ultra11y 5.33.1 → 5.34.1

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
       "name": "ultra11y",
       "source": "./",
       "description": "Automatic accessibility review of the code under change before every commit, push and pull request, plus full WCAG 2.2 AA / RGAA audits and conformance deliverables on demand.",
-      "version": "5.33.1",
+      "version": "5.34.1",
       "author": { "name": "maxgfr" },
       "homepage": "https://github.com/maxgfr/ultra11y",
       "repository": "https://github.com/maxgfr/ultra11y",

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "ultra11y",
   "description": "Accessibility for WCAG 2.2 AA and country standards (RGAA…): an automatic review of the code under change before every commit, push and pull request, plus full audits and conformance deliverables on demand.",
-  "version": "5.33.1",
+  "version": "5.34.1",
   "author": { "name": "maxgfr" },
   "homepage": "https://github.com/maxgfr/ultra11y",
   "repository": "https://github.com/maxgfr/ultra11y",

package/.codex-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "ultra11y",
   "description": "Accessibility for WCAG 2.2 AA and country standards (RGAA…): an automatic review of the code under change before every commit, push and pull request, plus full audits and conformance deliverables on demand.",
-  "version": "5.33.1",
+  "version": "5.34.1",
   "author": { "name": "maxgfr", "url": "https://github.com/maxgfr" },
   "homepage": "https://github.com/maxgfr/ultra11y",
   "repository": "https://github.com/maxgfr/ultra11y",

package/.opencode/plugins/ultra11y.js

@@ -27,7 +27,7 @@ import { fileURLToPath } from "node:url";
 // Ownership marker: `ultra11y install --opencode` refuses to overwrite a plugin file that
 // does not carry it, so a user's own ultra11y.js is never clobbered.
 const ULTRA11Y_OPENCODE_PLUGIN = "ULTRA11Y_OPENCODE_PLUGIN";
-const ULTRA11Y_PLUGIN_VERSION = "5.33.1";
+const ULTRA11Y_PLUGIN_VERSION = "5.34.1";
 
 const here = dirname(fileURLToPath(import.meta.url));
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ultra11y",
-  "version": "5.33.1",
+  "version": "5.34.1",
   "description": "Audit HTML/CSS/JSX against WCAG 2.2 AA and produce a dated compliance report, or author/review accessible markup (native-HTML-first). A skills.sh agent skill: an install-free static engine (bundled, no npm install) plus the agent's judgment, with check/verify gates against hallucinated non-conformities, and pluggable in-repo standards packs (RGAA France first; Section 508 & EN 301 549 next).",
   "type": "module",
   "packageManager": "pnpm@10.33.0",

package/scripts/ultra11y.mjs

@@ -6,7 +6,7 @@ import { join as join52, relative as relative5, resolve as resolve16, sep as sep
 import { fileURLToPath as fileURLToPath5, pathToFileURL as pathToFileURL3 } from "url";
 
 // src/types.ts
-var VERSION = "5.33.1";
+var VERSION = "5.34.1";
 var SCHEMA_VERSION = 2;
 var INAPPLICABLE_STATUS = "C";
 
