@event4u/agent-config 6.1.0 → 7.0.1

package/src/scripts/check_memory.ts

@@ -0,0 +1,845 @@
+#!/usr/bin/env tsx
+/**
+ * Engineering Memory validator.
+ *
+ * TypeScript twin of `src/scripts/check_memory.py` (ADR-200, Phase 4 /
+ * Wave 4a). The CLI contract mirrors the Python original EXACTLY — same
+ * flags (`--path`, `--format`, `--append-only`, `--base`), same exit
+ * codes, same stdout/stderr split,
+ * byte-identical finding messages, same scan scope and ordering. No
+ * behaviour changes — latent bugs are replicated and flagged in the
+ * porting report, not fixed.
+ *
+ * Validates YAML files under `agents/memory/<type>/**\/*.yml` against the
+ * schema documented in `guidelines/agent-infra/engineering-memory-data-format`.
+ *
+ * Checks:
+ *   * Required shared frontmatter: id, status, confidence, source, owner,
+ *     last_validated, review_after_days.
+ *   * Duplicate `id` within the same type.
+ *   * Basic redaction: obvious secrets, private URLs with credentials,
+ *     IP addresses tied to internal ranges.
+ *   * Staleness: entries where (today - last_validated) > review_after_days
+ *     are reported (informational, never hard fail).
+ *
+ *   * Append-only (--append-only): inspects `git diff` against a ref to
+ *     ensure intake JSONL files (`agents/memory/intake/*.jsonl`) only gained
+ *     lines at EOF. In-place edits, deletions, or reorderings fail the check.
+ *     See `road-to-memory-merge-safety.md` Phase 0.
+ *
+ * Exit codes: 0 = clean, 1 = violations, 2 = PyYAML missing, 3 = internal error.
+ *
+ * Usage:
+ *     check_memory                     # validate templates + agents/memory
+ *     check_memory --path agents/memory
+ *     check_memory --format json
+ *     check_memory --append-only       # CI: diff vs origin/main
+ *     check_memory --append-only --base HEAD~1
+ *
+ * Note on YAML date typing: the Python original relies on PyYAML's
+ * `safe_load` converting unquoted `YYYY-MM-DD` scalars to `datetime.date`
+ * (and `YYYY-MM-DD HH:MM:SS` to `datetime.datetime`). The `yaml` npm
+ * package parses every scalar to a string, so this twin re-implements
+ * PyYAML's timestamp implicit-resolver to convert matching *plain*
+ * (unquoted) scalars into a `PyDate` marker — preserving the
+ * `isinstance(value, datetime.date)` semantics the staleness checks rely
+ * on. Quoted dates stay strings, exactly as in PyYAML.
+ */
+
+import { spawnSync } from 'node:child_process';
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import process from 'node:process';
+import { pathToFileURL } from 'node:url';
+import YAML, { parseDocument } from 'yaml';
+
+type Severity = 'error' | 'warning' | 'info';
+
+const REQUIRED_KEYS: ReadonlySet<string> = new Set([
+    'id',
+    'status',
+    'confidence',
+    'source',
+    'owner',
+    'last_validated',
+    'review_after_days',
+]);
+const VALID_STATUS: ReadonlySet<string> = new Set(['active', 'deprecated', 'archived', 'superseded']);
+const VALID_CONFIDENCE: ReadonlySet<string> = new Set(['low', 'medium', 'high']);
+// `priority` is optional (default `normal`); enum is the smallest set that
+// solves the tier-0 surfacing use case. See the Phase 2 council brief for why
+// the `high` tier was rejected.
+const VALID_PRIORITY: ReadonlySet<string> = new Set(['critical', 'normal', 'low']);
+// Soft-cap on `priority: critical` entries per memory type. Tier-0 inflation
+// is the failure mode: when too many entries claim "always surface", the
+// slice loses signal. Warn (not fail) when the cap is exceeded so curators
+// notice without being blocked.
+const CRITICAL_WARN_THRESHOLD = 10;
+// Stale-critical guard: a `priority: critical` entry that hasn't been
+// re-validated in this many days emits a warning. Surfaced separately
+// from the generic `stale:` info so reviewers see it before merge.
