@event4u/agent-config 6.0.0 → 7.0.0

package/src/scripts/mcp_server/tools.ts

@@ -0,0 +1,1031 @@
+// MCP Server — Phase 4 tools layer + Phase 1 discovery stubs.
+//
+// A0 contract amendment: real handlers run only for the tools listed in
+// `ALLOWLIST` (`lint_skills` + `chat_history_append`). All other names
+// in `scripts/mcp_server/consumer_tool_catalog.json` are surfaced via
+// `tools/list` as discovery stubs; `tools/call` against them returns
+// the `not_implemented` envelope defined in
+// `docs/contracts/mcp-tool-stub-envelope.md` (a successful result with
+// `code: not_implemented`, an `install_hint` and an `alternative`).
+// Names that are neither implemented nor catalog-listed raise an
+// Error (rendered by the SDK as JSON-RPC error).
+//
+// Path-scoping is mandatory for any tool that writes: the resolved target
+// path must stay under `<consumer_root>` and within the allowlist of
+// filenames (`agents/runtime/.agent-chat-history` — current default;
+// `agents/.agent-chat-history` and `.agent-chat-history` — kept for
+// back-compat with older consumer installs that have not migrated yet).
+// Escape attempts surface as an Error before the underlying writer runs.
+//
+// This module deliberately does NOT spawn shells or HTTP clients. The
+// chat-history writer is a minimal faithful inline port of the five
+// functions `tools.py` lazily imports from `scripts/chat_history.py`
+// (SCHEMA_VERSION, init, append, read_header, read_entries); the wire
+// surface exposes no shell execution.
+//
+// Tools return an object from their handlers — the SDK wraps that in a
+// TextContent block with the JSON-serialized payload, so MCP clients
+// can render structured output.
+//
+// TS twin of tools.py (py2ts Phase 8). Mirrors the full public surface:
+//   STDIO_TRANSPORT, BuiltinTool, ToolHandler, ALLOWLIST, to_mcp_tool_meta,
+//   CATALOG_STUBS, STUB_NAMES, REGISTRY, ToolCache, boot_log_line.
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+
+import {
+    SCHEMA_VERSION as _CH_SCHEMA_VERSION,
+    append as _chAppendImpl,
+    init as _chInitImpl,
+    read_entries as _chReadEntriesImpl,
+    read_header as _chReadHeaderImpl,
+} from '../chat_history.js';
+import {
+    type CatalogEntry,
+    install_hint as _catalog_install_hint,
+    load_catalog,
+    not_implemented_envelope,
+} from './catalog.js';
+import { type Outcome, record_call } from './telemetry.js';
+import type { PromptCache } from './prompts.js';
+import type { ResourceCache } from './resources.js';
+
+// Stable transport tag for the stub envelope. Mirrored verbatim by
+// `internal/workers/mcp/src/stubs.ts` with `"worker"`.
+export const STDIO_TRANSPORT = 'stdio';
+
+// Allowlisted directories (relative to consumer_root) where tool writes
+// are permitted. `chat_history_append` resolves its path through this
+// guard before the underlying writer touches the filesystem.
+const _ALLOWED_WRITE_REL_PATHS: ReadonlySet<string> = new Set([
+    // Current default (Volatile Runtime policy — agents/runtime/ is
+    // local-only and ignored by git).
+    'agents/runtime/.agent-chat-history',
+    // Back-compat: older consumer installs still write to the flat
+    // location. Additive — both paths stay accepted until the next
+    // major tool version bump.
+    'agents/.agent-chat-history',
+    '.agent-chat-history',
+]);
+
+export type ToolHandler = (
+    args: Record<string, unknown>,
+    consumerRoot: string,
+) => Promise<Record<string, unknown>>;
+
+/**
+ * Static registration record for an allowlisted MCP tool.
+ *
+ * `input_schema` is a JSON-Schema object the SDK validates against on
+ * each `tools/call`. `handler` is an async function that receives the
+ * validated arguments + the resolved `consumer_root` path.
+ *
+ * Mirrors the Python frozen dataclass `BuiltinTool` (field order
+ * preserved).
+ */
+export interface BuiltinTool {
+    readonly name: string;
+    readonly description: string;
+    readonly input_schema: Record<string, unknown>;
+    readonly handler: ToolHandler;
+}
+
+/** Mirror Python `str.strip()` (also strips the same Unicode whitespace set we care about here). */
+function _strip(s: string): string {
+    return s.replace(/^\s+/, '').replace(/\s+$/, '');
+}
+
+/** Mirror Python `Path(...).resolve()` — absolutize + realpath, tolerating a missing tail. */
+function _resolvePath(p: string): string {
+    try {
+        return fs.realpathSync(p);
+    } catch {
+        return path.resolve(p);
+    }
+}
+
+/**
+ * Pick the consumer-project root.
