Codex Session Markdown

Search Rules

• Prefer a short but identifying prompt snippet from the user request.
• The script first tries exact normalized substring matching.
• If exact match fails, it falls back to keyword overlap scoring and picks the best candidate.
• Sessions from the current workspace get a preference bonus.
• If multiple sessions match at the same quality level, the newest one wins.
• If prompt matching is too weak, the script falls back to the newest session from the current workspace.
• If no good match exists, the script returns a hint that the prompt keywords may be too long or badly shaped and suggests shortening them.

Output

Default output path pattern:

.agent_cache/codex-session-markdown/<session-id>.md

The Markdown includes:

• session metadata
• ## User blocks
• ## Assistant blocks
• optional tool call and tool result sections

The skill also maintains an index cache at:

.agent_cache/codex-session-markdown/session-index.json

Useful Variants

Include tool activity:

python3 .agents/skills/codex-session-markdown/scripts/export_codex_session.py \
  --prompt "<prompt snippet>" \
  --include-tools

Write to a custom workspace path:

python3 .agents/skills/codex-session-markdown/scripts/export_codex_session.py \
  --prompt "<prompt snippet>" \
  --output runtime/session-review/latest.md

Inspect only the selected session without writing a file:

python3 .agents/skills/codex-session-markdown/scripts/export_codex_session.py \
  --prompt "<prompt snippet>" \
  --dry-run

Disable workspace-cwd fallback if you need strict prompt matching only:

python3 .agents/skills/codex-session-markdown/scripts/export_codex_session.py \
  --prompt "<prompt snippet>" \
  --no-cwd-fallback

Notes

• This skill reads from ~/.codex/sessions, but it only writes inside the current workspace.
• Search explanation is printed on every run so another agent can audit the decision path.
• The exported transcript is intentionally readable, not a byte-for-byte replay of every JSONL record.
• Hidden reasoning is not recoverable in full; the script only exports visible assistant text and optional reasoning summaries when requested.