agent-management

name	agent-management
description	You need to create a new agent, restart a crashed agent, change an agent's model or config, fix a Telegram bot token, troubleshoot why an agent is not responding, enable or disable an agent, spawn an agent for another user, manage PM2 process management, reset crash limits, or do anything that touches an agent's lifecycle, configuration, or credentials. This is the definitive guide for every agent operation in cortextOS.
triggers	["new agent","create agent","spawn agent","add agent","restart","soft restart","hard restart","disable agent","enable agent","change model","switch model","bot token","BotFather","agent not responding","agent crashed","agent down","crash limit","reset crashes","agent health","list agents","heartbeat","onboard","setup agent","configure agent",".env","config.json","pm2","ecosystem.config","cross-org","agent for someone else","agent management","agent lifecycle","[Truncated]"]

Agent Management

The definitive guide for managing cortextOS agent lifecycle. Every operation, every script, every protocol. Follow these EXACTLY - do not improvise.

---

CRITICAL RULES

1. ALWAYS use the CLI. Never manually edit state files or .env without using the proper command.
2. ALWAYS create .env before enabling. An agent without .env may inherit parent credentials.
3. ALWAYS write restart markers before /exit. Use cortextos bus self-restart, never raw /exit.
4. ALWAYS use cortextos enable to start agents. Never manually edit PM2 config.
5. NEVER share bot tokens between agents. Each agent gets its own bot from @BotFather.
6. NEVER hardcode chat IDs. Get them from the actual user via Telegram getUpdates.
7. ALWAYS ask the user which runtime (claude-code, codex-app-server, or opencode) before scaffolding a new agent. Default to claude-code only if the user has no preference. Never silently pick.

---

1. Creating a New Agent

For Yourself (Same User)

# Option A: CLI (recommended)

# STEP 0 — REQUIRED BEFORE SCAFFOLDING — Ask the user which runtime:
#   "Should this agent run on Claude Code (Anthropic), Codex (OpenAI gpt-5-codex),
#    or OpenCode (provider-agnostic TUI, default model openai/gpt-4.1-nano)?"
# Default to claude-code if the user has no preference. Never silently pick.
# Codex agents MUST use the agent-codex template; OpenCode agents MUST use the
# agent-opencode template. Passing --template agent auto-maps to the runtime's
# native bootstrap (agent-codex for codex, agent-opencode for opencode).
# orchestrator/analyst/m2c1-worker have no codex/opencode variant: codex is
# rejected outright with those templates; opencode is not blocked but would
# scaffold a Claude-oriented bootstrap under an opencode runtime, so avoid it.

# claude-code path (the common one):
cortextos add-agent <name> --template agent --org <org> --runtime claude-code

# codex-app-server path (gpt-5-codex via codex CLI app-server JSONRPC):
cortextos add-agent <name> --template agent-codex --org <org> --runtime codex-app-server

# opencode path (provider-agnostic OpenCode TUI; ships the context-handoff
# lifecycle, default-on at 60% of the context window; model set in config.json,
# default openai/gpt-4.1-nano). --template agent auto-maps to agent-opencode:
cortextos add-agent <name> --template agent-opencode --org <org> --runtime opencode

# Option B: Manual
TEMPLATE="agent"  # or "orchestrator" or "analyst"
AGENT_NAME="myagent"
ORG="myorg"

# Step 1: Copy template
cp -r "$CTX_FRAMEWORK_ROOT/templates/$TEMPLATE" \
      "$CTX_FRAMEWORK_ROOT/orgs/$ORG/agents/$AGENT_NAME"

# Step 2: Create Telegram bot
# Tell the user:
















 >  << 
 600 


node -e 


cortextos   --org 


cortextos status

For Another Person (Cross-User Agent)

AGENT_NAME="theiragent"
ORG="myorg"
THEIR_BOT_TOKEN="<token from THEIR BotFather bot>"
THEIR_CHAT_ID="<THEIR chat_id, NOT yours>"
THEIR_USER_ID="<THEIR user_id>"