+ *
+ * Default: the current working directory. Tests pass an explicit
+ * override; the stdio entrypoint relies on the CWD set by the
+ * `./agent-config mcp:run` wrapper.
+ */
+function _resolveConsumerRoot(override?: string | null): string {
+    if (override !== undefined && override !== null) {
+        return _resolvePath(override);
+    }
+    return _resolvePath(process.cwd());
+}
+
+/**
+ * Compute the POSIX relative path of `target` under `root`, or null
+ * when `target` is not under `root`. Mirrors `Path.relative_to`, which
+ * is purely lexical (no symlink chasing) on already-resolved paths.
+ */
+function _relativeUnder(target: string, root: string): string | null {
+    const rel = path.relative(root, target);
+    if (rel === '') {
+        return '';
+    }
+    if (rel.startsWith('..') || path.isAbsolute(rel)) {
+        return null;
+    }
+    return rel.split(path.sep).join('/');
+}
+
+/** Mirror Python `repr(sorted(...))` for the write-allowlist error text. */
+function _reprStrList(values: string[]): string {
+    return `[${values.map((v) => `'${v}'`).join(', ')}]`;
+}
+
+/**
+ * Resolve `raw` under `consumer_root` and assert it stays in tree.
+ *
+ * Returns the resolved target path. Raises when the path escapes the
+ * root or is not in the write allowlist. null / undefined / empty
+ * falls back to the default chat-history location.
+ */
+function _validateInTreePath(raw: string | null | undefined, consumerRoot: string): string {
+    const root = _resolvePath(consumerRoot);
+    let target: string;
+    if (raw === null || raw === undefined || raw === '') {
+        target = path.join(root, 'agents', 'runtime', '.agent-chat-history');
+    } else if (path.isAbsolute(raw)) {
+        target = _resolvePath(raw);
+    } else {
+        target = _resolvePath(path.join(root, raw));
+    }
+    const rel = _relativeUnder(target, root);
+    if (rel === null) {
+        throw new Error(`path escapes consumer_root: ${target} not under ${root}`);
+    }
+    if (!_ALLOWED_WRITE_REL_PATHS.has(rel)) {
+        throw new Error(
+            `path not in write allowlist: '${rel}' ` +
+                `(allowed: ${_reprStrList([..._ALLOWED_WRITE_REL_PATHS].sort())})`,
+        );
+    }
+    return target;
+}
+
+// ---------------------------------------------------------------------
+// Minimal chat-history surface — the five functions tools.py lazily
+// imports from scripts/chat_history.py (SCHEMA_VERSION, init, append,
+// read_header, read_entries) now come from the chat_history.ts twin
+// (the canonical single source of truth, byte-for-byte with the Python
+// writer: ensure_ascii=False default separators, trailing newline,
+// ISO-8601 second-precision `ts`). These thin shims keep the handler
+// call sites unchanged while delegating to the twin.
+// ---------------------------------------------------------------------
+
+function _chReadHeader(target: string): Record<string, unknown> | null {
+    return _chReadHeaderImpl(target);
+}
+
+function _chInit(target: string, freq = 'per_phase'): Record<string, unknown> {
+    return _chInitImpl(freq, { path: target });
+}
+
+function _chAppend(
+    entry: Record<string, unknown>,
+    target: string,
+    session: string | null | undefined,
+): void {
+    _chAppendImpl(entry, { path: target, session });
+}
+
+/** Mirror `read_entries(last, path, session)` (the subset the read handler uses). */
+function _chReadEntries(
+    target: string,
+    last: number | null | undefined,
+    session: string | null | undefined,
+): Record<string, unknown>[] {
+    return _chReadEntriesImpl({ path: target, last, session });
+}
+
+// ---------------------------------------------------------------------
+// Implemented handlers
+// ---------------------------------------------------------------------
+
+/**
+ * D2 — read-only wrapper around `skill_linter.lint_file`.
+ *
+ * Arguments:
+ *   paths: optional list of repo-relative paths to lint. Empty /
+ *       missing → lint the full `.agent-src.uncondensed/` tree
+ *       via `gather_all_candidate_files`.
+ *
+ * Never spawns `git` (no `--changed` mode); never writes; mirrors
+ * the JSON output format of `scripts/skill_linter.py --format json`.
+ */
+async function _lintSkillsHandler(
+    args: Record<string, unknown>,
+    consumerRoot: string,
+): Promise<Record<string, unknown>> {
+    // Import lazily so the loader-layer import-surface test stays clean.
