jorgex-stack 1.9.57 → 1.9.58

package/dist/codex-WSWHXYJ3.js

@@ -0,0 +1,11 @@
+#!/usr/bin/env node
+import {
+  codexAdapter,
+  verifyOfficialSetup
+} from "./chunk-GOLJ5PA3.js";
+import "./chunk-KB5IQEKL.js";
+import "./chunk-NQZVJFV2.js";
+export {
+  codexAdapter,
+  verifyOfficialSetup
+};

package/dist/opencode-NUDGK6J7.js

@@ -0,0 +1,31 @@
+#!/usr/bin/env node
+import {
+  checkOpencodeDuplicates,
+  checkOpencodeOfficialMcp,
+  checkOpencodeOfficialStatusline,
+  collectUnverifiableOpencodeConfigs,
+  inspectOpencodeEngramPlugin,
+  inspectOpencodePluginFile,
+  isOfficialOpencodePluginContent,
+  opencodeAdapter,
+  readOpencodeConfigFile,
+  shouldRetireLegacyEngram,
+  transferEngramOwnership,
+  verifyOfficialSetup
+} from "./chunk-M6QX2F3A.js";
+import "./chunk-KB5IQEKL.js";
+import "./chunk-NQZVJFV2.js";
+export {
+  checkOpencodeDuplicates,
+  checkOpencodeOfficialMcp,
+  checkOpencodeOfficialStatusline,
+  collectUnverifiableOpencodeConfigs,
+  inspectOpencodeEngramPlugin,
+  inspectOpencodePluginFile,
+  isOfficialOpencodePluginContent,
+  opencodeAdapter,
+  readOpencodeConfigFile,
+  shouldRetireLegacyEngram,
+  transferEngramOwnership,
+  verifyOfficialSetup
+};

package/dist/quality-verifier.js

@@ -3,10 +3,12 @@ import {
   QUALITY_PROFILES,
   canonicalJson,
   evaluateQualityPolicy,
-  samePath,
   sha256,
   validateQualityReceipt
-} from "./chunk-IO2FFZOH.js";
+} from "./chunk-N5IC2JBJ.js";
+import {
+  samePath
+} from "./chunk-NQZVJFV2.js";
 
 // src/lib/quality-verifier.ts
 import { createHash } from "crypto";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jorgex-stack",
-  "version": "1.9.57",
+  "version": "1.9.58",
   "description": "Harness multi-agente portable: instala la config JorgeX (agentes, skills, hooks, Engram, MCPs) en Claude Code, Codex CLI, OpenCode y Pi",
   "type": "module",
   "license": "MIT",

package/stack/plugins/opencode/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@jorgex/opencode-plugins",
   "version": "1.0.0",
-  "description": "OpenCode plugins del stack: engram (memoria persistente), hooks (puente de scripts por evento/tool) y worktree (gestión de git worktrees)",
+  "description": "OpenCode plugins del stack: hooks (puente de scripts por evento/tool) y worktree (gestión de git worktrees). Engram lo provee `engram setup opencode` (plugin oficial).",
   "type": "module",
   "main": "hooks.ts",
   "dependencies": {

