Files
kandev/docs/plans/external-question-answering/task-08-external-tool-catalog.md
T
2026-08-24 11:09:03 +00:00

4.4 KiB

id, title, status, wave, depends_on, plan, spec
id title status wave depends_on plan spec
08-external-tool-catalog External MCP tool catalog: add the two tools and close the drift done 1
plan.md ../../specs/integrations/requirements/external-question-answering.md

Task 08: External MCP tool catalog — add the two tools and close the drift

The settings page under-advertises the external endpoint by five tools, and its own drift note is itself stale. Add the two new tools, add the five missing ones, and delete the note rather than renumbering it.

  • Acceptance:

    1. The catalog lists 37 tools: the current 30, plus list_repositories_kandev, import_workflow_kandev, get_task_conversation_kandev, add_task_dependency_kandev, remove_task_dependency_kandev, plus list_pending_questions_kandev and answer_question_kandev in a new group about answering agent questions (C1, C2).
    2. The KNOWN DRIFT comment in the catalog and the corresponding NOTE in its test are deleted, not rewritten to a new number.
    3. Every new title and description key resolves in en/settings.json, and en and pseudo have identical key sets (C3).
  • Verification:

    cd apps && pnpm install --frozen-lockfile && \
      pnpm --filter @kandev/web test -- lib/settings/external-mcp-tools.test.ts
    cd apps/web && pnpm run i18n:pseudo && pnpm run i18n:check && \
      pnpm run i18n:ratchet && pnpm run typecheck && pnpm run lint
    

    The tool-count test pins the catalog, so it passes on its own. The backend's 37 only becomes true when task 05 lands; task 05's checks re-run this test for that reason.

  • Files likely touched:

    • apps/web/lib/settings/external-mcp-tools.ts (7 new entries, one new group, drift comment deleted)
    • apps/web/lib/settings/external-mcp-tools.test.ts (pinned count 30 → 37; NOTE deleted)
    • apps/web/src/locales/en/settings.json (new group title/description keys and 7 tool description keys)
    • apps/web/src/locales/pseudo/settings.json (regenerated, not hand-edited)
  • Dependencies: None.

  • Parallelism: parallel-safe with tasks 01 and 06 — disjoint files.

  • Inputs:

    • Spec C1, C2, C3, and § Verification notes → Tests that need deliberate updating.
    • Plan § Frontend and § Tests → Catalog.
    • Descriptions must be checked against the backend's own tool wording (the reason the drift was left alone before): read the mcp.WithDescription(...) text in apps/backend/internal/mcp/server/server.go for each of the five existing tools and the two new ones, and write the user-facing copy from that, not from the tool name.
    • Root AGENTS.md § Internationalization: no t() at module scope (the catalog is evaluated at import, which is why it stores keys), no Unicode em dash in any locale value, and pseudo is generated by pnpm run i18n:pseudo rather than edited.
    • Real locales (pt-pt, zh-*) only warn on parity and are translated out of band — do not hand-translate them here.
  • Output contract: summary, files changed, tests run with counts, blockers, risks, and the task/plan status update in the same conversation.

Results

Implemented in commit 31dcd3dff ("feat(web): close external MCP catalog drift, add question-answering tools"). Added list_repositories_kandev, import_workflow_kandev, get_task_conversation_kandev, add_task_dependency_kandev, and remove_task_dependency_kandev to their matching existing/new groups, added a new "Answering agent questions" group for list_pending_questions_kandev and answer_question_kandev, deleted the KNOWN DRIFT comment block (not renumbered), and added the 11 new en/settings.json keys plus the regenerated pseudo/settings.json. Pinned count moved 30 → 37; the stale NOTE in the test file was deleted.

Verified in this session: pnpm --filter @kandev/web test -- lib/settings/external-mcp-tools.test.ts passes (tool count 37, uniqueness, create_task_kandev inclusion, session-scoped-tool exclusion, non-empty groups, every key resolving in en/settings.json, every interpolated description supplied). pnpm run i18n:pseudo, pnpm run i18n:check, pnpm run i18n:ratchet, pnpm run typecheck, and pnpm run lint from apps/web are all clean.

No external side effects. The tool-count pin was 37 as soon as this commit landed since it only pins the catalog; the backend's own count reached 37 once task 05 landed, and task 05 re-ran this same catalog test to confirm.