@pugi/cli 0.1.0-beta.4 → 0.1.0-beta.41

package/dist/tools/powershell.js

@@ -0,0 +1,156 @@
+/**
+ * PowerShell tool — leak L6 (Pugi parity catch-up 2026-05-28).
+ *
+ * Windows operators cannot run native `*.ps1` scripts via the bash tool
+ * (which spawns `/bin/sh`). This tool spawns `pwsh -NoProfile -Command`
+ * на cross-platform PowerShell 7+ binary so Windows-first workflows are
+ * first-class на Pugi.
+ *
+ * Clean-room re-implementation. Surface mirrors bashTool's permission
+ * gate, env sanitiser, output cap, timeout, and exit-code propagation;
+ * the only difference is the shell binary selection. Per-platform
+ * resolution:
+ *   - All OS: try `pwsh` on $PATH first (PowerShell 7+ cross-platform).
+ *   - Windows fallback: `powershell.exe` (Windows PowerShell 5.1 baked-in).
+ *   - Other OS without pwsh: tool returns a clear "powershell binary
+ *     not found" error so the operator can install pwsh or fall back
+ *     к bash.
+ *
+ * Permission class: reuses the bash classifier — destructive patterns,
+ * sandbox detection, and additional-directories checks are command-string
+ * based and apply equally to pwsh and sh.
+ */
+import { spawnSync } from 'node:child_process';
+import { evaluateBashPermission } from '../core/permission.js';
+import { recordToolCall, recordToolResult } from '../core/session.js';
+export const POWERSHELL_OUTPUT_CAP_BYTES = 64 * 1024;
+export const POWERSHELL_DEFAULT_TIMEOUT_MS = 30_000;
+export const POWERSHELL_MAX_TIMEOUT_MS = 120_000;
+/** Cached binary path so repeated calls inside one session skip the probe. */
+let cachedShellBinary;
+function resolveShellBinary() {
+    if (cachedShellBinary !== undefined)
+        return cachedShellBinary;
+    // Try pwsh (cross-platform PowerShell 7+) first.
+    const pwshProbe = spawnSync('pwsh', ['-NoProfile', '-Command', 'exit 0'], {
+        encoding: 'utf8',
+        stdio: ['ignore', 'ignore', 'ignore'],
+        timeout: 3000,
+    });
+    if (pwshProbe.status === 0) {
+        cachedShellBinary = 'pwsh';
+        return 'pwsh';
+    }
+    // Windows fallback к the baked-in PowerShell 5.1.
+    if (process.platform === 'win32') {
+        const wpsProbe = spawnSync('powershell.exe', ['-NoProfile', '-Command', 'exit 0'], {
+            encoding: 'utf8',
+            stdio: ['ignore', 'ignore', 'ignore'],
+            timeout: 3000,
+        });
+        if (wpsProbe.status === 0) {
+            cachedShellBinary = 'powershell.exe';
+            return 'powershell.exe';
+        }
+    }
+    cachedShellBinary = null;
+    return null;
+}
+function sanitizeTimeout(value) {
+    if (value === undefined || !Number.isFinite(value) || value <= 0) {
+        return POWERSHELL_DEFAULT_TIMEOUT_MS;
+    }
+    return Math.min(value, POWERSHELL_MAX_TIMEOUT_MS);
+}
+function buildChildEnv() {
+    const env = { ...process.env };
+    delete env['PUGI_API_KEY'];
+    delete env['PUGI_LOGIN_TOKEN'];
+    return env;
+}
+/**
+ * Sync PowerShell dispatch. Mirrors bashToolSync shape so dispatchTool
+ * can call either tool with the same context shape.