package/upstreams.json

@@ -9,48 +9,55 @@
     }
   },
   "complements": {
-    "$comment": "Complementos oficiales externos (ejecutan código en el runtime): pin exacto al instalar, nunca latest ni rangos flotantes. `reviewed` = última revisión aceptada; `update --check` es discovery-only y avisa si el upstream se movió; la adopción es re-pin deliberado con verificación, nunca auto-update.",
+    "$comment": "Complementos oficiales externos (ejecutan código en el runtime): `strategy` distingue `exact` (pin obligatorio, ausencia es error; al instalar se usa el pin, nunca latest ni rangos) de `provider-managed` (el provider gestiona la versión: rolling/null/main/latest aceptados sin warning por ausencia de pin; update --check informa la versión observada/drift sin bloquear). `reviewed` = última revisión aceptada; `update --check` es discovery-only, nunca auto-update; la adopción exact es re-pin deliberado con verificación.",
     "gentle-engram": {
       "source": "npm:gentle-engram",
-      "version": "0.1.12",
-      "reviewed": "2026-09-17",
+      "version": "0.1.14",
+      "strategy": "provider-managed",
+      "reviewed": "2026-09-20",
       "status": "pendiente-adopcion",
-      "note": "última publicada en npm (dist-tag latest); el setup oficial en main referencia 0.1.13 aún no publicada — re-pinear al salir. Incluye auto-cierre en session_shutdown. Issue: #155"
+      "note": "publicado/latest en npm y usado por el setup oficial en main: provider gestiona la versión. Incluye auto-cierre en session_shutdown. Issue: #155"
     },
     "pi-mcp-adapter": {
       "source": "npm:pi-mcp-adapter",
       "version": null,
-      "reviewed": "2026-09-17",
+      "strategy": "provider-managed",
+      "reviewed": "2026-09-20",
       "status": "pendiente-pin",
-      "note": "upstream lo referencia sin versión (setup.go main): exigir pin exacto antes de adoptar. Issue: #155"
+      "note": "provider-managed: upstream lo referencia sin versión (setup.go main); rolling aceptado, sin warning por ausencia de pin. Último observado latest 2.34.0 como referencia informativa, sin fijarlo. Issue: #155"
     },
     "engram-claude-plugin": {
       "source": "github:Gentleman-Programming/engram#plugin/claude-code",
-      "version": "1.20.0",
-      "reviewed": "2026-09-17",
-      "note": "autocierre vía hook Stop (v1.20.0) / SessionEnd (main). Stack convive (detección + allowlists duales); falta instalación/verificación por Stack. Issue: #155"
+      "version": "2.0.0",
+      "strategy": "provider-managed",
+      "reviewed": "2026-09-20",
+      "note": "provider setup según upstream (2.0.0 solo referencia observada, no pin de instalación); autocierre vía hook Stop (v2.0.0) / SessionEnd (main). Stack convive (detección + allowlists duales); falta instalación/verificación por Stack. Issue: #155"
     },
     "engram-codex-plugin": {
       "source": "github:Gentleman-Programming/engram#plugin/codex",
-      "version": "1.20.0",
-      "reviewed": "2026-09-17",
-      "note": "autocierre vía hook Stop. Stack convive; falta instalación/verificación por Stack. Issue: #155"
+      "version": "main",
+      "strategy": "provider-managed",
+      "reviewed": "2026-09-20",
+      "note": "main rolling aprobado (provider-managed): autocierre vía hook Stop. Stack convive; falta instalación/verificación por Stack. Issue: #155"
     },
     "engram-opencode-plugin": {
       "source": "github:Gentleman-Programming/engram#plugin/opencode",
-      "version": "1.20.0",
-      "reviewed": "2026-09-17",
-      "note": "auto-cierre (endSessionInEngram) solo en main; la copia vendorizada de Stack no lo tiene — migrar con transferencia de ownership de plugins/engram.ts. Issue: #155"
+      "version": "2.0.0",
+      "strategy": "provider-managed",
+      "reviewed": "2026-09-20",
+      "note": "provider setup según upstream (2.0.0 solo referencia observada, no pin de instalación); auto-cierre (endSessionInEngram) solo en main; la copia vendorizada de Stack no lo tiene — migrar con transferencia de ownership de plugins/engram.ts. Issue: #155"
     },
     "chrome-devtools-mcp": {
       "source": "npm:chrome-devtools-mcp",
       "version": "1.6.0",
+      "strategy": "exact",
       "reviewed": "2026-09-17",
       "note": "opt-in, default-off; flags --isolated --redact-network-headers --no-performance-crux --no-usage-statistics"
     },
     "playwright-cli": {
       "source": "npm:@playwright/cli",
       "version": "0.1.18",
+      "strategy": "exact",
       "reviewed": "2026-09-17",
       "note": "CLI global compartido por la máquina; camino recomendado opt-in"
     }

