Gtm Demo Validation And Preview Gif Workflow

GTM Demo Validation and Preview GIF Workflow

Use when validating the 5 GTM demos in digitalmodel/examples/demos/gtm, especially for GitHub issue #2118 / #1809 style work.

When to use

• GTM demos need proof they actually run end-to-end
• --from-cache behavior is flaky or regressed
• reports need browser/render validation
• you need quick collateral GIFs from validated reports

Key finding

Demo 2 (demo_02_wall_thickness_multicode.py) can fail in --from-cache mode if the committed cache is a legacy JSON containing only:

• metadata
• results
• summary

Newer chart builders expect intermediate keys:

• lifecycle_phases
• min_wall_thickness
• weight_penalty

If those keys are missing, --from-cache may crash indirectly when Chart 1 falls back into engineering-code paths and PipeDefinition / related symbols were never initialized.

Verified fix pattern

In Demo 2:

1. Add a helper that checks whether the cache contains the required intermediate keys.
2. If --from-cache is requested but those keys are missing, print a warning and fall back to full recalculation.
3. In the true cache-fast path, still initialize code constants/modules needed by downstream chart builders.

Minimal pattern

CACHE_INTERMEDIATE_KEYS = (
    "lifecycle_phases",
    "min_wall_thickness",
    "weight_penalty",
)

def _cache_has_intermediate_data(cached):
    return all(cached.get(key) is not None for key in CACHE_INTERMEDIATE_KEYS)

Then in main():

• compute use_cache
• load cache
• if missing intermediates: use_cache = False
• if not use_cache: do full recalc
• if use_cache but globals/constants are needed: call _load_engineering_modules() and _init_code_constants()

Validation workflow

From digitalmodel/:

1. First run the suite

PYTHONPATH=examples/demos/gtm:src uv run pytest examples/demos/gtm/tests/test_gtm_demos.py -q

2. If Demo 2 cache path is suspect, isolate it

PYTHONPATH=examples/demos/gtm:src uv run pytest examples/demos/gtm/tests/test_gtm_demos.py -q -k wall_thickness

3. Run all 5 demos fresh

Use --force for approval-grade evidence:

PYTHONPATH=examples/demos/gtm:src uv run python examples/demos/gtm/demo_01_dnv_freespan_viv.py --force
PYTHONPATH=examples/demos/gtm:src uv run python examples/demos/gtm/demo_02_wall_thickness_multicode.py --force
PYTHONPATH=examples/demos/gtm:src uv run python examples/demos/gtm/demo_03_deepwater_mudmat_installation.py --force
PYTHONPATH=examples/demos/gtm:src uv run python examples/demos/gtm/demo_04_shallow_water_pipelay.py --force
PYTHONPATH=examples/demos/gtm:src uv run python examples/demos/gtm/demo_05_deepwater_rigid_jumper_installation.py --force

4. Verify regenerated JSON counts

Expected totals:

• Demo 1: 680
• Demo 2: 72
• Demo 3: 180
• Demo 4: 60
• Demo 5: 300

Use a small Python check against metadata.total_cases and the main data array length.

Browser validation workflow

Use a local static server rooted at the GTM folder, not repo root.

Correct server command

python -m http.server 8766 --directory /absolute/path/to/digitalmodel/examples/demos/gtm

Then open:

• http://127.0.0.1:8766/output/demo_01_freespan_report.html
• etc.

What to verify

Per report:

• title renders
• document.querySelectorAll('.chart-container').length == 5
• document.querySelectorAll('.js-plotly-plot').length == 5
• browser console has no JS errors

Hand-check pattern

Approval-grade evidence should include at least 2 simple spot checks.

Good picks that matched well

1. Demo 3 landing phase:

• formula: bearing_kpa = w_sub_kn / a_base_m2
• utilisation: bearing_kpa / bearing_limit_kpa

2. Demo 5 in-air bending:

• moment: M = w * L^2 / 8
• utilisation: bending_stress_mpa / allowable_stress_mpa

These are easy to recompute from JSON results and usually match within rounding tolerance.

Preview GIF workflow

This is useful for quick progress on #1809 before full workflow recordings exist.

Preconditions

• validated HTML reports exist
• local HTTP server is running against examples/demos/gtm
• npx playwright works
• ffmpeg is installed

Pattern

1. Use Playwright to open each report URL.
2. Scroll through the page in ~8 steps.
3. Capture PNG frames.
4. Convert frames to a compressed GIF with ffmpeg.
5. Clean up frame directories; keep GIFs only.

Output location

Store GIFs under:

examples/demos/gtm/media/

Verified naming used for interim assets

• demo_01_freespan_preview.gif
• demo_02_wall_thickness_preview.gif
• demo_03_mudmat_installation_preview.gif
• demo_04_shallow_pipelay_preview.gif
• demo_05_jumper_installation_preview.gif

These were all kept under ~1 MB using Playwright screenshots + ffmpeg palette generation.

