@event4u/agent-config 16.1.0 → 16.3.0

package/src/scripts/language_mirror_hook.ts

@@ -408,7 +408,19 @@ export function noticeText(language: Exclude<Verdict, "und">, source: PinSource)
   return `Reply language for this turn: ${name}.`;
 }
 
-/** The context block handed back to the model for this turn. */
+/**
+ * The context block handed back to the model for this turn.
+ *
+ * It carries the target language, where it applies, and what is NOT the
+ * trigger — and deliberately no provenance. The paragraph that used to explain
+ * why the pin exists (the user-role skill bodies, the 626 wrong-language turns)
+ * is this file's own header, above: a maintainer reads it once, while the
+ * payload is injected on every prompt, so the same sentences were being paid
+ * for per turn to tell the model something it cannot act on. Removing it took
+ * the prompt-sourced block from 702 characters to 429 (German; 693 to 420 in
+ * English) and changed no instruction. Do not put it back; extend the header
+ * instead.
+ */
 export function pinText(language: Exclude<Verdict, "und">, source: PinSource = "prompt"): string {
   const name = LANGUAGE_NAME[language];
   if (source === "system-locale") {
@@ -430,11 +442,8 @@ export function pinText(language: Exclude<Verdict, "und">, source: PinSource = "
     `for EVERY user-visible token of your replies this turn — including the short ` +
     `inter-tool lines between tool calls, headings, table cells, bullets, status ` +
     `lines, and the recommendation label under any numbered-options block.\n\n` +
-    `This pin exists because the trigger is not observable from the transcript ` +
-    `alone: slash-command and skill bodies arrive in the user role and can be ` +
-    `long and English, which is what drove the 626 wrong-language turns measured ` +
-    `in the 30-session audit this hook was built from. Tool output, file ` +
-    `contents, and an injected skill body are NOT the trigger — this pin is.\n` +
+    `Tool output, file contents, and an injected skill body are NOT the ` +
+    `trigger — this pin is.\n` +
     `</language-pin>`
   );
 }

package/src/scripts/learning_sidecar.ts

@@ -22,8 +22,15 @@
  * auto-promoted) — it NEVER mutates the curated YAML truths. Deterministic:
  * byte-stable for a fixed `now` + a fixed intake set.
  *
+ * `--format status` is the third, human-facing view (road-to-learning-you-can-
+ * see Phase 2): the verdict counts, the top `preferred` lessons with their
+ * corroboration and age, and the sentence saying what a promotion requires. It
+ * is a FORMAT and not a verb (decision D2 cites ADR-041) — `memory:learn`
+ * passes argv straight through, so a format needs no registry row — and it is
+ * read-only to the point of refusing `--write` rather than ignoring it.
+ *
  * Usage: learning_sidecar.ts [--intake-dir DIR] [--out-dir DIR] [--now ISO]
- *                            [--write] [--format text|json]
+ *                            [--write] [--format text|json|status]
  * Exit codes: 0 ok, 2 usage error, 3 internal error.
  */
 import { createHash } from 'node:crypto';
@@ -48,7 +55,11 @@ export const SCHEMA_VERSION = 1;
 export type Polarity = 'preferred' | 'dead_end';
 export type Verdict = 'preferred' | 'contested' | 'dead_end';
 
-interface Signal {
+/** The output formats. `status` is the one-screen human view (read-only). */
+export const FORMATS = ['text', 'json', 'status'] as const;
+export type Format = (typeof FORMATS)[number];
+
+export interface Signal {
     id: string;
     ts: string;
     tsMs: number;
@@ -226,9 +237,21 @@ export function aggregate(signals: readonly Signal[], nowMs: number): Lesson[] {
 }
 
 export function buildSidecar(intakeDir: string, nowIso: string): Sidecar {
+    return buildSidecarFromSignals(readSignals(intakeDir), nowIso);
+}
+
+/**
+ * The half of `buildSidecar` that does not read the disk.
+ *
+ * Split out for a caller that needs BOTH the input count and the output
+ * lessons from one pass — the session-end dogfood ledger records `signals_in`
+ * beside `lessons_out`, and reading the intake twice to get them would make the
+ * two numbers describe two different reads under the 2 s budget they exist to
+ * measure.
+ */
+export function buildSidecarFromSignals(signals: readonly Signal[], nowIso: string): Sidecar {
     const nowMs = _parseTsMs(nowIso) ?? 0;
-    const lessons = aggregate(readSignals(intakeDir), nowMs);
-    return { schema_version: SCHEMA_VERSION, generated_at: nowIso, lessons };
+    return { schema_version: SCHEMA_VERSION, generated_at: nowIso, lessons: aggregate(signals, nowMs) };
 }
 