package/stack/plugins/opencode/engram.ts

@@ -1,444 +0,0 @@
-/**
- * Engram — OpenCode plugin adapter
- *
- * Thin layer that connects OpenCode's event system to the Engram Go binary.
- * The Go binary runs as a local HTTP server and handles all persistence.
- *
- * Flow:
- *   OpenCode events → this plugin → HTTP calls → engram serve → SQLite
- *
- * Session resilience:
- *   Uses `ensureSession()` before any DB write. This means sessions are
- *   created on-demand — even if the plugin was loaded after the session
- *   started (restart, reconnect, etc.). The session ID comes from OpenCode's
- *   hooks (input.sessionID) rather than relying on a session.created event.
- */
-
-import type { Plugin } from "@opencode-ai/plugin"
-
-declare const Bun: {
-  which?: (bin: string) => string | null
-  spawnSync: (args: string[], options?: Record<string, unknown>) => { exitCode: number; stdout?: { toString(): string } | string }
-  spawn: (args: string[], options?: Record<string, unknown>) => unknown
-  file: (path: string) => { exists: () => Promise<boolean> }
-}
-
-// ─── Configuration ───────────────────────────────────────────────────────────
-
-const ENGRAM_PORT = parseInt(process.env.ENGRAM_PORT ?? "7437")
-const ENGRAM_URL = `http://127.0.0.1:${ENGRAM_PORT}`
-// "{{ENGRAM_BIN}}" lo resuelve el instalador con el binario detectado (D7).
-const ENGRAM_BIN = "{{ENGRAM_BIN}}"
-
-export function resolveEngramBin(installerBin = ENGRAM_BIN): string {
-  const envBin = process.env.ENGRAM_BIN
-  if (envBin) return envBin
-
-  const bun = globalThis as typeof globalThis & { Bun?: { which?: (bin: string) => string | null } }
-  const bunBin = bun.Bun?.which?.("engram")
-  if (bunBin) return bunBin
-
-  return installerBin !== "{{ENGRAM_BIN}}" ? installerBin : "engram"
-}
-
-// Engram's own MCP tools — don't count these as "tool calls" for session stats
-const ENGRAM_TOOLS = new Set([
-  "mem_search",
-  "mem_save",
-  "mem_update",
-  "mem_delete",
-  "mem_suggest_topic_key",
-  "mem_save_prompt",
-  "mem_session_summary",
-  "mem_context",
-  "mem_stats",
-  "mem_timeline",
-  "mem_get_observation",
-  "mem_session_start",
-  "mem_session_end",
-])
-
-// ─── Memory Instructions ─────────────────────────────────────────────────────
-// These get injected into the agent's context so it knows to call mem_save.
-// El placeholder ENGRAM_PROTOCOL lo resuelve el instalador con el contenido
-// del protocolo canónico (stack/system-prompt/engram-protocol.md) — única
-// fuente, alineada automáticamente con lo que reciben Claude Code y Codex.
-
-const MEMORY_INSTRUCTIONS = "{{ENGRAM_PROTOCOL}}"
-
-// ─── HTTP Client ─────────────────────────────────────────────────────────────
-
-async function engramFetch(
-  path: string,
-  opts: { method?: string; body?: any } = {}
-): Promise<any> {
-  try {
-    const res = await fetch(`${ENGRAM_URL}${path}`, {
-      method: opts.method ?? "GET",
-      headers: opts.body ? { "Content-Type": "application/json" } : undefined,
-      body: opts.body ? JSON.stringify(opts.body) : undefined,
-    })
-    if (!res.ok) return null
-    return await res.json()
-  } catch {
-    // Engram server not running — silently fail
-    return null
-  }
-}
-
-async function isEngramRunning(): Promise<boolean> {
-  try {
-    const res = await fetch(`${ENGRAM_URL}/health`, {
-      signal: AbortSignal.timeout(500),
-    })
-    return res.ok
-  } catch {
-    return false
-  }
-}
-
-// ─── Helpers ─────────────────────────────────────────────────────────────────
-
-function extractProjectName(directory: string): string {
-  // Try git remote origin URL
-  try {
-    const result = Bun.spawnSync(["git", "-C", directory, "remote", "get-url", "origin"], { stdout: "pipe", stderr: "ignore", stdin: "ignore" })
-    if (result.exitCode === 0) {
-      const url = result.stdout?.toString().trim()
-      if (url) {
-        const name = url.replace(/\.git$/, "").split(/[/:]/).pop()
-        if (name) return name
-      }
-    }
-  } catch {}
-
-  // Fallback: git root directory name (works in worktrees)
-  try {
-    const result = Bun.spawnSync(["git", "-C", directory, "rev-parse", "--show-toplevel"], { stdout: "pipe", stderr: "ignore", stdin: "ignore" })
-    if (result.exitCode === 0) {
-      const root = result.stdout?.toString().trim()
-      if (root) return root.split(/[\\/]/).pop() ?? "unknown"
