@hawkeyexl/manni 4.2.0 → 4.3.1

package/dist/cli.js

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   buildProgram
-} from "./chunk-BNHQWIAE.js";
+} from "./chunk-UBMMXS2Y.js";
 import {
   ALL_FILL_FIELDS,
   CONFIG_SECTION,
@@ -9,7 +9,6 @@ import {
   DEFAULT_CONFIG_FILENAME2 as DEFAULT_CONFIG_FILENAME,
   DEFAULT_EXTENSIONS,
   DocevalsError,
-  EXECUTION_GRANTS,
   EXPORT_TARGETS,
   GraphError,
   INFER_FORMATS,
@@ -52,6 +51,7 @@ import {
   renderFill,
   renderFill2,
   renderList,
+  renderList2,
   renderQuery,
   renderRemoveJson,
   renderRemovePretty,
@@ -72,12 +72,14 @@ import {
   runExport,
   runFill,
   runFill2,
+  runFill3,
   runGenerate,
   runInit,
   runKeyRotate,
   runKeySet,
   runLint,
   runList,
+  runList2,
   runPromote,
   runQuery,
   runRemove,
@@ -97,7 +99,19 @@ import {
   shortSrc,
   spellSource,
   validateGraph
-} from "./chunk-UYDRDFYF.js";
+} from "./chunk-V4KUVE47.js";
+import {
+  CALIBRATE_FORMATS,
+  REPORT_FORMATS as REPORT_FORMATS3,
+  SUMMARY_FORMATS as SUMMARY_FORMATS2,
+  parseFormat as parseFormat2,
+  runRun as runRun2
+} from "./chunk-KJR7OVDX.js";
+import "./chunk-5HNHEUD5.js";
+import {
+  allowExecutionMessage,
+  isExecutionGrant
+} from "./chunk-CFDUZRM2.js";
 import {
   COMMON_FORMAT_LIST,
   CONCEPT_FIELDS,
@@ -109,11 +123,10 @@ import {
   STDIN_LABEL,
   STDIN_TOKEN,
   TERM_RELATION_FIELDS,
-  ToolError,
   assertNonEmpty,
   collect,
+  colorFor,
   configOption,
-  errorMessage,
   escapeWorkflowCommandMessage,
   escapeWorkflowCommandProperty,
   explicitFalse,
@@ -151,10 +164,17 @@ import {
   valueAt,
   warn,
   writeTextAtomic
-} from "./chunk-TH565O2N.js";
+} from "./chunk-ZQBTK3DO.js";
+import {
+  TracevalsError
+} from "./chunk-3R23C5DT.js";
+import {
+  ToolError,
+  errorMessage
+} from "./chunk-IOYXPBCL.js";
 
 // src/cli.ts
-import { Command as Command8 } from "commander";
+import { Command as Command9 } from "commander";
 
 // src/a11y/cli.ts
 import { Command } from "commander";
@@ -601,12 +621,15 @@ function finishPage(page, floor) {
   const score = applicable === 0 ? null : Math.round(100 * page.passes / applicable);
   return { ...page, violations, score };
 }
+function fails(page) {
+  return page.error !== void 0 || page.violations.length > 0;
+}
 function summarize(results, { discovered, skipped: skipped2, duplicates, redirects, excluded }, sitemap, crawl2) {
   const bySeverity = { notice: 0, warning: 0, error: 0 };
   let failed = 0;
   let violations = 0;
   for (const page of results) {
-    if (page.error !== void 0 || page.violations.length > 0) failed += 1;
+    if (fails(page)) failed += 1;
     for (const violation of page.violations) {
       violations += 1;
       bySeverity[violation.severity] += 1;
@@ -632,7 +655,9 @@ function summarize(results, { discovered, skipped: skipped2, duplicates, redirec
 
 // src/a11y/core/analyzer.ts
 import { AxeBuilder } from "@axe-core/playwright";
-import { chromium } from "playwright-core";
+import {
+  chromium
+} from "playwright-core";
 var NO_BROWSER_MESSAGE = "No browser found. Install one with `npx playwright install chromium`, or install Google Chrome or Microsoft Edge.";
 var CHANNELS = ["chrome", "msedge", "chromium"];
 function launch(channel) {
@@ -649,19 +674,23 @@ async function launchAny() {
 }
 function createPlaywrightAnalyzer() {
   let session = null;
+  let current = null;
   const open = () => {
