@fiodos/cli 0.1.21 → 0.1.23

package/src/llm.js

@@ -1,144 +0,0 @@
-/**
- * Layer 3 — EXPLICIT LLM calls inside the pipeline.
- *
- * The static layers (1+2) extract facts; this layer makes judgment calls the
- * code cannot express, and every call is LOGGED so the final report can show
- * exactly which decisions came from the model vs from static analysis:
- *
- *   1. classifyApp        — app type vs the pattern catalog
- *   2. curateActions      — which detected callables deserve agent exposure,
- *                           category, sensitivity, parameters, label
- *   3. writeRouteMeta     — intent ids, labels and example phrases for routes
- *   4. patternGapFill     — given the matched pattern, propose what's missing
- *
- * Model: gpt-4o-mini via the OpenAI REST API (OPENAI_API_KEY env var).
- */
-'use strict';
-
-const LLM_LOG = [];
-
-function getLog() {
-  return LLM_LOG;
-}
-
-async function callLLM(taskName, system, user, { maxTokens = 2000 } = {}) {
-  const apiKey = process.env.OPENAI_API_KEY;
-  if (!apiKey) throw new Error('OPENAI_API_KEY is required for the LLM layer');
-
-  const t0 = Date.now();
-  const res = await fetch('https://api.openai.com/v1/chat/completions', {
-    method: 'POST',
-    headers: {
-      'Content-Type': 'application/json',
-      Authorization: `Bearer ${apiKey}`,
-    },
-    body: JSON.stringify({
-      model: process.env.AUTO_MANIFEST_MODEL || 'gpt-4o-mini',
-      messages: [
-        { role: 'system', content: system },
-        { role: 'user', content: user },
-      ],
-      temperature: 0.2,
-      max_tokens: maxTokens,
-      response_format: { type: 'json_object' },
-    }),
-  });
-  if (!res.ok) {
-    const body = await res.text();
-    throw new Error(`LLM call '${taskName}' failed: ${res.status} ${body.slice(0, 300)}`);
-  }
-  const data = await res.json();
-  const content = data.choices?.[0]?.message?.content || '{}';
-  let parsed;
-  try {
-    parsed = JSON.parse(content);
-  } catch {
-    throw new Error(`LLM call '${taskName}' returned non-JSON output`);
-  }
-  LLM_LOG.push({
-    task: taskName,
-    ms: Date.now() - t0,
-    promptChars: system.length + user.length,
-    usage: data.usage || null,
-  });
-  return parsed;
-}
-
-/** 1. Classify the app against the available pattern catalog. */
-async function classifyApp(appSummary, patternSummaries) {
-  const system =
-    'You classify a mobile app into one of the available integration patterns, or "none". ' +
-    'Answer JSON: {"pattern": "<id-or-none>", "confidence": 0..1, "appKind": "<free text>", "reason": "<short>"}';
-  const user =
-    `Available patterns:\n${patternSummaries.map((p) => `- id=${p.id}: ${p.summary}`).join('\n')}\n\n` +
-    `App evidence (from static analysis):\n${JSON.stringify(appSummary, null, 1).slice(0, 4000)}`;
-  return callLLM('classifyApp', system, user, { maxTokens: 300 });
-}
-
-/**
- * 2. Curate detected callables into agent actions.
- * Batched: large candidate lists are processed in chunks so no candidate is
- * silently dropped by prompt-size truncation.
- */
-async function curateActions(appKind, candidates) {
-  const BATCH = 14;
-  const all = [];
-  for (let i = 0; i < candidates.length; i += BATCH) {
-    const chunk = candidates.slice(i, i + BATCH);
-    const res = await curateActionsBatch(appKind, chunk);
-    all.push(...(res.actions || []));
-  }
-  return { actions: all };
-}
-
-async function curateActionsBatch(appKind, candidates) {
-  const system =
-    'You design the action surface of an in-app voice agent. Input: callables detected by static ' +
-    'analysis of a mobile app (handlers, store actions, buttons, alerts). For EACH candidate decide:\n' +
-    '- expose: should the voice agent be able to trigger it? (skip pure-UI plumbing like toggling pickers, ' +
-    'rendering callbacks, backdrop handlers, text-input onChange)\n' +
-    '- intent: snake_case id; label: short human label (English); category: one of discovery|workout|history|account|navigation|other\n' +
-    '- requireConfirmation: true only for destructive/sensitive operations (deleting data, losing progress, signing out, spending money)\n' +
-    '- confirmationMessage: required question if requireConfirmation\n' +
-    '- parameters: object of {name:{type,required,description}} ONLY when the user must provide a value by voice (e.g. a search query)\n' +
-    '- requiresVisibleContent: true when the action operates on an item the user is looking at\n' +
-    '- examples: 3-4 natural English voice phrases\n' +
-    '- evidence_used: which input evidence justified your decision\n' +
-    'Be conservative: when a candidate looks like internal plumbing, expose=false. ' +
-    'Sign-in/sign-up/verification flows requiring typed credentials should be expose=false ' +
-    '(a voice agent cannot dictate passwords safely).\n' +
-    'Return a decision for EVERY candidate. "intent" MUST be a plain string.\n' +
-    'Answer JSON: {"actions": [{...candidate fields above, "source_name": "<input name>"}]}';
-  const user =
-    `App kind: ${appKind}\n\nCandidates (with static evidence):\n` +
-    JSON.stringify(candidates, null, 1).slice(0, 14000);
-  return callLLM('curateActions', system, user, { maxTokens: 4000 });
-}
-
-/** 3. Route intents, labels, examples. */
-async function writeRouteMeta(appKind, routes) {
-  const system =
-    'You name navigation intents for an in-app voice agent. For each route give: ' +
-    'intent (snake_case), label (short English), examples (3-4 natural English voice phrases). ' +
-    'Skip auth screens the agent should not navigate to mid-session if any seem like sign-in/sign-up (mark skip=true). ' +
-    'For dynamic routes ([param]) mark dynamic=true and skip=true (they need a runtime parameter the static manifest cannot provide). ' +
-    'Answer JSON: {"routes": [{"routePath": "...", "intent": "...", "label": "...", "examples": [...], "skip": false, "skipReason": ""}]}';
-  const user = `App kind: ${appKind}\n\nRoutes:\n${JSON.stringify(routes, null, 1)}`;
-  return callLLM('writeRouteMeta', system, user, { maxTokens: 2500 });
-}
-
-/** 4. Pattern gap-fill: what does the matched pattern expect that we missed? */
-async function patternGapFill(patternId, patternText, draftManifestSummary) {
-  const system =
-    'You compare a draft voice-agent manifest against an integration pattern document. ' +
-    'List actions/routes the pattern suggests SHOULD usually exist for this app type but are missing or ' +
-    'under-specified in the draft. Only suggest things genuinely implied by the pattern; do not invent. ' +
-    'Answer JSON: {"suggestions": [{"kind": "action"|"route", "intent": "...", "label": "...", "why": "...", ' +
-    '"confidence": 0..1}], "patternUseful": true|false, "notes": "<honest assessment of how much this pattern helped>"}';
-  const user =
-    `Pattern (${patternId}):\n${patternText.slice(0, 5000)}\n\nDraft manifest summary:\n` +
-    JSON.stringify(draftManifestSummary, null, 1).slice(0, 5000);
-  return callLLM('patternGapFill', system, user, { maxTokens: 1200 });
-}
-
-module.exports = { classifyApp, curateActions, writeRouteMeta, patternGapFill, getLog };