@pugi/cli 0.1.0-beta.24 → 0.1.0-beta.26

package/dist/core/hooks/registry.js

@@ -0,0 +1,213 @@
+/**
+ * Pugi hooks MVP — registry (Leak L12, first pass).
+ *
+ * Reads `<home>/hooks-mvp.json` and validates its shape with Zod. The
+ * file uses the Claude Code-style nested config:
+ *
+ *   {
+ *     "hooks": {
+ *       "SessionStart": [{ "command": "echo session-start" }],
+ *       "PreToolUse":   [{ "matcher": "bash", "command": "echo bash-pre", "blocking": true }]
+ *     }
+ *   }
+ *
+ * Schema constraints:
+ *   - Each hook entry MUST have a `command` (non-empty string).
+ *   - `matcher` is optional. Defaults to `*` (any tool / any payload).
+ *     For tool events, `matcher` is compared against the tool name.
+ *     For non-tool events (SessionStart), `matcher` is ignored.
+ *   - `timeoutMs` is optional. Defaults to 30 000 ms (per task spec).
+ *     Capped at 60 000 ms to prevent operator-defined deadlocks.
+ *   - `blocking` is optional. When true AND the hook exits non-zero,
+ *     the registry surfaces an `anyBlocked: true` outcome so the
+ *     caller can refuse the originating action. Only honoured for
+ *     `PreToolUse` in the MVP — other events log but do not block.
+ *
+ * Failure modes:
+ *   - File missing -> the registry is `empty()`. `list()` returns []
+ *     and `fire()` is a no-op. This matches the Claude Code default
+ *     (hooks are opt-in).
+ *   - File present but invalid JSON / fails schema -> `load()` throws.
+ *     The CLI surface (`pugi hooks doctor`) reports the error
+ *     verbatim so the operator can fix the config.
+ *
+ * Brand voice: ASCII only, no emoji, no em-dashes.
+ */
+import { existsSync, readFileSync } from 'node:fs';
+import { homedir } from 'node:os';
+import { resolve } from 'node:path';
+import { z } from 'zod';
+import { ALL_HOOK_EVENTS_V2 } from './events.js';
+/** Default per-hook timeout when the operator does not set `timeoutMs`. */
+export const DEFAULT_HOOK_TIMEOUT_MS = 30_000;
+/** Hard upper bound on `timeoutMs`. Prevents config-defined deadlocks. */
+export const MAX_HOOK_TIMEOUT_MS = 60_000;
+const hookEntrySchema = z
+    .object({
+    /**
+     * Tool-name matcher. `*` matches any tool. Plain strings match
+     * exactly (no glob in the MVP — fast-follow widens to glob). Ignored
+     * for non-tool events such as `SessionStart`.
+     */
+    matcher: z.string().min(1).optional(),
+    /** Shell command. Spawned via `/bin/sh -c <command>`. */
+    command: z.string().min(1),
+    /** Per-hook timeout override. Defaults to 30 000 ms. */
+    timeoutMs: z.number().int().positive().max(MAX_HOOK_TIMEOUT_MS).optional(),
+    /**
+     * When true, a non-zero exit code from this hook blocks the
+     * originating action (currently `PreToolUse` only). Other events
+     * log the exit but do not block.
+     */
+    blocking: z.boolean().optional(),
+})
+    .strict();
+const hookEventEnum = z.enum([
+    'SessionStart',
+    'PreToolUse',
+    'PostToolUse',
+    'UserPromptSubmit',
+    'Stop',
+    'SubagentStop',
+    'PreCompact',
+    'Notification',
+]);
+const hooksFileSchema = z
+    .object({
+    hooks: z.record(hookEventEnum, z.array(hookEntrySchema)).default({}),
+})
+    .strict();
+/** Default config file location — `~/.pugi/hooks-mvp.json`. */
+export function defaultHooksMvpPath(home) {
+    const root = home ?? process.env.PUGI_HOME ?? resolve(homedir(), '.pugi');
+    return resolve(root, 'hooks-mvp.json');
+}
+/**
+ * In-memory snapshot of the operator's `hooks-mvp.json`. Construct
+ * via `loadHooksConfig(path)` — `new HooksConfig()` is intentionally
+ * not exported so all production code paths go through the loader.
