Jared Smith Founder & Engineering Leader
Book a call

Kit · Claude Code and Codex

Your agent says it's done. Are the tests green?

The Senior Engineer's Agent Pack is six pieces for Claude Code and Codex: a reviewer subagent, a test gate that runs your tests before the agent finishes, a format-on-edit hook, and skills for PR descriptions, root-cause debugging and change planning. Your agent sets them up for your test command, your formatter and your default branch.

Get the pack $19+ · pay what you want See the six pieces

$29 suggested. One license covers one person. 7-day refund.

The gaps it is aimed at

A coding agent that finishes with the tests red, leaves files unformatted, writes a PR description that claims checks nobody ran, or reviews a diff without pointing at a line, costs you the time it takes to notice.

This pack does not make your agent write better code. It runs your own checks at two points: when the agent finishes, and after each file it writes.

Your agent sets it up

  1. Unzip the pack, open Claude Code or Codex in the folder, and say: read README.md and set this up.
  2. The agent asks which agents and OS you use, which pieces you want, and for each project your test command, how long it takes, your formatter and your default branch.
  3. It shows you a plan with a dry run of every file it would change, and waits for your yes. Then it installs, backs up anything it replaces, and merges hook entries into your settings without touching your other keys or hooks.
  4. It runs verify.sh, which prints PASS, FAIL or SKIP for each piece. An answer you don't have becomes a TODO, and that piece stays off for that project. Nothing is guessed.

The six pieces

Each one installs on its own, so you can take the ones you want.

reviewer
A subagent that reviews a diff for correctness bugs only. Each finding has file:line, a severity, the input that triggers it and a suggested fix, and the review ends with what it did not check. In Claude Code a hook lets its shell run only eight git subcommands with plain arguments. In Codex it is read-only by instruction.
test-gate
A Stop hook. When the agent tries to finish, it runs your test command from the project folder. While the command fails, the agent is told to keep working and sees the last lines of output.
format-on-edit
A PostToolUse hook that runs your formatter on each file the agent writes. A failed format is reported to the agent. The edit itself is never blocked or undone.
pr-description
A skill that writes the description from the actual diff and commit log, with Summary, Changes, Testing (only what ran) and a Not verified section that is always present.
root-cause
A debugging skill: reproduce the failure, shrink it, test one hypothesis at a time, name the root cause before editing code, then prove the fix with the reproduction.
plan-check
A skill for multi-step changes. It searches for every caller, test, config and doc the change reaches, and writes a map where each cited file is marked as read or marked UNVERIFIED with what would settle it.

How the test gate stays out of a loop

A Stop hook that always blocks would keep an agent working forever. The limits are in the script:

  • At most two blocks in a row per session and project by default. Then the agent may stop, and you are told the tests still fail.
  • No re-run on a tree that already failed in this session, or that already passed.
  • A hung test command's process group is killed at the timeout, 300 seconds by default.
  • The log keeps the first 5 MB of output, so a flood cannot fill the disk.
  • One command turns it off for everything: touch ~/.config/agent-pack/test-gate.off.

What it does not do

It is not CI and not a sandbox. The test gate runs your command as you. The reviewer guard reads the command text before the shell runs it, and an allowed git command still runs whatever the repo's own git config names.

It does not edit your repos or commit. Codex runs new hooks only after you trust them in /hooks.

Who it is for

Engineers who use Claude Code or Codex every day and want the agent to run their tests and formatter, and to show its evidence, without wiring hooks by hand.

If you already have a test gate, a formatter hook and a review agent the way you like them, this pack will mostly duplicate them.

What was tested, and what was not

On Linux with Claude Code 2.1.321, 106 scripted tests pass in throwaway home folders whose paths contain a space. They cover the install and re-run, the settings merges, every test-gate limit, the formatter, the reviewer guard's allowed and blocked commands, and every verify.sh check.

In headless Claude Code sessions the hooks ran for real: the formatter on written files, and the test gate blocking a stop and then passing after the fix. A pre-sale reviewer ran the whole setup headless from the README on the build before the last fixes, and verify.sh printed 15 PASS.

Not tested

  • Codex end to end. The Codex wiring follows OpenAI's docs, and no Codex session ran these hooks or the reviewer.
  • A full setup of the current build, and the interview's wait for a yes.
  • The reviewer's findings on a real diff, and the three skills in a live session.
  • macOS and Windows.

Codex, macOS and Windows are in beta. The pack's SOURCES.md lists every vendor page it relies on, the date it was read, and the behaviors those pages do not confirm.

Built by Jared Smith

13+ years shipping production systems. I was CISO at Lavender, where I built the security program from zero through SOC 2 Type II and shipped production AI features for 100,000+ users. Before that: backend and identity at BlockFi, telemetry ingestion at AAMP Global, and co-founder of PopSocial.

Today I'm a founding engineer at an early-stage startup, where most of the code is typed by coding agents and every line is mine to answer for.

Questions

Who is The Senior Engineer's Agent Pack for?

Engineers who use Claude Code or Codex every day and want the agent to run their tests and formatter, and to show its evidence, without wiring hooks by hand. If you already have a test gate, a formatter hook and a review agent the way you like them, it will mostly duplicate them.

What do I need to run it?

Claude Code or Codex, sh, and git. To merge hook entries into your settings it uses jq or python3; without either, setup gives you the entries to add by hand. Setup makes no network calls.

Can the test gate trap my agent in a loop?

No. By default it blocks at most twice in a row per session and project, then lets the agent stop and tells you the tests still fail. It does not re-run on a tree that already failed or passed, it kills a hung test command at the timeout, and one file turns it off for everything.

Does it work with Codex?

Codex is in beta. The Codex wiring follows OpenAI's docs, but no Codex session ran these hooks or the reviewer end to end. Linux with Claude Code is the tested setup. macOS and Windows are in beta too.

Is the reviewer sandboxed?

No. In Claude Code a hook lets the reviewer's shell run only eight git subcommands with plain arguments. That is a guardrail, not a sandbox: an allowed git command still runs whatever the repo's own git config names. In Codex the reviewer is read-only by instruction.

Do I need the AI Vault?

No. It is a separate pack and does not need the Vault. If you run the Vault's secrets guard, every setup command in this pack passes it.

Can I use it for client work?

Yes. One license covers one person. Use it on your own machines and for client work, and keep the installed files in your own dotfiles. Don't resell or redistribute the pack.

How much does it cost?

Pay what you want: $19 minimum, $29 suggested.

Refunds?

7 days, no questions. Email jared@sublimecoding.com.

The Senior Engineer's AI Vault is a separate kit: a memory vault, handoffs and a secrets guard. The Agent Pack does not need it, and every setup command in this pack passes the Vault's secrets guard.

Get the pack

A zip of Markdown steps, hook scripts, agent templates and skills, delivered via Gumroad. No app, no account, no API keys.

Get the pack $19+ · pay what you want

$29 suggested. 7-day refund, no questions.