A small surface: find, impact, path and tests, answered with their code; CI runs the query cases

[swapnilpaliwal-sd]

A small surface. Agents were offered 8 MCP tools with 69 parameters and about 10,000 characters of tool descriptions. The CLI had 10 to 37 flags per verb. In practice agents asked two questions with no options, and after almost every answer that located code they opened the file to read it.

The offering is now four questions plus a setup verb, the same on the CLI and over MCP, with no options:

CLI	MCP	answers
axiomcode find "<question>"	find(question)	where the code for a task lives, including names the code calls that nothing declares
axiomcode impact <name>	impact(name)	callers, what a change reaches, and the tests
axiomcode impact	impact()	the same, for the declarations your uncommitted edits changed
axiomcode path <A> <B>	path(start, end)	the call chain
axiomcode tests	tests()	the tests your edits reach, and the command that runs them
axiomcode index	—	builds the graph (--lang, --src, --library)

Answer shape

• Each answer is a numbered list of places. Each place carries its enclosing function's code, so the agent does not have to open the file.
• There is one block per function, and at most 10 places.
• Rows that match only a word, and module-scope rows, are dropped.
• The shape applies at the front doors, the installed command and the MCP server, when no flag is passed. Calling the dispatcher directly, setting AXIOMCODE_RAW, or passing any flag gives the verb's own answer. So hooks, test suites and scripts are unchanged.
• The old verbs and flags still work; they are no longer advertised.

Docs and hooks. Help, SKILL.md, AGENTS.md, the Cursor rule, the Gemini copy, the README and the hooks' hints teach only these verbs. An MCP answer drops any clause that names an option the tools refuse; code blocks are never touched.

CI gate against regressions

• The engine (<lang>) job now runs tests/run.py --lang <lang> after the engine suite. Every case must pass, and a pending case that passes fails the run until its pending mark is removed. Five cases had regressed on 0.1.9 because this suite never ran in CI (fixed in impact: one row per TypeScript signature use; pipeline-array-dispatch is no longer pending #1812–tests: csharp unmodelled-entry case follows the now-modelled hosted-service registration #1815).
• The python leg also runs tests/front_door.py. It checks the answer shape through the installed command and MCP, and checks that direct calls and flags keep the old answers.

Validation (on the integration tip)

• tests/run.py: 1249/1249.
• 15 standalone suites, including front_door, surfaces, mcp, mcp_docs, manifests and latency: all pass.
• Release end-to-end harness (every verb, edit and rebuild rounds, five languages): 0 failures.
• MCP check (find, impact, path and tests on five languages): 35/35.