Antigravity Two-File Source Identity#

Status: Pending implementation — Antigravity-specific ingestion does not yet exist in the codebase as of 2026-09-05. The guidance below describes the required design for when it is added.

What Is This#

Antigravity CLI 1.1.22 creates two relevant files per session:

1. Statusline transcript path — reported by the Antigravity CLI, located at ~/.gemini/antigravity/.../transcript.jsonl
2. Actual transcript — the file that contains the real session content, written at ~/.gemini/antigravity-cli/.../transcript_full.jsonl

When decant ingests an Antigravity session, it must treat these two files as a single logical source: resolve the reported path to the actual transcript, and build the ingest fingerprint from both files.

Path Resolution#

The Antigravity CLI's statusline reports a path under ~/.gemini/antigravity/, but the actual transcript is written to the sibling directory ~/.gemini/antigravity-cli/ with the filename transcript_full.jsonl. Any Antigravity source reader must:

1. Receive (or discover) the legacy statusline path.
2. Rewrite it to the antigravity-cli counterpart to obtain the actual transcript content.

This mirrors how geminiSessionLayout() and geminiProjectPath() in 1 derive structured layout information from a raw file path before parsing — path resolution is a prerequisite to reading content.

Why the Fingerprint Must Include Sidecar Content#

Decant's current skip logic 2 decides whether to re-ingest a file by comparing three values stored in ingest_source:

• size (bytes)
• mtime (seconds)
• ingest_revision

This three-factor check is insufficient for Antigravity sessions because:

Same-size, same-timestamp sidecar replacement. If the sidecar (transcript_full.jsonl) is atomically replaced by a new file of identical byte length written in the same filesystem-timestamp second, the skip condition at 2 evaluates to true and the updated session is silently ignored.

The fix: include a hash of the sidecar's content in the fingerprint. The Antigravity ingest path should compute a combined SHA-256 fingerprint — e.g., hash the concatenation or XOR of the statusline transcript content and the transcript_full.jsonl content — so that any byte-level change in either file triggers re-ingest regardless of size and mtime collisions.

This is analogous to how Claude Code uses readClaudeSidecarMeta() 4 to pull in adjacent .meta.json data during session parsing, though for Antigravity the sidecar content must influence the fingerprint, not just the parse.

Implementation Guidance#

When adding Antigravity support:

Related Code#

• Skip / fingerprint logic: 2
• readClaudeSidecarMeta (sidecar pattern precedent): 4
• Gemini layout resolution (path-resolution precedent): 1
• Parser dispatch: 6
• Gemini source parser (structural template): 5