     session ??= (async () => {
       const found = await launchAny();
       if (found === null) throw new A11yError(NO_BROWSER_MESSAGE);
       const context = await found.browser.newContext();
-      const page = await context.newPage();
-      return { browser: found.browser, page };
+      return { browser: found.browser, context };
     })();
     return session;
   };
   return {
     async analyze(url, opts) {
-      const { page } = await open();
+      const { context } = await open();
+      const page = await context.newPage();
+      const previous = current;
+      current = page;
+      await previous?.close().catch(() => void 0);
       const navigations = [];
       const onRequest = (request) => {
         if (request.isNavigationRequest() && request.frame() === page.mainFrame() && request.redirectedFrom() === null) {
@@ -679,6 +708,7 @@ function createPlaywrightAnalyzer() {
       if (session === null) return;
       const pending = session;
       session = null;
+      current = null;
       try {
         const { browser } = await pending;
         await browser.close();
@@ -1375,7 +1405,7 @@ function renderCheckJunit(run) {
 // src/cite/cli.ts
 var UPDATE_FORMATS = ["pretty", "json"];
 var REMOVE_FORMATS = UPDATE_FORMATS;
-function colorFor(command, isTTY, env) {
+function colorFor2(command, isTTY, env) {
   const noColor = colorOwner(command).opts().color === false;
   return shouldColor({ noColor, isTTY, env });
 }
@@ -1502,7 +1532,7 @@ function buildProgram3() {
       switch (format) {
         case "pretty":
           text2 = renderCheckPretty(run, {
-            color: colorFor(command, process.stdout.isTTY),
+            color: colorFor2(command, process.stdout.isTTY),
             quiet: Boolean(options.quiet),
             showDiff: Boolean(options.showDiff),
             reveal: Boolean(options.reveal)
@@ -1666,7 +1696,7 @@ function buildProgram3() {
         dryRun
       });
       const text2 = format === "json" ? renderUpdateJson(run) : renderUpdatePretty(run, {
-        color: colorFor(command, process.stdout.isTTY),
+        color: colorFor2(command, process.stdout.isTTY),
         // The diffs are the point of a dry run; a real run prints
         // what it rewrote and leaves the file to say the rest.
         showDiff: dryRun
@@ -1741,7 +1771,7 @@ function buildProgram3() {
         dryRun
       });
       const text2 = format === "json" ? renderRemoveJson(run) : renderRemovePretty(run, {
-        color: colorFor(command, process.stdout.isTTY),
+        color: colorFor2(command, process.stdout.isTTY),
         // The diffs are the point of a dry run; a real run prints
         // what it removed and leaves the file to say the rest.
         showDiff: dryRun,
@@ -1766,28 +1796,14 @@ function buildProgram3() {
 
 // src/docevals/cli.ts
 import { Command as Command3 } from "commander";
-function colorFor2(command, isTTY, env) {
-  const noColor = colorOwner2(command).opts().color === false;
-  return shouldColor({ noColor, isTTY, env });
-}
-function colorOwner2(command) {
-  for (let c = command; c !== null; c = c.parent) {
-    if (c.options.some((o) => o.long === "--no-color")) return c;
-  }
-  return command;
-}
 function buildProgram4() {
   const program = new Command3();
   program.name("docevals").description(
     "Deterministic and LLM-as-judge evals for documentation pages, driven by frontmatter."
