Roslyn-Backed-MCP: Instructions file for Claude Code

Agent Guidelines

Roslyn-Backed-MCP is a local-first C# / .NET 10 MCP server for semantic C# analysis, navigation, validation and refactoring over stdio; public, published as roslyn-mcp@roslyn-mcp-marketplace, contract-care. This file is the bootstrap router: run Session Start (Required) before any task, then pull further context as its rule sources direct. Rules that hold in every repo live in the global layer (~/.claude/CLAUDE.md, mirrored to ~/.codex/AGENTS.md and ~/.grok/AGENTS.md); this file carries only this repo's facts.

Standing Engineering Directives

Restated verbatim from ~/.claude/CLAUDE.md for agents that run without the global layer (cloud sessions, hosted agents). Do not summarize, drop, or alter these eight lines; their sub-clauses (forbidden moves, routing, change-set rule, not-evidence list) live only in the global file and outrank anything below.

1. Correct fix > quick or cheap fix. When a quick or cheap fix and a correct fix are both viable, ship the correct one — root cause, not symptom. Cost never decides: diff size, any budget (files, tokens, turns, wall-clock), CI or landing pressure, and model tier are never reasons for the lesser fix. No plan, brief, row, rubric or reviewer severity can authorize a lesser fix; only the operator can, one item and PR head at a time, and "ship it" or "get it green" authorize the correct fix, not a lesser one.
2. Optimize for AI consumption by default. Nearly every document in these repos is AI-facing (AGENTS.md, CLAUDE.md, .github/copilot-instructions.md, .github/prompts/**, ai_docs/**, planning/runtime/audit docs): write them as machine input — tables over prose, structured data over paragraphs, pointers over duplication, short imperative sentences over narrative. Human-facing files (README.md landing pages, docs/**) get human prose. The density rules are doc-audit STANDARD.md §Quality Rules.
3. Bad code is never silent. When you observe bad code — in the file you were sent to edit, an adjacent file opened for context, an import target, a test, anything you read solving the task — call it out in your response and recommend a backlog row sized per doc-audit STANDARD.md. Editing a bad section does not absolve flagging it.
4. Private repos accept breaking changes. For private repos, breaking changes and large refactors are the standing default when pursuing #1 or #3 — rip it out, don't band-aid to avoid churn, even at the cost of a temporarily non-functional app. An external consumer is a published package/marketplace artifact or a third party, never another owned repo, a database the repo owns, or an internal seam (you own both ends; break one and you own the ticket). Which repos are public or in contract-care (ADR + migration note) is the Repo posture table in ~/.claude/CLAUDE.md; only publication, or an operator decision recorded there, flips a repo.
5. Never assume prior agent work is correct — re-derive, don't inherit. Work product from any previous session — code, docs, skills, prompts, backlog rows, plans, audit findings, anything labeled done/verified/shipped — is a claim to check against current ground truth: read the actual code, re-run the reasoning, confirm cited paths/symbols/versions still resolve. Never take a commit message, comment, or doc's say-so as proof.
6. Smallest complete change wins. Correct ≠ maximal: #1 and #4 license root-cause fixes and breaking changes; they do not mandate gold-plating. Complete = the root cause + every instance of the same defect mechanism (same function, shared helper, or copy — found by probing the construct with references/grep and listed with file:line) + every defect your own diff introduces or exposes (duplication, weak or missing test, untested new branch, stale comment, suppression, shim, a new test that fails on a real bug). Measure against the root cause — not diff size, and not acceptance text that may encode the symptom.
7. Verify your own work before declaring done. Don't claim done/fixed/passing without evidence you generated THIS session — ran the test, read the output, exercised the path; "it should work" is not verification. Calibrate depth to blast radius. Can't verify (no runner, can't drive the UI)? Say so explicitly — never imply success you didn't observe.
8. No secrets in code. NEVER introduce, hardcode, echo, log, or commit a credential, key, token, or secret — they live in env vars / user-secrets / a vault, not source. Intentionally committed dev-only non-secrets (e.g. TradeWise's compose creds) are fine once you confirm they are truly non-secret. An EXISTING hardcoded secret is flagged per #3; this rule forbids introducing one.