@danieltmn/tamnora 0.11.3 → 0.13.0

package/harness/packages/server/public/index.html

@@ -37,7 +37,7 @@
   </head>
   <body>
     <header class="topbar">
-      <div class="brand"><span class="logo"><img src="/icons/logo.png" alt="Tamnora" /></span> <span id="topbar-section" class="topbar-section">chat</span> <span class="ver">v0.11.3</span></div>
+      <div class="brand"><span class="logo"><img src="/icons/logo.png" alt="Tamnora" /></span> <span id="topbar-section" class="topbar-section">chat</span> <span class="ver">v0.13.0</span></div>
       <nav class="tabs">
         <button class="tab active" data-tab="chat">chat</button>
         <button class="tab" data-tab="providers">proveedores</button>
@@ -461,6 +461,7 @@
             <button type="button" class="ajustes-tab" data-panel="sesiones" role="tab">Sesiones</button>
             <button type="button" class="ajustes-tab" data-panel="mcp" role="tab">MCP</button>
             <button type="button" class="ajustes-tab" data-panel="skills" role="tab">Skills</button>
+            <button type="button" class="ajustes-tab" data-panel="memoria" role="tab">Memoria</button>
             <button type="button" class="ajustes-tab" data-panel="telegram" role="tab">Telegram</button>
           <button type="button" class="ajustes-tab" data-panel="voz" role="tab">Voz</button>
           <button type="button" class="ajustes-tab" data-panel="asesor" role="tab">Asesor</button>
@@ -486,6 +487,9 @@
               <label class="check" title="aprobar automáticamente tools con permiso ask">
                 <input id="yes" type="checkbox" /> Aprobar automáticamente las tools (auto)
               </label>
+              <label class="check" title="el agente hace exactamente lo pedido, sin mejoras ni cambios fuera del pedido">
+                <input id="strict-scope" type="checkbox" /> Hacer solo lo pedido (estricto)
+              </label>
             </div>
           </div>
           <div class="row">
@@ -580,6 +584,25 @@
           <ul id="skill-list" class="providers"></ul>
         </div>
 
+        <div class="panel ajustes-panel" data-panel="memoria">
+          <div class="row between">
+            <h2>Memoria</h2>
+            <button id="memory-new" class="btn">+ tema</button>
+          </div>
+          <p class="muted verify-line">
+            Lo que el proyecto recuerda entre sesiones (<code>.tamnora/memory/</code>). El agente la mantiene con la
+            herramienta <code>memory</code>; acá la podés revisar y editar.
+          </p>
+          <label class="field"><span>índice (MEMORY.md)</span>
+            <textarea id="memory-index" rows="6" spellcheck="false" placeholder="# Memoria del proyecto"></textarea>
+          </label>
+          <div class="row between settings-actions">
+            <span id="memory-status" class="summary"></span>
+            <button id="memory-save-index" class="btn" type="button">guardar índice</button>
+          </div>
+          <ul id="memory-list" class="providers"></ul>
+        </div>
+
         <div class="panel ajustes-panel" data-panel="telegram">
           <div class="row between">
             <h2>Telegram</h2>
@@ -953,6 +976,31 @@
       </form>
     </dialog>
 
+    <dialog id="memory-modal" class="modal">
+      <form id="memory-form" class="modal-card">
+        <div class="row between">
+          <h3 id="memory-title">nuevo tema</h3>
+          <button type="button" id="memory-close" class="btn icon" title="cerrar" aria-label="cerrar">
+            <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true">
+              <path d="M18 6 6 18M6 6l12 12" />
+            </svg>
+          </button>
+        </div>
+        <div class="row">
+          <label class="field"><span>tema (slug)</span><input id="mem-name" spellcheck="false" placeholder="arquitectura" /></label>
+          <label class="field grow"><span>título</span><input id="mem-title" spellcheck="false" placeholder="Arquitectura" /></label>
+        </div>
+        <label class="field"><span>contenido (markdown)</span>
+          <textarea id="mem-body" rows="10" spellcheck="false" placeholder="Qué recordar del proyecto…"></textarea>
+        </label>
+        <div class="row between settings-actions">
+          <button type="button" id="mem-cancel" class="btn">cancelar</button>
+          <button type="submit" class="btn primary">guardar</button>
+        </div>
+        <div id="mem-status" class="summary"></div>
+      </form>
+    </dialog>
+
     <dialog id="palette-modal" class="modal palette-modal">
       <div class="palette">
         <div class="palette-row">