+ */
+export class HooksConfig {
+    path;
+    entries;
+    constructor(path, entries) {
+        this.path = path;
+        this.entries = entries;
+    }
+    /** Absolute path of the config file this snapshot was loaded from. */
+    configPath() {
+        return this.path;
+    }
+    /** All hooks declared for a given event. Returns [] when none. */
+    list(event) {
+        return this.entries[event] ?? [];
+    }
+    /**
+     * Hooks that match the (event, toolName?) tuple. For tool events
+     * (`PreToolUse`), `matcher` is compared against the tool name with
+     * `*` matching any. For non-tool events, all entries are returned
+     * regardless of `matcher`.
+     */
+    listMatching(event, toolName) {
+        const all = this.list(event);
+        if (!isToolEvent(event))
+            return all;
+        return all.filter((entry) => matchesTool(entry.matcher, toolName));
+    }
+    /** Flat list of (event, entry) pairs across every configured event. */
+    flatten() {
+        const out = [];
+        for (const event of ALL_HOOK_EVENTS_V2) {
+            for (const entry of this.list(event)) {
+                out.push({ event, entry });
+            }
+        }
+        return out;
+    }
+    /** True iff at least one hook is registered for any event. */
+    isEmpty() {
+        return this.flatten().length === 0;
+    }
+    /** A no-op snapshot used when the config file is absent. */
+    static empty(path) {
+        return new HooksConfig(path, {});
+    }
+}
+/**
+ * Load + validate `hooks-mvp.json`. Returns a no-op snapshot when the
+ * file is absent. Throws on invalid JSON or schema violations — the
+ * caller is expected to surface the error to the operator via
+ * `pugi hooks doctor`.
+ *
+ * Contract (non-null invariant): this function ALWAYS returns a
+ * `HooksConfig` instance. It never returns `null` / `undefined`. When
+ * the config file is missing, callers receive `HooksConfig.empty(path)`
+ * — a truthy snapshot for which `isEmpty()` returns `true` and `list()`
+ * returns `[]`. Callers may safely chain `.isEmpty()` without a null
+ * guard. Asserted by `registry-empty.spec.ts`.
+ */
+export function loadHooksConfig(pathOverride) {
+    const path = pathOverride ?? defaultHooksMvpPath();
+    if (!existsSync(path)) {
+        return HooksConfig.empty(path);
+    }
+    let raw;
+    try {
+        raw = readFileSync(path, 'utf8');
+    }
+    catch (error) {
+        throw new Error(`pugi hooks: cannot read ${path}: ${error.message}`);
+    }
+    let parsed;
+    try {
+        parsed = JSON.parse(raw);
+    }
+    catch (error) {
+        throw new Error(`pugi hooks: ${path} is not valid JSON: ${error.message}`);
+    }
+    const result = hooksFileSchema.safeParse(parsed);
+    if (!result.success) {
+        const issues = result.error.issues
+            .map((issue) => `${issue.path.join('.') || '<root>'} ${issue.message}`)
+            .join('; ');
+        throw new Error(`pugi hooks: ${path} failed schema validation: ${issues}`);
+    }
+    // Zod's `z.record(enum, value)` returns `Partial<Record<...>>` shape
+    // — keys that the operator did not include are `undefined`. Coerce
+    // explicitly into the same shape `HooksConfig` expects.
+    const entries = {};
+    for (const event of ALL_HOOK_EVENTS_V2) {
+        const list = result.data.hooks[event];
+        if (list && list.length > 0) {
+            entries[event] = list;
+        }
+    }
+    return new HooksConfig(path, entries);
+}
+/**
+ * `isToolEvent(event)` -> true for events where `matcher` is compared
+ * against the tool name. SessionStart / Stop / Notification / etc. do
+ * not have an associated tool so matcher is ignored.
+ */
+export function isToolEvent(event) {
+    return event === 'PreToolUse' || event === 'PostToolUse';
+}
+/**
+ * Tool-name match grammar for the MVP. Intentionally narrow:
+ *   - matcher missing or `*` -> matches any tool name (and the
+ *     `bash`/`read`/... shape).
+ *   - matcher === toolName    -> exact match.