# Step 1: Add agent via CLI
cortextos add-agent "$AGENT_NAME" --template agent --org "$ORG"

# Step 2: Write THEIR .env (CRITICAL - must be THEIR credentials)
cat > "$CTX_FRAMEWORK_ROOT/orgs/$ORG/agents/$AGENT_NAME/.env" << EOF
BOT_TOKEN=$THEIR_BOT_TOKEN
CHAT_ID=$THEIR_CHAT_ID
ALLOWED_USER=$THEIR_USER_ID
EOF
chmod 600 "$CTX_FRAMEWORK_ROOT/orgs/$ORG/agents/$AGENT_NAME/.env"

# Step 3: Enable
cortextos enable "$AGENT_NAME" --org "$ORG"

# VERIFY: The new agent messages THEM, not you
# If messages come to you instead of them, the .env has wrong CHAT_ID

Common Mistake: If you skip the .env creation, the agent can inherit credentials from the parent environment. Messages meant for the new agent may route through the wrong identity. ALWAYS create .env BEFORE enabling.

---

2. Restarting Agents

Soft Restart (Preserves Conversation)

# Via bus command (preferred — writes marker file automatically)
cortextos bus self-restart --reason "<reason>"

# Restart a DIFFERENT agent
cortextos bus send-message <agent_name> high "soft-restart" "<reason>"

What it does:

1. Writes .user-restart marker (prevents false crash alert)
2. Sends /exit (Claude exits gracefully)
3. Daemon detects exit, finds marker, categorizes as "user_initiated"
4. Daemon relaunches with --continue (preserves conversation history)

Hard Restart (Fresh Session, Loses History)

cortextos bus hard-restart --reason "context exhaustion"

When to use: Context window full, conversation corrupted, need clean slate, or wiring a new hook (see Hook Reload Lifecycle below). Hard-restart sends an IPC restart-agent signal to the daemon, which kills the running PID and respawns it fresh. The respawn consumes the .force-fresh marker and skips --continue, so settings.json (including newly-wired hooks) is re-read from scratch.

Hook Reload Lifecycle

.claude/settings.json hooks (PreToolUse, PostToolUse, Stop, SessionStart, etc) are loaded once when the Claude CLI process starts. They are NOT reloaded by cortextos bus self-restart — soft restart relaunches Claude with --continue, which reuses the existing process settings cache. Project-level hooks added or changed in settings.json since the original boot DO NOT take effect under soft restart.

When wiring a new hook to a running agent, use a hard restart, not a soft restart:

# From inside the agent itself, or from another agent/terminal:
cortextos bus hard-restart --reason "loading new hook"

# Equivalent ops-side path:
cortextos stop <agent> && cortextos start <agent>

Both kill the existing PID and respawn a fresh process that re-reads settings.json from scratch. Verify by adding a one-line debug write to your hook script before testing — if the file shows up after a benign tool call, the hook is loaded.

If bus hard-restart writes restart markers without actually stopping the process, the running fork may be stale. Fall back to external stop && start until the fork is updated.

Restart from Another Agent

# Soft restart another agent via message bus
cortextos bus send-message assistant high "soft-restart" "goal refresh"

# Check status after restart
cortextos status

---

3. Changing Agent Model

AGENT="sentinel"
ORG="myorg"
NEW_MODEL="claude-sonnet-4-6"  # or "claude-opus-4-6" or "claude-haiku-4-5-20251001"

# Step 1: Update config.json
node -e "
const fs = require('fs');
const path = '$CTX_FRAMEWORK_ROOT/orgs/$ORG/agents/$AGENT/config.json';
const c = JSON.parse(fs.readFileSync(path));
c.model = '$NEW_MODEL';
fs.writeFileSync(path, JSON.stringify(c, null, 2));
"

# Step 2: Soft restart to pick up new model
cortextos bus send-message "$AGENT" high "soft-restart" "model change to $NEW_MODEL"

Available models:

• claude-opus-4-6 - Most capable, highest cost
• claude-sonnet-4-6 - Good balance, ~5x cheaper than Opus
• claude-haiku-4-5-20251001 - Fastest, cheapest, for simple tasks

Context window suffix: Append [1m] to any model ID (e.g., claude-opus-4-6[1m]) to enable the extended 1M token context window. Without it, agents get the default shorter context window and will compact much sooner. Recommended for orchestrators and any agent doing complex multi-step work.

