@aigne/afs-llm-bench 1.12.0-beta.5

package/dist/prompts.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"prompts.mjs","names":["parseYaml","dumpYaml"],"sources":["../src/prompts.ts"],"sourcesContent":["/**\n * Prompt loader for the LLM-AFS benchmark suite.\n *\n * Each prompt is a Markdown file with YAML frontmatter (`gray-matter`-parsed)\n * and a body sliced by H1 sections (`# Task`, `# System prompt`, etc.).\n *\n * Phase 0 only needs the load + parse path so that `list /bench/prompts` and\n * `read /bench/prompts/<id>` return real entries. Runner / patcher consume the\n * parsed shape in Phase 1.\n */\n\nimport { getPlatform } from \"@aigne/afs\";\nimport matter from \"gray-matter\";\nimport { dump as dumpYaml, load as parseYaml } from \"js-yaml\";\nimport { z } from \"zod\";\n\n// gray-matter@4 bundles `js-yaml.safeLoad`, which js-yaml@4 removed — calling\n// matter() with the default engine throws \"yaml.safeLoad is removed\". Supply an\n// explicit engine backed by js-yaml@4's load/dump so frontmatter parses.\nconst GRAY_MATTER_OPTIONS = {\n  engines: {\n    yaml: {\n      parse: (str: string) => parseYaml(str) as object,\n      stringify: (data: object) => dumpYaml(data),\n    },\n  },\n} as const;\n\nconst promptFrontmatterSchema = z.object({\n  id: z.string().min(1, \"frontmatter.id is required\"),\n  title: z.string().min(1, \"frontmatter.title is required\"),\n  domain: z.enum([\"chinese\", \"english\", \"coding\"]).optional(),\n  max_rounds: z.number().int().positive(\"max_rounds must be a positive integer\"),\n  weight_dim: z.string().optional(),\n});\n\nexport type PromptFrontmatter = z.infer<typeof promptFrontmatterSchema>;\n\nexport interface Prompt {\n  /** Stable id from frontmatter, used as the path segment under `/prompts/`. */\n  id: string;\n  /** Source filename relative to `promptsDir`. */\n  filename: string;\n  /** Validated frontmatter. */\n  frontmatter: PromptFrontmatter;\n  /** Raw markdown body (excludes frontmatter). */\n  body: string;\n  /** Body sliced by `# H1` headings — section name → section body. */\n  sections: Record<string, string>;\n}\n\n/**\n * Slice markdown body into sections keyed by H1 heading text.\n *\n * Section headings are case-sensitive and trimmed. Content before the first H1\n * is dropped — by convention every benchmark prompt starts with `# Task`.\n */\nexport function sliceBodyByH1(body: string): Record<string, string> {\n  const out: Record<string, string> = {};\n  const lines = body.split(/\\r?\\n/);\n  let currentName: string | undefined;\n  let currentBuf: string[] = [];\n\n  const flush = () => {\n    if (currentName !== undefined) {\n      out[currentName] = currentBuf.join(\"\\n\").trim();\n    }\n  };\n\n  for (const line of lines) {\n    const match = /^#\\s+(.+?)\\s*$/.exec(line);\n    if (match && !line.startsWith(\"##\")) {\n      flush();\n      currentName = match[1]!.trim();\n      currentBuf = [];\n    } else if (currentName !