@pugi/cli 0.1.0-beta.100 → 0.1.0-beta.101

package/dist/core/codegraph/watcher.js

@@ -0,0 +1,440 @@
+/**
+ * Pugi local symbol index - chokidar file watcher (PR L2).
+ *
+ * Live auto-sync layer на top of the parser (PR L1) и DB (PR L0). When
+ * the operator runs `pugi index --watch`, this module:
+ *
+ *   1. Spins up a chokidar watcher rooted at `workspaceRoot`.
+ *   2. Filters events to the v1 parseable extensions
+ *      (`PARSEABLE_EXTENSIONS` from parser.ts).
+ *   3. Coalesces bursts via a 2-second debounce (configurable; tests
+ *      pass `debounceMs: 50` to keep specs fast).
+ *   4. On flush: per file -> `deleteFile` cascade -> `parseFile` ->
+ *      `insertSymbols` -> resolve `pendingEdges` -> `insertEdges`.
+ *   5. On `unlink`: `deleteFile` only (cascade removes symbols + edges
+ *      via the schema's FK ON DELETE CASCADE).
+ *
+ * Architectural decisions:
+ *
+ *  - **No nested transactions.** `deleteFile` / `insertSymbols` /
+ *    `insertEdges` each open + commit their own BEGIN block. We call
+ *    them sequentially and rely on per-call atomicity. A higher-level
+ *    SAVEPOINT wrapper was considered и rejected: chokidar batches are
+ *    small (typically 1-5 files) and the per-file work is already
+ *    transactional, so the extra level adds complexity without a
+ *    customer-visible win.
+ *
+ *  - **Two-step edge resolution.** `parseFile` returns `pendingEdges`
+ *    keyed by name + scope. We resolve them after `insertSymbols`
+ *    populates ids, using the shared `resolvePendingEdges` helper.
+ *    Cross-file targets are looked up via a SQL `SELECT id FROM
+ *    symbols WHERE name = ?` closure - the live DB IS the cross-file
+ *    map for the watcher. Reindex builds an in-memory map because it
+ *    sees every file in one pass; the watcher sees files one delta at
+ *    a time.
+ *
+ *  - **Path normalization.** Chokidar emits absolute or cwd-relative
+ *    paths depending on options. We always work in workspace-relative
+ *    POSIX form (`path.relative(root, abs).split(sep).join('/')`)
+ *    because that is what `reindex.ts` writes to `Symbol.file` and
+ *    `files.path`. Mismatched separators on Windows would break the
+ *    FK + the `deleteFile` lookup.
+ *
+ *  - **Exclude list parity.** We use the EXACT same `PRUNE_DIRS` as
+ *    `reindex.ts` (extended slightly with the v1 extensions the spec
+ *    pinned). Divergence between the walker and the watcher would
+ *    silently produce different indexes depending on entry point.
+ *
+ *  - **In-flight queue at close.** SIGINT / SIGTERM trigger
+ *    `close()`. If a debounce timer is still pending, we flush it
+ *    synchronously before tearing down chokidar - dropping a pending
+ *    edit would leave the index out of sync until the next manual
+ *    `pugi index` run.
+ *
+ *  - **Initial scan is the operator's job.** `ignoreInitial: true`
+ *    means we do NOT re-index the workspace on startup. The CLI
+ *    command runs `reindexWorkspace` first (covered by PR L1) и THEN
+ *    starts the watcher; this avoids a double-pass over the
+ *    workspace.
+ */
+import { existsSync } from 'node:fs';
+import { readFile } from 'node:fs/promises';
+import { createHash } from 'node:crypto';
+import { relative, resolve, sep, extname } from 'node:path';
+import chokidar from 'chokidar';
+import { deleteFile, insertEdges, insertSymbols, upsertFile, } from './db.js';
+import { PARSEABLE_EXTENSIONS, parseFile } from './parser.js';
+import { buildPerFileNameMap, resolvePendingEdges, } from './resolve-edges.js';
+/**
+ * Directories pruned from the watch. EXACT mirror of
+ * `reindex.ts:PRUNE_DIRS` so the walker и watcher agree on what
+ * "workspace" means. Touching this list requires touching both
+ * sites.
