Groundhog

Relive the user's agent sessions to spot the patterns they repeat every day. The main agent orchestrates a sweep over all local Claude Code and Codex traces, fans out reader subagents to mine them for recurring workflows and friction, and synthesizes a numbered report of improvement opportunities. It then stops: no file edits, no installs, no config changes until the user replies with the numbers they want implemented.

Data Sources

• Claude Code: ~/.claude/projects/<flattened-cwd>/<session-uuid>.jsonl — one file per session. User messages carry promptSource, entrypoint, isSidechain, permissionMode, cwd, gitBranch, and timestamps.
• Codex: ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl — first line is a session_meta payload with originator, source, cwd, and model info. ~/.codex/history.jsonl holds raw prompt history.

Verify what actually exists before planning the sweep — trace formats and locations drift across CLI versions, and other agent CLIs (Gemini, opencode, …) may also have local traces worth including. A quick head of a few files beats assuming the layout above is still exact.

Inputs (all optional)

• Analysis budget. Default: up to 500 user messages (human-typed prompts, with their surrounding session context), walking sessions newest-first, going back at most 60 days. Whichever limit hits first ends the scope. The user may raise, lower, or drop either bound ("all history", "last 2 weeks", "2000 messages").
• Project filter. Default: all projects.
• Tool filter. Default: every agent CLI with local traces found.

Recent sessions matter most — they reflect current habits, current projects, and friction that is still live. Older sessions are only worth reading if the budget allows after the recent ones are covered.