-    }
-  } catch {}
-
-  // Final fallback: cwd basename
-  return directory.split(/[\\/]/).pop() ?? "unknown"
-}
-
-function truncate(str: string, max: number): string {
-  if (!str) return ""
-  return str.length > max ? str.slice(0, max) + "..." : str
-}
-
-/**
- * Strip <private>...</private> tags before sending to engram.
- * Double safety: the Go binary also strips, but we strip here too
- * so sensitive data never even hits the wire.
- */
-function stripPrivateTags(str: string): string {
-  if (!str) return ""
-  return str.replace(/<private>[\s\S]*?<\/private>/gi, "[REDACTED]").trim()
-}
-
-/**
- * Log errors and warnings to OpenCode's TUI log.
- * Errors in hooks must not propagate to the plugin surface — they are silently
- * swallowed here so the TUI does not crash. The error is visible only in OpenCode's
- * log stream (visible to the user via `opencode logs`).
- */
-async function logToOpenCode(client: unknown, level: string, message: string, extra?: unknown): Promise<void> {
-  if (typeof client !== "object" || client === null) return;
-  const app = (client as { app?: unknown }).app;
-  if (typeof app !== "object" || app === null) return;
-  const log = (app as { log?: unknown }).log;
-  if (typeof log !== "function") return;
-  const safeExtra = extra instanceof Error ? { message: extra.message, stack: extra.stack } : extra;
-  try {
-    await log.call(app, {
-      body: {
-        service: "engram",
-        level,
-        message,
-        extra: safeExtra,
-      },
-    });
-  } catch {
-    // Logging must never break the plugin path.
-  }
-}
-
-// ─── Plugin Export ───────────────────────────────────────────────────────────
-
-export const Engram: Plugin = async (ctx) => {
-  const oldProject = ctx.directory.split(/[\\/]/).pop() ?? "unknown"
-  const project = extractProjectName(ctx.directory)
-  const client = (ctx as { client?: unknown }).client
-  const engramBin = resolveEngramBin()
-
-  // Track tool counts per session (in-memory only, not critical)
-  const toolCounts = new Map<string, number>()
-
-  // Track which sessions we've already ensured exist in engram
-  const knownSessions = new Set<string>()
-
-  // Track sub-agent session IDs so we can suppress their tool-hook registrations.
-  // Sub-agents (Task() calls) have a parentID or a title ending in " subagent)".
-  // We must not register them as top-level Engram sessions — they cause session
-  // inflation (e.g. 170 sessions for 1 real conversation, issue #116).
-  const subAgentSessions = new Set<string>()
-
-  /**
-   * Ensure a session exists in engram. Idempotent — calls POST /sessions
-   * which uses INSERT OR IGNORE. Safe to call multiple times.
-   *
-   * Silently skips sub-agent sessions (tracked in `subAgentSessions`).
-   */
-  async function ensureSession(sessionId: string): Promise<boolean> {
-    if (!sessionId) return false
-    if (knownSessions.has(sessionId)) return true
-    // Do not register sub-agent sessions in Engram (issue #116).
-    if (subAgentSessions.has(sessionId)) return false
-    const session = await engramFetch("/sessions", {
-      method: "POST",
-      body: {
-        id: sessionId,
-        project,
-        directory: ctx.directory,
-      },
-    })
-
-    if (session === null) return false
-
-    knownSessions.add(sessionId)
-    return true
-  }
-
-  // Try to start engram server if not running
-  const running = await isEngramRunning()
-  if (!running) {
-    try {
-      Bun.spawn([engramBin, "serve"], {
-        stdout: "ignore",
-        stderr: "ignore",
-        stdin: "ignore",
-      })
-      await new Promise((r) => setTimeout(r, 500))
-    } catch {
-      // Binary not found or can't start — plugin will silently no-op
-    }
-  }
-
-  // Migrate project name if it changed (one-time, idempotent)
-  // Must run AFTER server startup to ensure the endpoint is available
-  if (oldProject !== project) {
-    await engramFetch("/projects/migrate", {
-      method: "POST",
-      body: { old_project: oldProject, new_project: project },
-    })
-  }
-
-  // Auto-import: if .engram/manifest.json exists in the project repo,
-  // run `engram sync --import` to load any new chunks into the local DB.
-  // This is how git-synced memories get loaded when cloning a repo or
-  // pulling changes. Each chunk is imported only once (tracked by ID).
-  try {
