Strip MCP server to a 3-tool warm-start set#99
Merged
Conversation
Reduce the advertised tool surface from 11 to 3 (OutlineCSharpFile, MinifyFile, TraceDiRegistrations) and reframe the project as a structural warm start for coding agents. - Demote the focus/trace family (FocusMethod, FocusMultipleMethods, FocusMethodsAcrossFiles, FocusType, FocusCallers, TraceCallers, TraceImplementors, MapProject) from MCP tools to internal API; the test suite still exercises them. - Delete legacy duplicate FocusedEmissionTool, which registered a second focus_method tool. - Drop the Markdown emitter. - Rewrite ServerInstructions for the 3-tool set; fixed per-session overhead drops 4,134 -> 1,415 tokens. - Reframe README, NuGet readme, viewer pages, and llms.txt around the warm start: it curbs wasteful reading, with the biggest win on smaller/cheaper models and large files; drop the "30-95% for everyone" framing. 170/170 tests pass; tools/list confirms 3 tools. Co-Authored-By: Claude Opus 4.8 <[email protected]>
This was referenced Jun 16, 2026
Merged
Byggarepop
added a commit
that referenced
this pull request
Jun 19, 2026
The 3-tool strip (#99) left the live MinifyFile and OutlineCSharpFile descriptions and the runtime low-saving nudge pointing at focus_method, and TraceDiRegistrations referencing trace_implementors -- none of which are exposed tools anymore. A model that follows that guidance reaches for a tool that does not exist and falls back to a whole-file Read. Repoint them to outline_c_sharp_file plus a narrow Read, and lead the two single-file tools with an imperative trigger so the host model picks them up before reading a large C#/VB file. Descriptions only; no behavior change. Claude-Session: https://claude.ai/code/session_0158rjCdskesQmgaVnW8x6yi Co-authored-by: Claude Opus 4.8 <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Reduces the advertised MCP tool surface from 11 tools to 3 and reframes the project as a structural warm start for AI coding agents.
Tools
Keeps
OutlineCSharpFile(warm-start primitive),MinifyFile(whole-file, all languages), andTraceDiRegistrations(DI wiring — the one discovery a text search can't do cleanly). The focus/trace family (FocusMethod,FocusMultipleMethods,FocusMethodsAcrossFiles,FocusType,FocusCallers,TraceCallers,TraceImplementors,MapProject) is demoted from MCP tools to internal API — the test suite still covers it, so nothing is lost and they're trivially restorable.Also deletes a legacy
FocusedEmissionToolthat was registering a duplicatefocus_methodtool, and drops the Markdown emitter.Overhead
ServerInstructionsrewritten for the 3-tool set; fixed per-session overhead drops 4,134 → 1,415 tokens (−66%). On a real Sonnet task this flips the session ledger from net −402 to net +1,460.Why
A controlled A/B (outline-only vs full toolset vs no-MCP) showed the focus/trace tools added no measured token value on scoped tasks — the durable win is a cheap structural warm start, after which
Grep+ a narrowReadof the outline's line-ranges does the rest. Measured end-to-end saving scales inversely with model capability: ~1% top-tier, ~5–7% mid-tier, ~8% on cheaper models; large files save the most per read.Docs
README, NuGet readme, viewer pages, and
llms.txtreframed around the warm start — biggest win on smaller/cheaper models and large files. The "30–95% for everyone" framing is dropped; the copy is explicit that it does not make a top model meaningfully cheaper and does not change the code the model writes.170/170 tests pass;
tools/listconfirms exactly 3 tools over the wire.