@agentprojectcontext/apx 1.34.0 → 1.35.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agentprojectcontext/apx",
-  "version": "1.34.0",
+  "version": "1.35.0",
   "description": "APX — unified CLI + daemon for the Agent Project Context (APC) standard.",
   "publishConfig": {
     "access": "public"

package/skills/apx/SKILL.md

@@ -3,7 +3,7 @@ name: apx
 description: >-
   APX CLI — local daemon that orchestrates agents, sessions, MCPs, and channels across CLIs.
   Use `apx exec "prompt"` for the local super-agent, or `apx run <agent> --runtime <claude-code|codex|opencode|aider|cursor-agent|gemini-cli|qwen-code> "prompt"` to hand the task to another CLI.
-  Activate on: 'apx', 'apx exec', 'apx run', 'apx daemon', 'pedile a codex/claude/opencode/gemini', 'que codex haga …', 'delegar a otro engine', 'ejecutar en codex/claude', 'run this in <runtime>'.
+  Activate on: 'apx', 'apx exec', 'apx run', 'apx daemon', 'ask codex/claude/opencode/gemini to …', 'have <runtime> do …', 'delegate to another runtime', 'run this in <runtime>', 'route this task to <runtime>'.
 homepage: https://github.com/agentprojectcontext/apx
 ---
 

package/src/core/agent/build-agent-system.js

@@ -1,16 +1,26 @@
+// System prompt for project agents (Cody, Sofía, etc.). Shares the agent-base
+// + action discipline with the super-agent, layered with a project-agent role
+// delta plus the agent's own profile fields.
+//
+// When a project agent answers through a real user channel (Telegram, web,
+// desktop), pass `channel` / `channelMeta` / `sender` so the channel-context,
+// relationship and voice-mode / segmenting blocks come along — without that
+// the agent has no idea HOW the user is talking to it.
 import fs from "node:fs";
-import path from "node:path";
-import { fileURLToPath } from "node:url";
 import { readAgentMemory } from "./memory.js";
 import { apcProjectFile, apcSkillFile } from "../apc/paths.js";
-
-// Anti-ghost-response rules injected into every agent system prompt. The text
-// lives next to the agent prompts (src/core/agent/prompts/action-discipline.md)
-// so it can be edited without touching code. Cached at module load.
-const __dirname = path.dirname(fileURLToPath(import.meta.url));
-const ACTION_DISCIPLINE_RULES = fs
-  .readFileSync(path.join(__dirname, "prompts", "action-discipline.md"), "utf8")
-  .trimEnd();
+import {
+  PROMPTS,
+  buildChannelContextBlock,
+  buildVoiceModeBlock,
+  buildRelationshipBlock,
+  buildUserContextBlock,
+  buildSegmentDiscipline,
+} from "./prompt-builder.js";
+
+// Cap the injected agent body so an over-long authored file can't blow the
+// token budget. Mirrors PROJECT_AGENTS_MAX_CHARS for AGENTS.md.
+const AGENT_BODY_MAX_CHARS = 6000;
 
 function listField(value) {
   if (Array.isArray(value)) return value.map(String).map((s) => s.trim()).filter(Boolean);
@@ -21,9 +31,9 @@ function projectName(project) {
   if (project?.name) return project.name;
   try {
     const meta = JSON.parse(fs.readFileSync(apcProjectFile(project.path), "utf8"));
-    return meta.name || path.basename(project.path);
+    return meta.name || project.path?.split("/").pop() || "";
   } catch {
-    return path.basename(project?.path || "");
+    return project?.path?.split("/").pop() || "";
   }
 }
 
@@ -31,80 +41,146 @@ export function agentSkills(agent) {
   return listField(agent?.fields?.Skills);
 }
 
+/**
+ * Build the system prompt for a project agent.
+ *
+ * @param project        { id, name, path, ... }
+ * @param agent          { slug, fields: { Description, Role, Language, Skills, Tools, ... }, body }
+ * @param opts
+ * @param opts.invocation  "engine" (direct LLM call), "telegram", "routine", etc.
+ * @param opts.runtime     external runtime name when relevant ("claude-code", …)
+ * @param opts.channel     surface the user is on (channels/<name>.md is layered in)
+ * @param opts.channelMeta meta for the channel template + `{voice: true}` flag
+ * @param opts.sender      resolved sender for the relationship block
+ * @param opts.caller      who invoked us (another agent slug, "user", "routine", …)
+ * @param opts.routine     routine name when invocation === "routine"
+ * @param opts.globalConfig used for user.language / user.locale / user.timezone
+ * @param opts.extraParts  additional blocks to append before discipline
+ */
 export function buildAgentSystem(project, agent, {
   invocation = "engine",
   runtime = null,
   channel = null,
+  channelMeta = {},
+  sender = null,
   caller = null,
   routine = null,
+  globalConfig = {},
   extraParts = [],
 } = {}) {
   const fields = agent.fields || {};
-  const parts = [
-    `You are APC agent "${agent.slug}".`,
-    `Project: ${projectName(project)} (${project.path}).`,
-  ];
+  const channelLow = String(channel || "").toLowerCase();
+  const voice = !!channelMeta?.voice || channelLow === "voice";
 
