@hawkeyexl/manni 4.5.0 → 4.7.0

package/dist/chunk-7EL5LLKD.js

@@ -0,0 +1,233 @@
+#!/usr/bin/env node
+import {
+  pathKey
+} from "./chunk-D6RO54QD.js";
+import {
+  DEFAULT_CAPTURE_DIR,
+  makeRedactor
+} from "./chunk-CU57HT2E.js";
+
+// src/tracevals/rules/ledger.ts
+import { createHash } from "node:crypto";
+import { mkdir, readFile, rename, rm, writeFile } from "node:fs/promises";
+import { dirname, isAbsolute, join, relative, resolve, sep } from "node:path";
+var LEDGER_ENTRIES = 10;
+var HISTORY_ENTRIES = 5;
+var NOTE_CHARS = 200;
+var FACT_TURNS = 30;
+var COMMAND_CHARS = 120;
+var LIMIT = { commands: 10, wrote: 10, read: 5, skills: 10, agents: 10 };
+var GAP_RUNS = 3;
+var sha256 = (s) => createHash("sha256").update(s).digest("hex");
+function idHash(id) {
+  return sha256(id).slice(0, 32);
+}
+function ledgerPath(projectDir, sessionId, agentId) {
+  const agent = agentId === null ? "" : `.${idHash(agentId)}`;
+  return join(projectDir, DEFAULT_CAPTURE_DIR, `${idHash(sessionId)}${agent}.ledger.json`);
+}
+var emptyLedger = () => ({ version: 1, turns: [], rules: {} });
+var OUTCOMES = /* @__PURE__ */ new Set(["followed", "broken", "needs-review", "not-applicable"]);
+function isEntry(v) {
+  if (typeof v !== "object" || v === null) return false;
+  const e = v;
+  return typeof e.turn === "number" && OUTCOMES.has(e.outcome) && typeof e.note === "string";
+}
+var isStrings = (v) => Array.isArray(v) && v.every((s) => typeof s === "string");
+var FACT_LISTS = ["sources", "commands", "wrote", "read", "skills", "agents"];
+function isFacts(v) {
+  if (typeof v !== "object" || v === null) return false;
+  const f = v;
+  return typeof f.turn === "number" && typeof f.inScope === "boolean" && FACT_LISTS.every((k) => isStrings(f[k]));
+}
+async function readLedger(path) {
+  let parsed;
+  try {
+    parsed = JSON.parse(await readFile(path, "utf-8"));
+  } catch {
+    return emptyLedger();
+  }
+  if (typeof parsed !== "object" || parsed === null) return emptyLedger();
+  const { version, turns, rules } = parsed;
+  if (version !== 1 || typeof rules !== "object" || rules === null) return emptyLedger();
+  const ledger = emptyLedger();
+  if (Array.isArray(turns)) ledger.turns = turns.filter(isFacts);
+  for (const [key, value] of Object.entries(rules)) {
+    if (typeof value !== "object" || value === null) continue;
+    const { text, entries } = value;
+    if (typeof text !== "string" || !Array.isArray(entries)) continue;
+    ledger.rules[key] = { text, entries: entries.filter(isEntry) };
+  }
+  return ledger;
+}
+var sourceOf = (key) => key.slice(0, key.lastIndexOf("#"));
+function keyEntries(entries) {
+  const broken = entries.filter((e) => e.outcome === "broken").reverse();
+  const picked = [entries.at(-1), entries.find((e) => e.outcome === "followed"), ...broken];
+  return [...new Set(picked.filter((e) => e !== void 0))].slice(0, HISTORY_ENTRIES).sort((a, b) => a.turn - b.turn);
+}
+function gaps(turns, source) {
+  const runs = [];
+  let open;
+  for (const t of turns) {
+    if (t.sources.includes(source)) {
+      open = void 0;
+    } else if (open !== void 0) {
+      open.to = t.turn;
+    } else {
+      open = { from: t.turn, to: t.turn };
+      runs.push(open);
+    }
+  }
+  return runs.slice(-GAP_RUNS);
+}
+function historyBlock(ledger, key, text, before) {
+  const rule = ledger.rules[key];
+  if (rule?.text !== sha256(text)) return "";
+  const shown = keyEntries(
+    rule.entries.filter((e) => e.turn < before && e.outcome !== "not-applicable")
+  );
+  if (shown.length === 0) return "";
+  const entries = shown.map((e) => ({
+    turn: e.turn,
+    line: `- turn ${String(e.turn)}: ${e.outcome}.${e.note !== "" ? ` ${e.note}` : ""}`
+  }));
+  const holes = gaps(
+    ledger.turns.filter((t) => t.turn < before),
+    sourceOf(key)
+  ).map((g) => ({
+    turn: g.from,
+    line: g.from === g.to ? `- (turn ${String(g.from)}: not in scope)` : `- (turns ${String(g.from)}\u2013${String(g.to)}: not in scope)`
+  }));
+  const lines = [...entries, ...holes].sort((a, b) => a.turn - b.turn).map((l) => l.line);
+  return `# This rule earlier in this session
+
+${lines.join("\n")}`;
+}
+var posix = (p) => p.split(sep).join("/");
+function projectPath(path, cwd, root) {
+  const abs = resolve(cwd, path);
+  const r = relative(root, abs);
+  return posix(r === "" || r.startsWith("..") || isAbsolute(r) ? abs : r);
+}
+var clipTo = (s, max) => s.length <= max ? s : `${s.slice(0, max - 1)}\u2026`;
+function turnFacts(trace, turn, sources, opts) {
+  const w = turn.window;
+  const scrub = makeRedactor(opts.redact);
+  const take = (items, n = Number.POSITIVE_INFINITY) => [...new Set(items)].slice(0, n);
+  const clip = (s) => clipTo(s, COMMAND_CHARS);
+  const rel = (path) => projectPath(path, opts.cwd, opts.root);
+  const commands = w.toolCalls.flatMap(
+    (c) => c.name === "Bash" && typeof c.input.command === "string" ? [clip(scrub(c.input.command.replace(/\s+/g, " ").trim()))] : []
+  );
+  const reads = new Set(
+    w.fileAccesses.filter((a) => a.op === "read").map((a) => pathKey(resolve(opts.cwd, a.path)))
+  );
+  return {
+    turn: turn.from,
+    inScope: sources.length > 0,
+    sources: take(sources.map((s) => s.displayPath)),
+    commands: take(commands, LIMIT.commands),
+    wrote: take(
+      w.fileAccesses.filter((a) => a.op !== "read").map((a) => scrub(rel(a.path))),
+      LIMIT.wrote
+    ),
+    read: take(
+      // A source with no file (the typed prompts, an inline plan) has path "",
+      // which would resolve to the cwd; it is never read.
+      sources.filter((s) => s.path !== "" && reads.has(pathKey(s.path))).map((s) => s.displayPath),
+      LIMIT.read
+    ),
+    skills: take(
+      w.skillInvocations.map((s) => s.name),
+      LIMIT.skills
+    ),
+    agents: w.empty ? [] : take(
+      trace.agentSpawns.filter((a) => a.index >= turn.from && a.index <= turn.to).map((a) => a.subagentType),
+      LIMIT.agents
+    )
+  };
+}
+function recordFacts(ledger, facts) {
+  const turns = [...ledger.turns.filter((t) => t.turn !== facts.turn), facts].sort((a, b) => a.turn - b.turn).slice(-FACT_TURNS);
+  return { ...ledger, turns };
+}
+function trim(entries) {
+  const first = entries.find((e) => e.outcome === "followed");
+  const out = [...entries];
+  while (out.length > LEDGER_ENTRIES) {
+    const last = out.length - 1;
+    const loose = out.findIndex((e, i) => e !== first && e.outcome !== "broken" && i !== last);
+    out.splice(loose >= 0 ? loose : out.findIndex((e) => e !== first), 1);
+  }
+  return out;
+}
+function recordTurn(ledger, turn, results) {
+  const next = { ...ledger, rules: { ...ledger.rules } };
+  for (const r of results) {
+    const hash = sha256(r.text);
+    const prior = next.rules[r.key];
+    const kept = prior?.text === hash ? prior.entries.filter((e) => e.turn !== turn) : [];
+    const entries = trim(
+      [...kept, { turn, outcome: r.outcome, note: r.note.slice(0, NOTE_CHARS) }].sort(
+        (a, b) => a.turn - b.turn
+      )
+    );
+    next.rules[r.key] = { text: hash, entries };
+  }
+  return next;
+}
+function rollup(ledger) {
+  const keys = Object.keys(ledger.rules);
+  if (ledger.turns.length === 0 && keys.length === 0) return null;
+  const rules = keys.flatMap((key) => {
+    const entries = ledger.rules[key]?.entries ?? [];
+    const last = entries.at(-1);
+    if (last === void 0) return [];
+    const source = sourceOf(key);
+    const n = (o) => entries.filter((e) => e.outcome === o).length;
+    return [
+      {
+        source,
+        rule: key.slice(source.length + 1),
+        followed: n("followed"),
+        broken: n("broken"),
+        repaired: entries.filter(
+          (e, i) => e.outcome === "broken" && entries.slice(i + 1).some((f) => f.outcome === "followed")
+        ).length,
+        needsReview: n("needs-review"),
+        notApplicable: n("not-applicable"),
+        notInScope: ledger.turns.filter((t) => !t.sources.includes(source)).length,
+        last: { turn: last.turn, outcome: last.outcome }
+      }
+    ];
+  });
+  return { turns: ledger.turns.length, rules };
+}
+async function writeLedger(path, ledger) {
+  await mkdir(dirname(path), { recursive: true });
+  const tmp = `${path}.${String(process.pid)}.tmp`;
+  await writeFile(tmp, `${JSON.stringify(ledger, null, 2)}
+`, "utf-8");
+  try {
+    await rename(tmp, path);
+  } catch (err) {
+    await rm(tmp, { force: true });
+    throw err;
+  }
+}
+
+export {
+  idHash,
+  ledgerPath,
+  readLedger,
+  historyBlock,
+  projectPath,
+  clipTo,
+  turnFacts,
+  recordFacts,
+  recordTurn,
+  rollup,
+  writeLedger
+};
+//# sourceMappingURL=chunk-7EL5LLKD.js.map

package/dist/chunk-7EL5LLKD.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/tracevals/rules/ledger.ts"],"sourcesContent":["/**\n * The session ledger (proposal 0079, \"The ledger\"). What each turn of one\n * session did, and what the in-loop judge found for each rule, so a rule\n * about a procedure that spans turns is judged with what earlier turns did.\n *\n * Only the hook writes it, one file per session and one per subagent, so\n * parallel SubagentStops never share a file. `check` by hand reads it and\n * never writes it. A missing or unreadable ledger is an empty history.\n */\nimport { createHash } from \"node:crypto\";\nimport { mkdir, readFile, rename, rm, writeFile } from \"node:fs/promises\";\nimport { dirname, isAbsolute, join, relative, resolve, sep } from \"node:path\";\nimport { DEFAULT_CAPTURE_DIR } from \"../capture/types.js\";\nimport { makeRedactor } from \"../judge/redact.js\";\nimport type { Trace } from \"../trace/types.js\";\nimport { pathKey, type RuleSource } from \"./sources.js\";\nimport type { TurnSlice } from \"./turn.js\";\n\nexport type LedgerOutcome = \"followed\" | \"broken\" | \"needs-review\" | \"not-applicable\";\n\nexport interface LedgerEntry {\n  /** The turn's first event, as an ordinal in the trace. */\n  turn: number;\n  outcome: LedgerOutcome;\n  /** The judge's reasoning, cut to `NOTE_CHARS`. */\n  note: string;\n}\n\n/** What one turn did, read from the trace with no model. */\nexport interface TurnFacts {\n  turn: number;\n  /** Whether any rule source governed the turn. */\n  inScope: boolean;\n  /** The rule sources in scope, by display path. */\n  sources: string[];\n  /** Bash commands, each clipped to `COMMAND_CHARS`. */\n  commands: string[];\n  /** Files written or edited, relative to the project. */\n  wrote: string[];\n  /** Rule sources the turn read. */\n  read: string[];\n  skills: string[];\n  /** Subagent types the turn spawned. */\n  agents: string[];\n}\n\nexport interface Ledger {\n  version: 1;\n  /** The last `FACT_TURNS` turns the hook saw, oldest first. */\n  turns: TurnFacts[];\n  /** Keyed `<source>#<id>`. `text` is the rule text's sha256, so an edited rule starts fresh. */\n  rules: Record<string, { text: string; entries: LedgerEntry[] }>;\n}\n\nexport const LEDGER_ENTRIES = 10;\nexport const HISTORY_ENTRIES = 5;\nexport const NOTE_CHARS = 200;\nexport const FACT_TURNS = 30;\nexport const COMMAND_CHARS = 120;\nconst LIMIT = { commands: 10, wrote: 10, read: 5, skills: 10, agents: 10 } as const;\n/** Runs of turns a rule's source missed, shown in its history. */\nconst GAP_RUNS = 3;\n\nconst sha256 = (s: string): string => createHash(\"sha256\").update(s).digest(\"hex\");\n\n/** A session or agent id, hashed so nothing in it can climb out of the directory. */\nexport function idHash(id: string): string {\n  return sha256(id).slice(0, 32);\n}\n\n/** `<project>/.manni/tracevals/sessions/<session hash>[.<agent id hash>].ledger.json` */\nexport function ledgerPath(projectDir: string, sessionId: string, agentId: string | null): string {\n  const agent = agentId === null ? \"\" : `.${idHash(agentId)}`;\n  return join(projectDir, DEFAULT_CAPTURE_DIR, `${idHash(sessionId)}${agent}.ledger.json`);\n}\n\nexport const emptyLedger = (): Ledger => ({ version: 1, turns: [], rules: {} });\n\nconst OUTCOMES = new Set<unknown>([\"followed\", \"broken\", \"needs-review\", \"not-applicable\"]);\n\nfunction isEntry(v: unknown): v is LedgerEntry {\n  if (typeof v !== \"object\" || v === null) return false;\n  const e = v as Record<string, unknown>;\n  return typeof e.turn === \"number\" && OUTCOMES.has(e.outcome) && typeof e.note === \"string\";\n}\n\nconst isStrings = (v: unknown): v is string[] =>\n  Array.isArray(v) && v.every((s) => typeof s === \"string\");\n\nconst FACT_LISTS = [\"sources\", \"commands\", \"wrote\", \"read\", \"skills\", \"agents\"] as const;\n\nfunction isFacts(v: unknown): v is TurnFacts {\n  if (typeof v !