Antigravity Usage and Cost Capture#

What is "Antigravity"?#

"Antigravity" is the internal code name for a Google coding CLI (analogous to Claude Code or Codex). In the decant codebase, Antigravity support is surfaced as Gemini CLI ingestion — the same tool referred to publicly as gemini in decant's source model 1.

Why Transcripts Alone Cannot Produce Exact Cost#

Gemini CLI stores sessions under ~/.gemini/tmp/<project-hash>/chats/session-*.jsonl 2. These JSONL event logs carry no model, token-usage, or cost fields at the file level; token data appears only on individual gemini-type message records, in a per-turn tokens object 3.

From the tokens record, decant extracts:

4

Gaps in the transcript-only picture:

• Cache creation tokens are never recorded by Gemini CLI; decant leaves cacheCreation at zero rather than guessing 5.
• No session-level result record exists equivalent to Claude Code's type: result aggregate. Claude's parser uses a terminal result record for authoritative session totals 6; Gemini has no such record, so decant sums per-turn usage instead 7.
• Cost is not embedded in any Gemini CLI transcript field. Decant computes it post-ingest from first-party published API rates via estimateCost(model, usage, pricing) 8.

Data Sources for Usage and Cost#

1. Per-Message Token Records (Primary — both interactive and headless)#

Gemini messages carry a tokens object on every assistant turn. Decant sums these across turns to produce session totals 7. This is the authoritative source for input/output/cache-read/reasoning tokens for all Gemini runs.

2. Status-Line cost.estimated Sidecar (Interactive Runs)#

Since Antigravity CLI 1.1.21, the CLI's custom status-line state payload exposes a field cost.estimated containing an unrounded cumulative session cost estimate, plus conversation_id, transcript_path, model, and context-window token counters. For interactive sessions, a sanitized per-conversation sidecar persisted from this payload is the recommended way to capture a local-first cost estimate without re-implementing Google's billing math.

Claude Code uses an analogous pattern: decant reads a .meta.json sidecar written beside each session JSONL file via readClaudeSidecarMeta(), then passes the parsed metadata as sidecarMeta to the session parser 9.

3. Terminal Result Events (Headless Runs)#

For headless (non-interactive) Antigravity runs, the CLI emits a terminal JSON or stream-json result event containing cumulative token counts: input_tokens, output_tokens, thinking_tokens, cache_read_tokens, and total_tokens. Cost is not included in this payload — it must still be derived from token counts × published rates. Persisting these terminal result events is the recommended capture path for headless pipelines.

Why Hooks Are Insufficient#

Antigravity hooks expose conversationId, transcriptPath, and modelName — enough to locate a session, but no token usage or cost. They cannot substitute for transcript-derived per-turn token records or a status-line sidecar.

Why the Per-Conversation SQLite Protobuf Metadata Is Not a Safe Contract#

Antigravity stores per-conversation data in a local SQLite database that embeds protobuf-encoded metadata. This metadata appears to contain usage and model information, but Google does not publish the protobuf schema. Relying on it as a primary cost/usage source is fragile: schema changes are unversioned and silent, and decoding requires reverse-engineering private binary formats. Decant's ingest pipeline deliberately avoids this path; all sources are ingested from their plaintext JSONL logs 10.

How Decant Computes Cost from Available Data#

Once token counts are captured, cost is computed deterministically:

1. Model canonicalization: estimateCost(model, usage, pricing) normalizes the model string to a pricing key; unrecognized models return cost 0 rather than guessing 8.
2. First-party rates: Gemini models are priced at published Google API rates, versioned by date in docs/pricing.md. Each flash-lite variant has its own rate; image/TTS/live/transcribe variants remain unpriced because they bill per output unit, not per token 11.
3. Cache token adjustments: Gemini's input token count already includes cached tokens. Decant subtracts cached from input before pricing, then prices cacheRead separately 12.

Key Source Files#