Roadmap

Fuse maintains a persistent .NET index, returns reduced task-scoped source, resolves framework wiring, and checks proposed C# files against compiler state. This page describes where we intend to go next. Nothing here is a commitment or a delivery date. Work that has shipped is recorded in the Changelog.

Resident verified-edit loop

Today, fuse_check can answer from a live resident workspace when one holds the compilation for your root. The next step is to make that path the default edit-time experience: warm once, then answer every proposed edit in tens of milliseconds instead of paying a full dotnet build round-trip.

That loop pairs speculative typecheck with blast-radius reads, covering-test selection, and verify-gated refactors so an agent can iterate on a change without writing broken code to disk. The resident workspace ships opt-in today (FUSE_RESIDENT=1); promotion to on-by-default follows measured warm latency and memory on real repos.

Multi-language syntax tier

C# structural analysis already runs through Roslyn. The syntax-tier indexer is provider-driven: each language registers an ILanguageSyntaxProvider that claims file extensions and extracts symbols and chunks without a compiler. Python and JavaScript providers ship today; community syntax providers are welcome through the same seam.

The typed semantic graph (DI wiring, request handlers, route maps) remains C#/Roslyn-only for now. Multi-language work expands what Fuse can index and retrieve lexically; it does not yet promise compiler-grade verification outside C#.

Honest retrieval limits

Open-ended task localization runs on deterministic lexical retrieval (BM25 over the persistent index plus dependency-centrality priors). On the pinned benchmark corpus it reaches about 38 percent changed-file recall at 21 percent precision with a git base absent; change review with a git base reaches much higher recall because the diff is the seed.

A richer build-exact graph improves verification (fuse_check, fuse_impact, covering tests) but does not materially lift open-ended recall on the same repos. Fuse will not ship retrieval features that benchmark suites do not support, and will not claim recall numbers the harness has not recorded.

Dense embedding reranking and learned per-repo ranking were evaluated and set aside: they added model dependency or tuning cost without measured recall lift on the corpus. The product bet is precise anchors (symbols, routes, services, git diffs) resolved through the graph, with ranked prose localization as fallback, not the headline.

Recently shipped

These moved to the Changelog when they landed:

• The 4.2 shared daemon, eager warm-on-start indexing, live incremental updates, bounded syntax-first cold reads, and compiler-state reuse.
• The semantic-language provider seam, focused index-store components, shared scoping model, and unified API-surface reference.
• Roslyn structural analysis for C# (skeleton, dependency graph, type location, outlines, route maps, semantic markers) as the default path; regex remains for content reduction, the project graph, and pattern detectors.
• Symbol-level scoping (Type.Member focus seeds and member-granularity query retrieval).
• A persistent SQLite store at .fuse/fuse.db for reduction output, analysis, and relevance tokens.

Next

Read the Changelog for shipped history, Benchmarks for measured results, or Contributing to propose or build planned work.