package/harness/packages/server/src/lib/agentRun.js

@@ -2,7 +2,7 @@
 // subagentes). La creacion de tools y el onEvent de cada canal siguen siendo
 // propios (los canales de salida son distintos); lo combinable vive aca.
 
-import { summarizeToolCall, previewToolResult } from './core.js';
+import { summarizeToolCall, previewToolResult, liveHistory } from './core.js';
 
 const FAILED_RE = /^Error|^Permiso denegado/i;
 
@@ -21,16 +21,15 @@ export function toolCallPayload(event) {
  * `approve` es el callback de aprobacion de las tools de administraación.
  */
 export function combineToolset({ local, admin = null, tasks = null, exposeAdmin = false, approve = null }) {
+  const adminAllowed = (name) => Boolean(admin?.handles(name)) && (exposeAdmin || admin.readOnly?.has(name));
   const definitions = [
     ...local.definitions,
-    ...(exposeAdmin && admin ? admin.definitions : []),
+    ...(admin ? admin.definitions.filter((d) => adminAllowed(d.function.name)) : []),
     ...(tasks ? tasks.definitions : []),
   ];
-  const execute = (name, args, context) => {
+  const execute = (name, args, context = {}) => {
     if (tasks?.handles(name)) return tasks.execute(name, args);
-    if (exposeAdmin && admin?.handles(name)) {
-      return admin.execute(name, args, approve ? { approve } : undefined);
-    }
+    if (adminAllowed(name)) return admin.execute(name, args, { approve, ...context });
     return local.execute(name, args, context);
   };
   return { definitions, execute };
@@ -67,12 +66,13 @@ export async function composeAgentMessages({ ctx, cwd, mode, history = [], userM
   const basePrompt = await ctx.systemPromptFor(cwd);
   const goal = sessionRecord?.meta?.goal;
   const summary = sessionRecord?.meta?.summary;
+  // Historial vigente: sin el asesor y sin lo ya resumido (si hay cursor).
+  const live = liveHistory(history, sessionRecord?.meta?.summaryCursor ?? null);
   return [
     { role: 'system', content: mode.systemPrompt ? `${basePrompt}\n\n${mode.systemPrompt}` : basePrompt },
     ...(goal ? [{ role: 'system', content: `Objetivo de la sesión: ${goal}` }] : []),
     ...(summary ? [{ role: 'system', content: `Resumen de la conversación anterior:\n${summary}` }] : []),
-    // El asesor no contamina el contexto del modelo: es una capa de comentario.
-    ...history.filter((message) => message.name !== 'advisor'),
+    ...live,
     userMessage,
   ];
 }

package/harness/packages/server/src/lib/agentTools.js

@@ -12,8 +12,14 @@ import {
   registryServerToConfig,
   normalizeMcpConfig,
   homeDir,
+  memorySlug,
+  listMemoryTopics,
+  readMemoryIndex,
+  readMemoryTopic,
+  writeMemory,
+  searchMemory,
 } from './core.js';
-import { SKILL_ID_PATTERN, skillsView, writeSkill, setSkillEnabled } from './skills.js';
+import { SKILL_ID_PATTERN, skillsView, writeSkill, setSkillEnabled, findSkill } from './skills.js';
 