   ).version(package_default.version).option("--no-color", "disable colored output").showHelpAfterError("(add --help for usage)").exitOverride();
   function collectGrant(value, previous) {
-    if (!EXECUTION_GRANTS.includes(value)) {
-      fail(
-        new DocevalsError(
-          `--allow-execution must be one of ${EXECUTION_GRANTS.join(" | ")}, got "${value}"`
-        )
-      );
+    if (!isExecutionGrant(value)) {
+      fail(new DocevalsError(allowExecutionMessage(value)));
     }
     return [...previous, value];
   }
@@ -1860,7 +1876,7 @@ function buildProgram4() {
         });
         console.log(
           renderList(run, opts.format, {
-            color: colorFor2(command, process.stdout.isTTY)
+            color: colorFor(command, process.stdout.isTTY)
           })
         );
         process.exitCode = run.exitCode;
@@ -1883,7 +1899,7 @@ function buildProgram4() {
     // frontmatter-commands docs/**` would silently swallow the glob and run
     // over the default file set instead. Repeat the flag to grant twice.
     "--allow-execution <kind>",
-    "Grant content-authored execution: frontmatter-commands",
+    "Run only these execution grants (repeatable): frontmatter-commands",
     collectGrant,
     []
   ).option("--no-execution", "Clear every execution grant for this run").option("--no-generate", "Do not generate scripts for command evals missing a command").option("--no-cache", "Bypass the judge response cache").option("--fail-on-review", "Exit 1 when any eval lands in the human-review zone").option("--provider <name>", "Judge provider: auto (default) | anthropic | openai | claude-cli | llama-cpp").option("--model <model>", "Model override; needs a named provider, from here or config").option("--local", LOCAL_FLAG_HELP).option("--runs <n>", "Ensemble runs per eval", parseIntArg("--runs")).option(
@@ -1898,6 +1914,9 @@ function buildProgram4() {
   ).option("--suite <name>", "Run only evals in this suite").option(
     "--since <ref>",
     "Evaluate only pages whose file or eval manifest changed between this git ref and HEAD"
+  ).option(
+    "--newer-than <duration>",
+    "Evaluate only pages whose file or eval manifest changed within this window, such as 30m, 24h, 7d or 2w; in CI, prefer --since"
   ).option(
     "--max-turns <n>",
     "Stop after this many ensemble runs (a cached ensemble costs none)",
@@ -1921,7 +1940,9 @@ function buildProgram4() {
         format: opts.format,
         deterministicOnly: opts.deterministicOnly,
         aiOnly: opts.aiOnly,
-        allowExecution: opts.allowExecution,
+        // Commander's default is `[]`, which is "no flag", not "run nothing":
+        // only a list someone typed narrows the run.
+        allowExecution: opts.allowExecution.length > 0 ? opts.allowExecution : void 0,
         execution: opts.execution,
         generate: opts.generate,
         cache: opts.cache,
@@ -1935,6 +1956,7 @@ function buildProgram4() {
         evalNames: opts.eval,
         suite: opts.suite,
         since: opts.since,
+        newerThan: opts.newerThan,
         // commander collapses `--baseline` to true and `--no-baseline` to
         // false on the same key; a string is an explicit path.
