Gemini CLI¶

{"projects": {<absolute_project_root>: <directory_name>}} — the registry mapping each project to the directory Gemini chose for it under both tmp/ and history/. It is the only lookup that covers every naming scheme: gemini_project_hash reproduces the sha256 scheme alone, and most directories are now basename slugs or run stamps. No prompt text, but the reverse index a project-scoped query wants.

Checkpointing shadow-git root, one directory per project holding a .project_root marker and, where checkpointing has run, a full .git/. File snapshots for restore, not conversation: despite the name it is not an archive of expired sessions, which sessionCleanup.ts hard-deletes. Catalogued so the name does not mislead a future adapter into reading it as prompt history.

Git worktrees created by -w/--worktree behind the experimental.worktrees setting, checked out inside the repository on a branch named worktree-<name> — the same convention as claude.worktrees. No index file; a worktree gets its own hashed tmp/ directory with no pointer back, so projects.json is the only mapping. Catalogued so the adapter never indexes a working tree as chat.

JSONL with mixed record types. Line 1 is a SessionMetadataRecord (sessionId, projectHash, startTime, lastUpdated, kind). Subsequent lines are MessageRecord turns (id, timestamp, type, content, optional toolCalls/thoughts/tokens/model/displayContent) interleaved with MetadataUpdateRecord updates ({$set: ...}). content is the searched (expanded) form; the occasional displayContent key is the UI-echo variant. Upstream types also declare RewindRecord and PartialMetadataRecord plus type values info/error/warning. The CLI does write info/error system records, but agentgrep surfaces only the user and gemini conversation turns and skips the system records. Adapter store field uses the underscore-flattened form gemini.tmp_chats.

Parsed by agentgrep via parse_gemini_chat_file (gemini.tmp_chats_jsonl.v1). When a gemini-typed record’s content is empty, the assistant’s prose is drawn from thoughts[*].subject/description and the tool-call context from toolCalls[*].name/description — concatenated into one SearchRecord per turn.

Single-file conversation snapshot written by the /chat save command. JSON object { history: Content[]; authType?: AuthType } where each Content is {role: 'user'|'model', parts: [...]}.

User-named snapshots vs. continuous transcript.

JSON array of LogEntry { sessionId, messageId, timestamp, type, message } — user-prompt audit log. Adapter store field uses the underscore-flattened form gemini.tmp_logs.

Parsed by agentgrep via parse_gemini_logs_file (gemini.tmp_logs_json.v1).

Pre-Feb 2026 single-file session format. JSON object with top-level sessionId, projectHash, startTime, lastUpdated, and a messages array carrying the same per-turn fields as the current JSONL format. Upstream still reads this shape via the isLegacyRecord discriminator. Adapter store field uses the underscore-flattened form gemini.tmp_chats_legacy.

Parsed by agentgrep via parse_gemini_chat_legacy_file (gemini.tmp_chats_legacy_json.v1). Covers sessions whose files predate the JSONL migration; upstream still handles them.

Skill definitions; not chat.

Configuration; not chat.

Global user-authored context/memory Markdown injected into Gemini CLI sessions — the Gemini analogue of Claude’s CLAUDE.md. Standing instructions, not chat; inspectable opt-in rather than searched by default.

Per-tool-call output text (run_shell_command / read_file / update_topic results) under tmp/<hash>/tool-outputs/session-<id>/. Tool output rather than user prompts (may echo file or command content), so inspectable opt-in, not searched by default.