+    const { format_json, gather_all_candidate_files, lint_file } = await import(
+        '../skill_linter.js'
+    );
+
+    const root = _resolvePath(consumerRoot);
+    const requested = args.paths ?? [];
+    if (!Array.isArray(requested)) {
+        throw new Error("'paths' must be a list of strings");
+    }
+
+    let paths: string[] = [];
+    if (requested.length > 0) {
+        for (const raw of requested) {
+            if (typeof raw !== 'string') {
+                throw new Error("'paths' entries must be strings");
+            }
+            const resolved = path.isAbsolute(raw)
+                ? _resolvePath(raw)
+                : _resolvePath(path.join(root, raw));
+            if (_relativeUnder(resolved, root) === null) {
+                throw new Error(`path escapes consumer_root: ${resolved}`);
+            }
+            if (fs.existsSync(resolved)) {
+                paths.push(resolved);
+            }
+        }
+    } else {
+        paths = gather_all_candidate_files(root);
+    }
+
+    const sortedUnique = [...new Set(paths)].sort(_pathlibCompare);
+    const results = sortedUnique.map((p) => lint_file(p, root));
+    const payload = JSON.parse(format_json(results)) as Record<string, unknown>;
+    return payload;
+}
+
+/**
+ * D3 — append one entry to the consumer's chat-history JSONL.
+ *
+ * Arguments mirror tools.py: text, entry_type, path, session, dry_run,
+ * min_schema_version.
+ */
+async function _chatHistoryAppendHandler(
+    args: Record<string, unknown>,
+    consumerRoot: string,
+): Promise<Record<string, unknown>> {
+    const text = args.text;
+    if (typeof text !== 'string' || !_strip(text)) {
+        throw new Error("'text' must be a non-empty string");
+    }
+    const entryTypeRaw = args.entry_type ?? 'note';
+    if (typeof entryTypeRaw !== 'string' || !_strip(entryTypeRaw)) {
+        throw new Error("'entry_type' must be a non-empty string");
+    }
+    const entryType = entryTypeRaw || 'note';
+    if (entryType === 'header') {
+        throw new Error("'entry_type' must not be 'header'");
+    }
+
+    const session = args.session;
+    if (session !== undefined && session !== null && typeof session !== 'string') {
+        throw new Error("'session' must be a string when provided");
+    }
+
+    const dryRun = _pyTruthy(args.dry_run ?? false);
+
+    const rawPath = (args.path ?? null) as string | null;
+    const target = _validateInTreePath(rawPath, consumerRoot);
+
+    const minSchema = args.min_schema_version;
+    if (minSchema !== undefined && minSchema !== null) {
+        if (typeof minSchema !== 'number' || !Number.isInteger(minSchema) || typeof minSchema === 'boolean') {
+            throw new Error("'min_schema_version' must be an integer");
+        }
+        const existingHeader = fs.existsSync(target) ? _chReadHeader(target) : null;
+        const observed =
+            existingHeader !== null && typeof existingHeader === 'object'
+                ? _pyInt(existingHeader.v ?? 0)
+                : _CH_SCHEMA_VERSION;
+        if (observed < minSchema) {
+            throw new Error(`chat-history schema ${observed} below required ${minSchema}`);
+        }
+    }
+
+    const entry: Record<string, unknown> = { t: entryType, text };
+
+    if (dryRun) {
+        return {
+            dry_run: true,
+            target_path: target,
+            entry,
+            session: session === undefined ? null : session,
+        };
+    }
+
+    // `append` requires the parent directory and a header line. Lazy-init
+    // the JSONL when the consumer hasn't run `agent-config chat:init` yet.
+    if (!fs.existsSync(target) || _chReadHeader(target) === null) {
+        fs.mkdirSync(path.dirname(target), { recursive: true });
+        _chInit(target);
+    }
+    _chAppend(entry, target, (session ?? null) as string | null);
+    return {
+        dry_run: false,
+        target_path: target,
+        entry,
+        session: session === undefined ? null : session,
+    };
+}
+
+/**
+ * Phase 3 L2 — read entries from the consumer's chat-history JSONL.
+ *
+ * Arguments mirror tools.py: last, session, entry_type, path.
+ */
+async function _chatHistoryReadHandler(
+    args: Record<string, unknown>,
+    consumerRoot: string,
+): Promise<Record<string, unknown>> {
+    const rawPath = (args.path ?? null) as string | null;
+    const target = _validateInTreePath(rawPath, consumerRoot);
+
+    const last = args.last;
+    if (
+        last !== undefined &&
+        last !== null &&
+        (typeof last !== 'number' || !Number.isInteger(last) || typeof last === 'boolean' || last < 1)
+    ) {
+        throw new Error("'last' must be a positive integer when provided");
+    }
+    const session = args.session;
+    if (session !== undefined && session !== null && typeof session !== 'string') {
+        throw new Error("'session' must be a string when provided");
+    }
+    const entryType = args.entry_type;
+    if (entryType !== undefined && entryType !== null && typeof entryType !== 'string') {
+        throw new Error("'entry_type' must be a string when provided");
+    }
+
+    if (!fs.existsSync(target)) {
+        return { path: target, entries: [], count: 0 };
+    }
+
+    let entries = _chReadEntries(
+        target,
+        (last ?? null) as number | null,
+        (session ?? null) as string | null,
+    );
+    if (entryType) {
+        entries = entries.filter((e) => e.t === entryType);
+    }
+    return { path: target, entries, count: entries.length };
+}
+
+/**
+ * Phase 3 L2 — file-backed memory retrieval over `agents/memory/`.