 function sanitizeId(value) {
   return String(value ?? '')
@@ -51,6 +57,12 @@ const DEFINITIONS = [
   ),
   definition('remove_mcp', 'Elimina un servidor MCP guardado desde la consola.', { id: { type: 'string' } }, ['id']),
   definition('list_skills', 'Lista las skills instaladas y si están habilitadas.', {}),
+  definition(
+    'skill',
+    'Carga el contenido completo de una skill por su id. Usala cuando una skill del índice aplique a la tarea.',
+    { name: { type: 'string', description: 'id de la skill (ver el índice de skills)' } },
+    ['name']
+  ),
   definition(
     'install_skill',
     'Crea o actualiza una skill en la casa. Pasá "content" (markdown, opcionalmente con frontmatter) o "url" para descargarla.',
@@ -65,6 +77,18 @@ const DEFINITIONS = [
     },
     []
   ),
+  definition(
+    'memory',
+    'Memoria durable del proyecto, persiste entre sesiones. Acciones: list (temas), read (índice o un tema), search (busca) y write (crea/actualiza un tema). Guardá ahí lo que valga la pena recordar del proyecto.',
+    {
+      action: { type: 'string', enum: ['list', 'read', 'search', 'write'], description: 'Acción a ejecutar' },
+      topic: { type: 'string', description: 'Tema (slug) para read/write' },
+      query: { type: 'string', description: 'Texto a buscar (action=search)' },
+      title: { type: 'string', description: 'Título del tema (action=write)' },
+      content: { type: 'string', description: 'Contenido markdown (action=write)' },
+    },
+    ['action']
+  ),
 ];
 
 async function fetchText(url) {
@@ -84,24 +108,26 @@ async function fetchText(url) {
   }
 }
 
-function decisionFor(ctx, name) {
+function decisionFor(ctx, name, args) {
   const perms = ctx.config.permissions ?? {};
   const value = perms[name];
   if (value === 'allow' || value === 'ask' || value === 'deny') return value;
-  return name.startsWith('list_') ? 'allow' : 'ask';
+  if (name === 'memory') return String(args?.action ?? 'read').toLowerCase() === 'write' ? 'ask' : 'allow';
+  return name.startsWith('list_') || name === 'skill' ? 'allow' : 'ask';
 }
 
 /**
- * Crea las tools de administracion. `execute(name, args, { approve })`, donde
- * `approve(name, args)` resuelve true/false para los permisos "ask".
+ * Crea las tools de administracion. `execute(name, args, { approve, cwd })`,
+ * donde `approve(name, args)` resuelve true/false para los permisos "ask" y
+ * `cwd` es la carpeta de la sesion (para la memoria del proyecto).
  */
 export function createAgentTools(ctx) {
   function handles(name) {
     return DEFINITIONS.some((d) => d.function.name === name);
   }
 
-  async function execute(name, args = {}, { approve } = {}) {
-    const decision = decisionFor(ctx, name);
+  async function execute(name, args = {}, { approve, cwd } = {}) {
+    const decision = decisionFor(ctx, name, args);
     if (decision === 'deny') return `Permiso denegado para "${name}".`;
     if (decision === 'ask') {
       const ok = approve ? await approve(name, args) : false;
@@ -197,6 +223,13 @@ export function createAgentTools(ctx) {
       case 'list_skills':
         return JSON.stringify(skillsView(ctx).map(({ id, name, enabled, description }) => ({ id, name, enabled, description })), null, 2);
 
+      case 'skill': {
+        const id = String(args.name ?? args.id ?? '').trim();
+        const skill = findSkill(ctx, id);
+        if (!skill) return `No existe la skill "${id}".`;
+        return `# ${skill.name || skill.id}\n\n${skill.body}`;
+      }
+
       case 'install_skill': {
         const ref = parseSkillRef(args.ref ?? args.url ?? '');
         if (ref && (args.ref || (args.url && !/\.md(\?|$)/i.test(String(args.url))))) {
@@ -235,10 +268,47 @@ export function createAgentTools(ctx) {
         return `Skill "${id}" instalada.`;
       }
 
