@pugi/cli 0.1.0-beta.1 → 0.1.0-beta.100

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: "${input.name}"`);
+    }
+    // β1a r1 : re-verify the on-disk skill payload against
+    // the trust manifest sha256 on EVERY invoke, not just at install
+    // time. Before this fix a post-install swap (malicious npm dep that
+    // touches `~/.pugi/skills/<name>/SKILL.md` after the operator
+    // approved the install) would bypass the trust gate — `listSkills`
+    // reads the body fresh from disk and the loader does no integrity
+    // check. The skill body lands directly in the model's tool result,
+    // so a mutated body is a prompt-injection vector against the agent
+    // loop's tool surface.
+    //
+    // Posture:
+    //  - `trusted`  → proceed (body is hash-pinned).
+    //  - `unsigned` → refuse: the operator never approved this skill.
+    //    This catches the case where a skill directory was dropped in
+    //    manually (no `pugi skills install`) and the loader picked it
+    //    up. Refusing is fail-closed.
+    //  - `mismatch` → refuse + surface the recorded vs actual hashes
+    //    so the operator can decide between re-trust and revoke.
+    //
+    // Performance: `hashSkillDir` walks the skill directory on every
+    // invoke. Skills are small (median 4-8 files, <50KB total) so the
+    // cost is sub-millisecond on warm cache. The β1a r1 spec exercises
+    // a mutated-body case; the existing skill-tool.spec.ts cases for
+    // happy-path use the `recordTrust` helper to seed the registry.
+    const actualHash = hashSkillDir(skill.dir);
+    const verdict = await verifyTrust('skill', skill.scope, skill.name, actualHash);
+    if (verdict.status === 'unsigned') {
+        throw new Error(`skill: refused to invoke "${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/sleep.js

@@ -0,0 +1,99 @@
+/**
+ * sleep tool — wall-clock pause primitive (tool gap pack).
+ *
+ * Closes a parity gap with the upstream tool's tool surface. The model calls
+ * this when it needs a fixed delay before its next action (waiting on
+ * a process the operator owns, throttling a poll loop). The call
+ * counts against `--max-turns` like every other tool dispatch, so the
+ * budget gate naturally caps abuse.
+ *
+ * Operator guidance: prefer a real poll loop (read + grep + retry) over
+ * blind sleep. The tool exists for the cases where polling is not an
+ * option (a fixed cooldown between API calls, a deterministic settle
+ * window for a build) — most agent flows do NOT want it.
+ *
+ * Wire shape:
+ *  args:  { seconds: number }
+ *          - integer in [1, 600]; non-integer / out-of-range rejects
+ *            at parse time with a sentinel string.
+ *  return: { ok: true, sleptMs: number } serialised JSON.
+ *
+ * No side effects beyond the wall-clock delay; nothing on disk, no
+ * subprocesses, no environment mutation.
+ *
+ * Brand voice: English only, no emoji, no banned words.
+ */
+/** Hard caps. The lower bound rejects zero / negative inputs at parse
+ * time so the model can self-correct; the upper bound matches the
+ * standard tool timeout budget used elsewhere in the CLI. */
+export const SLEEP_MIN_SECONDS = 1;
+export const SLEEP_MAX_SECONDS = 600;
+/** Sentinel prefix returned when input validation rejects the call. */
+export const SLEEP_INVALID_ARGS = 'SLEEP_INVALID_ARGS';
+/**
+ * Validate the raw arguments. Returns the typed value on success or a
+ * `SLEEP_INVALID_ARGS: ...` sentinel string. Non-integer values reject
+ * because partial seconds invite drift across platforms; the model
+ * should round explicitly at the call site.
+ */
+export function parseSleepArgs(raw) {
+    if (typeof raw !== 'object' || raw === null || Array.isArray(raw)) {
+        return `${SLEEP_INVALID_ARGS}: arguments must be a JSON object`;
+    }
+    const obj = raw;
+    const seconds = obj['seconds'];
+    if (typeof seconds !== 'number' || !Number.isFinite(seconds)) {
+        return `${SLEEP_INVALID_ARGS}: seconds must be a finite number`;
+    }
+    if (!Number.isInteger(seconds)) {
+        return `${SLEEP_INVALID_ARGS}: seconds must be an integer`;
+    }
+    if (seconds < SLEEP_MIN_SECONDS) {
+        return `${SLEEP_INVALID_ARGS}: seconds must be >= ${SLEEP_MIN_SECONDS}`;
+    }
+    if (seconds > SLEEP_MAX_SECONDS) {
+        return `${SLEEP_INVALID_ARGS}: seconds must be <= ${SLEEP_MAX_SECONDS}`;
+    }
+    return { seconds };
+}
+/**
+ * Dispatch entry point. Validates input, awaits the wall-clock delay,
+ * and returns the structured result envelope as JSON.
+ *
+ * On validation failure returns the sentinel string directly (no throw)
+ * so the engine adapter surfaces it as a recoverable tool result and
+ * the model can self-correct the arguments.
+ */
+export async function dispatchSleep(ctx, raw) {
+    const parsed = parseSleepArgs(raw);
+    if (typeof parsed === 'string') {
+        return parsed;
+    }
+    const ms = parsed.seconds * 1_000;
+    const timer = ctx.timer ?? ((cb, delay) => setTimeout(cb, delay));
+    await new Promise((resolveDelay) => {
+        timer(resolveDelay, ms);
+    });
+    const result = { ok: true, sleptMs: ms };
+    return JSON.stringify(result);
+}
+/**
+ * JSON-Schema fragment the schema builder advertises to the model.
+ * Hand-written for parity with the rest of the tool surface (see the
+ * note on `briefJsonSchema` for why we do not pull in zod-to-json-schema).
+ */
+export const sleepJsonSchema = {
+    type: 'object',
+    additionalProperties: false,
+    required: ['seconds'],
+    properties: {
+        seconds: {
+            type: 'integer',
+            minimum: SLEEP_MIN_SECONDS,
+            maximum: SLEEP_MAX_SECONDS,
+            description: `Wall-clock pause in seconds. Integer in [${SLEEP_MIN_SECONDS}, ${SLEEP_MAX_SECONDS}]. ` +
+                'Prefer a real poll loop over blind sleep; this tool counts against --max-turns.',
+        },
+    },
+};
+//# sourceMappingURL=sleep.js.map