No model set = default (Opus). Always set explicitly for cost control.

---

4. Managing Bot Tokens

Creating a New Bot

Guide the user through BotFather:

1. Open Telegram, message @BotFather
2. Send /newbot
3. Enter display name (e.g., "Assistant - MyOrg Bot")
4. Enter username (must end in bot, e.g., assistant_myorg_bot)
5. Copy the token (format: 1234567890:AAxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx)

Getting Chat ID

After the user messages the bot:

BOT_TOKEN="<the token>"
curl -s "https://api.telegram.org/bot${BOT_TOKEN}/getUpdates" | \
  node -e "const d=JSON.parse(require('fs').readFileSync('/dev/stdin','utf8')); console.log(d.result[0].message.chat.id)"

Note: If getUpdates returns empty, the user needs to send /start to the bot first.

Updating a Bot Token

AGENT="assistant"
ORG="myorg"

# Edit .env (replace BOT_TOKEN line)
sed -i '' "s/^BOT_TOKEN=.*/BOT_TOKEN=<new_token>/" \
  "$CTX_FRAMEWORK_ROOT/orgs/$ORG/agents/$AGENT/.env"

# Restart to pick up new token
cortextos bus send-message "$AGENT" high "soft-restart" "bot token updated"

---

5. Managing .env Files

Required Fields

BOT_TOKEN=<telegram bot token>
CHAT_ID=<telegram chat id for the user>
ALLOWED_USER=<telegram user id for security filtering>

File Permissions

chmod 600 "$CTX_FRAMEWORK_ROOT/orgs/$ORG/agents/$AGENT/.env"

Verifying .env

AGENT_ENV="$CTX_FRAMEWORK_ROOT/orgs/$ORG/agents/$AGENT/.env"
if [[ ! -f "$AGENT_ENV" ]]; then
    echo "ERROR: No .env file for $AGENT!"
elif ! grep -q "BOT_TOKEN=" "$AGENT_ENV"; then
    echo "ERROR: Missing BOT_TOKEN in $AGENT .env"
elif ! grep -q "CHAT_ID=" "$AGENT_ENV"; then
    echo "ERROR: Missing CHAT_ID in $AGENT .env"
elif ! grep -q "ALLOWED_USER=" "$AGENT_ENV"; then
    echo "WARNING: Missing ALLOWED_USER - agent will reject all Telegram messages"
fi

---

6. Managing Crons

Crons are daemon-managed and persisted to ${CTX_ROOT}/state/<agent>/crons.json. The daemon dispatches them automatically — no agent-side restoration needed. Use the bus commands; do NOT edit config.json or use /loop / CronCreate.

Adding a Cron

cortextos bus add-cron <agent> <name> <interval-or-cron-expr> "<prompt>"
# Example: cortextos bus add-cron sentinel new-cron 2h "Do the thing"

Removing a Cron

cortextos bus remove-cron <agent> <name>

Updating a Cron

cortextos bus update-cron <agent> <name> --interval <new>

Listing Crons

cortextos bus list-crons <agent>

---

7. Enabling / Disabling Agents

Enable

cortextos enable <agent> --org <org>

Disable

cortextos disable <agent> --org <org>

This stops the agent's PM2 process and marks the agent as disabled. Config and .env are preserved.

