@riddledc/riddle-proof 0.8.24 → 0.8.26

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (36) hide show
  1. package/README.md +17 -0
  2. package/dist/advanced/engine-harness.cjs +54 -1
  3. package/dist/advanced/engine-harness.js +2 -2
  4. package/dist/advanced/index.cjs +54 -1
  5. package/dist/advanced/index.d.cts +1 -1
  6. package/dist/advanced/index.d.ts +1 -1
  7. package/dist/advanced/index.js +4 -4
  8. package/dist/advanced/proof-run-core.cjs +54 -1
  9. package/dist/advanced/proof-run-core.js +1 -1
  10. package/dist/advanced/proof-run-engine.cjs +54 -1
  11. package/dist/advanced/proof-run-engine.d.cts +1 -1
  12. package/dist/advanced/proof-run-engine.d.ts +1 -1
  13. package/dist/advanced/proof-run-engine.js +2 -2
  14. package/dist/advanced/runner.js +2 -2
  15. package/dist/{chunk-P2RN2NYR.js → chunk-DZYH67J2.js} +1 -1
  16. package/dist/{chunk-KS3N5APP.js → chunk-EGOVHMBQ.js} +1 -1
  17. package/dist/{chunk-2DW2LBUD.js → chunk-EMVMXVJV.js} +1 -1
  18. package/dist/{chunk-FU73I4V3.js → chunk-RF72NWHM.js} +54 -1
  19. package/dist/{chunk-UTQJHWCQ.js → chunk-WULFU42E.js} +322 -2
  20. package/dist/cli/index.js +3 -3
  21. package/dist/cli.cjs +374 -1
  22. package/dist/cli.js +3 -3
  23. package/dist/engine-harness.cjs +54 -1
  24. package/dist/engine-harness.js +2 -2
  25. package/dist/index.cjs +54 -1
  26. package/dist/index.js +3 -3
  27. package/dist/proof-run-core.cjs +54 -1
  28. package/dist/proof-run-core.js +1 -1
  29. package/dist/{proof-run-engine-Vh9uESqh.d.ts → proof-run-engine-By7oLsF-.d.ts} +3 -3
  30. package/dist/{proof-run-engine-DI1qBmMf.d.cts → proof-run-engine-D80hVFMf.d.cts} +3 -3
  31. package/dist/proof-run-engine.cjs +54 -1
  32. package/dist/proof-run-engine.d.cts +1 -1
  33. package/dist/proof-run-engine.d.ts +1 -1
  34. package/dist/proof-run-engine.js +2 -2
  35. package/dist/runner.js +2 -2
  36. package/package.json +1 -1
package/dist/cli.cjs CHANGED
@@ -244,6 +244,36 @@ function normalizeCaptureScript(value) {
244
244
  const script = normalizeOptionalString(value) || "";
245
245
  return script ? guardProofEvidenceGlobalAssignments(script) : "";
246
246
  }
247
+ function compactCaptureScriptForHeuristics(script) {
248
+ return script.replace(/\/\*[\s\S]*?\*\//g, " ").replace(/(^|[^:])\/\/.*$/gm, "$1 ").replace(/\s+/g, " ").trim().toLowerCase();
249
+ }
250
+ function interactionCaptureScriptLooksPassive(script) {
251
+ const text = compactCaptureScriptForHeuristics(script);
252
+ if (!text) return true;
253
+ const actionPatterns = [
254
+ /\bpage\.(click|dblclick|tap|fill|press|type|check|uncheck|selectoption|dispatch(event)?|goto|reload)\s*\(/,
255
+ /\blocator\s*\([^)]*\)\s*\.\s*(click|dblclick|tap|fill|press|type|check|uncheck|selectoption|dispatch(event)?)\s*\(/,
256
+ /\b(getby(role|text|label|testid|placeholder|title)|getbyalttext)\s*\([^)]*\)\s*\.\s*(click|dblclick|tap|fill|press|type|check|uncheck|selectoption)\s*\(/,
257
+ /\bkeyboard\s*\.\s*(press|type|inserttext)\s*\(/,
258
+ /\bmouse\s*\.\s*(click|dblclick|down|up|move)\s*\(/,
259
+ /\btouchscreen\s*\.\s*tap\s*\(/
260
+ ];
261
+ if (actionPatterns.some((pattern) => pattern.test(text))) return false;
262
+ const evidencePatterns = [
263
+ /\breturn\s+[{[]/,
264
+ /\breturn\s+\w+/,
265
+ /__riddleproofevidence/,
266
+ /\bproof_evidence\b/,
267
+ /\brouteexpectationsource\b/,
268
+ /\bexpectedurl\b/,
269
+ /\bassertions?\b/
270
+ ];
271
+ return !evidencePatterns.some((pattern) => pattern.test(text));
272
+ }
273
+ function setStructuredInteractionCaptureFailure(state, summary) {
274
+ const existing = typeof state.structured_interaction_capture_failure_summary === "string" ? state.structured_interaction_capture_failure_summary.trim() : "";
275
+ if (!existing) state.structured_interaction_capture_failure_summary = summary;
276
+ }
247
277
  function appendProofSummaryLine(state, line) {
248
278
  const text = String(line || "").trim();
249
279
  if (!text) return;
@@ -484,6 +514,9 @@ function proofAssessmentHardBlockersForState(state = {}) {
484
514
  }
485
515
  add(state?.structured_interaction_capture_failure_summary);
486
516
  add(state?.structured_interaction_failure_summary);
517
+ if (isInteractionVerificationMode(normalizedVerificationMode(state)) && !stateHasProofEvidence(state)) {
518
+ add("interaction proof evidence is required before ready_to_ship; proof_evidence_present=false");
519
+ }
487
520
  const mergeRecommendation = String(state?.merge_recommendation || "").trim();
488
521
  if (mergeRecommendation === "do-not-merge" && blockers.length) {
489
522
  add("merge_recommendation=do-not-merge because the proof bundle contains hard blockers.");
@@ -519,7 +552,22 @@ function stateHasAfterEvidence(state = {}) {
519
552
  const observation = objectValue(after.observation);
520
553
  const supporting = objectValue(after.supporting_artifacts);
521
554
  return Boolean(
522
- observation.valid === true && (supporting.has_structured_payload === true || supporting.proof_evidence_present === true || observation.telemetry_ready === true || Object.keys(objectValue(bundle.proof_evidence)).length > 0 || Object.keys(objectValue(after.proof_evidence)).length > 0)
