Atlas Scout Preview 29 · Official release · August 17, 2026
A better index is not enough. Scout must know when not to answer.
A 120-run real-world coding campaign showed both sides of structural navigation. Scout saved substantial work on one task class, yet some correct solutions became more expensive when broad context exposed duplicate or unnecessary paths. Preview 29 turns that evidence into a stricter product contract: incompatible indexes cannot answer, compact context favors the implementation over its mirrors, and uncertain relationships stop before they become invented dependency chains.
Extractor generation · fail-closed admission · machine-readable recovery
Stale structure no longer gets the benefit of the doubt.
Source drift and incompatible extraction logic are different problems. Ordinary file changes can still use Scout’s bounded stale-while-revalidate behavior. An index produced by incompatible extraction code can no longer serve structural evidence at all. symbol_schema remains available for diagnosis while every evidence-serving tool returns atlas_scout.index_refresh_required until the refreshed generation is healthy.
The gate checks both a persisted extractor-projection fingerprint and the language and extractor versions already recorded in the database. A copied, partially upgraded, or damaged cache cannot quietly answer with incomplete old structure.
One compact orientation · canonical representatives · exact ranges
Broad context now points toward the implementation, not every copy of it.
Compact fast_context gives direct task evidence priority over identifiers derived from the prompt, discovers likely tests from the strongest production files, and hydrates read ranges around the best evidence. Exact header, vendor, fixture, and tool clones collapse to one canonical representative in compact mode. Expanded detail and exact symbol tools still expose every indexed path.
AT005 replay
Less duplicate context.
On the refreshed Linux fixture, compact context kept the production implementation, a relevant self-test, and the canonical UAPI header while suppressing two exact tools mirrors. The response was 43.6% smaller than expanded detail.
Agent guidance
Stop after sufficient evidence.
Broad work starts with one compact context call. Agents read the cited ranges before another broad navigation call, avoid unnecessary initial detail and limit expansion, and stop when the evidence answers the task.
Closed evidence classes · recomputed heuristics · bounded topology
A plausible name match is not a dependency chain.
Graphs, traces, paths, and edit-impact frontiers now extend only through receiver-typed or proven language-rule relationships. Name-only and unknown observations remain visible at the first hop, with their uncertainty intact, but cannot manufacture downstream topology.
Heuristic targets are revoked and recomputed after relevant index changes. Receiver registries stay inside one language family, Rust bare-import provenance is retained, and a typed receiver that does not select exactly one indexed method blocks later global fallback. New collisions remove old certainty instead of leaving a stale edge behind.
C/C++ recovery · safe call shape · explicit precision blockers
Deeper language resolution, without pretending to be a compiler.
C and C++
Recover more, claim less.
Kernel-style loop macros no longer hide declarations that follow them. Scout uses a position-preserving parse projection and marks the recovered file partial. Persisted call argument counts reject only provably incompatible fixed-arity overloads; defaults, variadics, operators, and uncertain signatures remain unresolved.
Python and PHP
Negative knowledge matters.
Curated runtime call names remain unresolved when their implementation normally lives outside the workspace. Exact import evidence or an intentional same-file shadow still wins, and reference results explain which precision guard blocked a resolution.
Signed tag · four native targets · independent public verification
The promoted bytes are the tested bytes.
- Signed tag
v1.0.0-preview.29- Release commit
fc52efaf15fd25f09e49d88bc1a0b2861c6456c2- Release manifest SHA-256
4aaba3a233fbcfa03b366f770922d0ee8ab498d26b5d98614018e06272fb1ca5- Candidate tests
- 514 passed per macOS/Linux lane · 483 passed on Windows · 0 failed
- Native archives
- Linux x86-64 · Linux ARM64 · macOS ARM64 · Windows x86-64
- Stable channel
- Untouched · no stable pointer published
Every target completed native preflight and candidate validation. The candidate was certified locally, staged privately without rebuilding, promoted through the owner-gated workflow, then downloaded again from the public origin. OpenPGP signatures, Sigstore provenance, exact archive sizes, and SHA-256 values all passed.
Public packages · live host checks · exact directory status
Preview 29 is available through every distribution path we can verify directly.
MCP distribution
Four native packages, one verified universal bundle.
Preview 29 is active and latest in the Official MCP Registry. The GitHub MCPB release exposes all four native targets. Smithery serves the unchanged universal bundle; a fresh download reproduced its frozen SHA-256 and all fourteen tool schemas are public.
Agent integrations
Signed wrappers, tested as real packages.
The signed Preview 29 Codex and Claude Code wrappers each passed a clean live-agent symbol_searchcheck against their public packages. The Preview 29 skills-only package is also approved and published in the OpenAI Plugins Directory. Anthropic’s community-directory review remains pending and is not presented as published before that directory says so.
Preview 29 is the current public preview
Upgrade the complete managed installation.
The updater verifies the signed release contract, checks pointer freshness, stages the runtime and integration payload, and atomically promotes the managed installation.
atlas-scout upgrade --check
atlas-scout upgrade