@pugi/cli 0.1.0-beta.9 → 0.1.0-beta.91

package/dist/core/agents/query-decomposer.js

@@ -0,0 +1,297 @@
+/**
+ * Query decomposer — primitive #3 of the 9-layer RAG architecture gap
+ * analysis (docs/research/2026-05-31-9-layer-rag-architecture-.md).
+ *
+ * Splits a user request into ordered atomic sub-queries the engine can
+ * route, plan, and execute step-by-step. This file is pure logic — it
+ * never opens a network connection itself. The caller injects a
+ * `transport` callback that maps a prompt string to a model response
+ * string, which keeps Anvil-engine wiring (Hiroshi persona, auth,
+ * retry, telemetry) out of the primitive and makes tests deterministic.
+ *
+ * Failure model is fail-closed to a single-step pass-through. Every
+ * malformed response, transport throw, schema violation, dependency
+ * cycle, or oversized step list collapses to:
+ *
+ *  { steps: [{ id: 'step-1', query, dependsOn: [], intent: 'unknown' }],
+ *    usedFallback: true, reason: '<why>' }
+ *
+ * The caller inspects `usedFallback` plus `reason` and decides whether
+ * the degradation is acceptable for the active route. Decomposition is
+ * never load-bearing for correctness — the engine can always treat the
+ * pass-through as the original query.
+ *
+ * Out of scope (separate PRs):
+ *  - Engine wire-up that binds Hiroshi persona to `transport`.
+ *  - Adaptive router (primitive #4) that consumes the step list.
+ *  - Cost telemetry for the decomposer call itself.
+ */
+import { estimateTokens } from '../compact/token-counter.js';
+/** Whitelisted intent categories. Anything outside collapses to 'unknown'. */
+const INTENT_VALUES = ['read', 'write', 'verify', 'plan', 'explain', 'unknown'];
+const DEFAULT_MAX_STEPS = 8;
+const DEFAULT_MAX_INPUT_TOKENS = 4000;
+const STEP_ID_PATTERN = /^step-\d+$/;
+/**
+ * Prompt template embedded as a constant so the primitive has zero
+ * filesystem dependencies. The `{query}` marker is substituted at call
+ * time. The instruction explicitly forbids markdown fences and prose
+ * around the JSON, which trims the most common parse failures observed
+ * with mid-tier models.
+ */
+export const DECOMPOSER_PROMPT_TEMPLATE = [
+    'You are a query decomposer for a software-engineering AI assistant.',
+    "Split the user's request into atomic ordered steps.",
+    '',
+    'Return JSON only, matching this exact schema:',
+    '{ "steps": [{ "id": "step-N", "query": "...", "dependsOn": ["step-M", ...], "intent": "read|write|verify|plan|explain|unknown" }] }',
+    '',
+    'Rules:',
+    '- Max 8 steps.',
+    '- Single-step queries return one step.',
+    '- "dependsOn" lists earlier step IDs this step needs. Empty if independent.',
+    '- "intent" is the top action verb category.',
+    '- No prose around the JSON. No markdown fences.',
+    '',
+    'USER QUERY:',
+    '{query}',
+].join('\n');
+/**
+ * Decompose a query into ordered atomic sub-queries.
+ *
+ * The function never throws. Every internal failure mode collapses to
+ * a single-step fallback result. Caller inspects `usedFallback` and
+ * `reason` to decide whether to surface a warning or treat the
+ * decomposition as opaque.
