Dicklesworthstone/coding_agent_session_search v0.9.0

on GitHub

cass v0.9.0

Sharded native ANN semantic search, a daemon that stays answerable under load, responsive transcript pages, and a round of fixes for large archives.

Fixes reporters were waiting on:

• #498: cass index --full on multi-million-document archives no longer fails at checkpoint_after_fold with "posting validation limit exceeded". Every lexical fold now stays within the per-term posting limit (da06837).
• #495 / #497: cass doctor --rebuild-canonical-fts --yes treats retiring an oversized FTS residue shadow as a completed repair (exit 0, not 13). It no longer rewrites markers on an already-retired archive, and it clears the repair-pending marker after a verified rebuild, so cass status stops reporting a pending repair (89e053f, 71b72ed).
• #496 / #498: the full-rebuild disk check no longer counts a failed rebuild's leftover .rebuild-staging generation twice. It now credits those bytes against the requirement, because the next rebuild resumes into them or clears them first: max(512 MiB, 2*db + 2*lexical - staging) (149ee21, 060807c).

Also fixed:

• Date-filtered search works again on long-lived indexes (#499; frankensearch-quill 0.3.4). Boolean queries follow the documented grammar in every search lane.
• A robot search that finds the lexical index missing or unusable on a large archive starts a detached background cass index instead of retrying an inline rebuild that could not finish (76306c0).
• macOS: a wedged footprint(1) can no longer stall indexing. All OS telemetry probes now have a 3 s deadline (8c9ba0b).
• GitHub Copilot Chat history stored only in VS Code's native chat stores (workspaceStorage/*/chatSessions, legacy state.vscdb) is now detected and indexed (cc16f0a).
• total_matches is exact up to 5M documents (was 50k).

Contract changes (pre-1.0 minor bump):

• When the remaining robot budget cannot fit semantic work, an explicit --mode semantic search now fails with exit 10 (timeout, retryable) instead of returning an empty exit-0 envelope. Lexical hits are never substituted.
• cass archive errors use distinct exit codes: 2 usage, 5 integrity, 7 busy, 14 I/O, 9 other.
• Search follow-ups use --message-index (a hit's line_number). --line addresses only a physical JSONL line. View payloads now name coordinate_space and content_source, and cass introspect describes both fields.
• An index run that cannot read a source keeps what it committed, then exits 9 (retryable) and names the incomplete scan.

The full list, including tests knowingly red at release, is in CHANGELOG.md.

Assets: Linux x86_64 and arm64 (glibc 2.28 floor), macOS arm64, Windows x86_64. Install with install.sh / install.ps1 from this release, or download an archive and check it against SHA256SUMS.