555
+ observation.valid === true && (supporting.has_structured_payload === true || stateHasProofEvidence(state) || observation.telemetry_ready === true)
556
+ );
557
+ }
558
+ function stateHasProofEvidence(state = {}) {
559
+ if (state?.proof_evidence_present === true) return true;
560
+ if (state?.proof_evidence !== void 0 && state?.proof_evidence !== null) {
561
+ if (typeof state.proof_evidence !== "object") return true;
562
+ if (Object.keys(objectValue(state.proof_evidence)).length > 0) return true;
563
+ }
564
+ const bundle = objectValue(state?.evidence_bundle);
565
+ const after = objectValue(bundle.after);
566
+ const supporting = objectValue(after.supporting_artifacts);
567
+ const request = objectValue(state?.proof_assessment_request);
568
+ const structuredEvidence = objectValue(request.structured_evidence);
569
+ return Boolean(
570
+ supporting.proof_evidence_present === true || structuredEvidence.proof_evidence_present === true || Object.keys(objectValue(bundle.proof_evidence)).length > 0 || Object.keys(objectValue(after.proof_evidence)).length > 0
523
571
  );
524
572
  }
525
573
  function validateShipGate(state = {}) {
@@ -725,6 +773,11 @@ function mergeStateFromParams(statePath, params) {
725
773
  state.supervisor_author_packet = parsed;
726
774
  if (typeof parsed?.proof_plan === "string") state.proof_plan = normalizeOptionalString(parsed.proof_plan) || "";
727
775
  if (typeof parsed?.capture_script === "string") state.capture_script = normalizeCaptureScript(parsed.capture_script);
776
+ if (isInteractionVerificationMode(state.verification_mode) && interactionCaptureScriptLooksPassive(state.capture_script || "")) {
777
+ const warning = "Interaction proof capture script appears passive: it does not perform a browser interaction or return structured proof evidence.";
778
+ appendStateWarning(state, "author_warnings", warning);
779
+ setStructuredInteractionCaptureFailure(state, warning);
780
+ }
728
781
  if (parsed?.baseline_understanding_used && typeof parsed.baseline_understanding_used === "object") {
729
782
  state.author_baseline_understanding_used = parsed.baseline_understanding_used;
730
783
  }
@@ -3156,6 +3209,7 @@ var init_proof_run_engine = __esm({
3156
3209
  });
3157
3210
 
3158
3211
  // src/cli.ts
3212
+ var import_node_child_process5 = require("child_process");
3159
3213
  var import_node_fs6 = require("fs");
3160
3214
  var import_node_path6 = __toESM(require("path"), 1);
3161
3215
 
@@ -16851,6 +16905,7 @@ var KNOWN_CLI_OPTIONS = /* @__PURE__ */ new Set([
16851
16905
  "intervalMs",
16852
16906
  "job",
16853
16907
  "jobId",
16908
+ "localCore",
16854
16909
  "maxIterations",
16855
16910
  "navigationTimeout",
16856
16911
  "output",
@@ -16860,6 +16915,8 @@ var KNOWN_CLI_OPTIONS = /* @__PURE__ */ new Set([
16860
16915
  "pollAttempts",
16861
16916
  "pollIntervalMs",
16862
16917
  "port",
16918
+ "pack",
16919
+ "packFile",
16863
16920
  "profile",
16864
16921
  "progressEveryMs",
16865
16922
  "quiet",
@@ -16910,6 +16967,7 @@ function usage() {
16910
16967
  " riddle-proof-loop run-profile --profile <file|json|-> --url <base-url> [--base-url <base-url>] [--runner riddle] [--viewport-name <name[,name...]>] [--strict true|false; default false] [--split-viewports true|false; default false] [--balance-preflight true|false; default true] [--poll-attempts n] [--output <dir>|--output-dir <dir>] [--result-format json|compact-json|summary|none; default json] [--quiet]",
16911
16968
  " riddle-proof-loop run-profile aggregate --profile <file|json|-> --url <base-url> [--base-url <base-url>] --input-dir <dir>|--inputs <path[,path...]> [--output <dir>|--output-dir <dir>] [--result-format json|compact-json|summary|none; default json]",
16912
16969
  " riddle-proof-loop run-profile recover --profile <file|json|-> --url <base-url> [--base-url <base-url>] --job <job-id> [--viewport-name <name[,name...]>] [--output <dir>|--output-dir <dir>] [--result-format json|compact-json|summary|none; default json]",
16970
+ " riddle-proof-loop regression-pack run [--pack oc-flow-regression|--pack-file <file>] [--local-core true|false; default true] [--format json|markdown|compact-json; default json] [--output <dir>|--output-dir <dir>]",
16913
16971
  " riddle-proof-loop profile-body-assertions --artifact <file|url|-> --candidates-json <file|json|-> [--required-json <file|json|->] [--format json|body-contains]",
16914
16972
  " riddle-proof-loop profile-http-status-preflight --profile <file|json|-> --url <base-url> [--format json|summary]",
16915
16973
  " riddle-proof-loop riddle-preview-deploy <build-dir> <label> [--framework spa|static]",
@@ -16988,6 +17046,27 @@ function runProfileViewportNamesOption(options) {
16988
17046
  }
16989
17047
  var DEFAULT_PROFILE_UNSUBMITTED_RETRY_TIMEOUT_MS = 9e4;
16990
17048
  var DEFAULT_PROFILE_UNSUBMITTED_RETRIES = 2;
17049
+ function cliPackageRoot() {
17050
+ const entryPath = process.argv[1] ? (() => {
17051
+ try {
17052