+ */
+export async function decomposeQuery(query, opts) {
+    const trimmed = query.trim();
+    if (trimmed.length === 0) {
+        return fallback(query, 'empty-query');
+    }
+    const maxSteps = clampMaxSteps(opts.maxSteps);
+    const maxInputTokens = clampMaxInputTokens(opts.maxInputTokens);
+    const safeQuery = truncateToTokenBudget(trimmed, maxInputTokens);
+    const prompt = DECOMPOSER_PROMPT_TEMPLATE.replace('{query}', safeQuery);
+    let raw;
+    try {
+        raw = await opts.transport(prompt);
+    }
+    catch {
+        return fallback(safeQuery, 'transport-error');
+    }
+    const parseResult = parseAndValidate(raw, maxSteps);
+    if (parseResult.ok) {
+        return { steps: parseResult.steps, usedFallback: false };
+    }
+    return fallback(safeQuery, parseResult.reason);
+}
+function parseAndValidate(raw, maxSteps) {
+    let parsed;
+    try {
+        parsed = JSON.parse(stripFences(raw));
+    }
+    catch {
+        return { ok: false, reason: 'parse-failed' };
+    }
+    if (!isRecord(parsed)) {
+        return { ok: false, reason: 'schema-violation' };
+    }
+    const stepsRaw = parsed.steps;
+    if (!Array.isArray(stepsRaw)) {
+        return { ok: false, reason: 'schema-violation' };
+    }
+    if (stepsRaw.length === 0) {
+        return { ok: false, reason: 'empty-step-list' };
+    }
+    if (stepsRaw.length > maxSteps) {
+        return { ok: false, reason: 'exceeds-max-steps' };
+    }
+    const steps = [];
+    const seenIds = new Set();
+    for (const candidate of stepsRaw) {
+        const built = buildStep(candidate);
+        if (!built.ok) {
+            return { ok: false, reason: built.reason };
+        }
+        if (seenIds.has(built.step.id)) {
+            return { ok: false, reason: 'duplicate-id' };
+        }
+        seenIds.add(built.step.id);
+        steps.push(built.step);
+    }
+    for (const step of steps) {
+        for (const dep of step.dependsOn) {
+            if (!seenIds.has(dep)) {
+                return { ok: false, reason: 'unknown-dependency' };
+            }
+        }
+    }
+    if (hasCycle(steps)) {
+        return { ok: false, reason: 'dependency-cycle' };
+    }
+    return { ok: true, steps };
+}
+function buildStep(candidate) {
+    if (!isRecord(candidate)) {
+        return { ok: false, reason: 'schema-violation' };
+    }
+    const id = candidate.id;
+    const query = candidate.query;
+    const dependsOn = candidate.dependsOn;
+    const intent = candidate.intent;
+    if (typeof id !== 'string' || !STEP_ID_PATTERN.test(id)) {
+        return { ok: false, reason: 'schema-violation' };
+    }
+    if (typeof query !== 'string' || query.trim().length === 0) {
+        return { ok: false, reason: 'schema-violation' };
+    }
+    const depsList = [];
+    if (dependsOn !== undefined) {
+        if (!Array.isArray(dependsOn)) {
+            return { ok: false, reason: 'schema-violation' };
+        }
+        for (const dep of dependsOn) {
+            if (typeof dep !== 'string' || !STEP_ID_PATTERN.test(dep)) {
+                return { ok: false, reason: 'schema-violation' };
+            }
+            if (dep === id) {
+                // self-dependency is a degenerate cycle, surface explicitly
+                return { ok: false, reason: 'dependency-cycle' };
+            }
+            depsList.push(dep);
+        }
+    }
+    let intentValue;
+    if (intent === undefined) {
+        intentValue = 'unknown';
+    }
+    else if (typeof intent === 'string' && isIntent(intent)) {
+        intentValue = intent;
+    }
+    else {
+        return { ok: false, reason: 'schema-violation' };
+    }
+    return {
+        ok: true,
+        step: {
+            id,
+            query: query.trim(),
+            dependsOn: depsList,
+            intent: intentValue,
+        },
+    };
+}
+/**
+ * Cycle detection via iterative depth-first search with coloring.
+ * `WHITE` = unvisited, `GRAY` = on current stack, `BLACK` = fully
+ * explored. A back edge to a GRAY node is a cycle.
+ */
+function hasCycle(steps) {
+    const WHITE = 0;
+    const GRAY = 1;
+    const BLACK = 2;
+    const colors = new Map();
+    for (const step of steps)
+        colors.set(step.id, WHITE);
+    const byId = new Map(steps.map((s) => [s.id, s]));
+    for (const root of steps) {
+        if (colors.get(root.id) !== WHITE)
+            continue;
+        const stack = [{ id: root.id, depIdx: 0 }];
+        colors.set(root.id, GRAY);
+        while (stack.length > 0) {
+            const frame = stack[stack.length - 1];
+            if (!frame)
+                break;
+            const node = byId.get(frame.id);
+            if (!node) {
+                stack.pop();
+                continue;
+            }
+            if (frame.depIdx >= node.dependsOn.length) {
+                colors.set(node.id, BLACK);
+                stack.pop();
+                continue;
+            }
+            const next = node.dependsOn[frame.depIdx];
+            frame.depIdx += 1;
+            if (next === undefined)
+                continue;
+            const c = colors.get(next);
+            if (c === GRAY)
+                return true;
+            if (c === WHITE) {
+                colors.set(next, GRAY);
+                stack.push({ id: next, depIdx: 0 });
+            }
+        }
+    }
+    return false;
+}
+function fallback(query, reason) {
+    return {
+        steps: [
+            {
+                id: 'step-1',
+                query: query.trim(),
+                dependsOn: [],
+                intent: 'unknown',
+            },
+        ],
+        usedFallback: true,
+        reason,
+    };
+}
+function clampMaxSteps(value) {
+    if (typeof value !== 'number' || !Number.isFinite(value) || value < 1) {
+        return DEFAULT_MAX_STEPS;
+    }
+    return Math.floor(value);
+}
+function clampMaxInputTokens(value) {
+    if (typeof value !== 'number' || !Number.isFinite(value) || value < 1) {
+        return DEFAULT_MAX_INPUT_TOKENS;
+    }
+    return Math.floor(value);
+}
+/**
+ * Truncate input to the token budget using the existing chars/4
+ * heuristic. We slice by characters proportional to the budget rather
+ * than by binary-searching `estimateTokens`, because the heuristic is
+ * monotonic in length and an over-shoot is acceptable — the prompt
+ * also carries the static template overhead.