+const CRITICAL_STALE_DAYS = 90;
+const KNOWN_TYPES: ReadonlySet<string> = new Set([
+    'domain-invariants',
+    'incident-learnings',
+    'product-rules',
+]);
+
+// Per-type soft entry cap (size-bounding without a decay engine). Over-cap →
+// warning, never a hard fail: the right answer to bloat is a consolidation pass
+// (prune archived, merge duplicates), not CI failure. See
+// road-to-memory-pipeline-consolidation.md Phase 7.
+const PER_TYPE_CAPS: Readonly<Record<string, number>> = {
+    ownership: 50,
+    'domain-invariants': 150,
+    'product-rules': 100,
+    'incident-learnings': 150,
+    'historical-patterns': 150,
+};
+const DEFAULT_TYPE_CAP = 150;
+// One-durable-fact-per-entry: a content field longer than this reads as a
+// transcript / narrative blob, not a single durable fact → warning.
+const ONE_FACT_MAX_CHARS = 600;
+const ONE_FACT_FIELDS = ['rule', 'pattern', 'statement', 'observation', 'body', 'decision', 'note'] as const;
+// Per-type entry tally, populated during validation, consumed by main().
+const _TYPE_COUNTS: Record<string, number> = {};
+
+// Redaction heuristics — plain-regex, deliberately conservative.
+// False positives are fixed by quoting the line differently; false
+// negatives are a curator responsibility.
+const REDACTION_PATTERNS: ReadonlyArray<readonly [RegExp, string]> = [
+    // Key=value secret with a clear credential-word prefix. The value must
+    // be a single token (no spaces, no "/" — skips filepaths and URLs).
+    [
+        /\b(api[_-]?key|auth[_-]?token|access[_-]?token|bearer|secret|password|passwd|private[_-]?key)\b\s*[:=]\s*[A-Za-z0-9+/=_\-]{8,}(?![/.\w])/i,
+        'inline credential',
+    ],
+    [/https?:\/\/[^\s"'/]*:[^\s"'/]*@/, 'url with credentials'],
+    [/\b10\.\d{1,3}\.\d{1,3}\.\d{1,3}\b/, 'internal ipv4 range'],
+    [/\b192\.168\.\d{1,3}\.\d{1,3}\b/, 'internal ipv4 range'],
+];
+
+// Date-discipline — relative-date phrases without an ISO YYYY-MM-DD anchor
+// within ±20 chars are rejected. Memory entries that say "yesterday" or
+// "last week" rot the moment the file is re-read on another day; the
+// anchor pins meaning.
+const RELATIVE_DATE_PATTERN =
+    /\b(yesterday|today|tomorrow|last\s+(?:week|month|year)|next\s+(?:week|month|year)|this\s+(?:week|month|year))\b/gi;
+const ISO_DATE_PATTERN = /\b\d{4}-\d{2}-\d{2}\b/;
+const DATE_ANCHOR_WINDOW = 20;
+
+// PyYAML timestamp implicit-resolver. A *plain* (unquoted) scalar matching
+// this is constructed as datetime.date / datetime.datetime — both pass
+// `isinstance(value, datetime.date)`. Source: yaml/resolver.py.
+const PYYAML_TIMESTAMP_RE =
+    /^(?:[0-9][0-9][0-9][0-9]-[0-9][0-9]-[0-9][0-9]|[0-9][0-9][0-9][0-9]-[0-9][0-9]?-[0-9][0-9]?(?:[Tt]|[ \t]+)[0-9][0-9]?:[0-9][0-9]:[0-9][0-9](?:\.[0-9]*)?(?:[ \t]*(?:Z|[-+][0-9][0-9]?(?::[0-9][0-9])?))?)$/;
+const PYYAML_DATE_ONLY_RE = /^[0-9]{4}-[0-9]{2}-[0-9]{2}$/;
+
+class PyDate {
+    constructor(
+        readonly year: number,
+        readonly month: number,
+        readonly day: number,
+    ) {}
+}
+
+class Finding {
+    constructor(
+        readonly file: string,
+        readonly line: number,
+        readonly severity: Severity,
+        readonly message: string,
+        readonly entry_id: string = '',
+    ) {}
+}
+
+type EntryValue = unknown;
+type Entry = Record<string, EntryValue>;
+
+/**
+ * Python-compatible JSON serializer.
+ *
+ * `json.dumps(..., indent=2)` uses item separator `,` + newline and key
+ * separator `": "`, and escapes every non-ASCII character (`ensure_ascii=True`
+ * default). `json.dumps(obj)` (no indent) uses `", "` / `": "`. JS's
+ * `JSON.stringify` does neither, so this reproduces both.
