Native Codex rollout JSONL, as written to ~/.codex/sessions/. The whole file
is the transcript string.
The adapter decodes session_meta (cwd, git branch, session id),
turn_context (model), event_msg reasoning, and response_item payloads:
message, reasoning, function_call, custom_tool_call, web_search_call,
tool_search_call, and their outputs. Results link to calls by native
call_id. Text blocks are joined with newlines, and input_image blocks in
messages and tool outputs become [image].
Each response_item reasoning summary part becomes one reasoning record.
Older writers also emit each summary part as an agent_reasoning event; those
events are decoded as before, and summary parts that repeat one are skipped.
Newer writers repeat summaries in item_completed Reasoning events instead,
which are ignored. Encrypted reasoning content is not decoded.
Codex rollout lines carry no per-record id, so identity anchors to the
append-only byte offset (kind byte), which stays stable across chunked
uploads via sourceContext.baseByteOffset. Canonical normalization requires a
resolved session id: include the session_meta row or pass
sourceContext.groupId.
System-injected user content (<environment_context>, <user_instructions>,
<permissions instructions>, <turn_context>) is dropped with an
injected_context_dropped diagnostic.
listTrajectories({ source: "codex" }) walks ~/.codex/sessions/ recursively
for rollout .jsonl files; each item's path is the transcript file.