Skip to content
Back to skills

Hecras Plan Execution Gpt Cmdr Ras Commander

ASecurity

Decision support for HEC-RAS execution strategy. Analyzes project inspector output to recommend which plans to run, execution mode, and optimal parameters. Provides structured execution plans with dependency ordering, prerequisite validation, and resource-aware parameter recommendations. Use when planning execution strategy, selecting execution mode, determining plan order, validating prerequisites, or optimizing parallel execution parameters. Triggers: plan execution, execution strategy, whi...

  • 666 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added February 9, 2026
documentationpythondebuggingapidocumentation

Works with

  • api

Security analysis

A100/100

Scanned February 12, 2026

npx -y skills add majiayu000/claude-skill-registry --skill hecras-plan-execution-gpt-cmdr-ras-commander --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Hecras Plan Execution Gpt Cmdr Ras Commander?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Hecras Plan Execution Gpt Cmdr Ras Commander
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/majiayu000-hecras-plan-execution-gpt-cmdr-ras-commander/badge)](https://www.skillsdirectory.com/skills/majiayu000-hecras-plan-execution-gpt-cmdr-ras-commander)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: hecras_plan_execution
description: |
  Decision support for HEC-RAS execution strategy. Analyzes project inspector output
  to recommend which plans to run, execution mode, and optimal parameters. Provides
  structured execution plans with dependency ordering, prerequisite validation, and
  resource-aware parameter recommendations.
  Use when planning execution strategy, selecting execution mode, determining plan
  order, validating prerequisites, or optimizing parallel execution parameters.
  Triggers: plan execution, execution strategy, which plans to run, execution order,
  parallel vs sequential, compute_plan, compute_parallel, mode selection, prerequisites,
  dependencies, runnable plans, blocked plans, execution plan, batch strategy.
---

# Planning HEC-RAS Execution

Decision support for selecting optimal HEC-RAS execution strategy. Analyzes Project Inspector output and generates actionable execution plans.

## Primary Sources

**DO NOT duplicate content from primary sources. This skill is a lightweight navigator.**

- **`.claude/rules/hec-ras/execution.md`** - Complete execution mode documentation
- **`.claude/agents/hecras-project-inspector.md`** - Project Intelligence Report schema
- **`.claude/skills/hecras_compute_plans/SKILL.md`** - Mode Selection Guide
- **`.claude/skills/hecras_compute_rascontrol/SKILL.md`** - Legacy HEC-RAS (< 6.0)

---

## Mode Selection Decision Tree

```
1. Is HEC-RAS version < 6.0?
   YES -> Use RasControl (hecras_compute_rascontrol skill)
   NO  -> Continue

2. How many runnable plans?
   ZERO    -> Report blockers, stop
   ONE     -> compute_plan() [+ stream_callback if monitoring needed]
   MULTIPLE -> Continue to step 3

3. Multiple Plan Decision:
   Debugging/diagnosing? -> compute_test_mode()
   Have remote workers?  -> compute_parallel_remote()
   Production/batch?     -> compute_parallel()
```

### Mode Selection Matrix

| Context | Plans | Mode | Rationale |
|---------|-------|------|-----------|
| Debugging | Any | `compute_test_mode()` | Sequential, single folder |
| Single plan, monitoring | 1 | `compute_plan()` + callback | Real-time output |
| Single plan, quick | 1 | `compute_plan()` | Simplest API |
| Multiple plans, production | 2+ | `compute_parallel()` | Fastest throughput |
| Distributed infrastructure | 2+ | `compute_parallel_remote()` | Scale-out |
| HEC-RAS < 6.0 | Any | `RasControl` | COM-based required |
| Resume interrupted | 2+ | `compute_parallel(skip_existing=True)` | Skips completed |

---

## Input: Project Intelligence Report

Expected input from Project Inspector Agent:

```markdown
## Plans Analysis
| Plan | Title | Status | Geometry | Flow | Issues |
|------|-------|--------|----------|------|--------|
| 01 | Baseline | Runnable | g01 | u01 | None |
| 02 | Future | Missing geom | g03 (MISSING) | u02 | Geometry file not found |

### Execution Readiness
- **Ready to Run**: 01, 03
- **Blocked**: 02 (missing geometry g03)
```