 /** Render the human dead-ends + contested ledger. */
@@ -261,22 +284,98 @@ export function renderLessonsMd(sidecar: Sidecar): string {
     return L.join('\n');
 }
 
+/** How many `preferred` lessons the status screen lists before it stops. */
+export const STATUS_TOP_N = 5;
+
+/**
+ * The human step. Printed as the LAST line of the status screen, because this
+ * tree writes no rule and no skill on its own: a lesson becomes a durable
+ * artefact only when a person carries it through `/memory:propose` or the
+ * `learning-to-rule-or-skill` skill. A list of preferred lessons is one flag
+ * away from `--apply`, and naming the hand-off on every screen is the cheapest
+ * standing answer to "why is there no button".
+ */
+export const PROMOTION_HANDOFF =
+    'Promote by hand: /memory:propose · src/skills/learning-to-rule-or-skill/SKILL.md';
+
+/**
+ * One screen answering "what has this learned?" — verdict counts, the top
+ * `preferred` lessons with their corroboration and age, and what a promotion
+ * requires. READ ONLY by construction: it takes a built sidecar and returns a
+ * string, so there is no path from this function to the disk.
+ */
+export function renderStatus(sidecar: Sidecar): string {
+    const byVerdict: Record<Verdict, number> = { preferred: 0, contested: 0, dead_end: 0 };
+    for (const l of sidecar.lessons) byVerdict[l.verdict] += 1;
+
+    const nowMs = _parseTsMs(sidecar.generated_at) ?? 0;
+    const L: string[] = [];
+    L.push(`${PROG}: ${String(sidecar.lessons.length)} lesson(s) as of ${sidecar.generated_at}`);
+    L.push(
+        `preferred: ${String(byVerdict.preferred)} · ` +
+            `contested: ${String(byVerdict.contested)} · ` +
+            `dead-end: ${String(byVerdict.dead_end)}`,
+    );
+    L.push('');
+
+    const preferred = sidecar.lessons
+        .filter((l) => l.verdict === 'preferred')
+        .sort((a, b) => b.score - a.score || (a.path < b.path ? -1 : 1))
+        .slice(0, STATUS_TOP_N);
+    L.push(`Top ${String(STATUS_TOP_N)} preferred:`);
+    if (preferred.length === 0) {
+        L.push('  (none — a lesson needs corroboration before it appears here)');
+    } else {
+        preferred.forEach((l, i) => {
+            // Age is floored, so a lesson seen hours ago reads `0d old` rather
+            // than rounding up into a day it has not lived.
+            const ageDays = Math.max(0, Math.floor((nowMs - (_parseTsMs(l.last_ts) ?? nowMs)) / 86400000));
+            L.push(
+                `  ${String(i + 1)}. ${l.entry_type} @ ${l.path} — ` +
+                    `${String(l.corroborations)} origins, ${String(ageDays)}d old`,
+            );
+            L.push(`     ${l.body}`);
+        });
+    }
+    L.push('');
+    L.push(
+        `A lesson is promoted at ${String(MIN_CORROBORATIONS)} distinct origins, ` +
+            `weighted by a ${String(HALF_LIFE_DAYS)}-day half-life; nothing here promotes itself.`,
+    );
+    L.push(PROMOTION_HANDOFF);
+    return `${L.join('\n')}\n`;
+}
+
 export function main(argv: string[]): number {
     let intakeDir = path.join('agents', 'memory', 'intake');
     let outDir = path.join('agents', 'memory');
     let nowIso = '';
     let write = false;
-    let format: 'text' | 'json' = 'text';
+    let format: Format = 'text';
     for (let i = 0; i < argv.length; i++) {
         const a = argv[i] as string;
         if (a === '--intake-dir') intakeDir = argv[++i] ?? intakeDir;
         else if (a === '--out-dir') outDir = argv[++i] ?? outDir;
         else if (a === '--now') nowIso = argv[++i] ?? '';
         else if (a === '--write') write = true;
-        else if (a === '--format') format = (argv[++i] as 'text' | 'json') ?? 'text';
-        else if (a === '-h' || a === '--help') {
+        else if (a === '--format') {
+            const raw = argv[++i] ?? '';
+            // Validated rather than cast. With two formats an unknown value
+            // fell through to `text` and looked like it had worked; with three,
+            // a typo in `status` would silently print the one-line text form
+            // and a reader would take that for the screen they asked for.
+            if (!(FORMATS as readonly string[]).includes(raw)) {
+                process.stderr.write(
+                    `${PROG}: error: argument --format: invalid choice: '${raw}' ` +
+                        `(choose from ${FORMATS.map((f) => `'${f}'`).join(', ')})\n`,
+                );
+                return 2;
+            }
+            format = raw as Format;
+        } else if (a === '-h' || a === '--help') {
             process.stdout.write(
-                `usage: ${PROG} [--intake-dir DIR] [--out-dir DIR] [--now ISO] [--write] [--format text|json]\n`,
+                `usage: ${PROG} [--intake-dir DIR] [--out-dir DIR] [--now ISO] [--write] ` +
+                    `[--format ${FORMATS.join('|')}]\n`,
             );
             return 0;
         } else {
@@ -284,6 +383,14 @@ export function main(argv: string[]): number {
             return 2;
         }
     }
+    if (format === 'status' && write) {
+        // Refused, not ignored. The acceptance criterion is that the screen
+        // leaves `git status` identical; ignoring the flag would honour that
+        // too, but it would do so silently, and a caller who asked for a write
+        // deserves to be told it did not happen.
+        process.stderr.write(`${PROG}: error: --format status is read-only; drop --write\n`);
+        return 2;
+    }
     if (!nowIso) {
         // Deterministic by contract: refuse to guess `now` from the clock when
         // writing, so output is reproducible. Callers pass --now explicitly.
@@ -310,6 +417,10 @@ export function main(argv: string[]): number {
         process.stdout.write(JSON.stringify(sidecar, null, 2) + '\n');
         return 0;
     }
+    if (format === 'status') {
+        process.stdout.write(renderStatus(sidecar));
+        return 0;
+    }
     const byVerdict = { preferred: 0, contested: 0, dead_end: 0 };
     for (const l of sidecar.lessons) byVerdict[l.verdict] += 1;
     process.stdout.write(