+ return (0, import_node_fs6.realpathSync)(process.argv[1]);
17053
+ } catch {
17054
+ return process.argv[1];
17055
+ }
17056
+ })() : "";
17057
+ const candidates = [
17058
+ entryPath ? import_node_path6.default.resolve(import_node_path6.default.dirname(entryPath), "..") : "",
17059
+ import_node_path6.default.resolve(process.cwd(), "packages", "riddle-proof"),
17060
+ process.cwd()
17061
+ ].filter(Boolean);
17062
+ for (const candidate of candidates) {
17063
+ if ((0, import_node_fs6.existsSync)(import_node_path6.default.join(candidate, "runtime", "tests", "trust_boundary_regression.py")) && (0, import_node_fs6.existsSync)(import_node_path6.default.join(candidate, "examples", "regression-packs"))) {
17064
+ return candidate;
17065
+ }
17066
+ }
17067
+ return candidates[0] || process.cwd();
17068
+ }
17069
+ var CLI_PACKAGE_ROOT = cliPackageRoot();
16991
17070
  function optionNumber(options, ...keys) {
16992
17071
  for (const key of keys) {
16993
17072
  const value = optionString(options, key);
@@ -17068,6 +17147,292 @@ function writeRunProfileResult(result, options) {
17068
17147
  process.stdout.write(`${JSON.stringify(result, null, 2)}
17069
17148
  `);
17070
17149
  }
17150
+ function regressionPackResultFormatOption(options) {
17151
+ const format = optionString(options, "format") ?? optionString(options, "resultFormat") ?? "json";
17152
+ if (format === "md" || format === "summary") return "markdown";
17153
+ if (format === "json" || format === "compact-json" || format === "markdown") return format;
17154
+ throw new Error("--format must be json, compact-json, or markdown.");
17155
+ }
17156
+ function regressionPackPathForCli(options) {
17157
+ const packFile = optionString(options, "packFile");
17158
+ if (packFile) return import_node_path6.default.resolve(packFile);
17159
+ const pack = optionString(options, "pack") || "oc-flow-regression";
17160
+ if ((0, import_node_fs6.existsSync)(pack)) return import_node_path6.default.resolve(pack);
17161
+ const normalized = pack.endsWith(".json") ? pack : `${pack}.json`;
17162
+ return import_node_path6.default.join(CLI_PACKAGE_ROOT, "examples", "regression-packs", normalized);
17163
+ }
17164
+ function readRegressionPackForCli(options) {
17165
+ const filePath = regressionPackPathForCli(options);
17166
+ if (!(0, import_node_fs6.existsSync)(filePath)) throw new Error(`Regression pack not found: ${filePath}`);
17167
+ const parsed = JSON.parse((0, import_node_fs6.readFileSync)(filePath, "utf-8"));
17168
+ if (parsed.version !== "riddle-proof.regression-pack.v1") {
17169
+ throw new Error(`${filePath} is not a riddle-proof.regression-pack.v1 manifest.`);
17170
+ }
17171
+ return { filePath, pack: parsed };
17172
+ }
17173
+ function regressionPackStringArray(value) {
17174
+ return Array.isArray(value) ? value.filter((item) => typeof item === "string" && item.trim().length > 0) : [];
17175
+ }
17176
+ function regressionPackRecord(value) {
17177
+ return value && typeof value === "object" && !Array.isArray(value) ? value : {};
17178
+ }
17179
+ function regressionPackCommandForLocalCore(pack) {
17180
+ const suite = regressionPackRecord(pack.local_core_suite);
17181
+ return cliString(suite.command) || "python3 packages/riddle-proof/runtime/tests/trust_boundary_regression.py";
17182
+ }
17183
+ function regressionPackLocalCoreScriptPath() {
17184
+ return import_node_path6.default.join(CLI_PACKAGE_ROOT, "runtime", "tests", "trust_boundary_regression.py");
17185
+ }
17186
+ function tailLines(text, limit = 40) {
17187
+ const lines = text.split(/\r?\n/).filter((line) => line.trim().length > 0);
17188
+ return lines.slice(-limit);
17189
+ }
17190
+ function parseRegressionPackLocalCoreStdout(stdout) {
17191
+ try {
17192
+ return JSON.parse(stdout);
17193
+ } catch {
17194
+ const start = stdout.indexOf("{");
17195
+ const end = stdout.lastIndexOf("}");
17196
+ if (start >= 0 && end > start) {
17197
+ return JSON.parse(stdout.slice(start, end + 1));
17198
+ }
17199
+ throw new Error("Local core suite did not emit parseable JSON.");
17200
+ }
17201
+ }
17202
+ function regressionPackCaseNames(result) {
17203
+ const results = Array.isArray(result.results) ? result.results : [];
17204
+ return results.map((item) => cliString(regressionPackRecord(item).name)).filter((item) => Boolean(item));
17205
+ }
17206
+ function regressionPackFailedCaseNames(result) {
17207
+ const failed = Array.isArray(result.failed) ? result.failed : [];
17208
+ const namedFailed = failed.map((item) => cliString(regressionPackRecord(item).name)).filter((item) => Boolean(item));
17209
+ if (namedFailed.length) return namedFailed;
17210
+ const results = Array.isArray(result.results) ? result.results : [];
17211
+ return results.map(regressionPackRecord).filter((item) => item.ok === false).map((item) => cliString(item.name)).filter((item) => Boolean(item));
17212
+ }
17213
+ function regressionPackMarkersSeen(result, markers) {
17214
+ const encoded = JSON.stringify(result);
17215
+ return markers.filter((marker) => encoded.includes(marker));
17216
+ }
17217
+ function runRegressionPackLocalCore(pack) {