@@ -2293,7 +2293,7 @@ var wcag_universe_default = {
   wcagVersion: "2.2",
   source: "https://www.w3.org/TR/WCAG22/",
   criteriaSource: "https://github.com/w3c/wcag",
-  provenance: "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-24 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
+  provenance: "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-25 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
   criteria: [
     {
       id: "1.1.1",
@@ -59367,6 +59367,12 @@ function auditFiles(audit2, cwd) {
     return /* @__PURE__ */ new Set();
   }
 }
+function offHarvestHint(criteriaId, evidence) {
+  const anchors = [...new Set(evidence.map((e) => e.selector?.trim() || `${e.file}:${e.line}`).filter(Boolean))];
+  const shown = anchors.slice(0, 4).join(", ");
+  const rest = anchors.length > 4 ? `, +${anchors.length - 4}` : "";
+  return anchors.length ? ` \u2014 this anchor is not among the evidence harvested for ${criteriaId}, so the snippet was verified literally rather than vouched for by the harvest. Cite one of: ${shown}${rest}.` : ` \u2014 ${criteriaId} was harvested no evidence, so nothing can vouch for a transcription here. If the criterion has no subject in scope, the verdict is NA with a justification and no citation to ground.`;
+}
 function tagOf2(x) {
   const fromSnippet = /<\s*([a-zA-Z][\w-]*)/.exec(x.snippet ?? "");
   if (fromSnippet) return fromSnippet[1].toLowerCase();
@@ -59538,8 +59544,8 @@ function applyAdjudication(audit2, adj, opts = {}) {
     }
     return engineNcCache.get(key2) ?? EMPTY_IDS;
   };
-  const toGround = (criteriaId, g, fallback) => {
-    const entry = { g, ...fallback ? { fallback } : {} };
+  const toGround = (criteriaId, g, fallback, offHarvest) => {
+    const entry = { g, ...fallback ? { fallback } : {}, ...offHarvest ? { offHarvest } : {} };
     const list = groundInputs.get(criteriaId);
     if (list) list.push(entry);
     else groundInputs.set(criteriaId, [entry]);
@@ -59584,7 +59590,7 @@ function applyAdjudication(audit2, adj, opts = {}) {
               anchor.representative ? { file: anchor.at.file, line: anchor.at.line, selector: anchor.at.selector ?? c2.selector, snippet: anchor.at.snippet } : { file: c2.file, line: c2.line, selector: anchor.at.selector ?? c2.selector }
             );
           } else {
-            toGround(it.criteriaId, cite);
+            toGround(it.criteriaId, cite, void 0, anchor ? void 0 : offHarvestHint(it.criteriaId, it.evidence));
           }
         }
       }
@@ -59644,7 +59650,7 @@ function applyAdjudication(audit2, adj, opts = {}) {
   }
   const grounding = { grounded: 0, moved: 0, failed: 0, issues: [] };
   for (const [criteriaId, inputs] of groundInputs) {
-    for (const { g, fallback } of inputs) {
+    for (const { g, fallback, offHarvest } of inputs) {
       let r = groundFinding(g, { cwd: opts.cwd });
       if (!r.ok && fallback) {
         const viaAnchor = groundFinding(fallback, { cwd: opts.cwd });
@@ -59656,8 +59662,9 @@ function applyAdjudication(audit2, adj, opts = {}) {
       } else {
         grounding.failed++;
         if (r.issue) {
-          grounding.issues.push(r.issue);
-          blame(criteriaId, r.issue);
+          const issue = offHarvest && r.issue.startsWith("cited snippet not found") ? `${r.issue}${offHarvest}` : r.issue;
+          grounding.issues.push(issue);
+          blame(criteriaId, issue);
         }
       }
     }
@@ -60378,7 +60385,9 @@ var VERDICT_KINDS = `Rule it (the apply gate is FAIL-CLOSED \u2014 a verdict mis
    - \`NC\` (non-conforming) \u2014 REQUIRES \`findings\`: at least one groundable \`{ file, line, selector?, message, snippet?, severity?, normativeRef }\` pointing at REAL source. The fold re-grounds every finding; an invented file:line is rejected, and so is a finding with no \`file\` at all. \`normativeRef\` MUST cite the precise failed test \u2014 under a country standard, one of the criterion's OWN numbered tests, listed in its brief under \xAB tests to rule on \xBB. A WCAG id looks alike, denotes an unrelated test, and is rejected.
    - \`NA\` (not applicable) \u2014 REQUIRES \`justification\`, AND \`citations[]\` whenever evidence WAS presented, to say which of those items fall outside the criterion's scope.
    - \`manual\` (still undecidable) \u2014 REQUIRES \`reason\`: \`needs-rendered-dom\` (only a rendered DOM can decide it, and no capture in this run carries its subject) or \`undecidable\` (the evidence cannot settle it either way).`;
-var ABSENCE_RULE = `AN NC SHAPED LIKE AN ABSENCE IS STILL ANCHORED. \xAB No second navigation system \xBB, \xAB no search engine \xBB, \xAB no error message suggests the expected format \xBB \u2014 an absence is OBSERVED somewhere: cite the element and the page you observed it on. And when the criterion's subject exists nowhere in the audited scope, the verdict is \`NA\` with its justification, never \`NC\`.`;
+var ABSENCE_RULE = `AN NC SHAPED LIKE AN ABSENCE IS STILL ANCHORED. \xAB No second navigation system \xBB, \xAB no search engine \xBB, \xAB no error message suggests the expected format \xBB \u2014 an absence is OBSERVED somewhere: cite the element and the page you observed it on. And when the criterion's subject exists nowhere in the audited scope, the verdict is \`NA\` with its justification, never \`NC\`.
+
+AND CITE FROM THIS CRITERION'S OWN ANCHORS, NOT THE THING YOU ARE RULING OUT. An absence pulls you toward the element you are arguing ABOUT \u2014 the search form that is not a site search engine, the menu that is not a second navigation system \u2014 and that element is, precisely because it is off-topic, absent from what this criterion was harvested. A citation on one of the criterion's own anchors is vouched for by the harvest; one outside it has its snippet verified LITERALLY, character for character, and a retyping then fails. Cite the region you inspected \u2014 the \`header\`, the \`nav\`, the \`footer\` the brief listed \u2014 and say in the justification what you did not find in it.`;
 var CAPTURE_RULE = `THE RENDERED PAGE MAY BE ON DISK. When a criterion's evidence is anchored under \`.ultra11y/pages/<id>/\`, the browser already ran: \`dom.html\`, \`styles.json\`, \`boxes.json\`, \`axtree.json\` and \`screen.png\` are there to read. \`needs-rendered-dom\` is refused on such a criterion \u2014 decide it from those files, or answer \`undecidable\` and say what the capture does not settle.`;
 var NEVER_GUESS_RULE = `Never guess. A criterion you cannot decide from real evidence stays \`manual\` with its reason \u2014 that is a valid, honest verdict, and it is worth more than a verdict the gate throws away.`;
 var SCOPE_RULE = `Rule ONLY on the criteria presented. Never introduce another \u2014 a verdict for a criterion nobody asked about is dropped, and under the fold it would otherwise overwrite what the deterministic engine already decided.`;
@@ -60588,6 +60597,7 @@ function applyRawVerdicts(items, verdicts) {
 var ALLOWED_TOOLS = "Read,Grep,Glob";
 var MAX_ATTEMPTS2 = 4;
 var DEFAULT_CLI_MODEL = "claude-haiku-4-5-20251001";
+var EFFORT_LEVELS = ["low", "medium", "high", "xhigh", "max"];
 var DEFAULT_TIMEOUT_MS2 = 10 * 6e4;
 function realSpawn(argv, input, timeoutMs) {
   return new Promise((resolve17, reject) => {
@@ -60664,6 +60674,7 @@ function cliArgv(opts, items) {
     opts.model ?? DEFAULT_CLI_MODEL
   ];
   if (opts.maxBudgetUsd !== void 0) argv.push("--max-budget-usd", String(opts.maxBudgetUsd));
+  if (opts.effort) argv.push("--effort", opts.effort);
   return argv;
 }
 function verdictsFromText(text) {
@@ -60726,6 +60737,7 @@ async function refuteBatchCli(items, prompt, opts, lang = "en") {
     opts.model ?? DEFAULT_CLI_MODEL
   ];
   if (opts.maxBudgetUsd !== void 0) argv.push("--max-budget-usd", String(opts.maxBudgetUsd));
+  if (opts.effort) argv.push("--effort", opts.effort);
   return runCli2(argv, prompt, opts, (env) => {
     const structured = env.structured_output?.verdicts;
     const raw = Array.isArray(structured) ? structured : verdictsArrayFromText(env.result ?? "");
@@ -68889,7 +68901,7 @@ Usage:
   ultra11y orchestrate --run <dir> [--phase adjudicate|verify-report] [--eco] [--list] [--lang auto|en|fr]
   ultra11y fix      <globs\u2026 | -> [--write] [--iterate] [--changed | --since <ref> | --staged] [--safe] [--include <glob>] [--exclude <glob>] [--ext <list>] [--only <ids>] [--jsx] [--json] [--lang auto|en|fr]
   ultra11y init     [--hook] [--ci] [--baseline] [--fail-on blocking|major|minor]
-  ultra11y judge    --in <audit.json> [--standard <pack>] [--runner api|cli] [--grain batch|criterion] [--max <n>] [--model <id>] [--max-budget-usd <n>] [--timeout <s>] [--out <dir>] [--apply]   (adjudicate the manual criteria with a model \u2014 api needs ANTHROPIC_API_KEY; cli runs the Claude Code CLI and takes CLAUDE_CODE_OAUTH_TOKEN too)
+  ultra11y judge    --in <audit.json> [--standard <pack>] [--runner api|cli] [--grain batch|criterion] [--max <n>] [--model <id>] [--effort <level>] [--max-budget-usd <n>] [--timeout <s>] [--out <dir>] [--apply]   (adjudicate the manual criteria with a model \u2014 api needs ANTHROPIC_API_KEY; cli runs the Claude Code CLI and takes CLAUDE_CODE_OAUTH_TOKEN too)
   ultra11y pack     check <pack.json> [--guidance <g.json>] [--json]  |  pack scaffold
   ultra11y scan     <url|file\u2026> [--runtime auto|local|docker] [--cwd <dir>] [--storage-state <file>] [--no-interact] [--interact-clicks] [--no-snapshot] [--merge <audit.json>] [--out <dir>] [--json]
   ultra11y scan     --sample [--runtime \u2026] [--cwd <dir>] [--storage-state <file>] [--merge <audit.json>] [--json]   (scan the .ultra11yrc.json page sample)
@@ -69122,6 +69134,8 @@ Options:
   --provider <id>    tickets: auto|github|gitlab|jira. 'auto' reads ULTRA11Y_TICKET_PROVIDER,
                      then .ultra11yrc.json, then the git remote (Jira is never auto-detected)
   --grain <mode>     tickets: what one ticket is \u2014 criterion (default) | page | page-criterion | single | file
+  --effort <level>   judge (--runner cli): reasoning effort per invocation \u2014 low | medium |
+                     high | xhigh | max. Unset leaves the CLI's own default
   --transport <t>    tickets: auto|cli|rest (mcp: stdio|http). 'auto' prefers the CLI (gh/glab),
                      falling back to REST when only a token is available. Jira is REST-only
   --max-tickets <n>  tickets: refuse to file more than n tickets in one run (default 200)
@@ -69325,6 +69339,7 @@ var VALUE_FLAGS2 = /* @__PURE__ */ new Set([
   "runner",
   "grain",
   "max-budget-usd",
+  "effort",
   "timeout",
   "concurrency",
   "out",
@@ -71460,6 +71475,15 @@ async function cmdJudge(p) {
     console.error(`ultra11y judge: --grain '${grain}' is not a grain \u2014 expected 'batch' or 'criterion'.`);
     return 2;
   }
+  const effort = typeof p.flags.effort === "string" ? p.flags.effort : void 0;
+  if (effort !== void 0 && !EFFORT_LEVELS.includes(effort)) {
+    console.error(`ultra11y judge: --effort '${effort}' is not a level \u2014 expected ${EFFORT_LEVELS.map((l) => `'${l}'`).join(", ")}.`);
+    return 2;
+  }
+  if (effort !== void 0 && runner !== "cli") {
+    console.error("ultra11y judge: --effort applies to `--runner cli` only \u2014 the Messages API has no session to set an effort on.");
+    return 2;
+  }
   const key2 = typeof p.flags["api-key"] === "string" && p.flags["api-key"] ? p.flags["api-key"] : apiKeyFromEnv();
   if (!key2 && runner === "api") {
     console.error(
@@ -71533,6 +71557,7 @@ async function cmdJudge(p) {
     // requests do. `--concurrency` overrides it for a runner with room.
     concurrency: runner === "cli" ? cliConcurrency : void 0,
     maxBudgetUsd: Number.isFinite(maxBudgetUsd) ? maxBudgetUsd : void 0,
+    effort,
     timeoutMs: Number.isFinite(timeoutMs) ? timeoutMs : void 0,
     onCost: (usd) => {
       spent += usd;

package/scripts/vendor/wcag-2.2-fr.json

@@ -1,6 +1,6 @@
 {
   "source": "https://www.w3.org/Translations/WCAG22-fr/",
-  "fetchedAt": "2026-08-24",
+  "fetchedAt": "2026-08-25",
   "principles": {
     "1": "Perceptible",
     "2": "Utilisable",