-    const manifestFile = `${ctx.directory}/.engram/manifest.json`
-    const file = Bun.file(manifestFile)
-    if (await file.exists()) {
-      Bun.spawn([engramBin, "sync", "--import"], {
-        cwd: ctx.directory,
-        stdout: "ignore",
-        stderr: "ignore",
-        stdin: "ignore",
-      })
-    }
-  } catch {
-    // Manifest doesn't exist or binary not found — silently skip
-  }
-
-  // ─── Safe Hook Wrapper ─────────────────────────────────────────
-  // Wraps every hook to catch errors and log them to OpenCode instead
-  // of letting them propagate. This prevents malformed payloads or
-  // missing fields (output.context, output.system, etc.) from crashing
-  // the TUI. Errors are logged and swallowed (promise resolves cleanly).
-  const safe = <A extends unknown[]>(name: string, fn: (...args: A) => Promise<void>) =>
-    async (...args: A): Promise<void> => {
-      try { await fn(...args) }
-      catch (error) { void logToOpenCode(client, "error", `Engram hook ${name} failed`, error) }
-    }
-
-  return {
-    // ─── Event Listeners ───────────────────────────────────────────
-
-    event: safe("event", async ({ event }: { event: { type: string; properties?: unknown } }) => {
-      // --- Session Created ---
-      if (event.type === "session.created") {
-        // Bug fix (#116): session data is nested under event.properties.info,
-        // not event.properties directly.
-        const info = (event.properties as any)?.info
-        const sessionId = info?.id
-        const parentID = info?.parentID
-        const title: string = info?.title ?? ""
-
-        // Sub-agent sessions (created via Task()) must NOT be registered as
-        // top-level Engram sessions. They cause massive session inflation
-        // (e.g. 170 sessions for 1 real conversation).
-        //
-        // Detection heuristics:
-        //   - parentID is set on all Task() sub-agent sessions
-        //   - title ends with " subagent)" as a secondary signal
-        const isSubAgent = !!parentID || title.endsWith(" subagent)")
-
-        if (sessionId && !isSubAgent) {
-          await ensureSession(sessionId)
-        } else if (sessionId && isSubAgent) {
-          // Remember this as a sub-agent session so tool-hook calls
-          // to ensureSession() are also suppressed for it.
-          subAgentSessions.add(sessionId)
-        }
-      }
-
-      // --- Session Deleted ---
-      if (event.type === "session.deleted") {
-        // Same properties.info path as session.created.
-        const info = (event.properties as any)?.info
-        const sessionId = info?.id
-        if (sessionId) {
-          toolCounts.delete(sessionId)
-          knownSessions.delete(sessionId)
-          subAgentSessions.delete(sessionId)
-        }
-      }
-
-    }),
-
-    // ─── User Prompt Capture ──────────────────────────────────────
-    // chat.message is called once per user message, before the LLM sees it.
-    // input.sessionID is always reliable here (no knownSessions workaround).
-    // output.message is typed as UserMessage (role:"user" already guaranteed).
-    // output.parts contains TextPart[] with the actual message text.
-
-    "chat.message": safe("chat.message", async (input: any, output: any) => {
-      // Skip sub-agent sessions — they inflate session counts (issue #116)
-      if (subAgentSessions.has(input.sessionID)) return
-
-      const sessionId = input.sessionID
-
-      // Extract text from parts (type:"text")
-      const content = output.parts
-        .filter((p: any) => p.type === "text")
-        .map((p: any) => p.text ?? "")
-        .join("\n")
-        .trim()
-
-      // Also fallback to summary if parts yield nothing
-      const fallback = !content && output.message.summary
-        ? `${output.message.summary.title ?? ""}\n${output.message.summary.body ?? ""}`.trim()
-        : ""
-
-      const finalContent = content || fallback
-
-      // Only capture non-trivial prompts (>10 chars)
-      if (finalContent.length > 10) {
-        const sessionReady = await ensureSession(sessionId)
-        if (sessionReady) {
-          await engramFetch("/prompts", {
-            method: "POST",
-            body: {
-              session_id: sessionId,
-              content: stripPrivateTags(truncate(finalContent, 2000)),
-              project,
-            },
-          })
