feat(wiki): expose the Path B draft tools over MCP (#579) - #581

ZETETIC-REVIEW: APPROVE

Summary

Registers the two pre-existing Path B handlers (wiki_get_draft, wiki_refine_draft) as MCP tools from a new mcp_server/tool_registry_wiki_drafts.py, wired into merged_schemas()/register_all(), with matching annotations, docs, and tool-count updates across 10+ files. Closes #579 by choosing "register" over "remove", as the issue required. Mergeable as-is.

Stakes classification (Move 7)

High. The diff modifies the public MCP tool surface (two tools become externally callable for the first time) and totals 507 changed lines (424 additions, 83 deletions), over the 400-line advisory threshold. I applied the full set of Moves 1 through 6 with runtime checks, not just static reading.

Move 0, ledger and seen-defect check

No formal "Completion Ledger" heading, but the PR's Test Plan and Audit Notes sections enumerate every new code path (wrapper registration, argument forwarding, one SQLite end-to-end call) with a named test and a mutation note, matching this repo's established PR convention; the immediately preceding merged PR #580 uses the same format. My own re-enumeration of the diff's paths found nothing unmapped: the only new logic is two wrapper functions with no branches; handler_get and handler_refine bodies are untouched, confirmed via git diff origin/main..HEAD -- mcp_server/handlers/wiki_refine.py, which shows annotations and docstrings only.

One claim needed correction. The PR states the 5 errors are tests_py/scripts/test_installer_*. Running the full suite here (UV_NO_SYNC=1 uv run pytest -q) produced 11 errors, not 5: 6 in test_installer_rerun_keeps_entries.py, 5 in test_installer_upgrade_single_version.py. The substance of the claim checks out: .venv/bin/python -m pip --version returns "No module named pip" in this worktree's venv, all 11 errors are pip-install subprocess failures inside those two files, and neither file is touched by this diff (git status --short stayed clean throughout my checks). Since the root cause is independently verified external to the diff, a typecheck-only venv with no pip module rather than a dismissed in-scope defect, I did not apply Move 0's REFUSED short-circuit. The "5" in the PR body should read "11" for accuracy; treated as a non-blocking nit.

Verification performed

1. Registration, runtime. Built a server with register_all(mcp, codebase=False, prd=False) and called list_tools(): both tools present, 54 total, correct annotations (wiki_get_draft read_only_hint=True; wiki_refine_draft read_only_hint=False, idempotent_hint=False). merged_schemas() also carries both; total is 57 with both upstream flags on.
2. Mutation testing, in a scratch copy, restored afterward with git status --short clean. Dropping _register_wiki_refine_draft(mcp) from register() fails 7 tests (test_wiki_draft_tools.py x3, test_main.py x4), including the pinned test_standalone_baseline_is_54_tools. Swapping the two annotation presets (READ_ONLY and NON_IDEMPOTENT_WRITE) in wiki_refine.py fails test_both_path_b_tools_are_registered on read_only_hint. Both mutations are caught; the new tests are not vacuous.
3. Wrapper correctness. mcp_server/tool_registry_wiki_drafts.py lines 36-90. Wrapper parameter sets and defaults match the handler schemas exactly (limit=20, list_pending=False, synth_model="claude_refine_v1"), and tests_py/handlers/test_tool_schema_parity.py already enforces this generically for every registered tool (widened by [bug] recall MCP tool does not expose tags_any, tags_all, include_low_signal, cross_domain, sa_mode #559), so this isn't resting solely on my manual comparison.
4. None-for-optional behavior. wiki_refine.py lines 279-310 (handler_refine) and pg_store_wiki_drafts.py lines 90-137 (update_draft): every optional field is only added to the SET clause when it is not None, so omitting title, lead, sections, or frontmatter is a true no-op for those columns, and passing None explicitly behaves identically to omission. Not new in this PR; handler_refine's body is unchanged by this diff, per the Move 0 section above.
5. wiki_refine_draft called with only draft_id. synth_model has a non-None default ("claude_refine_v1") both in the wrapper and in handler_refine's own args.get("synth_model", "claude_refine_v1"), so update_draft will set synth_model and the call will insert a refined_llm audit memo even though no content changed. This is pre-existing handler behavior, unchanged by this diff, now reachable by real clients for the first time. Not harmful: no data loss, confidence stays untouched per ADR-1065, and "updated" correctly reports the DB write. It does mean a no-op-looking call still writes an audit row and overwrites synth_model. Worth a follow-up issue if the maintainer wants that call shape to be a true no-op; non-blocking, out of this PR's stated scope of register-or-remove per [bug] wiki_get_draft and wiki_refine_draft (Path B) are defined but never registered as MCP tools #579.
6. Count consistency. scripts/check_doc_claims.py: OK. scripts/check_craftsmanship.py --base origin/main: OK. ruff check and ruff format --check: clean. .venv/bin/python -m pyright mcp_server/: 0 errors, 0 warnings. pytest tests_py/test_main.py tests_py/test_tool_profiles.py tests_py/handlers/: all passed. Full suite pytest -q: 8658 passed, 8 skipped, 11 errors (installer/pip related, covered above). Grepped for stale "52"/"55" tool-count claims outside CHANGELOG history entries and test fixtures, which legitimately keep historical numbers; found none live.
7. Layer check. tool_registry_wiki_drafts.py imports only mcp_server.handlers.wiki_refine, tool_meta, and tool_error_handler, matching every sibling tool_registry*.py module. No boundary violation.
8. SOLID audit. No new conditionals, no interface widening beyond the two intentional new tools, no concrete-infra-in-core issue. The registry module has one responsibility: register these two tools.
9. Complexity. New module is 91 lines, no method near the 40-line cap. Craftsmanship gate is clean, including the .craftsmanship-baseline.json shrink, one entry removed, matching the test_main.py method that was split and renamed, not gamed; verified the removed entry corresponds to the renamed test method and the baseline's other ~5700 lines are untouched.
10. Security and hygiene. No new I/O surface beyond what handler_get and handler_refine already had; SQL is parameterized via psycopg placeholders in pg_store_wiki_drafts.py. No secrets. Commit is a single conventional feat PR closing one issue, no scope creep.

Issues

Blocking: none.

Non-blocking:

• PR body should read 11 errors, not 5 (see Move 0 section above); doesn't change the conclusion.
• mcp_server/handlers/wiki_refine.py lines 266-334 (handler_refine, pre-existing, not touched by this diff): a wiki_refine_draft call with only draft_id still writes synth_model and inserts a refined_llm audit memo despite no content change, because synth_model always has a non-None default. Consider filing a follow-up issue if the maintainer wants that call shape to be a true no-op now that the tool is externally callable.

Hand-offs

None needed. No root-cause fix, no structural decomposition, no formal-correctness or threat-modeling gap, and no performance claim to verify.

Verdict

APPROVE