@devrik-tools/claude-gates 0.1.2 → 0.2.0

package/plugins/gates/hooks/gates/risk-level/index.mjs

@@ -1,203 +1,263 @@
-// risk-level — denies an implementation delegation that does not declare its risk
-// LEVEL (QUESTION|MICRO|STANDARD|HIGH-RISK), and denies a declared level that
-// contradicts a real high-impact signal in the prompt itself (anything but HIGH-RISK).
-//
-// ── What this does and does not decide ──────────────────────────────────────────────
-// Classifying a task correctly between QUESTION/MICRO/STANDARD stays the delegator's
-// judgment — this gate cannot tell whether STANDARD was the "right" level for a task
-// with no detectable high-impact signal. The only deterministic thing it can and does
-// impose: the declaration EXISTS, and it does not contradict a real, non-quoted signal
-// of high risk. The level vocabulary itself (the four tokens) is a fixed base, not a
-// project param — only which signals count as high-impact, and which subagents are
-// exempt, are configurable.
-//
-// ── Same two-stage intent pattern as intent-flow ────────────────────────────────────
-// Stage 1: cheap lexical prefilter (implementation verb + high-impact signal present
-// anywhere). Stage 2: strip quoted/templated text, require near co-occurrence between
-// the verb and the signal, discard a verb whose object is a documentary deliverable.
-// Over-declaring HIGH-RISK is never penalized — this gate only catches under-declaring.
-
-import { runGate, deny, TOOL_GROUPS } from '../../lib/hook-io.mjs';
-
-const GATE_ID = 'risk-level';
-const CONFIG_KEY = 'requireDeclaredRiskLevel';
-
-const DELEGATION_TOOLS = new Set(TOOL_GROUPS.delegation);
-
-const DEFAULT_READ_ONLY_SUBAGENTS = ['explore', 'claude-code-guide', 'plan'];
-const DEFAULT_HIGH_IMPACT_PATTERNS = [
-  'dinero|pago|cobro|money|payment',
-  'auth|autenticaci[oó]n|authentication|permiso|permission|credencial|credential|token|' +
-    'sesi[oó]n|session|seguridad|security|PII',
-  'borrar|delete|drop|migraci[oó]n|migration|schema',
-  'contrato|contract|irreversible|producci[oó]n|production|API p[uú]blica|public api',
-];
-
-const CO_OCCURRENCE_WINDOW = 100;
-const VERB_OBJECT_WINDOW = 60;
-// How much of an over-length prompt to quote back in a denial message.
-const PROMPT_EXCERPT_LENGTH = 80;
-
-function withUnicodeWordBoundary(alternatives) {
-  return new RegExp(
-    `(?<![\\p{L}\\p{N}_])(${alternatives})(?![\\p{L}\\p{N}_])`,
-    'iu',
-  );
-}
-
-const IMPLEMENTATION_VERBS = withUnicodeWordBoundary(
-  'implementa|implementar|implement(á|é)|agreg(a|á)|agregar|añad(e|í)|añadir|cre(a|á)|crear|' +
-    'arregl(a|á)|arreglar|cambi(a|á)|cambiar|migr(a|á)|migrar|escrib(í|e) c[oó]digo|escribir c[oó]digo|' +
-    'corrige|corregir|correg(í|ir)|constru(ye|í)|construir|modific(a|á)|modificar|' +
-    'refactoriz(a|á)|refactorizar|elimin(a|á)|eliminar|reescrib(e|í)|reescribir|desplieg(a|á)|desplegar|' +
-    'escrib(í|e)|escribir|implement|build|fix|migrate|modify|refactor|remove|rewrite|deploy|write',
-);
-
-const QUERY_VERBS = withUnicodeWordBoundary(
-  'explica|explic(á|ar)|qu[eé]|c[oó]mo|muestra|mostr(á|ar)|analiza|analiz(á|ar)|investiga|' +
-    'investig(á|ar)|revisa|revis(á|ar)|audita|audit(á|ar)|diagnostica|diagnostic(á|ar)|' +
-    'explain|what|how|show|analyze|investigate|review|audit|diagnose',
-);
-
-const DOCUMENTARY_DELIVERABLE_PATTERN = withUnicodeWordBoundary(
-  'documento|documentaci[oó]n|reporte|informe|diagrama|readme|wiki|changelog|' +
