grok-session-restore — list and load Grok CLI sessions for context restoration.

Session layout (under $GROK_HOME or ~/.grok): sessions///summary.json — index sessions///updates.jsonl — ACP conversation log sessions///chat_history.jsonl sessions///signals.json sessions///plan.json sessions///compaction/ — segment summaries sessions///subagents/ — child session ids active_sessions.json — live pids

Memory is a separate plane under ~/.grok/memory/ and is only path-listed.

Topic selection

summary.json's own generated_title/session_summary wins when non-empty. Otherwise the topic falls back to the first non-synthetic human prompt found in chat_history.jsonl (a bounded head scan, then the tail digest as a last resort) — never left blank when real content exists.

Digest is verbatim, not a summary

load's report never paraphrases the end of a session: the last human prompts, the last assistant texts, and the last tool operations (calls with their key argument, plus outputs and failures) are printed as direct quotes from updates.jsonl/chat_history.jsonl, in chronological order, newest last. summary.json's own last_turn_summary and any compaction segment body are the store's own LLM-written paraphrases — they are shown separately, clearly labeled, and never substituted for the verbatim digest. The compaction body itself is hidden by default; --full-summary opts in.

Reads are byte-budgeted, not full-file scans

Both updates.jsonl and chat_history.jsonl are read via a bounded tail window (DEFAULT_TAIL_BYTES), and the topic fallback uses a bounded head window (HEAD_FALLBACK_BYTES), so even a multi-hundred-megabyte session loads in a fraction of a second.