+ */
+function pyJsonDumps(value: unknown, indent: number | null): string {
+    const raw = indent === null ? JSON.stringify(value) : JSON.stringify(value, null, indent);
+    // Reproduce the default-separator form for the non-indented path:
+    // Python emits `", "` between items and `": "` after keys. JS already
+    // emits `,` / `:` with no spaces. Re-introduce the spaces only for the
+    // non-indented case (the indented case uses newlines + indentation).
+    let out = raw;
+    if (indent === null) {
+        out = pyCompactSeparators(raw);
+    }
+    return escapeNonAscii(out);
+}
+
+/** Re-introduce Python's `", "` / `": "` separators into a compact JSON string. */
+function pyCompactSeparators(json: string): string {
+    let result = '';
+    let inString = false;
+    let escaped = false;
+    for (let i = 0; i < json.length; i += 1) {
+        const ch = json[i] as string;
+        result += ch;
+        if (escaped) {
+            escaped = false;
+            continue;
+        }
+        if (ch === '\\') {
+            escaped = true;
+            continue;
+        }
+        if (ch === '"') {
+            inString = !inString;
+            continue;
+        }
+        if (!inString && (ch === ',' || ch === ':')) {
+            result += ' ';
+        }
+    }
+    return result;
+}
+
+/** Escape every non-ASCII char as \uXXXX, mirroring json.dumps ensure_ascii. */
+function escapeNonAscii(s: string): string {
+    let out = '';
+    for (const ch of s) {
+        const code = ch.codePointAt(0) ?? 0;
+        if (code > 0x7f) {
+            // Emit UTF-16 code units (surrogate pairs for astral chars), as
+            // Python's json does.
+            for (let i = 0; i < ch.length; i += 1) {
+                out += `\\u${ch.charCodeAt(i).toString(16).padStart(4, '0')}`;
+            }
+        } else {
+            out += ch;
+        }
+    }
+    return out;
+}
+
+/**
+ * Parse a YAML document the way PyYAML's `safe_load` would, for the subset
+ * of typing the validator inspects: unquoted timestamp scalars become
+ * `PyDate`. Everything else falls through to the `yaml` core schema.
+ * Throws on parse error (mirrors PyYAML raising YAMLError).
+ */
+function pyYamlSafeLoad(text: string): unknown {
+    const doc = parseDocument(text, { prettyErrors: false });
+    if (doc.errors.length > 0) {
+        const err = doc.errors[0];
+        throw new Error(err ? err.message : 'YAML parse error');
+    }
+    YAML.visit(doc, {
+        Scalar(_key, node) {
+            if (node.type === 'PLAIN' && typeof node.value === 'string') {
+                const m = PYYAML_TIMESTAMP_RE.exec(node.value);
+                if (m) {
+                    const dm = PYYAML_DATE_ONLY_RE.exec(node.value);
+                    if (dm) {
+                        const parts = node.value.split('-');
+                        node.value = new PyDate(
+                            Number(parts[0]),
+                            Number(parts[1]),
+                            Number(parts[2]),
+                        );
+                    } else {
+                        // Full timestamp — date component is what matters.
+                        const dateComponent = node.value.split(/[Tt \t]/)[0] as string;
+                        const parts = dateComponent.split('-');
+                        node.value = new PyDate(
+                            Number(parts[0]),
+                            Number(parts[1]),
+                            Number(parts[2]),
+                        );
+                    }
+                }
+            }
+        },
+    });
+    return doc.toJS({ mapAsMap: false });
+}
+
+/** Days between two PyDate values, mirroring (date.today() - lv).days. */
+function _dateDiffDays(today: PyDate, lv: PyDate): number {
+    const a = Date.UTC(today.year, today.month - 1, today.day);
+    const b = Date.UTC(lv.year, lv.month - 1, lv.day);
+    return Math.round((a - b) / 86_400_000);
+}
+
+function _today(): PyDate {
+    const now = new Date();
+    return new PyDate(now.getFullYear(), now.getMonth() + 1, now.getDate());
+}
+
+function _isPlainObject(v: unknown): v is Record<string, unknown> {
+    return typeof v === 'object' && v !== null && !Array.isArray(v) && !(v instanceof PyDate);
+}
+
+function _isInteger(v: unknown): v is number {
+    return typeof v === 'number' && Number.isInteger(v);
+}
+
+/** Mirror Python len(str) — code-point count, not UTF-16 unit count. */
+function _pyLen(s: string): number {
+    let n = 0;
+    for (const _ of s) {
+        n += 1;
+    }
+    return n;
+}
+
+function _memoryType(p: string): string {
+    // Supports `memory/<type>.yml`, `memory/<type>/<hash>.yml`, and
+    // `memory/<type>.example.yml` template filenames.
