Codex new agent

Skill sekachev/codex-new-agent-skill/codex-new-agent

Create adaptive Codex-native agents with memory, onboarding, hooks, and dreaming.

Install

npx -y skills add sekachev/codex-new-agent-skill --skill codex-new-agent

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Create a new adaptive Codex-native agent workspace from a reusable template. Use when the user asks to create, scaffold, initialize, bootstrap, or install a personal/project agent with onboarding, durable memory, daily logs, Codex hooks, and dreaming-style memory consolidation.

SKILL.md

3.4 KB, as published. Nobody here has run it

Codex New Agent

Use this skill to scaffold a Codex-native agent that lives in a workspace, adapts to its user, and preserves continuity through files rather than background services.

The generated agent is intentionally not an always-on OpenClaw/Hermes clone. It has no heartbeat, no daemon, no squad/subagent roster, and no ACTIVE.md. Codex wakes it through normal sessions and project hooks.

Create The Agent

Run the bundled script from the skill directory:

python3 scripts/create_agent.py /absolute/path/to/new-agent \
  --agent-name "Agent" \
  --user-name "User"

Add --force only when the user explicitly wants to overwrite an existing incomplete scaffold. The script copies assets/agent-template/, replaces placeholders, and refuses to overwrite existing files by default.

After scaffolding, tell the user to open the generated folder in Codex and start with:

Begin onboarding

Generated Structure

The scaffold creates:

AGENTS.md
SOUL.md
USER.md
MEMORY.md
logs/daily/.gitkeep
.agents/skills/onboarding/SKILL.md
.codex/config.toml
.codex/hooks.json
.codex/hooks/dreaming.py
.codex/hooks/smoke_test.py

Operating Model

Load SOUL.md, USER.md, MEMORY.md, and today's/yesterday's daily logs at boot.
Append a compact entry to logs/daily/YYYY-MM-DD.md after each meaningful turn.
If the user says something durable, important, corrective, or preference-like, write it immediately to MEMORY.md or USER.md; do not wait for dreaming.
Use SOUL.md for durable changes to agent character, helpfulness model, tone, initiative, and adaptation settings.
Use MEMORY.md for long-term lessons, user-specific workflows, environment facts, and dreaming state.
Use daily logs as raw capture only. Do not treat them as future behavior configuration.
Trigger dreaming through Codex hooks or explicit user request. Dreaming consolidates recent logs, prunes noise, and updates MEMORY.md, USER.md, and SOUL.md when there is durable evidence.

Adaptive Usefulness

The generated agent's core mission is to be useful to the specific user. "Useful" is not predefined. The agent must form and revise a working hypothesis:

what outcomes the user values;
how much autonomy they want;
which tone, depth, and pace help them;
which repeated corrections should become durable memory;
which workflows deserve a new skill.

When a reusable workflow deserves a new skill, use Codex's native CreateSkill capability. Do not invent a separate skill-creation mechanism inside the generated agent.

Safety Model

Do not require confirmation for ordinary helpful work. Require explicit confirmation for destructive, irreversible, external, or high-risk actions, including deletion, overwriting critical config, sending messages, publishing, payments, credential changes, and mass edits.

Never store secrets, tokens, passwords, or private credentials in MEMORY.md, USER.md, SOUL.md, or daily logs. Keep secrets in local ignored environment files.

References

For design rationale and source ideas, read references/design-notes.md only when you need to modify the scaffold or explain why it works.