Bug: Codex rollout writer error on every message

[machmikal79]

Bug: Codex rollout writer error on every message

When using the Codex agent in lil-agents, every response produces this error in the popover:

ERROR codex_core::session: failed to record rollout items: thread <uuid> not found

The error does NOT appear when running codex directly in Terminal.app with the same auth and the same Codex CLI version, so it seems specific to how CodexSession.swift invokes the CLI.

Likely cause

CodexSession.send(message:) spawns a fresh codex exec --json --full-auto ... process for every user turn, passing the full prior history as a synthetic prompt. This appears to conflict with recent versions of Codex CLI, where the rollout/session lifecycle now expects a longer-lived thread. Related upstream issues:

• TUI exits to shell on subagent/thread switch when startup skills/list fails; parent rollout stays empty openai/codex#16914 (rollout recorder "channel closed" / "failed to record rollout items")
• VSCode extension: "no rollout found" error on every new conversation (ChatGPT Plus, Windows 10) openai/codex#11997 ("no rollout found for thread")
• Desktop automations on Windows/WSL create automation runs but no rollout materializes, then resume fails with 'no rollout found' openai/codex#16994 (rollout never materializes)

The code comment says // Current Codex CLI: only "codex exec [OPTIONS] <PROMPT>" (resume/--last removed). — but Codex now exposes a JSON-RPC app-server with thread/start and thread/resume, which would be the correct long-lived integration.

Impact

• Visual noise (red error on every turn)
• Token waste: full history is re-sent in every prompt

Suggested fix

Either (a) filter the known-cosmetic stderr lines so they don't surface to the user, or (b) migrate CodexSession to use codex app-server (JSON-RPC over stdio) with a single long-lived process per session.

Happy to test fixes. Thanks for the project!