| name | codex-session-history-recovery |
| description | Locate, identify, and safely summarize prior local Codex conversations from session JSONL when a user asks whether an earlier task, decision, or implementation can be recovered. |
Codex Session History Recovery
Recover evidence from local Codex session logs without confusing injected instructions, tool output, or repeated context with actual user requests.
Use This Skill When
- The user asks whether Codex remembers or can find a previous conversation.
- A repository, URL, command, feature name, or outcome must be traced back to the session where it was discussed.
- A past implementation decision needs to be reconstructed from local conversation records.
Do not use this skill when the current thread already contains the needed exchange.
Workflow
- State the boundary: do not claim personal memory of another thread before locating evidence.
- Search the local session store broadly for distinctive, non-secret anchors such as repository names, URLs, feature names, or command names.
- Treat raw text matches only as candidates. The same anchor can occur in injected
AGENTS.md, tool output, quoted history, or a later summary.
- Parse candidate JSONL and isolate actual
response_item messages whose role is user or assistant.
- Compare the real user requests, timestamps, session ID, and assistant completion messages to identify the relevant session.
- Report the smallest useful result: session identity, approximate time, request, outcome, and a local source link when helpful.
- Warn before exposing a full transcript if it contains credentials, incident details, personal data, or other sensitive material.
Evidence Standard
A filename match is not enough. Confirm at least one genuine user message and one corroborating item such as:
- the assistant's completion message;
- a command result tied to the requested work;
- a commit hash or repository state recorded in the session;
- a later user message acknowledging the result.
Separate confirmed facts from reconstruction. If only a candidate match exists, label it as unconfirmed.
Search Discipline
- Prefer
rg for candidate discovery.
- Prefer structured JSON parsing over reading raw JSONL dumps.
- Search multiple anchors when common terms create noise.
- Do not print entire tool outputs merely to prove a match.
- Do not copy session files into a repository.
- Never include secrets in the response, even when they are expired.
See JSONL search recipes for compact extraction commands and false-positive checks.
Output
Lead with whether the conversation was found. Then provide only the identifiers and facts the user needs. Use a compact table when several exact fields must be compared; otherwise use prose.
If no session is confirmed, say which stores and anchors were checked without presenting guesses as history.
