[briangreunke]

Problem

ghostwriter-readonly pinned gql[aiohttp,websockets]>=3.0,<4.0. gql 3 requires websockets<12, and fastmcp 3.2+ requires websockets>=15.0.1, so uv run resolved the MCP server to fastmcp 2.12.5 and mcp 1.16.0. Both versions have published security advisories. mcp 1.16.0 also fails at import on current pydantic:

ImportError: cannot import name 'eval_type_backport' from 'pydantic._internal._typing_extra'

This applies the same pin change as #175 (mythic-c2-readonly).

Changes

• mcp/server.py, mcp/test_server.py: require fastmcp>=3.2.0 and gql[aiohttp,websockets]>=4.0.
• skills/ghostwriter-readonly/scripts/ghostwriter_read.py: require gql[aiohttp,websockets]>=4.0.
• _run_query (server) and run_query (skill script): set variable_values on the gql 4 request object instead of passing the deprecated variable_values= keyword to session.execute.
• mcp/test_server.py: replace the removed FastMCP.get_tools() with list_tools(). Both tests still expect 22 tools.
• mcp/test_server.py: add a test that runs _run_query against a local aiohttp stub. It checks that {"id": 7} reaches the request body as variables, that the stub's data is returned, and that no variable_values deprecation warning is raised.
• capability.yaml: 1.0.0 → 1.0.1.

Evidence

Base 9814433e93cb4850b3b570077f3ca04b1fb8e258, head 7c26724a697aca4121ac35496d30548712d6060e. Python 3.13, linux x86_64.

• uv tree --script on mcp/server.py and mcp/test_server.py: fastmcp v4.1.2, mcp v2.3.0, gql v4.4.0. On ghostwriter_read.py: gql v4.4.0.
• uv run --python 3.13 --script capabilities/ghostwriter-readonly/mcp/test_server.py: 27 passed. The new stub test fails when _run_query is reverted to the variable_values= keyword.
• uv run --python 3.13 --script .../scripts/ghostwriter_read.py --help: exit 0.
• git grep -n 'variable_values=' -- <server.py> <ghostwriter_read.py>: no matches.
• just validate strict=true: 26 validated, 0 failed (6 warnings for missing local binaries, unrelated).
• pre-commit run --files <changed>: all hooks pass except ruff-format, which already fails on server.py and test_server.py at base. Reformatting would touch about 140 unrelated lines, so it is left out of this PR. CI does not run ruff-format.

Not tested against a live GhostWriter instance. ghostwriter_read.py has no automated test; it uses the same request pattern as the server.

🤖 Generated with Claude Code

[briangreunke]

Automated agent review (not a human approval)

Verdict: APPROVE, at head 7c26724a697aca4121ac35496d30548712d6060e against base 9814433e93cb4850b3b570077f3ca04b1fb8e258. No blocking findings.

Checked on Python 3.13, linux x86_64:

• uv tree --script on mcp/server.py and mcp/test_server.py: fastmcp v4.1.2, mcp v2.3.0, gql v4.4.0. ghostwriter_read.py: gql v4.4.0.
• mcp/test_server.py: 27 passed. Restoring the variable_values= keyword makes the new stub test fail.
• Server started over stdio against a local GraphQL stub, listed 22 tools, and get_finding(id=7) sent "variables": {"id": 7}.
• ghostwriter_read.py clients --limit 3 against a stub with PYTHONWARNINGS=always: exit 0, variables sent, no deprecation warning.
• just validate strict=true: 0 failed.
• CI: all checks pass.

Non-blocking: the stub test has two overlapping warning assertions.