_{Analysis by GitHub Copilot (Claude Opus 4.8, VS Code), directed and reviewed by Terry.McGuinness. Implementation work described here is by Anton Fernando ([email protected]).}

Report date: 2026-09-22.

The branch feature/ecflow-c48-atm turns the forecast-only ecFlow scaffold — documented in the companion page GFS-Forecast-Only-ecFlow-and-Rocoto-Report — into a working, loadable ecFlow suite for the smallest atmosphere-only CI case, C48_ATM. Where the September 2025 scaffold left get_cycledefs() and write() as pass, this branch supplies a Rocoto-parallel task-class hierarchy, a shared .ecf script + header library, and an end-to-end case loader that stands the suite up on a running ecFlow server.

This is the first concrete implementation increment of NOAA-EMC epic #4106 "Dynamically Generate EcFlow Suites" (author AnnaSmoot-NOAA, open), whose stated goal is to remove the situation where "Rocoto is the only workflow engine option for developers and the EcFlow engine is hard coded for production."

All implementation on this branch is by Anton Fernando. The work is nine non-merge commits authored 2026-09-17 → 2026-09-21, adding 21 files and ~1,751 lines under dev/workflow/ecflow/ and dev/ecflow/.

• Inspected checkout: supported_repos/global-workflow_forked on branch feature/ecflow-c48-atm, tracking anton/feature/ecflow-c48-atm (https://github.com/AntonMFernando-NOAA/global-workflow.git).
• Merge-base with noaa-emc/develop: 5e6a11c60. The branch is 37 commits ahead, of which 9 are branch-authored and 28 are pass-through merges bringing develop forward. This report describes only the 9 authored commits; the develop merges are out of scope.
• Line/file counts are restricted to the two directories the branch owns (dev/workflow/ecflow/, dev/ecflow/) via git diff --numstat noaa-emc/develop...HEAD -- <dirs>.
• No source files were modified during this analysis; findings are read-only.
• EIB MCP graph/doc queries were issued against tenant gw, branch develop. Direct source and Git history are authoritative for provenance and implementation claims; MCP results are supporting context.

The design mirrors the existing dev/workflow/rocoto/ package so the two engines stay symmetrical and share the resource / forecast-hour / system plumbing. The only intended divergence is the output shape: dictionaries with ecFlow trigger strings instead of Rocoto XML fragments.

Nine task scripts under dev/ecflow/scripts/ — stage_ic, fcst, atmos_prod, tracker, genesis, metp, arch_vrfy, arch_tars, cleanup. Each is a thin ecFlow wrapper that %includes the shared headers, exports the ecFlow-substituted variables (%HOMEglobal%, %EXPDIR%, %COMROOT%, …), sources dev/ush/load_modules.sh run, invokes the existing J-Job (e.g. dev/jobs/JGLOBAL_FCST), and signals completion via tail.h.

Shared headers under dev/ecflow/utils/:

Notable fix (a91692b53): atmos_prod.ecf loops over a comma-separated %FHR_LIST% (one script serving a fanned-out forecast-hour family). The printf format had to be escaped as %%03d because ecFlow performs its own %VAR% substitution on the script before Slurm ever runs it — an unescaped %03d would be eaten by the ecFlow pre-processor.

• dev/ctests/cases/C48_S2SW* YAML and dev/ctests/CMakeLists.txt updated so the CI harness recognizes ecFlow-loaded cases.
• dev/workflow/setup_buildxml.py and ecflow_suite_factory.py updated so setup_workflow.py --engine ecflow yields a valid GFS forecast-only .def.
• dev/workflow/hosts/gaeac6.yaml one-line default the suite needs.
• .gitlab-ci.yml change to exercise the C48_ATM ecFlow load in CI.
• The last two commits add --pslot, --expdir, --comroot, and --stmp CLI overrides to load_ecflow_case.py, so multiple developers can run C48_ATM side-by-side on a shared box (URSA / Gaea-c6) without editing YAML.