== undefined) {\n      currentBuf.push(line);\n    }\n  }\n  flush();\n\n  return out;\n}\n\n/**\n * Parse a single prompt file. Throws `BadPromptError` (regular Error subclass)\n * if frontmatter is missing required fields.\n */\nexport class BadPromptError extends Error {\n  override readonly name = \"BadPromptError\";\n  constructor(\n    public readonly filename: string,\n    message: string,\n  ) {\n    super(`[${filename}] ${message}`);\n  }\n}\n\nexport function parsePrompt(filename: string, raw: string): Prompt {\n  const parsed = matter(raw, GRAY_MATTER_OPTIONS);\n  const fm = promptFrontmatterSchema.safeParse(parsed.data);\n  if (!fm.success) {\n    const issues = fm.error.issues.map((i) => `${i.path.join(\".\")}: ${i.message}`).join(\"; \");\n    throw new BadPromptError(filename, `invalid frontmatter — ${issues}`);\n  }\n\n  const body = parsed.content.replace(/^\\s+/, \"\");\n  return {\n    id: fm.data.id,\n    filename,\n    frontmatter: fm.data,\n    body,\n    sections: sliceBodyByH1(body),\n  };\n}\n\n/**\n * Load every `*.md` file under `dir` and parse it into a Prompt. Missing\n * directory → empty array (Phase 0 must not crash on unconfigured promptsDir).\n *\n * Files that fail frontmatter validation are skipped with their error returned\n * in `errors` so callers can surface them (Phase 1 runner, conformance probe).\n */\nexport async function loadPromptsFromDir(\n  dir: string,\n): Promise<{ prompts: Prompt[]; errors: BadPromptError[] }> {\n  const platform = getPlatform();\n  const fs = platform.fs;\n  const path = platform.path;\n  if (!fs) return { prompts: [], errors: [] };\n\n  if (!(await fs.exists(dir))) return { prompts: [], errors: [] };\n  const entries = await fs.readdir(dir);\n\n  const prompts: Prompt[] = [];\n  const errors: BadPromptError[] = [];\n  for (const filename of entries.slice().sort()) {\n    if (path.extname(filename).toLowerCase() !== \".md\") continue;\n    const raw = await fs.readTextFile(path.join(dir, filename));\n    try {\n      prompts.push(parsePrompt(filename, raw));\n    } catch (err) {\n      if (err instanceof BadPromptError) errors.push(err);\n      else throw err;\n    }\n  }\n\n  return { prompts, errors };\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAmBA,MAAM,sBAAsB,EAC1B,SAAS,EACP,MAAM;CACJ,QAAQ,QAAgBA,KAAU,IAAI;CACtC,YAAY,SAAiBC,KAAS,KAAK;CAC5C,EACF,EACF;AAED,MAAM,0BAA0B,EAAE,OAAO;CACvC,IAAI,EAAE,QAAQ,CAAC,IAAI,GAAG,6BAA6B;CACnD,OAAO,EAAE,QAAQ,CAAC,IAAI,GAAG,gCAAgC;CACzD,QAAQ,EAAE,KAAK;EAAC;EAAW;EAAW;EAAS,CAAC,CAAC,UAAU;CAC3D,YAAY,EAAE,QAAQ,CAAC,KAAK,CAAC,SAAS,wCAAwC;CAC9E,YAAY,EAAE,QAAQ,CAAC,UAAU;CAClC,CAAC;;;;;;;AAuBF,SAAgB,cAAc,MAAsC;CAClE,MAAM,MAA8B,EAAE;CACtC,MAAM,QAAQ,KAAK,MAAM,QAAQ;CACjC,IAAI;CACJ,IAAI,aAAuB,EAAE;CAE7B,MAAM,cAAc;AAClB,MAAI,gBAAgB,OAClB,KAAI,eAAe,WAAW,KAAK,KAAK,CAAC,MAAM;;AAInD,MAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,QAAQ,iBAAiB,KAAK,KAAK;AACzC,MAAI,SAAS,CAAC,KAAK,WAAW,KAAK,EAAE;AACnC,UAAO;AACP,iBAAc,MAAM,GAAI,MAAM;AAC9B,gBAAa,EAAE;aACN,gBAAgB,OACzB,YAAW,KAAK,KAAK;;AAGzB,QAAO;AAEP,QAAO;;;;;;AAOT,IAAa,iBAAb,cAAoC,MAAM;CACxC,AAAkB,OAAO;CACzB,YACE,AAAgB,UAChB,SACA;AACA,QAAM,IAAI,SAAS,IAAI,UAAU;EAHjB;;;AAOpB,SAAgB,YAAY,UAAkB,KAAqB;CACjE,MAAM,SAAS,OAAO,KAAK,oBAAoB;CAC/C,MAAM,KAAK,wBAAwB,UAAU,OAAO,KAAK;AACzD,KAAI,CAAC,GAAG,QAEN,OAAM,IAAI,eAAe,UAAU,yBADpB,GAAG,MAAM,OAAO,KAAK,MAAM,GAAG,EAAE,KAAK,KAAK,IAAI,CAAC,IAAI,EAAE,UAAU,CAAC,KAAK,KAAK,GACpB;CAGvE,MAAM,OAAO,OAAO,QAAQ,QAAQ,QAAQ,GAAG;AAC/C,QAAO;EACL,IAAI,GAAG,KAAK;EACZ;EACA,aAAa,GAAG;EAChB;EACA,UAAU,cAAc,KAAK;EAC9B;;;;;;;;;AAUH,eAAsB,mBACpB,KAC0D;CAC1D,MAAM,WAAW,aAAa;CAC9B,MAAM,KAAK,SAAS;CACpB,MAAM,OAAO,SAAS;AACtB,KAAI,CAAC,GAAI,QAAO;EAAE,SAAS,EAAE;EAAE,QAAQ,EAAE;EAAE;AAE3C,KAAI,CAAE,MAAM,GAAG,OAAO,IAAI,CAAG,QAAO;EAAE,SAAS,EAAE;EAAE,QAAQ,EAAE;EAAE;CAC/D,MAAM,UAAU,MAAM,GAAG,QAAQ,IAAI;CAErC,MAAM,UAAoB,EAAE;CAC5B,MAAM,SAA2B,EAAE;AACnC,MAAK,MAAM,YAAY,QAAQ,OAAO,CAAC,MAAM,EAAE;AAC7C,MAAI,KAAK,QAAQ,SAAS,CAAC,aAAa,KAAK,MAAO;EACpD,MAAM,MAAM,MAAM,GAAG,aAAa,KAAK,KAAK,KAAK,SAAS,CAAC;AAC3D,MAAI;AACF,WAAQ,KAAK,YAAY,UAAU,IAAI,CAAC;WACjC,KAAK;AACZ,OAAI,eAAe,eAAgB,QAAO,KAAK,IAAI;OAC9C,OAAM;;;AAIf,QAAO;EAAE;EAAS;EAAQ"}

package/dist/reporter.cjs

@@ -0,0 +1,322 @@
+const require_rolldown_runtime = require('./_virtual/rolldown_runtime.cjs');
+const require_errors = require('./errors.cjs');
+require('./runner.cjs');
+const require_patcher = require('./patcher.cjs');
+let node_fs = require("node:fs");
+
+//#region src/reporter.ts
+/**
+* Phase 2 (Run 4) — HTML / Markdown report generator (plan.md L1063-1068).
+*
+* Reads a results JSON and produces either a markdown or HTML report:
+*   - model × dimension heatmap (emoji thresholds for markdown, color-coded
+*     CSS classes for HTML)
+*   - per-policy ranking (computed from the same weight tables the patcher
+*     uses, resolved through `/dev/ai/score-weights/<name>` by the bench wiring)
+*   - delta vs a `previous` results JSON when supplied
+*   - per-domain table when domains are present
+*
+* Security invariants (plan.md L1098-1101):
+*   1. All user-controlled strings (model id, judge id, version, etc.) are
+*      HTML-escaped before insertion in HTML mode. Markdown mode also escapes
+*      `<>` so a model name like `<script>alert(1)<\/script>` cannot be smuggled
+*      into a downstream HTML viewer that renders the markdown.
+*   2. Vault credentials never end up in the report — `validateResultsShape`
+*      rejects sensitive field names before we ever reach the renderer.
+*   3. Default `includeRaw: false` — raw trace[] is never embedded unless the
+*      caller explicitly asks; even then we serialise via JSON.stringify with
+*      no key reordering so audit diffs stay clean.
+*   4. `assertWithinResultsDir` keeps the path inside the results root, same
+*      as patcher.ts.
+*
+* Read-only: never writes files, never mutates input.
