Repository navigation
What's new
This release is about efficiency: Headroom no longer spikes to gigabytes of memory for OpenCode users, and every provider scan now costs only as much as the data that changed.
- OpenCode: the token scan parsed whole message rows, up to hundreds of MB each, on every database write. It now reads just the head and tail of each row through SQLite blob I/O, and only rows written since the last scan.
- Codex: rollouts are tailed by byte offset instead of being re-read in full on every change.
- Grok, Codex and OpenCode state is kept in the ledger cache, so a relaunch re-reads nothing.
- Cold indexing is twice as fast, with a fixed-format parser for transcript timestamps.
Measured on one Mac against 1.5, two rounds each:
| Scenario | 1.5 | 1.6 |
|---|---|---|
| OpenCode scan on each database change (every 10 s while OpenCode runs) | 6.9 s, 5.1 GB peak | 17 ms, no measurable memory |
| Cold index of 2.6 GB of Claude Code transcripts, CPU | 21-24 s | 9 s |
| Cold index, peak memory | 4.4 GB | 200 MB |
| Warm launch | 4.8 s, 4.4 GB | 1.8 s, 270 MB |
| Codex re-scan after one record appended to a 100 MB rollout | 15 s | 1 ms |
The first launch after upgrading reads the OpenCode database once more to fill the cache (about 2.4 s, 400 MB on the same data), then stays at the 1.6 numbers.
Full changelog: v1.5...v1.6
Universal build for Apple Silicon and Intel, macOS 14 or newer. Signed and notarized by Apple.