+ */
+const PRUNE_DIRS = Object.freeze([
+    'node_modules',
+    '.git',
+    'dist',
+    'build',
+    '.next',
+    '.turbo',
+    '.cache',
+    'coverage',
+    '.pugi',
+]);
+/** Default debounce window between burst-of-events and flush. */
+const DEFAULT_DEBOUNCE_MS = 2000;
+/**
+ * Start a chokidar watcher rooted at `workspaceRoot` against the
+ * caller-owned IndexDB. Returns a handle the caller invokes к stop the
+ * watcher cleanly.
+ *
+ * The function is sync - the watcher is "started" the moment chokidar
+ * exists, even if it has не yet emitted `ready`. The first `add`/
+ * `change` events flow into the debounce queue immediately.
+ */
+export function startWatcher(opts) {
+    const root = resolve(opts.workspaceRoot);
+    if (!existsSync(root)) {
+        throw new Error(`startWatcher: workspaceRoot does not exist: ${root}`);
+    }
+    const debounceMs = Math.max(0, opts.debounceMs ?? DEFAULT_DEBOUNCE_MS);
+    const writeStderr = opts.writeStderr ?? ((text) => void process.stderr.write(text));
+    const watchFn = opts.chokidarWatchFn ?? chokidar.watch;
+    const quiet = opts.quiet === true;
+    const stats = {
+        added: 0,
+        changed: 0,
+        removed: 0,
+        reparsed: 0,
+        edgesInserted: 0,
+        edgesOrphaned: 0,
+        parseErrors: 0,
+        watcherErrors: 0,
+    };
+    /** Workspace-relative POSIX path. */
+    const toRelPosix = (abs) => {
+        const rel = relative(root, abs);
+        if (rel.length === 0 || rel.startsWith('..'))
+            return '';
+        return rel.split(sep).join('/');
+    };
+    /**
+     * Defense in depth: chokidar matches on directory globs, but a
+     * watched-parent-dir event still might fire для excluded subpaths
+     * on some platforms. Re-check the relative path against the prune
+     * list before doing any work.
+     */
+    const isExcluded = (relPosix) => {
+        if (relPosix.length === 0)
+            return true;
+        const parts = relPosix.split('/');
+        for (const p of parts) {
+            if (PRUNE_DIRS.includes(p))
+                return true;
+        }
+        return false;
+    };
+    const isParseable = (relPosix) => {
+        const ext = extname(relPosix).toLowerCase();
+        return PARSEABLE_EXTENSIONS.includes(ext);
+    };
+    // ---- queue + debounce -------------------------------------------------
+    /** Files queued для re-parse on the next flush. */
+    const dirty = new Set();
+    /** Files queued для unlink-cascade on the next flush. */
+    const removed = new Set();
+    let timer = null;
+    /** Tracks the in-progress flush so close() can await it. */
+    let flushing = null;
+    /** Once closed, ignore further events. */
+    let closed = false;
+    const armDebounce = () => {
+        if (closed)
+            return;
+        if (timer)
+            clearTimeout(timer);
+        timer = setTimeout(() => {
+            timer = null;
+            void flush();
+        }, debounceMs);
+    };
+    /**
+     * Drain the queue. Sequential per-file work - the parser is async,
+     * SQLite handles are sync but already wrapped in per-call BEGIN/
+     * COMMIT inside db.ts, so concurrency would only add lock contention
+     * without throughput gain on a single-process watcher.
+     */
+    const flush = async () => {
+        // If a flush is already in-flight, queue behind it. We serialize
+        // so the dirty Set drained by one flush never collides with one
+        // mutated by another.
+        if (flushing) {
+            await flushing;
+            // Nothing new since the previous flush picked it up? Bail.
+            if (dirty.size === 0 && removed.size === 0)
+                return;
+        }
+        flushing = doFlush().finally(() => {
+            flushing = null;
+        });
+        return flushing;
+    };
+    const doFlush = async () => {
+        const snapshotDirty = Array.from(dirty);
+        dirty.clear();
+        const snapshotRemoved = Array.from(removed);
+        removed.clear();
+        // Process removals first - if a file was deleted and re-added in
+        // the same window we want the add half to write fresh rows.
