Codex resume replays canonical tool names (Bash/Edit) instead of exec_command/apply_patch

[Drix10]

What happened?

Exporting a session to Codex (txcript continue <id> --with codex) writes the canonical tool names (Bash, Edit) into function_call lines. The
Codex runtime has no tools by those names — it expects exec_command (with {"cmd": ...}) and apply_patch (custom_tool_call) — so the replayed
history fails validation and the resumed session breaks. Same for results: patch calls need custom_tool_call_output, not function_call_output.

Expected: from_common emits the native Codex tool shapes, the exact reverse of what to_common normalizes (exec_command/shell → Bash,
apply_patch → Edit/Write).

How to reproduce

let common = Transcript { /* contains Tool::Bash { command: "ls", .. } and Tool::Edit { .. } */ };
let native = codex::Codex::from_common(&common).unwrap();
// native lines contain {"type":"function_call","name":"Bash",...}
// and {"type":"function_call","name":"Edit",...} — both unknown to Codex.

Or via CLI: convert any session containing shell/edit calls to Codex and inspect the response_item lines for "name":"Bash" / "name":"Edit".

Automated repro (fails on main, 3 failed): cargo test -p txcript --test integration codex::from_common — the from_common_denormalizes_* tests in
tests/integration/codex.rs.

Versions

• txcript: 0.14.4
• Agents and their versions: Codex CLI (any)
• Operating system: Windows (applies on all platforms)