+ *
+ * Fast-follow widens this to glob via `picomatch` so operators can
+ * write `mcp__*` patterns. Deliberately not pulling in a glob lib for
+ * the MVP — the narrow grammar is enough to land the surface and the
+ * test matrix stays small.
+ */
+export function matchesTool(matcher, toolName) {
+    if (!matcher || matcher === '*')
+        return true;
+    if (!toolName)
+        return false;
+    return matcher === toolName;
+}
+//# sourceMappingURL=registry.js.map

package/dist/core/hooks/runner.js

@@ -0,0 +1,236 @@
+/**
+ * Pugi hooks MVP — runner (Leak L12, first pass).
+ *
+ * Spawns the shell command declared in `hooks-mvp.json`, applies the
+ * timeout watchdog, captures stdout / stderr, and surfaces a
+ * structured result. Two events are wired in the MVP (SessionStart,
+ * PreToolUse); the runner itself is event-agnostic so the fast-follow
+ * PR can attach the remaining 6 events without changing this file.
+ *
+ * Safety properties:
+ *   - 30 s default timeout (per task spec); SIGTERM then SIGKILL with
+ *     a 2 s grace window.
+ *   - 1 MiB output cap per stream — a misbehaving hook (`yes`) cannot
+ *     OOM the parent CLI by buffering unbounded data.
+ *   - Spawn failures are caught + logged; the session never crashes
+ *     because of a missing binary or a syntax error in the command.
+ *   - Hook errors are atomic-appended to `<workspaceRoot>/.pugi/logs/
+ *     hooks.log`. Multiple sessions can write concurrently without
+ *     interleaving because `appendFileSync` opens with O_APPEND.
+ *
+ * Brand voice: ASCII only.
+ */
+import { spawn } from 'node:child_process';
+import { appendFileSync, existsSync, mkdirSync } from 'node:fs';
+import { resolve } from 'node:path';
+import { DEFAULT_HOOK_TIMEOUT_MS, isToolEvent, } from './registry.js';
+const HOOK_STREAM_CAP_BYTES = 1024 * 1024;
+const SIGKILL_GRACE_MS = 2_000;
+/**
+ * Fire every matching hook for `event` sequentially. Sequential (not
+ * parallel) is the intentional default — operators frequently chain
+ * `git add` -> `eslint --fix` style hooks that would race otherwise.
+ * Returns a `HookFireOutcome` with the per-invocation results.
+ */
+export async function fireHooks(opts) {
+    const { config, event, payload, toolName, workspaceRoot, env } = opts;
+    const matching = config.listMatching(event, toolName);
+    if (matching.length === 0) {
+        return { event, results: [], anyBlocked: false };
+    }
+    const logger = workspaceRoot ? new HookLogger(workspaceRoot) : undefined;
+    const results = [];
+    let anyBlocked = false;
+    for (const entry of matching) {
+        const timeoutMs = entry.timeoutMs ?? DEFAULT_HOOK_TIMEOUT_MS;
+        const result = await executeOne(entry.command, payload, timeoutMs, env);
+        // Blocking semantics only honored for PreToolUse in the MVP.
+        // Other events can declare `blocking: true` but the runner just
+        // logs that intent — it does NOT short-circuit. The fast-follow
+        // PR threads PostToolUse + UserPromptSubmit blocking through.