== \"object\" || v === null) return false;\n  const f = v as Record<string, unknown>;\n  return (\n    typeof f.turn === \"number\" &&\n    typeof f.inScope === \"boolean\" &&\n    FACT_LISTS.every((k) => isStrings(f[k]))\n  );\n}\n\n/** Never throws: anything but a well-formed ledger reads as an empty one. */\nexport async function readLedger(path: string): Promise<Ledger> {\n  let parsed: unknown;\n  try {\n    parsed = JSON.parse(await readFile(path, \"utf-8\"));\n  } catch {\n    return emptyLedger();\n  }\n  if (typeof parsed !== \"object\" || parsed === null) return emptyLedger();\n  const { version, turns, rules } = parsed as Record<string, unknown>;\n  if (version !== 1 || typeof rules !== \"object\" || rules === null) return emptyLedger();\n  const ledger = emptyLedger();\n  if (Array.isArray(turns)) ledger.turns = turns.filter(isFacts);\n  for (const [key, value] of Object.entries(rules)) {\n    if (typeof value !== \"object\" || value === null) continue;\n    const { text, entries } = value as Record<string, unknown>;\n    if (typeof text !== \"string\" || !Array.isArray(entries)) continue;\n    ledger.rules[key] = { text, entries: entries.filter(isEntry) };\n  }\n  return ledger;\n}\n\n/** The source half of a `<source>#<id>` key. */\nconst sourceOf = (key: string): string => key.slice(0, key.lastIndexOf(\"#\"));\n\n/**\n * The entries a rule's history keeps in view: its latest, its first\n * `followed`, then its `broken` ones from newest back, at most\n * `HISTORY_ENTRIES`, oldest first.\n */\nfunction keyEntries(entries: LedgerEntry[]): LedgerEntry[] {\n  const broken = entries.filter((e) => e.outcome === \"broken\").reverse();\n  const picked = [entries.at(-1), entries.find((e) => e.outcome === \"followed\"), ...broken];\n  return [...new Set(picked.filter((e): e is LedgerEntry => e !== undefined))]\n    .slice(0, HISTORY_ENTRIES)\n    .sort((a, b) => a.turn - b.turn);\n}\n\n/** The last `GAP_RUNS` runs of consecutive recorded turns whose sources leave `source` out. */\nfunction gaps(turns: TurnFacts[], source: string): { from: number; to: number }[] {\n  const runs: { from: number; to: number }[] = [];\n  let open: { from: number; to: number } | undefined;\n  for (const t of turns) {\n    if (t.sources.includes(source)) {\n      open = undefined;\n    } else if (open !== undefined) {\n      open.to = t.turn;\n    } else {\n      open = { from: t.turn, to: t.turn };\n      runs.push(open);\n    }\n  }\n  return runs.slice(-GAP_RUNS);\n}\n\n/**\n * The block a rule's item carries, or \"\" when the ledger holds nothing for it\n * before `before` other than turns it did not apply to. Runs of recorded\n * turns where its source was not in scope are marked, so a hole in the\n * history does not read as quiet.\n */\nexport function historyBlock(ledger: Ledger, key: string, text: string, before: number): string {\n  const rule = ledger.rules[key];\n  if (rule?.text !== sha256(text)) return \"\";\n  const shown = keyEntries(\n    rule.entries.filter((e) => e.turn < before && e.outcome !== \"not-applicable\"),\n  );\n  if (shown.length === 0) return \"\";\n  const entries = shown.map((e) => ({\n    turn: e.turn,\n    line: `- turn ${String(e.turn)}: ${e.outcome}.${e.note !== \"\" ? ` ${e.note}` : \"\"}`,\n  }));\n  const holes = gaps(\n    ledger.turns.filter((t) => t.turn < before),\n    sourceOf(key),\n  ).map((g) => ({\n    turn: g.from,\n    line:\n      g.from === g.to\n        ? `- (turn ${String(g.from)}: not in scope)`\n        : `- (turns ${String(g.from)}–${String(g.to)}: not in scope)`,\n  }));\n  const lines = [...entries, ...holes].sort((a, b) => a.turn - b.turn).map((l) => l.line);\n  return `# This rule earlier in this session\\n\\n${lines.join(\"\\n\")}`;\n}\n\nconst posix = (p: string): string => p.split(sep).join(\"/\");\n\n/** A path as the turn named it, relative to the project root when it is under it. */\nexport function projectPath(path: string, cwd: string, root: string): string {\n  const abs = resolve(cwd, path);\n  const r = relative(root, abs);\n  return posix(r === \"\" || r.startsWith(\"..\") || isAbsolute(r) ? abs : r);\n}\n\n/** `s` cut to at most `max` characters, the last an ellipsis when cut. */\nexport const clipTo = (s: string, max: number): string =>\n  s.length <= max ? s : `${s.slice(0, max - 1)}…`;\n\n/**\n * What the turn did, read from its window. Commands and paths pass through\n * the judge's redactor. `sources` are the rule sources in scope, none for a\n * turn that stopped before they were resolved.\n */\nexport function turnFacts(\n  trace: Trace,\n  turn: TurnSlice,\n  sources: Pick<RuleSource, \"path\" | \"displayPath\">[],\n  opts: { cwd: string; root: string; redact: readonly string[] },\n): TurnFacts {\n  const w = turn.window;\n  const scrub = makeRedactor(opts.redact);\n  const take = (items: string[], n = Number.POSITIVE_INFINITY): string[] =>\n    [...new Set(items)].slice(0, n);\n  const clip = (s: string): string => clipTo(s, COMMAND_CHARS);\n  const rel = (path: string): string => projectPath(path, opts.cwd, opts.root);\n  const commands = w.toolCalls.flatMap((c) =>\n    c.name === \"Bash\" && typeof c.input.command === \"string\"\n      ? [clip(scrub(c.input.command.replace(/\\s+/g, \" \").trim()))]\n      : [],\n  );\n  const reads = new Set(\n    w.fileAccesses.filter((a) => a.op === \"read\").map((a) => pathKey(resolve(opts.cwd, a.path))),\n  );\n  return {\n    turn: turn.from,\n    inScope: sources.length > 0,\n    sources: take(sources.map((s) => s.displayPath)),\n    commands: take(commands, LIMIT.commands),\n    wrote: take(\n      w.fileAccesses.filter((a) => a.op !== \"read\").map((a) => scrub(rel(a.path))),\n      LIMIT.wrote,\n    ),\n    read: take(\n      // A source with no file (the typed prompts, an inline plan) has path \"\",\n      // which would resolve to the cwd; it is never read.\n      sources.filter((s) => s.path !== \"\" && reads.has(pathKey(s.path))).map((s) => s.displayPath),\n      LIMIT.read,\n    ),\n    skills: take(\n      w.skillInvocations.map((s) => s.name),\n      LIMIT.skills,\n    ),\n    agents: w.empty\n      ? []\n      : take(\n          trace.agentSpawns\n            .filter((a) => a.index >= turn.from && a.index <= turn.to)\n            .map((a) => a.subagentType),\n          LIMIT.agents,\n        ),\n  };\n}\n\n/** The ledger with this turn's facts, replacing any earlier record of the same turn. */\nexport function recordFacts(ledger: Ledger, facts: TurnFacts): Ledger {\n  const turns = [...ledger.turns.filter((t) => t.turn !== facts.turn), facts]\n    .sort((a, b) => a.turn - b.turn)\n    .slice(-FACT_TURNS);\n  return { ...ledger, turns };\n}\n\nexport interface TurnResult {\n  key: string;\n  text: string;\n  outcome: LedgerOutcome;\n  note: string;\n}\n\n/**\n * Cut a rule's entries to `LEDGER_ENTRIES`. The first `followed`, every\n * `broken` and the latest are pinned, so the others go first, oldest first.\n * Only when every entry is pinned does the oldest `broken` go.\n */\nfunction trim(entries: LedgerEntry[]): LedgerEntry[] {\n  const first = entries.find((e) => e.outcome === \"followed\");\n  const out = [...entries];\n  while (out.length > LEDGER_ENTRIES) {\n    const last = out.length - 1;\n    const loose = out.findIndex((e, i) => e !== first && e.outcome !== \"broken\" && i !== last);\n    out.splice(loose >= 0 ? loose : out.findIndex((e) => e !== first), 1);\n  }\n  return out;\n}\n\n/** The ledger with this turn's results, replacing any earlier record of the same turn. */\nexport function recordTurn(ledger: Ledger, turn: number, results: TurnResult[]): Ledger {\n  const next: Ledger = { ...ledger, rules: { ...ledger.rules } };\n  for (const r of results) {\n    const hash = sha256(r.text);\n    const prior = next.rules[r.key];\n    const kept = prior?.text === hash ? prior.entries.filter((e) => e.turn !== turn) : [];\n    const entries = trim(\n      [...kept, { turn, outcome: r.outcome, note: r.note.slice(0, NOTE_CHARS) }].sort(\n        (a, b) => a.turn - b.turn,\n      ),\n    );\n    next.rules[r.key] = { text: hash, entries };\n  }\n  return next;\n}\n\nexport interface RuleRollup {\n  source: string;\n  rule: string;\n  followed: number;\n  broken: number;\n  /** Broken entries that a later `followed` entry came after. */\n  repaired: number;\n  needsReview: number;\n  notApplicable: number;\n  /** Recorded turns where the rule's source was not in scope. */\n  notInScope: number;\n  last: { turn: number; outcome: LedgerOutcome };\n}\n\nexport interface SessionRollup {\n  /** Turns the ledger records. */\n  turns: number;\n  rules: RuleRollup[];\n}\n\n/** Each rule's record over the session, or null when the ledger holds nothing. */\nexport function rollup(ledger: Ledger): SessionRollup | null {\n  const keys = Object.keys(ledger.rules);\n  if (ledger.turns.length === 0 && keys.length === 0) return null;\n  const rules = keys.flatMap((key): RuleRollup[] => {\n    const entries = ledger.rules[key]?.entries ?? [];\n    const last = entries.at(-1);\n    if (last === undefined) return [];\n    const source = sourceOf(key);\n    const n = (o: LedgerOutcome): number => entries.filter((e) => e.outcome === o).length;\n    return [\n      {\n        source,\n        rule: key.slice(source.length + 1),\n        followed: n(\"followed\"),\n        broken: n(\"broken\"),\n        repaired: entries.filter(\n          (e, i) =>\n            e.outcome === \"broken\" && entries.slice(i + 1).some((f) => f.outcome === \"followed\"),\n        ).length,\n        needsReview: n(\"needs-review\"),\n        notApplicable: n(\"not-applicable\"),\n        notInScope: ledger.turns.filter((t) => !t.sources.includes(source)).length,\n        last: { turn: last.turn, outcome: last.outcome },\n      },\n    ];\n  });\n  return { turns: ledger.turns.length, rules };\n}\n\n/** Temp file and rename, so a reader never sees half a ledger. Throws on failure. */\nexport async function writeLedger(path: string, ledger: Ledger): Promise<void> {\n  await mkdir(dirname(path), { recursive: true });\n  const tmp = `${path}.${String(process.pid)}.tmp`;\n  await writeFile(tmp, `${JSON.stringify(ledger, null, 2)}\\n`, \"utf-8\");\n  try {\n    await rename(tmp, path);\n  } catch (err) {\n    await rm(tmp, { force: true });\n    throw err;\n  }\n}\n"],"mappings":";;;;;;;;;;AASA,SAAS,kBAAkB;AAC3B,SAAS,OAAO,UAAU,QAAQ,IAAI,iBAAiB;AACvD,SAAS,SAAS,YAAY,MAAM,UAAU,SAAS,WAAW;AA2C3D,IAAM,iBAAiB;AACvB,IAAM,kBAAkB;AACxB,IAAM,aAAa;AACnB,IAAM,aAAa;AACnB,IAAM,gBAAgB;AAC7B,IAAM,QAAQ,EAAE,UAAU,IAAI,OAAO,IAAI,MAAM,GAAG,QAAQ,IAAI,QAAQ,GAAG;AAEzE,IAAM,WAAW;AAEjB,IAAM,SAAS,CAAC,MAAsB,WAAW,QAAQ,EAAE,OAAO,CAAC,EAAE,OAAO,KAAK;AAG1E,SAAS,OAAO,IAAoB;AACzC,SAAO,OAAO,EAAE,EAAE,MAAM,GAAG,EAAE;AAC/B;AAGO,SAAS,WAAW,YAAoB,WAAmB,SAAgC;AAChG,QAAM,QAAQ,YAAY,OAAO,KAAK,IAAI,OAAO,OAAO,CAAC;AACzD,SAAO,KAAK,YAAY,qBAAqB,GAAG,OAAO,SAAS,CAAC,GAAG,KAAK,cAAc;AACzF;AAEO,IAAM,cAAc,OAAe,EAAE,SAAS,GAAG,OAAO,CAAC,GAAG,OAAO,CAAC,EAAE;AAE7E,IAAM,WAAW,oBAAI,IAAa,CAAC,YAAY,UAAU,gBAAgB,gBAAgB,CAAC;AAE1F,SAAS,QAAQ,GAA8B;AAC7C,MAAI,OAAO,MAAM,YAAY,MAAM,KAAM,QAAO;AAChD,QAAM,IAAI;AACV,SAAO,OAAO,EAAE,SAAS,YAAY,SAAS,IAAI,EAAE,OAAO,KAAK,OAAO,EAAE,SAAS;AACpF;AAEA,IAAM,YAAY,CAAC,MACjB,MAAM,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,MAAM,OAAO,MAAM,QAAQ;AAE1D,IAAM,aAAa,CAAC,WAAW,YAAY,SAAS,QAAQ,UAAU,QAAQ;AAE9E,SAAS,QAAQ,GAA4B;AAC3C,MAAI,OAAO,MAAM,YAAY,MAAM,KAAM,QAAO;AAChD,QAAM,IAAI;AACV,SACE,OAAO,EAAE,SAAS,YAClB,OAAO,EAAE,YAAY,aACrB,WAAW,MAAM,CAAC,MAAM,UAAU,EAAE,CAAC,CAAC,CAAC;AAE3C;AAGA,eAAsB,WAAW,MAA+B;AAC9D,MAAI;AACJ,MAAI;AACF,aAAS,KAAK,MAAM,MAAM,SAAS,MAAM,OAAO,CAAC;AAAA,EACnD,QAAQ;AACN,WAAO,YAAY;AAAA,EACrB;AACA,MAAI,OAAO,WAAW,YAAY,WAAW,KAAM,QAAO,YAAY;AACtE,QAAM,EAAE,SAAS,OAAO,MAAM,IAAI;AAClC,MAAI,YAAY,KAAK,OAAO,UAAU,YAAY,UAAU,KAAM,QAAO,YAAY;AACrF,QAAM,SAAS,YAAY;AAC3B,MAAI,MAAM,QAAQ,KAAK,EAAG,QAAO,QAAQ,MAAM,OAAO,OAAO;AAC7D,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,KAAK,GAAG;AAChD,QAAI,OAAO,UAAU,YAAY,UAAU,KAAM;AACjD,UAAM,EAAE,MAAM,QAAQ,IAAI;AAC1B,QAAI,OAAO,SAAS,YAAY,CAAC,MAAM,QAAQ,OAAO,EAAG;AACzD,WAAO,MAAM,GAAG,IAAI,EAAE,MAAM,SAAS,QAAQ,OAAO,OAAO,EAAE;AAAA,EAC/D;AACA,SAAO;AACT;AAGA,IAAM,WAAW,CAAC,QAAwB,IAAI,MAAM,GAAG,IAAI,YAAY,GAAG,CAAC;AAO3E,SAAS,WAAW,SAAuC;AACzD,QAAM,SAAS,QAAQ,OAAO,CAAC,MAAM,EAAE,YAAY,QAAQ,EAAE,QAAQ;AACrE,QAAM,SAAS,CAAC,QAAQ,GAAG,EAAE,GAAG,QAAQ,KAAK,CAAC,MAAM,EAAE,YAAY,UAAU,GAAG,GAAG,MAAM