{ "markdown": "\n\n

\n \"Fuse\"\n

\n\n

\n \"CI\"\n \"NuGet\n \"NuGet\n \"License\"\n

\n\n# Fuse\n\nFuse is a local .NET tool with a persistent syntax index, typed-graph wiring\nresolution on demand, reduced task-scoped source, and pre-write compiler verification\nfor coding agents. It stores repository facts in `.fuse/fuse.db` and reuses them across\nagent turns instead of rediscovering the same structure through repeated file reads and\ntext searches.\n\nFrom a .NET project inside a Git repository:\n\n```bash\ndotnet tool install -g Fuse\nfuse mcp install\n```\n\nThe installer supports Claude Code, Cursor, GitHub Copilot, OpenCode, Kilo Code, Codex,\nand Grok Build. Use `--client ` to configure one client; the default `all` configures\nall seven for the selected scope. `fuse mcp install` writes MCP registration and a short,\nversioned managed block in the client's documented instruction file, such as `AGENTS.md` or\n`CLAUDE.md`. Pass `--no-rules` to register MCP without the managed block. `--with-hooks`\nseparately writes project-scoped Claude Code hooks. Run `fuse mcp doctor` after installation\nto inspect registration, managed guidance, daemon reachability, and index state. See [Connect\nyour coding agent](https://fuse.codes/docs/start/connect-your-ai)\nfor the exact file and scope matrix.\n\nReload your MCP client, then ask:\n\n```text\nResolve IOrderService to its implementation, then check the proposed OrderService.cs edit\nwith fuse_check before writing it.\n```\n\nWhen the MCP server starts, its shared local daemon owns one index job for the repository.\nA cold read starts a syntax index and reports its job state. Run `fuse index` to render\nprogress in a terminal, `fuse index --semantic` for explicit compiler analysis, and\n`fuse index status` or `fuse index cancel` to control the job. The installer adds `.fuse/`\nto `.gitignore` at project scope.\n\nEvery MCP operation except `fuse_reduce` requires a Git repository identity. Fuse walks upward to the nearest\n`.git` directory or file, so a call from a nested source or output folder uses the same\nrepository-root daemon, lock, and index. A non-Git folder returns\n`workspace_identity_unresolved:` without starting a daemon or writing an index;\n`fuse_reduce` remains available. Each warm index records its repository root and complete\nfile inventory. A missing or incomplete manifest rebuilds automatically even when the\ndatabase already contains file rows.\n\n

\n \"An\n

\n\nAnalysis runs locally and can work offline. Fuse walks a typed graph of DI registrations,\nhandlers, routes, and callers, emits reduced source for a selected scope, and lets a coding\nagent typecheck proposed single-file content before writing it. No model is required. The\noptional update check can contact NuGet, and build-grade operations can use the package\nfeeds configured for the repository.\n\n## Persistent Discovery and Reduced Context\n\nCoding agents can inspect a repository through file reads, grep, and regex. On a large\nsolution, those operations can rediscover the same symbols, references, registrations, and\nproject structure across several turns. Fuse performs that discovery through MSBuild and\nRoslyn, persists the result, and incrementally re-indexes files as they change.\n\nWhen a project loads semantically, the graph records DI registrations, request handlers,\nroutes, options bindings, and call edges. When it does not, Fuse falls back to syntax-level\nindexing for that project and reports the mode.\n\n- **Resolve wiring.** `fuse_find` traces a service, request, route, or configuration\n section to the code that handles it. Text search finds `IOrderService`; Fuse follows the\n registration to the implementation that runs.\n- **Pack branch context.** `fuse_review` seeds on the git diff and returns related callers,\n handlers, and tests with provenance. On 69 recorded pull requests the median response was\n 1,026 tokens at 93.4 percent precision (`review.json`).\n- **Return less source.** `fuse_context` reduces the selected files under a token budget\n and records why each file was included. Across four recorded repositories, skeleton\n reduction removed 38 to 44 percent of tokens while retaining every measured public and\n protected type name (`reduce.json`).\n- **Read warm.** On the recorded NodaTime run (semantic tier, 14,760 symbols), exact\n symbol lookup took 1.8 ms at the median, task localization 15.7 ms, and review planning\n 106.3 ms (`performance.json`; timings are environment-dependent).\n\n