package/scripts/vendor/wcag-2.2-sc-universe.json

@@ -2,7 +2,7 @@
   "wcagVersion": "2.2",
   "source": "https://www.w3.org/TR/WCAG22/",
   "criteriaSource": "https://github.com/w3c/wcag",
-  "provenance": "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-24 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
+  "provenance": "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-25 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
   "principles": [
     {
       "number": 1,

package/scripts/vendor/wcag-2.2-techniques.json

@@ -2,7 +2,7 @@
   "source": "https://github.com/w3c/wcag",
   "understanding": "https://www.w3.org/WAI/WCAG22/Understanding/",
   "license": "W3C Document License (code samples: W3C Software License)",
-  "fetchedAt": "2026-08-24",
+  "fetchedAt": "2026-08-25",
   "criteria": {
     "1.1.1": {
       "slug": "non-text-content",

package/scripts/vendor/wcag-2.2-text.json

@@ -1,7 +1,7 @@
 {
   "source": "https://github.com/w3c/wcag",
   "license": "W3C Document License",
-  "fetchedAt": "2026-08-24",
+  "fetchedAt": "2026-08-25",
   "criteria": {
     "1.1.1": "All non-text content that is presented to the user has a text alternative that serves the equivalent purpose, except for the situations listed below.\nControls, Input:\nIf non-text content is a control or accepts user input, then it has a name that describes its purpose. (Refer to Success Criterion 4.1.2 Name, Role, Value for additional requirements for controls and content that accepts user input.)\nTime-Based Media:\nIf non-text content is time-based media, then text alternatives at least provide descriptive identification of the non-text content. (Refer to Guideline 1.2 for additional requirements for media.)\nTest:\nIf non-text content is a test or exercise that would be invalid if presented in text, then text alternatives at least provide descriptive identification of the non-text content.\nSensory:\nIf non-text content is primarily intended to create a specific sensory experience, then text alternatives at least provide descriptive identification of the non-text content.\nCAPTCHA:\nIf the purpose of non-text content is to confirm that content is being accessed by a person rather than a computer, then text alternatives that identify and describe the purpose of the non-text content are provided, and alternative forms of CAPTCHA using output modes for different types of sensory perception are provided to accommodate different disabilities.\nDecoration, Formatting, Invisible:\nIf non-text content is pure decoration, is used only for visual formatting, or is not presented to users, then it is implemented in a way that it can be ignored by assistive technology.",
     "1.2.1": "For prerecorded audio-only and prerecorded video-only media, the following are true, except when the audio or video is a media alternative for text and is clearly labeled as such:\nPrerecorded Audio-only:\nAn alternative for time-based media is provided that presents equivalent information for prerecorded audio-only content.\nPrerecorded Video-only:\nEither an alternative for time-based media or an audio track is provided that presents equivalent information for prerecorded video-only content.",

package/skills/review-a11y/SKILL.md

@@ -5,7 +5,7 @@ when_to_use: "Invoke WITHOUT being asked whenever changed front-end code should
 allowed-tools: Bash(node ${CLAUDE_SKILL_DIR}/scripts/ultra11y.mjs *), Read, Grep, Glob
 license: MIT
 metadata:
-  version: 5.33.1
+  version: 5.34.1
 ---
 
 # review-a11y — accessibility review of changed code

package/skills/review-a11y/scripts/ultra11y.mjs

@@ -6,7 +6,7 @@ import { join as join52, relative as relative5, resolve as resolve16, sep as sep
 import { fileURLToPath as fileURLToPath5, pathToFileURL as pathToFileURL3 } from "url";
 
 // src/types.ts
-var VERSION = "5.33.1";
+var VERSION = "5.34.1";
 var SCHEMA_VERSION = 2;
 var INAPPLICABLE_STATUS = "C";
 