         baseline: opts.baseline,
@@ -1943,7 +1965,7 @@ function buildProgram4() {
       });
       console.log(
         render(report, opts.format, {
-          color: colorFor2(command, process.stdout.isTTY)
+          color: colorFor(command, process.stdout.isTTY)
         })
       );
       process.exitCode = report.exitCode;
@@ -1959,7 +1981,7 @@ function buildProgram4() {
   ).option("--provider <name>", "Provider: auto (default) | anthropic | openai | claude-cli | llama-cpp").option("--model <model>", "Model override; needs a named provider, from here or config").option("--local", LOCAL_FLAG_HELP).action(
     async (paths, opts, command) => {
       try {
-        const pc = palette(colorFor2(command, process.stdout.isTTY));
+        const pc = palette(colorFor(command, process.stdout.isTTY));
         const result = await runGenerate(paths, {
           ...documentOptions(opts),
           provider: opts.provider,
@@ -2038,7 +2060,7 @@ function buildProgram4() {
       });
       console.log(
         renderFill(report, opts.format, {
-          color: colorFor2(command, process.stdout.isTTY)
+          color: colorFor(command, process.stdout.isTTY)
         })
       );
       process.exitCode = report.exitCode;
@@ -2054,7 +2076,7 @@ function buildProgram4() {
   ).option("--write", "Apply promotions (write scripts and rewrite evals)").option("--provider <name>", "Provider: auto (default) | anthropic | openai | claude-cli | llama-cpp").option("--model <model>", "Model override; needs a named provider, from here or config").option("--local", LOCAL_FLAG_HELP).action(
     async (paths, opts, command) => {
       try {
-        const pc = palette(colorFor2(command, process.stdout.isTTY));
+        const pc = palette(colorFor(command, process.stdout.isTTY));
         const proposals = await runPromote(paths, {
           ...documentOptions(opts),
           write: opts.write,
@@ -2093,7 +2115,7 @@ function buildProgram4() {
   ).option("--no-cache", "Bypass the judge response cache").action(
     async (opts, command) => {
       try {
-        const color = colorFor2(command, process.stdout.isTTY);
+        const color = colorFor(command, process.stdout.isTTY);
         const pc = palette(color);
         if (opts.seed) {
           const seeded = seedGoldenCases({ golden: opts.golden });
@@ -2148,7 +2170,7 @@ function buildProgram4() {
         if (!file) {
           console.log(
             renderReviews(listReviews(), {
-              color: colorFor2(command, process.stdout.isTTY)
+              color: colorFor(command, process.stdout.isTTY)
             })
           );
           return;
@@ -6508,7 +6530,7 @@ async function runLint2(opts) {
 
 // src/term/commands/list.ts
 var LIST_FORMATS2 = ["pretty", "json", "csv"];
-async function runList2(opts) {
+async function runList3(opts) {
   const { set } = await loadTerms(opts);
   return { terms: set.terms, set };
 }
@@ -7135,7 +7157,7 @@ function buildProgram8() {
   ).action(async (paths, options) => {
     try {
       const format = formatChoice(options.format, LIST_FORMATS2);
-      const { terms } = await runList2(await coreOptions(paths, options));
+      const { terms } = await runList3(await coreOptions(paths, options));
       print(format === "json" ? renderListJson(terms) : format === "csv" ? renderListCsv(terms) : renderListPretty(terms));
       process.exitCode = 0;
     } catch (err) {
@@ -7236,13 +7258,376 @@ function buildProgram8() {
   return program;
 }
 
-// src/cli.ts
+// src/tracevals/cli.ts
+import { Command as Command8 } from "commander";
+var { version } = package_default;
 function buildProgram9() {
+  const program = new Command8();
+  program.name("tracevals").description(
+    "Deterministic and LLM-as-judge adherence evals for AI agent session traces."
+  ).version(version).option("--no-color", "disable colored output").showHelpAfterError("(add --help for usage)").exitOverride();
+  function collectGrant(value, previous) {
+    if (!isExecutionGrant(value)) fail(new TracevalsError(allowExecutionMessage(value)));
+    return [...previous, value];
+  }
+  function addConfigFlags(cmd) {
+    return cmd.option("-c, --config <path>", "Path to manni.config.yaml").option("--no-config", "Ignore any discovered config file");
+  }
+  function configOptions(opts) {
+    const { configPath, noConfig } = configOption(opts.config);
+    return {
+      ...configPath === void 0 ? {} : { config: configPath },
+      ...noConfig === true ? { noConfig } : {}
+    };
+  }
+  function parseFormatArg(name, allowed) {
+    return (value) => {
+      try {
+        return parseFormat2(value, allowed, name);
+      } catch (e) {
+        fail(e);
+      }
+    };
+  }
+  const range = (min, max) => Number.isFinite(max) ? `between ${min} and ${max}` : `of at least ${min}`;
+  function numeric(name, value, min, max = Number.POSITIVE_INFINITY) {
+    if (value === void 0) return;
+    if (Number.isFinite(value) && value >= min && value <= max) return;
+    throw new TracevalsError(
+      `${name} must be a number ${range(min, max)}, got ${value}`
+    );
+  }
+  function whole(name, value, min, max = Number.POSITIVE_INFINITY) {
+    if (value === void 0) return;
+    if (Number.isInteger(value) && value >= min && value <= max) return;
+    throw new TracevalsError(
+      `${name} must be a whole number ${range(min, max)}, got ${value}`
+    );
+  }
+  function sharedRunOptions(opts) {
+    whole("--runs", opts.runs, 1);
+    whole("--limit", opts.limit, 1);
+    whole("--max-turns", opts.maxTurns, 1);
+    return {
+      ...configOptions(opts),
+      project: opts.project,
+      provider: opts.provider,
+      model: opts.model,
+      local: opts.local,
+      // `undefined` rather than `false` when absent, so a future config
+      // `offline:` still decides on a run where the flag was never typed.