-    'p[aá]gina de documentaci[oó]n|archivo html|p[aá]gina html|markdown|document|documentation|report',
-);
-const DOCUMENTARY_EXTENSION_PATTERN = /\.(md|html?|adoc)\b/iu;
-
-/** Declared LEVEL: one of the four fixed tokens, near the word "level"/"classification"
- * (in Spanish or English) so a stray mention elsewhere in the prompt is not mistaken
- * for a declaration. */
-const DECLARED_LEVEL_PATTERN =
-  /(nivel|level|clasificaci[oó]n|classification)[^\n]{0,25}?\b(QUESTION|MICRO|STANDARD|HIGH-RISK)\b/iu;
-
-function isReadOnlySubagent(toolInput, readOnlySubagents) {
-  const type = String(
-    toolInput.subagent_type ?? toolInput.subagentType ?? '',
-  ).toLowerCase();
-  return new Set(readOnlySubagents.map((name) => name.toLowerCase())).has(type);
-}
-
-function isImplementationRequest(prompt) {
-  return IMPLEMENTATION_VERBS.test(prompt);
-}
-
-function isExemptQuery(prompt) {
-  if (isImplementationRequest(prompt)) return false;
-  return QUERY_VERBS.test(prompt);
-}
-
-function stripQuoted(text) {
-  return text
-    .replace(/```[\s\S]*?```/g, ' ')
-    .replace(/"[^"\n]{0,300}"/g, ' ')
-    .replace(/'[^'\n]{0,300}'/g, ' ');
-}
-
-function allMatches(pattern, text) {
-  const flags = pattern.flags.includes('g')
-    ? pattern.flags
-    : `${pattern.flags}g`;
-  return [...text.matchAll(new RegExp(pattern.source, flags))];
-}
-
-function verbProducesDocument(text, verbMatch) {
-  const from = verbMatch.index;
-  const to = from + verbMatch[0].length + VERB_OBJECT_WINDOW;
-  const objectOfVerb = text.slice(from, to);
-  return (
-    DOCUMENTARY_DELIVERABLE_PATTERN.test(objectOfVerb) ||
-    DOCUMENTARY_EXTENSION_PATTERN.test(objectOfVerb)
-  );
-}
-
-/** Returns the matched high-impact signal text if real (non-quoted) intent survives
- * stage 2, else null. Returns null immediately when stage 1 finds no candidate. */
-function realHighImpactSignal(prompt, highImpactPattern) {
-  const isCandidate =
-    IMPLEMENTATION_VERBS.test(prompt) && highImpactPattern.test(prompt);
-  if (!isCandidate) return null;
-
-  const text = stripQuoted(prompt);
-  const verbs = allMatches(IMPLEMENTATION_VERBS, text);
-  const signals = allMatches(highImpactPattern, text);
-
-  for (const verb of verbs) {
-    if (verbProducesDocument(text, verb)) continue;
-    for (const signal of signals) {
-      if (Math.abs(signal.index - verb.index) <= CO_OCCURRENCE_WINDOW)
-        return signal[0];
-    }
-  }
-  return null;
-}
-
-function declaredLevel(prompt) {
-  const match = DECLARED_LEVEL_PATTERN.exec(prompt);
-  return match ? match[2].toUpperCase() : null;
-}
-
-function denyNoLevelDeclared(prompt) {
-  const excerpt = prompt.slice(0, PROMPT_EXCERPT_LENGTH);
-  const ellipsis = prompt.length > PROMPT_EXCERPT_LENGTH ? '…' : '';
-  deny(
-    GATE_ID,
-    `This implementation delegation ("${excerpt}${ellipsis}") does not declare its risk LEVEL. Add a ` +
-      'line such as "LEVEL: STANDARD" (or QUESTION/MICRO/HIGH-RISK, whichever fits) before relaunching ' +
-      'this delegation.',
-  );
-}
-
-function denyLevelContradictsSignal(level, signal) {
-  deny(
-    GATE_ID,
-    `This delegation declares LEVEL: ${level}, but the request touches the risk signal "${signal}" near ` +
-      'an implementation verb (not quoted, not the topic of a documentary deliverable) — that requires ' +
-      `LEVEL: HIGH-RISK, not ${level}. Raise the declaration to HIGH-RISK before relaunching.`,
-  );
-}
-
-function buildHighImpactPattern(highImpactPatterns) {
