How to make Claude Code or Codex escalate a judgment call to your tech lead

An agent can settle any decision its instructions leave open. Here is how to decide which ones it should not make, tell it so, and get the question to the person who owns the answer, with enough context to decide quickly.

1. Decide which calls need a person

Most of what an agent decides is fine to leave with it: names, small refactors, which test to write first. The calls to take away from it are the ones that are expensive to undo or that other people will live with.

• architecture and module boundaries other code will depend on;
• public APIs, database schemas and migrations;
• security, authentication and access control;
• new dependencies, services, or anything that costs money;
• product behaviour a user would notice;
• anything that contradicts a decision already written down.

2. Write the rule where the agent reads it

Claude Code reads a CLAUDE.md file in your repository, and Codex reads AGENTS.md. Put the rule there, with who to ask and what a good question looks like. It is an instruction the agent reads, not a setting that enforces it, so state it plainly. Replace the bracketed parts with your own.

## Judgment calls
Unless the task already settles it, stop and ask before deciding:
- a change to a public API, the database schema or a migration
- security, authentication or access control
- a new dependency, service, or anything that costs money
- an architecture choice other modules will depend on
- product behaviour a user would notice
- anything that contradicts a decision already recorded

Ask [who, e.g. the tech lead] through [where, e.g. the team channel], and give:
- the question, in one sentence
- the background a reviewer needs
- two or more options, each with its pros and cons
- your recommendation, and why
- screenshots, if the choice is visual
Say whether you are blocked. Carry on with any work the answer does not change.

3. Know what an approval gate can and cannot do

Claude Code’s permission settings and hooks, and Codex’s approval and sandbox settings, decide what an agent may do without asking and what it can reach. They are the right tool for risky actions.

They answer a different question, though. The answer is yes or no, from the developer running the agent or a rule set in advance. An approval prompt does not carry a set of options to a tech lead who is somewhere else.

4. Get the question to the person who owns it

Without a tool for it, the developer is the courier: they copy the question into Slack or a ticket, chase the person who can answer, and paste the reply back into the session. It works, but it costs the developer’s time, the options and tradeoffs are easily trimmed on the way, and the reasoning ends up in a chat thread, away from the code it shaped.

This is the step CruxDeck is built for. The agent files a Decision Request through the CruxDeck MCP server, with the question, the options and its recommendation. If your team connects Slack or Microsoft Teams and chooses a channel, it is posted there, where the reviewer can decide from the thread; the assigned reviewer is notified either way. The agent reads the outcome back through MCP and carries on. CruxDeck is not yet available; until it is, the courier route with a good question is the honest alternative.

5. Keep the answer where the next agent will find it

A decision nobody can find tends to get made again, sometimes differently. Record each one with its reason somewhere your agents read: an architecture decision record in the repository, or a decision log your CLAUDE.md or AGENTS.md points to. CruxDeck keeps every resolved request, and agents can list and read past decisions through MCP, within the plan’s history window: on Free, resolved decisions older than 60 days are hidden until you upgrade.

What a good escalation contains

• The question, in one sentence.
• Why the agent should not decide it alone.
• Two or more real options, each with its pros and cons.
• The agent’s recommendation, and its reason.
• Whether the work is blocked while it waits.
• Screenshots, files or links, when the choice is easier to see than to describe.

Why a choice and an approval are different things: decision requests vs. approval gates. What a lead sees: for tech leads, architects and CTOs.