+ */
+function truncateToTokenBudget(text, maxTokens) {
+    const current = estimateTokens(text);
+    if (current <= maxTokens)
+        return text;
+    const ratio = maxTokens / current;
+    const targetLength = Math.max(1, Math.floor(text.length * ratio));
+    return text.slice(0, targetLength);
+}
+/**
+ * Tolerate the most common mid-tier-model deviation: wrapping JSON in
+ * a fenced code block. We strip a single leading and trailing fence
+ * marker when both are present; otherwise the raw text is returned
+ * unchanged and `JSON.parse` will reject it through the normal path.
+ */
+function stripFences(raw) {
+    const trimmed = raw.trim();
+    if (!trimmed.startsWith('```'))
+        return trimmed;
+    const firstNewline = trimmed.indexOf('\n');
+    if (firstNewline < 0)
+        return trimmed;
+    const closing = trimmed.lastIndexOf('```');
+    if (closing <= firstNewline)
+        return trimmed;
+    return trimmed.slice(firstNewline + 1, closing).trim();
+}
+function isRecord(value) {
+    return typeof value === 'object' && value !== null && !Array.isArray(value);
+}
+function isIntent(value) {
+    return INTENT_VALUES.includes(value);
+}
+//# sourceMappingURL=query-decomposer.js.map

package/dist/core/agents/registry.js

@@ -7,10 +7,10 @@
  * that mapping lives — keep it tight and explicit so persona drift never
  * leaks back into the dispatch surface.
  *
- * M1 closed set (ADR-0056 Sprint α5.1): 9 roles, all mapped to Tier 1
+ * M1 closed set : 9 roles, all mapped to Tier 1
  * Engineering Core or Tier 2 Specialist personas from THE_TEN. Tier 1
  * Missing Functions (Growth/Legal/Security/Sales/Support) are deferred
- * to α7.5; Sigma is intentionally absent because it is an OES Enterprise
+ * to ; Sigma is intentionally absent because it is an OES Enterprise
  * persona via Anvil triple-review proxy, not a Cyber-Zoo brand persona.
  */
 import { THE_TEN, getPersona } from '@pugi/personas';
@@ -30,7 +30,7 @@ function requirePersona(slug) {
 /**
  * CLI-only role-to-persona mapping. Roles are dispatcher-facing strings;
  * personas come from the brand-canonical THE_TEN. Vera (qa) intentionally
- * dual-roles as verifier + reviewer per ADR-0056 — the cabinet's review
+ * dual-roles as verifier + reviewer per  — the cabinet's review
  * pipeline already merges the two surfaces.
  */
 export const SUBAGENT_REGISTRY = [

package/dist/core/approvals/shortcut-resolver.js

@@ -0,0 +1,98 @@
+/**
+ * Numbered-shortcut + reject-with-reason resolver for approval UX .
+ *
+ * pattern: when an approval modal lists options, the user can
+ * press a digit (1-9) or single letter (y/n/a/d/q) to act without
+ * arrow-key navigation. Reject-with-reason lets the user type
+ * `n: didn't like the plan` so downstream callers know WHY и can
+ * potentially auto-revise.
+ *
+ * This module is the pure parsing layer — no Ink, no TUI. The TUI
+ * consumer wires `parseShortcut` to keyboard events и passes the
+ * result back к the engine.
+ */
+const ACCEPT_LETTERS = new Set(['y', 'a']);
+const REJECT_LETTERS = new Set(['n']);
+const ABORT_LETTERS = new Set(['q', 'x']);
+const DEFER_LETTERS = new Set(['d', 'l']);
+const MAX_DIGIT_SHORTCUT = 9;
+/**
+ * Parse a single keystroke or short input string into a decision.