-  if (fields.Description) parts.push(String(fields.Description));
-  if (fields.Role) parts.push(`Role: ${fields.Role}`);
-  if (fields.Language) parts.push(`Default language: ${fields.Language}`);
+  // Shared base + project-agent role delta (the "I'm scoped to one project" framing).
+  const roleBlock = [PROMPTS.AGENT_BASE, PROMPTS.PROJECT_AGENT_ROLE].join("\n\n");
 
-  const declaredTools = listField(fields.Tools);
-  if (declaredTools.length) {
-    parts.push(
-      [
-        "## Declared Tool Hints",
-        declaredTools.join(", "),
-        "These are agent-level tool expectations, not a guarantee. Actual callable tools depend on invocation surface.",
-      ].join("\n")
-    );
+  // Agent profile (its display name + description + Role + Language + owner from config).
+  const profileLines = [
+    `# Agent profile`,
+    `You are **${agent.slug}**, a project agent dedicated to **${projectName(project)}** (\`${project.path}\`).`,
+  ];
+  if (fields.Description) profileLines.push(fields.Description);
+  if (fields.Role) profileLines.push(`Role: ${fields.Role}`);
+  if (fields.Language) profileLines.push(`Default language: ${fields.Language}`);
+
+  // The agent's authored body (everything after the frontmatter in its
+  // `.apc/agents/<slug>.md`) is its real instruction set — persona, domain
+  // rules, API endpoints, tone, hard limits. Without injecting it the agent
+  // runs on its fields alone and loses everything its author actually wrote.
+  let customBody = String(agent.body || "").trim();
+  if (customBody.length > AGENT_BODY_MAX_CHARS) {
+    customBody = customBody.slice(0, AGENT_BODY_MAX_CHARS) + "\n\n…(instructions truncated)";
   }
+  const customInstructions = customBody ? `# Custom instructions\n${customBody}` : "";
 
-  parts.push(buildInvocationContext({ invocation, runtime, channel, caller, routine }));
-
-  const memory = readAgentMemory(project, agent.slug);
-  if (memory) parts.push("## Memory\n" + memory);
+  // User context (owner name, language, timezone) — same block the super-agent
+  // gets, so project agents know how to address the user.
+  const userContext = buildUserContextBlock(null, globalConfig, { agentName: agent.slug });
 
-  const apxSkill = apcSkillFile(project.path, "apx");
-  if (fs.existsSync(apxSkill)) parts.push("## APX\n" + fs.readFileSync(apxSkill, "utf8"));
+  // Channel context — the same channels/*.md the super-agent uses. Project
+  // agents talk through the same surfaces; they need the same formatting rules.
+  const channelBlock = buildChannelContextBlock(channel, channelMeta);
+  const voiceBlock = buildVoiceModeBlock(voice);
+  const segmentDiscipline = buildSegmentDiscipline({ channel: channelLow, voice });
 
-  for (const skill of agentSkills(agent)) {
-    const skillPath = apcSkillFile(project.path, skill);
-    if (fs.existsSync(skillPath)) parts.push(`## Skill: ${skill}\n` + fs.readFileSync(skillPath, "utf8"));
-  }
+  // Relationship block — "you're talking to <owner>" / "<contact>" / "<guest>".
+  const relationship = buildRelationshipBlock(sender);
 
-  for (const ep of extraParts) {
-    if (ep) parts.push(ep);
-  }
+  // Declared tool hints (informational — actual callables come from runtime).
+  const declaredTools = listField(fields.Tools);
+  const toolHints = declaredTools.length
+    ? [
+        "## Declared tool hints (agent-level expectations)",
+        declaredTools.join(", "),
+        "Actual callable tools depend on the invocation surface — use whatever the runtime sends this turn.",
+      ].join("\n")
+    : "";
 
-  // Always append action discipline rules last so they are close to the end
-  // of the system prompt and harder for the model to "forget".
-  parts.push(ACTION_DISCIPLINE_RULES);
+  // Invocation context (who called me, through what, for what).
+  const invocationCtx = buildInvocationContext({ invocation, runtime, caller, routine });
 
