ADR-0021: The command line with the agent reporter is the primary agent surface; MCP and LSP are additive

• Status: Accepted
• Date: 2026-09-20
• Derives from: design.md § Rules an agent can implement and follow, § Where it would be ignored (MCP over CLI), § Two front-ends that will matter more than the MCP server, § Where each lands, § How to know, rather than believe
• Constrains: architecture.md § Agent surface
• Implemented by: Wave 1 plan, Wave 2 plan, Wave 3 plan

Context

Agents in Claude Code reach for Bash first and use an MCP server when it is the only door. They live in compiler and linter output. A separate ritual beside tsc, eslint, vitest and dotnet build is skipped.

Decision

1. Wave 1: the agent reporter (token-budgeted, fix-cost ordered, with member references and decision links), fix, examples, expires, line and column, stable ids, receipts, test, explain and explain --plain, can-import, config lint, hooks install --claude-code (SessionStart brief, PreToolUse impact, Stop affected cruise), attest, and --require-comment-token.
2. Wave 2: docs (agents-md, contributing, skill), propose, impact, place, test --generate, decisions, the test adapters, sarif and junit, and eslint-plugin-rulebearing.
3. Wave 3: serve --mcp, serve --lsp, the Roslyn analyzer, guard --watch, --mode source for .NET. All are thin loops over the same commands and the same cached graph; none has its own rule file.
4. Adoption is measured from wave 1 by the six signals in the design (first-run pass rate above 90%, one-turn median to green, authoring guardrails catching rules, Stop-hook p95 under 2 s, fix on above 80% of rules, zero merged budget raises). If the first two do not move, the agent surface is cut back to the agent reporter and the hook.

Consequences

• Nothing in the agent surface may disagree with the gate, because it reads the same config and graph.
• The Stop hook's two-second budget is a performance requirement on --affected and the cache, not a nice-to-have.