Branch / tenant: dev/gfs.v17 (gw_v17) Path: ecf/scripts/gfs/forecast/jgfs_fcst_fsm.ecf Analysis date: 2026-06-11 Method: AWS Bedrock AgentCore agentcore-mcp-rag MCP cross-checked against the on-disk v17 worktree.

jgfs_fcst_fsm.ecf is the Forecast Segmentation Manager (FSM) task for the GFS forecast. It is a lightweight coordination job that monitors and manages the multi-segment GFS forecast execution. When the forecast is split into segments (for checkpointing and parallel post-processing), the FSM orchestrates the segment transitions, resubmitting forecast segments and coordinating with dependent tasks. It calls JGLOBAL_FSM, which drives exglobal_fsm.sh.

#PBS -S /bin/bash
#PBS -N gfs_fcst_fsm_%CYC%
#PBS -j oe
#PBS -q %QUEUE%
#PBS -A %PROJ%-%PROJENVIR%
#PBS -l walltime=06:00:00
#PBS -l select=1:ncpus=1:ompthreads=1:mem=20GB
#PBS -l place=vscatter:shared
#PBS -l debug=true

model=gfs
export cyc="%CYC%"
%include <head.h>
%include <envir-p1.h>

############################################################
# Load modules
############################################################
module load intel/${intel_ver}
module load craype/${craype_ver}

module list
############################################################
# CALL executable job script here
############################################################
export RJN="%RJN%"

export HOMEglobal=${HOMEgfs}
${HOMEgfs}/jobs/JGLOBAL_FSM
if [ $? -ne 0 ]; then
   ecflow_client --msg="***JOB ${ECF_NAME} ERROR RUNNING J-SCRIPT ***"
   ecflow_client --abort
   exit
fi

%include <tail.h>
%manual

%end

This is a single task per cycle (gfs_fcst_fsm_%CYC%). It runs alongside the main forecast task and manages segment transitions.

ecf/defs/gfs_prod.def : task jgfs_fcst_fsm
  +-- jgfs_fcst_fsm.ecf                       (PBS submit)
       +-- ${HOMEgfs}/jobs/JGLOBAL_FSM          (J-Job)
            +-- ${HOMEglobal}/scripts/exglobal_fsm.sh   (ex-script)
                 +-- monitors forecast segments
                 +-- resubmits/triggers next segment
                 +-- coordinates with ecFlow task dependencies

The FSM does not produce data products. Instead it:

1. Monitors the forecast job's progress through its segments.
2. Detects when a segment completes (by checking for restart files or sentinel files).
3. Triggers the next forecast segment submission.
4. Coordinates with ecFlow to release dependent tasks that can run on partial forecast output (e.g., post-processing for early hours while later hours are still integrating).

This enables overlap between forecast and post-processing, reducing the critical path wall-clock time.

Uses err_exit on non-zero return (EE2-correct pattern). The set -xue in the J-Job provides strict mode natively.

The J-Job does not use jjob_header.sh — it has a simplified setup:

• Directly sets EXPDIR, DATA, cycle, then sources PDY.
• No config files loaded (the FSM reads the forecast's config indirectly via exglobal_fsm.sh).

• No load_modules.sh: unlike all other v17 ECF scripts, this one loads modules directly (intel, craype) — minimal footprint for a coordination process.
• Long walltime: 6 hours matches the forecast walltime — the FSM must remain alive for the entire forecast duration.
• Low resources: 1 cpu, 20 GB — just enough to run the coordination logic.
• RJN ecFlow variable: links the FSM to its associated forecast task for ecFlow coordination.
• set -xue in J-Job: uses bash strict mode directly rather than the standard set_strict.sh sourcing pattern.
• Uses scripts/ path: calls ${HOMEglobal}/scripts/exglobal_fsm.sh rather than ${SCRglobal}/..., suggesting this is in the top-level scripts directory rather than dev/scripts.

The FSM is critical for the segmented forecast execution model. If it fails, the forecast may not advance beyond segment 0, stalling all downstream processing.

• On-disk v17 worktree: ecf/scripts/gfs/forecast/jgfs_fcst_fsm.ecf, dev/jobs/JGLOBAL_FSM.
• agentcore-mcp-rag MCP (tenant gw_v17): on-disk analysis only (FSM not in graph).