+        for (const relPath of snapshotRemoved) {
+            try {
+                deleteFile(opts.db, relPath);
+            }
+            catch (err) {
+                const message = err instanceof Error ? err.message : String(err);
+                if (!quiet) {
+                    writeStderr(`pugi index --watch: deleteFile(${relPath}) failed: ${message}\n`);
+                }
+            }
+        }
+        for (const relPath of snapshotDirty) {
+            // Drop файлы that were also removed in this batch - the unlink
+            // above already wiped them.
+            if (snapshotRemoved.includes(relPath))
+                continue;
+            const abs = resolve(root, relPath);
+            let bytes;
+            try {
+                bytes = await readFile(abs, 'utf8');
+            }
+            catch {
+                // Disappeared between event и flush - treat as unlink.
+                try {
+                    deleteFile(opts.db, relPath);
+                }
+                catch (err) {
+                    const message = err instanceof Error ? err.message : String(err);
+                    if (!quiet) {
+                        writeStderr(`pugi index --watch: post-disappear deleteFile(${relPath}) failed: ${message}\n`);
+                    }
+                }
+                continue;
+            }
+            let result;
+            try {
+                result = await parseFile(abs, {
+                    relPath,
+                    quiet: true,
+                    sourceOverride: bytes,
+                });
+            }
+            catch (err) {
+                stats.parseErrors += 1;
+                if (!quiet) {
+                    const message = err instanceof Error ? err.message : String(err);
+                    writeStderr(`pugi index --watch: parse(${relPath}) failed: ${message}\n`);
+                }
+                continue;
+            }
+            // Clear prior rows for this file before re-inserting. Cheap
+            // даже if the file is new - the lookup hits the `idx_symbols_file`
+            // index и returns zero rows.
+            try {
+                deleteFile(opts.db, relPath);
+            }
+            catch (err) {
+                const message = err instanceof Error ? err.message : String(err);
+                if (!quiet) {
+                    writeStderr(`pugi index --watch: deleteFile(${relPath}) failed: ${message}\n`);
+                }
+                continue;
+            }
+            if (result.symbols.length === 0) {
+                // Record the file fingerprint so a future "did this change?"
+                // query can short-circuit. Only when we got далеко enough to
+                // recognise the language - skipping unsupported extensions
+                // earlier means we never see them here.
+                if (result.language !== null) {
+                    try {
+                        upsertFile(opts.db, {
+                            path: relPath,
+                            sha256: sha256(bytes),
+                            lastIndexedAt: new Date().toISOString(),
+                            symbolCount: 0,
+                        });
+                    }
+                    catch (err) {
+                        const message = err instanceof Error ? err.message : String(err);
+                        if (!quiet) {
+                            writeStderr(`pugi index --watch: upsertFile(${relPath}) failed: ${message}\n`);
+                        }
+                    }
+                }
+                stats.reparsed += 1;
+                continue;
+            }
+            let ids;
+            try {
+                ids = insertSymbols(opts.db, result.symbols);
+            }
+            catch (err) {
+                const message = err instanceof Error ? err.message : String(err);
+                if (!quiet) {
+                    writeStderr(`pugi index --watch: insertSymbols(${relPath}) failed: ${message}\n`);
+                }
+                continue;
+            }
+            const perFile = buildPerFileNameMap(result.symbols, ids);
+            const { resolved, orphaned } = resolvePendingEdges(result.pendingEdges, perFile, (name) => crossFileLookup(opts.db, name));
+            stats.edgesOrphaned += orphaned;
+            if (resolved.length > 0) {
+                try {
+                    insertEdges(opts.db, resolved);
+                    stats.edgesInserted += resolved.length;
+                }
+                catch (err) {
+                    const message = err instanceof Error ? err.message : String(err);
+                    if (!quiet) {
+                        writeStderr(`pugi index --watch: insertEdges(${relPath}) failed: ${message}\n`);
+                    }
+                }
+            }
+            try {
+                upsertFile(opts.db, {
+                    path: relPath,
+                    sha256: sha256(bytes),
+                    lastIndexedAt: new Date().toISOString(),
+                    symbolCount: ids.length,
+                });
+            }
+            catch (err) {
+                const message = err instanceof Error ? err.message : String(err);
+                if (!quiet) {
+                    writeStderr(`pugi index --watch: upsertFile(${relPath}) failed: ${message}\n`);
+                }
+            }
+            stats.reparsed += 1;
+        }
+    };
+    // ---- chokidar wiring --------------------------------------------------
+    // We watch the workspace root recursively and let our own filter
+    // drop unparseable extensions + excluded directories. Chokidar's
+    // built-in `ignored` accepts glob/regex/function; we use a function
+    // form so we can re-use the prune list verbatim instead of
+    // duplicating it as a glob array.
