What Scout skips is now shown. What Scout says is now checked.

.atlasignore · .gitignore outside Git · ignored_paths

What Scout skips is no longer a secret.

An exclusion you cannot see is indistinguishable from a bug. When a directory is missing from results, the agent - or the person supervising one - cannot tell whether the index is broken, the language is unsupported, or an ignore rule somewhere decided the files should not exist. Preview 27 resolves that ambiguity in both directions: more control over what is excluded, and full visibility into what was.

Verified claims · real search patterns · guidance where it is read

The hooks stopped promising what they had not checked.

The Claude Code tool notice used to open with “Atlas Scout has this workspace indexed” - unconditionally. In an unindexed workspace that was a promise the first tool call could not keep, and a reader taught to distrust one notice distrusts them all. Both the tool notice and the new prompt reminder now verify that an index database actually exists at or above the hook’s working directory, and stay silent when it does not.

The notice also aimed at the wrong target. It matched only Claude’s Grep tool, but sessions search through Bash running rg or grep at least as often - and newer harnesses may not expose Grep at all. The hook now parses shell commands too, finds the first rg/grep-family invocation across command separators, and applies the same identifier test to its pattern. The failure mode is chosen deliberately: a mistaken parse costs a missed notice, never a wrong one, because the identifier gate still has the final word. The per-session notice budget rises from two to four to match the wider reach.

And the navigation policy moved to where it is actually read. The session-start description sits at the top of a conversation that only grows away from it, and models weight recent context far more heavily. A new UserPromptSubmit hook adds a short reminder of the structural tools on the first prompt of a session and every fifth prompt after - only in workspaces that hold an Atlas Scout index, and never twice for a turn that just paid for one.

The hooks remain advisory only: they inject context, never make a permission decision, and never deny, retry, or modify a tool call. Prose searches and regular expressions are still ignored.

One command · three managed surfaces · every upgrade

The integration now follows the runtime.

Claude Code copies a plugin into its install cache. That is the right design for Claude Code - and it means releasing a new Atlas Scout runtime silently left the old hooks, skill, and MCP declaration in place, speaking for a version that no longer existed. Cached guidance drifting behind the runtime is the same failure as an unverified index claim: the integration saying something the product cannot back.

atlas-scout claude-code sync

The command aligns three things independently and reports each: the managed Atlas Scout section of the user-global ~/.claude/CLAUDE.md - created for new users, appended to an existing memory, rewritten in place when a release changes the canonical text, and never touching content outside its begin/end markers - plus the managed marketplace registration and the installed plugin, registered, installed, enabled, or updated as the inspected state requires. A memory file with a begin marker but no end marker is reported and left untouched rather than guessed at.

You will rarely need to type it: atlas-scout upgradenow runs the upgraded binary’s sync after promotion, so each release carries its own texts forward, and doctor recommendations point at the command instead of describing manual steps.

The code-navigation skill also learned the phrasings users actually type. Its description listed capability nouns; it now also names the questions - where a symbol is defined, who calls or uses a function, what implements an interface, what breaks if code changes - so Claude Code’s description matching can find it from a natural request.

Fresh on arrival · space returned · never blocking

Maintenance that never asks for your time.

Signed release · four native targets · public verification

The release boundary, verified as always.

Signed tag

v1.0.0-preview.27

Release commit

2e7a23ef4739fc7e93b40e64a7b8bf353a4a8f6f

Release manifest SHA-256

2cb96953f35d265d2b51cc0a7b7ef6fbe2067c9c5f93087092619f0965191d8f

Release tests

487 passed per Linux and macOS lane · 456 on Windows · intentional ignores only · 0 failed on every target

Public immutable objects

24 objects · exact size and SHA-256 verified

Stable channel

Untouched · no stable pointer published

Every native lane built and tested on its target runner, completed the offline smoke test, produced its SBOM and Sigstore provenance, passed OpenPGP certification, and moved through private staging and owner-gated promotion without rebuilding. Independent public verification then downloaded every versioned object and checked its bytes, hash, and cache policy.