@@ -2293,7 +2293,7 @@ var wcag_universe_default = {
   wcagVersion: "2.2",
   source: "https://www.w3.org/TR/WCAG22/",
   criteriaSource: "https://github.com/w3c/wcag",
-  provenance: "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-24 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
+  provenance: "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-25 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
   criteria: [
     {
       id: "1.1.1",
@@ -59367,6 +59367,12 @@ function auditFiles(audit2, cwd) {
     return /* @__PURE__ */ new Set();
   }
 }
+function offHarvestHint(criteriaId, evidence) {
+  const anchors = [...new Set(evidence.map((e) => e.selector?.trim() || `${e.file}:${e.line}`).filter(Boolean))];
+  const shown = anchors.slice(0, 4).join(", ");
+  const rest = anchors.length > 4 ? `, +${anchors.length - 4}` : "";
+  return anchors.length ? ` \u2014 this anchor is not among the evidence harvested for ${criteriaId}, so the snippet was verified literally rather than vouched for by the harvest. Cite one of: ${shown}${rest}.` : ` \u2014 ${criteriaId} was harvested no evidence, so nothing can vouch for a transcription here. If the criterion has no subject in scope, the verdict is NA with a justification and no citation to ground.`;
+}
 function tagOf2(x) {
   const fromSnippet = /<\s*([a-zA-Z][\w-]*)/.exec(x.snippet ?? "");
   if (fromSnippet) return fromSnippet[1].toLowerCase();
@@ -59538,8 +59544,8 @@ function applyAdjudication(audit2, adj, opts = {}) {
     }
     return engineNcCache.get(key2) ?? EMPTY_IDS;
   };
-  const toGround = (criteriaId, g, fallback) => {
-    const entry = { g, ...fallback ? { fallback } : {} };
+  const toGround = (criteriaId, g, fallback, offHarvest) => {
+    const entry = { g, ...fallback ? { fallback } : {}, ...offHarvest ? { offHarvest } : {} };
     const list = groundInputs.get(criteriaId);
     if (list) list.push(entry);
     else groundInputs.set(criteriaId, [entry]);
@@ -59584,7 +59590,7 @@ function applyAdjudication(audit2, adj, opts = {}) {
               anchor.representative ? { file: anchor.at.file, line: anchor.at.line, selector: anchor.at.selector ?? c2.selector, snippet: anchor.at.snippet } : { file: c2.file, line: c2.line, selector: anchor.at.selector ?? c2.selector }
             );
           } else {
-            toGround(it.criteriaId, cite);
+            toGround(it.criteriaId, cite, void 0, anchor ? void 0 : offHarvestHint(it.criteriaId, it.evidence));
           }
         }
       }
@@ -59644,7 +59650,7 @@ function applyAdjudication(audit2, adj, opts = {}) {
   }
   const grounding = { grounded: 0, moved: 0, failed: 0, issues: [] };
   for (const [criteriaId, inputs] of groundInputs) {
-    for (const { g, fallback } of inputs) {
+    for (const { g, fallback, offHarvest } of inputs) {
       let r = groundFinding(g, { cwd: opts.cwd });
       if (!r.ok && fallback) {
         const viaAnchor = groundFinding(fallback, { cwd: opts.cwd });
@@ -59656,8 +59662,9 @@ function applyAdjudication(audit2, adj, opts = {}) {
       } else {
         grounding.failed++;
         if (r.issue) {
-          grounding.issues.push(r.issue);
-          blame(criteriaId, r.issue);
+          const issue = offHarvest && r.issue.startsWith("cited snippet not found") ? `${r.issue}${offHarvest}` : r.issue;
+          grounding.issues.push(issue);
+          blame(criteriaId, issue);
         }
       }
     }
@@ -60378,7 +60385,9 @@ var VERDICT_KINDS = `Rule it (the apply gate is FAIL-CLOSED \u2014 a verdict mis
    - \`NC\` (non-conforming) \u2014 REQUIRES \`findings\`: at least one groundable \`{ file, line, selector?, message, snippet?, severity?, normativeRef }\` pointing at REAL source. The fold re-grounds every finding; an invented file:line is rejected, and so is a finding with no \`file\` at all. \`normativeRef\` MUST cite the precise failed test \u2014 under a country standard, one of the criterion's OWN numbered tests, listed in its brief under \xAB tests to rule on \xBB. A WCAG id looks alike, denotes an unrelated test, and is rejected.
    - \`NA\` (not applicable) \u2014 REQUIRES \`justification\`, AND \`citations[]\` whenever evidence WAS presented, to say which of those items fall outside the criterion's scope.
    - \`manual\` (still undecidable) \u2014 REQUIRES \`reason\`: \`needs-rendered-dom\` (only a rendered DOM can decide it, and no capture in this run carries its subject) or \`undecidable\` (the evidence cannot settle it either way).`;
-var ABSENCE_RULE = `AN NC SHAPED LIKE AN ABSENCE IS STILL ANCHORED. \xAB No second navigation system \xBB, \xAB no search engine \xBB, \xAB no error message suggests the expected format \xBB \u2014 an absence is OBSERVED somewhere: cite the element and the page you observed it on. And when the criterion's subject exists nowhere in the audited scope, the verdict is \`NA\` with its justification, never \`NC\`.`;
+var ABSENCE_RULE = `AN NC SHAPED LIKE AN ABSENCE IS STILL ANCHORED. \xAB No second navigation system \xBB, \xAB no search engine \xBB, \xAB no error message suggests the expected format \xBB \u2014 an absence is OBSERVED somewhere: cite the element and the page you observed it on. And when the criterion's subject exists nowhere in the audited scope, the verdict is \`NA\` with its justification, never \`NC\`.
+
+AND CITE FROM THIS CRITERION'S OWN ANCHORS, NOT THE THING YOU ARE RULING OUT. An absence pulls you toward the element you are arguing ABOUT \u2014 the search form that is not a site search engine, the menu that is not a second navigation system \u2014 and that element is, precisely because it is off-topic, absent from what this criterion was harvested. A citation on one of the criterion's own anchors is vouched for by the harvest; one outside it has its snippet verified LITERALLY, character for character, and a retyping then fails. Cite the region you inspected \u2014 the \`header\`, the \`nav\`, the \`footer\` the brief listed \u2014 and say in the justification what you did not find in it.`;
 var CAPTURE_RULE = `THE RENDERED PAGE MAY BE ON DISK. When a criterion's evidence is anchored under \`.ultra11y/pages/<id>/\`, the browser already ran: \`dom.html\`, \`styles.json\`, \`boxes.json\`, \`axtree.json\` and \`screen.png\` are there to read. \`needs-rendered-dom\` is refused on such a criterion \u2014 decide it from those files, or answer \`undecidable\` and say what the capture does not settle.`;