-  return new RegExp(
-    (highImpactPatterns ?? []).map((source) => `(?:${source})`).join('|'),
-    'iu',
-  );
-}
-
-runGate(
-  {
-    id: GATE_ID,
-    configKey: CONFIG_KEY,
-    enabledByDefault: false,
-    defaultParams: {
-      highImpactPatterns: DEFAULT_HIGH_IMPACT_PATTERNS,
-      readOnlySubagents: DEFAULT_READ_ONLY_SUBAGENTS,
-    },
-  },
-  ({ toolName, toolInput, parameters }) => {
-    if (!DELEGATION_TOOLS.has(toolName)) return;
-
-    const prompt = String(
-      toolInput.prompt ?? toolInput.description ?? toolInput.task ?? '',
-    );
-    if (!prompt.trim()) return;
-    if (isReadOnlySubagent(toolInput, parameters.readOnlySubagents)) return;
-    if (isExemptQuery(prompt)) return;
-    if (!isImplementationRequest(prompt)) return;
-
-    const highImpactPattern = buildHighImpactPattern(
-      parameters.highImpactPatterns,
-    );
-
-    const level = declaredLevel(prompt);
-    const signal = realHighImpactSignal(prompt, highImpactPattern);
-
-    if (!level) {
-      denyNoLevelDeclared(prompt.trim());
-    }
-
-    if (signal && level !== 'HIGH-RISK') {
-      denyLevelContradictsSignal(level, signal);
-    }
-  },
-);
+// risk-level — denies an implementation delegation that does not declare its risk
+// LEVEL (QUESTION|MICRO|STANDARD|HIGH-RISK), and denies a declared level that
+// contradicts a real high-impact signal in the prompt itself (anything but HIGH-RISK).
+//
+// ── What this does and does not decide ──────────────────────────────────────────────
+// Classifying a task correctly between QUESTION/MICRO/STANDARD stays the delegator's
+// judgment — this gate cannot tell whether STANDARD was the "right" level for a task
+// with no detectable high-impact signal. The only deterministic thing it can and does
+// impose: the declaration EXISTS, and it does not contradict a real, non-quoted signal
+// of high risk. The level vocabulary itself (the four tokens) is a fixed base, not a
+// project param — only which signals count as high-impact, and which subagents are
+// exempt, are configurable.
+//
+// ── Same two-stage intent pattern as intent-flow ────────────────────────────────────
+// Stage 1: cheap lexical prefilter (implementation verb + high-impact signal present
+// anywhere). Stage 2: strip quoted/templated text, require near co-occurrence between
+// the verb and the signal, discard a verb whose object is a documentary deliverable.
+// Over-declaring HIGH-RISK is never penalized — this gate only catches under-declaring.
+//
+// ── Which declaration governs ────────────────────────────────────────────────────────
+// A prompt can mention the LEVEL token more than once (a decoy/reference to a previous
+// task, then the real declaration for THIS task). Taking the first match lets an early,
+// irrelevant mention govern instead of the operative one. This gate instead takes the
+// LAST declaration as operative (a delegator revising a decoy note downward writes the
+// real value last); when multiple declarations disagree, it denies and asks for a single
+// unambiguous LEVEL rather than guessing which one is real.
+//
+// ── readOnlySubagents is a declared label, not a verified capability ────────────────
+// This hook cannot check what tools a named subagent actually has — the whitelist
+// exemption is void whenever the prompt itself carries a mutation-risk signal (money/
+// auth/data/write/deploy): the signal in the text outranks the label on the call.