;AACxF,SAAO,CAAC,GAAG,IAAI,IAAI,OAAO,OAAO,CAAC,MAAwB,MAAM,MAAS,CAAC,CAAC,EACxE,MAAM,GAAG,eAAe,EACxB,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;AACnC;AAGA,SAAS,KAAK,OAAoB,QAAgD;AAChF,QAAM,OAAuC,CAAC;AAC9C,MAAI;AACJ,aAAW,KAAK,OAAO;AACrB,QAAI,EAAE,QAAQ,SAAS,MAAM,GAAG;AAC9B,aAAO;AAAA,IACT,WAAW,SAAS,QAAW;AAC7B,WAAK,KAAK,EAAE;AAAA,IACd,OAAO;AACL,aAAO,EAAE,MAAM,EAAE,MAAM,IAAI,EAAE,KAAK;AAClC,WAAK,KAAK,IAAI;AAAA,IAChB;AAAA,EACF;AACA,SAAO,KAAK,MAAM,CAAC,QAAQ;AAC7B;AAQO,SAAS,aAAa,QAAgB,KAAa,MAAc,QAAwB;AAC9F,QAAM,OAAO,OAAO,MAAM,GAAG;AAC7B,MAAI,MAAM,SAAS,OAAO,IAAI,EAAG,QAAO;AACxC,QAAM,QAAQ;AAAA,IACZ,KAAK,QAAQ,OAAO,CAAC,MAAM,EAAE,OAAO,UAAU,EAAE,YAAY,gBAAgB;AAAA,EAC9E;AACA,MAAI,MAAM,WAAW,EAAG,QAAO;AAC/B,QAAM,UAAU,MAAM,IAAI,CAAC,OAAO;AAAA,IAChC,MAAM,EAAE;AAAA,IACR,MAAM,UAAU,OAAO,EAAE,IAAI,CAAC,KAAK,EAAE,OAAO,IAAI,EAAE,SAAS,KAAK,IAAI,EAAE,IAAI,KAAK,EAAE;AAAA,EACnF,EAAE;AACF,QAAM,QAAQ;AAAA,IACZ,OAAO,MAAM,OAAO,CAAC,MAAM,EAAE,OAAO,MAAM;AAAA,IAC1C,SAAS,GAAG;AAAA,EACd,EAAE,IAAI,CAAC,OAAO;AAAA,IACZ,MAAM,EAAE;AAAA,IACR,MACE,EAAE,SAAS,EAAE,KACT,WAAW,OAAO,EAAE,IAAI,CAAC,oBACzB,YAAY,OAAO,EAAE,IAAI,CAAC,SAAI,OAAO,EAAE,EAAE,CAAC;AAAA,EAClD,EAAE;AACF,QAAM,QAAQ,CAAC,GAAG,SAAS,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,MAAM,EAAE,IAAI;AACtF,SAAO;AAAA;AAAA,EAA0C,MAAM,KAAK,IAAI,CAAC;AACnE;AAEA,IAAM,QAAQ,CAAC,MAAsB,EAAE,MAAM,GAAG,EAAE,KAAK,GAAG;AAGnD,SAAS,YAAY,MAAc,KAAa,MAAsB;AAC3E,QAAM,MAAM,QAAQ,KAAK,IAAI;AAC7B,QAAM,IAAI,SAAS,MAAM,GAAG;AAC5B,SAAO,MAAM,MAAM,MAAM,EAAE,WAAW,IAAI,KAAK,WAAW,CAAC,IAAI,MAAM,CAAC;AACxE;AAGO,IAAM,SAAS,CAAC,GAAW,QAChC,EAAE,UAAU,MAAM,IAAI,GAAG,EAAE,MAAM,GAAG,MAAM,CAAC,CAAC;AAOvC,SAAS,UACd,OACA,MACA,SACA,MACW;AACX,QAAM,IAAI,KAAK;AACf,QAAM,QAAQ,aAAa,KAAK,MAAM;AACtC,QAAM,OAAO,CAAC,OAAiB,IAAI,OAAO,sBACxC,CAAC,GAAG,IAAI,IAAI,KAAK,CAAC,EAAE,MAAM,GAAG,CAAC;AAChC,QAAM,OAAO,CAAC,MAAsB,OAAO,GAAG,aAAa;AAC3D,QAAM,MAAM,CAAC,SAAyB,YAAY,MAAM,KAAK,KAAK,KAAK,IAAI;AAC3E,QAAM,WAAW,EAAE,UAAU;AAAA,IAAQ,CAAC,MACpC,EAAE,SAAS,UAAU,OAAO,EAAE,MAAM,YAAY,WAC5C,CAAC,KAAK,MAAM,EAAE,MAAM,QAAQ,QAAQ,QAAQ,GAAG,EAAE,KAAK,CAAC,CAAC,CAAC,IACzD,CAAC;AAAA,EACP;AACA,QAAM,QAAQ,IAAI;AAAA,IAChB,EAAE,aAAa,OAAO,CAAC,MAAM,EAAE,OAAO,MAAM,EAAE,IAAI,CAAC,MAAM,QAAQ,QAAQ,KAAK,KAAK,EAAE,IAAI,CAAC,CAAC;AAAA,EAC7F;AACA,SAAO;AAAA,IACL,MAAM,KAAK;AAAA,IACX,SAAS,QAAQ,SAAS;AAAA,IAC1B,SAAS,KAAK,QAAQ,IAAI,CAAC,MAAM,EAAE,WAAW,CAAC;AAAA,IAC/C,UAAU,KAAK,UAAU,MAAM,QAAQ;AAAA,IACvC,OAAO;AAAA,MACL,EAAE,aAAa,OAAO,CAAC,MAAM,EAAE,OAAO,MAAM,EAAE,IAAI,CAAC,MAAM,MAAM,IAAI,EAAE,IAAI,CAAC,CAAC;AAAA,MAC3E,MAAM;AAAA,IACR;AAAA,IACA,MAAM;AAAA;AAAA;AAAA,MAGJ,QAAQ,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM,MAAM,IAAI,QAAQ,EAAE,IAAI,CAAC,CAAC,EAAE,IAAI,CAAC,MAAM,EAAE,WAAW;AAAA,MAC3F,MAAM;AAAA,IACR;AAAA,IACA,QAAQ;AAAA,MACN,EAAE,iBAAiB,IAAI,CAAC,MAAM,EAAE,IAAI;AAAA,MACpC,MAAM;AAAA,IACR;AAAA,IACA,QAAQ,EAAE,QACN,CAAC,IACD;AAAA,MACE,MAAM,YACH,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,QAAQ,EAAE,SAAS,KAAK,EAAE,EACxD,IAAI,CAAC,MAAM,EAAE,YAAY;AAAA,MAC5B,MAAM;AAAA,IACR;AAAA,EACN;AACF;AAGO,SAAS,YAAY,QAAgB,OAA0B;AACpE,QAAM,QAAQ,CAAC,GAAG,OAAO,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM,IAAI,GAAG,KAAK,EACvE,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI,EAC9B,MAAM,CAAC,UAAU;AACpB,SAAO,EAAE,GAAG,QAAQ,MAAM;AAC5B;AAcA,SAAS,KAAK,SAAuC;AACnD,QAAM,QAAQ,QAAQ,KAAK,CAAC,MAAM,EAAE,YAAY,UAAU;AAC1D,QAAM,MAAM,CAAC,GAAG,OAAO;AACvB,SAAO,IAAI,SAAS,gBAAgB;AAClC,UAAM,OAAO,IAAI,SAAS;AAC1B,UAAM,QAAQ,IAAI,UAAU,CAAC,GAAG,MAAM,MAAM,SAAS,EAAE,YAAY,YAAY,MAAM,IAAI;AACzF,QAAI,OAAO,SAAS,IAAI,QAAQ,IAAI,UAAU,CAAC,MAAM,MAAM,KAAK,GAAG,CAAC;AAAA,EACtE;AACA,SAAO;AACT;AAGO,SAAS,WAAW,QAAgB,MAAc,SAA+B;AACtF,QAAM,OAAe,EAAE,GAAG,QAAQ,OAAO,EAAE,GAAG,OAAO,MAAM,EAAE;AAC7D,aAAW,KAAK,SAAS;AACvB,UAAM,OAAO,OAAO,EAAE,IAAI;AAC1B,UAAM,QAAQ,KAAK,MAAM,EAAE,GAAG;AAC9B,UAAM,OAAO,OAAO,SAAS,OAAO,MAAM,QAAQ,OAAO,CAAC,MAAM,EAAE,SAAS,IAAI,IAAI,CAAC;AACpF,UAAM,UAAU;AAAA,MACd,CAAC,GAAG,MAAM,EAAE,MAAM,SAAS,EAAE,SAAS,MAAM,EAAE,KAAK,MAAM,GAAG,UAAU,EAAE,CAAC,EAAE;AAAA,QACzE,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE;AAAA,MACvB;AAAA,IACF;AACA,SAAK,MAAM,EAAE,GAAG,IAAI,EAAE,MAAM,MAAM,QAAQ;AAAA,EAC5C;AACA,SAAO;AACT;AAuBO,SAAS,OAAO,QAAsC;AAC3D,QAAM,OAAO,OAAO,KAAK,OAAO,KAAK;AACrC,MAAI,OAAO,MAAM,WAAW,KAAK,KAAK,WAAW,EAAG,QAAO;AAC3D,QAAM,QAAQ,KAAK,QAAQ,CAAC,QAAsB;AAChD,UAAM,UAAU,OAAO,MAAM,GAAG,GAAG,WAAW,CAAC;AAC/C,UAAM,OAAO,QAAQ,GAAG,EAAE;AAC1B,QAAI,SAAS,OAAW,QAAO,CAAC;AAChC,UAAM,SAAS,SAAS,GAAG;AAC3B,UAAM,IAAI,CAAC,MAA6B,QAAQ,OAAO,CAAC,MAAM,EAAE,YAAY,CAAC,EAAE;AAC/E,WAAO;AAAA,MACL;AAAA,QACE;AAAA,QACA,MAAM,IAAI,MAAM,OAAO,SAAS,CAAC;AAAA,QACjC,UAAU,EAAE,UAAU;AAAA,QACtB,QAAQ,EAAE,QAAQ;AAAA,QAClB,UAAU,QAAQ;AAAA,UAChB,CAAC,GAAG,MACF,EAAE,YAAY,YAAY,QAAQ,MAAM,IAAI,CAAC,EAAE,KAAK,CAAC,MAAM,EAAE,YAAY,UAAU;AAAA,QACvF,EAAE;AAAA,QACF,aAAa,EAAE,cAAc;AAAA,QAC7B,eAAe,EAAE,gBAAgB;AAAA,QACjC,YAAY,OAAO,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,QAAQ,SAAS,MAAM,CAAC,EAAE;AAAA,QACpE,MAAM,EAAE,MAAM,KAAK,MAAM,SAAS,KAAK,QAAQ;AAAA,MACjD;AAAA,IACF;AAAA,EACF,CAAC;AACD,SAAO,EAAE,OAAO,OAAO,MAAM,QAAQ,MAAM;AAC7C;AAGA,eAAsB,YAAY,MAAc,QAA+B;AAC7E,QAAM,MAAM,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AAC9C,QAAM,MAAM,GAAG,IAAI,IAAI,OAAO,QAAQ,GAAG,CAAC;AAC1C,QAAM,UAAU,KAAK,GAAG,KAAK,UAAU,QAAQ,MAAM,CAAC,CAAC;AAAA,GAAM,OAAO;AACpE,MAAI;AACF,UAAM,OAAO,KAAK,IAAI;AAAA,EACxB,SAAS,KAAK;AACZ,UAAM,GAAG,KAAK,EAAE,OAAO,KAAK,CAAC;AAC7B,UAAM;AAAA,EACR;AACF;","names":[]}

package/dist/{chunk-C3KI45RF.js → chunk-7XZ6INKO.js}

@@ -15,10 +15,10 @@ import {
 import {
   REPORT_FORMAT_LIST,
   renderJunit
-} from "./chunk-TCC737IS.js";
+} from "./chunk-5VWDIWOH.js";
 import {
   isErrorSeverity
-} from "./chunk-527HHQPS.js";
+} from "./chunk-MSMB6XOS.js";
 import {
   palette
 } from "./chunk-RNVYGWIC.js";
@@ -28,7 +28,7 @@ import {
 } from "./chunk-MNH7QO4Z.js";
 import {
   package_default
-} from "./chunk-6XTQDFGP.js";
+} from "./chunk-YSCZL4U5.js";
 
 // src/lint/core/rule-id.ts
 var RULE_ID_PREFIX = "manni:lint";
@@ -633,4 +633,4 @@ export {
   renderTemplates,
   renderTools
 };
-//# sourceMappingURL=chunk-C3KI45RF.js.map
+//# sourceMappingURL=chunk-7XZ6INKO.js.map

package/dist/{chunk-ZX43SRZR.js → chunk-AMEILOVV.js}

@@ -1,16 +1,18 @@
 #!/usr/bin/env node
 import {
-  assertProviderSelection,
   renderCalibration,
   resolveBatchTraces,
-  runBatch,
-  selectProvider
-} from "./chunk-SJO2FTPG.js";
+  runBatch
+} from "./chunk-QP5X66TO.js";
 import {
   discoverConfig,
   labelKey,
   loadLabels
-} from "./chunk-3R5Y3XOL.js";
+} from "./chunk-TPVOOTMN.js";
+import {
+  assertProviderSelection,
+  selectProvider
+} from "./chunk-BWCWTHYS.js";
 import {
   TracevalsError
 } from "./chunk-UKZNLVJW.js";
@@ -376,4 +378,4 @@ export {
   score,
   runCalibrate
 };
-//# sourceMappingURL=chunk-ZX43SRZR.js.map
+//# sourceMappingURL=chunk-AMEILOVV.js.map

package/dist/{chunk-ZX43SRZR.js.map → chunk-AMEILOVV.js.map}

@@ -1 +1 @@
-{"version":3,"sources":["../src/tracevals/commands/calibrate.ts","../src/tracevals/calibrate/score.ts"],"sourcesContent":["/**\n * `manni tracevals calibrate <traces...> --labels <file>` (ADR 01022).\n *\n * The tool could report what it decided; it could not report whether it was\n * right. This command runs the batch (ADR 01018), joins the results against a\n * human's answers, and reports the three numbers the calibration page used to\n * ask people to count by hand — false passes, false fails, and review volume —\n * plus which eval disagreed.\n *\n * `--sweep` costs no inference beyond the one run it already made: a judged\n * eval's consensus carries the whole `JudgeRun[]`, so a different zone\n * threshold or a smaller ensemble is arithmetic over verdicts already on disk.\n * Read-only throughout — labels, traces, and artifacts are never written.\n */\nimport { writeFile } from \"node:fs/promises\";\nimport { basename, resolve } from \"node:path\";\nimport { discoverConfig } from \"../core/config.js\";\nimport { assertProviderSelection, selectProvider } from \"../judge/provider.js\";\nimport { renderCalibration, type ReportFormat } from \"../reporters/index.js\";\nimport { TracevalsError, type EvalResult } from \"../types.js\";\nimport { loadLabels, type EvalLabel } from \"../calibrate/labels.js\";\nimport { joinLabels, score, type AmbiguousLabel } from \"../calibrate/score.js\";\nimport type {\n  CalibrationGate,\n  CalibrationReport,\n  JudgeSetting,\n  SweepCell,\n} from \"../calibrate/types.js\";\nimport {\n  resolveBatchTraces,\n  runBatch,\n  type BatchCommandOptions,\n} from \"./batch.js\";\n\nexport interface CalibrateCommandOptions extends BatchCommandOptions {\n  /** Overrides `calibrate.labels`; resolved against the working directory. */\n  labels?: string;\n  /** Re-score the corpus across the configured grid. */\n  sweep?: boolean;\n  /** Thresholds; undefined defers to the config, which defaults to no gate. */\n  maxFalsePass?: number;\n  maxFalseFail?: number;\n  maxReview?: number;\n}\n\nexport interface CalibrateCommandResult {\n  report: CalibrationReport;\n  rendered: string;\n}\n\n/** Largest ensemble the sweep needs cached, so every cell can be re-scored. */\nfunction sweepDepth(grid: number[], baseline: number): number {\n  return Math.max(baseline, ...grid);\n}\n\nexport async function runCalibrate(\n  options: CalibrateCommandOptions,\n): Promise<CalibrateCommandResult> {\n  const start = Date.now();\n  const { config, dir: configDir } = await discoverConfig(\n    options.configDir ?? process.cwd(),\n    {\n      ...(options.config === undefined ? {} : { configPath: options.config }),\n      ...(options.noConfig === undefined ? {} : { noConfig: options.noConfig }),\n    },\n  );\n\n  // Before the labels file is read: a contradictory `--local --provider`, an\n  // unknown provider name, or a model with no provider to own it is a usage\n  // error, and a usage error must not arrive behind an unrelated one about a\n  // file the run would never have needed.\n  assertProviderSelection(\n    selectProvider(config, {\n      ...(options.provider !== undefined ? { provider: options.provider } : {}),\n      ...(options.model !== undefined ? { model: options.model } : {}),\n      ...(options.local !== undefined ? { local: options.local } : {}),\n    }),\n  );\n\n  // A flag overrides the config rather than bypassing it (CLAUDE.md,\n  // \"Config <-> CLI flags\"). An explicit --labels is a working-directory path,\n  // the way every other path argument on the CLI is; the config default is a\n  // project path and resolves against the config file's directory.\n  const labelsFile =\n    options.labels !== undefined\n      ? resolve(options.labels)\n      : resolve(configDir, config.calibrate.labels);\n  const labels = await loadLabels(labelsFile);\n\n  // Resolve the corpus once and hand the list to the batch, so the membership\n  // check below cannot disagree with what actually ran.\n  const corpus = (await resolveBatchTraces(options)).map((f) => resolve(f));\n  assertLabelsInCorpus(labels, corpus, labelsFile);\n\n  const baseline: JudgeSetting = {\n    ensembleRuns: options.runs ?? config.judge.ensembleRuns,\n    zones: config.judge.zones,\n  };\n  const grid = config.calibrate.sweep;\n  // With --sweep the corpus is judged once at the deepest ensemble the grid\n  // asks for, and every cell — the baseline included — is re-scored from those\n  // runs. Judging at the configured depth instead would make the larger cells\n  // unscoreable, and judging per cell would multiply the bill by the grid.\n  const depth =\n    options.sweep === true\n      ? sweepDepth(grid.ensembleRuns, baseline.ensembleRuns)\n      : baseline.ensembleRuns;\n\n  const batch = await runBatch({\n    ...options,\n    traces: corpus,\n    allProjects: undefined,\n    newerThan: undefined,\n    limit: undefined,\n    runs: depth,\n    // The calibration report is the deliverable; rendering the batch as well\n    // would write the wrong thing to --output.\n    format: \"json\",\n    output: undefined,\n  });\n\n  const all: EvalResult[] = [];\n  for (const outcome of batch.outcomes) {\n    if (\"report\" in outcome) all.push(...outcome.report.evalResults);\n  }\n  const { joined, unmatched, ambiguous } = joinLabels(batch.outcomes, labels);\n  assertNoUnmatched(unmatched, batch, labelsFile);\n  assertNoAmbiguous(ambiguous, labelsFile);\n\n  // Without a sweep the reported outcomes are used verbatim, so the headline\n  // numbers are exactly the verdicts `run` produced. With one, the corpus was\n  // judged deeper than the configured ensemble, so the baseline row has to be\n  // re-scored back down to it.