The iterative workflow

An AI coding agent does not start a task by editing. It starts by finding the right code: which concrete type runs, which handler processes a request, what a change will break. On a large .NET solution, directory listings and blind file reads make that search expensive. Fuse provides targeted calls against a warm persistent index. This page is the mental model for how an agent strings those calls together. For what one call does internally, see How Fuse works.

Where Fuse Fits

Without a semantic index, the agent reads, decides, and reads again. Fuse supplies a compact map and typed evidence for the parts of that loop that the local index can answer:

• Tokens. Every file opened to orient burns context window that could have gone to the change. Much of it is comments, using directives, and boilerplate the agent does not need to reason about structure.
• Repeated context. Discovery is sequential, and later calls can overlap files already returned. A shared session elides unchanged repeats.
• Lost wiring. Reading files in isolation, the agent reconstructs the DI registrations, the request and route handlers, and the public surface by hand. Grep finds the interface but not the type that runs.

The Measured Boundary

The benchmark corpus measures two distinct paths. Open-ended localization from a title alone recalls 37.7 percent of changed files at a median 1,348 tokens. Review starts from a Git diff, keeps those changed files as must-keep seeds, and packs the branch context at 93.4 percent precision in a median 1,026 tokens. Review does not discover the changed files.

The loop benchmark does not show fewer build and test calls. The Fuse arm records a mean 3.1 agent-visible dotnet build plus dotnet test calls, compared with 3.2 for native. Its measured result is higher test-verified pass@1 (89 versus 82 percent) and one fewer false-done rollout (8 versus 9), not a build-round-trip reduction.

The Verified-Edit Loop

Across a task, an agent can use targeted calls against the warm index:

Survey the workspace cheaply with fuse_workspace (action=map), localize a task to candidate files with fuse_find (kind=task), resolve the wiring a task names with fuse_find (kind=service|request|route|config), plan and emit the context with fuse_context, and review a branch with fuse_review. Pass a sessionId across calls so later calls do not resend files already returned. See Sessions and deltas for the refinement step and Context for an agent for a worked sequence.

When a git base is available, fuse_review is the routed default: it seeds on the changed files and returns the blast radius and packed context in one call. When the task is open-ended, start with fuse_find (kind=task); when it names a route, interface, request, or config section, start with fuse_find (kind=service|request|route|config).

Honest Boundary

File count is not a round-trip count: an agent can read several files in one tool call or revisit one file across several calls. The loop suite counts actual agent-visible build and test calls and finds no material reduction. Review metrics describe compact Git-seeded branch context; localization metrics describe open-ended discovery.

Run the Loop

Start with the strongest anchor available. After the first edit, send the proposed file content to fuse_check. Before changing a public signature, call fuse_impact. Run fuse_test for the edited symbol, then call fuse_review against the branch base before handoff.