package/dist/tools/synthetic-output.js

@@ -0,0 +1,133 @@
+/**
+ * synthetic_output tool — operator-readable echo helper (tool gap
+ * pack). EXPERIMENTAL / engine-only.
+ *
+ * Test fixture cousin of `brief`. Whereas `brief` persists a structured
+ * record to `.pugi/briefs/<session>.jsonl`, `synthetic_output` writes
+ * the supplied text VERBATIM к the requested stream (`stdout` or
+ * `stderr`). The use case is the engine-side test harness: a spec wants
+ * to assert that "the loop ran this tool with this payload" without
+ * spinning up the entire dispatch pipeline, AND wants to observe the
+ * write the same way an operator would. Production agent flows should
+ * prefer `brief` because the structured record survives crashes — the
+ * stream write does not.
+ *
+ * The tool is NOT exposed as a slash command. The schema is advertised
+ * to the engine only when the executor is built with an
+ * `allowSyntheticOutput: true` opt-in. Customer CLIs leave it off so
+ * the model cannot use it as a side-channel to bypass the normal tool
+ * result envelope (which is logged, classified, hooked, etc.).
+ *
+ * Wire shape:
+ *  args:  { stream: 'stdout'|'stderr', text: string }
+ *          - text is capped at 16 KiB UTF-8; over-cap rejects with
+ *            the SYNTHETIC_OUTPUT_TOO_LARGE sentinel.
+ *  side:  raw `process.stdout.write(text)` or `process.stderr.write(text)`.
+ *  return: short JSON envelope { ok, stream, bytes } so the engine
+ *          loop can audit the call without re-reading the streamed
+ *          bytes (which it does not buffer).
+ *
+ * Atomicity: writes go to the live process stream; partial writes are
+ * the stream's responsibility, not the tool's. The tool does NOT add a
+ * trailing newline — the model controls the exact bytes. This is the
+ * design point that makes it a useful test fixture (specs assert on
+ * verbatim bytes).
+ *
+ * Brand voice: English only, no emoji, no banned words.
+ */
+/** Per-call byte cap. Mirrors the bash tool's stdout/stderr cap so the
+ * model cannot blow either pane with a single synthetic flush. */
+export const SYNTHETIC_OUTPUT_MAX_BYTES = 16 * 1_024;
+/** Canonical stream values. */
+export const SYNTHETIC_OUTPUT_STREAMS = ['stdout', 'stderr'];
+/** Sentinel returned when input validation rejects the call. */
+export const SYNTHETIC_OUTPUT_INVALID_ARGS = 'SYNTHETIC_OUTPUT_INVALID_ARGS';
+/** Sentinel returned when the encoded text exceeds the per-call cap.
+ * Distinct from the args-schema failure so the model knows to shorten
+ * the payload rather than change the shape. */
+export const SYNTHETIC_OUTPUT_TOO_LARGE = 'SYNTHETIC_OUTPUT_TOO_LARGE';
+/**
+ * Validate the raw arguments. Returns the typed value on success or a
+ * `SYNTHETIC_OUTPUT_INVALID_ARGS: ...` sentinel string.
+ */
+export function parseSyntheticOutputArgs(raw) {
+    if (typeof raw !== 'object' || raw === null || Array.isArray(raw)) {