+        const blockable = entry.blocking === true && event === 'PreToolUse';
+        const blocked = blockable && !result.ok;
+        if (blocked) {
+            anyBlocked = true;
+            result.blocked = true;
+            result.blockSentinel = `HOOK_BLOCKED: ${truncate(entry.command, 80)} exited ${result.exitCode}`;
+        }
+        if (logger && !result.ok) {
+            logger.recordFailure(event, entry.command, result);
+        }
+        results.push(result);
+    }
+    return { event, results, anyBlocked };
+}
+async function executeOne(command, payload, timeoutMs, env) {
+    const startedAt = Date.now();
+    return new Promise((resolvePromise) => {
+        const payloadJson = JSON.stringify(payload);
+        const childEnv = {
+            ...(env ?? process.env),
+            PUGI_HOOK_PAYLOAD: payloadJson,
+            PUGI_HOOK_EVENT: payload.event,
+            PUGI_HOOK_SESSION_ID: payload.sessionId,
+        };
+        const child = spawn('/bin/sh', ['-c', command], {
+            env: childEnv,
+            stdio: ['pipe', 'pipe', 'pipe'],
+        });
+        const state = {
+            stdout: '',
+            stderr: '',
+            killedForTimeout: false,
+            killedForStreamCap: false,
+        };
+        const escalateKill = () => {
+            if (state.sigKillTimer)
+                return;
+            state.sigKillTimer = setTimeout(() => {
+                if (!child.killed)
+                    child.kill('SIGKILL');
+            }, SIGKILL_GRACE_MS);
+            if (state.sigKillTimer.unref)
+                state.sigKillTimer.unref();
+        };
+        const enforceStreamCap = () => {
+            if (state.killedForStreamCap)
+                return;
+            if (state.stdout.length + state.stderr.length <= HOOK_STREAM_CAP_BYTES)
+                return;
+            state.killedForStreamCap = true;
+            child.kill('SIGTERM');
+            escalateKill();
+        };
+        child.stdout?.on('data', (chunk) => {
+            if (state.killedForStreamCap)
+                return;
+            state.stdout += chunk.toString('utf8');
+            enforceStreamCap();
+        });
+        child.stderr?.on('data', (chunk) => {
+            if (state.killedForStreamCap)
+                return;
+            state.stderr += chunk.toString('utf8');
+            enforceStreamCap();
+        });
+        // Best-effort stdin payload — hook scripts that want to read it can
+        // (e.g. `jq .`); scripts that ignore stdin will EPIPE on our write
+        // which we swallow because the env var carries the same data.
+        if (child.stdin) {
+            child.stdin.on('error', () => {
+                // EPIPE is benign — see above.
+            });
+            child.stdin.end(payloadJson);
+        }
+        const timer = setTimeout(() => {
+            state.killedForTimeout = true;
+            child.kill('SIGTERM');
+            escalateKill();
+        }, timeoutMs);
+        if (timer.unref)
+            timer.unref();
+        child.on('error', (error) => {
+            clearTimeout(timer);
+            if (state.sigKillTimer)
+                clearTimeout(state.sigKillTimer);
+            resolvePromise({
+                command: truncate(command, 200),
+                exitCode: -1,
+                stdoutBytes: state.stdout.length,
+                stderrBytes: state.stderr.length,
+                elapsedMs: Date.now() - startedAt,
+                ok: false,
+                blocked: false,
+                timedOut: false,
+                // No blockSentinel here — spawn errors are not the same as
+                // blocking-failure semantics. The caller logs them generically.
+            });
+        });
+        child.on('close', (code, signal) => {
+            clearTimeout(timer);
+            if (state.sigKillTimer)
+                clearTimeout(state.sigKillTimer);
+            let exitCode;
+            if (code !== null) {
+                exitCode = code;
+            }
+            else if (signal === 'SIGTERM') {
+                exitCode = -15;
+            }
+            else if (signal === 'SIGKILL') {
+                exitCode = -9;
+            }
+            else {
+                exitCode = -1;
+            }
+            const ok = exitCode === 0 &&
+                !state.killedForTimeout &&
+                !state.killedForStreamCap;
+            resolvePromise({
+                command: truncate(command, 200),
+                exitCode,
+                stdoutBytes: state.stdout.length,
+                stderrBytes: state.stderr.length,
+                elapsedMs: Date.now() - startedAt,
+                ok,
+                blocked: false,
+                timedOut: state.killedForTimeout,
+            });
+        });
+    });
+}
+/**
+ * Append-only failure log at `<workspaceRoot>/.pugi/logs/hooks.log`.
+ * Each line is a JSON record so log scrapers can `jq` over it.
