handofflisted

# Handoff Mirror a live conversation into the other agent's history, so closing one tool and opening the other finds the work already there. ## Why this is possible at all Claude Code and Codex converged on the same two mechanisms, which is the whole reason a bridge is cheap: - **The same hook schema.** `~/.codex/hooks.json` uses Claude Code's shape — `PostToolUse`, `Stop`, `matcher`, `hooks[].command`, even `$CLAUDE_TOOL_ARG_*`. One script registers on both sides. - **The same storage idea.** Both append one JSON object per line as the conversation runs. Neither writes a conversation only at the end, so a hook can read what has accumulated and translate just that. They disagree on every field name, and that is all `scripts/formats.py` does: | Claude Code | Codex | |---|---| | no header; `cwd` on every line | `session_meta`, first line only | | `type:user`, `message.content` | `response_item/message` role=user | | `type:assistant`, `content[text]` | `response_item/message` role=assistant | | `content[thinking]` | `response_item/reasoning` | | `content[tool_use]` | `response_item/function_call` | | `type:user`, `content[tool_result]` | `response_item/function_call_output` | | `parentUuid` linked list | file order, `turn_id` per item | ## Two things the existing converters get wrong Worth knowing before reading their source, because both mistakes look like success. **A mirror must be its own conversation.** Appending into a history the other tool has open races that to