Atlas Scout Preview 22 · Official release · July 30, 2026
Smaller on disk. Lighter in memory. Faster to ready.
Preview 22 changes how Scout represents millions of code relationships and how expanded indexing work moves through the writer. On Firefox and Linux, the result is a materially smaller index with less memory pressure and less disk writing—without giving up correctness, warm-query performance, or complete-readiness time.
Firefox and Linux · same frozen readiness boundary
The database stopped repeating the expensive part
Large C and C++ repositories can produce millions of relationships. Preview 21 stored a long stable source-symbol identifier with each one and repeated that value again in the relationship key. Preview 22 keeps the public stable identity agents rely on, but stores the internal source reference as a compact SQLite row ID.
Swipe the chart and exact-value table sideways to inspect every value.
More repository. Less machine.
- p50 of 3 fresh runs
- empty isolated product cache
- same frozen corpora
- same asserted readiness boundary
- lower is better
| Corpus | Metric | Preview 21 p50 | Preview 22 p50 | Change |
|---|---|---|---|---|
| Linux | Final cache | 7.016 GiB | 4.891 GiB | −30.28% |
| Firefox | Final cache | 7.724 GiB | 5.744 GiB | −25.64% |
| Linux | Peak memory | 11.563 GiB | 8.701 GiB | −24.75% |
| Firefox | Peak memory | 12.309 GiB | 9.619 GiB | −21.85% |
| Linux | Physical writes | 20.882 GiB | 15.073 GiB | −27.82% |
| Firefox | Physical writes | 23.214 GiB | 17.459 GiB | −24.79% |
Complete readiness improved too
Two focused changes · one disposable local index
Scout keeps the useful structures and removes avoidable duplication
Compact relationship sources
Stable outside. Smaller inside.
Tools, agents, and source coordinates continue to use the same stable symbol IDs. Inside schema 6, relationship sources use the owning symbol’s compact database row ID. Incoming and outgoing navigation keeps the indexes required for low-latency traversal.
No queued expanded batch
Overlap remains. The memory pile-up does not.
Extraction still overlaps with the serial SQLite writer, but a fully expanded batch can no longer wait in memory while another batch is being produced. This matters when a bounded slice of C or C++ source expands into millions of observations.
Same map · same query result · less storage
The reduction did not come from deleting useful evidence
Linux produced identical content before and after the optimization. The relationship target, target-name, and source-file indexes remain; the source representation changed, not the facts Scout can navigate.
Warm-query guard
The exact 230-request formal Ollama MCP stream was replayed five times. Aggregate p50 remained 0.49 seconds for both Preview 21 and Preview 22 at the measured resolution.
Database schema 6 · atomic migration
Upgrading does not hold readiness hostage to a giant rewrite
Existing schema-5 indexes migrate atomically, preserve symbol row identities, recreate required indexes, and roll back cleanly if migration cannot complete. Scout does not force a synchronous VACUUM during upgrade.
- A fresh Preview 22 index receives the smaller physical representation immediately.
- An upgraded index keeps freed pages available for reuse instead of blocking startup on a full database rewrite.
- API-visible symbol identities, source paths, target identities, and navigation behavior stay stable.
Modern protocol metadata · legacy initialization preserved
MCP 2026-07-28, without abandoning existing hosts
Preview 22 adopts MCP 2026-07-28 with stateless discovery, per-request protocol-version metadata, modern result and cache metadata, and explicit errors for unsupported protocol versions. Existing initialization and response shapes remain available for clients that still depend on them.
The faster path uses the machine differently
The honest cost: somewhat more CPU utilization
Average index CPU rose from 167.319% to 200.445% on Linux and from 199.186% to 211.108% on Firefox, where one logical core equals 100%. Peak task count stayed at 40. The candidate used available CPU more effectively while wall time, memory, writes, and final cache all fell.
One Firefox observation peaked at 12.985 GiB—above the old 12.309 GiB median. The other two Preview 22 runs measured 9.608 and 9.619 GiB, making 9.619 GiB the retained p50. No observation was excluded.
These are formal measurements on one documented workstation, two frozen repositories, and three repetitions per corpus with uncontrolled operating-system page cache. They support the scoped changes above, not one universal resource guarantee.
Locked inputs · immutable baseline · deterministic analysis
The result is bound to the exact candidate that was measured
- Campaign lock SHA-256
c0ac9533de9be4c853999d82cf16ae6ae99ea72a06ee818c93ec0cdab82e4530- Candidate executable SHA-256
05379b9551c743617c0730acf370526855539dc1069849014ba65cd05fee745e- Candidate production diff SHA-256
46270c01c688d0d30025ee6d0b5024def100032f71cf8115830f2aa31b862bba- Formal analysis SHA-256
6b14aabc2e8ea3423fb8a6ecce424a2bcb3b214ccbb29f9008957b860c7f4742
Preview 22 is the current public preview
Upgrade the complete managed installation
The native updater moves the runtime pair, integrations, documentation, and legal payload together after verifying the release.
atlas-scout upgrade --check
atlas-scout upgrade