+ *
+ * Wraps `scripts/memory_lookup.retrieve_v1` to keep the v1 envelope on
+ * the wire. Retrieval is entirely file-backed (no external backend).
+ */
+async function _memoryLookupHandler(
+    args: Record<string, unknown>,
+    consumerRoot: string,
+): Promise<Record<string, unknown>> {
+    const memoryLookup = await import('../memory_lookup.js');
+
+    const types = args.types;
+    if (
+        !Array.isArray(types) ||
+        types.length === 0 ||
+        !types.every((t) => typeof t === 'string')
+    ) {
+        throw new Error("'types' must be a non-empty list of strings");
+    }
+    const keys = args.keys ?? [];
+    if (!Array.isArray(keys) || !keys.every((k) => typeof k === 'string')) {
+        throw new Error("'keys' must be a list of strings");
+    }
+    const limitRaw = args.limit ?? 5;
+    if (
+        typeof limitRaw !== 'number' ||
+        !Number.isInteger(limitRaw) ||
+        typeof limitRaw === 'boolean' ||
+        limitRaw < 1
+    ) {
+        throw new Error("'limit' must be a positive integer");
+    }
+
+    // Mirror the Python `os.chdir(consumer_root)` scoping: memory_lookup
+    // resolves `agents/memory` relative to CWD, so the consumer root is
+    // applied via a chdir window that is always restored.
+    const prevCwd = process.cwd();
+    let envelope: Record<string, unknown>;
+    try {
+        process.chdir(consumerRoot);
+        envelope = memoryLookup.retrieve_v1(
+            [...(types as string[])],
+            [...(keys as string[])],
+            limitRaw,
+        );
+    } finally {
+        process.chdir(prevCwd);
+    }
+    return envelope;
+}
+
+/** Phase 3 L2 — surface `scripts/memory_status.status()` as JSON. */
+async function _memoryStatusHandler(
+    _args: Record<string, unknown>,
+    consumerRoot: string,
+): Promise<Record<string, unknown>> {
+    const memoryStatus = await import('../memory_status.js');
+
+    const prevCwd = process.cwd();
+    let result: ReturnType<typeof memoryStatus.status>;
+    try {
+        process.chdir(consumerRoot);
+        result = memoryStatus.status();
+    } finally {
+        process.chdir(prevCwd);
+    }
+    // Python `asdict(result)` over the constant file-backend dataclass →
+    // {status, backend, reason, elapsed_ms}. memory_status.ts.status()
+    // already returns that plain object, so it IS the dict.
+    return result as unknown as Record<string, unknown>;
+}
+
+// Module-level prompt / resource caches reused across handler calls so
+// repeated `list_*` / `read_resource_body` calls share mtime tracking.
+const _PROMPT_CACHES = new Map<string, unknown>();
+const _RESOURCE_CACHES = new Map<string, unknown>();
+
+async function _getPromptCache(consumerRoot: string): Promise<PromptCache> {
+    const { PromptCache } = await import('./prompts.js');
+    const key = _resolvePath(consumerRoot);
+    let cache = _PROMPT_CACHES.get(key) as PromptCache | undefined;
+    if (cache === undefined) {
+        cache = new PromptCache(consumerRoot);
+        _PROMPT_CACHES.set(key, cache);
+    }
+    return cache;
+}
+
+async function _getResourceCache(
+    consumerRoot: string,
+): Promise<ResourceCache> {
+    const { ResourceCache } = await import('./resources.js');
+    const key = _resolvePath(consumerRoot);
+    let cache = _RESOURCE_CACHES.get(key) as ResourceCache | undefined;
+    if (cache === undefined) {
+        cache = new ResourceCache(consumerRoot);
+        _RESOURCE_CACHES.set(key, cache);
+    }
+    return cache;
+}
+
+/** Phase 3 L2 — enumerate skill prompts (kind=='skill'). */
+async function _listSkillsHandler(
+    _args: Record<string, unknown>,
+    consumerRoot: string,
+): Promise<Record<string, unknown>> {
+    const { to_mcp_prompt_meta } = await import('./prompts.js');
+    const cache = await _getPromptCache(consumerRoot);
+    const [prompts, errors] = cache.get();
+    const items = prompts
+        .filter((p) => p.kind === 'skill')
+        .map((p) => ({
+            name: p.name,
+            description: p.description,
+            source: p.source,
+            wire_name: to_mcp_prompt_meta(p).name as string,
+        }));
+    items.sort((a, b) => _strCmp(a.name, b.name));
+    return { count: items.length, skills: items, errors: [...errors] };
+}
+
+/** Phase 3 L2 — enumerate command prompts (kind=='command'). */
+async function _listCommandsHandler(
+    _args: Record<string, unknown>,
+    consumerRoot: string,
+): Promise<Record<string, unknown>> {
+    const { to_mcp_prompt_meta } = await import('./prompts.js');
+    const cache = await _getPromptCache(consumerRoot);
+    const [prompts, errors] = cache.get();
+    const items = prompts
+        .filter((p) => p.kind === 'command')
+        .map((p) => ({
+            name: p.name,
+            description: p.description,
+            source: p.source,
+            wire_name: to_mcp_prompt_meta(p).name as string,
+        }));
+    items.sort((a, b) => _strCmp(a.name, b.name));
+    return { count: items.length, commands: items, errors: [...errors] };
+}
+
+/** Phase 3 L2 — enumerate rule resources (kind=='rule'). */
+async function _listRulesHandler(