+        return `${SYNTHETIC_OUTPUT_INVALID_ARGS}: arguments must be a JSON object`;
+    }
+    const obj = raw;
+    const issues = [];
+    const stream = obj['stream'];
+    if (typeof stream !== 'string') {
+        issues.push('stream: must be a string');
+    }
+    else if (!SYNTHETIC_OUTPUT_STREAMS.includes(stream)) {
+        issues.push(`stream: must be one of ${SYNTHETIC_OUTPUT_STREAMS.join('|')}`);
+    }
+    const text = obj['text'];
+    if (typeof text !== 'string') {
+        issues.push('text: must be a string');
+    }
+    if (issues.length > 0) {
+        return `${SYNTHETIC_OUTPUT_INVALID_ARGS}: ${issues.join('; ')}`;
+    }
+    return {
+        stream: stream,
+        text: text,
+    };
+}
+/**
+ * Dispatch entry point. Validates input, writes the text verbatim to
+ * the requested stream, and returns the structured envelope as JSON.
+ *
+ * Returns sentinel strings (no throw) on recoverable failures so the
+ * engine adapter surfaces them as tool results.
+ */
+export function dispatchSyntheticOutput(ctx, raw) {
+    const parsed = parseSyntheticOutputArgs(raw);
+    if (typeof parsed === 'string') {
+        return parsed;
+    }
+    const bytes = Buffer.byteLength(parsed.text, 'utf8');
+    if (bytes > SYNTHETIC_OUTPUT_MAX_BYTES) {
+        return `${SYNTHETIC_OUTPUT_TOO_LARGE}: encoded text exceeds ${SYNTHETIC_OUTPUT_MAX_BYTES} bytes`;
+    }
+    // Resolve writers once per call so a spec can override stdout while
+    // letting stderr fall through to the real process stream.
+    const writer = parsed.stream === 'stdout'
+        ? (ctx.stdoutWrite ?? defaultStdoutWrite)
+        : (ctx.stderrWrite ?? defaultStderrWrite);
+    writer(parsed.text);
+    const result = {
+        ok: true,
+        stream: parsed.stream,
+        bytes,
+    };
+    return JSON.stringify(result);
+}
+function defaultStdoutWrite(chunk) {
+    process.stdout.write(chunk);
+}
+function defaultStderrWrite(chunk) {
+    process.stderr.write(chunk);
+}
+/**
+ * JSON-Schema fragment the schema builder advertises to the model.
+ * Hand-written for parity with the rest of the tool surface.
+ */
+export const syntheticOutputJsonSchema = {
+    type: 'object',
+    additionalProperties: false,
+    required: ['stream', 'text'],
+    properties: {
+        stream: {
+            type: 'string',
+            enum: [...SYNTHETIC_OUTPUT_STREAMS],
+            description: 'Destination stream: stdout or stderr.',
+        },
+        text: {
+            type: 'string',
+            description: `Verbatim bytes to write. Capped at ${SYNTHETIC_OUTPUT_MAX_BYTES} bytes UTF-8.`,
+        },
+    },
+};
+//# sourceMappingURL=synthetic-output.js.map

package/dist/tools/tasks.js

@@ -0,0 +1,208 @@
+/**
+ * task_* tool family — β1 T1/T6 (TodoWrite + agent task ledger).
+ *
+ * Mirrors the standard tool'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 : 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