package/src/scripts/lint_deny_text.ts

@@ -0,0 +1,419 @@
+#!/usr/bin/env tsx
+/**
+ * A deny never names its own kill switch.
+ *
+ * road-to-a-kernel-that-guards-its-plumbing 1.3.
+ *
+ * A blocking concern refuses a tool call and then prints a message to the one
+ * party that wanted the call to succeed. If that message names the environment
+ * variable, the `export` line or the settings key that turns the guard off, the
+ * refusal is a speed bump with instructions attached. The reader most motivated
+ * to follow them is the agent that was just refused.
+ *
+ * The rule is therefore about the OUTPUT, not about the mechanism: a kill
+ * switch may exist, may be documented in the concern's header, may be listed in
+ * the manifest and in the contract. It may not appear in the text the refusal
+ * prints. Two of the nine blocking concerns carry a kill switch today and both
+ * name it only in a comment — this gate is what keeps that true, rather than
+ * true-so-far.
+ *
+ * WHAT IS SCANNED, stated precisely because the boundary is the whole design.
+ *
+ * The corpus is the `severity: blocking` concerns in `hook_manifest.yaml`, read
+ * from the manifest rather than listed here, so a guard added there joins the
+ * corpus without an edit. Within each, the unit is a MESSAGE-SHAPED string
+ * literal — comments stripped first, then any literal containing whitespace.
+ *
+ * The prose test is the discriminator, and it is what makes the settings-key
+ * rule usable at all: `'hooks.enabled'` as a lookup key is how a concern reads
+ * its own configuration and is not a message, while
+ * `'set hooks.enabled: false to skip'` is. A rule that fired on the bare key
+ * would fire on every concern that reads a setting, which is the false-positive
+ * shape that gets a young gate disabled.
+ *
+ * WHAT IT DOES NOT SEE, so nobody reads more into a green than is there:
+ * a message assembled at runtime from non-literal parts, a message in a module
+ * the concern imports, and a kill switch named in a way none of the three
+ * patterns matches. This is a pattern gate over source text, not a proof about
+ * emitted bytes.
+ *
+ * Exit codes: 0 clean · 1 a deny names a bypass · 2 the corpus could not be read.
+ */
+
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { fileURLToPath, pathToFileURL } from 'node:url';
+
+import * as yaml from 'js-yaml';
+
+import { GateLedger } from './_lib/gate_ledger.js';
+import { runGateCli, runSelfTest } from './_lib/gate_self_test.js';
+import { assertScanned, reportScanned } from './_lib/scan_scope.js';
+
+const _HERE = fileURLToPath(import.meta.url);
+const ROOT = path.resolve(path.dirname(_HERE), '..', '..');
+
+export const GATE = 'lint_deny_text';
+
+const MANIFEST_REL = 'src/scripts/hook_manifest.yaml';
+
+/**
+ * Top-level settings keys, so `hooks.enabled` reads as a settings path and
+ * `payload.tool_input` does not.
+ *
+ * A fixed list rather than a parse of the template: the template is a large
+ * file with placeholder substitutions, and the gate needs the discriminator to
+ * be stable and reviewable in a diff. Adding a top-level settings section is
+ * the moment to add it here; a missing entry costs a miss, never a false
+ * refusal.
+ */
+export const SETTINGS_ROOTS: readonly string[] = [
+    'ai_team',
+    'augment',
+    'chat_history',
+    'code_style',
+    'commands',
+    'consistency',
+    'continuity',
+    'cost',
+    'decision_engine',
+    'delivery',
+    'design',
+    'emergency',
+    'execution',
+    'github',
+    'hooks',
+    'knowledge',
+    'memory',
+    'model',
+    'onboarding',
+    'personal',
+    'pipelines',
+    'planning',
+    'profile',
+    'project',
+    'projection',
+    'quality',
+    'reasoning',
+    'roadmap',
+    'screenshots',
+    'subagents',
+    'telegraph',
+    'tokens',
+    'update_check',
+    'verbosity',
+];
+
+export interface DenyFinding {
+    readonly concern: string;