+      case 'memory': {
+        const root = cwd ?? ctx.cwd;
+        const action = String(args.action ?? 'read').toLowerCase();
+        if (action === 'list') {
+          const topics = listMemoryTopics(root);
+          if (!topics.length) return 'No hay memoria guardada todavía para este proyecto.';
+          return JSON.stringify(topics.map(({ name, title, bytes }) => ({ name, title, bytes })), null, 2);
+        }
+        if (action === 'search') {
+          const query = String(args.query ?? '').trim();
+          if (!query) return 'Pasá "query" para buscar en la memoria.';
+          const hits = searchMemory(root, query);
+          if (!hits.length) return `Sin resultados para "${query}".`;
+          return hits.map((h) => `### ${h.title} (${h.name})\n${h.snippet}`).join('\n\n');
+        }
+        if (action === 'write') {
+          const content = String(args.content ?? '').trim();
+          if (!content) return 'Pasá "content" para guardar en memoria.';
+          try {
+            const saved = await writeMemory(root, { topic: args.topic ?? args.title, title: args.title, content });
+            return `Guardado en memoria: ${saved.name} (${saved.title}).`;
+          } catch (error) {
+            return `No se pudo guardar la memoria: ${error.message}`;
+          }
+        }
+        const topic = String(args.topic ?? '').trim();
+        if (topic) {
+          const name = topic.endsWith('.md') ? topic : `${memorySlug(topic)}.md`;
+          const read = readMemoryTopic(root, name);
+          if (!read) return `No hay memoria del tema "${topic}".`;
+          return read.body;
+        }
+        return readMemoryIndex(root).trim() || 'No hay índice de memoria todavía.';
+      }
+
       default:
         return `Herramienta desconocida: ${name}`;
     }
   }
 
-  return { definitions: DEFINITIONS, handles, execute };
+  // Tools de admin que no modifican nada (lístar/leer): disponibles en todo modo.
+  const readOnly = new Set(['list_mcp', 'list_skills', 'skill']);
+  return { definitions: DEFINITIONS, handles, execute, readOnly };
 }

package/harness/packages/server/src/lib/approvals.js

@@ -48,6 +48,13 @@ export function createApprovalBroker({ timeoutMs = 300000 } = {}) {
       for (const id of ids) settle(id, false);
     },
 
+    /** Deny every pending request of a session (used when the user stops the run). */
+    cancelForSession(sessionId) {
+      for (const [id, entry] of pending) {
+        if (entry.sessionId === sessionId) settle(id, false);
+      }
+    },
+
     /** Deny everything still pending (used on shutdown). */
     cancelAll() {
       for (const id of [...pending.keys()]) settle(id, false);

package/harness/packages/server/src/lib/channels/telegram.js

@@ -25,6 +25,7 @@ import {
   toolCallPayload,
 } from '../agentRun.js';
 import { mp3AOgg } from '../ffmpeg.js';
+import { createRulesInjector } from '../rules.js';
 import { resolveTts } from '../ttsSettings.js';
 import { transcribeAudio } from '../transcribe.js';
 