 var NEVER_GUESS_RULE = `Never guess. A criterion you cannot decide from real evidence stays \`manual\` with its reason \u2014 that is a valid, honest verdict, and it is worth more than a verdict the gate throws away.`;
 var SCOPE_RULE = `Rule ONLY on the criteria presented. Never introduce another \u2014 a verdict for a criterion nobody asked about is dropped, and under the fold it would otherwise overwrite what the deterministic engine already decided.`;
@@ -60588,6 +60597,7 @@ function applyRawVerdicts(items, verdicts) {
 var ALLOWED_TOOLS = "Read,Grep,Glob";
 var MAX_ATTEMPTS2 = 4;
 var DEFAULT_CLI_MODEL = "claude-haiku-4-5-20251001";
+var EFFORT_LEVELS = ["low", "medium", "high", "xhigh", "max"];
 var DEFAULT_TIMEOUT_MS2 = 10 * 6e4;
 function realSpawn(argv, input, timeoutMs) {
   return new Promise((resolve17, reject) => {
@@ -60664,6 +60674,7 @@ function cliArgv(opts, items) {
     opts.model ?? DEFAULT_CLI_MODEL
   ];
   if (opts.maxBudgetUsd !== void 0) argv.push("--max-budget-usd", String(opts.maxBudgetUsd));
+  if (opts.effort) argv.push("--effort", opts.effort);
   return argv;
 }
 function verdictsFromText(text) {
@@ -60726,6 +60737,7 @@ async function refuteBatchCli(items, prompt, opts, lang = "en") {
     opts.model ?? DEFAULT_CLI_MODEL
   ];
   if (opts.maxBudgetUsd !== void 0) argv.push("--max-budget-usd", String(opts.maxBudgetUsd));
+  if (opts.effort) argv.push("--effort", opts.effort);
   return runCli2(argv, prompt, opts, (env) => {
     const structured = env.structured_output?.verdicts;
     const raw = Array.isArray(structured) ? structured : verdictsArrayFromText(env.result ?? "");
@@ -68889,7 +68901,7 @@ Usage:
   ultra11y orchestrate --run <dir> [--phase adjudicate|verify-report] [--eco] [--list] [--lang auto|en|fr]
   ultra11y fix      <globs\u2026 | -> [--write] [--iterate] [--changed | --since <ref> | --staged] [--safe] [--include <glob>] [--exclude <glob>] [--ext <list>] [--only <ids>] [--jsx] [--json] [--lang auto|en|fr]
   ultra11y init     [--hook] [--ci] [--baseline] [--fail-on blocking|major|minor]
-  ultra11y judge    --in <audit.json> [--standard <pack>] [--runner api|cli] [--grain batch|criterion] [--max <n>] [--model <id>] [--max-budget-usd <n>] [--timeout <s>] [--out <dir>] [--apply]   (adjudicate the manual criteria with a model \u2014 api needs ANTHROPIC_API_KEY; cli runs the Claude Code CLI and takes CLAUDE_CODE_OAUTH_TOKEN too)
+  ultra11y judge    --in <audit.json> [--standard <pack>] [--runner api|cli] [--grain batch|criterion] [--max <n>] [--model <id>] [--effort <level>] [--max-budget-usd <n>] [--timeout <s>] [--out <dir>] [--apply]   (adjudicate the manual criteria with a model \u2014 api needs ANTHROPIC_API_KEY; cli runs the Claude Code CLI and takes CLAUDE_CODE_OAUTH_TOKEN too)
   ultra11y pack     check <pack.json> [--guidance <g.json>] [--json]  |  pack scaffold
   ultra11y scan     <url|file\u2026> [--runtime auto|local|docker] [--cwd <dir>] [--storage-state <file>] [--no-interact] [--interact-clicks] [--no-snapshot] [--merge <audit.json>] [--out <dir>] [--json]
   ultra11y scan     --sample [--runtime \u2026] [--cwd <dir>] [--storage-state <file>] [--merge <audit.json>] [--json]   (scan the .ultra11yrc.json page sample)
@@ -69122,6 +69134,8 @@ Options:
   --provider <id>    tickets: auto|github|gitlab|jira. 'auto' reads ULTRA11Y_TICKET_PROVIDER,
                      then .ultra11yrc.json, then the git remote (Jira is never auto-detected)
   --grain <mode>     tickets: what one ticket is \u2014 criterion (default) | page | page-criterion | single | file
+  --effort <level>   judge (--runner cli): reasoning effort per invocation \u2014 low | medium |
+                     high | xhigh | max. Unset leaves the CLI's own default
   --transport <t>    tickets: auto|cli|rest (mcp: stdio|http). 'auto' prefers the CLI (gh/glab),
                      falling back to REST when only a token is available. Jira is REST-only
   --max-tickets <n>  tickets: refuse to file more than n tickets in one run (default 200)
@@ -69325,6 +69339,7 @@ var VALUE_FLAGS2 = /* @__PURE__ */ new Set([
   "runner",
   "grain",
   "max-budget-usd",
+  "effort",
   "timeout",
   "concurrency",
   "out",
@@ -71460,6 +71475,15 @@ async function cmdJudge(p) {
     console.error(`ultra11y judge: --grain '${grain}' is not a grain \u2014 expected 'batch' or 'criterion'.`);
     return 2;
   }
+  const effort = typeof p.flags.effort === "string" ? p.flags.effort : void 0;
+  if (effort !== void 0 && !EFFORT_LEVELS.includes(effort)) {
+    console.error(`ultra11y judge: --effort '${effort}' is not a level \u2014 expected ${EFFORT_LEVELS.map((l) => `'${l}'`).join(", ")}.`);
+    return 2;
+  }
+  if (effort !== void 0 && runner !== "cli") {
+    console.error("ultra11y judge: --effort applies to `--runner cli` only \u2014 the Messages API has no session to set an effort on.");
+    return 2;
+  }
   const key2 = typeof p.flags["api-key"] === "string" && p.flags["api-key"] ? p.flags["api-key"] : apiKeyFromEnv();
   if (!key2 && runner === "api") {
     console.error(
@@ -71533,6 +71557,7 @@ async function cmdJudge(p) {
     // requests do. `--concurrency` overrides it for a runner with room.
     concurrency: runner === "cli" ? cliConcurrency : void 0,
     maxBudgetUsd: Number.isFinite(maxBudgetUsd) ? maxBudgetUsd : void 0,
+    effort,
     timeoutMs: Number.isFinite(timeoutMs) ? timeoutMs : void 0,
     onCost: (usd) => {
       spent += usd;

package/skills/ultra11y/SKILL.md

@@ -4,7 +4,7 @@ description: "Use to AUDIT existing HTML/CSS/JSX against WCAG 2.2 AA accessibili
 when_to_use: "Invoke when the user asks for an accessibility AUDIT or a formal deliverable: audit a repo, site or page against WCAG 2.2 AA or a country standard (RGAA, Section 508, EN 301 549); produce a dated conformance report, a criterion grid, a PRD backlog, or tickets (GitHub/GitLab/Jira); author accessible markup from scratch; wire the repo gate (init --hook/--ci) or a rendered-DOM capture pipeline. For a review of the code UNDER CHANGE (diff, branch, PR), use the `review-a11y` skill instead."
 license: MIT
 metadata:
-  version: 5.33.1
+  version: 5.34.1
 ---
 