+ */
+export function powerShellToolSync(input, ctx) {
+    const cmd = input.cmd ?? '';
+    const additionalDirectories = ctx.additionalDirectories ?? [];
+    const source = ctx.source ?? 'agent';
+    const toolCallId = recordToolCall(ctx.session, 'powershell', cmd);
+    const decision = evaluateBashPermission(cmd, ctx.settings.permissions.mode, {
+        workspaceRoot: ctx.root,
+        additionalDirectories,
+        source,
+    });
+    if (decision.decision !== 'allow') {
+        const reason = `Permission ${decision.decision}: ${decision.reason}`;
+        recordToolResult(ctx.session, toolCallId, 'error', reason);
+        return {
+            stdout: '',
+            stderr: `Permission denied: ${decision.reason}`,
+            exitCode: 126,
+            truncated: false,
+            timedOut: false,
+            shellBinary: 'unresolved',
+        };
+    }
+    const shellBinary = resolveShellBinary();
+    if (shellBinary === null) {
+        const reason = 'powershell binary not found (tried pwsh' +
+            (process.platform === 'win32' ? ', powershell.exe' : '') +
+            '). Install PowerShell 7+ from https://aka.ms/powershell or use the bash tool instead.';
+        recordToolResult(ctx.session, toolCallId, 'error', reason);
+        return {
+            stdout: '',
+            stderr: reason,
+            exitCode: 127,
+            truncated: false,
+            timedOut: false,
+            shellBinary: 'unavailable',
+        };
+    }
+    const timeoutMs = sanitizeTimeout(input.timeoutMs);
+    const childEnv = buildChildEnv();
+    const cwd = input.cwd ?? ctx.root;
+    const result = spawnSync(shellBinary, ['-NoProfile', '-Command', cmd], {
+        cwd,
+        env: childEnv,
+        encoding: 'utf8',
+        stdio: ['ignore', 'pipe', 'pipe'],
+        timeout: timeoutMs,
+        maxBuffer: 10 * 1024 * 1024,
+    });
+    const stdoutFull = (result.stdout ?? '').toString();
+    const stderrFull = (result.stderr ?? '').toString();
+    const combined = stdoutFull.length + stderrFull.length;
+    const truncated = combined > POWERSHELL_OUTPUT_CAP_BYTES;
+    let stdoutOut = stdoutFull;
+    let stderrOut = stderrFull;
+    if (truncated) {
+        const halfCap = POWERSHELL_OUTPUT_CAP_BYTES / 2;
+        stdoutOut = stdoutFull.slice(0, halfCap);
+        stderrOut = stderrFull.slice(0, halfCap);
+    }
+    const timedOut = result.error?.code === 'ETIMEDOUT' ||
+        result.signal === 'SIGTERM';
+    const exitCode = timedOut ? 124 : result.status ?? 1;
+    if (timedOut) {
+        recordToolResult(ctx.session, toolCallId, 'error', `powershell timed out after ${timeoutMs}ms`);
+    }
+    else {
+        recordToolResult(ctx.session, toolCallId, 'success', `powershell exit=${exitCode} bytes=${combined} binary=${shellBinary}`);
+    }
+    return {
+        stdout: stdoutOut,
+        stderr: stderrOut,
+        exitCode,
+        truncated,
+        timedOut,
+        shellBinary,
+    };
+}
+/** Visible-for-spec helper: forces a re-probe on next call. */
+export function _resetShellBinaryCacheForSpec() {
+    cachedShellBinary = undefined;
+}
+//# sourceMappingURL=powershell.js.map

package/dist/tools/registry.js

@@ -1,8 +1,36 @@
 const registry = [
+    // α7.7: unified-diff patch apply. Routes through the same security
+    // gate as Layer A/B/C, so the risk class matches `edit`/`write`
+    // (medium — writes inside the workspace, never to protected files).
+    { name: 'apply_patch', permission: 'edit', risk: 'medium', concurrencySafe: false, m1: true },
+    // Leak L5 (2026-05-27): structured multi-choice clarifier tool. Risk =
+    // low because the dispatch is a pure UI surface — no file writes, no
+    // shell, no network. Permission = none (no workspace access required).
+    // concurrencySafe = true because the prompt-budget gate runs in the
+    // engine loop, not via tool-side mutex (one prompt per turn is enforced
+    // by the persona system prompt + the engine's tool_calls budget).
+    { name: 'ask_user_question', permission: 'none', risk: 'low', concurrencySafe: true, m1: true },
     { name: 'bash', permission: 'bash', risk: 'high', concurrencySafe: false, m1: true },
     { name: 'edit', permission: 'edit', risk: 'medium', concurrencySafe: false, m1: true },
     { name: 'glob', permission: 'read', risk: 'low', concurrencySafe: true, m1: true },
     { name: 'grep', permission: 'read', risk: 'low', concurrencySafe: true, m1: true },
+    // α7.7: LSP read-only surface. Server runs locally, no Anvil
+    // round-trip. Concurrency-safe because every operation reads
+    // server state without mutating workspace files.