---

## Output: Execution Plan

Generate structured execution plans in this format:

```markdown
# Execution Plan

## Recommended Mode
`compute_parallel()` - Multiple runnable plans, production context

## Execution Order
1. Plan 01 (Baseline) - Ready
2. Plan 03 (Modified) - Ready
   [Skip: Plan 02 - blocked by missing geometry g03]

## Parameters
```python
RasCmdr.compute_parallel(
    plans_to_run=["01", "03"],
    max_workers=2,
    num_cores=4,
    force_rerun=False,
    verify=True
)
```

## Prerequisites
- None for runnable plans

## Blockers (Cannot Run)
| Plan | Issue | Resolution |
|------|-------|------------|
| 02 | Missing geometry g03 | Create geometry file or fix reference |

## Warnings
- Plan 02 blocked until g03 is created
```

---

## Parameter Recommendation

### Worker/Core Allocation

| System Cores | Model Type | `max_workers` | `num_cores` |
|--------------|------------|---------------|-------------|
| 4-8 | 1D | 2-4 | 2 |
| 8-16 | 2D | 2-4 | 4 |
| 16+ | Mixed | 4-6 | 4 |

**Formula**: `max_workers = min(num_plans, total_cores // num_cores)`

### Geometry Flags

| Scenario | `clear_geompre` | `force_geompre` |
|----------|-----------------|-----------------|
| Geometry unchanged | `False` | `False` |
| Minor edits (Manning's n) | `True` | `False` |
| Major changes (new mesh) | `False` | `True` |
| Geometry corruption | `False` | `True` |

### Skip/Rerun Flags

| Scenario | `force_rerun` | `skip_existing` |
|----------|---------------|-----------------|
| Normal (smart skip) | `False` | `False` |
| Force fresh results | `True` | `False` |
| Resume interrupted | `False` | `True` |

---

## Dependency Analysis

### Check for Blockers

```python
def get_runnable_plans(plan_df):
    """Identify which plans can execute."""
    runnable, blocked = [], {}

    for _, plan in plan_df.iterrows():
        plan_num = plan['plan_number']
        geom_ok = Path(plan['Geom Path']).exists() if plan.get('Geom Path') else True
        flow_ok = Path(plan['Flow Path']).exists() if plan.get('Flow Path') else True

        if geom_ok and flow_ok:
            runnable.append(plan_num)
        else:
            blocked[plan_num] = "Missing geometry" if not geom_ok else "Missing flow"

    return runnable, blocked
```

### Dependencies to Consider

1. **Shared Geometry** - Plans using same geometry can run in parallel
2. **Upstream/Downstream** - Cascade models need sequential execution
3. **DSS Dependencies** - Plans reading same DSS may conflict

---

## Context-Specific Templates

### Debugging Context

```python
RasCmdr.compute_test_mode(
    plans_to_run=["01", "02", "03"],
    dest_folder_suffix="[Debug]",
    num_cores=2,
    verify=True
)
```

### Production Context

```python
RasCmdr.compute_parallel(
    plans_to_run=["01", "02", "03"],
    max_workers=4,
    num_cores=4,
    verify=True
)
```

### Monitoring Context

```python
from ras_commander.callbacks import ConsoleCallback

RasCmdr.compute_plan(
    "01",
    stream_callback=ConsoleCallback(verbose=True),
    verify=True
)
```

---

## Common Blockers

| Blocker | Resolution |
|---------|------------|
| Missing geometry file | Create geometry or fix .prj reference |
| Missing flow file | Create flow file or fix reference |
| Missing DSS file | Download/create DSS or update path |
| Geometry HDF outdated | Run with `force_geompre=True` |
| Results already current | Use `force_rerun=True` if needed |
| HEC-RAS < 6.0 | Use RasControl skill instead |

---

## Integration

### Upstream Skills
- **hecras-project-inspector** - Generates intelligence report
- **dss_read_boundary-data** - Validates boundary conditions
- **qa_repair_geometry** - Fixes geometry blockers

### Downstream Skills
- **hecras_compute_plans** - Runs the execution plan
- **hecras_compute_remote** - Distributed execution
- **hecras_extract_results** - Post-execution analysis
- **hecras_parse_compute-messages** - Interpret execution results

---

## See Also

- **`.claude/rules/hec-ras/execution.md`** - Complete parameter documentation
- **`.claude/agents/hecras-project-inspector.md`** - Input format details
- **`examples/113_parallel_execution.ipynb`** - Working parallel example

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.