+    const ignored = (p) => {
+        // chokidar may pass absolute or workspace-relative paths depending
+        // on platform и `cwd` option; normalize then reuse the predicate.
+        const abs = resolve(p);
+        const rel = toRelPosix(abs);
+        if (rel.length === 0)
+            return false; // root itself - keep watching
+        return isExcluded(rel);
+    };
+    const watcher = watchFn(root, {
+        ignored,
+        ignoreInitial: true,
+        persistent: true,
+        followSymlinks: false,
+        awaitWriteFinish: {
+            stabilityThreshold: Math.min(50, Math.max(10, Math.floor(debounceMs / 4))),
+            pollInterval: 10,
+        },
+    });
+    const onChange = (kind, abs) => {
+        if (closed)
+            return;
+        const rel = toRelPosix(abs);
+        if (isExcluded(rel) || !isParseable(rel))
+            return;
+        if (kind === 'add')
+            stats.added += 1;
+        else
+            stats.changed += 1;
+        dirty.add(rel);
+        armDebounce();
+    };
+    const onUnlink = (abs) => {
+        if (closed)
+            return;
+        const rel = toRelPosix(abs);
+        if (isExcluded(rel) || !isParseable(rel))
+            return;
+        stats.removed += 1;
+        // If the file is also pending как dirty, drop the dirty marker -
+        // the unlink wins.
+        dirty.delete(rel);
+        removed.add(rel);
+        armDebounce();
+    };
+    watcher.on('add', (p) => onChange('add', p));
+    watcher.on('change', (p) => onChange('change', p));
+    watcher.on('unlink', (p) => onUnlink(p));
+    watcher.on('error', (err) => {
+        stats.watcherErrors += 1;
+        if (!quiet) {
+            const message = err instanceof Error ? err.message : String(err);
+            writeStderr(`pugi index --watch: chokidar error: ${message}\n`);
+        }
+    });
+    const close = async () => {
+        if (closed)
+            return;
+        closed = true;
+        if (timer) {
+            clearTimeout(timer);
+            timer = null;
+        }
+        // Drain any pending events before closing the watcher.
+        if (dirty.size > 0 || removed.size > 0) {
+            try {
+                await flush();
+            }
+            catch (err) {
+                if (!quiet) {
+                    const message = err instanceof Error ? err.message : String(err);
+                    writeStderr(`pugi index --watch: close-time flush failed: ${message}\n`);
+                }
+            }
+        }
+        // Wait for any in-progress flush even if the queue was empty.
+        if (flushing) {
+            try {
+                await flushing;
+            }
+            catch {
+                /* already surfaced via writeStderr above */
+            }
+        }
+        await watcher.close();
+    };
+    const flushNow = async () => {
+        if (timer) {
+            clearTimeout(timer);
+            timer = null;
+        }
+        await flush();
+    };
+    return {
+        close,
+        flushNow,
+        stats: () => ({ ...stats }),
+    };
+}
+/**
+ * Cross-file target resolver backed by the live `symbols` table.
+ * Returns the first matching row id - deterministic stability via the
+ * autoincrement order, mirrors `reindex.ts`'s "first-write-wins"
+ * convention. `null` when no symbol carries the requested name.