+    { name: 'lsp_definition', permission: 'read', risk: 'low', concurrencySafe: true, m1: true },
+    { name: 'lsp_diagnostics', permission: 'read', risk: 'low', concurrencySafe: true, m1: true },
+    { name: 'lsp_hover', permission: 'read', risk: 'low', concurrencySafe: true, m1: true },
+    { name: 'lsp_references', permission: 'read', risk: 'low', concurrencySafe: true, m1: true },
+    // β7 L5+T11: multi_edit dispatches an ordered batch of Layer A edits
+    // as a single transaction. Risk = medium (same chokepoints as `edit`).
+    // concurrencySafe = false because the journal serialises one dispatch
+    // per session.
+    { name: 'multi_edit', permission: 'edit', risk: 'medium', concurrencySafe: false, m1: true },
+    // Leak L6 (2026-05-28): PowerShell tool for Windows-first workflows. Same
+    // bash permission class — destructive-pattern classification fires the
+    // same gate. concurrencySafe = false because spawn-shell child cwd /
+    // env carry-over could race across parallel agent calls.
+    { name: 'powershell', permission: 'bash', risk: 'high', concurrencySafe: false, m1: false },
     { name: 'question', permission: 'none', risk: 'low', concurrencySafe: false, m1: true },
     { name: 'read', permission: 'read', risk: 'low', concurrencySafe: true, m1: true },
     { name: 'skill', permission: 'read', risk: 'low', concurrencySafe: true, m1: true },
@@ -10,7 +38,30 @@ const registry = [
     { name: 'task_get', permission: 'none', risk: 'low', concurrencySafe: true, m1: true },
     { name: 'task_list', permission: 'none', risk: 'low', concurrencySafe: true, m1: true },
     { name: 'task_update', permission: 'none', risk: 'low', concurrencySafe: false, m1: true },
+    // Leak L16 (2026-05-27): batch TodoWrite. Mirrors Claude Code's upstream
+    // surface — full board snapshot, single-in-progress invariant, atomic
+    // tmp+rename persistence to `.pugi/todos.json`. `concurrencySafe = false`
+    // because two concurrent writes could lose the loser's snapshot (the
+    // rename is atomic but the read-modify-write loop is not). Risk = low
+    // because the only filesystem mutation lands inside `.pugi/todos.json`,
+    // which is metadata, not source.
+    { name: 'todo_write', permission: 'none', risk: 'low', concurrencySafe: false, m1: true },
     { name: 'web_fetch', permission: 'network', risk: 'medium', concurrencySafe: true, m1: true },
+    // α7.7: scratch worktree management. `worktree_create` writes nothing
+    // dangerous (a clone under `.pugi/worktrees/`); `worktree_promote`
+    // applies a diff back to the main tree, so it shares the `edit`
+    // risk class. `worktree_drop` is the cleanup primitive.
+    //
+    // R1 fix (2026-05-26, PR #413 r1, Fix 9): raised `worktree_create`
+    // and `worktree_drop` from `low` to `medium`. `worktree_drop` runs
+    // `rmSync` on its target — even with the new path-containment gate
+    // in `core/edits/worktree.ts::dropWorktree`, a destructive primitive
+    // belongs in `medium` so the permission FSM prompts on every call.
+    // `worktree_create` is raised for disk-pressure parity (a runaway
+    // agent loop could fill the disk with abandoned scratch worktrees).
+    { name: 'worktree_create', permission: 'edit', risk: 'medium', concurrencySafe: false, m1: true },
+    { name: 'worktree_drop', permission: 'edit', risk: 'medium', concurrencySafe: false, m1: true },
+    { name: 'worktree_promote', permission: 'edit', risk: 'medium', concurrencySafe: false, m1: true },
     { name: 'write', permission: 'edit', risk: 'medium', concurrencySafe: false, m1: true },
 ];
 export const toolRegistry = registry.sort((a, b) => a.name.localeCompare(b.name));

package/dist/tools/skill-tool.js

@@ -0,0 +1,96 @@
+import { listSkills } from '../core/skills/loader.js';
+import { hashSkillDir, verifyTrust } from '../core/skills/trust.js';
+export const SKILL_BODY_CAP_BYTES = 32 * 1024;
+export const SKILL_LIST_CAP = 100;
+export function skillList(ctx, input) {
+    const scope = input.scope ?? 'all';
+    const all = [];
+    if (scope === 'all' || scope === 'global') {
+        all.push(...listSkills('global', ctx.workspaceRoot));
+    }
+    if (scope === 'all' || scope === 'workspace') {
+        all.push(...listSkills('workspace', ctx.workspaceRoot));
+    }
+    // Dedup by name, prefer workspace scope when both exist (workspace
+    // overrides global per skills loader convention).
