Edt MCP Ready To Deploy

EDT-MCP — ready to deploy (final checklist)

Run this when a piece of work is finished — it is the gate that proves the whole thing still works. Walk it top to bottom and stop on the first red; each step proves a layer the previous one can't. Companion skills do the "how": edt-mcp-build-test, edt-mcp-e2e-testing, edt-mcp-testing, edt-mcp-bilingual, edt-mcp-tool-conventions.

A green lower tier does NOT prove a higher one. "Compiles" ≠ "tests pass" ≠ "works live" ≠ "protocol-conformant". Don't declare done on review/grep alone.

The ordered gate

1. Code hygiene

   • Surface is English only (tool descriptions, errors, README, skills) — no Russian prose, no transliteration.
   • Cyrillic in code only where it is real 1C/BSL data the code matches or documents (type tokens, BSL keywords, example FQNs) — keep it; regexes use \uXXXX. No raw Russian-prose comments. Fix the 1С→1C homoglyph (Cyrillic Es). (See CLAUDE.md don't #7, edt-mcp-bilingual.)
   • Parameters lowerCamelCase; errors via ToolResult.error(...).toJson(); model touched only inside a transaction boundary; markdown table cells escaped.

2. Tests written (the ratchets fail the build otherwise)

   • New/changed tool → an XxxToolTest (unit) and tests/e2e/tools/test_<tool>.py (e2e).
   • Changed metadata/code resolution → a bilingual case (English Name, Russian Name, synonym). (edt-mcp-bilingual.)

3. Build + unit (Tier 1) — bash source/compile.sh → BUILD SUCCESS + every ratchet green (BuiltInToolTestCoverageTest, ToolContractConsistencyTest, e2e coverage ratchet). Toolchain isn't on PATH — pass --java-home/--maven-home.

   • Then run the agent-pack validator: python docs/validate_agent_skills.py. It is a SEPARATE CI job (Validate business-project skills) that compile.sh does not cover, and it owns its own ratchets — the tool-docs↔registrar match and the single-word tool-name list. On Windows ignore its CR/CRLF is not allowed lines: those are the local checkout's line endings, and the committed blobs are LF (git show HEAD:<file> | grep -c $'\r' → 0).

   • Touched the target platform, or a call into an EDT/platform API? Also compile against the older EDTs, once per installation — each with its own flags:

     • 2026.1: bash source/verify-oldest-platform.sh <2026.1-install-dir> (the defaults: JavaSE-17, SVG dropped);
     • 2026.2.0: bash source/verify-oldest-platform.sh <2026.2.0-install-dir> --ee JavaSE-25 --keep-svg — without the flags it dies in platform resolution (osgi.ee 25 missing) before testing any API.

     The build compiles against the NEWEST platform, and neither the BREE nor the lower-bound imports can express "no signature that only the newer release has" — that binds fine and fails at the call site with NoSuchMethodError on the older EDT. The script points the target platform at a local installation and compiles; it restores the working tree on every exit path. Run it against 2026.1 (the oldest supported major) and against 2026.2.0 (the previous service release of the current major). The second one is not optional cover: 1C deletes the previous service release from its channel, so CI builds only against what ruby/2026.2/ serves today and a local install is the only remaining proof that the single artifact still loads on 2026.2.0.

4. README updated — bump the tool count (both places), add/adjust the group table, the flat tool table, and the detailed section; parameters must match the schema.

5. Live redeploy (Tier 2) — redeploy the freshly built bundle to the non-elevated dev EDT copy (edt-redeploy.ps1; signal MCP server UP on 8765, exit 1 ≠ failure). Anti-stale: confirm the DEPLOYED jar contains your change (unzip -p <jar> path/Class.class | grep <literal>). Wait for both projects ready. If EDT wedges → kill + -clean relaunch.

6. Golden — if tools/list changed (new tool / description / schema / annotations), regenerate and commit the golden (EDT_MCP_UPDATE_GOLDEN=1 python tests/e2e/run_all.py --project TestConfiguration --filter tools_list); the golden snapshot test must pass.

7. Full e2e (Tier 3) — python tests/e2e/run_all.py --project TestConfiguration → all pass (a handful of env-gated skips are normal), fixture clean: True. (edt-mcp-e2e-testing.)

8. Protocol conformance — run the official suite against the live server and pin known gaps: npx @modelcontextprotocol/conformance@latest server --url http://127.0.0.1:8765/mcp --spec-version 2025-11-25 --expected-failures tests/conformance/baseline.yml → no NEW failures beyond the committed baseline. A new failure = a real protocol regression → fix the server. (tests/conformance/README.md explains the layer split: e2e = business logic, conformance = protocol.)

9. Clean git tree — git status clean: no stray files, no leftover fixture churn (the e2e run reverts its fixtures; if not, sync them). The diff should read as exactly the intended change.

10. Commit — adversarially review the diff, then commit locally on the working branch. Do NOT push or merge unless explicitly asked.