+
+import {
+  runGate,
+  deny,
+  toolInGroups,
+  delegationPromptOf,
+} from '../../lib/hook-io.mjs';
+import { MUTATION_RISK_SIGNAL } from '../../lib/signals.mjs';
+
+const GATE_ID = 'risk-level';
+const CONFIG_KEY = 'requireDeclaredRiskLevel';
+
+const DEFAULT_READ_ONLY_SUBAGENTS = ['explore', 'claude-code-guide', 'plan'];
+const DEFAULT_HIGH_IMPACT_PATTERNS = [
+  'dinero|pago|cobro|money|payment',
+  'auth|autenticaci[oó]n|authentication|permiso|permission|credencial|credential|token|' +
+    'sesi[oó]n|session|seguridad|security|PII',
+  'borrar|delete|drop|migraci[oó]n|migration|schema',
+  'contrato|contract|irreversible|producci[oó]n|production|API p[uú]blica|public api',
+];
+
+const CO_OCCURRENCE_WINDOW = 100;
+const VERB_OBJECT_WINDOW = 60;
+// How much of an over-length prompt to quote back in a denial message.
+const PROMPT_EXCERPT_LENGTH = 80;
+
+function withUnicodeWordBoundary(alternatives) {
+  return new RegExp(
+    `(?<![\\p{L}\\p{N}_])(${alternatives})(?![\\p{L}\\p{N}_])`,
+    'iu',
+  );
+}
+
+// A prompt-level signal that a whitelisted read-only subagent name should NOT be
+// trusted to exempt this call: the name is a declared label, never a verified
+// capability this hook can check, and real mutation risk in the text must win over it.
+// Centralized in lib/signals.mjs (ES+EN) — see its header for the class this covers.
+const MUTATION_RISK_SIGNAL_PATTERN = MUTATION_RISK_SIGNAL;
+
+const IMPLEMENTATION_VERBS = withUnicodeWordBoundary(
+  'implementa|implementar|implement(á|é)|agreg(a|á)|agregar|añad(e|í)|añadir|cre(a|á)|crear|' +
+    'arregl(a|á)|arreglar|cambi(a|á)|cambiar|migr(a|á)|migrar|escrib(í|e) c[oó]digo|escribir c[oó]digo|' +
+    'corrige|corregir|correg(í|ir)|constru(ye|í)|construir|modific(a|á)|modificar|' +
+    'refactoriz(a|á)|refactorizar|elimin(a|á)|eliminar|reescrib(e|í)|reescribir|desplieg(a|á)|desplegar|' +
+    'escrib(í|e)|escribir|implement|build|fix|migrate|modify|refactor|remove|rewrite|deploy|write',
+);
+
+const QUERY_VERBS = withUnicodeWordBoundary(
+  'explica|explic(á|ar)|qu[eé]|c[oó]mo|muestra|mostr(á|ar)|analiza|analiz(á|ar)|investiga|' +
+    'investig(á|ar)|revisa|revis(á|ar)|audita|audit(á|ar)|diagnostica|diagnostic(á|ar)|' +
+    'explain|what|how|show|analyze|investigate|review|audit|diagnose',
+);
+
+const DOCUMENTARY_DELIVERABLE_PATTERN = withUnicodeWordBoundary(
+  'documento|documentaci[oó]n|reporte|informe|diagrama|readme|wiki|changelog|' +
+    'p[aá]gina de documentaci[oó]n|archivo html|p[aá]gina html|markdown|document|documentation|report',
+);
+const DOCUMENTARY_EXTENSION_PATTERN = /\.(md|html?|adoc)\b/iu;
+
+/** Declared LEVEL: one of the four fixed tokens, near the word "level"/"classification"
+ * (in Spanish or English) so a stray mention elsewhere in the prompt is not mistaken
+ * for a declaration. */
+const DECLARED_LEVEL_PATTERN =
+  /(nivel|level|clasificaci[oó]n|classification)[^\n]{0,25}?\b(QUESTION|MICRO|STANDARD|HIGH-RISK)\b/iu;
+
+function isReadOnlySubagentName(toolInput, readOnlySubagents) {
+  const type = String(
+    toolInput.subagent_type ?? toolInput.subagentType ?? '',
+  ).toLowerCase();
+  return new Set(readOnlySubagents.map((name) => name.toLowerCase())).has(type);
+}
+
+/** A whitelisted subagent name exempts a call ONLY when the prompt carries no
+ * mutation-risk signal. The name is a declared label, never a verified capability this
+ * hook can check — a real risk signal in the text must win over it. */