+    const byName = new Map();
+    for (const skill of all) {
+        const prev = byName.get(skill.name);
+        if (!prev || skill.scope === 'workspace') {
+            byName.set(skill.name, skill);
+        }
+    }
+    return Array.from(byName.values())
+        .slice(0, SKILL_LIST_CAP)
+        .map((skill) => ({
+        name: skill.name,
+        description: skill.frontmatter.description,
+        scope: skill.scope,
+    }));
+}
+export async function skillInvoke(ctx, input) {
+    if (!input.name || typeof input.name !== 'string') {
+        throw new Error('skill: name is required');
+    }
+    // Defense-in-depth: skill loader already validates slugs but the
+    // tool surface is operator-controlled.
+    if (!/^[a-zA-Z0-9_-]{1,128}$/.test(input.name)) {
+        throw new Error(`skill: invalid skill name shape: "${input.name}"`);
+    }
+    // Workspace scope wins over global (operator override). Mirrors
+    // SkillLoader convention.
+    const workspace = listSkills('workspace', ctx.workspaceRoot).find((s) => s.name === input.name);
+    const global = workspace
+        ? null
+        : listSkills('global', ctx.workspaceRoot).find((s) => s.name === input.name);
+    const skill = workspace ?? global;
+    if (!skill) {
+        throw new Error(`skill: not found: "${input.name}"`);
+    }
+    // β1a r1 (2026-05-26): re-verify the on-disk skill payload against
+    // the trust manifest sha256 on EVERY invoke, not just at install
+    // time. Before this fix a post-install swap (malicious npm dep that
+    // touches `~/.pugi/skills/<name>/SKILL.md` after the operator
+    // approved the install) would bypass the trust gate — `listSkills`
+    // reads the body fresh from disk and the loader does no integrity
+    // check. The skill body lands directly in the model's tool result,
+    // so a mutated body is a prompt-injection vector against the agent
+    // loop's tool surface.
+    //
+    // Posture:
+    //   - `trusted`   → proceed (body is hash-pinned).
+    //   - `unsigned`  → refuse: the operator never approved this skill.
+    //     This catches the case where a skill directory was dropped in
+    //     manually (no `pugi skills install`) and the loader picked it
+    //     up. Refusing is fail-closed.
+    //   - `mismatch`  → refuse + surface the recorded vs actual hashes
+    //     so the operator can decide between re-trust and revoke.
+    //
+    // Performance: `hashSkillDir` walks the skill directory on every
+    // invoke. Skills are small (median 4-8 files, <50KB total) so the
+    // cost is sub-millisecond on warm cache. The β1a r1 spec exercises
+    // a mutated-body case; the existing skill-tool.spec.ts cases for
+    // happy-path use the `recordTrust` helper to seed the registry.
+    const actualHash = hashSkillDir(skill.dir);
+    const verdict = await verifyTrust('skill', skill.scope, skill.name, actualHash);
+    if (verdict.status === 'unsigned') {
+        throw new Error(`skill: refused to invoke "${skill.name}" — no trust entry (run \`pugi skills trust ${skill.name}\` to approve)`);
+    }
+    if (verdict.status === 'mismatch') {
+        throw new Error(`skill: refused to invoke "${skill.name}" — sha256 mismatch (recorded ${verdict.recorded.slice(0, 12)}…, actual ${verdict.actual.slice(0, 12)}…). Re-trust via \`pugi skills trust ${skill.name}\`.`);
+    }
+    const body = skill.body;
+    const truncated = Buffer.byteLength(body, 'utf8') > SKILL_BODY_CAP_BYTES;
+    const cappedBody = truncated
+        ? body.slice(0, SKILL_BODY_CAP_BYTES) +
+            `\n\n(... truncated at ${SKILL_BODY_CAP_BYTES} bytes — see \`pugi skills info ${skill.name}\` for full text)`
+        : body;
+    return {
+        name: skill.name,
+        scope: skill.scope,
+        description: skill.frontmatter.description,
+        body: cappedBody,
+        truncated,
+    };
+}
+//# sourceMappingURL=skill-tool.js.map

package/dist/tools/tasks.js

@@ -0,0 +1,208 @@
+/**
+ * task_* tool family — β1 T1/T6 (TodoWrite + agent task ledger).