+    const parts = p.split(path.sep);
+    if (!parts.includes('memory')) {
+        return _stem(parts[parts.length - 1] as string);
+    }
+    const idx = parts.indexOf('memory');
+    const nxt = idx + 1 < parts.length ? (parts[idx + 1] as string) : '';
+    const stem = _stem(nxt);
+    // Strip `.example` suffix used in template files.
+    return stem.endsWith('.example') ? stem.slice(0, -'.example'.length) : stem;
+}
+
+/** Mirror pathlib `Path(name).stem`: basename with the final suffix removed. */
+function _stem(name: string): string {
+    const base = path.basename(name);
+    const dot = base.lastIndexOf('.');
+    if (dot <= 0) {
+        return base;
+    }
+    return base.slice(0, dot);
+}
+
+function _validateEntry(
+    entry: Entry,
+    p: string,
+    seenIds: Set<string>,
+    findings: Finding[],
+    criticalCounts: Record<string, number> | null,
+): void {
+    const eid = typeof entry['id'] === 'string' ? (entry['id'] as string) : entry['id'] != null ? String(entry['id']) : '';
+    const eidStr = entry['id'] == null ? '' : eid;
+    const present = new Set(Object.keys(entry));
+    const missing: string[] = [];
+    for (const key of REQUIRED_KEYS) {
+        if (!present.has(key)) {
+            missing.push(key);
+        }
+    }
+    missing.sort();
+    for (const key of missing) {
+        findings.push(new Finding(p, 0, 'error', `missing required field: ${key}`, eidStr));
+    }
+    const status = entry['status'];
+    if (status && typeof status === 'string' && !VALID_STATUS.has(status)) {
+        findings.push(new Finding(p, 0, 'error', `invalid status '${status}'`, eidStr));
+    }
+    const confidence = entry['confidence'];
+    if (confidence && typeof confidence === 'string' && !VALID_CONFIDENCE.has(confidence)) {
+        findings.push(new Finding(p, 0, 'error', `invalid confidence '${confidence}'`, eidStr));
+    }
+    // Priority is optional (defaults to `normal` at read time). When present
+    // it MUST be one of the three-tier enum — see VALID_PRIORITY for the
+    // rationale on rejecting a fourth `high` tier.
+    const priority = entry['priority'];
+    if (priority !== undefined && priority !== null && !(typeof priority === 'string' && VALID_PRIORITY.has(priority))) {
+        const sortedPri = [...VALID_PRIORITY].sort();
+        findings.push(
+            new Finding(
+                p,
+                0,
+                'error',
+                `invalid priority '${_pyStr(priority)}' (expected one of [${sortedPri.map((x) => `'${x}'`).join(', ')}])`,
+                eidStr,
+            ),
+        );
+    }
+    const sources = (entry['source'] ?? []) as unknown;
+    if (!Array.isArray(sources) || sources.length < 1) {
+        findings.push(new Finding(p, 0, 'error', 'source must be a list with ≥1 entry', eidStr));
+    }
+    if (eidStr && seenIds.has(eidStr)) {
+        findings.push(new Finding(p, 0, 'error', `duplicate id '${eidStr}'`, eidStr));
+    }
+    seenIds.add(eidStr);
+    // Staleness.
+    const lv = entry['last_validated'];
+    const days = entry['review_after_days'];
+    if (lv instanceof PyDate && _isInteger(days)) {
+        const age = _dateDiffDays(_today(), lv);
+        if (age > days && entry['status'] === 'active') {
+            findings.push(new Finding(p, 0, 'info', `stale: last_validated ${age} days ago (limit ${days})`, eidStr));
+        }
+    }
+    // Critical-stale guard: a `priority: critical` entry that has not been
+    // re-validated within CRITICAL_STALE_DAYS surfaces as a warning, even
+    // when the entry's own `review_after_days` is more lenient. Critical
+    // entries surface on every /memory:load — they have a tighter SLA.