+ */
+class HookLogger {
+    path;
+    prepared = false;
+    constructor(workspaceRoot) {
+        this.path = resolve(workspaceRoot, '.pugi', 'logs', 'hooks.log');
+    }
+    recordFailure(event, command, result) {
+        this.prepareDir();
+        const line = JSON.stringify({
+            ts: new Date().toISOString(),
+            event,
+            command: truncate(command, 200),
+            exitCode: result.exitCode,
+            timedOut: result.timedOut,
+            elapsedMs: result.elapsedMs,
+            stdoutBytes: result.stdoutBytes,
+            stderrBytes: result.stderrBytes,
+            toolEvent: isToolEvent(event),
+        });
+        try {
+            appendFileSync(this.path, `${line}\n`, 'utf8');
+        }
+        catch {
+            // Logging is best-effort — the session must not crash when the
+            // disk is full or the directory is read-only. The runner has
+            // already returned the result; dropping the log line is the
+            // safe fallback.
+        }
+    }
+    prepareDir() {
+        if (this.prepared)
+            return;
+        const dir = resolve(this.path, '..');
+        if (!existsSync(dir)) {
+            try {
+                mkdirSync(dir, { recursive: true });
+            }
+            catch {
+                // ignored — appendFileSync will surface a fresh error on the
+                // write path, which we also swallow.
+            }
+        }
+        this.prepared = true;
+    }
+}
+function truncate(value, max) {
+    if (value.length <= max)
+        return value;
+    return `${value.slice(0, max - 3)}...`;
+}
+//# sourceMappingURL=runner.js.map

package/dist/core/lsp/cache.js

@@ -0,0 +1,105 @@
+/**
+ * Per-process LSP client cache — Leak L15.
+ *
+ * The α7.7 `runtime/commands/lsp.ts` CLI surface spawns one LSP server
+ * per invocation and stops it at the end. That is correct for the
+ * one-shot `pugi lsp hover ...` shape but wrong for L15's
+ * post-edit auto-diagnostics: every successful `edit`/`write` would
+ * otherwise pay the ~2-3s cold-start of `typescript-language-server`,
+ * which is unusable inside an agent loop.
+ *
+ * This module owns a singleton map keyed by `LspLanguage` with lazy
+ * initialization (`getOrStart`). The first edit of a TS file in a
+ * session pays cold-start; every subsequent edit of any TS/TSX file
+ * in the same workspace reuses the warm client.
+ *
+ * Lifecycle:
+ *   - `getOrStart(lang, cwd)` — spawn if missing, return cached otherwise.
+ *   - `stopAll()` — graceful shutdown of every cached client. Called from
+ *     `runCli` exit so a Ctrl-C never leaves zombie LSP processes behind.
+ *   - `reset()` — test-only escape hatch, drops the cache without
+ *     touching child processes (specs inject stubs that own their own
+ *     lifecycle).
+ *
+ * Failure handling: a startup failure is NOT cached. The next call
+ * tries again. This keeps the cache from poisoning a session when the
+ * operator installs the missing LSP binary mid-session and re-edits.
+ *
+ * Brand voice: ASCII only, no emoji, no banned words.
+ */
+import { isLspLanguageDisabled, startLspClient, } from './client.js';
+const cache = new Map();
+/**
+ * Return a warm client for `lang`, starting one if needed. The
+ * workspace `cwd` is captured at cache-insert time; if a subsequent
+ * call asks for the same language with a different `cwd` we tear
+ * down the old client and start a fresh one. This handles the
+ * agent-worktree case where the same process hops between workspace
+ * roots inside one Node lifetime.
+ */
+export async function getOrStartLspClient(lang, opts) {
+    // β7 L9: respect the per-language disable toggle BEFORE we attempt to
+    // spawn. The check is cheap and keeps the disabled path from paying
+    // the `npx --yes` warmup cost on first use.
+    if (isLspLanguageDisabled(lang, opts.lspSettings)) {
+        return {
+            ok: false,
+            reason: 'lsp_disabled',
+            detail: `${lang} is disabled via .pugi/settings.json::lsp`,
+        };
+    }
+    const existing = cache.get(lang);
+    if (existing && existing.cwd === opts.cwd) {
+        return { ok: true, client: existing.client };
+    }
+    if (existing && existing.cwd !== opts.cwd) {
+        // Workspace switched — stop the old client and fall through to spawn.