-        }
-      }
-    }),
-
-    // ─── Tool Execution Hook ─────────────────────────────────────
-    // Count tool calls per session (for session end stats).
-    // Also ensures the session exists — handles plugin reload / reconnect.
-    // Passive capture: when a Task tool completes, POST its output to
-    // the passive capture endpoint so the server extracts learnings.
-
-    "tool.execute.after": safe("tool.execute.after", async (input: any, output: any) => {
-      if (ENGRAM_TOOLS.has(input.tool.toLowerCase())) return
-
-      // input.sessionID comes from OpenCode — always available
-      const sessionId = input.sessionID
-      const sessionReady = sessionId ? await ensureSession(sessionId) : false
-      if (sessionReady && sessionId) {
-        toolCounts.set(sessionId, (toolCounts.get(sessionId) ?? 0) + 1)
-      }
-
-      // Passive capture: extract learnings from Task tool output
-      // (OpenCode reports the tool name in lowercase: "task")
-      if (input.tool.toLowerCase() === "task" && output && sessionId) {
-        const text = typeof output === "string" ? output : JSON.stringify(output)
-        if (text.length > 50 && sessionReady) {
-          await engramFetch("/observations/passive", {
-            method: "POST",
-            body: {
-              session_id: sessionId,
-              content: stripPrivateTags(text),
-              project,
-              source: "task-complete",
-            },
-          })
-        }
-      }
-    }),
-
-    // ─── System Prompt: Always-on memory instructions ──────────
-    // Injects MEMORY_INSTRUCTIONS into the system prompt of every message.
-    // This ensures the agent ALWAYS knows about Engram, even after compaction.
-    //
-    // We append to the last existing system entry instead of pushing a new one.
-    // Some models (Qwen3.5, Mistral/Ministral via llama.cpp) reject multiple
-    // system messages — their Jinja chat templates only allow a single system
-    // block at the beginning. By concatenating, we avoid adding extra system
-    // messages that would break these models. See: GitHub issue #23.
-
-    "experimental.chat.system.transform": safe("experimental.chat.system.transform", async (_input: any, output: any) => {
-      if (output.system.length > 0) {
-        output.system[output.system.length - 1] += "\n\n" + MEMORY_INSTRUCTIONS
-      } else {
-        output.system.push(MEMORY_INSTRUCTIONS)
-      }
-    }),
-
-    // ─── Compaction Hook: Persist memory + inject context ──────────
-    // Compaction is triggered by the system (not the agent) when context
-    // gets too long. The old agent "dies" and a new one starts with the
-    // compacted summary. This is our chance to:
-    // 1. Auto-save a session checkpoint (the agent can't do this itself)
-    // 2. Inject context from previous sessions into the compaction prompt
-    // 3. Tell the compressor to remind the new agent to save memories
-
-    "experimental.session.compacting": safe("experimental.session.compacting", async (input: any, output: any) => {
-      if (input.sessionID) {
-        await ensureSession(input.sessionID)
-      }
-
-      // Inject context from previous sessions
-      const data = await engramFetch(
-        `/context?project=${encodeURIComponent(project)}`
-      )
-      if (data?.context) {
-        output.context.push(data.context)
-      }
-
-      // Tell the compressor to instruct the new agent to persist the
-      // compacted summary to Engram. The new agent reads the compacted
-      // summary and this instruction, then saves it as a session summary.
-      output.context.push(
-        `CRITICAL INSTRUCTION FOR COMPACTED SUMMARY:\n` +
-        `The agent has access to Engram persistent memory via MCP tools.\n` +
-        `You MUST include the following instruction at the TOP of the compacted summary:\n\n` +
-        `"FIRST ACTION REQUIRED: Call mem_session_summary with the content of this compacted summary. ` +
-        `Use project '${project}' if the tool accepts it. This preserves what was accomplished before compaction. ` +
-        `Do this BEFORE any other work. After that, delegate to the global engram subagent to recover only the relevant previous context."\n\n` +
-        `This is NOT optional. Without this, everything done before compaction is lost from memory.`
-      )
-    }),
-  }
-}