+    if (priority === 'critical' && entry['status'] === 'active' && lv instanceof PyDate) {
+        const critAge = _dateDiffDays(_today(), lv);
+        if (critAge > CRITICAL_STALE_DAYS) {
+            findings.push(
+                new Finding(
+                    p,
+                    0,
+                    'warning',
+                    `critical-stale: last_validated ${critAge} days ago (critical SLA is ${CRITICAL_STALE_DAYS} days)`,
+                    eidStr,
+                ),
+            );
+        }
+    }
+    // One-durable-fact-per-entry: reject transcript/narrative blobs. A single
+    // content field over ONE_FACT_MAX_CHARS is the bloat signal.
+    for (const fld of ONE_FACT_FIELDS) {
+        const val = entry[fld];
+        if (typeof val === 'string' && _pyLen(val) > ONE_FACT_MAX_CHARS) {
+            findings.push(
+                new Finding(
+                    p,
+                    0,
+                    'warning',
+                    `one-fact: \`${fld}\` is ${_pyLen(val)} chars (limit ${ONE_FACT_MAX_CHARS}) — split into separate durable facts, not a narrative blob`,
+                    eidStr,
+                ),
+            );
+            break;
+        }
+    }
+    // Tier-0 inflation tracking — increment per memory type. The aggregate
+    // warning is emitted in main() after all files are validated.
+    if (criticalCounts !== null && priority === 'critical' && entry['status'] === 'active') {
+        const mtype = _memoryType(p);
+        criticalCounts[mtype] = (criticalCounts[mtype] ?? 0) + 1;
+    }
+    // Per-type entry-count tracking — aggregate cap warning in main().
+    if (criticalCounts !== null) {
+        const mt = _memoryType(p);
+        _TYPE_COUNTS[mt] = (_TYPE_COUNTS[mt] ?? 0) + 1;
+    }
+}
+
+/** Mirror Python's `str(value)` for the priority error message. */
+function _pyStr(value: unknown): string {
+    if (value === null) {
+        return 'None';
+    }
+    if (value === true) {
+        return 'True';
+    }
+    if (value === false) {
+        return 'False';
+    }
+    if (value instanceof PyDate) {
+        const mm = String(value.month).padStart(2, '0');
+        const dd = String(value.day).padStart(2, '0');
+        return `${value.year}-${mm}-${dd}`;
+    }
+    return String(value);
+}
+
+function _checkRedaction(p: string, findings: Finding[]): void {
+    const text = fs.readFileSync(p, 'utf-8');
+    const lines = _splitlines(text);
+    lines.forEach((line, idx) => {
+        const lineNo = idx + 1;
+        // Skip comments — redaction warnings in example/commentary lines are noise.
+        if (line.trimStart().startsWith('#')) {
+            return;
+        }
+        for (const [pattern, label] of REDACTION_PATTERNS) {
+            if (pattern.test(line)) {
+                findings.push(new Finding(p, lineNo, 'error', `possible leak: ${label}`));
+            }
+        }
+    });
+}
+
+function _checkDateDiscipline(p: string, findings: Finding[]): void {
+    // Reject relative-date phrases without an ISO YYYY-MM-DD anchor.
+    const text = fs.readFileSync(p, 'utf-8');
+    const lines = _splitlines(text);
+    lines.forEach((line, idx) => {
+        const lineNo = idx + 1;
+        const stripped = line.trimStart();
+        if (stripped.startsWith('#') || stripped.startsWith('last_validated')) {
+            return;
+        }
+        RELATIVE_DATE_PATTERN.lastIndex = 0;
+        let match: RegExpExecArray | null;
+        while ((match = RELATIVE_DATE_PATTERN.exec(line)) !== null) {
+            const start = Math.max(0, match.index - DATE_ANCHOR_WINDOW);
+            const end = Math.min(line.length, match.index + match[0].length + DATE_ANCHOR_WINDOW);
+            const windowStr = line.slice(start, end);
+            if (ISO_DATE_PATTERN.test(windowStr)) {
+                continue;
+            }
+            const phrase = match[0];
+            findings.push(
+                new Finding(
+                    p,
+                    lineNo,
+                    'error',
+                    `relative date '${phrase}' without an ISO YYYY-MM-DD anchor within ±${DATE_ANCHOR_WINDOW} chars (re-anchor before commit)`,
+                ),
+            );
+        }
+    });
+}
+
+/** Mirror Python str.splitlines() for the line-numbering used above. */
+function _splitlines(text: string): string[] {
+    if (text === '') {
+        return [];
+    }
+    // Python splitlines splits on \n, \r, \r\n (and more) and drops a final
+    // trailing newline. For YAML/text files \n and \r\n cover the cases.