\n \"Fuse\n

\n\n## Compiler Checks and Change Verification\n\n- `fuse_check` checks the proposed content of one file and returns compiler diagnostics\n without changing the working tree. Oracle grade reuses compiler state captured from the\n real build. Build grade runs a scoped `dotnet build` for the owning project when captured\n state is unavailable. Supported API-shape errors can include a repair packet; in the\n recorded run, the top suggestion repaired 20 of 20 near-miss member and type errors\n (`diagbench.json`).\n- Before changing a public method, `fuse_impact` finds callers, implementations, and\n referencing types. Given a package id and two versions, it returns the break set for\n that NuGet upgrade.\n- When a signature must change, `fuse_refactor` stages the refactor as a diff and\n returns it only when the compiler reports no new diagnostic.\n- After an edit, `fuse_test` selects and runs the test types that reach the changed\n symbol, grouped by owning project, instead of starting with the whole suite.\n\nEvery answer names how it was produced. Fuse calls this the **verification grade**:\noracle grade checks against the compilation captured from the real build, build grade\nruns a scoped `dotnet build`, and when neither compiler path can answer, Fuse abstains\nand names the missing prerequisite instead of guessing.\n\n## What the Recorded Results Cover\n\nEvery result below comes from `tests/benchmarks/results` and has a reproduction command\non the [benchmarks page](https://fuse.codes/docs/project/benchmarks).\n\n- Across 1,000 compiler-labeled edits in the recorded OrderingApp test app (500 breaking,\n 500 neutral), `fuse_check` reported zero broken edits as clean and rejected zero valid\n edits (`checkgate.json`).\n- In the same app, Fuse matched all 24 expected .NET wiring links with no extra matches\n (`semantics.json`).\n- Across 69 real pull requests, branch review retained every git-changed file by\n construction at 93.4 percent precision with a median size of 1,026 tokens (`review.json`).\n- In the reduced-scope agent-loop run, the Fuse arm's edits passed the project's own\n tests on the first attempt in 89 percent of scored rollouts versus 82 percent for\n native tools, with overlapping confidence intervals. The Fuse arm declared success on a\n failing edit 8 times versus 9 for native. Build and test calls were essentially equal at\n 3.1 versus 3.2 (`loop.json`).\n- Across four recorded repositories, skeleton reduction removed 38 to 44 percent of\n tokens while keeping every public and protected type name and 96.3 to 99.4 percent of\n method names (`reduce.json`).\n\nThe opt-in resident workspace answered repeated `fuse_check` calls in 31.2 ms at the\nmedian on the recorded NodaTime run (`resident-latency.json`). That path is faster than a\nfull build for speculative checks; it does not replace normal builds or tests before merge.\n\nThese measurements have defined limits. Read the\n[full methods and results](https://fuse.codes/docs/project/benchmarks) before comparing\ntools or applying the figures to another repository.\n\n## Local and Write-Safe\n\nFuse reads source, compiler state, git metadata, and its local `.fuse/fuse.db` index.\nRead, check, impact, refactor, and review operations do not write the working tree.\n`fuse_workspace` with `action=apply` is the one explicit tree-write path, and it is a dry\nrun unless `write=true`.\n\nCompiler-backed wiring analysis is .NET-only. Other languages receive syntax-level search\nand reduction.\n\n## Related Code-Intelligence Tools\n\nRepository indexes, code graphs, and language-server tools already exist.\n[CodeGraphContext](https://github.com/CodeGraphContext/CodeGraphContext) provides a local\nmulti-language graph, [Serena](https://github.com/oraios/serena) exposes\nlanguage-server-backed symbol tools, and [Sourcegraph](https://sourcegraph.com/code-search)\ncovers multi-repository search and code intelligence. Fuse concentrates on local .NET work\nthrough MSBuild and Roslyn: framework wiring, reduced scoped context, compiler-backed\nproposed-file checks, change impact, and covering-test selection. It can run alongside the\nindex or search built into a coding client.\n\nThe [peer comparison](https://fuse.codes/docs/project/benchmarks#peer-comparison-fuse-versus-codegraph-coa-codesearch-and-serena)\nrecords a bounded, dated experiment and its sampling limits. It is not a general ranking of\ncode-intelligence tools.\n\n## Start Here\n\n- [Quickstart](https://fuse.codes/docs/start/quickstart)\n- [Connect your coding agent](https://fuse.codes/docs/start/connect-your-ai)\n- [How Fuse works](https://fuse.codes/docs/concepts/how-fuse-works)\n- [Tool reference](https://fuse.codes/docs/reference/mcp-tools)\n- [Install options](https://fuse.codes/docs/start/install)\n- [Contributing](https://fuse.codes/docs/project/contributing)\n\nApache 2.0. Copyright (c) 2026 Litenova Solutions. See [LICENSE](LICENSE) and\n[NOTICE](NOTICE).\n", "bytes": 10467, "sha": "a0a2f99afb25a260f75e995dafa58985d2103c6474ae6dfc97cf8c694055d6b6", "repo_slug": "litenova-solutions/fuse", "fonte": "repo", "truncated": false, "api": "https://agentalog.com/api/listings/mcp_io_github_litenova_solutions_fuse_4cfc0783/readme" }