\n  const scored = score(\n    { joined, all },\n    options.sweep === true ? baseline : undefined,\n  );\n\n  const sweep =\n    options.sweep === true\n      ? buildSweep({ joined, all }, baseline, grid)\n      : undefined;\n\n  const gates = buildGates(scored.counts, {\n    falsePass: options.maxFalsePass ?? config.calibrate.maxFalsePass,\n    falseFail: options.maxFalseFail ?? config.calibrate.maxFalseFail,\n    review: options.maxReview ?? config.calibrate.maxReview,\n  });\n\n  const warnings = [...batch.report.warnings];\n  if (scored.counts.insufficient > 0) {\n    warnings.push(\n      `${scored.counts.insufficient} labelled eval(s) had fewer cached runs than the setting asks for and were left at their reported outcome`,\n    );\n  }\n\n  // A measurement that measured nothing must not report a pass.\n  //\n  // Every label joining to a `skipped` result — `--deterministic-only` over\n  // ai-graded labels, a budget exhausted on the first trace — leaves `scored`\n  // at 0, `agreement` at null, and each threshold reading \"0 of at most N, not\n  // exceeded\". The gates are not lying; the denominator behind them is empty.\n  // Saying so is the report's job, and exit 0 would say the opposite.\n  const measuredNothing = scored.counts.scored === 0;\n  if (measuredNothing) {\n    warnings.push(\n      `no labelled eval produced evidence: ${scored.counts.labels} label(s) all joined to a skipped result, ` +\n        `so agreement and every --max-* threshold are computed over an empty denominator`,\n    );\n  }\n\n  const report: CalibrationReport = {\n    labelsFile,\n    corpus,\n    setting: baseline,\n    counts: scored.counts,\n    disagreements: scored.disagreements,\n    unscored: scored.unscored,\n    ...(sweep !== undefined ? { sweep } : {}),\n    gates,\n    batch: batch.report,\n    warnings,\n    // A calibration run is a measurement, not a gate: disagreement is the\n    // finding, not a failure. Exit 1 only when a threshold was asked for and\n    // missed, or when the measurement itself is incomplete — a trace lost, the\n    // shared judge budget cut the corpus short, or nothing scored at all. An\n    // incomplete number presented as a clean one is worse than no number.\n    exitCode:\n      gates.some((g) => g.exceeded) ||\n      batch.report.summary.tracesErrored > 0 ||\n      batch.report.budget !== undefined ||\n      measuredNothing\n        ? 1\n        : 0,\n    turns: batch.report.turns,\n    durationMs: Date.now() - start,\n  };\n\n  const rendered = renderCalibration(report, options.format ?? \"pretty\", {\n    color: options.color === true,\n  });\n  if (options.output) await writeFile(options.output, rendered, \"utf-8\");\n  return { report, rendered };\n}\n\n/**\n * Every label must name a trace this run evaluates. Checked before the batch,\n * because the alternative is paying for a corpus and then being told the file\n * was misspelled.\n */\nfunction assertLabelsInCorpus(\n  labels: EvalLabel[],\n  corpus: string[],\n  labelsFile: string,\n): void {\n  const known = new Set(corpus);\n  const missing = [...new Set(labels.filter((l) => !known.has(l.traceFile)).map((l) => l.trace))];\n  if (missing.length === 0) return;\n  throw new TracevalsError(\n    `${labelsFile}: label(s) name trace(s) not in the corpus: ${missing.join(\", \")}. ` +\n      `The corpus is: ${corpus.map((f) => basename(f)).join(\", \")}`,\n  );\n}\n\n/**\n * A label that matched nothing is an error, not a quiet omission.\n *\n * It has to be: a typo'd eval id contributes to no count, so the report would\n * read as a cleaner corpus than it is — the same silent-nothing failure the\n * empty-selector rule exists to prevent (ADR 01018).\n */\nfunction assertNoUnmatched(\n  unmatched: EvalLabel[],\n  batch: { outcomes: { file: string; report?: { evalResults: EvalResult[] } }[] },\n  labelsFile: string,\n): void {\n  if (unmatched.length === 0) return;\n  const lines = unmatched.map((label) => {\n    const outcome = batch.outcomes.find((o) => o.file === label.traceFile);\n    const results = outcome?.report?.evalResults ?? [];\n    const sameArtifact = results.filter(\n      (r) => r.artifactName === label.artifact,\n    );\n    const hint =\n      sameArtifact.length > 0\n        ? `that trace evaluated ${sameArtifact.map((r) => r.evalName).sort().join(\", \")} for \"${label.artifact}\"`\n        : `that trace resolved no artifact named \"${label.artifact}\" — it evaluated ${\n            [...new Set(results.map((r) => r.artifactName))].sort().join(\", \") ||\n            \"nothing\"\n          }`;\n    return `  ${label.artifact} › ${label.eval} on ${basename(label.traceFile)} (${hint})`;\n  });\n  throw new TracevalsError(\n    `${labelsFile}: label(s) matched no eval result:\\n${lines.join(\"\\n\")}`,\n  );\n}\n\n/**\n * A label that omits `type:` over a name two artifact kinds both answer cannot\n * be resolved, and must not be guessed. Binding it to whichever result was\n * indexed last would score agreement against a verdict belonging to a\n * different artifact — the report would read as measured when it is wrong,\n * which is the one thing this command exists not to do.\n *\n * Operational, like `assertNoUnmatched`: the labels file is under-specified,\n * so this is exit 2 (bad input), not exit 1 (a check failed).\n */\nfunction assertNoAmbiguous(\n  ambiguous: AmbiguousLabel[],\n  labelsFile: string,\n): void {\n  if (ambiguous.length === 0) return;\n  const lines = ambiguous.map(({ label, candidates }) => {\n    const kinds = [...new Set(candidates.map((r) => r.artifactType))].sort();\n    return `  ${label.artifact} › ${label.eval} on ${basename(label.traceFile)} matches ${kinds.join(\" and \")} — add \\`type:\\` to say which`;\n  });\n  throw new TracevalsError(\n    `${labelsFile}: label(s) name an artifact two kinds answer:\\n${lines.join(\"\\n\")}`,\n  );\n}\n\nfunction buildGates(\n  counts: { falsePass: number; falseFail: number; reviewVolume: number },\n  limits: { falsePass?: number; falseFail?: number; review?: number },\n): CalibrationGate[] {\n  const gates: CalibrationGate[] = [];\n  const add = (\n    name: CalibrationGate[\"name\"],\n    limit: number | undefined,\n    actual: number,\n  ): void => {\n    if (limit === undefined) return;\n    gates.push({ name, limit, actual, exceeded: actual > limit });\n  };\n  add(\"falsePass\", limits.falsePass, counts.falsePass);\n  add(\"falseFail\", limits.falseFail, counts.falseFail);\n  add(\"review\", limits.review, counts.reviewVolume);\n  return gates;\n}\n\n/**\n * One axis at a time, every other knob held at its configured value.\n *\n * A full cross-product would be dozens of rows and would answer a question\n * nobody asked; calibration is \"move one knob, re-run, recount\", and the\n * per-axis shape is what makes a row readable as *the effect of that knob*.\n */\nfunction buildSweep(\n  input: Parameters<typeof score>[0],\n  baseline: JudgeSetting,\n  grid: { ensembleRuns: number[]; autoPass: number[]; autoFail: number[] },\n): SweepCell[] {\n  const cell = (\n    axis: SweepCell[\"axis\"],\n    value: number,\n    setting: JudgeSetting,\n  ): SweepCell => ({\n    axis,\n    value,\n    setting,\n    counts: score(input, setting).counts,\n  });\n\n  const cells: SweepCell[] = [\n    cell(\"baseline\", baseline.ensembleRuns, baseline),\n  ];\n  for (const runs of [...grid.ensembleRuns].sort((a, b) => a - b)) {\n    cells.push(\n      cell(\"ensembleRuns\", runs, { ensembleRuns: runs, zones: baseline.zones }),\n    );\n  }\n  for (const autoPass of [...grid.autoPass].sort((a, b) => a - b)) {\n    cells.push(\n      cell(\"zones.autoPass\", autoPass, {\n        ensembleRuns: baseline.ensembleRuns,\n        zones: { autoPass, autoFail: baseline.zones.autoFail },\n      }),\n    );\n  }\n  for (const autoFail of [...grid.autoFail].sort((a, b) => a - b)) {\n    cells.push(\n      cell(\"zones.autoFail\", autoFail, {\n        ensembleRuns: baseline.ensembleRuns,\n        zones: { autoPass: baseline.zones.autoPass, autoFail },\n      }),\n    );\n  }\n  return cells;\n}\n\nexport type { ReportFormat };\n","/**\n * Joining labels to results, and re-scoring cached verdicts (ADR 01022).\n *\n * Everything here is pure: results and labels in, counts out. That is what\n * makes a sweep free. A judged eval's `ConsensusResult` carries the whole\n * `JudgeRun[]` it was built from, so a different zone threshold — or a smaller\n * ensemble — is a fresh call to `computeConsensus`/`zoneFor` over runs the\n * corpus already produced, not a fresh call to a provider.\n */\nimport { computeConsensus, zoneFor } from \"@hawkeyexl/inference\";\nimport type { BatchOutcome } from \"../aggregate.js\";\nimport type { EvalResult, Outcome } from \"../types.js\";\nimport { labelKey, type EvalLabel, type LabelOutcome } from \"./labels.js\";\nimport type {\n  CalibrationCounts,\n  Disagreement,\n  JoinKind,\n  JudgeSetting,\n} from \"./types.js\";\n\n/**\n * How one label met one result, over the full (expected × actual) grid.\n *\n * The order matters. `skipped` and `error` are tested *first*, so a labelled\n * eval that produced no evidence can never be counted as agreement — the same\n * rule that keeps an errored judge run from becoming a silent pass.\n */\nexport function classify(expected: LabelOutcome, actual: Outcome): JoinKind {\n  if (actual === \"skipped\") return \"skipped\";\n  if (actual === \"error\") return \"error\";\n  if (actual === expected) return \"agree\";\n  if (expected === \"fail\" && actual === \"pass\") return \"false-pass\";\n  if (expected === \"pass\" && actual === \"fail\") return \"false-fail\";\n  if (actual === \"needs-review\") return \"review\";\n  return \"missed-review\";\n}\n\nexport interface JoinedLabel {\n  label: EvalLabel;\n  result: EvalResult;\n}\n\nexport interface JoinResult {\n  joined: JoinedLabel[];\n  /**\n   * Labels that matched no result. Never silently dropped: a typo'd eval id\n   * would otherwise deflate every count on the report by one.\n   */\n  unmatched: EvalLabel[];\n  /**\n   * Labels that omit `type:` over a name two artifact kinds both answer. Never\n   * bound to one of them: a skill and the slash command that types it can share\n   * a name in one trace, and picking whichever was indexed last would measure\n   * agreement against a verdict belonging to something else — a false pass or\n   * false fail reported with total confidence.\n   */\n  ambiguous: AmbiguousLabel[];\n}\n\n/** An untyped label and every result its name matched. */\nexport interface AmbiguousLabel {\n  label: EvalLabel;\n  candidates: EvalResult[];\n}\n\n/**\n * Match each label to the result its trace produced. The key is\n * `(trace file, artifact name, eval id)`, with the artifact *type* joining too\n * when the label states one — needed only for a skill and an agent sharing a\n * name, which is legal and does happen.\n */\nexport function joinLabels(\n  outcomes: BatchOutcome[],\n  labels: EvalLabel[],\n): JoinResult {\n  // Typed keys are unique by construction. Untyped ones are not: two artifact\n  // kinds may answer the same name, so keep every candidate rather than\n  // letting the last one indexed overwrite the rest.\n  const byType = new Map<string, EvalResult>();\n  const byName = new Map<string, EvalResult[]>();\n  for (const outcome of outcomes) {\n    if (!(\"report\" in outcome)) continue;\n    for (const result of outcome.report.evalResults) {\n      byType.set(\n        labelKey(\n          outcome.file,\n          result.artifactName,\n          result.evalName,\n          result.artifactType,\n        ),\n        result,\n      );\n      const key = labelKey(outcome.file, result.artifactName, result.evalName);\n      const seen = byName.get(key);\n      if (seen === undefined) byName.set(key, [result]);\n      else seen.push(result);\n    }\n  }\n\n  const joined: JoinedLabel[] = [];\n  const unmatched: EvalLabel[] = [];\n  const ambiguous: AmbiguousLabel[] = [];\n  for (const label of labels) {\n    if (label.type !