@@ -46,6 +47,7 @@ const HELP = [
   '/model [id] — cambia el modelo (sin id: lista con botones)',
   '/provider [id] — cambia el proveedor (sin id: lista)',
   '/mode [plan|build] — cambia el modo (sin arg: panel)',
+  '/estricto [on|off] — el agente hace solo lo pedido (sin arg: alterna)',
   '/stop — cancela la respuesta en curso',
   '/help — esta ayuda',
   '',
@@ -504,10 +506,13 @@ export function startTelegramChannel(
         messages,
         tools: definitions,
         execute,
+        toolContext: { cwd, sessionId },
+        onLifecycle: createRulesInjector({ cwd, patterns: ctx.config.rules ?? [] }) ?? undefined,
         signal: controller.signal,
         maxRounds: ctx.effectiveMaxRounds?.() ?? ctx.config.maxRounds,
         maxTokens: ctx.effectiveMaxTokens?.() ?? 0,
         maxDurationMs: ctx.effectiveMaxDurationMs?.() ?? 0,
+        maxOutputTokens: ctx.effectiveMaxOutputTokens?.() ?? ctx.config.maxOutputTokens ?? 0,
         onEvent(event) {
           if (event.type === 'text') {
             full = event.full;
@@ -659,7 +664,7 @@ export function startTelegramChannel(
       const session = sessionId ? ctx.store.getSession(sessionId) : null;
       return send(
         chatId,
-        `proyecto: ${session ? projectLabel(session) : '—'}\nproveedor: ${session?.provider || ctx.effectiveProviderId()}\nmodelo: ${session?.model || ctx.effectiveModel() || '—'}\nmodo: ${ctx.effectiveMode()}\nsesión: ${session?.title || '—'}`
+        `proyecto: ${session ? projectLabel(session) : '—'}\nproveedor: ${session?.provider || ctx.effectiveProviderId()}\nmodelo: ${session?.model || ctx.effectiveModel() || '—'}\nmodo: ${ctx.effectiveMode()}\nsolo lo pedido: ${ctx.store.getSetting('strictScope') === true ? 'sí' : 'no'}\nsesión: ${session?.title || '—'}`
       );
     }
     if (cmd === 'resumen' || cmd === 'resume') {
@@ -698,6 +703,12 @@ export function startTelegramChannel(
       controllers.get(String(chatId))?.abort();
       return send(chatId, 'Cancelando…');
     }
+    if (cmd === 'estricto' || cmd === 'strict') {
+      const current = ctx.store.getSetting('strictScope') === true;
+      const value = arg ? /^(on|si|sí|1|true)$/i.test(arg) : !current;
+      ctx.store.setSetting('strictScope', value);
+      return send(chatId, `solo lo pedido: ${value ? 'activado' : 'desactivado'}`);
+    }
     return send(chatId, `Comando desconocido: /${cmd}.\n\n${HELP}`);
   }
 

package/harness/packages/server/src/lib/compact.js

@@ -0,0 +1,77 @@
+// Compactación de contexto: resumen no destructivo, compactación manual
+// (/compact) y auto-compactación antes de armar el prompt de un turno.
+
+import { runAgent, liveHistory, planCompaction, COMPACT_INSTRUCTION, SUMMARY_INSTRUCTION } from './core.js';
+
+export { COMPACT_INSTRUCTION, SUMMARY_INSTRUCTION };
+
+function transcript(messages) {
+  return messages
+    .map((m) => `${m.role}: ${typeof m.content === 'string' ? m.content : JSON.stringify(m.content)}`)
+    .join('\n');
+}
+
+function providerFor(ctx, session) {
+  const providerId = session.provider || ctx.effectiveProviderId();
+  const model = session.model || ctx.effectiveModel();
+  if (!model) return { error: 'Falta el modelo para resumir.', status: 400 };
+  const { provider, config } = ctx.resolveProvider({
+    ...ctx.providerSettings(providerId),
+    apiKey: ctx.apiKeyFor(providerId),
+  });
+  return { provider, config, model };
+}
+
+/** Corre una consulta de resumen sobre `messages` (opcionalmente con el resumen previo). */
+async function summarize(ctx, session, messages, instruction, previous) {
+  const resolved = providerFor(ctx, session);
+  if (resolved.error) return null;
+  const parts = [];
+  if (previous) parts.push(`RESUMEN PREVIO:\n${previous}`);
+  parts.push(`CONVERSACIÓN:\n${transcript(messages).slice(0, 60000)}`);
+  try {
+    const result = await runAgent({
+      provider: resolved.provider,
+      config: resolved.config,
+      model: resolved.model,
+      messages: [
+        { role: 'system', content: instruction },
+        { role: 'user', content: parts.join('\n\n') },
+      ],
+      tools: [],
+    });
+    return result.content?.trim() || null;
+  } catch {
+    return null;
+  }
+}
+
+/** Resumen no destructivo de toda la sesión (no toca el historial). */
+export async function summarizeSession(ctx, session, instruction) {
+  const messages = ctx.store.getMessages(session.id);
+  if (!messages.length) return { error: 'La sesión no tiene mensajes para resumir.', status: 400 };
+  const resolved = providerFor(ctx, session);
+  if (resolved.error) return resolved;
+  const summary = await summarize(ctx, session, messages, instruction);
+  if (!summary) return { error: 'El modelo no devolvió un resumen.', status: 502 };
+  return { summary };
+}
+
+/**
+ * Auto-compactación: si el historial vigente supera el umbral del contexto,
+ * resume lo viejo y lo reemplaza por `meta.summary` + `meta.summaryCursor`.
+ * Devuelve `{ summary, cursor, summarized }` o `null` si no hace falta.
+ */
+export async function maybeAutoCompact(ctx, session, history, { contextTokens } = {}) {
+  const cfg = ctx.config.compact ?? {};
+  if (cfg.enabled === false) return null;
+  const threshold = Number(cfg.threshold) > 0 && Number(cfg.threshold) < 1 ? Number(cfg.threshold) : 0.75;
+  const keepLast = Number.isFinite(Number(cfg.keepLast)) ? Number(cfg.keepLast) : 8;
+  const live = liveHistory(history, session.meta?.summaryCursor ?? null);
+  const plan = planCompaction(live, { contextTokens, threshold, keepLast });
+  if (!plan) return null;
+  const summary = await summarize(ctx, session, plan.toSummarize, COMPACT_INSTRUCTION, session.meta?.summary);
+  if (!summary) return null;
+  ctx.store.touchSession(session.id, { meta: { summary, summaryCursor: plan.cursor } });
+  return { summary, cursor: plan.cursor, summarized: plan.toSummarize.length };
+}