+ *
+ * Mirrors Claude Code's TodoWrite tool surface so a model trained on
+ * the upstream tool grammar speaks Pugi's variant verbatim. Four ops:
+ *
+ *   - `task_create` — append a new task to the session's todo ledger.
+ *     Returns the assigned id.
+ *   - `task_get`    — fetch a single task by id.
+ *   - `task_list`   — list every task in the current session, ordered
+ *                     by createdAt ascending.
+ *   - `task_update` — mutate status/title/notes of an existing task.
+ *                     Append-only journal — every mutation lands as a
+ *                     fresh JSONL line and the latest line per id wins
+ *                     on `task_list` / `task_get` reads.
+ *
+ * Persistence: append-only JSONL at
+ * `.pugi/sessions/<sessionId>/tasks.jsonl`. Append-only keeps crash
+ * recovery trivial — a partial write at the end of the file is the
+ * worst case and the parser drops the malformed tail line.
+ *
+ * Scope: this is the local-side ledger surface. Anvil-side mirror
+ * (cabinet `/projects/[id]/tasks` page) ships in β5 once the session-
+ * memory hook lands; until then the ledger is purely local.
+ */
+import { appendFileSync, chmodSync, existsSync, mkdirSync, readFileSync, } from 'node:fs';
+import { dirname, join } from 'node:path';
+import { randomUUID } from 'node:crypto';
+function ledgerPath(ctx) {
+    // Defense-in-depth: the sessionId is supposed to be a UUID minted by
+    // openSession() but the tool surface is operator-facing. Validate the
+    // shape before composing a path — refuse anything that contains
+    // separators or shell wildcards.
+    if (!/^[a-zA-Z0-9_-]{1,128}$/.test(ctx.sessionId)) {
+        throw new Error(`task_*: invalid sessionId shape: "${ctx.sessionId}"`);
+    }
+    return join(ctx.workspaceRoot, '.pugi', 'sessions', ctx.sessionId, 'tasks.jsonl');
+}
+function nowIso(ctx) {
+    return (ctx.now ? ctx.now() : new Date()).toISOString();
+}
+function ensureDir(path) {
+    // β1a r1 (2026-05-26): switched from POSIX-only
+    // `path.slice(0, path.lastIndexOf('/'))` to `path.dirname()` so
+    // Windows path separators (`\`) work. Also chmod the per-session
+    // directory to 0o700 — the tasks ledger carries operator-confidential
+    // brief text, status notes, and timing metadata that should not be
+    // world-readable through an inherited umask.
+    const dir = dirname(path);
+    if (!existsSync(dir)) {
+        mkdirSync(dir, { recursive: true });
+        try {
+            chmodSync(dir, 0o700);
+        }
+        catch {
+            // Best-effort. POSIX permission setting is a no-op on Windows
+            // NTFS, and the dir-creation race with another concurrent task
+            // tool call is the only realistic failure case. The 0o600 mode
+            // on the JSONL file itself remains the primary guard; the dir
+            // chmod is defense in depth for tools that walk `.pugi/`.
+        }
+    }
+}
+function readJournal(ctx) {
+    const path = ledgerPath(ctx);
+    if (!existsSync(path))
+        return [];
+    const raw = readFileSync(path, 'utf8');
+    const out = [];
+    for (const line of raw.split('\n')) {
+        if (!line.trim())
+            continue;
+        try {
+            const parsed = JSON.parse(line);
+            if ((parsed.op === 'create' || parsed.op === 'update') &&
+                typeof parsed.id === 'string' &&
+                typeof parsed.at === 'string') {
+                out.push(parsed);
+            }
+        }
+        catch {
+            // Drop malformed line (partial-write tail or external corruption).