== undefined) {\n      const result = byType.get(\n        labelKey(label.traceFile, label.artifact, label.eval, label.type),\n      );\n      if (result === undefined) unmatched.push(label);\n      else joined.push({ label, result });\n      continue;\n    }\n    const candidates =\n      byName.get(labelKey(label.traceFile, label.artifact, label.eval)) ?? [];\n    const kinds = new Set(candidates.map((r) => r.artifactType));\n    if (candidates.length === 0) unmatched.push(label);\n    else if (kinds.size > 1) ambiguous.push({ label, candidates });\n    else joined.push({ label, result: candidates[0] as EvalResult });\n  }\n  return { joined, unmatched, ambiguous };\n}\n\nexport interface Rescored {\n  outcome: Outcome;\n  /** The cache held fewer runs than the setting asks for. */\n  insufficient: boolean;\n  /**\n   * The arithmetic **at this setting**, not at the setting the corpus was\n   * judged with. A sweep judges deeper than most of its cells, so reporting\n   * the full cached ensemble beside a re-scored verdict would show a reader\n   * evidence that does not add up to the outcome next to it.\n   */\n  consensus?: Disagreement[\"consensus\"];\n}\n\n/** The numbers a reader needs from a consensus, without the run bodies. */\nconst summarise = (\n  consensus: NonNullable<EvalResult[\"consensus\"]>,\n): Disagreement[\"consensus\"] => ({\n  votes: consensus.votes,\n  agreement: consensus.agreement,\n  meanConfidence: consensus.meanConfidence,\n  runs: consensus.runs.length,\n});\n\n/**\n * What this result would have been at a different setting.\n *\n * A result with no `consensus` — deterministic, `human`, skipped, errored — is\n * returned untouched: zones and ensemble size do not reach it, and pretending\n * otherwise would make a sweep look like it moved things it cannot move.\n *\n * A smaller `ensembleRuns` re-scores the **first k** of the cached runs. That\n * is a sub-sample of the ensemble that was actually drawn, not a fresh k-run\n * ensemble; at temperature 0 the runs are independent samples of one prompt,\n * so it is an honest estimate, and the alternative is paying per grid cell.\n */\nexport function rescore(result: EvalResult, setting: JudgeSetting): Rescored {\n  const consensus = result.consensus;\n  if (consensus === undefined) {\n    return { outcome: result.outcome, insufficient: false };\n  }\n  const runs = consensus.runs;\n  if (runs.length < setting.ensembleRuns) {\n    // Never scale up by guessing. Report the setting as unreachable and leave\n    // the row at what it actually was.\n    return { outcome: result.outcome, insufficient: true };\n  }\n  const slice = runs.slice(0, setting.ensembleRuns);\n  const base = computeConsensus(slice);\n  const zone = zoneFor(base, setting.zones);\n  return {\n    outcome:\n      zone === \"auto-pass\" ? \"pass\" : zone === \"auto-fail\" ? \"fail\" : \"needs-review\",\n    insufficient: false,\n    consensus: {\n      votes: base.votes,\n      agreement: base.agreement,\n      meanConfidence: base.meanConfidence,\n      runs: slice.length,\n    },\n  };\n}\n\nconst emptyCounts = (): CalibrationCounts => ({\n  labels: 0,\n  scored: 0,\n  agree: 0,\n  falsePass: 0,\n  falseFail: 0,\n  review: 0,\n  missedReview: 0,\n  errored: 0,\n  skipped: 0,\n  reviewVolume: 0,\n  insufficient: 0,\n  agreement: null,\n});\n\nexport interface ScoreInput {\n  joined: JoinedLabel[];\n  /** Every eval result in the corpus, for the corpus-wide review volume. */\n  all: EvalResult[];\n}\n\nexport interface ScoreResult {\n  counts: CalibrationCounts;\n  disagreements: Disagreement[];\n  unscored: Disagreement[];\n}\n\n/**\n * Score the joined corpus, optionally at a setting other than the one it ran\n * at. Omit `setting` and the reported outcomes are used verbatim, so the\n * headline numbers of a plain `calibrate` are exactly the verdicts `run`\n * would have produced — no re-derivation standing between the two commands.\n */\nexport function score(input: ScoreInput, setting?: JudgeSetting): ScoreResult {\n  const counts = emptyCounts();\n  const disagreements: Disagreement[] = [];\n  const unscored: Disagreement[] = [];\n\n  for (const result of input.all) {\n    const outcome =\n      setting === undefined ? result.outcome : rescore(result, setting).outcome;\n    if (outcome === \"needs-review\") counts.reviewVolume += 1;\n  }\n\n  for (const { label, result } of input.joined) {\n    const scored: Rescored =\n      setting === undefined\n        ? { outcome: result.outcome, insufficient: false }\n        : rescore(result, setting);\n    if (scored.insufficient) counts.insufficient += 1;\n    const kind = classify(label.expected, scored.outcome);\n    counts.labels += 1;\n\n    const row: Disagreement = {\n      trace: label.traceFile,\n      artifactType: result.artifactType,\n      artifactName: result.artifactName,\n      evalName: result.evalName,\n      grader: result.grader,\n      expected: label.expected,\n      actual: scored.outcome,\n      kind,\n      ...(label.note !== undefined ? { note: label.note } : {}),\n      ...(result.skipReason !== undefined\n        ? { skipReason: result.skipReason }\n        : {}),\n      // At the setting being scored, falling back to the run's own numbers\n      // when nothing was re-scored.\n      ...(() => {\n        const c =\n          scored.consensus ??\n          (result.consensus !== undefined ? summarise(result.consensus) : undefined);\n        return c !== undefined ? { consensus: c } : {};\n      })(),\n    };\n\n    switch (kind) {\n      case \"skipped\":\n        counts.skipped += 1;\n        unscored.push(row);\n        break;\n      case \"agree\":\n        counts.agree += 1;\n        counts.scored += 1;\n        break;\n      case \"false-pass\":\n        counts.falsePass += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n      case \"false-fail\":\n        counts.falseFail += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n      case \"review\":\n        counts.review += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n      case \"missed-review\":\n        counts.missedReview += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n      case \"error\":\n        counts.errored += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n    }\n  }\n\n  counts.agreement = counts.scored > 0 ? counts.agree / counts.scored : null;\n  // Byte comparison, so the ubuntu and windows legs order a report the same\n  // way and two runs over one corpus stay diffable.\n  const byRow = (a: Disagreement, b: Disagreement): number => {\n    const ka = `${a.kind} ${a.trace} ${a.artifactName} ${a.evalName}`;\n    const kb = `${b.kind} ${b.trace} ${b.artifactName} ${b.evalName}`;\n    return ka < kb ? -1 : ka > kb ? 1 : 0;\n  };\n  disagreements.sort(byRow);\n  unscored.sort(byRow);\n  return { counts, disagreements, unscored };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAcA,SAAS,iBAAiB;AAC1B,SAAS,UAAU,eAAe;;;ACNlC,SAAS,kBAAkB,eAAe;AAkBnC,SAAS,SAAS,UAAwB,QAA2B;AAC1E,MAAI,WAAW,UAAW,QAAO;AACjC,MAAI,WAAW,QAAS,QAAO;AAC/B,MAAI,WAAW,SAAU,QAAO;AAChC,MAAI,aAAa,UAAU,WAAW,OAAQ,QAAO;AACrD,MAAI,aAAa,UAAU,WAAW,OAAQ,QAAO;AACrD,MAAI,WAAW,eAAgB,QAAO;AACtC,SAAO;AACT;AAoCO,SAAS,WACd,UACA,QACY;AAIZ,QAAM,SAAS,oBAAI,IAAwB;AAC3C,QAAM,SAAS,oBAAI,IAA0B;AAC7C,aAAW,WAAW,UAAU;AAC9B,QAAI,EAAE,YAAY,SAAU;AAC5B,eAAW,UAAU,QAAQ,OAAO,aAAa;AAC/C,aAAO;AAAA,QACL;AAAA,UACE,QAAQ;AAAA,UACR,OAAO;AAAA,UACP,OAAO;AAAA,UACP,OAAO;AAAA,QACT;AAAA,QACA;AAAA,MACF;AACA,YAAM,MAAM,SAAS,QAAQ,MAAM,OAAO,cAAc,OAAO,QAAQ;AACvE,YAAM,OAAO,OAAO,IAAI,GAAG;AAC3B,UAAI,SAAS,OAAW,QAAO,IAAI,KAAK,CAAC,MAAM,CAAC;AAAA,UAC3C,MAAK,KAAK,MAAM;AAAA,IACvB;AAAA,EACF;AAEA,QAAM,SAAwB,CAAC;AAC/B,QAAM,YAAyB,CAAC;AAChC,QAAM,YAA8B,CAAC;AACrC,aAAW,SAAS,QAAQ;AAC1B,QAAI,MAAM,SAAS,QAAW;AAC5B,YAAM,SAAS,OAAO;AAAA,QACpB,SAAS,MAAM,WAAW,MAAM,UAAU,MAAM,MAAM,MAAM,IAAI;AAAA,MAClE;AACA,UAAI,WAAW,OAAW,WAAU,KAAK,KAAK;AAAA,UACzC,QAAO,KAAK,EAAE,OAAO,OAAO,CAAC;AAClC;AAAA,IACF;AACA,UAAM,aACJ,OAAO,IAAI,SAAS,MAAM,WAAW,MAAM,UAAU,MAAM,IAAI,CAAC,KAAK,CAAC;AACxE,UAAM,QAAQ,IAAI,IAAI,WAAW,IAAI,CAAC,MAAM,EAAE,YAAY,CAAC;AAC3D,QAAI,WAAW,WAAW,EAAG,WAAU,KAAK,KAAK;AAAA,aACxC,MAAM,OAAO,EAAG,WAAU,KAAK,EAAE,OAAO,WAAW,CAAC;AAAA,QACxD,QAAO,KAAK,EAAE,OAAO,QAAQ,WAAW,CAAC,EAAgB,CAAC;AAAA,EACjE;AACA,SAAO,EAAE,QAAQ,WAAW,UAAU;AACxC;AAgBA,IAAM,YAAY,CAChB,eAC+B;AAAA,EAC/B,OAAO,UAAU;AAAA,EACjB,WAAW,UAAU;AAAA,EACrB,gBAAgB,UAAU;AAAA,EAC1B,MAAM,UAAU,KAAK;AACvB;AAcO,SAAS,QAAQ,QAAoB,SAAiC;AAC3E,QAAM,YAAY,OAAO;AACzB,MAAI,cAAc,QAAW;AAC3B,WAAO,EAAE,SAAS,OAAO,SAAS,cAAc,MAAM;AAAA,EACxD;AACA,QAAM,OAAO,UAAU;AACvB,MAAI,KAAK,SAAS,QAAQ,cAAc;AAGtC,WAAO,EAAE,SAAS,OAAO,SAAS,cAAc,KAAK;AAAA,EACvD;AACA,QAAM,QAAQ,KAAK,MAAM,GAAG,QAAQ,YAAY;AAChD,QAAM,OAAO,iBAAiB,KAAK;AACnC,QAAM,OAAO,QAAQ,MAAM,QAAQ,KAAK;AACxC,SAAO;AAAA,IACL,SACE,SAAS,cAAc,SAAS,SAAS,cAAc,SAAS;AAAA,IAClE,cAAc;AAAA,IACd,WAAW;AAAA,MACT,OAAO,KAAK;AAAA,MACZ,WAAW,KAAK;AAAA,MAChB,gBAAgB,KAAK;AAAA,MACrB,MAAM,MAAM;AAAA,IACd;AAAA,EACF;AACF;AAEA,IAAM,cAAc,OAA0B;AAAA,EAC5C,QAAQ;AAAA,EACR,QAAQ;AAAA,EACR,OAAO;AAAA,EACP,WAAW;AAAA,EACX,WAAW;AAAA,EACX,QAAQ;AAAA,EACR,cAAc;AAAA,EACd,SAAS;AAAA,EACT,SAAS;AAAA,EACT,cAAc;AAAA,EACd,cAAc;AAAA,EACd,WAAW;AACb;AAoBO,SAAS,MAAM,OAAmB,SAAqC;AAC5E,QAAM,SAAS,YAAY;AAC3B,QAAM,gBAAgC,CAAC;AACvC,QAAM,WAA2B,CAAC;AAElC,aAAW,UAAU,MAAM,KAAK;AAC9B,UAAM,UACJ,YAAY,SAAY,OAAO,UAAU,QAAQ,QAAQ,OAAO,EAAE;AACpE,QAAI,YAAY,eAAgB,QAAO,gBAAgB;AAAA,EACzD;AAEA,aAAW,EAAE,OAAO,OAAO,KAAK,MAAM,QAAQ;AAC5C,UAAM,SACJ,YAAY,SACR,EAAE,SAAS,OAAO,SAAS,cAAc,MAAM,IAC/C,QAAQ,QAAQ,OAAO;AAC7B,QAAI,OAAO,aAAc,QAAO,gBAAgB;AAChD,UAAM,OAAO,SAAS,MAAM,UAAU,OAAO,OAAO;AACpD,WAAO,UAAU;AAEjB,UAAM,MAAoB;AAAA,MACxB,OAAO,MAAM;AAAA,MACb,cAAc,OAAO;AAAA,MACrB,cAAc,OAAO;AAAA,MACrB,UAAU,OAAO;AAAA,MACjB,QAAQ,OAAO;AAAA,MACf,UAAU,MAAM;AAAA,MAChB,QAAQ,OAAO;AAAA,MACf;AAAA,MACA,GAAI,MAAM,SAAS,SAAY,EAAE,MAAM,MAAM,KAAK,IAAI,CAAC;AAAA,MACvD,GAAI,OAAO,eAAe,SACtB,EAAE,YAAY,OAAO,WAAW,IAChC,CAAC;AAAA;AAAA;AAAA,MAGL,IAAI,MAAM;AACR,cAAM,IACJ,OAAO,cACN,OAAO,cAAc,SAAY,UAAU,OAAO,SAAS,IAAI;AAClE,eAAO,MAAM,SAAY,EAAE,WAAW,EAAE,IAAI,CAAC;AAAA,MAC/C,GAAG;AAAA,IACL;AAEA,YAAQ,MAAM;AAAA,MACZ,KAAK;AACH,eAAO,WAAW;AAClB,iBAAS,KAAK,GAAG;AACjB;AAAA,MACF,KAAK;AACH,eAAO,SAAS;AAChB,eAAO,UAAU;AACjB;AAAA,MACF,KAAK;AACH,eAAO,aAAa;AACpB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,MACF,KAAK;AACH,eAAO,aAAa;AACpB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,MACF,KAAK;AACH,eAAO,UAAU;AACjB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,MACF,KAAK;AACH,eAAO,gBAAgB;AACvB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,MACF,KAAK;AACH,eAAO,WAAW;AAClB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,IACJ;AAAA,EACF;AAEA,SAAO,YAAY,OAAO,SAAS,IAAI,OAAO,QAAQ,OAAO,SAAS;AAGtE,QAAM,QAAQ,CAAC,GAAiB,MAA4B;AAC1D,UAAM,KAAK,GAAG,EAAE,IAAI,IAAI,EAAE,KAAK,IAAI,EAAE,YAAY,IAAI,EAAE,QAAQ;AAC/D,UAAM,KAAK,GAAG,EAAE,IAAI,IAAI,EAAE,KAAK,IAAI,EAAE,YAAY,IAAI,EAAE,QAAQ;AAC/D,WAAO,KAAK,KAAK,KAAK,KAAK,KAAK,IAAI;AAAA,EACtC;AACA,gBAAc,KAAK,KAAK;AACxB,WAAS,KAAK,KAAK;AACnB,SAAO,EAAE,QAAQ,eAAe,SAAS;AAC3C;;;ADhQA,SAAS,WAAW,MAAgB,UAA0B;AAC5D,SAAO,KAAK,IAAI,UAAU,GAAG,IAAI;AACnC;AAEA,eAAsB,aACpB,SACiC;AACjC,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,EAAE,QAAQ,KAAK,UAAU,IAAI,MAAM;AAAA,IACvC,QAAQ,aAAa,QAAQ,IAAI;AAAA,IACjC;AAAA,MACE,GAAI,QAAQ,WAAW,SAAY,CAAC,IAAI,EAAE,YAAY,QAAQ,OAAO;AAAA,MACrE,GAAI,QAAQ,aAAa,SAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;AAAA,IACzE;AAAA,EACF;AAMA;AAAA,IACE,eAAe,QAAQ;AAAA,MACrB,GAAI,QAAQ,aAAa,SAAY,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;AAAA,MACvE,GAAI,QAAQ,UAAU,SAAY,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;AAAA,MAC9D,GAAI,QAAQ,UAAU,SAAY,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;AAAA,IAChE,CAAC;AAAA,EACH;AAMA,QAAM,aACJ,QAAQ,WAAW,SACf,QAAQ,QAAQ,MAAM,IACtB,QAAQ,WAAW,OAAO,UAAU,MAAM;AAChD,QAAM,SAAS,MAAM,WAAW,UAAU;AAI1C,QAAM,UAAU,MAAM,mBAAmB,OAAO,GAAG,IAAI,CAAC,MAAM,QAAQ,CAAC,CAAC;AACxE,uBAAqB,QAAQ,QAAQ,UAAU;AAE/C,QAAM,WAAyB;AAAA,IAC7B,cAAc,QAAQ,QAAQ,OAAO,MAAM;AAAA,IAC3C,OAAO,OAAO,MAAM;AAAA,EACtB;AACA,QAAM,OAAO,OAAO,UAAU;AAK9B,QAAM,QACJ,QAAQ,UAAU,OACd,WAAW,KAAK,cAAc,SAAS,YAAY,IACnD,SAAS;AAEf,QAAM,QAAQ,MAAM,SAAS;AAAA,IAC3B,GAAG;AAAA,IACH,QAAQ;AAAA,IACR,aAAa;AAAA,IACb,WAAW;AAAA,IACX,OAAO;AAAA,IACP,MAAM;AAAA;AAAA;AAAA,IAGN,QAAQ;AAAA,IACR,QAAQ;AAAA,EACV,CAAC;AAED,QAAM,MAAoB,CAAC;AAC3B,aAAW,WAAW,MAAM,UAAU;AACpC,QAAI,YAAY,QAAS,KAAI,KAAK,GAAG,QAAQ,OAAO,WAAW;AAAA,EACjE;AACA,QAAM,EAAE,QAAQ,WAAW,UAAU,IAAI,WAAW,MAAM,UAAU,MAAM;AAC1E,oBAAkB,WAAW,OAAO,UAAU;AAC9C,oBAAkB,WAAW,UAAU;AAMvC,QAAM,SAAS;AAAA,IACb,EAAE,QAAQ,IAAI;AAAA,IACd,QAAQ,UAAU,OAAO,WAAW;AAAA,EACtC;AAEA,QAAM,QACJ,QAAQ,UAAU,OACd,WAAW,EAAE,QAAQ,IAAI,GAAG,UAAU,IAAI,IAC1C;AAEN,QAAM,QAAQ,WAAW,OAAO,QAAQ;AAAA,IACtC,WAAW,QAAQ,gBAAgB,OAAO,UAAU;AAAA,IACpD,WAAW,QAAQ,gBAAgB,OAAO,UAAU;AAAA,IACpD,QAAQ,QAAQ,aAAa,OAAO,UAAU;AAAA,EAChD,CAAC;AAED,QAAM,WAAW,CAAC,GAAG,MAAM,OAAO,QAAQ;AAC1C,MAAI,OAAO,OAAO,eAAe,GAAG;AAClC,aAAS;AAAA,MACP,GAAG,OAAO,OAAO,YAAY;AAAA,IAC/B;AAAA,EACF;AASA,QAAM,kBAAkB,OAAO,OAAO,WAAW;AACjD,MAAI,iBAAiB;AACnB,aAAS;AAAA,MACP,uCAAuC,OAAO,OAAO,MAAM;AAAA,IAE7D;AAAA,EACF;AAEA,QAAM,SAA4B;AAAA,IAChC;AAAA,IACA;AAAA,IACA,SAAS;AAAA,IACT,QAAQ,OAAO;AAAA,IACf,eAAe,OAAO;AAAA,IACtB,UAAU,OAAO;AAAA,IACjB,GAAI,UAAU,SAAY,EAAE,MAAM,IAAI,CAAC;AAAA,IACvC;AAAA,IACA,OAAO,MAAM;AAAA,IACb;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAMA,UACE,MAAM,KAAK,CAAC,MAAM,EAAE,QAAQ,KAC5B,MAAM,OAAO,QAAQ,gBAAgB,KACrC,MAAM,OAAO,WAAW,UACxB,kBACI,IACA;AAAA,IACN,OAAO,MAAM,OAAO;AAAA,IACpB,YAAY,KAAK,IAAI,IAAI;AAAA,EAC3B;AAEA,QAAM,WAAW,kBAAkB,QAAQ,QAAQ,UAAU,UAAU;AAAA,IACrE,OAAO,QAAQ,UAAU;AAAA,EAC3B,CAAC;AACD,MAAI,QAAQ,OAAQ,OAAM,UAAU,QAAQ,QAAQ,UAAU,OAAO;AACrE,SAAO,EAAE,QAAQ,SAAS;AAC5B;AAOA,SAAS,qBACP,QACA,QACA,YACM;AACN,QAAM,QAAQ,IAAI,IAAI,MAAM;AAC5B,QAAM,UAAU,CAAC,GAAG,IAAI,IAAI,OAAO,OAAO,CAAC,MAAM,CAAC,MAAM,IAAI,EAAE,SAAS,CAAC,EAAE,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC;AAC9F,MAAI,QAAQ,WAAW,EAAG;AAC1B,QAAM,IAAI;AAAA,IACR,GAAG,UAAU,+CAA+C,QAAQ,KAAK,IAAI,CAAC,oBAC1D,OAAO,IAAI,CAAC,MAAM,SAAS,CAAC,CAAC,EAAE,KAAK,IAAI,CAAC;AAAA,EAC/D;AACF;AASA,SAAS,kBACP,WACA,OACA,YACM;AACN,MAAI,UAAU,WAAW,EAAG;AAC5B,QAAM,QAAQ,UAAU,IAAI,CAAC,UAAU;AACrC,UAAM,UAAU,MAAM,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,MAAM,SAAS;AACrE,UAAM,UAAU,SAAS,QAAQ,eAAe,CAAC;AACjD,UAAM,eAAe,QAAQ;AAAA,MAC3B,CAAC,MAAM,EAAE,iBAAiB,MAAM;AAAA,IAClC;AACA,UAAM,OACJ,aAAa,SAAS,IAClB,wBAAwB,aAAa,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,KAAK,IAAI,CAAC,SAAS,MAAM,QAAQ,MACpG,0CAA0C,MAAM,QAAQ,yBACtD,CAAC,GAAG,IAAI,IAAI,QAAQ,IAAI,CAAC,MAAM,EAAE,YAAY,CAAC,CAAC,EAAE,KAAK,EAAE,KAAK,IAAI,KACjE,SACF;AACN,WAAO,KAAK,MAAM,QAAQ,WAAM,MAAM,IAAI,OAAO,SAAS,MAAM,SAAS,CAAC,KAAK,IAAI;AAAA,EACrF,CAAC;AACD,QAAM,IAAI;AAAA,IACR,GAAG,UAAU;AAAA,EAAuC,MAAM,KAAK,IAAI,CAAC;AAAA,EACtE;AACF;AAYA,SAAS,kBACP,WACA,YACM;AACN,MAAI,UAAU,WAAW,EAAG;AAC5B,QAAM,QAAQ,UAAU,IAAI,CAAC,EAAE,OAAO,WAAW,MAAM;AACrD,UAAM,QAAQ,CAAC,GAAG,IAAI,IAAI,WAAW,IAAI,CAAC,MAAM,EAAE,YAAY,CAAC,CAAC,EAAE,KAAK;AACvE,WAAO,KAAK,MAAM,QAAQ,WAAM,MAAM,IAAI,OAAO,SAAS,MAAM,SAAS,CAAC,YAAY,MAAM,KAAK,OAAO,CAAC;AAAA,EAC3G,CAAC;AACD,QAAM,IAAI;AAAA,IACR,GAAG,UAAU;AAAA,EAAkD,MAAM,KAAK,IAAI,CAAC;AAAA,EACjF;AACF;AAEA,SAAS,WACP,QACA,QACmB;AACnB,QAAM,QAA2B,CAAC;AAClC,QAAM,MAAM,CACV,MACA,OACA,WACS;AACT,QAAI,UAAU,OAAW;AACzB,UAAM,KAAK,EAAE,MAAM,OAAO,QAAQ,UAAU,SAAS,MAAM,CAAC;AAAA,EAC9D;AACA,MAAI,aAAa,OAAO,WAAW,OAAO,SAAS;AACnD,MAAI,aAAa,OAAO,WAAW,OAAO,SAAS;AACnD,MAAI,UAAU,OAAO,QAAQ,OAAO,YAAY;AAChD,SAAO;AACT;AASA,SAAS,WACP,OACA,UACA,MACa;AACb,QAAM,OAAO,CACX,MACA,OACA,aACe;AAAA,IACf;AAAA,IACA;AAAA,IACA;AAAA,IACA,QAAQ,MAAM,OAAO,OAAO,EAAE;AAAA,EAChC;AAEA,QAAM,QAAqB;AAAA,IACzB,KAAK,YAAY,SAAS,cAAc,QAAQ;AAAA,EAClD;AACA,aAAW,QAAQ,CAAC,GAAG,KAAK,YAAY,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC,GAAG;AAC/D,UAAM;AAAA,MACJ,KAAK,gBAAgB,MAAM,EAAE,cAAc,MAAM,OAAO,SAAS,MAAM,CAAC;AAAA,IAC1E;AAAA,EACF;AACA,aAAW,YAAY,CAAC,GAAG,KAAK,QAAQ,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC,GAAG;AAC/D,UAAM;AAAA,MACJ,KAAK,kBAAkB,UAAU;AAAA,QAC/B,cAAc,SAAS;AAAA,QACvB,OAAO,EAAE,UAAU,UAAU,SAAS,MAAM,SAAS;AAAA,MACvD,CAAC;AAAA,IACH;AAAA,EACF;AACA,aAAW,YAAY,CAAC,GAAG,KAAK,QAAQ,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC,GAAG;AAC/D,UAAM;AAAA,MACJ,KAAK,kBAAkB,UAAU;AAAA,QAC/B,cAAc,SAAS;AAAA,QACvB,OAAO,EAAE,UAAU,SAAS,MAAM,UAAU,SAAS;AAAA,MACvD,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../src/tracevals/commands/calibrate.ts","../src/tracevals/calibrate/score.ts"],"sourcesContent":["/**\n * `manni tracevals calibrate <traces...> --labels <file>` (ADR 01022).\n *\n * The tool could report what it decided; it could not report whether it was\n * right. This command runs the batch (ADR 01018), joins the results against a\n * human's answers, and reports the three numbers the calibration page used to\n * ask people to count by hand — false passes, false fails, and review volume —\n * plus which eval disagreed.\n *\n * `--sweep` costs no inference beyond the one run it already made: a judged\n * eval's consensus carries the whole `JudgeRun[]`, so a different zone\n * threshold or a smaller ensemble is arithmetic over verdicts already on disk.\n * Read-only throughout — labels, traces, and artifacts are never written.\n */\nimport { writeFile } from \"node:fs/promises\";\nimport { basename, resolve } from \"node:path\";\nimport { discoverConfig } from \"../core/config.js\";\nimport { assertProviderSelection, selectProvider } from \"../judge/provider.js\";\nimport { renderCalibration, type ReportFormat } from \"../reporters/index.js\";\nimport { TracevalsError, type EvalResult } from \"../types.js\";\nimport { loadLabels, type EvalLabel } from \"../calibrate/labels.js\";\nimport { joinLabels, score, type AmbiguousLabel } from \"../calibrate/score.js\";\nimport type {\n  CalibrationGate,\n  CalibrationReport,\n  JudgeSetting,\n  SweepCell,\n} from \"../calibrate/types.js\";\nimport {\n  resolveBatchTraces,\n  runBatch,\n  type BatchCommandOptions,\n} from \"./batch.js\";\n\nexport interface CalibrateCommandOptions extends BatchCommandOptions {\n  /** Overrides `calibrate.labels`; resolved against the working directory. */\n  labels?: string;\n  /** Re-score the corpus across the configured grid. */\n  sweep?: boolean;\n  /** Thresholds; undefined defers to the config, which defaults to no gate. */\n  maxFalsePass?: number;\n  maxFalseFail?: number;\n  maxReview?: number;\n}\n\nexport interface CalibrateCommandResult {\n  report: CalibrationReport;\n  rendered: string;\n}\n\n/** Largest ensemble the sweep needs cached, so every cell can be re-scored. */\nfunction sweepDepth(grid: number[], baseline: number): number {\n  return Math.max(baseline, ...grid);\n}\n\nexport async function runCalibrate(\n  options: CalibrateCommandOptions,\n): Promise<CalibrateCommandResult> {\n  const start = Date.now();\n  const { config, dir: configDir } = await discoverConfig(\n    options.configDir ?? process.cwd(),\n    {\n      ...(options.config === undefined ? {} : { configPath: options.config }),\n      ...(options.noConfig === undefined ? {} : { noConfig: options.noConfig }),\n    },\n  );\n\n  // Before the labels file is read: a contradictory `--local --provider`, an\n  // unknown provider name, or a model with no provider to own it is a usage\n  // error, and a usage error must not arrive behind an unrelated one about a\n  // file the run would never have needed.\n  assertProviderSelection(\n    selectProvider(config, {\n      ...(options.provider !== undefined ? { provider: options.provider } : {}),\n      ...(options.model !== undefined ? { model: options.model } : {}),\n      ...(options.local !== undefined ? { local: options.local } : {}),\n    }),\n  );\n\n  // A flag overrides the config rather than bypassing it (CLAUDE.md,\n  // \"Config <-> CLI flags\"). An explicit --labels is a working-directory path,\n  // the way every other path argument on the CLI is; the config default is a\n  // project path and resolves against the config file's directory.\n  const labelsFile =\n    options.labels !== undefined\n      ? resolve(options.labels)\n      : resolve(configDir, config.calibrate.labels);\n  const labels = await loadLabels(labelsFile);\n\n  // Resolve the corpus once and hand the list to the batch, so the membership\n  // check below cannot disagree with what actually ran.\n  const corpus = (await resolveBatchTraces(options)).map((f) => resolve(f));\n  assertLabelsInCorpus(labels, corpus, labelsFile);\n\n  const baseline: JudgeSetting = {\n    ensembleRuns: options.runs ?? config.judge.ensembleRuns,\n    zones: config.judge.zones,\n  };\n  const grid = config.calibrate.sweep;\n  // With --sweep the corpus is judged once at the deepest ensemble the grid\n  // asks for, and every cell — the baseline included — is re-scored from those\n  // runs. Judging at the configured depth instead would make the larger cells\n  // unscoreable, and judging per cell would multiply the bill by the grid.\n  const depth =\n    options.sweep === true\n      ? sweepDepth(grid.ensembleRuns, baseline.ensembleRuns)\n      : baseline.ensembleRuns;\n\n  const batch = await runBatch({\n    ...options,\n    traces: corpus,\n    allProjects: undefined,\n    newerThan: undefined,\n    limit: undefined,\n    runs: depth,\n    // The calibration report is the deliverable; rendering the batch as well\n    // would write the wrong thing to --output.\n    format: \"json\",\n    output: undefined,\n  });\n\n  const all: EvalResult[] = [];\n  for (const outcome of batch.outcomes) {\n    if (\"report\" in outcome) all.push(...outcome.report.evalResults);\n  }\n  const { joined, unmatched, ambiguous } = joinLabels(batch.outcomes, labels);\n  assertNoUnmatched(unmatched, batch, labelsFile);\n  assertNoAmbiguous(ambiguous, labelsFile);\n\n  // Without a sweep the reported outcomes are used verbatim, so the headline\n  // numbers are exactly the verdicts `run` produced. With one, the corpus was\n  // judged deeper than the configured ensemble, so the baseline row has to be\n  // re-scored back down to it.\n  const scored = score(\n    { joined, all },\n    options.sweep === true ? baseline : undefined,\n  );\n\n  const sweep =\n    options.sweep === true\n      ? buildSweep({ joined, all }, baseline, grid)\n      : undefined;\n\n  const gates = buildGates(scored.counts, {\n    falsePass: options.maxFalsePass ?? config.calibrate.maxFalsePass,\n    falseFail: options.maxFalseFail ?? config.calibrate.maxFalseFail,\n    review: options.maxReview ?? config.calibrate.maxReview,\n  });\n\n  const warnings = [...batch.report.warnings];\n  if (scored.counts.insufficient > 0) {\n    warnings.push(\n      `${scored.counts.insufficient} labelled eval(s) had fewer cached runs than the setting asks for and were left at their reported outcome`,\n    );\n  }\n\n  // A measurement that measured nothing must not report a pass.\n  //\n  // Every label joining to a `skipped` result — `--deterministic-only` over\n  // ai-graded labels, a budget exhausted on the first trace — leaves `scored`\n  // at 0, `agreement` at null, and each threshold reading \"0 of at most N, not\n  // exceeded\". The gates are not lying; the denominator behind them is empty.\n  // Saying so is the report's job, and exit 0 would say the opposite.\n  const measuredNothing = scored.counts.scored === 0;\n  if (measuredNothing) {\n    warnings.push(\n      `no labelled eval produced evidence: ${scored.counts.labels} label(s) all joined to a skipped result, ` +\n        `so agreement and every --max-* threshold are computed over an empty denominator`,\n    );\n  }\n\n  const report: CalibrationReport = {\n    labelsFile,\n    corpus,\n    setting: baseline,\n    counts: scored.counts,\n    disagreements: scored.disagreements,\n    unscored: scored.unscored,\n    ...(sweep !== undefined ? { sweep } : {}),\n    gates,\n    batch: batch.report,\n    warnings,\n    // A calibration run is a measurement, not a gate: disagreement is the\n    // finding, not a failure. Exit 1 only when a threshold was asked for and\n    // missed, or when the measurement itself is incomplete — a trace lost, the\n    // shared judge budget cut the corpus short, or nothing scored at all. An\n    // incomplete number presented as a clean one is worse than no number.\n    exitCode:\n      gates.some((g) => g.exceeded) ||\n      batch.report.summary.tracesErrored > 0 ||\n      batch.report.budget !== undefined ||\n      measuredNothing\n        ? 1\n        : 0,\n    turns: batch.report.turns,\n    durationMs: Date.now() - start,\n  };\n\n  const rendered = renderCalibration(report, options.format ?? \"pretty\", {\n    color: options.color === true,\n  });\n  if (options.output) await writeFile(options.output, rendered, \"utf-8\");\n  return { report, rendered };\n}\n\n/**\n * Every label must name a trace this run evaluates. Checked before the batch,\n * because the alternative is paying for a corpus and then being told the file\n * was misspelled.\n */\nfunction assertLabelsInCorpus(\n  labels: EvalLabel[],\n  corpus: string[],\n  labelsFile: string,\n): void {\n  const known = new Set(corpus);\n  const missing = [...new Set(labels.filter((l) => !known.has(l.traceFile)).map((l) => l.trace))];\n  if (missing.length === 0) return;\n  throw new TracevalsError(\n    `${labelsFile}: label(s) name trace(s) not in the corpus: ${missing.join(\", \")}. ` +\n      `The corpus is: ${corpus.map((f) => basename(f)).join(\", \")}`,\n  );\n}\n\n/**\n * A label that matched nothing is an error, not a quiet omission.\n *\n * It has to be: a typo'd eval id contributes to no count, so the report would\n * read as a cleaner corpus than it is — the same silent-nothing failure the\n * empty-selector rule exists to prevent (ADR 01018).