+function isReadOnlySubagent(toolInput, prompt, readOnlySubagents) {
+  if (!isReadOnlySubagentName(toolInput, readOnlySubagents)) return false;
+  return !MUTATION_RISK_SIGNAL_PATTERN.test(prompt);
+}
+
+function isImplementationRequest(prompt) {
+  return IMPLEMENTATION_VERBS.test(prompt);
+}
+
+function isExemptQuery(prompt) {
+  if (isImplementationRequest(prompt)) return false;
+  return QUERY_VERBS.test(prompt);
+}
+
+function stripQuoted(text) {
+  return text
+    .replace(/```[\s\S]*?```/g, ' ')
+    .replace(/"[^"\n]{0,300}"/g, ' ')
+    .replace(/'[^'\n]{0,300}'/g, ' ');
+}
+
+function allMatches(pattern, text) {
+  const flags = pattern.flags.includes('g')
+    ? pattern.flags
+    : `${pattern.flags}g`;
+  return [...text.matchAll(new RegExp(pattern.source, flags))];
+}
+
+function verbProducesDocument(text, verbMatch) {
+  const from = verbMatch.index;
+  const to = from + verbMatch[0].length + VERB_OBJECT_WINDOW;
+  const objectOfVerb = text.slice(from, to);
+  return (
+    DOCUMENTARY_DELIVERABLE_PATTERN.test(objectOfVerb) ||
+    DOCUMENTARY_EXTENSION_PATTERN.test(objectOfVerb)
+  );
+}
+
+/** Returns the matched high-impact signal text if real (non-quoted) intent survives
+ * stage 2, else null. Returns null immediately when stage 1 finds no candidate. */
+function realHighImpactSignal(prompt, highImpactPattern) {
+  const isCandidate =
+    IMPLEMENTATION_VERBS.test(prompt) && highImpactPattern.test(prompt);
+  if (!isCandidate) return null;
+
+  const text = stripQuoted(prompt);
+  const verbs = allMatches(IMPLEMENTATION_VERBS, text);
+  const signals = allMatches(highImpactPattern, text);
+
+  for (const verb of verbs) {
+    if (verbProducesDocument(text, verb)) continue;
+    for (const signal of signals) {
+      if (Math.abs(signal.index - verb.index) <= CO_OCCURRENCE_WINDOW)
+        return signal[0];
+    }
+  }
+  return null;
+}
+
+/** Every LEVEL declaration in the prompt, in order of appearance. */
+function declaredLevels(prompt) {
+  const withGlobal = new RegExp(
+    DECLARED_LEVEL_PATTERN.source,
+    `${DECLARED_LEVEL_PATTERN.flags}g`,
+  );
+  return [...prompt.matchAll(withGlobal)].map((match) => match[2].toUpperCase());
+}
+
+/** The operative LEVEL: the LAST declaration in the prompt (a delegator who corrects an
+ * earlier decoy/reference mention writes the real value last). Returns `{ level }` when
+ * every declaration agrees or there is exactly one; returns `{ ambiguous: true }` when
+ * two or more DIFFERENT levels are declared — the gate cannot know which one governs,
+ * so it asks for a single unambiguous LEVEL rather than silently picking one. */
+function operativeLevel(prompt) {
+  const levels = declaredLevels(prompt);
+  if (levels.length === 0) return { level: null };
+  const distinct = new Set(levels);
+  if (distinct.size > 1) return { ambiguous: true, levels };
+  return { level: levels[levels.length - 1] };
+}
+
+function denyNoLevelDeclared(prompt) {
+  const excerpt = prompt.slice(0, PROMPT_EXCERPT_LENGTH);
+  const ellipsis = prompt.length > PROMPT_EXCERPT_LENGTH ? '…' : '';
+  deny(
+    GATE_ID,
+    `This implementation delegation ("${excerpt}${ellipsis}") does not declare its risk LEVEL. Add a ` +
+      'line such as "LEVEL: STANDARD" (or QUESTION/MICRO/HIGH-RISK, whichever fits) before relaunching ' +
+      'this delegation.',
+  );
+}
+
+function denyAmbiguousLevel(levels) {
+  deny(
+    GATE_ID,