+        try {
+            await existing.client.stop();
+        }
+        catch {
+            // best effort; stop() is idempotent + swallow-safe
+        }
+        cache.delete(lang);
+    }
+    const result = await startLspClient(lang, opts);
+    if (!result.ok) {
+        return { ok: false, reason: result.reason, detail: result.detail };
+    }
+    cache.set(lang, { client: result.value, cwd: opts.cwd });
+    return { ok: true, client: result.value };
+}
+/** Look up the cached client without starting one. Returns undefined when missing. */
+export function peekLspClient(lang) {
+    return cache.get(lang)?.client;
+}
+/** Snapshot of currently-cached languages — used by `pugi lsp status` debug output. */
+export function listCachedLanguages() {
+    return Array.from(cache.keys());
+}
+/**
+ * Stop every cached client and clear the cache. Called from `runCli`
+ * exit and from specs that own the lifecycle of their stub servers.
+ */
+export async function stopAllLspClients() {
+    const snapshot = Array.from(cache.values());
+    cache.clear();
+    await Promise.all(snapshot.map(async (entry) => {
+        try {
+            await entry.client.stop();
+        }
+        catch {
+            // best effort — shutting down anyway
+        }
+    }));
+}
+/**
+ * Test-only: drop the cache map WITHOUT calling stop on the children.
+ * Specs that inject stub servers manage the stub lifecycle themselves;
+ * this lets a spec swap a stub mid-test without the cache holding a
+ * stale reference to a torn-down process.
+ */
+export function __resetLspCacheForTests() {
+    cache.clear();
+}
+//# sourceMappingURL=cache.js.map

package/dist/core/lsp/language-detect.js

@@ -0,0 +1,66 @@
+/**
+ * Language-from-extension detection — Leak L15.
+ *
+ * Single source of truth for "given a file path, which `LspLanguage`
+ * slug do we route to". The α7.7 `runtime/commands/lsp.ts` shipped its
+ * own inline `inferLanguage` switch; L15 (post-edit auto-diagnostics)
+ * needs the same lookup from `core/engine/tool-bridge.ts`, so we lift
+ * the table into a dedicated module to avoid a second copy drifting
+ * out of sync.
+ *
+ * Returning `undefined` is the calling code's signal to silently skip
+ * LSP — an unsupported extension is NOT an error, it just means "no
+ * diagnostics for this file". The tool-bridge hook treats this as a
+ * no-op envelope tail.
+ *
+ * Adding a new language requires THREE coordinated changes:
+ *   1. Add the `LspLanguage` slug + server descriptor in `client.ts`.
+ *   2. Map its extensions here.
+ *   3. Add a `lsp-language-matrix` spec row exercising the new ext.
+ *
+ * Brand voice: ASCII only, no emoji, no banned words.
+ */
+import { extname } from 'node:path';
+/**
+ * Lower-case extension (including the dot) → LSP language slug.
+ * Mirror of the switch in `runtime/commands/lsp.ts::inferLanguage`.
+ * The table form lets tests assert coverage and lets new languages
+ * land with one edit instead of two.
+ */
+export const EXTENSION_TO_LANGUAGE = {
+    '.ts': 'ts',
+    '.tsx': 'ts',
+    '.mts': 'ts',
+    '.cts': 'ts',
+    '.js': 'js',
+    '.jsx': 'js',
+    '.mjs': 'js',
+    '.cjs': 'js',
+    '.py': 'py',
+    '.pyi': 'py',
+    '.go': 'go',
+    '.rs': 'rust',
+};
+/**
+ * Infer the `LspLanguage` for a workspace-relative or absolute path.
+ * Returns `undefined` for unmapped extensions — the caller decides
+ * whether that is silently skipped (post-edit hook) or surfaced as
+ * `language_unsupported` (`pugi lsp` CLI).
+ */
+export function languageForFile(file) {
+    const ext = extname(file).toLowerCase();
+    if (!ext)
+        return undefined;
+    return EXTENSION_TO_LANGUAGE[ext];
+}
+/**
+ * Return every extension currently mapped to the given language.
+ * Used by the matrix spec to assert coverage without re-typing the
+ * extension list.
+ */
+export function extensionsForLanguage(lang) {
+    return Object.entries(EXTENSION_TO_LANGUAGE)
+        .filter(([, value]) => value === lang)
+        .map(([ext]) => ext);
+}
+//# sourceMappingURL=language-detect.js.map