+ *
+ * Supported forms (case-insensitive on letters):
+ *  "1" .. "9"              → accept option by 1-based index
+ *  "y" / "a"               → accept first option
+ *  "n"                     → reject (no reason)
+ *  "n: didn't like the AST" → reject with reason
+ *  "q" / "x"               → abort
+ *  "d" / "l"               → defer (decide later)
+ *  anything else           → unknown
+ *
+ * Whitespace around the input is trimmed. Reject-reason syntax accepts
+ * both `n:` и `n -` separators, plus the longer `reject: <reason>`.
+ */
+export function parseShortcut(input, options = []) {
+    const raw = input.trim();
+    if (raw.length === 0)
+        return { kind: 'unknown' };
+    // reject:<reason> или reject - <reason>
+    const rejectVerbose = /^reject\s*[:\-]\s*(.*)$/i.exec(raw);
+    if (rejectVerbose) {
+        const reason = (rejectVerbose[1] ?? '').trim();
+        return { kind: 'reject', reason: reason.length > 0 ? reason : null };
+    }
+    // n: <reason> / n - <reason>
+    const rejectShort = /^n\s*[:\-]\s*(.*)$/i.exec(raw);
+    if (rejectShort) {
+        const reason = (rejectShort[1] ?? '').trim();
+        return { kind: 'reject', reason: reason.length > 0 ? reason : null };
+    }
+    // Single digit 1..9 mapped to option index (1-based)
+    if (/^\d$/.test(raw)) {
+        const idx = Number.parseInt(raw, 10);
+        if (idx >= 1 && idx <= MAX_DIGIT_SHORTCUT) {
+            // If options provided, validate в range; else trust the index
+            if (options.length === 0 || idx <= options.length) {
+                return { kind: 'accept', index: idx };
+            }
+        }
+        return { kind: 'unknown' };
+    }
+    const lowered = raw.toLowerCase();
+    // Explicit per-option shortcut letter (e.g., "y" wired to option 1)
+    for (const opt of options) {
+        if (opt.shortcut && opt.shortcut.toLowerCase() === lowered) {
+            return { kind: 'accept', index: opt.index };
+        }
+    }
+    // Generic letters
+    if (ACCEPT_LETTERS.has(lowered))
+        return { kind: 'accept', index: 1 };
+    if (REJECT_LETTERS.has(lowered))
+        return { kind: 'reject', reason: null };
+    if (ABORT_LETTERS.has(lowered))
+        return { kind: 'abort' };
+    if (DEFER_LETTERS.has(lowered))
+        return { kind: 'defer' };
+    return { kind: 'unknown' };
+}
+/**
+ * Pre-render a help line listing shortcuts. Used by the TUI footer.
+ * Format: "[1-N] accept · [n] reject · [q] abort · [d] defer".
+ */
+export function formatShortcutHelp(options) {
+    const parts = [];
+    if (options.length > 0) {
+        if (options.length === 1) {
+            parts.push('[1] accept');
+        }
+        else {
+            parts.push(`[1-${Math.min(options.length, MAX_DIGIT_SHORTCUT)}] accept`);
+        }
+    }
+    parts.push('[n] reject ([n: reason] для details)');
+    parts.push('[d] defer');
+    parts.push('[q] abort');
+    return parts.join(' · ');
+}
+//# sourceMappingURL=shortcut-resolver.js.map

package/dist/core/artifact-chain/dispatcher.js

@@ -0,0 +1,148 @@
+/**
+ * Artifact chain dispatcher — Pugi  .
+ *
+ * Bridges the chain state machine (`state.ts`) to the existing
+ * `pugi delegate` surface. Each step renders a brief from the
+ * step's template + chain context, dispatches the brief to the
+ * step's persona, and writes the persona's response verbatim to the
+ * artifact file on disk.
+ *
+ * Mock-friendly by contract — the dispatcher accepts an injected
+ * `dispatch` function so specs can drive every branch without
+ * standing up a real runtime + Anvil round-trip. The real wire-up
+ * binds `dispatch` to the existing `submitDelegate` SDK helper +
+ * the SSE waiter; that wiring lives in `runtime/commands/chain.ts`
+ * so this module stays pure with respect to the network.
+ *
+ * Module contract:
+ *
+ *  - `dispatchStep` is the ONLY entry point. It reads the chain,
+ *    renders the brief, calls the injected dispatcher, persists the
+ *    artifact, and flips the step state. Callers do not touch the
+ *    state machine directly — that surface is intentionally narrow.