+    readonly file: string;
+    readonly kind: 'env-assignment' | 'export-line' | 'settings-key';
+    readonly text: string;
+}
+
+/**
+ * Characters after which a `/` begins a REGEX literal rather than a division.
+ *
+ * Regex handling is not a nicety here, it is the difference between a working
+ * gate and a nonsense one — measured on the first run against the real tree.
+ * `turn_end_gate_hook.ts` carries `/\bI (wi)ll\b/i`; without this the
+ * apostrophe inside a pattern like `/\bI'll\b/` opens a phantom string literal
+ * that swallows the next few hundred characters of real code and comments, and
+ * the gate then reports `export function …` as deny text. Three such
+ * false positives, all in one file, all from one unhandled token class.
+ */
+const REGEX_PRECEDERS = new Set('(,=:[!&|?{};+-*%~^<>'.split(''));
+
+/** Strip `//` and block comments, preserving string literals and regexes. */
+export function stripComments(src: string): string {
+    let out = '';
+    let i = 0;
+    let quote: string | null = null;
+    let lastSignificant = '';
+    while (i < src.length) {
+        const c = src[i] as string;
+        const next = src[i + 1];
+        if (quote !== null) {
+            out += c;
+            if (c === '\\') {
+                out += src[i + 1] ?? '';
+                i += 2;
+                continue;
+            }
+            if (c === quote) {
+                quote = null;
+            }
+            i += 1;
+            continue;
+        }
+        if (c === "'" || c === '"' || c === '`') {
+            quote = c;
+            out += c;
+            i += 1;
+            lastSignificant = c;
+            continue;
+        }
+        if (c === '/' && next === '/') {
+            while (i < src.length && src[i] !== '\n') i += 1;
+            continue;
+        }
+        if (c === '/' && next === '*') {
+            i += 2;
+            while (i < src.length && !(src[i] === '*' && src[i + 1] === '/')) i += 1;
+            i += 2;
+            continue;
+        }
+        if (c === '/' && (lastSignificant === '' || REGEX_PRECEDERS.has(lastSignificant))) {
+            // Regex literal — consumed whole, emitted as a placeholder so the
+            // scanner below never sees its contents as a message.
+            i += 1;
+            let inClass = false;
+            while (i < src.length) {
+                const r = src[i] as string;
+                if (r === '\\') {
+                    i += 2;
+                    continue;
+                }
+                if (r === '[') inClass = true;
+                else if (r === ']') inClass = false;
+                else if (r === '/' && !inClass) break;
+                else if (r === '\n') break;
+                i += 1;
+            }
+            i += 1;
+            out += '0';
+            lastSignificant = '0';
+            continue;
+        }
+        out += c;
+        i += 1;
+        if (!/\s/.test(c)) {
+            lastSignificant = c;
+        }
+    }
+    return out;
+}
+
+/** Message-shaped string literals — comments already stripped. */
+export function messageLiterals(src: string): string[] {
+    const out: string[] = [];
+    const re = /'((?:[^'\\]|\\.)*)'|"((?:[^"\\]|\\.)*)"|`((?:[^`\\]|\\.)*)`/g;
+    for (const m of src.matchAll(re)) {
+        const body = m[1] ?? m[2] ?? m[3] ?? '';
+        if (/\s/.test(body)) {
+            out.push(body);
+        }
+    }
+    return out;
+}
+
+const ENV_ASSIGNMENT_RE = /AGENT_CONFIG_[A-Z0-9_]+\s*=/;
+const EXPORT_RE = /(^|[\s`'"(])export\s/;
+
+function settingsKeyIn(text: string): boolean {
+    for (const root of SETTINGS_ROOTS) {
+        const re = new RegExp(`(^|[^A-Za-z0-9_.])${root}\\.[a-z][a-z0-9_]*`);
+        if (re.test(text)) {
+            return true;
+        }
+    }
+    return false;
+}
+
+/** Findings for one concern's source text. */
+export function scanSource(concern: string, file: string, src: string): DenyFinding[] {
