session-retracelisted
Install: claude install-skill arm-pz/caffeinity-skills
# session-retrace
Each supported agent keeps its conversations as plaintext JSONL on disk, but their session pickers have no content search. This skill finds the session by what was said in it.
| Agent | Default location | Notes |
| --- | --- | --- |
| `qoder` | `~/.qoder/projects/<slug>/<id>.jsonl` | full turns; sibling `state.json` is encrypted — never read it |
| `claude` | `~/.claude/projects/<slug>/<id>.jsonl` | full turns; Claude Code writes the same event schema as Qoder |
| `codex` | `~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl` | full turns, plus `~/.codex/history.jsonl` for prompts whose rollout is gone |
| `gemini` | `~/.gemini/tmp/<alias>/chats/session-*.jsonl` | snapshot-format messages |
Environment overrides: `QODER_CONFIG_DIR`, `CLAUDE_CONFIG_DIR`, `CODEX_HOME`; `--root` points one `--provider` at a non-standard tree. Cursor (SQLite), GitHub Copilot (logs only) and OpenCode (no usable transcripts on this machine) are not covered — say so instead of pretending to search them.
## Workflow
1. Turn the memory into search terms. Drop filler ("the one where", "I think"), keep proper nouns and tool names verbatim, and add one synonym for a vague noun ("login problem" → redirect, auth, 401). If the user gives one generic word, ask once for the project, agent, or rough date — one question, not an interview.
2. Search every agent by default (resolve `<skill>` to the base directory you were given):
```
python <skill>/scripts/session_retrace.py search "spott