+ *
+ *  - The dispatcher never auto-advances. After a step lands the
+ *    cursor stays on the same step until the operator runs
+ *    `pugi chain next` again (or until `markComplete` is invoked
+ *    by the CLI handler after operator approval).
+ */
+import { mkdirSync, writeFileSync } from 'node:fs';
+import { join } from 'node:path';
+import { chainDir, markComplete, markDispatched, markError, readChain, } from './state.js';
+import { findStep, renderBrief, } from './steps.js';
+/**
+ * Dispatch one step in the chain. The function is non-throwing — every
+ * failure mode returns a structured result so the CLI handler can map
+ * to an exit code without a try/catch wrapper.
+ */
+export async function dispatchStep(options) {
+    const now = options.now ?? (() => new Date());
+    const state = readChain(options.workspaceCwd, options.chainId);
+    if (!state) {
+        // Gemini P1 fix (PR): the docblock above promises a non-
+        // throwing surface, but the historical implementation threw here.
+        // Return a structured result keyed on `reason: 'chain_not_found'`
+        // so callers can switch on the failure family deterministically.
+        return {
+            ok: false,
+            stepId: 'prd',
+            error: `chain ${options.chainId} not found`,
+            reason: 'chain_not_found',
+            state: null,
+        };
+    }
+    const targetStepId = options.stepId ?? state.nextStep;
+    if (!targetStepId) {
+        return {
+            ok: false,
+            stepId: 'code',
+            error: 'chain is finalised — every step is already complete',
+            state,
+        };
+    }
+    const step = findStep(targetStepId);
+    if (!step) {
+        return {
+            ok: false,
+            stepId: targetStepId,
+            error: `unknown step id '${targetStepId}'`,
+            state,
+        };
+    }
+    const stepRecord = state.steps.find((s) => s.id === targetStepId);
+    if (stepRecord?.status === 'complete') {
+        return {
+            ok: false,
+            stepId: targetStepId,
+            error: `step '${targetStepId}' is already complete`,
+            state,
+        };
+    }
+    const brief = renderBrief(step.briefTemplate, {
+        chainId: state.id,
+        intent: state.intent,
+    });
+    let outcome;
+    try {
+        outcome = await options.dispatch({ chainId: state.id, step, brief });
+    }
+    catch (err) {
+        const message = err instanceof Error ? err.message : String(err);
+        const errorState = markError(options.workspaceCwd, options.chainId, targetStepId, message, { now });
+        return {
+            ok: false,
+            stepId: targetStepId,
+            error: message,
+            state: errorState,
+        };
+    }
+    if (outcome.kind === 'failed') {
+        const errorState = markError(options.workspaceCwd, options.chainId, targetStepId, outcome.error, { now });
+        return {
+            ok: false,
+            stepId: targetStepId,
+            error: outcome.error,
+            state: errorState,
+        };
+    }
+    // Success path: persist the artifact, flip the step to dispatched
+    // (or complete when autoApprove is set).
+    const dir = chainDir(options.workspaceCwd, state.id);
+    mkdirSync(dir, { recursive: true });
+    const artifactPath = join(dir, step.artifactFilename);
+    writeFileSync(artifactPath, ensureTrailingNewline(outcome.artifact), 'utf8');
+    let finalState = markDispatched(options.workspaceCwd, options.chainId, targetStepId, outcome.dispatchId, { now });
+    if (options.autoApprove) {
+        finalState = markComplete(options.workspaceCwd, options.chainId, targetStepId, { now });
+    }
+    return {
+        ok: true,
+        stepId: targetStepId,
+        dispatchId: outcome.dispatchId,
+        artifactPath,
+        state: finalState,
+    };
+}
+/**
+ * Synthesize a dispatcher that captures every call for inspection.
+ * Test-only — the real wire-up never uses this.
+ */
+export function createRecordingDispatcher(outcomes) {
+    const calls = [];
+    let index = 0;
+    const fn = async ({ chainId, step, brief }) => {
+        calls.push({ chainId, stepId: step.id, brief });
+        const outcome = outcomes[index];
+        index += 1;
+        if (!outcome) {
+            throw new Error(`recording dispatcher exhausted at call #${calls.length}`);
+        }
+        return outcome;
+    };
+    fn.calls = calls;
+    return fn;
+}
+function ensureTrailingNewline(text) {
+    return text.endsWith('\n') ? text : `${text}\n`;
+}
+//# sourceMappingURL=dispatcher.js.map