+    const out: DenyFinding[] = [];
+    for (const text of messageLiterals(stripComments(src))) {
+        if (ENV_ASSIGNMENT_RE.test(text)) {
+            out.push({ concern, file, kind: 'env-assignment', text });
+            continue;
+        }
+        if (EXPORT_RE.test(text)) {
+            out.push({ concern, file, kind: 'export-line', text });
+            continue;
+        }
+        if (settingsKeyIn(text)) {
+            out.push({ concern, file, kind: 'settings-key', text });
+        }
+    }
+    return out;
+}
+
+/** `severity: blocking` concerns, as `[name, script]`, read from the manifest. */
+export function blockingConcerns(manifestText: string): Array<[string, string]> {
+    const doc = yaml.load(manifestText) as { concerns?: Record<string, Record<string, unknown>> };
+    const concerns = doc?.concerns ?? {};
+    const out: Array<[string, string]> = [];
+    for (const [name, body] of Object.entries(concerns)) {
+        if (String(body['severity'] ?? '') !== 'blocking') continue;
+        const script = body['script'];
+        if (typeof script === 'string') out.push([name, script]);
+    }
+    return out.sort((a, b) => a[0].localeCompare(b[0]));
+}
+
+export function scan(root: string): { findings: DenyFinding[]; scanned: number } {
+    const manifest = path.join(root, MANIFEST_REL);
+    if (!fs.existsSync(manifest)) {
+        return { findings: [], scanned: 0 };
+    }
+    const findings: DenyFinding[] = [];
+    let scanned = 0;
+    for (const [name, script] of blockingConcerns(fs.readFileSync(manifest, 'utf8'))) {
+        const abs = path.join(root, script);
+        if (!fs.existsSync(abs)) continue;
+        scanned += 1;
+        findings.push(...scanSource(name, script, fs.readFileSync(abs, 'utf8')));
+    }
+    return { findings, scanned };
+}
+
+const REASON: Readonly<Record<DenyFinding['kind'], string>> = {
+    'env-assignment': 'names an environment kill switch, with its value',
+    'export-line': 'carries an `export` line the reader can paste',
+    'settings-key': 'names a settings key path',
+};
+
+export function main(argv: readonly string[]): number {
+    if (argv.includes('--self-test')) return selfTest();
+    const rootIdx = argv.indexOf('--root');
+    const root = rootIdx === -1 ? ROOT : (argv[rootIdx + 1] ?? ROOT);
+
+    const { findings, scanned } = scan(root);
+
+    const ledger = new GateLedger(GATE);
+    ledger.plan(findings.map((f) => `${f.concern}:${f.kind}:${f.text.slice(0, 40)}`));
+    for (const f of findings) {
+        ledger.fail(`${f.concern}:${f.kind}:${f.text.slice(0, 40)}`, REASON[f.kind]);
+    }
+
+    // A gate that read nothing has not passed. BEFORE the verdict: a manifest
+    // that moved, or a `severity:` key that was renamed, would otherwise
+    // certify an empty corpus clean.
+    assertScanned({
+        gate: GATE,
+        scanned,
+        units: 'blocking concern(s)',
+        roots: [MANIFEST_REL],
+    });
+    reportScanned({
+        gate: GATE,
+        scanned,
+        units: 'blocking concern(s)',
+        roots: [MANIFEST_REL],
+    });
+    ledger.report();
+
+    if (findings.length > 0) {
+        for (const f of findings) {
+            process.stdout.write(
+                `❌  ${GATE}: ${f.file} (${f.concern}) ${REASON[f.kind]}:\n      ${f.text.slice(0, 200)}\n`,
+            );
+        }
+        process.stdout.write(
+            '\nA blocking concern may HAVE a kill switch. Its refusal message may not name it.\n' +
+                "Move the sentence to the concern's header comment, the manifest entry, or\n" +
+                'docs/contracts/hook-architecture-v1.md — all three are read by people deciding\n' +
+                'whether the guard should exist, and none is read by the agent it just refused.\n',
+        );
+        return 1;
+    }
+
+    process.stdout.write(
+        `✅  ${GATE}: no blocking concern names its own bypass ` +