 # ultra11y — audit WCAG 2.2 AA and write accessible markup

package/skills/ultra11y/scripts/ultra11y.mjs

@@ -6,7 +6,7 @@ import { join as join52, relative as relative5, resolve as resolve16, sep as sep
 import { fileURLToPath as fileURLToPath5, pathToFileURL as pathToFileURL3 } from "url";
 
 // src/types.ts
-var VERSION = "5.33.1";
+var VERSION = "5.34.1";
 var SCHEMA_VERSION = 2;
 var INAPPLICABLE_STATUS = "C";
 
@@ -2293,7 +2293,7 @@ var wcag_universe_default = {
   wcagVersion: "2.2",
   source: "https://www.w3.org/TR/WCAG22/",
   criteriaSource: "https://github.com/w3c/wcag",
-  provenance: "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-24 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
+  provenance: "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-25 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
   criteria: [
     {
       id: "1.1.1",
@@ -59367,6 +59367,12 @@ function auditFiles(audit2, cwd) {
     return /* @__PURE__ */ new Set();
   }
 }
+function offHarvestHint(criteriaId, evidence) {
+  const anchors = [...new Set(evidence.map((e) => e.selector?.trim() || `${e.file}:${e.line}`).filter(Boolean))];
+  const shown = anchors.slice(0, 4).join(", ");
+  const rest = anchors.length > 4 ? `, +${anchors.length - 4}` : "";
+  return anchors.length ? ` \u2014 this anchor is not among the evidence harvested for ${criteriaId}, so the snippet was verified literally rather than vouched for by the harvest. Cite one of: ${shown}${rest}.` : ` \u2014 ${criteriaId} was harvested no evidence, so nothing can vouch for a transcription here. If the criterion has no subject in scope, the verdict is NA with a justification and no citation to ground.`;
+}
 function tagOf2(x) {
   const fromSnippet = /<\s*([a-zA-Z][\w-]*)/.exec(x.snippet ?? "");
   if (fromSnippet) return fromSnippet[1].toLowerCase();
@@ -59538,8 +59544,8 @@ function applyAdjudication(audit2, adj, opts = {}) {
     }
     return engineNcCache.get(key2) ?? EMPTY_IDS;
   };
-  const toGround = (criteriaId, g, fallback) => {
-    const entry = { g, ...fallback ? { fallback } : {} };
+  const toGround = (criteriaId, g, fallback, offHarvest) => {
+    const entry = { g, ...fallback ? { fallback } : {}, ...offHarvest ? { offHarvest } : {} };
     const list = groundInputs.get(criteriaId);
     if (list) list.push(entry);
     else groundInputs.set(criteriaId, [entry]);
@@ -59584,7 +59590,7 @@ function applyAdjudication(audit2, adj, opts = {}) {
               anchor.representative ? { file: anchor.at.file, line: anchor.at.line, selector: anchor.at.selector ?? c2.selector, snippet: anchor.at.snippet } : { file: c2.file, line: c2.line, selector: anchor.at.selector ?? c2.selector }
             );
           } else {
-            toGround(it.criteriaId, cite);
+            toGround(it.criteriaId, cite, void 0, anchor ? void 0 : offHarvestHint(it.criteriaId, it.evidence));
           }
         }
       }
@@ -59644,7 +59650,7 @@ function applyAdjudication(audit2, adj, opts = {}) {
   }
   const grounding = { grounded: 0, moved: 0, failed: 0, issues: [] };
   for (const [criteriaId, inputs] of groundInputs) {
-    for (const { g, fallback } of inputs) {
+    for (const { g, fallback, offHarvest } of inputs) {
       let r = groundFinding(g, { cwd: opts.cwd });
       if (!r.ok && fallback) {
         const viaAnchor = groundFinding(fallback, { cwd: opts.cwd });
@@ -59656,8 +59662,9 @@ function applyAdjudication(audit2, adj, opts = {}) {
       } else {
         grounding.failed++;
         if (r.issue) {
-          grounding.issues.push(r.issue);
-          blame(criteriaId, r.issue);
+          const issue = offHarvest && r.issue.startsWith("cited snippet not found") ? `${r.issue}${offHarvest}` : r.issue;
+          grounding.issues.push(issue);
+          blame(criteriaId, issue);
         }
       }
     }
@@ -60378,7 +60385,9 @@ var VERDICT_KINDS = `Rule it (the apply gate is FAIL-CLOSED \u2014 a verdict mis
    - \`NC\` (non-conforming) \u2014 REQUIRES \`findings\`: at least one groundable \`{ file, line, selector?, message, snippet?, severity?, normativeRef }\` pointing at REAL source. The fold re-grounds every finding; an invented file:line is rejected, and so is a finding with no \`file\` at all. \`normativeRef\` MUST cite the precise failed test \u2014 under a country standard, one of the criterion's OWN numbered tests, listed in its brief under \xAB tests to rule on \xBB. A WCAG id looks alike, denotes an unrelated test, and is rejected.
    - \`NA\` (not applicable) \u2014 REQUIRES \`justification\`, AND \`citations[]\` whenever evidence WAS presented, to say which of those items fall outside the criterion's scope.
    - \`manual\` (still undecidable) \u2014 REQUIRES \`reason\`: \`needs-rendered-dom\` (only a rendered DOM can decide it, and no capture in this run carries its subject) or \`undecidable\` (the evidence cannot settle it either way).`;
-var ABSENCE_RULE = `AN NC SHAPED LIKE AN ABSENCE IS STILL ANCHORED. \xAB No second navigation system \xBB, \xAB no search engine \xBB, \xAB no error message suggests the expected format \xBB \u2014 an absence is OBSERVED somewhere: cite the element and the page you observed it on. And when the criterion's subject exists nowhere in the audited scope, the verdict is \`NA\` with its justification, never \`NC\`.`;
+var ABSENCE_RULE = `AN NC SHAPED LIKE AN ABSENCE IS STILL ANCHORED. \xAB No second navigation system \xBB, \xAB no search engine \xBB, \xAB no error message suggests the expected format \xBB \u2014 an absence is OBSERVED somewhere: cite the element and the page you observed it on. And when the criterion's subject exists nowhere in the audited scope, the verdict is \`NA\` with its justification, never \`NC\`.
+
+AND CITE FROM THIS CRITERION'S OWN ANCHORS, NOT THE THING YOU ARE RULING OUT. An absence pulls you toward the element you are arguing ABOUT \u2014 the search form that is not a site search engine, the menu that is not a second navigation system \u2014 and that element is, precisely because it is off-topic, absent from what this criterion was harvested. A citation on one of the criterion's own anchors is vouched for by the harvest; one outside it has its snippet verified LITERALLY, character for character, and a retyping then fails. Cite the region you inspected \u2014 the \`header\`, the \`nav\`, the \`footer\` the brief listed \u2014 and say in the justification what you did not find in it.`;
 var CAPTURE_RULE = `THE RENDERED PAGE MAY BE ON DISK. When a criterion's evidence is anchored under \`.ultra11y/pages/<id>/\`, the browser already ran: \`dom.html\`, \`styles.json\`, \`boxes.json\`, \`axtree.json\` and \`screen.png\` are there to read. \`needs-rendered-dom\` is refused on such a criterion \u2014 decide it from those files, or answer \`undecidable\` and say what the capture does not settle.`;
 var NEVER_GUESS_RULE = `Never guess. A criterion you cannot decide from real evidence stays \`manual\` with its reason \u2014 that is a valid, honest verdict, and it is worth more than a verdict the gate throws away.`;
 var SCOPE_RULE = `Rule ONLY on the criteria presented. Never introduce another \u2014 a verdict for a criterion nobody asked about is dropped, and under the fold it would otherwise overwrite what the deterministic engine already decided.`;
@@ -60588,6 +60597,7 @@ function applyRawVerdicts(items, verdicts) {
 var ALLOWED_TOOLS = "Read,Grep,Glob";
 var MAX_ATTEMPTS2 = 4;
 var DEFAULT_CLI_MODEL = "claude-haiku-4-5-20251001";
+var EFFORT_LEVELS = ["low", "medium", "high", "xhigh", "max"];
 var DEFAULT_TIMEOUT_MS2 = 10 * 6e4;
 function realSpawn(argv, input, timeoutMs) {
   return new Promise((resolve17, reject) => {
@@ -60664,6 +60674,7 @@ function cliArgv(opts, items) {
     opts.model ?? DEFAULT_CLI_MODEL
   ];
   if (opts.maxBudgetUsd !== void 0) argv.push("--max-budget-usd", String(opts.maxBudgetUsd));
+  if (opts.effort) argv.push("--effort", opts.effort);
   return argv;
 }
 function verdictsFromText(text) {
@@ -60726,6 +60737,7 @@ async function refuteBatchCli(items, prompt, opts, lang = "en") {
     opts.model ?? DEFAULT_CLI_MODEL
   ];
   if (opts.maxBudgetUsd !== void 0) argv.push("--max-budget-usd", String(opts.maxBudgetUsd));
+  if (opts.effort) argv.push("--effort", opts.effort);
   return runCli2(argv, prompt, opts, (env) => {
     const structured = env.structured_output?.verdicts;
     const raw = Array.isArray(structured) ? structured : verdictsArrayFromText(env.result ?? "");
@@ -68889,7 +68901,7 @@ Usage:
   ultra11y orchestrate --run <dir> [--phase adjudicate|verify-report] [--eco] [--list] [--lang auto|en|fr]
   ultra11y fix      <globs\u2026 | -> [--write] [--iterate] [--changed | --since <ref> | --staged] [--safe] [--include <glob>] [--exclude <glob>] [--ext <list>] [--only <ids>] [--jsx] [--json] [--lang auto|en|fr]
   ultra11y init     [--hook] [--ci] [--baseline] [--fail-on blocking|major|minor]
-  ultra11y judge    --in <audit.json> [--standard <pack>] [--runner api|cli] [--grain batch|criterion] [--max <n>] [--model <id>] [--max-budget-usd <n>] [--timeout <s>] [--out <dir>] [--apply]   (adjudicate the manual criteria with a model \u2014 api needs ANTHROPIC_API_KEY; cli runs the Claude Code CLI and takes CLAUDE_CODE_OAUTH_TOKEN too)
+  ultra11y judge    --in <audit.json> [--standard <pack>] [--runner api|cli] [--grain batch|criterion] [--max <n>] [--model <id>] [--effort <level>] [--max-budget-usd <n>] [--timeout <s>] [--out <dir>] [--apply]   (adjudicate the manual criteria with a model \u2014 api needs ANTHROPIC_API_KEY; cli runs the Claude Code CLI and takes CLAUDE_CODE_OAUTH_TOKEN too)
   ultra11y pack     check <pack.json> [--guidance <g.json>] [--json]  |  pack scaffold
   ultra11y scan     <url|file\u2026> [--runtime auto|local|docker] [--cwd <dir>] [--storage-state <file>] [--no-interact] [--interact-clicks] [--no-snapshot] [--merge <audit.json>] [--out <dir>] [--json]
   ultra11y scan     --sample [--runtime \u2026] [--cwd <dir>] [--storage-state <file>] [--merge <audit.json>] [--json]   (scan the .ultra11yrc.json page sample)
@@ -69122,6 +69134,8 @@ Options:
   --provider <id>    tickets: auto|github|gitlab|jira. 'auto' reads ULTRA11Y_TICKET_PROVIDER,
                      then .ultra11yrc.json, then the git remote (Jira is never auto-detected)
   --grain <mode>     tickets: what one ticket is \u2014 criterion (default) | page | page-criterion | single | file
+  --effort <level>   judge (--runner cli): reasoning effort per invocation \u2014 low | medium |
+                     high | xhigh | max. Unset leaves the CLI's own default
   --transport <t>    tickets: auto|cli|rest (mcp: stdio|http). 'auto' prefers the CLI (gh/glab),
                      falling back to REST when only a token is available. Jira is REST-only
   --max-tickets <n>  tickets: refuse to file more than n tickets in one run (default 200)
@@ -69325,6 +69339,7 @@ var VALUE_FLAGS2 = /* @__PURE__ */ new Set([
   "runner",
   "grain",
   "max-budget-usd",
+  "effort",
   "timeout",
   "concurrency",
   "out",
@@ -71460,6 +71475,15 @@ async function cmdJudge(p) {
     console.error(`ultra11y judge: --grain '${grain}' is not a grain \u2014 expected 'batch' or 'criterion'.`);
     return 2;
   }
+  const effort = typeof p.flags.effort === "string" ? p.flags.effort : void 0;
+  if (effort !== void 0 && !EFFORT_LEVELS.includes(effort)) {
+    console.error(`ultra11y judge: --effort '${effort}' is not a level \u2014 expected ${EFFORT_LEVELS.map((l) => `'${l}'`).join(", ")}.`);
+    return 2;
+  }
+  if (effort !== void 0 && runner !== "cli") {
+    console.error("ultra11y judge: --effort applies to `--runner cli` only \u2014 the Messages API has no session to set an effort on.");
+    return 2;
+  }
   const key2 = typeof p.flags["api-key"] === "string" && p.flags["api-key"] ? p.flags["api-key"] : apiKeyFromEnv();
   if (!key2 && runner === "api") {
     console.error(
@@ -71533,6 +71557,7 @@ async function cmdJudge(p) {
     // requests do. `--concurrency` overrides it for a runner with room.
     concurrency: runner === "cli" ? cliConcurrency : void 0,
     maxBudgetUsd: Number.isFinite(maxBudgetUsd) ? maxBudgetUsd : void 0,
+    effort,
     timeoutMs: Number.isFinite(timeoutMs) ? timeoutMs : void 0,
     onCost: (usd) => {
       spent += usd;