+    const lines = text.split(/\r\n|\r|\n/);
+    if (lines.length > 0 && lines[lines.length - 1] === '') {
+        lines.pop();
+    }
+    return lines;
+}
+
+function _validateFile(
+    p: string,
+    findings: Finding[],
+    criticalCounts: Record<string, number> | null,
+): void {
+    const mtype = _memoryType(p);
+    if (!KNOWN_TYPES.has(mtype)) {
+        findings.push(new Finding(p, 0, 'warning', `unknown memory type '${mtype}'`));
+    }
+    _checkRedaction(p, findings);
+    _checkDateDiscipline(p, findings);
+    let data: unknown;
+    try {
+        data = pyYamlSafeLoad(fs.readFileSync(p, 'utf-8'));
+        if (data === null || data === undefined) {
+            data = {};
+        }
+    } catch (exc) {
+        // mirror Python `except Exception` → report class name. PyYAML raises
+        // a YAMLError; the `yaml` package surfaces a parse error.
+        findings.push(new Finding(p, 0, 'error', `YAML parse error: ${_excClassName(exc)}`));
+        return;
+    }
+    if (!_isPlainObject(data) || !('entries' in data)) {
+        findings.push(new Finding(p, 0, 'error', "missing top-level 'entries' key"));
+        return;
+    }
+    const seenIds: Set<string> = new Set();
+    const entries = (data['entries'] ?? []) as unknown;
+    if (Array.isArray(entries)) {
+        for (const entry of entries) {
+            if (_isPlainObject(entry)) {
+                _validateEntry(entry as Entry, p, seenIds, findings, criticalCounts);
+            }
+        }
+    }
+}
+
+/** Mirror Python `exc.__class__.__name__` for the YAML-parse-error message. */
+function _excClassName(_exc: unknown): string {
+    // PyYAML errors surface as YAMLError. The TS twin always reports the same
+    // class name so the byte-identical "YAML parse error: <Name>" message holds
+    // for malformed YAML inputs.
+    return 'YAMLError';
+}
+
+const INTAKE_GLOB = 'agents/memory/intake/*.jsonl';
+
+function _git(cmd: string[]): [number, string] {
+    const result = spawnSync(cmd[0] as string, cmd.slice(1), { encoding: 'utf-8' });
+    if (result.error && (result.error as NodeJS.ErrnoException).code === 'ENOENT') {
+        return [127, 'git not found'];
+    }
+    const code = result.status ?? 1;
+    return [code, (result.stdout || '') + (result.stderr || '')];
+}
+
+function _resolveBase(explicit: string | null): string | null {
+    // CI: GITHUB_BASE_REF is set on PRs → origin/<base>.
+    if (explicit) {
+        return explicit;
+    }
+    const baseRef = process.env['GITHUB_BASE_REF'];
+    if (baseRef) {
+        return `origin/${baseRef}`;
+    }
+    // Local fallback: origin/main if it exists.
+    const [rc] = _git(['git', 'rev-parse', '--verify', 'origin/main']);
+    if (rc === 0) {
+        return 'origin/main';
+    }
+    return null;
+}
+
+function _checkAppendOnly(base: string | null, findings: Finding[]): void {
+    const ref = _resolveBase(base);
+    if (ref === null) {
+        findings.push(
+            new Finding(
+                INTAKE_GLOB,
+                0,
+                'warning',
+                'append-only: no base ref resolved (set --base or GITHUB_BASE_REF)',
+            ),
+        );
+        return;
+    }
+    const [rc, diff] = _git([
+        'git',
+        'diff',
+        '--unified=0',
+        '--no-color',
+        ref,
+        '--',
+        'agents/memory/intake/',
+    ]);
+    if (rc !== 0) {
+        findings.push(new Finding(INTAKE_GLOB, 0, 'warning', `append-only: git diff failed vs ${ref}`));
+        return;
+    }
+    if (diff.trim() === '') {
+        return; // nothing changed, nothing to check
+    }
+    let currentFile = '';
+    let oldPath = '';
+    for (const line of diff.split('\n')) {
+        if (line.startsWith('diff --git')) {
+            const parts = line.split(/\s+/);
+            currentFile = parts.length >= 4 ? (parts[parts.length - 1] as string).slice(2) : '';
+            oldPath = '';
+        } else if (line.startsWith('--- ')) {
+            oldPath = line.slice(4).trim();
+        } else if (line.startsWith('@@')) {
+            // Hunk header: @@ -oldStart,oldCount +newStart,newCount @@
+            const m = /^@@ -(\d+)(?:,(\d+))? \+(\d+)(?:,(\d+))? @@/.exec(line);
+            if (!m) {
+                continue;
+            }
+            const oldStart = parseInt(m[1] as string, 10);
+            const oldCount = parseInt(m[2] ?? '1', 10);
+            // Any hunk that removes >0 lines from the old file = in-place edit.