+      offline: opts.offline ? true : void 0,
+      runs: opts.runs,
+      deterministicOnly: opts.deterministicOnly,
+      noCache: opts.cache === false,
+      maxTurns: opts.maxTurns,
+      format: opts.format ?? "pretty",
+      output: opts.output,
+      // Undefined when neither spelling is passed, so the config still decides.
+      // Both spellings are declared for precisely that reason: commander
+      // defaults a lone `--no-x` to `true`, which makes the config's own value
+      // unreachable from the command line.
+      failOnReview: opts.failOnReview,
+      // Commander's default is `[]`, which is "no flag", not "run nothing": only
+      // a list someone typed narrows the run, and `--no-execution` runs none.
+      ...opts.execution === false ? { allowExecution: [] } : (opts.allowExecution ?? []).length > 0 ? { allowExecution: opts.allowExecution } : {},
+      reportUnusedArtifacts: opts.reportUnusedArtifacts,
+      ...opts.require !== void 0 ? { require: opts.require } : {},
+      ...opts.allProjects !== void 0 ? { allProjects: opts.allProjects } : {},
+      ...opts.newerThan !== void 0 ? { newerThan: opts.newerThan } : {},
+      ...opts.limit !== void 0 ? { limit: opts.limit } : {}
+    };
+  }
+  async function executeRun(traces, opts, color) {
+    const shared = {
+      ...sharedRunOptions(opts),
+      color,
+      // Run-only, both of them: `--history` is a point in one session's
+      // timeline, and a manifest belongs to exactly one session (ADR 01024).
+      history: opts.history,
+      ...opts.manifest !== void 0 ? { manifest: opts.manifest } : {}
+    };
+    const selecting = opts.allProjects === true || opts.newerThan !== void 0 || opts.limit !== void 0;
+    if (traces.length > 1 || selecting) {
+      if (opts.manifest !== void 0) {
+        throw new TracevalsError(
+          "--manifest names one session's manifest, so it cannot be used with a corpus; drop it and each trace will find its own"
+        );
+      }
+      const { runBatch } = await import("./batch-A5HZKYIW.js");
+      const { report: report2, rendered: rendered3 } = await runBatch({
+        ...shared,
+        ...traces.length > 0 ? { traces } : {}
+      });
+      console.log(rendered3);
+      process.exitCode = report2.exitCode;
+      return;
+    }
+    let trace = traces[0];
+    if (trace === void 0) {
+      if (!process.stdout.isTTY || !process.stdin.isTTY) {
+        throw new TracevalsError(
+          "no trace given; pass a trace file or run `manni tracevals list`"
+        );
+      }
+      const { pickTrace } = await import("./picker-6HRMETE7.js");
+      trace = await pickTrace();
+    }
+    const { report, rendered: rendered2 } = await runRun2({ ...shared, tracePath: trace });
+    console.log(rendered2);
+    process.exitCode = report.exitCode;
+  }
+  function addRunFlags(cmd, options = {}) {
+    const withHistory = options.history !== false;
+    const formats = options.formats ?? REPORT_FORMATS3;
+    const base = addConfigFlags(cmd).option(
+      "--project <dir>",
+      "artifact-lookup root (overrides the trace's recorded cwd)"
+    ).option(
+      "--provider <name>",
+      "Judge provider: auto (default) | anthropic | openai | claude-cli | llama-cpp"
+    ).option(
+      "--model <model>",
+      "Model override; needs a named provider, from here or config"
+    ).option("--local", LOCAL_FLAG_HELP).option("--runs <n>", "ensemble runs per eval", (v) => Number(v)).option("--deterministic-only", "skip LLM-judged evals").option(
+      "--offline",