+*/
+const DIMENSIONS = [
+	"exploration",
+	"tool_reliability",
+	"self_stop",
+	"parallelism",
+	"efficiency",
+	"instruction_following"
+];
+const THRESHOLDS = {
+	good: .75,
+	ok: .5
+};
+async function generateReport(input) {
+	const format = input.format ?? "markdown";
+	if (format !== "markdown" && format !== "html") throw new require_errors.ValidationError(`unsupported format "${format}"; use "markdown" or "html"`);
+	const current = await loadResults(input.resultsPath, input.resultsRoot);
+	const previous = input.previousPath ? await loadResults(input.previousPath, input.resultsRoot) : void 0;
+	const weightsTable = await input.loadPolicyWeights();
+	if (!weightsTable || typeof weightsTable !== "object" || Array.isArray(weightsTable)) throw new require_errors.ValidationError("loadPolicyWeights must return a plain object of { policy: weights }");
+	const generatedAt = (/* @__PURE__ */ new Date()).toISOString();
+	const data = assembleReport(current, previous, weightsTable);
+	return {
+		content: format === "markdown" ? renderMarkdown(data, generatedAt, input.includeRaw === true ? current.raw : void 0) : renderHtml(data, generatedAt, input.includeRaw === true ? current.raw : void 0),
+		format,
+		generatedAt
+	};
+}
+async function loadResults(p, root) {
+	const safe = require_patcher.assertWithinResultsDir(p, root);
+	let text;
+	try {
+		text = await node_fs.promises.readFile(safe, "utf8");
+	} catch (err) {
+		throw new require_errors.ValidationError(`failed to read results: ${err instanceof Error ? err.message : String(err)}`);
+	}
+	let parsed;
+	try {
+		parsed = JSON.parse(text);
+	} catch (err) {
+		throw new require_errors.ValidationError(`results is not valid JSON: ${err instanceof Error ? err.message : String(err)}`);
+	}
+	return {
+		...require_patcher.validateResultsShape(parsed),
+		raw: parsed
+	};
+}
+function assembleReport(current, previous, weightsTable) {
+	const judges = Array.isArray(current.judge_model) ? current.judge_model : current.judge_model ? [current.judge_model] : [];
+	const modelIds = Object.keys(current.results).sort();
+	const models = modelIds.map((m) => {
+		const r = current.results[m];
+		const scores = {};
+		for (const d of DIMENSIONS) scores[d] = r.scores[d];
+		return {
+			model: m,
+			scores,
+			overall: meanOf(Object.values(r.scores).filter(Number.isFinite))
+		};
+	});
+	const dims = models.some((m) => typeof m.scores.instruction_following === "number") ? DIMENSIONS : DIMENSIONS.filter((d) => d !== "instruction_following");
+	const policies = Object.entries(weightsTable).map(([policy, weights]) => {
+		return {
+			policy,
+			rows: modelIds.map((m) => ({
+				model: m,
+				score: computeWeightedScore(current.results[m], weights)
+			})).sort((a, b) => b.score - a.score)
+		};
+	}).sort((a, b) => a.policy.localeCompare(b.policy));
+	const domains = modelIds.map((m) => ({
+		model: m,
+		domains: current.results[m]?.domains ?? {}
+	})).filter((row) => Object.keys(row.domains).length > 0);
+	const delta = previous ? computeDelta(current, previous) : [];
+	return {
+		version: current.version,
+		judges,
+		dims,
+		models,
+		policies,
+		domains,
+		delta
+	};
+}
+function meanOf(xs) {
+	if (xs.length === 0) return void 0;
+	let s = 0;
+	for (const x of xs) s += x;
+	return s / xs.length;
+}
+function computeDelta(current, previous) {
+	const rows = [];
+	const models = new Set([...Object.keys(current.results), ...Object.keys(previous.results)]);
+	for (const m of [...models].sort()) {
+		const cur = current.results[m]?.scores ?? {};
+		const prev = previous.results[m]?.scores ?? {};
+		for (const dim of DIMENSIONS) {
+			const c = cur[dim];
+			const p = prev[dim];
+			if (typeof c !== "number" && typeof p !== "number") continue;
+			if (c === p) continue;
+			const d = typeof c === "number" && typeof p === "number" ? c - p : void 0;
+			rows.push({
+				model: m,
+				dimension: dim,
+				prev: p,
+				curr: c,
+				delta: d
