Context for an agent on a large codebase

Give an AI agent the context it needs by starting wide and narrowing, so each request stays inside the token budget.

Goal: hand an AI agent the context for a task on a large .NET codebase without spending its window on files the task does not touch.

The principle is to start wide and narrow: a cheap map first, then a scoped drill-in. Each step is one MCP tool call that returns scoped, reduced context. The payoff is measured: in the Layer 4 scenario, one scoped call acquired a task's context in about 40,000 tokens against roughly 512,000 for a generic packer dump at the same single call, while a blind agent must read each needed file at least once (a lower bound of about six round-trips over the 24 pull requests tested). Read that token number with its recall: one scoped query call reaches 51 percent of the changed files, and the change-scoping mode reaches 88 percent when a git base is available.

The sequence

Survey. Call fuse_toc first. It returns a directory tree with per-file token costs and a symbol outline, a low-token map of the whole codebase that also tells the agent what fetching each file would cost. Use fuse_skeleton instead when you want signatures rather than an outline.
Drill in. When the agent knows the area by name, call fuse_focus with the seed. When it has a topic but not a name, call fuse_search with the query. Both expand from the seed through the dependency graph.
Review changes. For a pull request, call fuse_changes with the base branch as the git ref, and set review to true for diff hunks and direct callers per changed file.
Full control. When you need options the workflow tools do not expose, call fuse_dotnet.

Example call sequence

fuse_skeleton(path="C:/Projects/MyApp/src", maxTokens=80000)
fuse_focus(path="C:/Projects/MyApp/src", focus="OrderService", depth=1, maxTokens=150000)
fuse_changes(path="C:/Projects/MyApp/src", changedSince="origin/main", maxTokens=100000)

The first call orients the agent, the second pulls the order-processing area and its dependencies, and the third scopes a later review to the branch under review.

Budget guidance

Stage	Tool	Typical tokens
Survey	`fuse_toc`	5,000 to 30,000
Architecture map	`fuse_skeleton`	50,000 to 100,000
Focused drill-in	`fuse_focus` or `fuse_search`	100,000 to 200,000
Change review	`fuse_changes`	50,000 to 150,000
Full fusion	`fuse_dotnet level=aggressive`	200,000 to 800,000

Shortcut: ask once

fuse_ask collapses the survey and drill-in: give it a task and a token budget and it chooses skeleton, focus, or search and packs to budget. See Ask one question. Across calls in one task, pass a session id to fuse_focus and fuse_search so files already returned are not sent again.

When to use it

Use this progressive flow on any codebase too large to hand an agent whole. On a small project, one fuse_dotnet call is simpler.

The tool parameters are in the MCP tools reference, and the modes behind them in Scoping.