[e2720pjk]

Hi ~ Thanks for creating Insights. I use agy myself, so I added support for it.
It’s been working well for me, and I thought I’d share it in case others find it useful.
Happy to have this merged if you think it’s a good fit!

Summary

• Add local Antigravity CLI token-usage collection to Herdr Insights.
• Prefer SQLite generation metadata for model attribution and complete input
  accounting, with compact transcript exports as a fallback.

What

• Add a bounded protobuf decoder and read-only SQLite scanner in
  core/antigravity.mjs, with schema/accounting notes alongside the code.
• Select one authoritative file per conversation, detect DB/WAL changes,
  deduplicate response IDs, and retain accepted history when replacement fails
  or stops before completion.
• Add Antigravity to collector and UI defaults, with conservative migration of
  untouched previous defaults. Update the README's source/privacy lines and the
  rebuild recipe; keep implementation details out of the README.
• Add synthetic regression coverage without shipping personal session data,
  additional dependencies, benchmark tooling or UI redesigns.

Why

Compact transcript exports alone omit generation model IDs and can omit system
input tokens. The conversation database contains generation-level usage and model
metadata. Reading it locally uses the existing collector/store contracts without
requiring authentication, a running agent service or network access.

Testing

Validated commit c4bb8b0 on macOS 26.6.1 (arm64). Package and tooling checks used
Node.js 24.22.2 and npm 12.0.0. No source, lockfile or test changes were needed.

Insights runtime matrix

Ran the underlying test:insights and check:insights commands with each runtime:

Node.js	Tests	Distribution checks
22.15.0	233 passed, 3 skipped, 0 failed	Passed
22.23.4	233 passed, 3 skipped, 0 failed	Passed
24.22.2	233 passed, 3 skipped, 0 failed	Passed

The three skips are conditional alternate-runtime tests: a runtime without zstd
and two installer/uninstaller checks requiring a discoverable older Node.

Package regression checks

The repository's change detector selects all five Pi packages for this PR.
For each package, npm ci --ignore-scripts --no-audit --no-fund, lint,
typecheck, test:unit, test:integration and build completed successfully.
Builds included tarball validation, clean-consumer installation and Pi
registration checks; nothing was published.

Package	Unit tests	Integration tests	Lint / types / build
pi-ask	120 passed	31 passed	Passed
pi-automode	33 passed	22 passed	Passed
pi-slate	531 passed, 1 skipped	90 passed	Passed
pi-subagents	148 passed	25 passed	Passed
pi-web-search	391 passed	45 passed	Passed

Slate's skipped test requires an operator host that is not installed in this
environment.

Compatibility

• No new dependencies, permissions, network calls or authentication flow. SQLite
  uses the existing Node runtime requirement and is opened read-only.
• Default roots are ~/.gemini/antigravity-cli/conversations and
  ~/.gemini/antigravity-cli/brain; both appear under Antigravity in Settings.
• Only exact, untouched three-source or four-source legacy defaults are upgraded.
  Customized/disabled/reordered/empty source lists remain unchanged; those users
  can add the antigravity adapter through the existing configuration editor.
• This does not decode retired Gemini CLI history or Antigravity IDE .pb stores.
  Compact transcript fallback retains unknown metadata when it is absent.

Risk

• The protobuf schema is reverse-engineered, not a stable vendor API. Mapping and
  reference links live in the scanner header. Malformed/oversized records are
  skipped with diagnostics; schema changes can therefore reduce reported usage.
• Thinking is included in output, not added twice. Prefilled system-only records
  without measured usage are excluded. Provider denotes the model publisher,
  not the billing gateway; this does not estimate charges or account limits.
• Step timestamps are used only when their response IDs match. Otherwise daily
  attribution falls back to session creation time, then DB mtime, with a note.
• Missing response IDs fall back to conversation/generation identity with a note.
  SQLite takes precedence over exports; configured source/root order selects
  among copies of the same kind. Divergent copies are not merged.
• Database and transcript refreshes are atomic and roll back on read errors,
  cancellation or deadline. Files exceeding a foreground slice use the existing
  background indexer; this adds no worker or protobuf-row checkpoint.

Checklist

• The change is focused; unrelated changes are excluded.
• Tests cover behavior changes, or the testing section explains why not.
• Relevant documentation is updated, or no documentation changes are needed.

[GaganSD]

@greptile-apps review

[GaganSD]

Thanks for the PR, merged!