flowchart TD
    A[AppConfig<br/>experiment config] --> B[EcFlowTasks / GFSEcFlowTasks<br/>+ ecflow_tasks_factory]
    B -->|task dicts w/ triggers| C[GFSForecastOnlyEcFlowSuite<br/>gfs_forecast_only_ecflow.py]
    C -->|ecFlow Python API| D[.def suite definition]
    D --> E[load_ecflow_case.py<br/>load + begin]
    E --> F[(ecFlow server<br/>ECF_HOST:ECF_PORT)]
    G[dev/ecflow/scripts/*.ecf<br/>+ utils/*.h] -->|ECF_FILES via manifest| F
    H[sync_ecf_scripts.sh] -.hot-refresh.-> G

The equivalent Rocoto path is AppConfig → rocoto.tasks + tasks_factory → Rocoto XML. The branch deliberately keeps the two engines symmetric so shared resource/config logic lives in rocoto.tasks.Tasks (the ecFlow base inherits from it).

# On URSA / gaeac6, with an ecflow_server already up and ECF_HOST/ECF_PORT set:
python3 dev/workflow/ecflow/c48_atm_ecflow.py \
    --pslot my_c48_atm --expdir "$STMP/EXPDIR" --overwrite

This runs setup_expt, setup_workflow (ecflow engine), copies the .ecf files into EXPDIR/…/ecf_scripts/, writes ecf_scripts.manifest (later used by sync_ecf_scripts.sh), loads the .def into the server, and begins the suite.

• gefs, sfs, gcafs task classes are registered-but-commented — reserved for later increments of epic #4106.
• No cycled DA support — forecast-only only.
• No wave / ocean / ice tasks in the C48_ATM DAG (those remain exercised on the Rocoto side by the C48_S2SW cases).

The companion report GFS-Forecast-Only-ecFlow-and-Rocoto-Report documents that, as of the June 2026 checkout, GFSForecastOnlyEcFlowSuite.get_cycledefs() and .write() were both pass — an extension point with zero of two generation hooks implemented. This branch is the direct sequel: it implements those hooks (the file grows +385/−19), adds the task-dict producers they consume, and supplies the .ecf payloads the generated .def references. Reading the two reports together gives the full arc from scaffold (Rahul Mahajan, Sep 2025) to first working case (Anton Fernando, Sep 2026).

Agent: GitHub Copilot running Claude Opus 4.8 (VS Code). The agent selected and issued all tool calls; Terry.McGuinness directed the investigation and reviewed results. Server: eib-mcp-localhost (Docker MCP Gateway via pw tunnel, tenant gw, branch develop). Git history and direct source were used as the authoritative cross-check.

The inventory below is complete: every OCA MCP call issued while drafting this report is listed. Per the originating request it carries an evaluation rating (Usefulness), a separate Assessment column explaining that rating, and an inline Recommendation to Improve column for the calls that under-performed.

Tally: 5 calls — 2 high-value (search_issues #2 and #5), 2 medium, 1 low, plus 1 cross-cutting branch-staleness caveat. As the Assessment column shows, the graph/doc tools mostly corroborated structure the source already exposed, while the issue search delivered the context Git could not — the epic that frames the whole branch. Net: the OCA MCP was most valuable for provenance and strategic framing (why this branch exists), and least valuable for code-internal detail (which direct source read better).

• dev/workflow/ecflow/ecflow_tasks.py — EcFlowTasks base and task-dict contract
• dev/workflow/ecflow/gfs_ecflow_tasks.py — per-task GFS definitions (best entry point for adding a task)
• dev/workflow/ecflow/ecflow_tasks_factory.py — engine registration
• dev/workflow/ecflow/gfs_forecast_only_ecflow.py — forecast-only suite generator (the filled-in scaffold)
• dev/workflow/ecflow/load_ecflow_case.py — generic case loader
• dev/workflow/ecflow/c48_atm_ecflow.py — C48_ATM entry point
• dev/ecflow/scripts/*.ecf, dev/ecflow/utils/*.h — script + header library
• NOAA-EMC #4106 — umbrella epic
• Companion: GFS-Forecast-Only-ecFlow-and-Rocoto-Report