+      "never fetch a remote manifest; read external metadata from local manifests"
+    ).option("--no-cache", "bypass the judge cache").option(
+      "--max-turns <n>",
+      "Stop after this many ensemble runs (a cached ensemble costs none)",
+      (v) => Number(v)
+    ).option(
+      "-f, --format <format>",
+      `Output format: ${formats.join(" | ")}`,
+      parseFormatArg("--format", formats),
+      "pretty"
+    ).option("-o, --output <file>", "also write the report to a file").option(
+      "--report-unused-artifacts",
+      "list every skill and agent the session was offered and never used"
+    ).option("--fail-on-review", "Exit 1 when any eval lands in the human-review zone (the default)").option("--no-fail-on-review", "Do not fail the run on evals left for human review").option(
+      "--allow-execution <kind>",
+      "Run only these execution grants (repeatable): frontmatter-commands",
+      collectGrant,
+      []
+    ).option("--no-execution", "Clear every execution grant for this run").option(
+      "--require <module>",
+      "load a grader plugin; repeatable, and added to config plugins",
+      collect,
+      []
+    ).option("--all-projects", "evaluate every project's traces in the session store").option(
+      "--newer-than <duration>",
+      "only traces modified within this window, such as 30m, 24h, 7d or 2w"
+    ).option("--limit <n>", "maximum traces to evaluate", (v) => Number(v));
+    return withHistory ? base.option(
+      "--history",
+      "append to history and compare with the previous run"
+    ) : base;
+  }
+  addRunFlags(
+    // No `isDefault` (proposal 0034): `run` is a verb, so a bare
+    // `manni tracevals` prints the help and exits 2, and a trace path with no
+    // verb is commander's unknown-command refusal rather than a silent run.
+    program.command("run [traces...]").description(
+      "Evaluate one or more traces against the skills and instructions they used"
+    )
+  ).option(
+    "--manifest <file>",
+    "session manifest to compare artifacts against; without it one is looked for beside the trace and under the project"
+  ).action(async (traces, opts, command) => {
+    try {
+      await executeRun(traces, opts, colorFor(command, process.stdout.isTTY));
+    } catch (e) {
+      fail(e);
+    }
+  });
+  addRunFlags(
+    program.command("calibrate [traces...]").description(
+      "Measure judged and graded verdicts against a labels file: false passes, false fails, and review volume"
+    ).option(
+      "--labels <file>",
+      "calibration labels sidecar (default: calibrate.labels)"
+    ).option(
+      "--sweep",
+      "re-score the corpus across the configured grid of zones and ensemble sizes, from cached verdicts"
+    ).option(
+      "--max-false-pass <n>",
+      "exit 1 above this many false passes",
+      (v) => Number(v)
+    ).option(
+      "--max-false-fail <n>",
+      "exit 1 above this many false fails",
+      (v) => Number(v)
+    ).option(
+      "--max-review <n>",
+      "exit 1 above this many needs-review outcomes",
+      (v) => Number(v)
+    ),
+    { history: false, formats: CALIBRATE_FORMATS }
+  ).action(
+    async (traces, opts, command) => {
+      try {
+        const shared = {
+          ...sharedRunOptions(opts),
+          color: colorFor(command, process.stdout.isTTY)
+        };
+        whole("--max-false-pass", opts.maxFalsePass, 0);
+        whole("--max-false-fail", opts.maxFalseFail, 0);
+        whole("--max-review", opts.maxReview, 0);