+            `(${String(scanned)} concern(s) scanned).\n`,
+    );
+    return 0;
+}
+
+function selfTest(): number {
+    const plant = (name: string, body: string): string => {
+        const dir = fs.mkdtempSync(path.join(os.tmpdir(), `deny-text-${name}-`));
+        fs.mkdirSync(path.join(dir, 'src', 'scripts', 'hooks'), { recursive: true });
+        fs.writeFileSync(
+            path.join(dir, 'src', 'scripts', 'hook_manifest.yaml'),
+            [
+                'schema_version: 1',
+                'concerns:',
+                '  probe-guard:',
+                '    script: src/scripts/hooks/probe_guard.ts',
+                '    severity: blocking',
+                '  probe-advisory:',
+                '    script: src/scripts/hooks/probe_advisory.ts',
+                '    severity: advisory',
+                '',
+            ].join('\n'),
+            'utf-8',
+        );
+        fs.writeFileSync(path.join(dir, 'src', 'scripts', 'hooks', 'probe_guard.ts'), body, 'utf-8');
+        fs.writeFileSync(
+            path.join(dir, 'src', 'scripts', 'hooks', 'probe_advisory.ts'),
+            "process.stderr.write('set AGENT_CONFIG_X=1 to skip this advice');\n",
+            'utf-8',
+        );
+        return dir;
+    };
+    const emit = (msg: string): string => `process.stderr.write(${JSON.stringify(msg)});\n`;
+    const run = (root: string): number =>
+        runGateCli(ROOT, path.join('src', 'scripts', 'lint_deny_text.ts'), ['--root', root], ROOT);
+
+    return runSelfTest({
+        gate: GATE,
+        minCases: 7,
+        minRejectCases: 3,
+        cases: [
+            {
+                name: 'an empty corpus is refused, not certified green',
+                expect: 'reject',
+                run: () => run(fs.mkdtempSync(path.join(os.tmpdir(), 'deny-text-empty-'))),
+            },
+            {
+                name: 'a deny naming an env kill switch with its value fails',
+                expect: 'reject',
+                run: () => run(plant('env', emit('BLOCKED — set AGENT_CONFIG_X=1 to skip'))),
+            },
+            {
+                name: 'a deny carrying an export line fails',
+                expect: 'reject',
+                run: () => run(plant('export', emit('BLOCKED — run export AC_OFF=1 first'))),
+            },
+            {
+                name: 'a deny naming a settings key path fails',
+                expect: 'reject',
+                run: () => run(plant('settings', emit('BLOCKED — flip hooks.enabled to false'))),
+            },
+            {
+                name: 'a deny that names no bypass passes',
+                expect: 'accept',
+                run: () => run(plant('clean', emit('BLOCKED — rebuild it from its source'))),
+            },
+            {
+                name: 'the kill switch named in a COMMENT passes — the rule is about output',
+                expect: 'accept',
+                run: () =>
+                    run(
+                        plant(
+                            'comment',
+                            '// Kill switch: AGENT_CONFIG_X=1.\n/* or AGENT_CONFIG_Y=1 */\n' +
+                                emit('BLOCKED — rebuild it from its source'),
+                        ),
+                    ),
+            },
+            {
+                name: 'a bare settings key used as a lookup passes — it is not a message',
+                expect: 'accept',
+                run: () => run(plant('lookup', "const k = 'hooks.enabled';\nvoid k;\n")),
+            },
+            {
+                name: 'an ADVISORY concern naming its switch passes — only denials are scanned',
+                expect: 'accept',
+                run: () => run(plant('advisory-only', emit('BLOCKED — rebuild it'))),
+            },
+        ],
+    });
+}
+
+if (process.argv[1] && import.meta.url === pathToFileURL(path.resolve(process.argv[1])).href) {
+    process.exit(main(process.argv.slice(2)));
+}