+            if (oldCount > 0 && oldPath !== '/dev/null') {
+                findings.push(
+                    new Finding(
+                        currentFile || INTAKE_GLOB,
+                        oldStart,
+                        'error',
+                        `append-only violation: ${oldCount} existing line(s) removed or modified (ref=${ref})`,
+                    ),
+                );
+            }
+        }
+    }
+}
+
+interface ParsedArgs {
+    path: string;
+    format: string;
+    append_only: boolean;
+    base: string | null;
+}
+
+/** Minimal argparse-compatible parser for this script's surface. */
+function _parseArgs(argv: string[]): ParsedArgs {
+    const args: ParsedArgs = {
+        path: 'agents/memory',
+        format: 'text',
+        append_only: false,
+        base: null,
+    };
+    for (let i = 0; i < argv.length; i += 1) {
+        const a = argv[i] as string;
+        if (a === '--path') {
+            args.path = argv[++i] as string;
+        } else if (a.startsWith('--path=')) {
+            args.path = a.slice('--path='.length);
+        } else if (a === '--format') {
+            args.format = _checkChoice(argv[++i] as string, ['text', 'json'], '--format');
+        } else if (a.startsWith('--format=')) {
+            args.format = _checkChoice(a.slice('--format='.length), ['text', 'json'], '--format');
+        } else if (a === '--append-only') {
+            args.append_only = true;
+        } else if (a === '--base') {
+            args.base = argv[++i] as string;
+        } else if (a.startsWith('--base=')) {
+            args.base = a.slice('--base='.length);
+        } else if (a === '-h' || a === '--help') {
+            _printUsage();
+            process.exit(0);
+        } else {
+            process.stderr.write(`check_memory: error: unrecognized arguments: ${a}\n`);
+            process.exit(2);
+        }
+    }
+    return args;
+}
+
+function _checkChoice(value: string | undefined, choices: string[], flag: string): string {
+    if (value === undefined || !choices.includes(value)) {
+        process.stderr.write(
+            `check_memory: error: argument ${flag}: invalid choice: '${value ?? ''}' (choose from ${choices.map((c) => `'${c}'`).join(', ')})\n`,
+        );
+        process.exit(2);
+    }
+    return value;
+}
+
+function _printUsage(): void {
+    process.stdout.write(
+        'usage: check_memory [-h] [--path PATH] [--format {text,json}] [--append-only] [--base BASE]\n',
+    );
+}
+
+function main(): number {
+    const args = _parseArgs(process.argv.slice(2));
+    const root = args.path;
+    const findings: Finding[] = [];
+    if (args.append_only) {
+        _checkAppendOnly(args.base, findings);
+        // Append-only mode is standalone by design: it inspects git state, not
+        // files on disk, and is meant to run as a fast CI gate. Skip YAML
+        // validation.
+        return _emit(findings, args.format);
+    }
+    if (!fs.existsSync(root)) {
+        if (args.format === 'json') {
+            process.stdout.write(`${pyJsonDumps({ findings: [], note: `${root} not found` }, null)}\n`);
+        } else {
+            process.stdout.write(`ℹ️  ${root} not found — nothing to validate\n`);
+        }
+        return 0;
+    }
+    const criticalCounts: Record<string, number> = {};
+    for (const key of Object.keys(_TYPE_COUNTS)) {
+        delete _TYPE_COUNTS[key];
+    }
+    for (const yml of _rglobYmlSorted(root)) {
+        _validateFile(yml, findings, criticalCounts);
+    }
+    // Tier-0 inflation warning — soft cap on `priority: critical` per type.