+ */
+function crossFileLookup(db, name) {
+    const row = db.conn
+        .prepare('SELECT id FROM symbols WHERE name = ? ORDER BY id ASC LIMIT 1')
+        .get(name);
+    return row?.id ?? null;
+}
+function sha256(s) {
+    return createHash('sha256').update(s).digest('hex');
+}
+//# sourceMappingURL=watcher.js.map

package/dist/core/diagnostics/probes/sandbox.js

@@ -24,21 +24,16 @@ export function probeSandbox(ctx) {
             extraWritePaths,
         });
         if (state.armed) {
-            // Discipline-gap honesty (Trust Sprint thesis): the adapter
-            // probes ok, but spawn-wrap is NOT yet wired into the bash
-            // runner (that file is owned by another agent on PUGI-VERIFY-
-            // GATE). Reporting status=ok would overstate the posture — an
-            // operator reading 'armed' would assume their bash calls were
-            // jailed when they still run with full process privileges. We
-            // surface 'warn' with a precise reason instead and flip к 'ok'
-            // when the runner indirection lands.
+            // Phase 1 #302 — bash runner indirection is now wired (the
+            // sandbox wrap fires for every bash spawn / spawnSync site). We
+            // can finally promote 'warn' → 'ok' when the adapter probes
+            // armed. Operators get an unambiguous signal: "the OS sandbox
+            // really is wrapping your bash calls right now."
             return {
                 name: 'SANDBOX',
-                status: 'warn',
-                detail: `configured (mode=${state.mode}) but spawn-wrap not yet wired — bash dispatches still run with full process privileges. ` +
+                status: 'ok',
+                detail: `armed (mode=${state.mode}). Bash dispatches go through the OS sandbox. ` +
                     `Adapter posture: ${state.details.join('; ')}`,
-                remediation: 'The seatbelt adapter is in-tree and exercised by tests; the bash runner indirection that consumes it lands in a follow-up. ' +
-                    'Bash classifier denylist + permission FSM remain in force in the meantime.',
             };
         }
         // Not armed — distinguish "operator chose none" from "configured

package/dist/core/engine/prompts.js

@@ -69,6 +69,36 @@ const CUSTOMER_APP_DEFAULTS = [
     ' - Locale: default to en-US for date/number formatting unless overridden.',
     'The operator\'s CHAT language is NOT the target market language. A brief sent in Russian or Ukrainian does NOT imply the generated app should ship in Russian / Ukrainian / ₽ / ₴. If the target market is ambiguous, ask one clarifying question OR ship the USD + English default and surface the choice in your final answer so the operator can override.',
 ].join('\n');
+/**
+ * PR N (2026-06-05) - simplicity criterion for customer apps (task #112).
+ *
+ * Hiroshi (dev persona) tends к overengineer customer apps. The
+ * Pizza Finder dogfood demo shipped 8 features + multi-filter UI
+ * when "MVP" would have been 3 pizzas + name search. CEO directive
+ * 2026-06-05: bake a simplicity criterion into the code/build
+ * prompt so the first turn ships the smallest shape that proves
+ * the brief.
+ *
+ * Pattern reference: Karpathy speedrun (smallest end-to-end loop
+ * first, then add capacity). Shipping the maximalist version on
+ * turn 1 wastes tokens AND makes the demo harder к reason about;
+ * ship the simplest shape, ask "want me to add X next?" only if
+ * context strongly hints the brief implies it.
+ *
+ * Memory pointer: backlog task #112 (P1, Simplicity criterion baked
+ * into Hiroshi system prompt).
+ */
+const SIMPLICITY_CRITERION = [
+    '# Simplicity criterion',
+    'When implementing a customer feature, default to the SIMPLEST shape that proves the brief:',
+    ' - Prefer 3 sample items over 8. Operator can ask for more.',
+    ' - Prefer 1 filter over 5. Operator can ask for more.',
+    ' - Prefer 1 sort option over 3. Operator can ask for more.',
+    ' - Hardcoded data over fetching, until the operator names a backend.',
+    ' - Inline CSS over Tailwind / library imports, for single-file demos.',
+    ' - Single file over multi-file, until the operator names structure.',
+    'The brief tells you WHAT to build. The operator\'s NEXT brief tells you HOW MUCH to add. Shipping the maximalist version on turn 1 wastes tokens AND makes the demo harder to reason about. Ship the simplest shape; ask "want me to add X next?" only if context strongly hints the brief implies it.',
+].join('\n');
 /**
  * PR I (2026-06-05) — server-kill guidance.
