Grok CLI¶

JSONL per-project user-prompt audit log. Keys: timestamp (ISO-8601 nanosecond), session_id (UUIDv7), prompt (text), is_bash (bool).

JSONL full session transcripts. type discriminates system/user/assistant/reasoning/tool_result/backend_tool_call. Assistant tool calls live in a tool_calls array on the assistant record; backend_tool_call records host-side calls. reasoning records carry a readable summary array of {type: summary_text, text} blocks plus an opaque encrypted_content blob; agentgrep does not surface them because the adapter reads only content, which reasoning records omit. content is text or a content-blocks array. No record of any type carries a timestamp key, so the adapter backfills the source mtime. Agent-injected turns are written as user records tagged with synthetic_reason (system_reminder, project_instructions); the adapter does not yet read that key, so they are indexed as user prompts.

Full per-session transcript with tool calls; grok.prompt_history is the user-prompts-only audit log.

SQLite with FTS5. Table session_docs: session_id, cwd, updated_at (unix seconds), title (generated), content (full-text index), content_hash, last_indexed_offset. A meta table carries session_search_schema_version (4) and last_bootstrap_at; PRAGMA user_version stays 0.

Pre-indexed session titles and content for fast lookup. De-duplicate against grok.sessions by session_id.

Per-subagent dispatch record. One JSON object per delegated subagent: prompt (the delegated instruction), description, subagent_type, tool_calls, turns, and parent/child session linkage. effective_model_id and child_cwd name the model and working directory the child ran under. The child session keeps its own chat_history.jsonl, read as grok.sessions; this record is the delegation itself. tool_calls is a count, not a list.

Subagent dispatch prompts are conversation content; the child’s own turns are in its session transcript. Parity with claude.projects.subagent and cursor-cli.subagent_transcripts.

Per-session event stream. type values include turn_started, turn_ended, loop_started, first_token, phase_changed, tool_started, tool_completed, permission_requested/resolved, and an mcp_* lifecycle family (illustrative, not exhaustive). A per-record schema_version (“1.0”) rides on turn_started.

Per-session summary: id, cwd, session_summary, created_at, updated_at, num_messages, current_model_id, git metadata, generated_title, agent_name.

Persistent memory Markdown managed by Grok’s memory system. Covers the flat memory/MEMORY.md and the per-project memory/<project-slug>-<hash8>/MEMORY.md subtree. The companion index.sqlite is not separately enumerated; it is a chunk store (chunks) with both an FTS5 mirror (chunks_fts) and a sqlite-vec embedding index (chunks_vec), not an FTS index alone. Grok derives the directory name from the git remote URL, so every clone and worktree of one repository shares a memory directory. Inspectable opt-in.

Structured application logs: ts, src, pid, lvl, msg, ctx. Debugging diagnostics, not chat content.

SQLite database tracking git worktrees created by Grok.

TOML configuration file.

Per-session plan-mode Markdown — the agent’s working plan for the session. Inspectable, parity with claude.plans and cursor-cli.plans; not searched by default.

The rendered system prompt for the session (agent instructions plus injected context). Agent-side boilerplate, largely shared across sessions; documented for inventory, not searched.

Session prompt-context metadata: working_directory, agents_md_files, persona_summaries, os_name, current_date, prompt_mode, audience, version. Configuration, not chat.

Edit-attribution JSONL (filePath, hunkStart/End, linesAdded/Removed, authorType, promptIndex, eventType). Code-change telemetry, no prompt payload; documented, not searched.

ACP-style session/update notification stream (method, params.sessionId, update payloads). Protocol traffic, not chat; documented, not searched.

Per-tool-call terminal stdout/stderr logs (thousands per active project). Tool output, not chat, and high-volume; documented for inventory and deliberately not searched.

skills/<name>/SKILL.md skill-instruction files a user authors. The directory is present but empty at grok 1.0.25: the bundled set lives under bundled/skills/ and is no longer mirrored here. Parity with claude.skills and cursor-cli.skills.

Pin garbage-collection state: an orphans list. Not conversation content.

Holds dismissed_ids. Not conversation content.

SQLite database with members and meta tables, beside its -wal and -shm sidecars. Dashboard state, not conversation content.

Shell scripts and .trace output files for background jobs the agent started. They can quote commands and their output, so they stay unread.

The second-generation memory tree, beside memory/, which Grok Build no longer writes. Each of global/ and workspaces/<workspace>/ holds a generated MEMORY.md index, topics/, archive/, observations/_inbox/<session>__t<turn>-<turn>__n<index>.md notes, and the SQLite files index.sqlite and memory_state.sqlite. Inbox and archived notes open with YAML front matter (schema_version: 2, type, topic_hint, session_id, and others). agentgrep reads every Markdown note as one text record; the SQLite files are not read.

Pager preferences in a [scrollback.scroll] table. Not conversation content.

A signed copy of server-side settings: a payload string and its signature. Not conversation content.