Delegating the sweep to dotnet-explore

This plugin ships one read-only navigator subagent, dotnet-explore. It turns a prose task into the symbols it is about: symbolIds, use sites, affected files, and how far a change would reach. It never judges code and it cannot change it — the invoking agent decides what to do with the map.

Host adapter

Claude Code can launch the bundled dotnet-explore custom agent directly. Codex plugins do not currently install custom-agent registrations. In Codex, launch the built-in read-only explorer and start its brief with: "Read <pluginRoot>/agents/dotnet-explore.md and follow its Markdown body; ignore only the Claude-specific YAML frontmatter." The file under agents/ remains the one role contract; do not copy its body into this skill or a Codex-only agent file.

Why delegate rather than fan out yourself. A wide search_index plus a get_references per candidate is the single most expensive shape of read this server produces, and almost all of those tokens are scaffolding you throw away once you know the answer. The agent pays them in its own context and returns a report a fraction of the size.

When to invoke it — and when not to

Invoke it when the symbol set is genuinely unknown: an unfamiliar subsystem, "how would this feature land", "what breaks if I change this", a review or refactor whose boundary you cannot yet name.

Skip it when:

• The symbol is already known. A two-call lookup is not what this is for; use dotnet-read.
• You need the text, not the map. The agent never reads a .cs file; it reports declarationSites, and the ranged Read over them is yours to make.
• The answer is in a non-C# file. .csproj, .json, .md, .editorconfig are out of its scope; it will say so and stop.

Step 0 — check readiness before launching

Call workspace_status yourself, before spawning. The agent checks too, but a cold or degraded workspace wastes an entire agent run:

• Still loading, or index_only → semantic results are unavailable, not empty. Wait, or accept a syntax-tier map and say so.
• degraded (projects failed to load) → the map may be silently wrong. Fix the build and reload_workspace before spawning, or don't spawn.
• Just did a git pull/checkout/rebase, or added or deleted a .cs file → reload_workspace(scope: "all") first.

What to tell it

Everything you state is context it does not have to re-derive, and re-derivation is the expensive part of a cold start.

• The task in prose (required): what someone is about to do to this codebase — not "find FooService", but "we want to add cancellation to the training loop".
• Any symbol, file, or project you already know is involved. It will start from there instead of guessing search terms.
• Depth: "where is this" (a one-hop map, ~8 calls) vs. "what breaks if I change this" (full blast radius, up to its 20-call ceiling). Transitive reach costs real tokens — ask for it only when the question is about ripple.
• A taskId you want it to use, if you plan to read its exact cost back with get_retrieval_metrics(groupBy: "task"). Otherwise it mints its own and reports it.

Parallelism

One instance per independent question, launched in a single message. Do not split one question across instances — the second would re-derive the first's search terms from scratch, and each cold start re-pays the whole baseline. Two genuinely separate subsystems is two instances; one subsystem explored "twice as thoroughly" is one.

Reading its report

Fixed sections: Target, Entry points, Blast radius, Affected files, What would need to change, Suggested next calls, Not covered.

• symbolIds are the payload — paste them straight into get_symbol, get_references or rename_symbol. A symidx_ prefix is provisional (syntax tier) and a symfb_ one is not a fetch target at all (a local function, a lambda, a symbol with no doc-comment id); the report flags those on the row.
• Read Not covered first, not last. It is mandatory and carries limitedBy verbatim, budget stops, non-C# files the answer touches, and any ambiguity it resolved on your behalf. A map that says where it ends is useful; one that looks complete is dangerous.
• Suggested next calls names the write path but never performs it. Verify a suggested call against the actual schema before running it.
• It reports locations, not conclusions. No quality judgments, no design opinions, no proposed diffs. If it offered one anyway, treat it as noise.

Then what

The map is the input to a change, not the change. Invoke dotnet-write next: its step 2 — Read the exact declaration span, then Edit anchored on that text — is exactly the step the agent deliberately did not do for you.