Repository navigation
Cache OpenCode Go scans and discover browser sessions in Auto - #888
Conversation
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📝 Walkthrough
Merge Risk: ⚪ Minimal · up to
|
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📝 Walkthrough
Merge Risk: ⚪ Minimal · up to
|
Summary
Ports upstream CodexBar v0.75.0 OpenCode Go changes (steipete#4410, steipete#4445; commits 8770f85, 976d8a0, 79b212d).
opencode.dbpath plus size and modification time of the database and its-walfile (upstream's key). The fingerprint is checked before and after each scan, so a scan that overlaps a commit is not cached. One entry, at most 4 MiB. Errors and a missing database are never cached. The provider fetch and the cost-history scans share the cache.local+web, authoritative). If the session is rejected, the local estimate is kept. Other web errors keep the estimate and still try the Zen balance. With an API result, discovery is only used for the Zen balance (upstream). OpenCode Go now owns browser cookie resolution, so the shell no longer reads browsers every refresh.pace_authoritativeto true.Skipped: macOS cookie cache/Keychain/Safe Storage gates (no Windows equivalent); the upstream SQLite metadata refactor (internal only).
Tests
cargo test --manifest-path rust/Cargo.toml: 4625 passed, 0 failed (after merging main)cargo test --manifest-path apps/desktop-tauri/src-tauri/Cargo.toml: 729 passed-D warningson both manifests: cleanNo UI change, so no CUA proof. Live browser discovery is covered only through injectable test seams.
Summary by CodeRabbit