17218
+ const script = regressionPackLocalCoreScriptPath();
17219
+ const suite = regressionPackRecord(pack.local_core_suite);
17220
+ const requiredCases = regressionPackStringArray(suite.required_cases);
17221
+ const forbiddenMarkers = regressionPackStringArray(pack.forbidden_terminal_markers);
17222
+ const command = regressionPackCommandForLocalCore(pack);
17223
+ const startedAt = (/* @__PURE__ */ new Date()).toISOString();
17224
+ if (!(0, import_node_fs6.existsSync)(script)) {
17225
+ return {
17226
+ requested: true,
17227
+ ok: false,
17228
+ command,
17229
+ started_at: startedAt,
17230
+ finished_at: (/* @__PURE__ */ new Date()).toISOString(),
17231
+ error: `Local core suite script not found: ${script}`,
17232
+ required_cases: requiredCases,
17233
+ missing_required_cases: requiredCases,
17234
+ forbidden_terminal_markers_seen: []
17235
+ };
17236
+ }
17237
+ const child = (0, import_node_child_process5.spawnSync)("python3", [script], {
17238
+ cwd: CLI_PACKAGE_ROOT,
17239
+ encoding: "utf-8",
17240
+ timeout: 12e4
17241
+ });
17242
+ const finishedAt = (/* @__PURE__ */ new Date()).toISOString();
17243
+ let parsed = null;
17244
+ let parseError;
17245
+ try {
17246
+ parsed = parseRegressionPackLocalCoreStdout(child.stdout || "");
17247
+ } catch (error) {
17248
+ parseError = error instanceof Error ? error.message : String(error);
17249
+ }
17250
+ const caseNames = parsed ? regressionPackCaseNames(parsed) : [];
17251
+ const missingRequiredCases = requiredCases.filter((caseId) => !caseNames.includes(caseId));
17252
+ const failedCases = parsed ? regressionPackFailedCaseNames(parsed) : [];
17253
+ const markersSeen = parsed ? regressionPackMarkersSeen(parsed, forbiddenMarkers) : [];
17254
+ const ok = child.status === 0 && parsed?.ok === true && !missingRequiredCases.length && !failedCases.length && !markersSeen.length;
17255
+ return {
17256
+ requested: true,
17257
+ ok,
17258
+ command,
17259
+ executed: {
17260
+ binary: "python3",
17261
+ args: [import_node_path6.default.relative(CLI_PACKAGE_ROOT, script)],
17262
+ cwd: CLI_PACKAGE_ROOT
17263
+ },
17264
+ exit_code: child.status,
17265
+ signal: child.signal,
17266
+ started_at: startedAt,
17267
+ finished_at: finishedAt,
17268
+ suite: cliString(parsed?.suite) || null,
17269
+ case_count: typeof parsed?.case_count === "number" ? parsed.case_count : caseNames.length,
17270
+ passed_case_count: caseNames.length - failedCases.length,
17271
+ required_cases: requiredCases,
17272
+ observed_cases: caseNames,
17273
+ missing_required_cases: missingRequiredCases,
17274
+ failed_cases: failedCases,
17275
+ forbidden_terminal_markers_seen: markersSeen,
17276
+ parse_error: parseError,
17277
+ stderr_tail: tailLines(child.stderr || ""),
17278
+ stdout_tail: ok ? void 0 : tailLines(child.stdout || "")
17279
+ };
17280
+ }
17281
+ function openClawHandoffPromptForRegressionPack(pack, input) {
17282
+ const minimumVersions = regressionPackRecord(pack.minimum_versions);
17283
+ const runtimeGate = regressionPackRecord(pack.runtime_gate);
17284
+ const liveSuite = regressionPackRecord(pack.openclaw_live_suite);
17285
+ const target = regressionPackRecord(liveSuite.target);
17286
+ const cases = Array.isArray(liveSuite.cases) ? liveSuite.cases.map(regressionPackRecord) : [];
17287
+ const fields = regressionPackStringArray(liveSuite.result_log_fields);
17288
+ const forbiddenMarkers = regressionPackStringArray(pack.forbidden_terminal_markers);
17289
+ const lines = [
17290
+ "Run the Riddle Proof OC flow regression pack in small serial chunks.",
17291
+ "",
17292
+ `First call ${cliString(runtimeGate.tool) || "riddle_proof_status"} and count only fresh loaded-runtime runs. Required loaded versions are at least:`,
17293
+ ...Object.entries(minimumVersions).map(([name, version]) => `- ${name}: ${version}`),
17294
+ "",
17295
+ "Target/default flags:",
17296
+ ...Object.entries(target).map(([key, value]) => `- ${key}: ${JSON.stringify(value)}`),
17297
+ "",
17298
+ "Rules:",
17299
+ "- Run cases serially, not as one broad parallel batch.",
17300
+ "- If loaded metadata is stale, stop and restart/reload the gateway before counting results.",
17301
+ "- If any generic lifecycle marker appears, report the exact marker and artifact, then stop the counted batch.",
17302
+ `- Forbidden terminal markers: ${forbiddenMarkers.join(", ") || "none"}.`,
17303
+ fields.length ? `- Log fields for every counted run: ${fields.join(", ")}.` : "",
17304
+ "",
17305
+ "Cases:",
17306
+ ...cases.map((testCase, index) => {
17307
+ const expect = regressionPackRecord(testCase.expect);
17308
+ return [
17309
+ `${index + 1}. ${cliString(testCase.id) || "unnamed-case"}: ${cliString(testCase.intent) || "no intent"}`,
17310
+ ` Expect: ${JSON.stringify(expect)}`
17311
+ ].join("\n");
17312
+ }),
17313
+ "",
17314
+ input.localCoreOk ? "Local generic core suite is green, so OC should only be validating wrapper/runtime behavior." : "Local generic core suite is not green or was not run; do not count OC failures as wrapper-only until local core is green."