-  return parts.join("\n\n");
+  // Per-agent memory (lives under <project>/.apc/agents/<slug>/memory.md).
+  const memory = readAgentMemory(project, agent.slug);
+  const memoryBlock = memory ? "# Memory\n" + memory : "";
+
+  // Project's APX skill + agent's declared skills (loaded as full bodies — they're
+  // small and specific to this agent).
+  const projectSkills = buildProjectSkills(project, agent);
+
+  return [
+    roleBlock,
+    profileLines.join("\n"),
+    customInstructions,
+    userContext,
+    memoryBlock,
+    relationship,
+    channelBlock,
+    toolHints,
+    invocationCtx,
+    projectSkills,
+    ...extraParts.filter(Boolean),
+    voiceBlock,
+    PROMPTS.ACTION_DISCIPLINE,
+    segmentDiscipline,
+  ]
+    .filter(Boolean)
+    .join("\n\n");
 }
 
-function buildInvocationContext({ invocation, runtime, channel, caller, routine }) {
-  const lines = [
-    "## Invocation Context",
-    `invocation: ${invocation}`,
-  ];
+function buildInvocationContext({ invocation, runtime, caller, routine }) {
+  const lines = [`## Invocation`, `invocation: ${invocation}`];
   if (runtime) lines.push(`runtime: ${runtime}`);
-  if (channel) lines.push(`channel: ${channel}`);
   if (caller) lines.push(`caller: ${caller}`);
   if (routine) lines.push(`routine: ${routine}`);
 
   if (runtime) {
     lines.push(
-      "You are running inside the named external runtime. Use only tools and permissions that runtime actually exposes."
+      "You're running inside the named external runtime. Use only tools and permissions that runtime exposes."
     );
-  } else if (invocation === "engine") {
-    lines.push("You are a direct LLM call through APX. Do not claim shell, file, MCP, or Telegram tools unless APX explicitly provided them.");
-  } else if (invocation === "telegram") {
-    lines.push("You are replying through Telegram. Keep responses brief, plain text, and matched to the user's language.");
   } else if (invocation === "routine") {
-    lines.push("You were invoked by an APX routine. Complete the requested work now; do not say you will do it later.");
+    lines.push(
+      "You were invoked by an APX routine. Complete the requested work now; don't say you will do it later."
+    );
+  } else if (invocation === "engine") {
+    lines.push(
+      "You're a direct LLM call through APX. Don't claim shell, file, MCP, or Telegram tools unless the runtime explicitly sent them this turn."
+    );
   }
-
   return lines.join("\n");
 }
+
+function buildProjectSkills(project, agent) {
+  const parts = [];
+  const apxSkill = apcSkillFile(project.path, "apx");
+  if (fs.existsSync(apxSkill)) {
+    parts.push("## APX\n" + fs.readFileSync(apxSkill, "utf8").trim());
+  }
+  for (const skill of agentSkills(agent)) {
+    const skillPath = apcSkillFile(project.path, skill);
+    if (fs.existsSync(skillPath)) {
+      parts.push(`## Skill: ${skill}\n` + fs.readFileSync(skillPath, "utf8").trim());
+    }
+  }
+  return parts.join("\n\n");
+}

package/src/core/agent/channels/voice-context.js

@@ -86,13 +86,13 @@ export function buildVoiceChannelContext(channel, { projectId, language = "es" }
   switch (channel) {
     case "voice":
       return { ...base, contextNote: dynamicNote, systemSuffix: SUGGESTIONS_INSTRUCTION, wantsSuggestions: true, channel: CHANNELS.DECK, channelMeta: { voice: true } };
-    case "deck":
+    case CHANNELS.DECK:
       return { ...base, contextNote: dynamicNote, systemSuffix: SUGGESTIONS_INSTRUCTION, wantsSuggestions: true, channel: CHANNELS.DECK, channelMeta: {} };
-    case "desktop":
+    case CHANNELS.DESKTOP:
       return { ...base, contextNote: dynamicNote, systemSuffix: SUGGESTIONS_INSTRUCTION, wantsSuggestions: true, channel: CHANNELS.DESKTOP, channelMeta: { voice: true } };
-    case "telegram":
+    case CHANNELS.TELEGRAM:
       return { ...base, contextNote: dynamicNote, channel: CHANNELS.TELEGRAM, channelMeta: {} };
     default:
-      return { ...base, contextNote: dynamicNote, channel: channel || "api", channelMeta: {} };
+      return { ...base, contextNote: dynamicNote, channel: channel || CHANNELS.API, channelMeta: {} };
   }
 }