package/src/adjudicate.ts

@@ -484,6 +484,22 @@ function auditFiles(audit: AuditResult, cwd?: string): Set<string> {
  *  criterion was never given. And it is only half the gate — `groundFinding` still has to find
  *  the cited content in the real source, so a plausible-looking file:line in the right
  *  neighbourhood proves nothing on its own. */
+/** Why a citation was held to the literal check, and what to cite instead.
+ *
+ *  Appended to a `cited snippet not found` refusal when the citation missed THIS criterion's
+ *  harvest. Without it the message names a symptom — a transcription that did not match — and
+ *  hides the cause, which is that the citation left the evidence the criterion was shown. The
+ *  two are fixed differently: one asks the adjudicator to copy more carefully, the other asks
+ *  it to cite something else entirely. */
+export function offHarvestHint(criteriaId: string, evidence: Evidence[]): string {
+  const anchors = [...new Set(evidence.map((e) => e.selector?.trim() || `${e.file}:${e.line}`).filter(Boolean))];
+  const shown = anchors.slice(0, 4).join(", ");
+  const rest = anchors.length > 4 ? `, +${anchors.length - 4}` : "";
+  return anchors.length
+    ? ` — this anchor is not among the evidence harvested for ${criteriaId}, so the snippet was verified literally rather than vouched for by the harvest. Cite one of: ${shown}${rest}.`
+    : ` — ${criteriaId} was harvested no evidence, so nothing can vouch for a transcription here. If the criterion has no subject in scope, the verdict is NA with a justification and no citation to ground.`;
+}
+
 /** The tag a citation or an anchor is about, from its snippet first and its selector second.
  *  Lowercased — HTML tag names are case-insensitive. Undefined when neither says. */
 function tagOf(x: { snippet?: string; selector?: string }): string | undefined {
@@ -789,7 +805,7 @@ export function applyAdjudication(
   // flat list) for the same reason as `blame`: a failed citation has to condemn its own
   // criterion and no other.
   type Ground = { file: string; line: number; selector?: string; snippet?: string };
-  const groundInputs = new Map<string, { g: Ground; fallback?: Ground }[]>();
+  const groundInputs = new Map<string, { g: Ground; fallback?: Ground; offHarvest?: string }[]>();
   // Memoised: only a citation that missed its criterion's own anchors ever asks for it.
   let scopeCache: Set<string> | undefined;
   const { citationDrift } = adjudicationLimits(opts.cwd);
@@ -819,8 +835,9 @@ export function applyAdjudication(
     criteriaId: string,
     g: { file: string; line: number; selector?: string; snippet?: string },
     fallback?: { file: string; line: number; selector?: string; snippet?: string },
+    offHarvest?: string,
   ) => {
-    const entry = { g, ...(fallback ? { fallback } : {}) };
+    const entry = { g, ...(fallback ? { fallback } : {}), ...(offHarvest ? { offHarvest } : {}) };
     const list = groundInputs.get(criteriaId);
     if (list) list.push(entry);
     else groundInputs.set(criteriaId, [entry]);
@@ -948,7 +965,30 @@ export function applyAdjudication(
                 : { file: c.file, line: c.line, selector: anchor.at.selector ?? c.selector },
             );
           } else {
-            toGround(it.criteriaId, cite);
+            // WHY THE STRICT PATH APPLIED, said at the point it applies.
+            //
+            // The citation belongs (it is in a file this audit read) but it is not on an anchor
+            // THIS criterion was shown, so the paragraph above cannot let the harvest vouch for
+            // the transcription — and the refusal that follows says only `cited snippet not
+            // found`, which reads as « you mistyped » when the actionable fact is « you cited
+            // outside this criterion's evidence, so your retyping was verified literally ».
+            //
+            // Measured: that message sent a reader diagnosing RGAA 12.5 to the wrong conclusion
+            // — a fold bug — with this source open. The risk is not the lost criterion, it is
+            // the « fix » that would follow: relaxing the strict path is exactly the hole the
+            // two bounds above exist to close. So the cause travels with the refusal.
+            //
+            // It bites hardest where the criterion's subject is an ABSENCE. 12.1 and 12.5 are
+            // the repeat offenders the adjudicator prompt already names: asked to prove there is
+            // no search engine, a model cites the search form it is arguing about — which is, by
+            // definition, not among the anchors harvested for a criterion about navigation.
+            // ONLY when the citation is genuinely off-harvest. Landing on an anchor and then
+            // claiming a different KIND of element (an <img> cited as a <video>) also reaches
+            // this branch, and telling that reader « not among the harvested evidence » would
+            // be false: the anchor is right there, it is the element that is wrong. That case
+            // keeps the bare refusal, which already says the honest thing — this markup is not
+            // in the file.
+            toGround(it.criteriaId, cite, undefined, anchor ? undefined : offHarvestHint(it.criteriaId, it.evidence));
           }
         }
       }
@@ -1069,7 +1109,7 @@ export function applyAdjudication(
   // flat `groundItems` produced.
   const grounding: GroundingSummary = { grounded: 0, moved: 0, failed: 0, issues: [] };
   for (const [criteriaId, inputs] of groundInputs) {
-    for (const { g, fallback } of inputs) {
+    for (const { g, fallback, offHarvest } of inputs) {
       let r = groundFinding(g, { cwd: opts.cwd });
       // The harvested anchor vouches for a citation the file could not confirm on its own —
       // an adjudicator's retyping of an element it did read. `moved` either way: what
@@ -1084,8 +1124,12 @@ export function applyAdjudication(
       } else {
         grounding.failed++;
         if (r.issue) {
-          grounding.issues.push(r.issue);
-          blame(criteriaId, r.issue);
+          // The cause travels with the refusal, and only when it IS the cause: a citation that
+          // never reached the literal check (a missing file, a line out of range) is not an
+          // off-harvest problem, and saying so would send the reader chasing the wrong thing.
+          const issue = offHarvest && r.issue.startsWith("cited snippet not found") ? `${r.issue}${offHarvest}` : r.issue;
+          grounding.issues.push(issue);
+          blame(criteriaId, issue);
         }
       }
     }

package/src/agent-cli.ts

@@ -39,6 +39,12 @@ const MAX_ATTEMPTS = 4;
  *  states a model rather than inheriting one. */
 export const DEFAULT_CLI_MODEL = "claude-haiku-4-5-20251001";
 
+/** The levels `claude --effort` accepts. Enumerated here so a caller's typo is refused with a
+ *  message instead of forwarded: this CLI treats an unrecognised value the way it treats an
+ *  unrecognised flag — silently — and a ceiling that is silently nothing is the exact failure
+ *  `--max-turns` already cost this repo once. */
+export const EFFORT_LEVELS = ["low", "medium", "high", "xhigh", "max"];
+
 /** What `claude -p --output-format json` puts on stdout. Only the fields this reads. */
 interface CliEnvelope {
   /** Where `--json-schema` puts the validated object. Present since 2.1.x; `result` carries
@@ -168,6 +174,9 @@ export function cliArgv(opts: LlmOptions, items: AdjudicationItem[]): string[] {
   // a turn budget and be an unbounded run. The bounds that are real: `--max-budget-usd`, which
   // the CLI documents, and the wall-clock kill in `judgeBatchCli`, which is ours.
   if (opts.maxBudgetUsd !== undefined) argv.push("--max-budget-usd", String(opts.maxBudgetUsd));
+  // `--effort` IS documented by this CLI, unlike `--max-turns` above — so it is passed, and a
+  // caller can raise how hard the model thinks without also having to change the model.
+  if (opts.effort) argv.push("--effort", opts.effort);
   return argv;
 }
 
@@ -274,6 +283,9 @@ export async function refuteBatchCli(items: VerifyItem[], prompt: string, opts:
     opts.model ?? DEFAULT_CLI_MODEL,
   ];
   if (opts.maxBudgetUsd !== undefined) argv.push("--max-budget-usd", String(opts.maxBudgetUsd));
+  // `--effort` IS documented by this CLI, unlike `--max-turns` above — so it is passed, and a
+  // caller can raise how hard the model thinks without also having to change the model.
+  if (opts.effort) argv.push("--effort", opts.effort);
   return runCli(argv, prompt, opts, (env) => {
     const structured = (env.structured_output as { verdicts?: unknown } | undefined)?.verdicts;
     const raw = Array.isArray(structured) ? structured : verdictsArrayFromText(env.result ?? "");