+        const { runCalibrate: runCalibrate2 } = await import("./calibrate-S5K6QHUY.js");
+        const { report, rendered: rendered2 } = await runCalibrate2({
+          ...shared,
+          ...traces.length > 0 ? { traces } : {},
+          ...opts.labels !== void 0 ? { labels: opts.labels } : {},
+          ...opts.sweep !== void 0 ? { sweep: opts.sweep } : {},
+          ...opts.maxFalsePass !== void 0 ? { maxFalsePass: opts.maxFalsePass } : {},
+          ...opts.maxFalseFail !== void 0 ? { maxFalseFail: opts.maxFalseFail } : {},
+          ...opts.maxReview !== void 0 ? { maxReview: opts.maxReview } : {}
+        });
+        console.log(rendered2);
+        process.exitCode = report.exitCode;
+      } catch (e) {
+        fail(e);
+      }
+    }
+  );
+  addConfigFlags(
+    program.command("fill [paths...]").description(
+      "Propose evals for skills, agent definitions, and project rules, and write those above the confidence threshold"
+    )
+  ).option("--project <dir>", "project root to scan (default: current directory)").option("--exclude <glob>", "Glob to exclude (repeatable)", collect, []).option("--dry-run", "report proposals without writing them").option(
+    "--confidence <n>",
+    "minimum confidence to write (0-1)",
+    (v) => parseFloat(v)
+  ).option(
+    "--max-evals <n>",
+    "maximum evals per artifact, including existing ones",
+    (v) => Number(v)
+  ).option(
+    "--max-turns <n>",
+    "Stop after this many inference calls (a cached artifact costs none)",
+    (v) => Number(v)
+  ).option("--no-cache", "bypass the proposal cache").option(
+    "--provider <name>",
+    "Provider: auto (default) | anthropic | openai | claude-cli | llama-cpp"
+  ).option(
+    "--model <model>",
+    "Model override; needs a named provider, from here or config"
+  ).option("--local", LOCAL_FLAG_HELP).option(
+    "--offline",
+    "never fetch a remote manifest; read external metadata from local manifests"
+  ).option(
+    "--require <module>",
+    "load a grader plugin; repeatable, and added to config plugins",
+    collect,
+    []
+  ).option(
+    "-f, --format <format>",
+    `Output format: ${SUMMARY_FORMATS2.join(" | ")}`,
+    parseFormatArg("--format", SUMMARY_FORMATS2),
+    "pretty"
+  ).action(async (paths, opts, command) => {
+    try {
+      numeric("--confidence", opts.confidence, 0, 1);
+      whole("--max-evals", opts.maxEvals, 1);
+      whole("--max-turns", opts.maxTurns, 1);
+      const { report, rendered: rendered2 } = await runFill3({
+        ...configOptions(opts),
+        ...paths.length > 0 ? { paths } : {},
+        ...opts.project !== void 0 ? { project: opts.project } : {},
+        ...opts.exclude !== void 0 ? { exclude: opts.exclude } : {},
+        ...opts.dryRun !== void 0 ? { dryRun: opts.dryRun } : {},
+        ...opts.confidence !== void 0 ? { confidence: opts.confidence } : {},
+        ...opts.maxEvals !== void 0 ? { maxEvals: opts.maxEvals } : {},
+        ...opts.maxTurns !== void 0 ? { maxTurns: opts.maxTurns } : {},
+        ...opts.cache === false ? { noCache: true } : {},
+        ...opts.provider !== void 0 ? { provider: opts.provider } : {},
+        ...opts.model !== void 0 ? { model: opts.model } : {},
+        ...opts.local !== void 0 ? { local: opts.local } : {},
+        ...opts.offline === true ? { offline: true } : {},
+        ...opts.require !== void 0 ? { require: opts.require } : {},
+        // P1 (proposal 0047): offered only on a terminal, where there is
+        // someone to answer. Off one the block goes to the artifact and the
+        // W1 line says so.