+			});
+		}
+	}
+	return rows;
+}
+function computeWeightedScore(entry, weights) {
+	let total = 0;
+	for (const [key, w] of Object.entries(weights)) {
+		const v = getNested(entry, key);
+		if (typeof v === "number" && Number.isFinite(v)) total += v * w;
+	}
+	return total;
+}
+function getNested(obj, p) {
+	let cur = obj;
+	for (const k of p.split(".")) {
+		if (cur === null || typeof cur !== "object") return void 0;
+		cur = cur[k];
+	}
+	return cur;
+}
+function renderMarkdown(data, generatedAt, raw) {
+	const lines = [];
+	lines.push("# LLM AFS Benchmark Report");
+	lines.push("");
+	lines.push(`**Generated:** ${escMd(generatedAt)}`);
+	lines.push(`**Run version:** ${escMd(data.version)}`);
+	if (data.judges.length > 0) lines.push(`**Judge(s):** ${data.judges.map(escMd).join(", ")}`);
+	lines.push(`**Models:** ${data.models.length}`);
+	lines.push("");
+	lines.push("## Model × Dimension Heatmap");
+	lines.push("");
+	lines.push(`| model | ${data.dims.join(" | ")} | overall |`);
+	lines.push(`|---|${data.dims.map(() => "---").join("|")}|---|`);
+	for (const row of data.models) {
+		const cells = data.dims.map((d) => formatCellMd(row.scores[d]));
+		const overall = typeof row.overall === "number" ? row.overall.toFixed(2) : "-";
+		lines.push(`| ${escMd(row.model)} | ${cells.join(" | ")} | ${overall} |`);
+	}
+	lines.push("");
+	if (data.policies.length > 0) {
+		lines.push("## Per-Policy Ranking");
+		lines.push("");
+		for (const p of data.policies) {
+			lines.push(`### ${escMd(p.policy)}`);
+			lines.push("");
+			lines.push("| rank | model | score |");
+			lines.push("|---|---|---|");
+			p.rows.forEach((r, i) => {
+				lines.push(`| ${i + 1} | ${escMd(r.model)} | ${r.score.toFixed(3)} |`);
+			});
+			lines.push("");
+		}
+	}
+	if (data.domains.length > 0) {
+		const allDomains = collectDomainKeys(data.domains);
+		lines.push("## Domains");
+		lines.push("");
+		lines.push(`| model | ${allDomains.join(" | ")} |`);
+		lines.push(`|---|${allDomains.map(() => "---").join("|")}|`);
+		for (const row of data.domains) {
+			const cells = allDomains.map((d) => formatCellMd(row.domains[d]));
+			lines.push(`| ${escMd(row.model)} | ${cells.join(" | ")} |`);
+		}
+		lines.push("");
+	}
+	if (data.delta.length > 0) {
+		lines.push("## Delta vs previous run");
+		lines.push("");
+		lines.push("| model | dimension | prev | curr | delta |");
+		lines.push("|---|---|---|---|---|");
+		for (const d of data.delta) lines.push(`| ${escMd(d.model)} | ${escMd(d.dimension)} | ${formatNum(d.prev)} | ${formatNum(d.curr)} | ${formatDelta(d.delta)} |`);
+		lines.push("");
+	}
+	if (raw !== void 0) {
+		lines.push("## Raw trace");
+		lines.push("");
+		lines.push("```json");
+		lines.push(JSON.stringify(raw, null, 2));
+		lines.push("```");
+		lines.push("");
+	}
+	return lines.join("\n");
+}
+function escMd(s) {
+	return String(s).replace(/[<>|]/g, (c) => c === "<" ? "&lt;" : c === ">" ? "&gt;" : "\\|");
+}
+function formatCellMd(v) {
+	if (typeof v !== "number" || !Number.isFinite(v)) return "-";
+	return `${v >= THRESHOLDS.good ? "✅" : v >= THRESHOLDS.ok ? "⚠️" : "❌"} ${v.toFixed(2)}`;
+}
+function formatNum(v) {
+	return typeof v === "number" && Number.isFinite(v) ? v.toFixed(2) : "-";
+}
+function formatDelta(v) {
+	if (typeof v !== "number" || !Number.isFinite(v)) return "-";
+	return `${v > 0 ? "+" : ""}${v.toFixed(2)}`;
+}
+function renderHtml(data, generatedAt, raw) {
+	const out = [];
+	out.push("<!doctype html>");
+	out.push("<html lang=\"en\">");
+	out.push("<head>");
+	out.push("<meta charset=\"utf-8\">");
+	out.push("<title>LLM AFS Benchmark Report</title>");
+	out.push("<style>");
+	out.push("body{font-family:system-ui,sans-serif;margin:2rem;color:#222}", "table{border-collapse:collapse;margin-bottom:1.5rem}", "th,td{border:1px solid #ccc;padding:.4rem .8rem;text-align:left}", "th{background:#f4f4f4}", ".good{background:#c6f6d5}.ok{background:#fef3c7}.bad{background:#fee2e2}", ".pos{color:#15803d}.neg{color:#b91c1c}", "h2{margin-top:2rem}");