package/harness/packages/server/src/lib/prompt.js

@@ -1,4 +1,4 @@
-import { loadInstructions } from './core.js';
+import { loadInstructions, loadRules, buildStaticRulesBlock, memoryPrompt } from './core.js';
 
 export const DEFAULT_SYSTEM_PROMPT =
   'Eres Tamnora, un asistente útil. Responde en el idioma del usuario y usa las herramientas cuando aporten valor.';
@@ -33,14 +33,22 @@ function workspacePreamble(cwd) {
  * @param {string[]} opts.files - instruction file names (e.g. ['AGENTS.md'])
  * @param {() => string} [opts.profileText] - user profile block, re-read on every call
  * @param {() => string} [opts.skillsText] - skills block, re-read on every call
+ * @param {() => string} [opts.scopeText] - strict-scope block, re-read on every call
+ * @param {string[]} [opts.ruleFiles] - glob patterns of project rule files
+ * @param {boolean} [opts.memory] - inject the project memory block
  */
-export function createPromptBuilder({ cwd, files, profileText, skillsText }) {
+export function createPromptBuilder({ cwd, files, profileText, skillsText, scopeText, ruleFiles = [], memory = false }) {
   const cache = new Map();
 
   async function buildFor(workspace) {
     const instructions = (await loadInstructions(workspace, files)).trim();
     const base = workspacePreamble(workspace);
-    return instructions ? `${base}\n\n---\n\n${instructions}` : base;
+    let text = instructions ? `${base}\n\n---\n\n${instructions}` : base;
+    if (ruleFiles.length) {
+      const block = buildStaticRulesBlock(await loadRules(workspace, ruleFiles).catch(() => []));
+      if (block) text = `${text}\n\n---\n\n${block}`;
+    }
+    return text;
   }
 
   return async function systemPromptFor(workspace = cwd) {
@@ -52,6 +60,12 @@ export function createPromptBuilder({ cwd, files, profileText, skillsText }) {
     if (profile) parts.push(profile);
     const extra = typeof skillsText === 'function' ? String(skillsText() ?? '').trim() : '';
     if (extra) parts.push(extra);
+    if (memory) {
+      const mem = memoryPrompt(key);
+      if (mem) parts.push(mem);
+    }
+    const scope = typeof scopeText === 'function' ? String(scopeText() ?? '').trim() : '';
+    if (scope) parts.push(scope);
     return parts.join('\n\n---\n\n');
   };
 }