\n */\nfunction assertNoUnmatched(\n  unmatched: EvalLabel[],\n  batch: { outcomes: { file: string; report?: { evalResults: EvalResult[] } }[] },\n  labelsFile: string,\n): void {\n  if (unmatched.length === 0) return;\n  const lines = unmatched.map((label) => {\n    const outcome = batch.outcomes.find((o) => o.file === label.traceFile);\n    const results = outcome?.report?.evalResults ?? [];\n    const sameArtifact = results.filter(\n      (r) => r.artifactName === label.artifact,\n    );\n    const hint =\n      sameArtifact.length > 0\n        ? `that trace evaluated ${sameArtifact.map((r) => r.evalName).sort().join(\", \")} for \"${label.artifact}\"`\n        : `that trace resolved no artifact named \"${label.artifact}\" — it evaluated ${\n            [...new Set(results.map((r) => r.artifactName))].sort().join(\", \") ||\n            \"nothing\"\n          }`;\n    return `  ${label.artifact} › ${label.eval} on ${basename(label.traceFile)} (${hint})`;\n  });\n  throw new TracevalsError(\n    `${labelsFile}: label(s) matched no eval result:\\n${lines.join(\"\\n\")}`,\n  );\n}\n\n/**\n * A label that omits `type:` over a name two artifact kinds both answer cannot\n * be resolved, and must not be guessed. Binding it to whichever result was\n * indexed last would score agreement against a verdict belonging to a\n * different artifact — the report would read as measured when it is wrong,\n * which is the one thing this command exists not to do.\n *\n * Operational, like `assertNoUnmatched`: the labels file is under-specified,\n * so this is exit 2 (bad input), not exit 1 (a check failed).\n */\nfunction assertNoAmbiguous(\n  ambiguous: AmbiguousLabel[],\n  labelsFile: string,\n): void {\n  if (ambiguous.length === 0) return;\n  const lines = ambiguous.map(({ label, candidates }) => {\n    const kinds = [...new Set(candidates.map((r) => r.artifactType))].sort();\n    return `  ${label.artifact} › ${label.eval} on ${basename(label.traceFile)} matches ${kinds.join(\" and \")} — add \\`type:\\` to say which`;\n  });\n  throw new TracevalsError(\n    `${labelsFile}: label(s) name an artifact two kinds answer:\\n${lines.join(\"\\n\")}`,\n  );\n}\n\nfunction buildGates(\n  counts: { falsePass: number; falseFail: number; reviewVolume: number },\n  limits: { falsePass?: number; falseFail?: number; review?: number },\n): CalibrationGate[] {\n  const gates: CalibrationGate[] = [];\n  const add = (\n    name: CalibrationGate[\"name\"],\n    limit: number | undefined,\n    actual: number,\n  ): void => {\n    if (limit === undefined) return;\n    gates.push({ name, limit, actual, exceeded: actual > limit });\n  };\n  add(\"falsePass\", limits.falsePass, counts.falsePass);\n  add(\"falseFail\", limits.falseFail, counts.falseFail);\n  add(\"review\", limits.review, counts.reviewVolume);\n  return gates;\n}\n\n/**\n * One axis at a time, every other knob held at its configured value.\n *\n * A full cross-product would be dozens of rows and would answer a question\n * nobody asked; calibration is \"move one knob, re-run, recount\", and the\n * per-axis shape is what makes a row readable as *the effect of that knob*.\n */\nfunction buildSweep(\n  input: Parameters<typeof score>[0],\n  baseline: JudgeSetting,\n  grid: { ensembleRuns: number[]; autoPass: number[]; autoFail: number[] },\n): SweepCell[] {\n  const cell = (\n    axis: SweepCell[\"axis\"],\n    value: number,\n    setting: JudgeSetting,\n  ): SweepCell => ({\n    axis,\n    value,\n    setting,\n    counts: score(input, setting).counts,\n  });\n\n  const cells: SweepCell[] = [\n    cell(\"baseline\", baseline.ensembleRuns, baseline),\n  ];\n  for (const runs of [...grid.ensembleRuns].sort((a, b) => a - b)) {\n    cells.push(\n      cell(\"ensembleRuns\", runs, { ensembleRuns: runs, zones: baseline.zones }),\n    );\n  }\n  for (const autoPass of [...grid.autoPass].sort((a, b) => a - b)) {\n    cells.push(\n      cell(\"zones.autoPass\", autoPass, {\n        ensembleRuns: baseline.ensembleRuns,\n        zones: { autoPass, autoFail: baseline.zones.autoFail },\n      }),\n    );\n  }\n  for (const autoFail of [...grid.autoFail].sort((a, b) => a - b)) {\n    cells.push(\n      cell(\"zones.autoFail\", autoFail, {\n        ensembleRuns: baseline.ensembleRuns,\n        zones: { autoPass: baseline.zones.autoPass, autoFail },\n      }),\n    );\n  }\n  return cells;\n}\n\nexport type { ReportFormat };\n","/**\n * Joining labels to results, and re-scoring cached verdicts (ADR 01022).\n *\n * Everything here is pure: results and labels in, counts out. That is what\n * makes a sweep free. A judged eval's `ConsensusResult` carries the whole\n * `JudgeRun[]` it was built from, so a different zone threshold — or a smaller\n * ensemble — is a fresh call to `computeConsensus`/`zoneFor` over runs the\n * corpus already produced, not a fresh call to a provider.\n */\nimport { computeConsensus, zoneFor } from \"@hawkeyexl/inference\";\nimport type { BatchOutcome } from \"../aggregate.js\";\nimport type { EvalResult, Outcome } from \"../types.js\";\nimport { labelKey, type EvalLabel, type LabelOutcome } from \"./labels.js\";\nimport type {\n  CalibrationCounts,\n  Disagreement,\n  JoinKind,\n  JudgeSetting,\n} from \"./types.js\";\n\n/**\n * How one label met one result, over the full (expected × actual) grid.\n *\n * The order matters. `skipped` and `error` are tested *first*, so a labelled\n * eval that produced no evidence can never be counted as agreement — the same\n * rule that keeps an errored judge run from becoming a silent pass.\n */\nexport function classify(expected: LabelOutcome, actual: Outcome): JoinKind {\n  if (actual === \"skipped\") return \"skipped\";\n  if (actual === \"error\") return \"error\";\n  if (actual === expected) return \"agree\";\n  if (expected === \"fail\" && actual === \"pass\") return \"false-pass\";\n  if (expected === \"pass\" && actual === \"fail\") return \"false-fail\";\n  if (actual === \"needs-review\") return \"review\";\n  return \"missed-review\";\n}\n\nexport interface JoinedLabel {\n  label: EvalLabel;\n  result: EvalResult;\n}\n\nexport interface JoinResult {\n  joined: JoinedLabel[];\n  /**\n   * Labels that matched no result. Never silently dropped: a typo'd eval id\n   * would otherwise deflate every count on the report by one.\n   */\n  unmatched: EvalLabel[];\n  /**\n   * Labels that omit `type:` over a name two artifact kinds both answer. Never\n   * bound to one of them: a skill and the slash command that types it can share\n   * a name in one trace, and picking whichever was indexed last would measure\n   * agreement against a verdict belonging to something else — a false pass or\n   * false fail reported with total confidence.\n   */\n  ambiguous: AmbiguousLabel[];\n}\n\n/** An untyped label and every result its name matched. */\nexport interface AmbiguousLabel {\n  label: EvalLabel;\n  candidates: EvalResult[];\n}\n\n/**\n * Match each label to the result its trace produced. The key is\n * `(trace file, artifact name, eval id)`, with the artifact *type* joining too\n * when the label states one — needed only for a skill and an agent sharing a\n * name, which is legal and does happen.\n */\nexport function joinLabels(\n  outcomes: BatchOutcome[],\n  labels: EvalLabel[],\n): JoinResult {\n  // Typed keys are unique by construction. Untyped ones are not: two artifact\n  // kinds may answer the same name, so keep every candidate rather than\n  // letting the last one indexed overwrite the rest.\n  const byType = new Map<string, EvalResult>();\n  const byName = new Map<string, EvalResult[]>();\n  for (const outcome of outcomes) {\n    if (!(\"report\" in outcome)) continue;\n    for (const result of outcome.report.evalResults) {\n      byType.set(\n        labelKey(\n          outcome.file,\n          result.artifactName,\n          result.evalName,\n          result.artifactType,\n        ),\n        result,\n      );\n      const key = labelKey(outcome.file, result.artifactName, result.evalName);\n      const seen = byName.get(key);\n      if (seen === undefined) byName.set(key, [result]);\n      else seen.push(result);\n    }\n  }\n\n  const joined: JoinedLabel[] = [];\n  const unmatched: EvalLabel[] = [];\n  const ambiguous: AmbiguousLabel[] = [];\n  for (const label of labels) {\n    if (label.type !== undefined) {\n      const result = byType.get(\n        labelKey(label.traceFile, label.artifact, label.eval, label.type),\n      );\n      if (result === undefined) unmatched.push(label);\n      else joined.push({ label, result });\n      continue;\n    }\n    const candidates =\n      byName.get(labelKey(label.traceFile, label.artifact, label.eval)) ?? [];\n    const kinds = new Set(candidates.map((r) => r.artifactType));\n    if (candidates.length === 0) unmatched.push(label);\n    else if (kinds.size > 1) ambiguous.push({ label, candidates });\n    else joined.push({ label, result: candidates[0] as EvalResult });\n  }\n  return { joined, unmatched, ambiguous };\n}\n\nexport interface Rescored {\n  outcome: Outcome;\n  /** The cache held fewer runs than the setting asks for. */\n  insufficient: boolean;\n  /**\n   * The arithmetic **at this setting**, not at the setting the corpus was\n   * judged with. A sweep judges deeper than most of its cells, so reporting\n   * the full cached ensemble beside a re-scored verdict would show a reader\n   * evidence that does not add up to the outcome next to it.\n   */\n  consensus?: Disagreement[\"consensus\"];\n}\n\n/** The numbers a reader needs from a consensus, without the run bodies. */\nconst summarise = (\n  consensus: NonNullable<EvalResult[\"consensus\"]>,\n): Disagreement[\"consensus\"] => ({\n  votes: consensus.votes,\n  agreement: consensus.agreement,\n  meanConfidence: consensus.meanConfidence,\n  runs: consensus.runs.length,\n});\n\n/**\n * What this result would have been at a different setting.\n *\n * A result with no `consensus` — deterministic, `human`, skipped, errored — is\n * returned untouched: zones and ensemble size do not reach it, and pretending\n * otherwise would make a sweep look like it moved things it cannot move.\n *\n * A smaller `ensembleRuns` re-scores the **first k** of the cached runs. That\n * is a sub-sample of the ensemble that was actually drawn, not a fresh k-run\n * ensemble; at temperature 0 the runs are independent samples of one prompt,\n * so it is an honest estimate, and the alternative is paying per grid cell.\n */\nexport function rescore(result: EvalResult, setting: JudgeSetting): Rescored {\n  const consensus = result.consensus;\n  if (consensus === undefined) {\n    return { outcome: result.outcome, insufficient: false };\n  }\n  const runs = consensus.runs;\n  if (runs.length < setting.ensembleRuns) {\n    // Never scale up by guessing. Report the setting as unreachable and leave\n    // the row at what it actually was.\n    return { outcome: result.outcome, insufficient: true };\n  }\n  const slice = runs.slice(0, setting.ensembleRuns);\n  const base = computeConsensus(slice);\n  const zone = zoneFor(base, setting.zones);\n  return {\n    outcome:\n      zone === \"auto-pass\" ? \"pass\" : zone === \"auto-fail\" ? \"fail\" : \"needs-review\",\n    insufficient: false,\n    consensus: {\n      votes: base.votes,\n      agreement: base.agreement,\n      meanConfidence: base.meanConfidence,\n      runs: slice.length,\n    },\n  };\n}\n\nconst emptyCounts = (): CalibrationCounts => ({\n  labels: 0,\n  scored: 0,\n  agree: 0,\n  falsePass: 0,\n  falseFail: 0,\n  review: 0,\n  missedReview: 0,\n  errored: 0,\n  skipped: 0,\n  reviewVolume: 0,\n  insufficient: 0,\n  agreement: null,\n});\n\nexport interface ScoreInput {\n  joined: JoinedLabel[];\n  /** Every eval result in the corpus, for the corpus-wide review volume. */\n  all: EvalResult[];\n}\n\nexport interface ScoreResult {\n  counts: CalibrationCounts;\n  disagreements: Disagreement[];\n  unscored: Disagreement[];\n}\n\n/**\n * Score the joined corpus, optionally at a setting other than the one it ran\n * at. Omit `setting` and the reported outcomes are used verbatim, so the\n * headline numbers of a plain `calibrate` are exactly the verdicts `run`\n * would have produced — no re-derivation standing between the two commands.