+	out.push("</style>");
+	out.push("</head>");
+	out.push("<body>");
+	out.push("<h1>LLM AFS Benchmark Report</h1>");
+	out.push(`<p><strong>Generated:</strong> ${esc(generatedAt)}</p>`);
+	out.push(`<p><strong>Run version:</strong> ${esc(data.version)}</p>`);
+	if (data.judges.length > 0) out.push(`<p><strong>Judge(s):</strong> ${data.judges.map(esc).join(", ")}</p>`);
+	out.push(`<p><strong>Models:</strong> ${data.models.length}</p>`);
+	out.push("<h2>Model × Dimension Heatmap</h2>");
+	out.push("<table>");
+	out.push(`<thead><tr><th>model</th>${data.dims.map((d) => `<th>${esc(d)}</th>`).join("")}<th>overall</th></tr></thead>`);
+	out.push("<tbody>");
+	for (const row of data.models) {
+		const cells = data.dims.map((d) => htmlCell(row.scores[d])).join("");
+		const overall = typeof row.overall === "number" ? row.overall.toFixed(2) : "-";
+		out.push(`<tr><td>${esc(row.model)}</td>${cells}<td>${overall}</td></tr>`);
+	}
+	out.push("</tbody></table>");
+	if (data.policies.length > 0) {
+		out.push("<h2>Per-Policy Ranking</h2>");
+		for (const p of data.policies) {
+			out.push(`<h3>${esc(p.policy)}</h3>`);
+			out.push("<table>");
+			out.push("<thead><tr><th>rank</th><th>model</th><th>score</th></tr></thead><tbody>");
+			p.rows.forEach((r, i) => {
+				out.push(`<tr><td>${i + 1}</td><td>${esc(r.model)}</td><td>${r.score.toFixed(3)}</td></tr>`);
+			});
+			out.push("</tbody></table>");
+		}
+	}
+	if (data.domains.length > 0) {
+		const allDomains = collectDomainKeys(data.domains);
+		out.push("<h2>Domains</h2>");
+		out.push("<table>");
+		out.push(`<thead><tr><th>model</th>${allDomains.map((d) => `<th>${esc(d)}</th>`).join("")}</tr></thead>`);
+		out.push("<tbody>");
+		for (const row of data.domains) {
+			const cells = allDomains.map((d) => htmlCell(row.domains[d])).join("");
+			out.push(`<tr><td>${esc(row.model)}</td>${cells}</tr>`);
+		}
+		out.push("</tbody></table>");
+	}
+	if (data.delta.length > 0) {
+		out.push("<h2>Delta vs previous run</h2>");
+		out.push("<table>");
+		out.push("<thead><tr><th>model</th><th>dimension</th><th>prev</th><th>curr</th><th>delta</th></tr></thead><tbody>");
+		for (const d of data.delta) {
+			const sign = d.delta === void 0 ? "" : d.delta > 0 ? "pos" : d.delta < 0 ? "neg" : "";
+			out.push(`<tr><td>${esc(d.model)}</td><td>${esc(d.dimension)}</td><td>${formatNum(d.prev)}</td><td>${formatNum(d.curr)}</td><td class="${sign}">${formatDelta(d.delta)}</td></tr>`);
+		}
+		out.push("</tbody></table>");
+	}
+	if (raw !== void 0) {
+		out.push("<h2>Raw trace</h2>");
+		out.push(`<pre>${esc(JSON.stringify(raw, null, 2))}</pre>`);
+	}
+	out.push("</body></html>");
+	return out.join("\n");
+}
+function htmlCell(v) {
+	if (typeof v !== "number" || !Number.isFinite(v)) return "<td>-</td>";
+	return `<td class="${v >= THRESHOLDS.good ? "good" : v >= THRESHOLDS.ok ? "ok" : "bad"}">${v.toFixed(2)}</td>`;
+}
+function esc(s) {
+	return String(s).replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;").replace(/"/g, "&quot;").replace(/'/g, "&#39;");
+}
+function collectDomainKeys(rows) {
+	const set = /* @__PURE__ */ new Set();
+	for (const r of rows) for (const k of Object.keys(r.domains)) set.add(k);
+	return [...set].sort();
+}
+
+//#endregion
+exports.generateReport = generateReport;

package/dist/reporter.d.cts

@@ -0,0 +1,27 @@
+import { PolicyWeightsTable } from "./patcher.cjs";
+
+//#region src/reporter.d.ts
+type ReportFormat = "markdown" | "html";
+interface GenerateReportInput {
+  resultsPath: string;
+  resultsRoot: string;
+  format?: ReportFormat;
+  previousPath?: string;
+  includeRaw?: boolean;
+  /**
+   * Resolve every policy → its weight map. Bench wires this through AFS
+   * (`/dev/ai/score-weights/<name>`) so the reporter stays decoupled from
+   * ai-device internals — same pattern as patcher.ts. Returning an empty
+   * object means the report skips the per-policy ranking section.