package/harness/packages/server/src/lib/rules.js

@@ -0,0 +1,43 @@
+// Reglas de proyecto por path inyectadas en runtime. Se engancha al seam de
+// ciclo de vida del agente: cuando una tool toca un archivo que matchea el
+// `paths:` de una regla, se agrega al contexto de las proximas rondas (efimera,
+// no se persiste). Dedup: una instancia por run, cada regla una vez.
+
+import { loadRules, rulesForPath, buildPathRulesBlock } from './core.js';
+
+const PATH_TOOLS = new Set(['read', 'write', 'edit', 'list']);
+
+/**
+ * Crea un handler `onLifecycle(event, payload)` para las reglas por path.
+ * Devuelve `null` si no hay patrones configurados.
+ */
+export function createRulesInjector({ cwd, patterns = [] }) {
+  if (!patterns.length) return null;
+  let cached = null;
+  let loading = null;
+  const injected = new Set();
+
+  async function all() {
+    if (cached) return cached;
+    if (!loading) {
+      loading = loadRules(cwd, patterns)
+        .catch(() => [])
+        .then((rules) => {
+          cached = rules;
+          return rules;
+        });
+    }
+    return loading;
+  }
+
+  return async function onLifecycle(event) {
+    if (event?.type !== 'tool.after' || !PATH_TOOLS.has(event.name)) return undefined;
+    const rel = event.args?.path ?? event.args?.file;
+    if (!rel || typeof rel !== 'string') return undefined;
+    const rules = await all();
+    const matches = rulesForPath(rules, rel).filter((rule) => !injected.has(rule.id));
+    if (!matches.length) return undefined;
+    for (const rule of matches) injected.add(rule.id);
+    return { inject: buildPathRulesBlock(matches, rel) };
+  };
+}

package/harness/packages/server/src/lib/skills.js

@@ -1,6 +1,6 @@
 import path from 'node:path';
 import { mkdir, rm, writeFile } from 'node:fs/promises';
-import { homeDir, loadSkills, mergeSkills, buildSkillsPrompt } from './core.js';
+import { homeDir, loadSkills, mergeSkills, buildSkillsPrompt, buildSkillsIndex } from './core.js';
 
 export const SKILL_ID_PATTERN = /^[a-z0-9][a-z0-9_-]{0,39}$/i;
 
@@ -46,6 +46,17 @@ export function skillsPrompt(ctx) {
   return buildSkillsPrompt(enabledSkills(ctx));
 }
 
+/** Índice compacto de skills (para carga on-demand con la tool `skill`). */
+export function skillsIndexPrompt(ctx) {
+  return buildSkillsIndex(enabledSkills(ctx));
+}
+
+/** Bloque de skills según `config.skills.mode`: `lazy` (default) o `eager`. */
+export function skillsBlock(ctx) {
+  const eager = ctx.config?.skills?.mode === 'eager';
+  return eager ? skillsPrompt(ctx) : skillsIndexPrompt(ctx);
+}
+
 export async function writeSkill(id, { name, description, when, body }) {
   const dir = path.join(homeDir(), 'skills', id);
   await mkdir(dir, { recursive: true });

package/harness/packages/server/src/lib/tasks.js

@@ -67,7 +67,7 @@ export function createSubagentTool(ctx, options) {
       Array.isArray(args.tools) && args.tools.length
         ? args.tools.filter((tool) => !exclude.includes(tool))
         : mode.tools ?? undefined;
-    const local = createTools({ root: cwd, enabled, permissions, yes: false, mcp });
+    const local = createTools({ root: cwd, enabled, permissions, yes: false, mcp, signal });
     const basePrompt = await ctx.systemPromptFor(cwd);
     const messages = [
       { role: 'system', content: `${basePrompt}\n\n${SUBAGENT_PROMPT}\n\nTarea: ${description}` },
@@ -88,6 +88,7 @@ export function createSubagentTool(ctx, options) {
         execute: local.execute,
         signal,
         maxRounds: ctx.config.agents?.maxRounds ?? 6,
+        maxOutputTokens: ctx.config.agents?.maxOutputTokens ?? ctx.config.maxOutputTokens ?? 0,
         onEvent(event) {
           if (event.type === 'text') emit('text', { full: event.full, agentId }, agentId);
           else if (event.type === 'reasoning') emit('reasoning', { full: event.full, agentId }, agentId);

