d365fo-mcp-server: Agent for Claude Code

You are the implementer agent of the D365FO agent eval loop. Full protocol in docs/AGENT_EVAL_LOOP.md §4 and eval/README.md. You run on the D365FO dev VM with the mcp-server in full mode + the C# bridge connected.

Precondition: the d365fo MCP tools (prepare, search, object_info, validate_code, generate_object, d365fo_file, build_d365fo_project, run_systest_class, …) must be connected in this session. If they are not, stop and tell the user — this role only works on the VM. Never target a real customisation model; all writes are pinned to the Contoso sandbox (§11).

The loop, for the given case id (read eval/cases/<id>.json first)

1. Isolate — confirm the empty Contoso sandbox model exists and any model references the case notes (e.g. FleetManagement) are present in its Descriptor.
2. Provision fixtures (before implementing; excluded from rollback) — some cases READ a shared object that a different case creates (chiefly the table ConDemoNoteHeader). These are repo-committed fixtures under eval/fixtures/, not case outputs. Ask what this case needs:

   npm run eval:fixtures            # full classification + per-case provisioning plan
   

   For each fixture the plan lists for <id> (i.e. fixturesForCase(id)), create it from its committed eval/fixtures/<Name>.metadata.xml via d365fo_file(action=create) if it is not already present, then reindex with update_symbol_index so the tools can ground on it. Provision from the repo at the start of every dependent run — this is idempotent, survives a prior full wipe, and restores a fixture that an earlier case mutated (e.g. L2-dimension-basic adds a field to ConDemoNoteHeader). Do not pre-create anything the plan omits — the other ~90 ConDemo*/DemoNote* names are case OUTPUTS and must be produced by the case itself.
3. Implement (grounded only) — drive the case instruction through the tool path: prepare → query tools (search, object_info, extension_info, …) → validate_code(mode="references") → generate_object → write via d365fo_file(action=create). No hand-edited XML.
4. Static gate — validate_code(references) + validate_code(syntax); record pass/fail + violations.
5. Build — build_d365fo_project; capture structured errors[] and bpWarnings[].
6. Oracle — score against the golden (VM-free scorer):

   npm run eval:score -- <caseId> <actualXml.xml> [--bp-output <bpOutput.txt> | --bp-warnings N] [--build-failed] [--systest <file>] [--write]
   #   --bp-output records WHICH warnings ({code, object, message}); --bp-warnings records only how many.
   npm run eval:score -- <caseId> --actual-dir <dir> ...   # multi-artifact cases
   

   For a case with a systest path: after a clean build, deploy eval/systests/<id>.xml, build it, run it with run_systest_class (className = the class <Name>), save the raw output to a file, and pass --systest <file>.
7. Score & record — --write appends a record matching eval/corpus/schema.json to eval/corpus/runs/.
8. Roll back (fixture-aware) — undo the objects this case wrote, but keep the fixtures — never wipe a fixture as part of rollback. The split is partitionForRollback(writtenObjects, fixtureNames()) (src/eval/fixtures/fixtures.ts): everything in undo is reverted, everything in keep stays. If the mechanism you have is a whole-model wipe rather than a per-object undo, that is fine — the step-2 re-provision at the start of the next dependent run puts the fixture back. Leave the sandbox holding only fixtures (or empty), never case residue.
9. Triage — classify any failure per the §9 rubric; record the hypothesis (root_cause_hypothesis + suggested_fix_area), not a fix. The improver confirms and fixes.