+    _args: Record<string, unknown>,
+    consumerRoot: string,
+): Promise<Record<string, unknown>> {
+    const cache = await _getResourceCache(consumerRoot);
+    const [resources, errors] = cache.get();
+    const items = resources
+        .filter((r) => r.kind === 'rule')
+        .map((r) => ({
+            uri: r.uri,
+            name: r.name,
+            description: r.description,
+            source: r.source,
+        }));
+    items.sort((a, b) => _strCmp(a.uri, b.uri));
+    return { count: items.length, rules: items, errors: [...errors] };
+}
+
+/** Phase 3 L2 — fetch the rendered body of a resource URI. */
+async function _readResourceBodyHandler(
+    args: Record<string, unknown>,
+    consumerRoot: string,
+): Promise<Record<string, unknown>> {
+    const uri = args.uri;
+    if (typeof uri !== 'string' || !uri) {
+        throw new Error("'uri' must be a non-empty string");
+    }
+    const cache = await _getResourceCache(consumerRoot);
+    const resource = cache.lookup(uri);
+    if (resource === null) {
+        throw new Error(`resource not found: ${uri}`);
+    }
+    return {
+        uri: resource.uri,
+        name: resource.name,
+        description: resource.description,
+        mime_type: resource.mime_type,
+        kind: resource.kind,
+        source: resource.source,
+        body: resource.body,
+    };
+}
+
+// ---------------------------------------------------------------------
+// Allowlist — hardcoded per AI Council Q1-a verdict (2026-05-10),
+// extended Phase 3 L2 (2026-05-12) under the council-waiver verdict.
+// Adding a tool here is a code-review event; settings cannot enable an
+// unlisted tool. Boot-time stderr log enumerates the registered set.
+// ---------------------------------------------------------------------
+
+export const ALLOWLIST: Record<string, BuiltinTool> = {
+    lint_skills: {
+        name: 'lint_skills',
+        description:
+            'Lint skill, rule, command, guideline, and persona markdown ' +
+            'files for frontmatter and structural errors. Use before ' +
+            'committing or opening a PR that adds or edits any of those ' +
+            'artifacts, to catch schema violations early. Read-only — ' +
+            'never writes files or spawns git. Returns the ' +
+            '`scripts/skill_linter.py --format json` payload: a `summary` ' +
+            'object (pass / pass_with_warnings / fail / total counts) and ' +
+            'a per-file `results` array with severity-tagged findings. ' +
+            'Pass `paths` to lint a subset; omit for a full tree scan.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                paths: {
+                    type: 'array',
+                    items: { type: 'string' },
+                    description:
+                        'Repo-relative paths to lint (files or ' +
+                        'directories). Empty or missing → full tree scan ' +
+                        'via gather_all_candidate_files.',
+                },
+            },
+            additionalProperties: false,
+        },
+        handler: _lintSkillsHandler,
+    },
+    chat_history_append: {
+        name: 'chat_history_append',
+        description:
+            "Append one structured entry to the consumer project's " +
+            'chat-history log (a JSONL file). Use to record a decision, ' +
+            'note, or phase marker that should persist into a later ' +
+            'session or be distilled by `mine_session`. Writes to the ' +
+            'filesystem (`agents/runtime/.agent-chat-history` by default; ' +
+            '`agents/.agent-chat-history` and `.agent-chat-history` ' +
+            'accepted for back-compat) and returns the written entry plus ' +
+            'its resolved target path. Path-scoped: a `path` outside the ' +
+            'allowlist, or any traversal escaping the project root, raises ' +
+            'an error before writing. Set `dry_run: true` to preview the ' +
+            'entry and target path without touching disk.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                text: {
+                    type: 'string',
+                    description: 'The entry body to record.',
+                },
+                entry_type: {
+                    type: 'string',
+                    description:
+                        'Short ``t`` tag categorising the entry (e.g. ' +
+                        'note, decision, phase). Defaults to ``note``.',
+                },
+                path: {
+                    type: 'string',
+                    description:
+                        'Optional path override. Must resolve to ' +
+                        '`agents/runtime/.agent-chat-history` ' +
+                        '(current default), ' +
+                        '`agents/.agent-chat-history`, or ' +
+                        '`.agent-chat-history` under consumer_root.',
+                },
+                session: {
+                    type: 'string',
+                    description:
+                        'Optional 16-char session id to group the entry ' +
