Regional Ocean Debugger

MCP server for debugging regional ocean (MOM6 + CESM) model runs.

Combines practical log analysis tools with a searchable knowledge base of expert interview transcripts, giving an LLM agent both the ability to inspect model output files and access to domain-expert guidance.

Tools

Tool	Purpose
read_run_log(case_dir, component, lines)	Read recent CESM/MOM6 run log lines
find_errors(case_dir)	Scan all logs for FATAL/NaN/CFL/etc.
classify_error(error_text)	Match error against known MOM6 failure modes
read_mom6_params(case_dir, filter)	Read MOM_input + MOM_override
diff_params(case_dir_a, case_dir_b)	Compare MOM_input between two cases
check_cfl(case_dir)	Evaluate CFL stability from MOM_input timestep settings
suggest_timestep(resolution_deg)	Recommend DT and DT_THERM for a given resolution
read_diag_table(case_dir)	Parse enabled MOM6 diagnostics
suggest_diagnostics(symptoms)	Suggest useful diagnostics for a given problem
query_domain_knowledge(question)	Search expert interview knowledge base
get_parameter_advice(param_name)	Look up guidance for a specific MOM6 parameter

Resources

URI	Content
ocean://case/{case_dir}/params	Parsed MOM_input as KEY=VALUE
ocean://case/{case_dir}/errors	Extracted errors from recent logs
ocean://knowledge/index	Summary of knowledge base contents

Knowledge Base

Domain knowledge comes from expert interviews on regional MOM6 modelling (private repo: AidanJanney/RegionalMOM6_InterviewTranscripts). Topics include: OBC setup, tidal forcing, bathymetry, stability, BGC, grid design, and regional ocean dynamics.

The transcripts are a git submodule in data/transcripts/. After cloning this repo:

git submodule update --init data/transcripts

Setup

pip install -e .
# with the transcripts submodule:
git submodule update --init data/transcripts

Running

regional-ocean-debugger-mcp
# or
python server.py

Add to .mcp.json:

{
  "mcpServers": {
    "ocean-debugger": {
      "command": "python",
      "args": ["/path/to/MCP_regional_ocean_debugger/server.py"]
    }
  }
}

Design

• Stateless: reads from the filesystem per call
• No ML dependencies: knowledge search is keyword-based (fast, transparent, works without GPU)
• Composable: use find_errors → classify_error → query_domain_knowledge in sequence for triage
• Private data: transcripts submodule is private; do not make public without authorisation

CESM / MOM6 paths on Derecho

Cases: ~/croc_cases/
MOM_input: <case_dir>/run/MOM_input or <case_dir>/Buildconf/momconf/MOM_input
diag_table: <case_dir>/run/diag_table