17315
+ ].filter((line) => line !== "");
17316
+ return lines.join("\n");
17317
+ }
17318
+ function compactRegressionPackRunResult(result) {
17319
+ const localCore = regressionPackRecord(result.local_core);
17320
+ const openClaw = regressionPackRecord(result.openclaw_live_suite);
17321
+ return {
17322
+ version: result.version,
17323
+ pack_id: result.pack_id,
17324
+ ok: result.ok,
17325
+ local_core: {
17326
+ requested: localCore.requested,
17327
+ ok: localCore.ok,
17328
+ command: localCore.command,
17329
+ case_count: localCore.case_count,
17330
+ missing_required_cases: localCore.missing_required_cases,
17331
+ failed_cases: localCore.failed_cases,
17332
+ forbidden_terminal_markers_seen: localCore.forbidden_terminal_markers_seen
17333
+ },
17334
+ openclaw_live_case_count: openClaw.case_count,
17335
+ output_dir: result.output_dir
17336
+ };
17337
+ }
17338
+ function regressionPackRunMarkdown(result) {
17339
+ const localCore = regressionPackRecord(result.local_core);
17340
+ const runtimeGate = regressionPackRecord(result.runtime_gate);
17341
+ const minimumVersions = regressionPackRecord(result.minimum_versions);
17342
+ const openClaw = regressionPackRecord(result.openclaw_live_suite);
17343
+ const lines = [
17344
+ `# ${cliString(result.public_name) || cliString(result.pack_id) || "Riddle Proof Regression Pack"}`,
17345
+ "",
17346
+ `Status: ${result.ok ? "passed" : "failed"}`,
17347
+ `Pack: ${cliString(result.pack_id) || "unknown"}`,
17348
+ "",
17349
+ "## Local Core",
17350
+ "",
17351
+ `- requested: ${localCore.requested === true}`,
17352
+ `- ok: ${localCore.ok === true}`,
17353
+ `- command: ${cliString(localCore.command) || "n/a"}`,
17354
+ `- cases: ${localCore.case_count ?? "n/a"}`,
17355
+ `- missing required: ${regressionPackStringArray(localCore.missing_required_cases).join(", ") || "none"}`,
17356
+ `- failed cases: ${regressionPackStringArray(localCore.failed_cases).join(", ") || "none"}`,
17357
+ `- forbidden markers seen: ${regressionPackStringArray(localCore.forbidden_terminal_markers_seen).join(", ") || "none"}`,
17358
+ "",
17359
+ "## Runtime Gate",
17360
+ "",
17361
+ `- tool: ${cliString(runtimeGate.tool) || "n/a"}`,
17362
+ ...Object.entries(minimumVersions).map(([name, version]) => `- ${name}: ${version}`),
17363
+ "",
17364
+ "## OpenClaw Live Suite",
17365
+ "",
17366
+ `- case count: ${openClaw.case_count ?? "n/a"}`,
17367
+ `- result log fields: ${regressionPackStringArray(openClaw.result_log_fields).join(", ") || "n/a"}`,
17368
+ "",
17369
+ "## OC Handoff Prompt",
17370
+ "",
17371
+ "```text",
17372
+ cliString(result.openclaw_handoff_prompt) || "",
17373
+ "```",
17374
+ ""
17375
+ ];
17376
+ return `${lines.join("\n")}
17377
+ `;
17378
+ }
17379
+ function writeRegressionPackOutput(outputDir, result) {
17380
+ if (!outputDir) return;
17381
+ (0, import_node_fs6.mkdirSync)(outputDir, { recursive: true });
17382
+ (0, import_node_fs6.writeFileSync)(import_node_path6.default.join(outputDir, "regression-pack-result.json"), `${JSON.stringify(result, null, 2)}
17383
+ `);
17384
+ (0, import_node_fs6.writeFileSync)(import_node_path6.default.join(outputDir, "summary.md"), regressionPackRunMarkdown(result));
17385
+ (0, import_node_fs6.writeFileSync)(import_node_path6.default.join(outputDir, "oc-handoff.md"), `${cliString(result.openclaw_handoff_prompt) || ""}
17386
+ `);
17387
+ }
17388
+ function runRegressionPackForCli(options) {
17389
+ const { filePath, pack } = readRegressionPackForCli(options);
17390
+ const localCoreRequested = optionBoolean(options, "localCore") ?? true;
17391
+ const localCore = localCoreRequested ? runRegressionPackLocalCore(pack) : { requested: false, ok: true, command: regressionPackCommandForLocalCore(pack) };
17392
+ const liveSuite = regressionPackRecord(pack.openclaw_live_suite);
17393
+ const liveCases = Array.isArray(liveSuite.cases) ? liveSuite.cases : [];
17394
+ const localCoreRecord = regressionPackRecord(localCore);
17395
+ const localCoreValidated = localCoreRecord.requested === true && localCoreRecord.ok === true;
17396
+ const ok = localCoreRequested ? localCoreValidated : true;
17397
+ const result = {
17398
+ version: "riddle-proof.regression-pack-run-result.v1",
17399
+ ok,
17400
+ local_core_validated: localCoreValidated,
17401
+ generated_at: (/* @__PURE__ */ new Date()).toISOString(),
17402
+ pack_path: filePath,
17403
+ pack_id: cliString(pack.pack_id) || null,
17404
+ public_name: cliString(pack.public_name) || null,
17405
+ description: cliString(pack.description) || null,
17406
+ minimum_versions: regressionPackRecord(pack.minimum_versions),
17407
+ runtime_gate: regressionPackRecord(pack.runtime_gate),