+                        'under. Defaults to the current session.',
+                },
+                dry_run: {
+                    type: 'boolean',
+                    default: false,
+                    description:
+                        'When true, return the entry and resolved target ' +
+                        'path without writing to disk.',
+                },
+                min_schema_version: {
+                    type: 'integer',
+                    description:
+                        'Refuse to write if the on-disk history schema is ' +
+                        'older than this version.',
+                },
+            },
+            required: ['text'],
+            additionalProperties: false,
+        },
+        handler: _chatHistoryAppendHandler,
+    },
+    chat_history_read: {
+        name: 'chat_history_read',
+        description:
+            "Read recent entries back from the consumer project's " +
+            'chat-history JSONL ' +
+            '(`agents/runtime/.agent-chat-history`; ' +
+            '`agents/.agent-chat-history` accepted for back-compat). Use ' +
+            'to recover context from an earlier session — decisions, ' +
+            'notes, phase markers — at the start of a new task. ' +
+            'Read-only. Returns the resolved file path plus a list of ' +
+            'matching entries (newest last). Combine `session`, `last`, ' +
+            'and `entry_type` to narrow the result.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                last: {
+                    type: 'integer',
+                    minimum: 1,
+                    description:
+                        'Return only the most recent N entries, after ' +
+                        'other filters apply.',
+                },
+                session: {
+                    type: 'string',
+                    description: 'Filter to a single 16-char session id.',
+                },
+                entry_type: {
+                    type: 'string',
+                    description: 'Filter by the `t` tag (e.g. note, decision, ' + 'phase).',
+                },
+                path: {
+                    type: 'string',
+                    description:
+                        'Optional history-file path override; defaults to ' +
+                        'the standard chat-history location under the ' +
+                        'project root.',
+                },
+            },
+            additionalProperties: false,
+        },
+        handler: _chatHistoryReadHandler,
+    },
+    memory_lookup: {
+        name: 'memory_lookup',
+        description:
+            'Retrieve engineering-memory entries for one or more memory ' +
+            'types, optionally narrowed to specific anchor paths. Use ' +
+            'before editing a security-sensitive or historically buggy ' +
+            'file to surface prior incidents, ownership, and patterns ' +
+            'tied to it. Reads `agents/memory/<type>/*.yml` plus the ' +
+            '`agents/memory/intake/*.jsonl` signal log. Read-only. ' +
+            'Returns the v1 retrieval envelope: a `status` field plus ' +
+            'per-type `slices` carrying the matched entries.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                types: {
+                    type: 'array',
+                    items: { type: 'string' },
+                    minItems: 1,
+                    description:
+                        'Memory types to scan, e.g. `historical-patterns`, ' +
+                        '`incident-learnings`, `ownership`. At least one ' +
+                        'required.',
+                },
+                keys: {
+                    type: 'array',
+                    items: { type: 'string' },
+                    description:
+                        'Optional anchor paths or globs to match entries ' +
+                        'against (e.g. a file you are about to edit).',
+                },
+                limit: {
+                    type: 'integer',
+                    minimum: 1,
+                    default: 5,
+                    description: 'Maximum entries to return per type. Defaults to 5.',
+                },
+            },
+            required: ['types'],
+            additionalProperties: false,
+        },
+        handler: _memoryLookupHandler,
+    },
+    memory_status: {
+        name: 'memory_status',
+        description:
+            'Report the memory backend status. Memory is entirely ' +
+            'file-backed (`agents/memory/`); there is no external backend. ' +
+            'Read-only, takes no arguments. Returns a `status` (`file`), ' +
+            'the active `backend` (`file`), and a short `reason`.',
+        input_schema: {
+            type: 'object',
+            properties: {},
+            additionalProperties: false,
+        },
+        handler: _memoryStatusHandler,
+    },
+    list_skills: {
+        name: 'list_skills',
+        description:
+            'Enumerate every skill the server currently exposes as a ' +
+            'prompt, each with its name, description, and source. Use to ' +
