v0.51.28 — Release E2 (MCP server Option A rewrite + WebUI /goal command)

Two-PR contributor batch shipping the MCP server rewrite and the new /goal command.

Highlights

/goal command — track multi-turn goals with budget enforcement

New /goal slash command lets you set a multi-turn objective and track progress automatically:

/goal "build a counter app in react" budget=15
/goal status
/goal pause
/goal resume
/goal clear

The agent emits goal and goal_continue SSE events after each completed turn, so the UI can show progress and queue continuation prompts when the goal isn't finished. SQLite-backed state per profile, fully integrated with the streaming chat path.

By @Michaelyklam — closes #1808.

MCP server Option A rewrite

mcp_server.py is now a clean 567-LOC implementation that imports api.models and api.profiles canonically rather than carrying duplicate slug-matching helpers. The _profiles_match helper now lives in api/profiles.py as the single source of truth (mcp_server.py and api/routes.py both import the canonical version — re-introducing a local copy in either trips a parity test immediately).

Notable improvements:

• Env-aware WEBUI_URL (HERMES_WEBUI_HOST / HERMES_WEBUI_PORT)
• Data-loss safety: delete_project now refuses to touch session JSONs when HERMES_WEBUI_PASSWORD is unset, returning {ok: true, unassigned_sessions: 0, warning: "…"} instead. Prevents data corruption when an MCP client tries to delete a project on an unauthenticated server.
• 53-test coverage in tests/test_mcp_server.py including HTTP wire-format integration tests, profile-scoped isolation, legacy untagged row visibility, and --profile foo CLI ordering regression.

By @samuelgudi — closes #1616.

Test isolation infrastructure

The MCP test suite addition exposed three separate sys.modules pollution bugs in the broader test infrastructure. All three are fixed in this release:

1. mock.patch.dict(sys.modules, ...) deletes lazy submodules — tests/test_issue1857_usage_overwrite.py was using patch.dict, which on __exit__ deletes any keys added during the patched scope. Lazily-imported pydantic submodules (e.g. pydantic.root_model) were silently evicted, producing KeyError: 'pydantic.root_model' in unrelated downstream tests. Replaced with manual save/restore using a sentinel.

2. Module-attribute restoration — tests/test_mcp_server.py mutated module-level constants on api.config/api.models/mcp_server without restoring; downstream tests read deleted tmpdirs. Now snapshot/restore on first _reimport_mcp() call.

3. Parent-package attribute leak — test_profiles_match_single_source_of_truth re-imports api.routes/api.profiles for the canonical-helper identity check. When restoring sys.modules only, fresh modules still leaked through because import api.routes as r resolves via sys.modules['api'].routes (parent-package attribute), NOT directly via sys.modules['api.routes']. Fix: also restore parent-package attributes.

CI updates

Added pytest-asyncio to CI deps and best-effort mcp install. tests/test_mcp_server.py uses pytest.importorskip as a safety net so the matrix stays green even if the optional mcp package fails to install on a future Python version.

Verification

• Tests: 4898 → 4947 passing (+49 net new)
• Browser API harness (port 8789): all 11 endpoints + 20 QA tests PASS
• node -c on all modified static/*.js files: clean
• Opus advisor pre-release verdict: SHIP-READY
• v0.51.27 carry-overs preserved: _strip_workspace_prefix, on_interim_assistant, _max_iterations_cfg, input-token guard

Follow-up items filed

• #1932 — Goal hook fires on every assistant turn including unrelated user messages
• #1933 — Goal command runtime status strings need i18n keys

Full diff

16 files, +2692/-105.

Full Changelog: v0.51.27...v0.51.28