package/src/cli.ts

@@ -64,7 +64,7 @@ import {
   type AdjudicationFile,
 } from "./adjudicate.js";
 import { entriesFrom, isLedger, ledgerPath, mergeLedger, readLedger, replayLedger, unreadableCaptures, type VerdictLedger, writeLedger } from "./ledger.js";
-import { DEFAULT_CLI_MODEL, judgeBatchCli, refuteBatchCli } from "./agent-cli.js";
+import { DEFAULT_CLI_MODEL, EFFORT_LEVELS, judgeBatchCli, refuteBatchCli } from "./agent-cli.js";
 import { BATCH_SIZE, apiKeyFromEnv, applyRawVerdicts, judgeAll, modelFromEnv, type LlmOptions } from "./llm.js";
 import { runScan, runScanMany, runCrawlScan, runSampleScan, mergeDynamic, mergeSnapshotAudit, cleanDynamic, dockerAvailable } from "./scan.js";
 import { runScanLocal, runScanManyLocal, runCrawlScanLocal, runSampleScanLocal, localAvailable, localTierStatus } from "./scan-local.js";
@@ -142,7 +142,7 @@ Usage:
   ultra11y orchestrate --run <dir> [--phase adjudicate|verify-report] [--eco] [--list] [--lang auto|en|fr]
   ultra11y fix      <globs… | -> [--write] [--iterate] [--changed | --since <ref> | --staged] [--safe] [--include <glob>] [--exclude <glob>] [--ext <list>] [--only <ids>] [--jsx] [--json] [--lang auto|en|fr]
   ultra11y init     [--hook] [--ci] [--baseline] [--fail-on blocking|major|minor]
-  ultra11y judge    --in <audit.json> [--standard <pack>] [--runner api|cli] [--grain batch|criterion] [--max <n>] [--model <id>] [--max-budget-usd <n>] [--timeout <s>] [--out <dir>] [--apply]   (adjudicate the manual criteria with a model — api needs ANTHROPIC_API_KEY; cli runs the Claude Code CLI and takes CLAUDE_CODE_OAUTH_TOKEN too)
+  ultra11y judge    --in <audit.json> [--standard <pack>] [--runner api|cli] [--grain batch|criterion] [--max <n>] [--model <id>] [--effort <level>] [--max-budget-usd <n>] [--timeout <s>] [--out <dir>] [--apply]   (adjudicate the manual criteria with a model — api needs ANTHROPIC_API_KEY; cli runs the Claude Code CLI and takes CLAUDE_CODE_OAUTH_TOKEN too)
   ultra11y pack     check <pack.json> [--guidance <g.json>] [--json]  |  pack scaffold
   ultra11y scan     <url|file…> [--runtime auto|local|docker] [--cwd <dir>] [--storage-state <file>] [--no-interact] [--interact-clicks] [--no-snapshot] [--merge <audit.json>] [--out <dir>] [--json]
   ultra11y scan     --sample [--runtime …] [--cwd <dir>] [--storage-state <file>] [--merge <audit.json>] [--json]   (scan the .ultra11yrc.json page sample)
@@ -375,6 +375,8 @@ Options:
   --provider <id>    tickets: auto|github|gitlab|jira. 'auto' reads ULTRA11Y_TICKET_PROVIDER,
                      then .ultra11yrc.json, then the git remote (Jira is never auto-detected)
   --grain <mode>     tickets: what one ticket is — criterion (default) | page | page-criterion | single | file
+  --effort <level>   judge (--runner cli): reasoning effort per invocation — low | medium |
+                     high | xhigh | max. Unset leaves the CLI's own default
   --transport <t>    tickets: auto|cli|rest (mcp: stdio|http). 'auto' prefers the CLI (gh/glab),
                      falling back to REST when only a token is available. Jira is REST-only
   --max-tickets <n>  tickets: refuse to file more than n tickets in one run (default 200)
@@ -582,6 +584,7 @@ const VALUE_FLAGS = new Set([
   "runner",
   "grain",
   "max-budget-usd",
+  "effort",
   "timeout",
   "concurrency",
   "out",
@@ -3561,6 +3564,21 @@ async function cmdJudge(p: ParsedArgs): Promise<number> {
     return 2;
   }
 
+  // HOW HARD THE MODEL IS ASKED TO THINK, which on this worklist is not a detail: what reaches
+  // this tier is what no engine could decide. Validated rather than forwarded, because the CLI
+  // SWALLOWS AN UNKNOWN VALUE the same way it swallows an unknown flag — `--effort hgih` would
+  // read as a setting and be nothing at all, which is the failure mode this repo already ate
+  // once with `--max-turns`. Refused here, loudly, while the message can still reach a human.
+  const effort = typeof p.flags.effort === "string" ? (p.flags.effort as string) : undefined;
+  if (effort !== undefined && !EFFORT_LEVELS.includes(effort)) {
+    console.error(`ultra11y judge: --effort '${effort}' is not a level — expected ${EFFORT_LEVELS.map((l) => `'${l}'`).join(", ")}.`);
+    return 2;
+  }
+  if (effort !== undefined && runner !== "cli") {
+    console.error("ultra11y judge: --effort applies to `--runner cli` only — the Messages API has no session to set an effort on.");
+    return 2;
+  }
+
   const key = typeof p.flags["api-key"] === "string" && p.flags["api-key"] ? (p.flags["api-key"] as string) : apiKeyFromEnv();
   if (!key && runner === "api") {
     console.error(
@@ -3684,6 +3702,7 @@ async function cmdJudge(p: ParsedArgs): Promise<number> {
     // requests do. `--concurrency` overrides it for a runner with room.
     concurrency: runner === "cli" ? cliConcurrency : undefined,
     maxBudgetUsd: Number.isFinite(maxBudgetUsd) ? maxBudgetUsd : undefined,
+    effort,
     timeoutMs: Number.isFinite(timeoutMs) ? timeoutMs : undefined,
     onCost: (usd) => {
       spent += usd;

package/src/data/wcag-universe.json

@@ -2,7 +2,7 @@
   "wcagVersion": "2.2",
   "source": "https://www.w3.org/TR/WCAG22/",
   "criteriaSource": "https://github.com/w3c/wcag",
-  "provenance": "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-24 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
+  "provenance": "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-25 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
   "criteria": [
     {
       "id": "1.1.1",

package/src/llm.ts

@@ -71,6 +71,15 @@ export interface LlmOptions {
   /** Dollar ceiling handed to the CLI backend — the bound that survives whatever a turn
    *  happens to cost, which is what a CI job actually wants to cap. */
   maxBudgetUsd?: number;
+  /** Reasoning effort handed to the CLI backend (`--effort`: low, medium, high, xhigh, max).
+   *
+   *  The criteria this tier rules on are the ones no engine can decide — alt relevance, link
+   *  purpose in context, reading order — so how hard the model is asked to think is a real
+   *  lever on the verdict, and one a caller should be able to move without also changing the
+   *  model. Unlike `--max-turns`, `--effort` IS a flag of this CLI, so it is passed rather
+   *  than refused. The Messages backend ignores it: effort is a session notion there, not a
+   *  request parameter. */
+  effort?: string;
   /** Called with each invocation's real cost, so a run can report what it spent instead of
    *  leaving the reader to find it in a provider dashboard. */
   onCost?: (usd: number) => void;

package/src/types.ts

@@ -1,6 +1,6 @@
 // Single source of truth for shared types + the embedded version. sync-version.mjs
 // rewrites VERSION at release time (kept in lockstep with package.json + SKILL.md).
-export const VERSION = "5.33.1";
+export const VERSION = "5.34.1";
 export const SCHEMA_VERSION = 2;
 
 // Lang = the UI FRAME's language — the `L` localization tables each module (report,

package/src/verdict-rules.ts

@@ -27,7 +27,9 @@ export const VERDICT_KINDS = `Rule it (the apply gate is FAIL-CLOSED — a verdi
    - \`manual\` (still undecidable) — REQUIRES \`reason\`: \`needs-rendered-dom\` (only a rendered DOM can decide it, and no capture in this run carries its subject) or \`undecidable\` (the evidence cannot settle it either way).`;
 
 /** An absence is observed somewhere. Measured: 12.1 and 12.5 lost to this on one run. */
-export const ABSENCE_RULE = `AN NC SHAPED LIKE AN ABSENCE IS STILL ANCHORED. « No second navigation system », « no search engine », « no error message suggests the expected format » — an absence is OBSERVED somewhere: cite the element and the page you observed it on. And when the criterion's subject exists nowhere in the audited scope, the verdict is \`NA\` with its justification, never \`NC\`.`;
+export const ABSENCE_RULE = `AN NC SHAPED LIKE AN ABSENCE IS STILL ANCHORED. « No second navigation system », « no search engine », « no error message suggests the expected format » — an absence is OBSERVED somewhere: cite the element and the page you observed it on. And when the criterion's subject exists nowhere in the audited scope, the verdict is \`NA\` with its justification, never \`NC\`.
+
+AND CITE FROM THIS CRITERION'S OWN ANCHORS, NOT THE THING YOU ARE RULING OUT. An absence pulls you toward the element you are arguing ABOUT — the search form that is not a site search engine, the menu that is not a second navigation system — and that element is, precisely because it is off-topic, absent from what this criterion was harvested. A citation on one of the criterion's own anchors is vouched for by the harvest; one outside it has its snippet verified LITERALLY, character for character, and a retyping then fails. Cite the region you inspected — the \`header\`, the \`nav\`, the \`footer\` the brief listed — and say in the justification what you did not find in it.`;
 
 /** `needs-rendered-dom` is a claim about the RUN, not a way out of a hard criterion. */
 export const CAPTURE_RULE = `THE RENDERED PAGE MAY BE ON DISK. When a criterion's evidence is anchored under \`.ultra11y/pages/<id>/\`, the browser already ran: \`dom.html\`, \`styles.json\`, \`boxes.json\`, \`axtree.json\` and \`screen.png\` are there to read. \`needs-rendered-dom\` is refused on such a criterion — decide it from those files, or answer \`undecidable\` and say what the capture does not settle.`;