+        confirm: terminalConfirm(),
+        onNotice: notice,
+        onRelocated: (result) => {
+          const manifests = result.manifests.map((m) => m.file).join(", ");
+          notice(`moved ${String(result.summary.moved)} value(s) into ${manifests}.`);
+        },
+        color: colorFor(command, process.stdout.isTTY)
+      });
+      console.log(
+        opts.format === "json" ? JSON.stringify(report, null, 2) : rendered2
+      );
+      process.exitCode = report.exitCode;
+    } catch (e) {
+      fail(e);
+    }
+  });
+  addConfigFlags(
+    program.command("capture").description(
+      "Record a session manifest: sha256 of every instruction artifact plus the git SHA, read from a Claude Code hook payload on stdin"
+    )
+  ).option("--project <dir>", "project root to scan (default: the payload's cwd)").option("--session-id <id>", "session id (default: the payload's session_id)").option("-o, --out <file>", "write here instead of the configured directory").option(
+    "-f, --format <format>",
+    `Output format: ${SUMMARY_FORMATS2.join(" | ")}`,
+    parseFormatArg("--format", SUMMARY_FORMATS2),
+    "pretty"
+  ).action(async (opts) => {
+    try {
+      const { runCapture } = await import("./capture-OZEZDABQ.js");
+      const result = await runCapture({
+        version,
+        ...configOptions(opts),
+        ...opts.project !== void 0 ? { project: opts.project } : {},
+        ...opts.sessionId !== void 0 ? { sessionId: opts.sessionId } : {},
+        ...opts.out !== void 0 ? { out: opts.out } : {},
+        format: opts.format ?? "pretty"
+      });
+      if (result.stdout !== "") console.log(result.stdout);
+      if (result.stderr !== "") notice(result.stderr);
+      process.exitCode = result.exitCode;
+    } catch (e) {
+      fail(e);
+    }
+  });
+  program.command("list").description("List discoverable traces (Claude Code session files)").option(
+    "--project <dir>",
+    "project directory to scope to (default: current directory)"
+  ).option("--all-projects", "scan every project in the session store").option(
+    "--newer-than <duration>",
+    "only traces modified within this window, such as 30m, 24h, 7d or 2w"
+  ).option("--limit <n>", "maximum traces to list", (v) => Number(v)).option(
+    "-f, --format <format>",
+    `Output format: ${SUMMARY_FORMATS2.join(" | ")}`,
+    parseFormatArg("--format", SUMMARY_FORMATS2),
+    "pretty"
+  ).action(async (opts, command) => {
+    try {
+      whole("--limit", opts.limit, 1);
+      const run = await runList2({
+        project: opts.project,
+        allProjects: opts.allProjects,
+        ...opts.newerThan !== void 0 ? { newerThan: opts.newerThan } : {},
+        limit: opts.limit
+      });
+      console.log(
+        opts.format === "json" ? JSON.stringify(run, null, 2) : renderList2(run, { color: colorFor(command, process.stdout.isTTY) })
+      );
+    } catch (e) {
+      fail(e);
+    }
+  });
+  return program;
+}
+
+// src/cli.ts
+function buildProgram10() {
   const meta = buildProgram().name("meta").description(
     "Validate, read, query and fill document metadata against JSON Schema"
   );
   const metaCommands = new Set(meta.commands.map((c) => c.name()));
-  const program = new Command8();
+  const program = new Command9();
   program.name("manni").description("Tools for documentation that is meant to be checked.").version(package_default.version, "-V, --version").showHelpAfterError("(add --help for usage)").exitOverride().configureOutput({
     outputError: (str, write) => {
       write(str);
@@ -7285,13 +7670,18 @@ function buildProgram9() {
       "Deterministic knowledge graphs derived from documentation frontmatter and formatting"
     )
   );
+  program.addCommand(
+    buildProgram9().name("tracevals").description(
+      "Deterministic and LLM-as-judge adherence evals for AI agent session traces"
+    )
+  );
   program.addCommand(
     buildProgram5().name("key").description("Set and rotate the family key that encrypted values are encrypted with.")
   );
   return program;
 }
-runIfMain(import.meta.url, buildProgram9);
+runIfMain(import.meta.url, buildProgram10);
 export {
-  buildProgram9 as buildProgram
+  buildProgram10 as buildProgram
 };
 //# sourceMappingURL=cli.js.map