feat(gemini): support Antigravity SQLite conversation database parsing - #1669

Check if this issue is valid — if so, understand the root cause and fix it. At rust/adapters/gemini/src/parser.rs, line 223:

<comment>When a database cannot be opened, queried, or fully iterated, this path reports successful empty or partial usage instead of a read failure. Propagate or log those errors so users do not receive an undiagnosed zero-usage report.</comment>

<file context>
@@ -218,6 +218,158 @@ pub(super) fn parse_jsonl_file(path: &Path) -> Result<Vec<GeminiUsageEvent>> {
 
+pub(super) fn parse_sqlite_file(path: &Path) -> Result<Vec<GeminiUsageEvent>> {
+    let fallback_timestamp = file_modified_timestamp(path);
+    let Ok(connection) =
+        sqlite::Connection::open_with_flags(path, sqlite::OpenFlags::new().with_read_only())
+    else {
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At rust/adapters/gemini/src/loader.rs, line 26:

<comment>Bound SQLite parsing concurrency instead of sending `.db` files through the generic CPU-parallel reader. `parse_sqlite_file` copies each BLOB into a `Vec<u8>`, so several large databases can be resident concurrently and exhaust the CLI’s memory.</comment>

<file context>
@@ -21,10 +21,10 @@ fn load_entries_inner(shared: &SharedArgs, pricing: &PricingMap) -> Result<Vec<L
-            parse_json_file(file)
+        let parsed = match file.extension().and_then(|extension| extension.to_str()) {
+            Some("jsonl") => parse_jsonl_file(file),
+            Some("db") => parse_sqlite_file(file),
+            _ => parse_json_file(file),
         };
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At rust/adapters/gemini/README.md, line 21:

<comment>The rewritten sentence attributes the walking, size-balanced chunking, and ordered parallel reads to both `ccusage-adapter-common` **and** `sqlite` (the singular "which handles..." reads as applying to both). Those capabilities belong to `ccusage-adapter-common`; the `sqlite` crate only opens the read-only DB connection. It also stretches the line to ~128 characters. Rewrite, e.g.: "Reads plain files through `ccusage-adapter-common`, which handles walking, size-balanced chunking, and ordered parallel reads, and reads SQLite databases through `sqlite`."</comment>

<file context>
@@ -16,8 +16,9 @@ Anything that is not specific to this source belongs in `ccusage-core` or
+- `${GEMINI_DATA_DIR:-~/.gemini}/antigravity/conversations/*.db`
 
-Reads plain files through `ccusage-adapter-common`, which handles walking, size-balanced
+Reads plain files and SQLite databases through `ccusage-adapter-common` and `sqlite`, which handles walking, size-balanced
 chunking, and ordered parallel reads.
 
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At rust/adapters/gemini/Cargo.toml, line 16:

<comment>This dev-dependency enables the `fetch-litellm-pricing` feature on ccusage-core, but nothing in the gemini crate calls that feature's fetch/minreq path — the tests use `PricingMap::load_embedded()` against the embedded snapshot, not live pricing fetch. The feature pulls in minreq's rustls stack, which ccusage-core's own Cargo.toml calls the workspace's most expensive build-dependency, so it only adds test-build time. Since ccusage-core is already a normal dependency, drop this dev-dependency entry unless the crate actually uses live pricing fetching.</comment>

<file context>
@@ -10,6 +10,8 @@ ccusage-core.workspace = true
+sqlite.workspace = true
 
 [dev-dependencies]
+ccusage-core = { workspace = true, features = ["fetch-litellm-pricing"] }
 ccusage-test-support.workspace = true
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At rust/adapters/gemini/src/loader.rs, line 95:

<comment>The SQLite fixture claims to encode timestamp `1779000000`, but these bytes decode to `1779161536`; correct the varint or comment so the test documents the value it exercises.</comment>

<file context>
@@ -75,4 +75,58 @@ mod tests {
+            0x28, 20, // 5: varint 20
+            0x48, 10, // 9: varint 10
+        ];
+        // 1.9.4.1 = 1779000000 (0x6A0BB9C0 -> varint: C0 B3 AF D0 06)
+        let ts_inner = vec![0x08, 0xC0, 0xB3, 0xAF, 0xD0, 0x06];
+        let ts_field9 = vec![0x22, ts_inner.len() as u8];
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At rust/adapters/gemini/src/parser.rs, line 331:

<comment>When Antigravity reports `Gemini 3 Pro`, this mapping emits an unpriced model ID. Map `gemini 3 pro` to `gemini-3-pro` so those sessions receive their bundled pricing.</comment>

<file context>
@@ -305,6 +313,45 @@ pub(super) fn parse_sqlite_file(path: &Path) -> Result<Vec<GeminiUsageEvent>> {
+
+    match base {
+        "gemini 3.6 flash" | "gemini 3 flash" => "gemini-3.6-flash".to_string(),
+        "gemini 3.6 pro" | "gemini 3 pro" => "gemini-3.6-pro".to_string(),
+        "gemini 2.5 flash" => "gemini-2.5-flash".to_string(),
+        "gemini 2.5 pro" => "gemini-2.5-pro".to_string(),
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At rust/adapters/gemini/src/loader.rs, line 170:

<comment>This position-based assertion is order-dependent because `parse_sqlite_file` does not order the `gen_metadata` query. With an unordered/reversed scan, the continuation is parsed before the model row and this same-timestamp fixture fails; order rows by `idx` in the parser before testing stateful model inheritance.</comment>

<file context>
@@ -124,20 +124,52 @@ mod tests {
         assert!(entries[0].cost > 0.0);
+
+        // Row 2: Inherited from active session model
+        assert_eq!(entries[1].model.as_deref(), Some("gemini-2.5-flash"));
+        assert_eq!(entries[1].data.message.usage.input_tokens, 400);
+        assert_eq!(entries[1].data.message.usage.output_tokens, 100);
</file context>