past the first reply; 8 never met their condition (the parent asks what the plan costs); 4 have no conversation yet.
+28 not measured: 16 need a conversation past the first reply; 8 never met their condition (the parent asks what the plan costs); 4 have no conversation yet.
 3 findings stand in the 12 situations you can read.
 2 findings stand in 3 situations kept back from you. You cannot read them, and a fix is graded on those too.
 ```
@@ -408,8 +408,8 @@ Production is not connected.
 
 ```
 Production, last 30 days: 4,812 conversations, 4,790 read.
-Rule checks passed: 93% of 61,204 (1,120 too close to call). Resolved 71%, frustrated 6%, asked for a human 2%, unanswered 4%.
+Rule checks held: 93% of 61,204 (1,120 unsure). Resolved 71%, frustrated 6%, asked for a human 2%, unanswered 4%.
 Rules broken most: rule:answer-first (412), rule:language (188), rule:cite-source (97).
-By journey: homework help 3,102 conversations, 94% passed; billing question 410 conversations, 88% passed.
+By journey: homework help 3,102 conversations, 94% held; billing question 410 conversations, 88% held.
 For the person: https://cortad.com/lab#field
 ```