17408
+ forbidden_terminal_markers: regressionPackStringArray(pack.forbidden_terminal_markers),
17409
+ local_core: localCore,
17410
+ openclaw_live_suite: {
17411
+ target: regressionPackRecord(liveSuite.target),
17412
+ result_log_fields: regressionPackStringArray(liveSuite.result_log_fields),
17413
+ case_count: liveCases.length,
17414
+ case_ids: liveCases.map((item) => cliString(regressionPackRecord(item).id)).filter(Boolean)
17415
+ },
17416
+ openclaw_handoff_prompt: openClawHandoffPromptForRegressionPack(pack, { localCoreOk: localCoreValidated }),
17417
+ output_dir: profileOutputDirOption(options) || null
17418
+ };
17419
+ writeRegressionPackOutput(profileOutputDirOption(options), result);
17420
+ return result;
17421
+ }
17422
+ function writeRegressionPackRunResult(result, options) {
17423
+ const format = regressionPackResultFormatOption(options);
17424
+ if (format === "markdown") {
17425
+ process.stdout.write(regressionPackRunMarkdown(result));
17426
+ return;
17427
+ }
17428
+ if (format === "compact-json") {
17429
+ process.stdout.write(`${JSON.stringify(compactRegressionPackRunResult(result), null, 2)}
17430
+ `);
17431
+ return;
17432
+ }
17433
+ process.stdout.write(`${JSON.stringify(result, null, 2)}
17434
+ `);
17435
+ }
17071
17436
  function previewFrameworkOption(options) {
17072
17437
  const framework = optionString(options, "framework") ?? "static";
17073
17438
  if (framework === "spa" || framework === "static") return framework;
@@ -20732,6 +21097,14 @@ async function main() {
20732
21097
  process.exitCode = profileStatusExitCode(profile, result.status);
20733
21098
  return;
20734
21099
  }
21100
+ if (command === "regression-pack") {
21101
+ const action = positional[1] || "run";
21102
+ if (action !== "run") throw new Error("Only `regression-pack run` is supported.");
21103
+ const result = runRegressionPackForCli(options);
21104
+ writeRegressionPackRunResult(result, options);
21105
+ process.exitCode = result.ok ? 0 : 1;
21106
+ return;
21107
+ }
20735
21108
  if (command === "profile-http-status-preflight") {
20736
21109
  const profile = normalizeProfileForCli(options);
20737
21110
  const result = await preflightRiddleProofProfileHttpStatusChecks(profile);
package/dist/cli.js CHANGED
@@ -1,11 +1,11 @@
1
1
  #!/usr/bin/env node
2
- import "./chunk-UTQJHWCQ.js";
2
+ import "./chunk-WULFU42E.js";
3
3
  import "./chunk-PEWAIEER.js";
4
4
  import "./chunk-TWTEUS7R.js";
5
- import "./chunk-2DW2LBUD.js";
5
+ import "./chunk-EMVMXVJV.js";
6
6
  import "./chunk-YZUVEJ5B.js";
7
7
  import "./chunk-FMOYUYH2.js";
8
- import "./chunk-FU73I4V3.js";
8
+ import "./chunk-RF72NWHM.js";
9
9
  import "./chunk-4FOHZ7JG.js";
10
10
  import "./chunk-JFQXAJH2.js";
11
11
  import "./chunk-EEIYUZXE.js";
@@ -244,6 +244,36 @@ function normalizeCaptureScript(value) {
244
244
  const script = normalizeOptionalString(value) || "";
245
245
  return script ? guardProofEvidenceGlobalAssignments(script) : "";
246
246
  }
247
+ function compactCaptureScriptForHeuristics(script) {
248
+ return script.replace(/\/\*[\s\S]*?\*\//g, " ").replace(/(^|[^:])\/\/.*$/gm, "$1 ").replace(/\s+/g, " ").trim().toLowerCase();
249
+ }
250
+ function interactionCaptureScriptLooksPassive(script) {
251
+ const text = compactCaptureScriptForHeuristics(script);
252
+ if (!text) return true;
253
+ const actionPatterns = [
254
+ /\bpage\.(click|dblclick|tap|fill|press|type|check|uncheck|selectoption|dispatch(event)?|goto|reload)\s*\(/,
255
+ /\blocator\s*\([^)]*\)\s*\.\s*(click|dblclick|tap|fill|press|type|check|uncheck|selectoption|dispatch(event)?)\s*\(/,
256
+ /\b(getby(role|text|label|testid|placeholder|title)|getbyalttext)\s*\([^)]*\)\s*\.\s*(click|dblclick|tap|fill|press|type|check|uncheck|selectoption)\s*\(/,
257
+ /\bkeyboard\s*\.\s*(press|type|inserttext)\s*\(/,
258
+ /\bmouse\s*\.\s*(click|dblclick|down|up|move)\s*\(/,
259
+ /\btouchscreen\s*\.\s*tap\s*\(/
260
+ ];
261
+ if (actionPatterns.some((pattern) => pattern.test(text))) return false;
262
+ const evidencePatterns = [
263
+ /\breturn\s+[{[]/,
264
+ /\breturn\s+\w+/,
265
+ /__riddleproofevidence/,
266
+ /\bproof_evidence\b/,
267
+ /\brouteexpectationsource\b/,
268
+ /\bexpectedurl\b/,
269
+ /\bassertions?\b/
270
+ ];
271
+ return !evidencePatterns.some((pattern) => pattern.test(text));
272
+ }
273
+ function setStructuredInteractionCaptureFailure(state, summary) {
274
+ const existing = typeof state.structured_interaction_capture_failure_summary === "string" ? state.structured_interaction_capture_failure_summary.trim() : "";
275
+ if (!existing) state.structured_interaction_capture_failure_summary = summary;
276
+ }
247
277
  function appendProofSummaryLine(state, line) {