+    `This delegation declares multiple different risk levels (${[...new Set(levels)].join(', ')}) — ` +
+      'it is not clear which one governs this task. Declare a single, unambiguous LEVEL for this ' +
+      'delegation (remove any decoy/reference mention of a different level) before relaunching.',
+  );
+}
+
+function denyLevelContradictsSignal(level, signal) {
+  deny(
+    GATE_ID,
+    `This delegation declares LEVEL: ${level}, but the request touches the risk signal "${signal}" near ` +
+      'an implementation verb (not quoted, not the topic of a documentary deliverable) — that requires ' +
+      `LEVEL: HIGH-RISK, not ${level}. Raise the declaration to HIGH-RISK before relaunching.`,
+  );
+}
+
+function buildHighImpactPattern(highImpactPatterns) {
+  return new RegExp(
+    (highImpactPatterns ?? []).map((source) => `(?:${source})`).join('|'),
+    'iu',
+  );
+}
+
+runGate(
+  {
+    id: GATE_ID,
+    configKey: CONFIG_KEY,
+    enabledByDefault: false,
+    defaultParams: {
+      highImpactPatterns: DEFAULT_HIGH_IMPACT_PATTERNS,
+      readOnlySubagents: DEFAULT_READ_ONLY_SUBAGENTS,
+    },
+  },
+  ({ toolName, toolInput, parameters }) => {
+    if (!toolInGroups(toolName, ['delegation'])) return;
+
+    const prompt = delegationPromptOf(toolInput);
+    if (!prompt.trim()) return;
+    if (isReadOnlySubagent(toolInput, prompt, parameters.readOnlySubagents))
+      return;
+    if (isExemptQuery(prompt)) return;
+    if (!isImplementationRequest(prompt)) return;
+
+    const highImpactPattern = buildHighImpactPattern(
+      parameters.highImpactPatterns,
+    );
+
+    const { level, ambiguous, levels } = operativeLevel(prompt);
+    const signal = realHighImpactSignal(prompt, highImpactPattern);
+
+    if (ambiguous) {
+      denyAmbiguousLevel(levels);
+    }
+
+    if (!level) {
+      denyNoLevelDeclared(prompt.trim());
+    }
+
+    if (signal && level !== 'HIGH-RISK') {
+      denyLevelContradictsSignal(level, signal);
+    }
+  },
+);

package/plugins/gates/hooks/gates/risk-level/risk-level.edge.test.mjs

@@ -0,0 +1,116 @@
+// Edge-case audit for risk-level. Each test demonstrates a confirmed BUG or an OK.
+// Run: node --test risk-level.edge.test.mjs
+import assert from 'node:assert/strict';
+import { execFileSync } from 'node:child_process';
+import { mkdtempSync, mkdirSync, writeFileSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { dirname, join } from 'node:path';
+import { test } from 'node:test';
+import { fileURLToPath } from 'node:url';
+
+const GATE = join(dirname(fileURLToPath(import.meta.url)), 'index.mjs');
+
+function runGate(payload, { config } = {}) {
+  const project = mkdtempSync(join(tmpdir(), 'risk-level-edge-'));
+  mkdirSync(join(project, '.git'));
+  if (config) {
+    mkdirSync(join(project, '.ai'));
+    writeFileSync(join(project, '.ai', 'config.json'), JSON.stringify(config));
+  }
+  const out = execFileSync(process.execPath, [GATE], {
+    input: JSON.stringify(payload),
+    encoding: 'utf8',
+    cwd: project,
+    env: { ...process.env, HOME: project, USERPROFILE: project },
+  });
+  return out.trim() ? JSON.parse(out.trim()) : null;
+}
+
+function isDeny(result) {
+  return result?.hookSpecificOutput?.permissionDecision === 'deny';
+}
+
+const ENABLED = { config: { gates: { requireDeclaredRiskLevel: true } } };
+
+const SENSITIVE_NO_LEVEL =
+  'Implementa el cobro del pago con la nueva pasarela de dinero para el checkout.';
+