+            'discover which skills are available before suggesting or ' +
+            'invoking one. Read-only manifest view, takes no arguments. ' +
+            'Returns a `count` plus a `skills` array.',
+        input_schema: {
+            type: 'object',
+            properties: {},
+            additionalProperties: false,
+        },
+        handler: _listSkillsHandler,
+    },
+    list_commands: {
+        name: 'list_commands',
+        description:
+            'Enumerate every slash command the server currently exposes ' +
+            'as a prompt, each with its name and description. Use to ' +
+            'discover available commands before routing a user request to ' +
+            'one. Read-only manifest view, takes no arguments. Returns a ' +
+            '`count` plus a `commands` array.',
+        input_schema: {
+            type: 'object',
+            properties: {},
+            additionalProperties: false,
+        },
+        handler: _listCommandsHandler,
+    },
+    list_rules: {
+        name: 'list_rules',
+        description:
+            'Enumerate every behavioral rule the server exposes as a ' +
+            'resource, each with its URI, name, and description. Use to ' +
+            'discover which rules are in effect, then fetch a body with ' +
+            '`read_resource_body` or `resources/read`. Read-only manifest ' +
+            'view, takes no arguments. Returns a `count` plus a `rules` ' +
+            'array.',
+        input_schema: {
+            type: 'object',
+            properties: {},
+            additionalProperties: false,
+        },
+        handler: _listRulesHandler,
+    },
+    read_resource_body: {
+        name: 'read_resource_body',
+        description:
+            'Fetch the rendered body of a single resource URI (rule, ' +
+            'guideline, or context document) in one call, without the ' +
+            'two-step `resources/list` + `resources/read` handshake. Use ' +
+            'when you already know the URI and want to inline its content ' +
+            'into a tool-call result. Read-only. Returns the resource ' +
+            '`uri`, `name`, `description`, and full text `body`.',
+        input_schema: {
+            type: 'object',
+            properties: {
+                uri: {
+                    type: 'string',
+                    description:
+                        'Resource URI to fetch, e.g. `rule://commit-policy`, ' +
+                        '`guideline://php/patterns/events`, or ' +
+                        '`context://authority/scope-mechanics`.',
+                },
+            },
+            required: ['uri'],
+            additionalProperties: false,
+        },
+        handler: _readResourceBodyHandler,
+    },
+};
+
+/** Render a `BuiltinTool` as kwargs for `mcp.types.Tool`. */
+export function to_mcp_tool_meta(tool: BuiltinTool): Record<string, unknown> {
+    return {
+        name: tool.name,
+        description: tool.description,
+        inputSchema: tool.input_schema,
+    };
+}
+
+// ---------------------------------------------------------------------
+// Phase 1 discovery stubs — catalog entries with no real handler.
+// Loaded at module import time. The Worker reads the same catalog via
+// `content.json` so `tools/list` returns identical metadata on both
+// transports apart from `implemented_on`.
+// ---------------------------------------------------------------------
+
+/** Closure that returns the `not_implemented` envelope for a stub. */
+function _makeStubHandler(entry: CatalogEntry, installHintValue: string): ToolHandler {
+    return async (
+        _args: Record<string, unknown>,
+        _consumerRoot: string,
+    ): Promise<Record<string, unknown>> =>
+        not_implemented_envelope(entry.name, {
+            transport: STDIO_TRANSPORT,
+            install_hint_value: installHintValue,
+        });
+}
+
+/**
+ * Build the stub registry from the catalog. ALLOWLIST wins on overlap.
+ *
+ * Returns [registry, stub_names]. `registry` contains every catalog
+ * entry not already in ALLOWLIST, each wired to a closure that emits the
+ * envelope.
+ */
+function _buildCatalogRegistry(): [Record<string, BuiltinTool>, ReadonlySet<string>] {
+    const installHintValue = _catalog_install_hint();
+    const entries = load_catalog();
+    const registry: Record<string, BuiltinTool> = {};
+    const stubNames = new Set<string>();
+    for (const entry of entries) {
+        if (entry.name in ALLOWLIST) {
+            continue;
+        }
+        registry[entry.name] = {
+            name: entry.name,
+            description: entry.description,
+            input_schema: entry.input_schema,
+            handler: _makeStubHandler(entry, installHintValue),
+        };
+        stubNames.add(entry.name);
+    }
+    return [registry, stubNames];
+}
+
+const _catalogBuild = _buildCatalogRegistry();
+export const CATALOG_STUBS: Record<string, BuiltinTool> = _catalogBuild[0];
+export const STUB_NAMES: ReadonlySet<string> = _catalogBuild[1];
+
+// Full wire-surface registry — implemented + stubs. `tools/list` reads
+// from this; `tools/call` dispatches against it.