248
278
  const text = String(line || "").trim();
249
279
  if (!text) return;
@@ -484,6 +514,9 @@ function proofAssessmentHardBlockersForState(state = {}) {
484
514
  }
485
515
  add(state?.structured_interaction_capture_failure_summary);
486
516
  add(state?.structured_interaction_failure_summary);
517
+ if (isInteractionVerificationMode(normalizedVerificationMode(state)) && !stateHasProofEvidence(state)) {
518
+ add("interaction proof evidence is required before ready_to_ship; proof_evidence_present=false");
519
+ }
487
520
  const mergeRecommendation = String(state?.merge_recommendation || "").trim();
488
521
  if (mergeRecommendation === "do-not-merge" && blockers.length) {
489
522
  add("merge_recommendation=do-not-merge because the proof bundle contains hard blockers.");
@@ -519,7 +552,22 @@ function stateHasAfterEvidence(state = {}) {
519
552
  const observation = objectValue(after.observation);
520
553
  const supporting = objectValue(after.supporting_artifacts);
521
554
  return Boolean(
522
- observation.valid === true && (supporting.has_structured_payload === true || supporting.proof_evidence_present === true || observation.telemetry_ready === true || Object.keys(objectValue(bundle.proof_evidence)).length > 0 || Object.keys(objectValue(after.proof_evidence)).length > 0)
555
+ observation.valid === true && (supporting.has_structured_payload === true || stateHasProofEvidence(state) || observation.telemetry_ready === true)
556
+ );
557
+ }
558
+ function stateHasProofEvidence(state = {}) {
559
+ if (state?.proof_evidence_present === true) return true;
560
+ if (state?.proof_evidence !== void 0 && state?.proof_evidence !== null) {
561
+ if (typeof state.proof_evidence !== "object") return true;
562
+ if (Object.keys(objectValue(state.proof_evidence)).length > 0) return true;
563
+ }
564
+ const bundle = objectValue(state?.evidence_bundle);
565
+ const after = objectValue(bundle.after);
566
+ const supporting = objectValue(after.supporting_artifacts);
567
+ const request = objectValue(state?.proof_assessment_request);
568
+ const structuredEvidence = objectValue(request.structured_evidence);
569
+ return Boolean(
570
+ supporting.proof_evidence_present === true || structuredEvidence.proof_evidence_present === true || Object.keys(objectValue(bundle.proof_evidence)).length > 0 || Object.keys(objectValue(after.proof_evidence)).length > 0
523
571
  );
524
572
  }
525
573
  function validateShipGate(state = {}) {
@@ -725,6 +773,11 @@ function mergeStateFromParams(statePath, params) {
725
773
  state.supervisor_author_packet = parsed;
726
774
  if (typeof parsed?.proof_plan === "string") state.proof_plan = normalizeOptionalString(parsed.proof_plan) || "";
727
775
  if (typeof parsed?.capture_script === "string") state.capture_script = normalizeCaptureScript(parsed.capture_script);
776
+ if (isInteractionVerificationMode(state.verification_mode) && interactionCaptureScriptLooksPassive(state.capture_script || "")) {
777
+ const warning = "Interaction proof capture script appears passive: it does not perform a browser interaction or return structured proof evidence.";
778
+ appendStateWarning(state, "author_warnings", warning);
779
+ setStructuredInteractionCaptureFailure(state, warning);
780
+ }
728
781
  if (parsed?.baseline_understanding_used && typeof parsed.baseline_understanding_used === "object") {
729
782
  state.author_baseline_understanding_used = parsed.baseline_understanding_used;
730
783
  }
@@ -2,10 +2,10 @@ import {
2
2
  createDisabledRiddleProofAgentAdapter,
3
3
  readRiddleProofRunStatus,
4
4
  runRiddleProofEngineHarness
5
- } from "./chunk-2DW2LBUD.js";
5
+ } from "./chunk-EMVMXVJV.js";
6
6
  import "./chunk-YZUVEJ5B.js";
7
7
  import "./chunk-FMOYUYH2.js";
8
- import "./chunk-FU73I4V3.js";
8
+ import "./chunk-RF72NWHM.js";
9
9
  import "./chunk-4FOHZ7JG.js";
10
10
  import "./chunk-VY4Y5U57.js";
11
11
  import "./chunk-MLKGABMK.js";
package/dist/index.cjs CHANGED
@@ -244,6 +244,36 @@ function normalizeCaptureScript(value) {
244
244
  const script = normalizeOptionalString(value) || "";
245
245
  return script ? guardProofEvidenceGlobalAssignments(script) : "";
246
246
  }
247
+ function compactCaptureScriptForHeuristics(script) {
248
+ return script.replace(/\/\*[\s\S]*?\*\//g, " ").replace(/(^|[^:])\/\/.*$/gm, "$1 ").replace(/\s+/g, " ").trim().toLowerCase();
249
+ }
250
+ function interactionCaptureScriptLooksPassive(script) {
251
+ const text = compactCaptureScriptForHeuristics(script);
252
+ if (!text) return true;
253
+ const actionPatterns = [
254
+ /\bpage\.(click|dblclick|tap|fill|press|type|check|uncheck|selectoption|dispatch(event)?|goto|reload)\s*\(/,
255