+test('FIXED: a delegation tool name outside the native list is now caught via toolInGroups (MCP delegation signal)', () => {
+  const result = runGate(
+    { tool_name: 'mcp__orchestrator__spawn_agent', tool_input: { prompt: SENSITIVE_NO_LEVEL } },
+    ENABLED,
+  );
+  assert.ok(isDeny(result), 'gate now denies the no-level prompt under the MCP tool name too');
+});
+
+test('FIXED: prompt carried in a field other than prompt/description/task is now read via delegationPromptOf', () => {
+  const result = runGate(
+    { tool_name: 'Agent', tool_input: { instructions: SENSITIVE_NO_LEVEL } },
+    ENABLED,
+  );
+  assert.ok(isDeny(result), 'gate now sees the prompt under "instructions" and denies it for missing LEVEL');
+});
+
+test('FIXED: a mutation-risk signal in the prompt overrides a whitelisted read-only subagent name', () => {
+  const result = runGate(
+    { tool_name: 'Agent', tool_input: { prompt: SENSITIVE_NO_LEVEL, subagent_type: 'plan' } },
+    ENABLED,
+  );
+  assert.ok(
+    isDeny(result),
+    'gate no longer exempts a real money-mutation delegation solely because subagent_type says "plan"',
+  );
+});
+
+test('FIXED: the LAST level declaration governs, not the first (a decoy early mention no longer wins)', () => {
+  // Previously DECLARED_LEVEL_PATTERN.exec(prompt) (no /g) took only the first match, so
+  // a decoy HIGH-RISK mention near the top let a real MICRO-declared money mutation
+  // through. operativeLevel() now takes the LAST declaration as operative: MICRO governs
+  // here and the real money-mutation signal contradicts it, so the gate denies.
+  const prompt =
+    'NIVEL: HIGH-RISK (nota: eso era para la tarea anterior, ignorar). Para esta tarea: NIVEL: MICRO. ' +
+    'Implementa el cobro del pago con la nueva pasarela de dinero para el checkout.';
+  const result = runGate({ tool_name: 'Agent', tool_input: { prompt } }, ENABLED);
+  assert.ok(
+    isDeny(result),
+    'gate now denies: the operative (last) declaration is MICRO, which contradicts the real money-mutation signal',
+  );
+});
+
+test('FIXED: two DIFFERENT levels that both look operative are denied as ambiguous, not silently resolved', () => {
+  const prompt =
+    'NIVEL: MICRO. Mas adelante, NIVEL: HIGH-RISK. Arregla el boton que no cambia de color en la pagina.';
+  const result = runGate({ tool_name: 'Agent', tool_input: { prompt } }, ENABLED);
+  assert.ok(
+    isDeny(result),
+    'gate denies and asks for a single unambiguous LEVEL when multiple different levels are declared',
+  );
+});
+
+test('OK: repeating the SAME level twice is not treated as ambiguous', () => {
+  const result = runGate(
+    {
+      tool_name: 'Agent',
+      tool_input: {
+        prompt:
+          'NIVEL: STANDARD. Como se declaro, NIVEL: STANDARD. Arregla el boton que no cambia de color.',
+      },
+    },
+    ENABLED,
+  );
+  assert.equal(result, null);
+});
+
+test('OK: same sensitive prompt without a read-only subagent name and no level is denied', () => {
+  const result = runGate({ tool_name: 'Agent', tool_input: { prompt: SENSITIVE_NO_LEVEL } }, ENABLED);
+  assert.ok(isDeny(result));
+});
+
+test('OK: Task and invoke_subagent tool names are both covered', () => {
+  assert.ok(isDeny(runGate({ tool_name: 'Task', tool_input: { prompt: SENSITIVE_NO_LEVEL } }, ENABLED)));
+  assert.ok(
+    isDeny(
+      runGate({ tool_name: 'invoke_subagent', tool_input: { prompt: SENSITIVE_NO_LEVEL } }, ENABLED),
+    ),
+  );
+});