| name | reading-codex-session |
| description | Use when reading Codex .jsonl session transcripts — covers locating files (incl. by UUID or project), filtering rewinds and compactions, focused skimming, and subagent transcripts. |
Every line is {timestamp, type, payload}. Use sed -n Np file | jq to expand a specific line.
Locate the file
Sessions live under ~/.codex/sessions/YYYY/MM/DD/rollout-<ISO-ts>-<uuid>.jsonl. The filename contains the session UUID.
-
By UUID (user pastes one): find ~/.codex/sessions -name "*<uuid>*".
-
By project (forward index): codex maintains a sqlite index at ~/.codex/state_*.sqlite (the N in state_N.sqlite is the schema version — always pick the highest-numbered file). The threads table has cwd, rollout_path, first_user_message, updated_at_ms, archived, etc., with a covering index on (archived, cwd, updated_at_ms):
DB=$(ls -1 ~/.codex/state_*.sqlite | sort -V | tail -1)
sqlite3 -separator $'\t' "$DB" "
SELECT datetime(updated_at_ms/1000,'unixepoch','localtime'),
id, substr(first_user_message,1,60), rollout_path
FROM threads
WHERE cwd = '/abs/path' AND archived = 0
ORDER BY updated_at_ms DESC LIMIT 20;"
first_user_message is denormalized into the row, so you get a label without opening any jsonl. (~/.codex/session_index.jsonl is unrelated — it's an id↔name log for codex resume <name>, no cwd.)
-
By keyword: just grep -l "<keyword>" across all jsonl files.
Fallback when sqlite is absent: scan head -1 of each jsonl and filter on payload.cwd.
Entry types you care about
response_item.message — user/assistant text. payload.content[].text is what was said.response_item.function_call — tool call. payload.{name, arguments, call_id}.response_item.function_call_output — tool result string. Match by call_id.response_item.reasoning — model reasoning. summary often empty, encrypted_content is opaque. Usually skip.event_msg.task_started / task_complete / turn_aborted — turn boundaries.event_msg.thread_rolled_back { num_turns: N } — rewind, see below.compacted { message, replacement_history } and event_msg.context_compacted — compaction, see below.session_meta (line 1) — cwd, id, model_provider, originator (codex-tui = interactive, codex_exec = subagent), thread_source.
Ignore unless asked: event_msg.token_count, turn_context.
Filter rollbacks
/undo (or backspace-rewind) writes an event_msg with payload.type == "thread_rolled_back" and num_turns: N. The N most recent turns before this event are abandoned but stay in the file.
A turn is bounded by task_started and one of task_complete / turn_aborted. To get the live conversation: walk forward, group response_item / event_msg into turns by task_started, and when you see thread_rolled_back num_turns=N, drop the last N grouped turns. Then continue.
Compaction
When the context is summarized, two entries are written close together:
-
compacted (top-level type) — payload.message is the summary text, payload.replacement_history is the array of {type:"message", role, content} items that replace all prior items in the model's view from this point on.
-
event_msg.context_compacted — UI notification only, no payload data.
-
To read what the model sees after a compaction: latest compacted.payload.message + replacement_history is the entire pre-compact context, prior items in the file are no longer visible to the model.
-
To read the full human-visible history: iterate the file in line order; the pre-compact items are still there.
What to read closely vs skim
- Read closely:
response_item.message (user/assistant text).
- Skim:
function_call.arguments and function_call_output.output. Truncate to ~80-char preview with {line, json-path} ref, expand a specific line only when surrounding discussion points to it. Some tool inputs (e.g. patch bodies, file writes) can be huge — skim unless the content matters.
- Treat any unfamiliar
type / payload.type as a tool call — built-in or MCP tools (web search, computer use, etc.) introduce shapes not listed above. Skim by default; expand only if it matters.
- Skip:
reasoning.encrypted_content, token_count, turn_context.
Long sessions drift
Plans, file names, and decisions change mid-conversation. Don't trust early statements as still-true — follow the user/assistant exchange forward, and let later turns override earlier ones. When the user asks "what did we decide about X", scan from the end backward.
Subagent transcripts
Subagents (if any are dispatched) run as separate session files in the same date tree. Their session_meta.originator is codex_exec (vs codex-tui for interactive sessions), and thread_source: subagent marks nested ones.
The dispatching tool call's function_call_output typically returns the subagent's UUID ({"agent_id":"<uuid>", ...}). The subagent's session file is:
~/.codex/sessions/YYYY/MM/DD/rollout-<ts>-<agent_id>.jsonl
Easiest lookup: find ~/.codex/sessions -name "*<agent_id>*.jsonl". Same format and filtering rules; subagents can dispatch further subagents — recurse identically.