package/harness/packages/server/src/routes/chat.js

@@ -18,6 +18,8 @@ import {
 } from '../lib/agentRun.js';
 import { fraseInicio, fraseFin, fraseError, createFiltroTools } from '../lib/announcer.js';
 import { resolveAdvisor, advisorShouldRun, consultAdvisor } from '../lib/advisor.js';
+import { maybeAutoCompact } from '../lib/compact.js';
+import { createRulesInjector } from '../lib/rules.js';
 
 /**
  * Emit streaming partials at most once per `ms` window, keeping only the
@@ -75,6 +77,9 @@ export function registerChat(route, ctx) {
   route('POST', '/api/sessions/:id/stop', ({ res, params }) => {
     if (!ctx.runs.isActive(params.id)) return sendJson(res, 200, { ok: true, stopped: false });
     ctx.runs.cancel(params.id);
+    // Un run esperando una aprobacion quedaria colgado: la denegamos para que
+    // el bucle del agente pueda desenrollarse y termine.
+    ctx.approvals.cancelForSession(params.id);
     sendJson(res, 200, { ok: true, stopped: true });
   });
 
@@ -200,6 +205,7 @@ export function registerChat(route, ctx) {
       yes: autoApprove,
       onAsk,
       mcp,
+      signal: ctx.runs.signal(sessionId),
     });
     // En modo plan (solo lectura) no exponemos las tools de administración.
     const exposeAdmin = !mode.tools;
@@ -225,7 +231,19 @@ export function registerChat(route, ctx) {
 
     const userMessage = buildUserMessage(input, images);
 
-    const sessionRecord = ctx.store.getSession(sessionId);
+    let sessionRecord = ctx.store.getSession(sessionId);
+    const contextTokens = Number(body.contextTokens) || ctx.config.contextTokens || 0;
+    // Auto-compactación: si el historial vigente ya no entra en el contexto,
+    // resumimos lo viejo y seguimos con el resumen + lo reciente. No borra nada.
+    try {
+      const compacted = await maybeAutoCompact(ctx, sessionRecord, history, { contextTokens });
+      if (compacted) {
+        ctx.runs.emit(sessionId, 'compacted', { summarized: compacted.summarized, contextTokens });
+        sessionRecord = ctx.store.getSession(sessionId);
+      }
+    } catch (error) {
+      ctx.runs.emit(sessionId, 'warning', { message: `auto-compact: ${error?.message ?? error}` });
+    }
     const messages = await composeAgentMessages({
       ctx,
       cwd: sessionCwd,
@@ -251,12 +269,16 @@ export function registerChat(route, ctx) {
           messages,
           tools: definitions,
           execute,
+          toolContext: { cwd: sessionCwd, sessionId },
+          onLifecycle: createRulesInjector({ cwd: sessionCwd, patterns: ctx.config.rules ?? [] }) ?? undefined,
           signal: ctx.runs.signal(sessionId),
           maxRounds: Number(body.maxRounds) || ctx.effectiveMaxRounds?.() || MAX_TOOL_ROUNDS,
           maxTokens: Number(body.maxTokens) || ctx.effectiveMaxTokens?.() || 0,
           maxDurationMs: Number(body.maxDurationMs) || ctx.effectiveMaxDurationMs?.() || 0,
+          maxOutputTokens: Number(body.maxOutputTokens) || ctx.effectiveMaxOutputTokens?.() || 0,
+          promptCaching: ctx.config.promptCaching === true,
           stallWarnMs: body.stallWarnMs ?? ctx.config.stallWarnMs,
-          contextTokens: Number(body.contextTokens) || ctx.config.contextTokens || 0,
+          contextTokens,
           onEvent(event) {
             if (event.type === 'text') partials.push('text', { delta: event.delta, full: event.full });
             else if (event.type === 'reasoning') partials.push('reasoning', { delta: event.delta, full: event.full });