Read repository-root AGENTS.md completely.
Read handoff/current-status-and-roadmap.md.
Read the assigned tasks/*.md file and every relevant document it names.
Use handoff/prompts/claude-working-style.md when starting a Claude development session.
Inspect unanswered mailbox files before implementation.

Mailboxes

Each message is a new Markdown file. Never append to, rename, or rewrite another agent's message.

handoff/messages/claude-to-codex/questions/  Claude writes questions here
handoff/messages/codex-to-claude/answers/    Codex writes answers here
handoff/messages/codex-to-claude/questions/  Codex writes questions here
handoff/messages/claude-to-codex/answers/    Claude writes answers here

Filename format:

Question: YYYYMMDD-HHMM-TASK-###-short-topic.md
Answer:   <exact-question-stem>-answer.md

Common question template

# Question: short title

- From: Claude | Codex
- To: Codex | Claude
- Task: TASK-###
- Baseline commit:
- Blocking: yes | no

## Context and evidence

## Question

## Proposed acceptance condition

Common answer template

# Answer: short title

- From: Claude | Codex
- To: Codex | Claude
- Responds to: <exact repository-relative question path>
- Task: TASK-###
- Decision authority: task contract | existing ADR | human decision required

## Answer

## Authorized scope, if changed

## Required checks or next action

An answer uses the exact question stem with -answer.md appended and repeats the exact repository-relative question path under Responds to. A correction is a new normally named file with Supersedes: <exact repository-relative path> in its metadata; the old file remains unchanged. One file contains one bounded question or answer. Secrets, tokens, authorization codes, cookies, personal data, and source-file contents from private repositories must never be copied into these files.

Ownership

Question and answer files are immutable after creation. Corrections use a new file referencing the superseded filename.
The active task owner is the only production-code writer.
The reviewer stays read-only and reports findings without editing implementation files.
A mailbox answer clarifies a task; it does not silently expand writable scope or authorize a security-sensitive product decision.
Neither direction may expand writable scope unless the authorized acceptance task is updated first.