+export const REGISTRY: Record<string, BuiltinTool> = { ...ALLOWLIST, ...CATALOG_STUBS };
+
+/**
+ * Registry view backing the MCP `tools/*` handlers.
+ *
+ * Default registry is `REGISTRY` (implemented + catalog stubs). Tests
+ * can pass a narrower object (e.g. `ALLOWLIST` alone) to isolate the
+ * implemented surface.
+ */
+export class ToolCache {
+    private _registry: Record<string, BuiltinTool>;
+
+    constructor(registry?: Record<string, BuiltinTool> | null) {
+        this._registry = { ...(registry !== undefined && registry !== null ? registry : REGISTRY) };
+    }
+
+    names(): string[] {
+        return Object.keys(this._registry).sort(_strCmp);
+    }
+
+    list(): BuiltinTool[] {
+        return this.names().map((name) => this._registry[name]!);
+    }
+
+    get(name: string): BuiltinTool | null {
+        return this._registry[name] ?? null;
+    }
+
+    /** True when `name` is a catalog stub on this cache. */
+    is_stub(name: string): boolean {
+        return STUB_NAMES.has(name) && name in this._registry;
+    }
+
+    /** Subset of `names()` whose handlers run real logic. */
+    implemented_names(): string[] {
+        return Object.keys(this._registry)
+            .filter((n) => n in ALLOWLIST)
+            .sort(_strCmp);
+    }
+
+    async dispatch(
+        name: string,
+        args: Record<string, unknown>,
+        consumerRoot?: string | null,
+    ): Promise<Record<string, unknown>> {
+        const root = _resolveConsumerRoot(consumerRoot);
+        const tool = this.get(name);
+        if (tool === null) {
+            // Sonnet's latent-demand pattern: log the unknown name before
+            // surfacing the JSON-RPC error so Phase 2 can rank the gap.
+            ToolCache._record(name, 'latent_demand', root);
+            throw new Error(`Unknown tool: ${name}`);
+        }
+        const outcome: Outcome = this.is_stub(name) ? 'stub' : 'implemented';
+        ToolCache._record(name, outcome, root);
+        return tool.handler(args ?? {}, root);
+    }
+
+    /** Best-effort JSONL write — failures never break the wire surface. */
+    private static _record(toolName: string, outcome: Outcome, consumerRoot: string): void {
+        record_call({
+            tool_name: toolName,
+            outcome,
+            transport: STDIO_TRANSPORT,
+            consumer_root: consumerRoot,
+        });
+    }
+}
+
+/** Single-line stderr enumeration of the registered tools. */
+export function boot_log_line(cache: ToolCache): string {
+    const total = cache.names().length;
+    const implemented = cache.implemented_names().length;
+    const stubs = total - implemented;
+    return (
+        `mcp-server: registered ${total} tools ` +
+        `(${implemented} implemented, ${stubs} stubs): ${_pyReprList(cache.names())}`
+    );
+}
+
+// ---------------------------------------------------------------------
+// Parity primitives
+// ---------------------------------------------------------------------
+
+/** Mirror Python truthiness for the small set of values handlers test. */
+function _pyTruthy(value: unknown): boolean {
+    if (value === undefined || value === null || value === false) {
+        return false;
+    }
+    if (value === 0 || value === '') {
+        return false;
+    }
+    if (Array.isArray(value)) {
+        return value.length > 0;
+    }
+    if (typeof value === 'object') {
+        return Object.keys(value as Record<string, unknown>).length > 0;
+    }
+    return Boolean(value);
+}
+
+/** Mirror Python `int(x)` for the header `v` field (int-or-stringy-int). */
+function _pyInt(value: unknown): number {
+    if (typeof value === 'number') {
+        return Math.trunc(value);
+    }
+    if (typeof value === 'string') {
+        const n = Number.parseInt(value, 10);
+        return Number.isNaN(n) ? 0 : n;
+    }
+    return 0;
+}
+
+/** Mirror Python `str.__lt__` (UTF-16 code-unit ordering matches CPython for BMP). */
+function _strCmp(a: string, b: string): number {
+    return a < b ? -1 : a > b ? 1 : 0;
+}
+
+/** pathlib `Path` ordering: compare path COMPONENTS lexicographically. */
+function _pathlibCompare(a: string, b: string): number {
+    const pa = a.split(path.sep);
+    const pb = b.split(path.sep);
+    const n = Math.min(pa.length, pb.length);
+    for (let i = 0; i < n; i++) {
+        const c = _strCmp(pa[i]!, pb[i]!);
+        if (c !== 0) {
+            return c;
+        }
+    }
+    return pa.length - pb.length;
+}
+
+/** Mirror Python `repr(list_of_str)` for the boot-log enumeration. */
+function _pyReprList(values: string[]): string {
+    return `[${values.map((v) => `'${v}'`).join(', ')}]`;
+}