---

8. Health Checks

Check All Agents

cortextos status
cortextos bus read-all-heartbeats

Check Specific Agent Heartbeat

cat "$HOME/.cortextos/default/state/$AGENT/heartbeat.json"

List All Agents

cortextos bus list-agents --format json

Check PM2 Process Status

pm2 list
pm2 jlist | node -e "
const d=JSON.parse(require('fs').readFileSync('/dev/stdin','utf8'));
d.forEach(p => console.log(p.name, p.pm2_env.status));
"

---

9. Crash Recovery

Reset Crash Counter

rm -f "$HOME/.cortextos/default/state/$AGENT/.crash_count_today"

Force Fresh Start (Lose Conversation)

echo "" > "$HOME/.cortextos/default/state/$AGENT/.force-fresh"
cortextos enable "$AGENT" --org "$ORG" --restart

---

10. Troubleshooting

Agent Not Responding to Telegram

1. Check .env exists and has BOT_TOKEN + CHAT_ID + ALLOWED_USER
2. Check fast-checker is running: ps aux | grep fast-checker | grep $AGENT
3. Check agent output: tail -50 $HOME/.cortextos/default/logs/$AGENT/stdout.log and tail -20 $HOME/.cortextos/default/logs/$AGENT/crashes.log
4. Check agent status: cortextos status

Messages Going to Wrong Person

1. Check .env CHAT_ID - is it the right person's chat ID?
2. Check .env BOT_TOKEN - is it the right bot?
3. If agent was spawned by another agent, the parent's env vars may have leaked
4. Fix: rewrite .env with correct credentials, soft restart

Agent Keeps Crashing

1. Check crash count: cat $HOME/.cortextos/default/state/$AGENT/.crash_count_today
2. Check crashes/restarts: tail -20 $HOME/.cortextos/default/logs/$AGENT/crashes.log and tail -20 $HOME/.cortextos/default/logs/$AGENT/stdout.log
3. Common causes: rate limit, auth expired, context exhaustion
4. Fix: reset crash count, fix root cause, cortextos enable <agent> --restart

PM2 Not Restarting Agent

1. Check PM2 status: pm2 list
2. Check PM2 logs: pm2 logs <agent-process-name>
3. Regenerate ecosystem config: cortextos ecosystem then pm2 restart ecosystem.config.js
4. If exit code shows throttling, wait 10s then cortextos enable <agent> --restart

New Hook Not Firing After Wiring It Up

1. Confirm settings.json is valid JSON and the hook block is in the right shape (matcher, type=command, etc).
2. Try a benign tool call that should trigger the hook. If nothing happens, the hook is not loaded.
3. Do NOT soft-restart — --continue does not reload project hooks (see Hook Reload Lifecycle in Section 2).
4. Hard-restart the agent: cortextos bus hard-restart --reason "load new hook". The daemon kills the PID and respawns a fresh process that re-reads settings.json.
5. If the hook still does not fire after a fresh PID, instrument the script with an unconditional write to /tmp/<agent>-hook-test.log at the very top of main(), run a benign tool, and check whether the file appears. If yes, the hook is firing but the script is returning early. If no, the hook is wired wrong in settings.json.

---

Quick Reference

I need to...	Command
Create new agent	cortextos add-agent <name> --template agent --org <org> --runtime claude-code (or --template agent-codex --runtime codex-app-server, or --template agent-opencode --runtime opencode, after asking the user which runtime)
Enable agent	cortextos enable <agent> --org <org>
Disable agent	cortextos disable <agent> --org <org>
Soft restart (self)	cortextos bus self-restart --reason "<reason>"
Hard restart (self)	cortextos bus hard-restart --reason "<reason>"
Restart another agent	cortextos bus send-message <agent> high "soft-restart" "<reason>"
Change model	Edit config.json model field + soft restart
Update bot token	Edit .env BOT_TOKEN + soft restart
Add cron	cortextos bus add-cron <agent> <name> <interval> "<prompt>"
Check health	cortextos status or cortextos bus read-all-heartbeats
List agents	cortextos bus list-agents --format json
Check PM2	pm2 list
Reset crash count	rm ~/.cortextos/default/state/<agent>/.crash_count_today
Force fresh start	Write .force-fresh + cortextos enable --restart