+            // The append-only design guarantees only the LAST line can be bad
+            // — everything before it is whole.
+        }
+    }
+    return out;
+}
+function fold(journal) {
+    const out = new Map();
+    for (const entry of journal) {
+        if (entry.op === 'create') {
+            if (!entry.title)
+                continue;
+            out.set(entry.id, {
+                id: entry.id,
+                title: entry.title,
+                status: entry.status ?? 'pending',
+                ...(entry.notes !== undefined ? { notes: entry.notes } : {}),
+                createdAt: entry.at,
+                updatedAt: entry.at,
+            });
+        }
+        else {
+            const prev = out.get(entry.id);
+            if (!prev)
+                continue; // update before create — drop silently
+            out.set(entry.id, {
+                ...prev,
+                ...(entry.title !== undefined ? { title: entry.title } : {}),
+                ...(entry.status !== undefined ? { status: entry.status } : {}),
+                ...(entry.notes !== undefined ? { notes: entry.notes } : {}),
+                updatedAt: entry.at,
+            });
+        }
+    }
+    return out;
+}
+function appendEntry(ctx, entry) {
+    const path = ledgerPath(ctx);
+    ensureDir(path);
+    appendFileSync(path, `${JSON.stringify(entry)}\n`, {
+        encoding: 'utf8',
+        mode: 0o600,
+    });
+}
+export function taskCreate(ctx, input) {
+    const title = input.title?.trim();
+    if (!title) {
+        throw new Error('task_create: title is required');
+    }
+    if (title.length > 2_000) {
+        throw new Error('task_create: title exceeds 2000 char cap');
+    }
+    const status = input.status ?? 'pending';
+    if (!isValidStatus(status)) {
+        throw new Error(`task_create: invalid status "${status}"`);
+    }
+    const id = `task-${randomUUID()}`;
+    const at = nowIso(ctx);
+    const entry = {
+        op: 'create',
+        id,
+        title,
+        status,
+        at,
+        ...(input.notes !== undefined ? { notes: input.notes } : {}),
+    };
+    appendEntry(ctx, entry);
+    return {
+        id,
+        title,
+        status,
+        ...(input.notes !== undefined ? { notes: input.notes } : {}),
+        createdAt: at,
+        updatedAt: at,
+    };
+}
+export function taskGet(ctx, id) {
+    if (typeof id !== 'string' || id.length === 0) {
+        throw new Error('task_get: id is required');
+    }
+    const folded = fold(readJournal(ctx));
+    return folded.get(id) ?? null;
+}
+export function taskList(ctx) {
+    const folded = fold(readJournal(ctx));
+    return Array.from(folded.values()).sort((a, b) => a.createdAt.localeCompare(b.createdAt));
+}
+export function taskUpdate(ctx, input) {
+    if (!input.id)
+        throw new Error('task_update: id is required');
+    const folded = fold(readJournal(ctx));
+    const existing = folded.get(input.id);
+    if (!existing) {
+        throw new Error(`task_update: unknown id "${input.id}"`);
+    }
+    if (input.status !== undefined && !isValidStatus(input.status)) {
+        throw new Error(`task_update: invalid status "${input.status}"`);
+    }
+    if (input.title !== undefined && input.title.trim().length === 0) {
+        throw new Error('task_update: title cannot be empty');
+    }
+    const at = nowIso(ctx);
+    const entry = {
+        op: 'update',
+        id: input.id,
+        at,
+        ...(input.title !== undefined ? { title: input.title } : {}),
+        ...(input.status !== undefined ? { status: input.status } : {}),
+        ...(input.notes !== undefined ? { notes: input.notes } : {}),
+    };
+    appendEntry(ctx, entry);
+    return {
+        ...existing,
+        ...(input.title !== undefined ? { title: input.title } : {}),
+        ...(input.status !== undefined ? { status: input.status } : {}),
+        ...(input.notes !== undefined ? { notes: input.notes } : {}),
+        updatedAt: at,
+    };
+}
+function isValidStatus(status) {
+    return (status === 'pending' ||
+        status === 'in_progress' ||
+        status === 'completed' ||
+        status === 'cancelled');
+}
+//# sourceMappingURL=tasks.js.map