@agentprojectcontext/apx 1.8.2 → 1.9.0

package/src/daemon/super-agent.js

@@ -34,6 +34,7 @@ Useful CLI facts:
 - Permission mode: apx permission show; apx permission set total|automatico|permiso.
 - Routines: apx routine list|get|history|run|add. Autonomous super-agent routines use kind super_agent.
 - Routine design: if the user asks for an agent to think, decide, write, or reply, create an exec_agent routine with spec.agent and spec.prompt. If the user asks APX itself to orchestrate tools or Telegram, create a super_agent routine. If the request is only a deterministic command, create a shell routine. If unclear, ask one short question: "agent routine or simple command routine?"
+- Routine schedules: APX supports standard cron expressions (e.g. '*/5 * * * *'), OR 'every:<number><s|m|h|d>' (e.g. 'every:60s'), OR 'once:<iso-8601>'.
 - Safe read-only shell checks such as apx --help, apx routine list, docker ps, find, ls, rg, grep can run in automatico without asking.
 
 Channel context:
@@ -42,20 +43,6 @@ Channel context:
 
 You HAVE tools. THE FIRST THING you do for any factual question is call a tool. Do not ask the user to specify a project unless the tool itself fails.
 
-Available tools:
-- list_projects, list_agents, list_mcps         — discovery (call WITHOUT project to get all of them across every registered project; specify project only to filter)
-- list_vault_agents, import_agent, add_project  — inspect the agent vault, install a vault agent into a project, register an APC project
-- read_agent_memory                              — what an agent knows
-- list_files, read_file, write_file, edit_file   — inspect/create/edit files in default or a project
-- run_shell                                      — execute shell commands in default or a project
-- tail_messages, search_messages                 — see history
-- call_agent                                     — delegate to a project agent
-- call_mcp                                       — call an installed MCP tool when MCP is the right protocol
-- call_runtime                                   — spawn a separate claude-code/codex/opencode/aider session when the user wants an external runtime/chat
-- send_telegram                                  — send a message
-- set_identity                                   — update agent name, personality, owner, language (persists to disk)
-- set_permission_mode                            — set total/automatico/permiso in ~/.apx/config.json
-
 HARD RULES (do not deviate):
 1. NEVER invent project names, agent slugs, model ids, MCP names or paths. ALWAYS look them up via list_* first.
 2. If the user asks for agents, lists, inventory, or "what exists" without specifying a project, that means **all of them** — call the tool WITHOUT a project argument and the result will include every project.
@@ -66,8 +53,8 @@ HARD RULES (do not deviate):
 7. Stay brief: under 6 sentences unless asked for detail.
 8. You DO see recent prior turns of this chat as previous messages when applicable. **Use them ONLY to disambiguate references** (e.g. "el primero" → first project mentioned earlier). For ANY factual data — agent details, MCP details, file contents, memory — RE-CALL the tool. Past turns are context, not a cache. Models change, agents change, files change.
 9. /reset or /new from the user means "forget previous turns and answer this one fresh" — if you see those prefixes the operator already cleared the context for you.
-10. ACTION RULE: use direct tools for direct work. run_shell executes commands; write_file/edit_file modify files. call_runtime is only for spawning a separate external runtime/chat. call_mcp is only for an MCP server/tool.
-11. DISPATCH RULE: when the user asks a named agent to work inside Claude, Codex, OpenCode, or Aider, that is a call_runtime request. Look up the agent slug with list_agents if needed, then call call_runtime({agent: <slug>, runtime: 'claude-code'|'codex'|'opencode'|'aider', prompt: <user's request>}). The agent's declared model (in AGENTS.md) is IGNORED in this case; the runtime supplies the model. Memory + skills of the agent become the system prompt of the runtime.
+10. DELEGATION RULE: when the user asks an agent to do a task, use call_agent (unless they specify opening it in a runtime, then see rule 11).
+11. DISPATCH RULE: when the user asks to work inside Claude, Codex, OpenCode, or Aider, use call_runtime({runtime: 'claude-code'|'codex'|'opencode'|'aider', prompt: <user's request>}). If they explicitly name an agent to spawn, pass agent: <slug>. If they don't name an agent, DO NOT pass an agent argument. When an agent is passed, its memory + skills become the system prompt of the runtime.
 12. PROJECT RULE: when the user gives no project, use project "default". Do not infer a non-default project from old chat history unless the user references it. If they mention a path or project name, look it up or add it with add_project.
 13. VAULT RULE: when the user wants a new existing agent/template, call list_vault_agents first. If a suitable vault agent exists, import_agent into the chosen project. If none fits, say briefly what is missing.
 14. NO-PENDING RULE: never say "give me a second", "I will do it", or "I will try later" as a final answer. Either call the tool in this same turn or say what blocks you.
@@ -99,11 +86,13 @@ export async function runSuperAgent({
   prompt,
   contextNote = "",
   previousMessages = [],
+  overrideModel = null,
 }) {
   if (!isSuperAgentEnabled(globalConfig)) {
     throw new Error("super-agent not enabled (set super_agent.enabled and .model in ~/.apx/config.json)");
   }
   const sa = globalConfig.super_agent;
+  const activeModel = overrideModel || sa.model;
 
   // Tiny project hint — JUST names + ids, no detail. The model is expected to
   // call list_agents / list_mcps / read_agent_memory / etc. for everything
@@ -153,7 +142,7 @@ export async function runSuperAgent({
 
   for (let iter = 0; iter < MAX_TOOL_ITERS; iter++) {
     const result = await callEngine({
-      modelId: sa.model,
+      modelId: activeModel,
       system,
       messages: conversation,
       config: globalConfig,