+ /\blocator\s*\([^)]*\)\s*\.\s*(click|dblclick|tap|fill|press|type|check|uncheck|selectoption|dispatch(event)?)\s*\(/,
256
+ /\b(getby(role|text|label|testid|placeholder|title)|getbyalttext)\s*\([^)]*\)\s*\.\s*(click|dblclick|tap|fill|press|type|check|uncheck|selectoption)\s*\(/,
257
+ /\bkeyboard\s*\.\s*(press|type|inserttext)\s*\(/,
258
+ /\bmouse\s*\.\s*(click|dblclick|down|up|move)\s*\(/,
259
+ /\btouchscreen\s*\.\s*tap\s*\(/
260
+ ];
261
+ if (actionPatterns.some((pattern) => pattern.test(text))) return false;
262
+ const evidencePatterns = [
263
+ /\breturn\s+[{[]/,
264
+ /\breturn\s+\w+/,
265
+ /__riddleproofevidence/,
266
+ /\bproof_evidence\b/,
267
+ /\brouteexpectationsource\b/,
268
+ /\bexpectedurl\b/,
269
+ /\bassertions?\b/
270
+ ];
271
+ return !evidencePatterns.some((pattern) => pattern.test(text));
272
+ }
273
+ function setStructuredInteractionCaptureFailure(state, summary) {
274
+ const existing = typeof state.structured_interaction_capture_failure_summary === "string" ? state.structured_interaction_capture_failure_summary.trim() : "";
275
+ if (!existing) state.structured_interaction_capture_failure_summary = summary;
276
+ }
247
277
  function appendProofSummaryLine(state, line) {
248
278
  const text = String(line || "").trim();
249
279
  if (!text) return;
@@ -484,6 +514,9 @@ function proofAssessmentHardBlockersForState(state = {}) {
484
514
  }
485
515
  add(state?.structured_interaction_capture_failure_summary);
486
516
  add(state?.structured_interaction_failure_summary);
517
+ if (isInteractionVerificationMode(normalizedVerificationMode(state)) && !stateHasProofEvidence(state)) {
518
+ add("interaction proof evidence is required before ready_to_ship; proof_evidence_present=false");
519
+ }
487
520
  const mergeRecommendation = String(state?.merge_recommendation || "").trim();
488
521
  if (mergeRecommendation === "do-not-merge" && blockers.length) {
489
522
  add("merge_recommendation=do-not-merge because the proof bundle contains hard blockers.");
@@ -519,7 +552,22 @@ function stateHasAfterEvidence(state = {}) {
519
552
  const observation = objectValue(after.observation);
520
553
  const supporting = objectValue(after.supporting_artifacts);
521
554
  return Boolean(
522
- observation.valid === true && (supporting.has_structured_payload === true || supporting.proof_evidence_present === true || observation.telemetry_ready === true || Object.keys(objectValue(bundle.proof_evidence)).length > 0 || Object.keys(objectValue(after.proof_evidence)).length > 0)
555
+ observation.valid === true && (supporting.has_structured_payload === true || stateHasProofEvidence(state) || observation.telemetry_ready === true)
556
+ );
557
+ }
558
+ function stateHasProofEvidence(state = {}) {
559
+ if (state?.proof_evidence_present === true) return true;
560
+ if (state?.proof_evidence !== void 0 && state?.proof_evidence !== null) {
561
+ if (typeof state.proof_evidence !== "object") return true;
562
+ if (Object.keys(objectValue(state.proof_evidence)).length > 0) return true;
563
+ }
564
+ const bundle = objectValue(state?.evidence_bundle);
565
+ const after = objectValue(bundle.after);
566
+ const supporting = objectValue(after.supporting_artifacts);
567
+ const request = objectValue(state?.proof_assessment_request);
568
+ const structuredEvidence = objectValue(request.structured_evidence);
569
+ return Boolean(
570
+ supporting.proof_evidence_present === true || structuredEvidence.proof_evidence_present === true || Object.keys(objectValue(bundle.proof_evidence)).length > 0 || Object.keys(objectValue(after.proof_evidence)).length > 0
523
571
  );
524
572
  }
525
573
  function validateShipGate(state = {}) {
@@ -725,6 +773,11 @@ function mergeStateFromParams(statePath, params) {
725
773
  state.supervisor_author_packet = parsed;
726
774
  if (typeof parsed?.proof_plan === "string") state.proof_plan = normalizeOptionalString(parsed.proof_plan) || "";
727
775
  if (typeof parsed?.capture_script === "string") state.capture_script = normalizeCaptureScript(parsed.capture_script);
776
+ if (isInteractionVerificationMode(state.verification_mode) && interactionCaptureScriptLooksPassive(state.capture_script || "")) {
777
+ const warning = "Interaction proof capture script appears passive: it does not perform a browser interaction or return structured proof evidence.";
778
+ appendStateWarning(state, "author_warnings", warning);
779
+ setStructuredInteractionCaptureFailure(state, warning);
780
+ }
728
781
  if (parsed?.baseline_understanding_used && typeof parsed.baseline_understanding_used === "object") {
729
782
  state.author_baseline_understanding_used = parsed.baseline_understanding_used;
730
783
  }