+   */
+  loadPolicyWeights: () => Promise<PolicyWeightsTable>;
+}
+interface GeneratedReport {
+  content: string;
+  format: ReportFormat;
+  generatedAt: string;
+}
+declare function generateReport(input: GenerateReportInput): Promise<GeneratedReport>;
+//#endregion
+export { GenerateReportInput, GeneratedReport, ReportFormat, generateReport };
+//# sourceMappingURL=reporter.d.cts.map

package/dist/reporter.d.cts.map

@@ -0,0 +1 @@
+{"version":3,"file":"reporter.d.cts","names":[],"sources":["../src/reporter.ts"],"mappings":";;;KAsCY,YAAA;AAAA,UAEK,mBAAA;EACf,WAAA;EACA,WAAA;EACA,MAAA,GAAS,YAAA;EACT,YAAA;EACA,UAAA;EAaA;;;AAqBF;;;EA3BE,iBAAA,QAAyB,OAAA,CAAQ,kBAAA;AAAA;AAAA,UAGlB,eAAA;EACf,OAAA;EACA,MAAA,EAAQ,YAAA;EACR,WAAA;AAAA;AAAA,iBAqBoB,cAAA,CAAe,KAAA,EAAO,mBAAA,GAAsB,OAAA,CAAQ,eAAA"}

package/dist/reporter.d.mts

@@ -0,0 +1,27 @@
+import { PolicyWeightsTable } from "./patcher.mjs";
+
+//#region src/reporter.d.ts
+type ReportFormat = "markdown" | "html";
+interface GenerateReportInput {
+  resultsPath: string;
+  resultsRoot: string;
+  format?: ReportFormat;
+  previousPath?: string;
+  includeRaw?: boolean;
+  /**
+   * Resolve every policy → its weight map. Bench wires this through AFS
+   * (`/dev/ai/score-weights/<name>`) so the reporter stays decoupled from
+   * ai-device internals — same pattern as patcher.ts. Returning an empty
+   * object means the report skips the per-policy ranking section.
+   */
+  loadPolicyWeights: () => Promise<PolicyWeightsTable>;
+}
+interface GeneratedReport {
+  content: string;
+  format: ReportFormat;
+  generatedAt: string;
+}
+declare function generateReport(input: GenerateReportInput): Promise<GeneratedReport>;
+//#endregion
+export { GenerateReportInput, GeneratedReport, ReportFormat, generateReport };
+//# sourceMappingURL=reporter.d.mts.map

package/dist/reporter.d.mts.map

@@ -0,0 +1 @@
+{"version":3,"file":"reporter.d.mts","names":[],"sources":["../src/reporter.ts"],"mappings":";;;KAsCY,YAAA;AAAA,UAEK,mBAAA;EACf,WAAA;EACA,WAAA;EACA,MAAA,GAAS,YAAA;EACT,YAAA;EACA,UAAA;EAaA;;;AAqBF;;;EA3BE,iBAAA,QAAyB,OAAA,CAAQ,kBAAA;AAAA;AAAA,UAGlB,eAAA;EACf,OAAA;EACA,MAAA,EAAQ,YAAA;EACR,WAAA;AAAA;AAAA,iBAqBoB,cAAA,CAAe,KAAA,EAAO,mBAAA,GAAsB,OAAA,CAAQ,eAAA"}