read-codex-history

Run the bundled list_codex_user_inputs.py script with --recent N for a global window or --session-id for specific sessions. It reads history.jsonl directly and returns exact stored inputs grouped by session, preserving duplicates, wording, and timestamps without summarization.

Use analyze_sessions.py search with --codex-only plus optional date, project, and exclusion filters. The scan covers live and archived rollout files with a time budget, and it fails visibly on timeout or unreadable files rather than presenting partial results as complete.

No. The Skill is strictly read-only: it never runs codex resume, never archives, renames, deletes, or repairs anything. To continue work after reading, the evidence is handed to a separate continue-codex-work skill.

The state database rollout_path is only a candidate; the authoritative identity is the session_meta.id inside the rollout file. The reader verifies this internal ID and rejects mismatches, and fork ancestry requires exact history_base byte offsets rather than filename or title matching.

It proves only that some process held Codex's canonical per-thread advisory lock during the snapshot. It does not identify the holder, prove the session is running, and an unmarked row does not prove a session stopped.

Broad scans have a stop-loss time budget and raise an error instead of returning incomplete results. Content absent from both raw records and compaction replacement_history cannot be recovered, and image-only messages cannot be restored to original attachments.