\n */\nexport function score(input: ScoreInput, setting?: JudgeSetting): ScoreResult {\n  const counts = emptyCounts();\n  const disagreements: Disagreement[] = [];\n  const unscored: Disagreement[] = [];\n\n  for (const result of input.all) {\n    const outcome =\n      setting === undefined ? result.outcome : rescore(result, setting).outcome;\n    if (outcome === \"needs-review\") counts.reviewVolume += 1;\n  }\n\n  for (const { label, result } of input.joined) {\n    const scored: Rescored =\n      setting === undefined\n        ? { outcome: result.outcome, insufficient: false }\n        : rescore(result, setting);\n    if (scored.insufficient) counts.insufficient += 1;\n    const kind = classify(label.expected, scored.outcome);\n    counts.labels += 1;\n\n    const row: Disagreement = {\n      trace: label.traceFile,\n      artifactType: result.artifactType,\n      artifactName: result.artifactName,\n      evalName: result.evalName,\n      grader: result.grader,\n      expected: label.expected,\n      actual: scored.outcome,\n      kind,\n      ...(label.note !== undefined ? { note: label.note } : {}),\n      ...(result.skipReason !== undefined\n        ? { skipReason: result.skipReason }\n        : {}),\n      // At the setting being scored, falling back to the run's own numbers\n      // when nothing was re-scored.\n      ...(() => {\n        const c =\n          scored.consensus ??\n          (result.consensus !== undefined ? summarise(result.consensus) : undefined);\n        return c !== undefined ? { consensus: c } : {};\n      })(),\n    };\n\n    switch (kind) {\n      case \"skipped\":\n        counts.skipped += 1;\n        unscored.push(row);\n        break;\n      case \"agree\":\n        counts.agree += 1;\n        counts.scored += 1;\n        break;\n      case \"false-pass\":\n        counts.falsePass += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n      case \"false-fail\":\n        counts.falseFail += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n      case \"review\":\n        counts.review += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n      case \"missed-review\":\n        counts.missedReview += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n      case \"error\":\n        counts.errored += 1;\n        counts.scored += 1;\n        disagreements.push(row);\n        break;\n    }\n  }\n\n  counts.agreement = counts.scored > 0 ? counts.agree / counts.scored : null;\n  // Byte comparison, so the ubuntu and windows legs order a report the same\n  // way and two runs over one corpus stay diffable.\n  const byRow = (a: Disagreement, b: Disagreement): number => {\n    const ka = `${a.kind} ${a.trace} ${a.artifactName} ${a.evalName}`;\n    const kb = `${b.kind} ${b.trace} ${b.artifactName} ${b.evalName}`;\n    return ka < kb ? -1 : ka > kb ? 1 : 0;\n  };\n  disagreements.sort(byRow);\n  unscored.sort(byRow);\n  return { counts, disagreements, unscored };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAcA,SAAS,iBAAiB;AAC1B,SAAS,UAAU,eAAe;;;ACNlC,SAAS,kBAAkB,eAAe;AAkBnC,SAAS,SAAS,UAAwB,QAA2B;AAC1E,MAAI,WAAW,UAAW,QAAO;AACjC,MAAI,WAAW,QAAS,QAAO;AAC/B,MAAI,WAAW,SAAU,QAAO;AAChC,MAAI,aAAa,UAAU,WAAW,OAAQ,QAAO;AACrD,MAAI,aAAa,UAAU,WAAW,OAAQ,QAAO;AACrD,MAAI,WAAW,eAAgB,QAAO;AACtC,SAAO;AACT;AAoCO,SAAS,WACd,UACA,QACY;AAIZ,QAAM,SAAS,oBAAI,IAAwB;AAC3C,QAAM,SAAS,oBAAI,IAA0B;AAC7C,aAAW,WAAW,UAAU;AAC9B,QAAI,EAAE,YAAY,SAAU;AAC5B,eAAW,UAAU,QAAQ,OAAO,aAAa;AAC/C,aAAO;AAAA,QACL;AAAA,UACE,QAAQ;AAAA,UACR,OAAO;AAAA,UACP,OAAO;AAAA,UACP,OAAO;AAAA,QACT;AAAA,QACA;AAAA,MACF;AACA,YAAM,MAAM,SAAS,QAAQ,MAAM,OAAO,cAAc,OAAO,QAAQ;AACvE,YAAM,OAAO,OAAO,IAAI,GAAG;AAC3B,UAAI,SAAS,OAAW,QAAO,IAAI,KAAK,CAAC,MAAM,CAAC;AAAA,UAC3C,MAAK,KAAK,MAAM;AAAA,IACvB;AAAA,EACF;AAEA,QAAM,SAAwB,CAAC;AAC/B,QAAM,YAAyB,CAAC;AAChC,QAAM,YAA8B,CAAC;AACrC,aAAW,SAAS,QAAQ;AAC1B,QAAI,MAAM,SAAS,QAAW;AAC5B,YAAM,SAAS,OAAO;AAAA,QACpB,SAAS,MAAM,WAAW,MAAM,UAAU,MAAM,MAAM,MAAM,IAAI;AAAA,MAClE;AACA,UAAI,WAAW,OAAW,WAAU,KAAK,KAAK;AAAA,UACzC,QAAO,KAAK,EAAE,OAAO,OAAO,CAAC;AAClC;AAAA,IACF;AACA,UAAM,aACJ,OAAO,IAAI,SAAS,MAAM,WAAW,MAAM,UAAU,MAAM,IAAI,CAAC,KAAK,CAAC;AACxE,UAAM,QAAQ,IAAI,IAAI,WAAW,IAAI,CAAC,MAAM,EAAE,YAAY,CAAC;AAC3D,QAAI,WAAW,WAAW,EAAG,WAAU,KAAK,KAAK;AAAA,aACxC,MAAM,OAAO,EAAG,WAAU,KAAK,EAAE,OAAO,WAAW,CAAC;AAAA,QACxD,QAAO,KAAK,EAAE,OAAO,QAAQ,WAAW,CAAC,EAAgB,CAAC;AAAA,EACjE;AACA,SAAO,EAAE,QAAQ,WAAW,UAAU;AACxC;AAgBA,IAAM,YAAY,CAChB,eAC+B;AAAA,EAC/B,OAAO,UAAU;AAAA,EACjB,WAAW,UAAU;AAAA,EACrB,gBAAgB,UAAU;AAAA,EAC1B,MAAM,UAAU,KAAK;AACvB;AAcO,SAAS,QAAQ,QAAoB,SAAiC;AAC3E,QAAM,YAAY,OAAO;AACzB,MAAI,cAAc,QAAW;AAC3B,WAAO,EAAE,SAAS,OAAO,SAAS,cAAc,MAAM;AAAA,EACxD;AACA,QAAM,OAAO,UAAU;AACvB,MAAI,KAAK,SAAS,QAAQ,cAAc;AAGtC,WAAO,EAAE,SAAS,OAAO,SAAS,cAAc,KAAK;AAAA,EACvD;AACA,QAAM,QAAQ,KAAK,MAAM,GAAG,QAAQ,YAAY;AAChD,QAAM,OAAO,iBAAiB,KAAK;AACnC,QAAM,OAAO,QAAQ,MAAM,QAAQ,KAAK;AACxC,SAAO;AAAA,IACL,SACE,SAAS,cAAc,SAAS,SAAS,cAAc,SAAS;AAAA,IAClE,cAAc;AAAA,IACd,WAAW;AAAA,MACT,OAAO,KAAK;AAAA,MACZ,WAAW,KAAK;AAAA,MAChB,gBAAgB,KAAK;AAAA,MACrB,MAAM,MAAM;AAAA,IACd;AAAA,EACF;AACF;AAEA,IAAM,cAAc,OAA0B;AAAA,EAC5C,QAAQ;AAAA,EACR,QAAQ;AAAA,EACR,OAAO;AAAA,EACP,WAAW;AAAA,EACX,WAAW;AAAA,EACX,QAAQ;AAAA,EACR,cAAc;AAAA,EACd,SAAS;AAAA,EACT,SAAS;AAAA,EACT,cAAc;AAAA,EACd,cAAc;AAAA,EACd,WAAW;AACb;AAoBO,SAAS,MAAM,OAAmB,SAAqC;AAC5E,QAAM,SAAS,YAAY;AAC3B,QAAM,gBAAgC,CAAC;AACvC,QAAM,WAA2B,CAAC;AAElC,aAAW,UAAU,MAAM,KAAK;AAC9B,UAAM,UACJ,YAAY,SAAY,OAAO,UAAU,QAAQ,QAAQ,OAAO,EAAE;AACpE,QAAI,YAAY,eAAgB,QAAO,gBAAgB;AAAA,EACzD;AAEA,aAAW,EAAE,OAAO,OAAO,KAAK,MAAM,QAAQ;AAC5C,UAAM,SACJ,YAAY,SACR,EAAE,SAAS,OAAO,SAAS,cAAc,MAAM,IAC/C,QAAQ,QAAQ,OAAO;AAC7B,QAAI,OAAO,aAAc,QAAO,gBAAgB;AAChD,UAAM,OAAO,SAAS,MAAM,UAAU,OAAO,OAAO;AACpD,WAAO,UAAU;AAEjB,UAAM,MAAoB;AAAA,MACxB,OAAO,MAAM;AAAA,MACb,cAAc,OAAO;AAAA,MACrB,cAAc,OAAO;AAAA,MACrB,UAAU,OAAO;AAAA,MACjB,QAAQ,OAAO;AAAA,MACf,UAAU,MAAM;AAAA,MAChB,QAAQ,OAAO;AAAA,MACf;AAAA,MACA,GAAI,MAAM,SAAS,SAAY,EAAE,MAAM,MAAM,KAAK,IAAI,CAAC;AAAA,MACvD,GAAI,OAAO,eAAe,SACtB,EAAE,YAAY,OAAO,WAAW,IAChC,CAAC;AAAA;AAAA;AAAA,MAGL,IAAI,MAAM;AACR,cAAM,IACJ,OAAO,cACN,OAAO,cAAc,SAAY,UAAU,OAAO,SAAS,IAAI;AAClE,eAAO,MAAM,SAAY,EAAE,WAAW,EAAE,IAAI,CAAC;AAAA,MAC/C,GAAG;AAAA,IACL;AAEA,YAAQ,MAAM;AAAA,MACZ,KAAK;AACH,eAAO,WAAW;AAClB,iBAAS,KAAK,GAAG;AACjB;AAAA,MACF,KAAK;AACH,eAAO,SAAS;AAChB,eAAO,UAAU;AACjB;AAAA,MACF,KAAK;AACH,eAAO,aAAa;AACpB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,MACF,KAAK;AACH,eAAO,aAAa;AACpB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,MACF,KAAK;AACH,eAAO,UAAU;AACjB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,MACF,KAAK;AACH,eAAO,gBAAgB;AACvB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,MACF,KAAK;AACH,eAAO,WAAW;AAClB,eAAO,UAAU;AACjB,sBAAc,KAAK,GAAG;AACtB;AAAA,IACJ;AAAA,EACF;AAEA,SAAO,YAAY,OAAO,SAAS,IAAI,OAAO,QAAQ,OAAO,SAAS;AAGtE,QAAM,QAAQ,CAAC,GAAiB,MAA4B;AAC1D,UAAM,KAAK,GAAG,EAAE,IAAI,IAAI,EAAE,KAAK,IAAI,EAAE,YAAY,IAAI,EAAE,QAAQ;AAC/D,UAAM,KAAK,GAAG,EAAE,IAAI,IAAI,EAAE,KAAK,IAAI,EAAE,YAAY,IAAI,EAAE,QAAQ;AAC/D,WAAO,KAAK,KAAK,KAAK,KAAK,KAAK,IAAI;AAAA,EACtC;AACA,gBAAc,KAAK,KAAK;AACxB,WAAS,KAAK,KAAK;AACnB,SAAO,EAAE,QAAQ,eAAe,SAAS;AAC3C;;;ADhQA,SAAS,WAAW,MAAgB,UAA0B;AAC5D,SAAO,KAAK,IAAI,UAAU,GAAG,IAAI;AACnC;AAEA,eAAsB,aACpB,SACiC;AACjC,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,EAAE,QAAQ,KAAK,UAAU,IAAI,MAAM;AAAA,IACvC,QAAQ,aAAa,QAAQ,IAAI;AAAA,IACjC;AAAA,MACE,GAAI,QAAQ,WAAW,SAAY,CAAC,IAAI,EAAE,YAAY,QAAQ,OAAO;AAAA,MACrE,GAAI,QAAQ,aAAa,SAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;AAAA,IACzE;AAAA,EACF;AAMA;AAAA,IACE,eAAe,QAAQ;AAAA,MACrB,GAAI,QAAQ,aAAa,SAAY,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;AAAA,MACvE,GAAI,QAAQ,UAAU,SAAY,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;AAAA,MAC9D,GAAI,QAAQ,UAAU,SAAY,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;AAAA,IAChE,CAAC;AAAA,EACH;AAMA,QAAM,aACJ,QAAQ,WAAW,SACf,QAAQ,QAAQ,MAAM,IACtB,QAAQ,WAAW,OAAO,UAAU,MAAM;AAChD,QAAM,SAAS,MAAM,WAAW,UAAU;AAI1C,QAAM,UAAU,MAAM,mBAAmB,OAAO,GAAG,IAAI,CAAC,MAAM,QAAQ,CAAC,CAAC;AACxE,uBAAqB,QAAQ,QAAQ,UAAU;AAE/C,QAAM,WAAyB;AAAA,IAC7B,cAAc,QAAQ,QAAQ,OAAO,MAAM;AAAA,IAC3C,OAAO,OAAO,MAAM;AAAA,EACtB;AACA,QAAM,OAAO,OAAO,UAAU;AAK9B,QAAM,QACJ,QAAQ,UAAU,OACd,WAAW,KAAK,cAAc,SAAS,YAAY,IACnD,SAAS;AAEf,QAAM,QAAQ,MAAM,SAAS;AAAA,IAC3B,GAAG;AAAA,IACH,QAAQ;AAAA,IACR,aAAa;AAAA,IACb,WAAW;AAAA,IACX,OAAO;AAAA,IACP,MAAM;AAAA;AAAA;AAAA,IAGN,QAAQ;AAAA,IACR,QAAQ;AAAA,EACV,CAAC;AAED,QAAM,MAAoB,CAAC;AAC3B,aAAW,WAAW,MAAM,UAAU;AACpC,QAAI,YAAY,QAAS,KAAI,KAAK,GAAG,QAAQ,OAAO,WAAW;AAAA,EACjE;AACA,QAAM,EAAE,QAAQ,WAAW,UAAU,IAAI,WAAW,MAAM,UAAU,MAAM;AAC1E,oBAAkB,WAAW,OAAO,UAAU;AAC9C,oBAAkB,WAAW,UAAU;AAMvC,QAAM,SAAS;AAAA,IACb,EAAE,QAAQ,IAAI;AAAA,IACd,QAAQ,UAAU,OAAO,WAAW;AAAA,EACtC;AAEA,QAAM,QACJ,QAAQ,UAAU,OACd,WAAW,EAAE,QAAQ,IAAI,GAAG,UAAU,IAAI,IAC1C;AAEN,QAAM,QAAQ,WAAW,OAAO,QAAQ;AAAA,IACtC,WAAW,QAAQ,gBAAgB,OAAO,UAAU;AAAA,IACpD,WAAW,QAAQ,gBAAgB,OAAO,UAAU;AAAA,IACpD,QAAQ,QAAQ,aAAa,OAAO,UAAU;AAAA,EAChD,CAAC;AAED,QAAM,WAAW,CAAC,GAAG,MAAM,OAAO,QAAQ;AAC1C,MAAI,OAAO,OAAO,eAAe,GAAG;AAClC,aAAS;AAAA,MACP,GAAG,OAAO,OAAO,YAAY;AAAA,IAC/B;AAAA,EACF;AASA,QAAM,kBAAkB,OAAO,OAAO,WAAW;AACjD,MAAI,iBAAiB;AACnB,aAAS;AAAA,MACP,uCAAuC,OAAO,OAAO,MAAM;AAAA,IAE7D;AAAA,EACF;AAEA,QAAM,SAA4B;AAAA,IAChC;AAAA,IACA;AAAA,IACA,SAAS;AAAA,IACT,QAAQ,OAAO;AAAA,IACf,eAAe,OAAO;AAAA,IACtB,UAAU,OAAO;AAAA,IACjB,GAAI,UAAU,SAAY,EAAE,MAAM,IAAI,CAAC;AAAA,IACvC;AAAA,IACA,OAAO,MAAM;AAAA,IACb;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAMA,UACE,MAAM,KAAK,CAAC,MAAM,EAAE,QAAQ,KAC5B,MAAM,OAAO,QAAQ,gBAAgB,KACrC,MAAM,OAAO,WAAW,UACxB,kBACI,IACA;AAAA,IACN,OAAO,MAAM,OAAO;AAAA,IACpB,YAAY,KAAK,IAAI,IAAI;AAAA,EAC3B;AAEA,QAAM,WAAW,kBAAkB,QAAQ,QAAQ,UAAU,UAAU;AAAA,IACrE,OAAO,QAAQ,UAAU;AAAA,EAC3B,CAAC;AACD,MAAI,QAAQ,OAAQ,OAAM,UAAU,QAAQ,QAAQ,UAAU,OAAO;AACrE,SAAO,EAAE,QAAQ,SAAS;AAC5B;AAOA,SAAS,qBACP,QACA,QACA,YACM;AACN,QAAM,QAAQ,IAAI,IAAI,MAAM;AAC5B,QAAM,UAAU,CAAC,GAAG,IAAI,IAAI,OAAO,OAAO,CAAC,MAAM,CAAC,MAAM,IAAI,EAAE,SAAS,CAAC,EAAE,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC;AAC9F,MAAI,QAAQ,WAAW,EAAG;AAC1B,QAAM,IAAI;AAAA,IACR,GAAG,UAAU,+CAA+C,QAAQ,KAAK,IAAI,CAAC,oBAC1D,OAAO,IAAI,CAAC,MAAM,SAAS,CAAC,CAAC,EAAE,KAAK,IAAI,CAAC;AAAA,EAC/D;AACF;AASA,SAAS,kBACP,WACA,OACA,YACM;AACN,MAAI,UAAU,WAAW,EAAG;AAC5B,QAAM,QAAQ,UAAU,IAAI,CAAC,UAAU;AACrC,UAAM,UAAU,MAAM,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,MAAM,SAAS;AACrE,UAAM,UAAU,SAAS,QAAQ,eAAe,CAAC;AACjD,UAAM,eAAe,QAAQ;AAAA,MAC3B,CAAC,MAAM,EAAE,iBAAiB,MAAM;AAAA,IAClC;AACA,UAAM,OACJ,aAAa,SAAS,IAClB,wBAAwB,aAAa,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,KAAK,IAAI,CAAC,SAAS,MAAM,QAAQ,MACpG,0CAA0C,MAAM,QAAQ,yBACtD,CAAC,GAAG,IAAI,IAAI,QAAQ,IAAI,CAAC,MAAM,EAAE,YAAY,CAAC,CAAC,EAAE,KAAK,EAAE,KAAK,IAAI,KACjE,SACF;AACN,WAAO,KAAK,MAAM,QAAQ,WAAM,MAAM,IAAI,OAAO,SAAS,MAAM,SAAS,CAAC,KAAK,IAAI;AAAA,EACrF,CAAC;AACD,QAAM,IAAI;AAAA,IACR,GAAG,UAAU;AAAA,EAAuC,MAAM,KAAK,IAAI,CAAC;AAAA,EACtE;AACF;AAYA,SAAS,kBACP,WACA,YACM;AACN,MAAI,UAAU,WAAW,EAAG;AAC5B,QAAM,QAAQ,UAAU,IAAI,CAAC,EAAE,OAAO,WAAW,MAAM;AACrD,UAAM,QAAQ,CAAC,GAAG,IAAI,IAAI,WAAW,IAAI,CAAC,MAAM,EAAE,YAAY,CAAC,CAAC,EAAE,KAAK;AACvE,WAAO,KAAK,MAAM,QAAQ,WAAM,MAAM,IAAI,OAAO,SAAS,MAAM,SAAS,CAAC,YAAY,MAAM,KAAK,OAAO,CAAC;AAAA,EAC3G,CAAC;AACD,QAAM,IAAI;AAAA,IACR,GAAG,UAAU;AAAA,EAAkD,MAAM,KAAK,IAAI,CAAC;AAAA,EACjF;AACF;AAEA,SAAS,WACP,QACA,QACmB;AACnB,QAAM,QAA2B,CAAC;AAClC,QAAM,MAAM,CACV,MACA,OACA,WACS;AACT,QAAI,UAAU,OAAW;AACzB,UAAM,KAAK,EAAE,MAAM,OAAO,QAAQ,UAAU,SAAS,MAAM,CAAC;AAAA,EAC9D;AACA,MAAI,aAAa,OAAO,WAAW,OAAO,SAAS;AACnD,MAAI,aAAa,OAAO,WAAW,OAAO,SAAS;AACnD,MAAI,UAAU,OAAO,QAAQ,OAAO,YAAY;AAChD,SAAO;AACT;AASA,SAAS,WACP,OACA,UACA,MACa;AACb,QAAM,OAAO,CACX,MACA,OACA,aACe;AAAA,IACf;AAAA,IACA;AAAA,IACA;AAAA,IACA,QAAQ,MAAM,OAAO,OAAO,EAAE;AAAA,EAChC;AAEA,QAAM,QAAqB;AAAA,IACzB,KAAK,YAAY,SAAS,cAAc,QAAQ;AAAA,EAClD;AACA,aAAW,QAAQ,CAAC,GAAG,KAAK,YAAY,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC,GAAG;AAC/D,UAAM;AAAA,MACJ,KAAK,gBAAgB,MAAM,EAAE,cAAc,MAAM,OAAO,SAAS,MAAM,CAAC;AAAA,IAC1E;AAAA,EACF;AACA,aAAW,YAAY,CAAC,GAAG,KAAK,QAAQ,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC,GAAG;AAC/D,UAAM;AAAA,MACJ,KAAK,kBAAkB,UAAU;AAAA,QAC/B,cAAc,SAAS;AAAA,QACvB,OAAO,EAAE,UAAU,UAAU,SAAS,MAAM,SAAS;AAAA,MACvD,CAAC;AAAA,IACH;AAAA,EACF;AACA,aAAW,YAAY,CAAC,GAAG,KAAK,QAAQ,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC,GAAG;AAC/D,UAAM;AAAA,MACJ,KAAK,kBAAkB,UAAU;AAAA,QAC/B,cAAc,SAAS;AAAA,QACvB,OAAO,EAAE,UAAU,SAAS,MAAM,UAAU,SAAS;AAAA,MACvD,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO;AACT;","names":[]}

package/dist/{chunk-SNEHX4VE.js → chunk-B5CMM57M.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   loadTerms
-} from "./chunk-NN4ASVTY.js";
+} from "./chunk-IIFADWWI.js";
 
 // src/term/commands/list.ts
 var LIST_FORMATS = ["pretty", "json", "csv"];
@@ -14,4 +14,4 @@ export {
   LIST_FORMATS,
   runList
 };
-//# sourceMappingURL=chunk-SNEHX4VE.js.map
+//# sourceMappingURL=chunk-B5CMM57M.js.map