Summary

Add _try_gemini_json parser for Gemini CLI session files stored at ~/.gemini/tmp/{project_hash}/chats/session-{timestamp}-{id}.json. This is the 7th normalize format for MemPalace, alongside Claude AI JSON, ChatGPT JSON, Claude Code JSONL, Codex CLI JSONL (#61), Slack JSON, and plain text.

Gemini CLI session format

Gemini CLI auto-saves every conversation as a single JSON file per session. Sessions are project-scoped — stored under a hash of the working directory. Retention defaults to 30 days / 100 sessions (configurable via settings.json).

Path: ~/.gemini/tmp/{project_hash}/chats/session-{timestamp}-{short_id}.json

Structure:

{
  "sessionId": "xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx",
  "projectHash": "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx...",
  "startTime": "2026-03-30T10:28:04.070Z",
  "lastUpdated": "2026-03-30T10:28:16.793Z",
  "messages": [
    {
      "id": "xxxxxxxx-...",
      "timestamp": "2026-03-30T10:28:04.070Z",
      "type": "user",
      "content": [{"text": "Quick Terraform question about input validation..."}]
    },
    {
      "id": "xxxxxxxx-...",
      "timestamp": "2026-03-30T10:28:16.793Z",
      "type": "gemini",
      "content": "Yes, the validation is **worth adding**..."
    }
  ],
  "kind": "main"
}

Message types

Other message types (model changes, tool calls, etc.) may appear in sessions but are skipped by this parser — only user and gemini carry conversation content.

Design decisions

Custom content extraction instead of shared _extract_content

Gemini user content blocks are {"text": "..."} without a "type" field. The shared _extract_content helper in normalize.py expects {"type": "text", "text": "..."} (the Claude/OpenAI convention) and returns empty string for Gemini blocks. Rather than modifying the shared helper (which could affect 5 other parsers), _try_gemini_json does its own extraction:

• Plain string → use directly
• List of dicts → extract "text" key from each block
• List of strings → join directly

Fingerprints on sessionId + messages keys

The parser requires both sessionId and messages in the top-level dict. This prevents false positives on:

• ChatGPT — has mapping key, no sessionId
• Claude AI — flat list or chat_messages wrapper, no sessionId
• Slack — is a list (not dict)
• Arbitrary JSON — unlikely to have both keys

Single JSON file (not JSONL)

Unlike Codex (JSONL per line) and Claude Code (JSONL per line), Gemini stores the entire session as one JSON object with a messages array. This means the parser registers in the _try_normalize_json dispatcher alongside the other JSON parsers (after json.loads), not in the JSONL section.

What's NOT handled (and why)

• Checkpoints / forked sessions: Gemini supports /resume save <tag> for manual checkpoints and conversation forking. These may create additional session files. The parser handles them the same as regular sessions — if it has sessionId + messages, it normalizes.
• Aborted/empty sessions: Sessions with fewer than 2 messages return None (same threshold as all other parsers).
• Tool call details: Only "user" and "gemini" message types are extracted. Tool calls, model changes, and thinking level changes are skipped — they're operational metadata, not conversation content.
• /chat share exports: Gemini can export conversations to Markdown or JSON via /chat share. The exported JSON format may differ from the auto-saved session format. This parser targets auto-saved sessions only.

Prior art

• PR docs: add Gemini CLI setup guide #106 (merged) added a Gemini CLI setup guide (docs only, no normalizer)
• Issue docs: Support Gemini CLI integration and hooks #107 (open) tracks Gemini CLI integration and hooks
• Format verified against real session files on disk and confirmed via Gemini CLI session management docs and Context7 (google-gemini/gemini-cli library)

Changes

1 file changed (mempalace/normalize.py), 47 insertions:

• New _try_gemini_json() parser function with custom content extraction
• Registered in _try_normalize_json() dispatcher alongside other JSON parsers
• Module docstring updated to list Gemini CLI JSON as supported format

Test plan

• ruff check mempalace/normalize.py passes clean
• ruff format --check already formatted
• python3 -m py_compile mempalace/normalize.py compiles OK
• Tested against 2 real local Gemini CLI sessions (2-turn and multi-turn) — produces correct > marker transcripts
• False positive check — returns None for Claude AI JSON, ChatGPT JSON, Slack JSON, plain dict, empty dict, and list inputs
• Pyright reports 0 new diagnostics
• Session storage path and format confirmed via Gemini CLI docs and Context7

Refs: #59