+    // Council convergence (Phase 2 B2): warn rather than block, because the
+    // right answer to "too many criticals" is curator review, not CI failure.
+    for (const [mtype, count] of Object.entries(criticalCounts).sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0))) {
+        if (count > CRITICAL_WARN_THRESHOLD) {
+            findings.push(
+                new Finding(
+                    `agents/memory/${mtype}`,
+                    0,
+                    'warning',
+                    `tier-0 inflation: ${count} active 'priority: critical' entries (threshold ${CRITICAL_WARN_THRESHOLD}) — review whether all still warrant always-surface treatment`,
+                ),
+            );
+        }
+    }
+    // Per-type entry-count cap (size-bounding, Phase 7). Warn, never block —
+    // over-cap signals a consolidation pass is due (prune archived, merge dups).
+    for (const [mtype, count] of Object.entries(_TYPE_COUNTS).sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0))) {
+        const cap = PER_TYPE_CAPS[mtype] ?? DEFAULT_TYPE_CAP;
+        if (count > cap) {
+            findings.push(
+                new Finding(
+                    `agents/memory/${mtype}`,
+                    0,
+                    'warning',
+                    `entry-cap: ${count} entries (soft cap ${cap}) — run a consolidation pass (prune archived, merge duplicates)`,
+                ),
+            );
+        }
+    }
+    return _emit(findings, args.format);
+}
+
+/** Mirror `sorted(root.rglob("*.yml"))` — absolute paths, sorted by POSIX string. */
+function _rglobYmlSorted(root: string): string[] {
+    const out: string[] = [];
+    const rootAbs = path.resolve(root);
+    const walk = (dir: string): void => {
+        let entries: fs.Dirent[];
+        try {
+            entries = fs.readdirSync(dir, { withFileTypes: true });
+        } catch {
+            return;
+        }
+        for (const ent of entries) {
+            const full = path.join(dir, ent.name);
+            if (ent.name.endsWith('.yml')) {
+                out.push(full);
+            }
+            if (ent.isDirectory() || (ent.isSymbolicLink() && _isDirSafe(full))) {
+                walk(full);
+            }
+        }
+    };
+    walk(rootAbs);
+    // Python's Path.rglob result is sorted by the Path objects; sorting the
+    // resolved absolute paths by string reproduces that order on POSIX hosts.
+    // But the original passes a relative root (e.g. "agents/memory") and
+    // emits str(Path) of those relative paths. Reproduce by re-relativizing.
+    const cwd = process.cwd();
+    const rel = out.map((p) => {
+        const r = path.relative(cwd, p);
+        // Match str(Path(root) / ...) which keeps the given root form. When the
+        // user passed an absolute root, keep absolute.
+        return path.isAbsolute(root) ? p : r;
+    });
+    rel.sort();
+    return rel;
+}
+
+function _isDirSafe(p: string): boolean {
+    try {
+        return fs.statSync(p).isDirectory();
+    } catch {
+        return false;
+    }
+}
+
+function _emit(findings: Finding[], fmt: string): number {
+    const errors = findings.filter((f) => f.severity === 'error');
+    if (fmt === 'json') {
+        process.stdout.write(
+            `${pyJsonDumps(
+                {
+                    findings: findings.map((f) => ({
+                        file: f.file,
+                        line: f.line,
+                        severity: f.severity,
+                        message: f.message,
+                        entry_id: f.entry_id,
+                    })),
+                },
+                2,
+            )}\n`,
+        );
+    } else {
+        const icon: Record<Severity, string> = { error: '❌', warning: '⚠️', info: 'ℹ️' };
+        for (const f of findings) {
+            const loc = f.line ? `${f.file}:${f.line}` : f.file;
+            const suffix = f.entry_id ? ` [${f.entry_id}]` : '';
+            process.stdout.write(`  ${icon[f.severity]}  ${loc}${suffix}  ${f.message}\n`);
+        }
+        const warnings = findings.filter((f) => f.severity === 'warning').length;
+        const infos = findings.filter((f) => f.severity === 'info').length;
+        process.stdout.write(`\nSummary: ${errors.length} error(s), ${warnings} warning(s), ${infos} info\n`);
+    }
+    return errors.length > 0 ? 1 : 0;
+}
+
+const _isMain =
+    process.argv[1] !== undefined && import.meta.url === pathToFileURL(process.argv[1]).href;
+if (_isMain) {
+    process.exit(main());
+}
+
+export { main };