  *
@@ -193,6 +223,7 @@ function baseSystemPromptFor(kind) {
                 'Command: `pugi code`. The operator gave you a feature request or refactor. Implement it end-to-end.',
                 EDIT_FLOW_RULES,
                 CUSTOMER_APP_DEFAULTS,
+                SIMPLICITY_CRITERION,
                 SERVER_KILL_HINT,
                 'If the request is ambiguous, ask one clarifying question by returning a text answer instead of editing.',
             ].join('\n\n');
@@ -224,6 +255,7 @@ function baseSystemPromptFor(kind) {
                 'Command: `pugi build`. The operator wants you to scaffold a feature across multiple files.',
                 EDIT_FLOW_RULES,
                 CUSTOMER_APP_DEFAULTS,
+                SIMPLICITY_CRITERION,
                 SERVER_KILL_HINT,
                 'Group related edits, run lint/test via bash where it adds confidence, and list every file you created or modified in the final answer.',
             ].join('\n\n');

package/dist/core/eval/v1/ledger.js

@@ -0,0 +1,83 @@
+/**
+ * Append-only TSV ledger for pugi-eval-v1 results.
+ *
+ * Pattern source: backlog #110 (Karpathy autoresearch). The ledger
+ * is git-tracked, never rewritten in place, never edited by hand. The
+ * column set is frozen as of schema v1; new columns must land in
+ * eval-v2 with a separate ledger file.
+ *
+ * Columns (tab-separated, in this exact order):
+ *
+ *   timestamp        UTC ISO 8601
+ *   git_sha          short sha of repo HEAD at run start
+ *   task_id          frozen task id (`NN-slug`)
+ *   model            engine model identifier or `(default)`
+ *   status           pass | fail | budget_exhausted | timeout | engine_error
+ *   pugi_score       per-task score, 2 decimal places
+ *   tokens           tokens reported by engine
+ *   turns            engine turn count
+ *   tool_calls       tool calls executed
+ *   wall_ms          wall-clock duration (ms)
+ *   exit_code        CLI subprocess exit code
+ *   verifications    `<passed>/<total>`
+ *
+ * Bytes containing TAB or NEWLINE are stripped from string fields
+ * before write so a single line is always one TSV record.
+ */
+import { appendFileSync, existsSync, mkdirSync, writeFileSync } from 'node:fs';
+import { dirname } from 'node:path';
+export const LEDGER_COLUMNS = [
+    'timestamp',
+    'git_sha',
+    'task_id',
+    'model',
+    'status',
+    'pugi_score',
+    'tokens',
+    'turns',
+    'tool_calls',
+    'wall_ms',
+    'exit_code',
+    'verifications',
+];
+export const LEDGER_HEADER = LEDGER_COLUMNS.join('\t');
+function safe(s) {
+    return s.replace(/[\t\r\n]+/g, ' ');
+}
+export function formatLedgerLine(row) {
+    const { timestamp, gitSha, model, result } = row;
+    const passed = result.verifications.filter((v) => v.passed).length;
+    const total = result.verifications.length;
+    const cells = [
+        safe(timestamp),
+        safe(gitSha),
+        safe(result.taskId),
+        safe(model),
+        safe(result.status),
+        result.pugiScore.toFixed(2),
+        String(result.tokensUsed),
+        String(result.turnsUsed),
+        String(result.toolCallCount),
+        String(result.wallClockMs),
+        String(result.exitCode),
+        `${passed}/${total}`,
+    ];
+    return cells.join('\t');
+}
+/**
+ * Append a single row. Creates the file with the header if it does
+ * not yet exist; otherwise appends a single line. Never rewrites
+ * existing content.
+ */
+export function appendLedgerRow(ledgerPath, row) {
+    mkdirSync(dirname(ledgerPath), { recursive: true });
+    if (!existsSync(ledgerPath)) {
+        writeFileSync(ledgerPath, `${LEDGER_HEADER}\n`, { mode: 0o644 });
+    }
+    appendFileSync(ledgerPath, `${formatLedgerLine(row)}\n`);
+}
+export function appendLedgerRows(ledgerPath, rows) {
+    for (const row of rows)
+        appendLedgerRow(ledgerPath, row);
+}
+//# sourceMappingURL=ledger.js.map