Summary

Add _try_codex_jsonl parser for OpenAI Codex CLI session files stored at ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl. This is the 6th normalize format supported by MemPalace, alongside Claude AI JSON, ChatGPT JSON, Claude Code JSONL, Slack JSON, and plain text.

Codex JSONL format

Codex CLI stores session transcripts as JSONL files with one event per line. Each line has:

{"timestamp": "...", "type": "<event_type>", "payload": {...}}

Relevant event types for conversation extraction:

Design decisions

Only event_msg entries are extracted

The parser uses event_msg with user_message / agent_message subtypes as the sole source of conversation turns. These represent the canonical user-authored prompts and assistant replies.

response_item entries are intentionally skipped

Codex rollout files can contain response_item entries with role: "user" that are not real user input — they include auto-injected <environment_context> blocks and other synthetic setup context. The same assistant reply can also appear both as an event_msg/agent_message and as a response_item with role: "assistant", leading to duplicated turns if both are extracted. Skipping response_item entirely avoids both problems.

session_meta header required

The parser only recognizes a file as a Codex rollout if it contains at least one session_meta event. This prevents false-positive matches on Claude Code JSONL or other JSONL formats that happen to contain event_msg-like structures.

Defensive payload handling

payload.message is checked with isinstance(msg, str) before .strip() to avoid AttributeError if the field is null or non-string in an unexpected rollout variant.

Prior art

• PR feat: add OpenAI Codex CLI JSONL normalizer #5 (self-closed by author as "not ready") proposed a Codex normalizer that only handled response_item with a role field. That approach would pull in synthetic environment context as real user input and miss the primary event_msg conversation events entirely.
• Format based on Codex source tests at codex-rs/rollout/src/recorder_tests.rs. This is a conservative interpretation — it extracts conversation messages without attempting full rollout reconstruction.

Changes

1 file changed (mempalace/normalize.py), 53 insertions:

• New _try_codex_jsonl() parser function
• Registered in _try_normalize_json() dispatcher after Claude Code JSONL
• Module docstring updated to list Codex CLI JSONL as a supported format

Test plan

• ruff check mempalace/normalize.py passes clean
• ruff format --check already formatted
• python3 -m py_compile mempalace/normalize.py compiles OK
• Pyright reports 0 new diagnostics
• JSONL event structure based on Codex source tests (recorder_tests.rs)
• session_meta gating verified to prevent false positives on Claude Code JSONL
• response_item exclusion prevents synthetic context pollution and message duplication

Refs: #59