Converting interim GIFs into issue-aligned deliverables

For issue #1809-style work, a fast next step is to normalize the preview names into the requested deliverable names rather than regenerating the 5 demo GIFs.

Verified rename/copy pattern

Create these final names under examples/demos/gtm/media/:

• demo_01_freespan.gif
• demo_02_wall_thickness.gif
• demo_03_mudmat_installation.gif
• demo_04_shallow_pipelay.gif
• demo_05_jumper_installation.gif

If the preview GIF content is already acceptable, copy/rename them first, then remove the _preview versions once references are checked. Git will usually detect these as renames if content is unchanged.

Comparison-matrix GIF pattern

The missing sixth deliverable for #1809 can be generated from:

• examples/demos/gtm/results/vessel_comparison_matrix.json
• examples/demos/gtm/results/structure_comparison_matrix.json

Reusable approach

1. Build a lightweight self-contained HTML page summarizing both matrices.
2. Use Playwright to render the page at 1024x640.
3. Capture a full-page screenshot once.
4. Slice it into viewport-height frames to simulate smooth scrolling.
5. Assemble frames into demo_comparison_matrix.gif with Pillow or ffmpeg.
6. Add a short media/README.md documenting the files and regeneration command.

Environment-specific finding

On this machine, the default python3 pointed to a 3.13 environment that did not have the Playwright Python module, while python3.12 did.

Working pattern:

python3.12 -c "from playwright.sync_api import sync_playwright; print('OK')"
python3.12 -m playwright install chromium
python3.12 examples/demos/gtm/media/generate_comparison_matrix_gif.py

If Playwright is available via npx but not in the default Python interpreter, do not assume python3 -m playwright will work. Probe python3.12 explicitly before rewriting the workflow.

Proven output set after normalization

• demo_01_freespan.gif
• demo_02_wall_thickness.gif
• demo_03_mudmat_installation.gif
• demo_04_shallow_pipelay.gif
• demo_05_jumper_installation.gif
• demo_comparison_matrix.gif
• media/README.md
• optional generator: media/generate_comparison_matrix_gif.py

Interactive Claude Code execution pattern for scoped media work

When the user explicitly wants implementation via interactive Claude Code, use tmux and pass the full issue context in the prompt rather than assuming Claude can read the parent repo issue from a nested repo.

Why

In the nested digitalmodel/ repo, gh issue view 1809 may fail or resolve against the wrong repository context. Passing the exact issue requirements in the prompt avoids this ambiguity.

Verified launch pattern

tmux new-session -d -s claude-1809 -x 160 -y 48
cd /mnt/local-analysis/workspace-hub/digitalmodel && \
claude --setting-sources user --dangerously-skip-permissions "$(cat /tmp/claude_1809_prompt.txt)"

Then handle dialogs in order:

1. Trust dialog: Enter
2. Bypass-permissions warning: Down, then Enter

Monitoring pattern

Use repeated pane captures while Claude works:

tmux capture-pane -t claude-1809 -p -S -200

This worked well for a scoped media-only task where Claude:

• copied/renamed the preview GIFs
• wrote generate_comparison_matrix_gif.py
• adapted to the local Python/Playwright mismatch
• generated demo_comparison_matrix.gif
• wrote media/README.md
• committed and pushed the work

Git hygiene

Running validation regenerates tracked HTML/JSON files. If the run is intended as evidence, commit and push the refreshed artifacts. If not, revert them before finishing.

Typical commit sequence used successfully:

1. bugfix commit for Demo 2 cache handling
2. artifact refresh commit after full validation
3. preview GIF commit

GitHub issue choreography

If a previously closed issue is contradicted by live validation, reopen it immediately with evidence. After the verified fix lands, close it again and link the validation evidence.

Effective pattern used:

• reopen regression issue (#2087)
• comment with exact failing command and error
• fix + validate
• comment with commit + green test output
• close issue
• separately update umbrella/validation issue (#2118) and tracker (#2016)

Pitfalls

• Serving from the repo root causes report URLs to 404 or point at the wrong directory.
• Legacy Demo 2 JSON may look valid because counts are present, but still be unusable for chart-only regeneration.
• Full validation runs modify tracked output artifacts; do not leave the nested repo dirty accidentally.
• Preview GIFs from reports are useful progress, but they are not automatically equivalent to a stricter “NL prompt -> agent navigation -> run -> report interaction” screencast spec.

Exit condition

This workflow is complete when:

• GTM pytest suite is green
• all 5 demos were rerun with --force
• reports render in browser with 5 Plotly graphs each and no JS errors
• at least 2 hand-checks are documented
• refreshed artifacts are committed/pushed
• any interim GIF assets are committed/pushed if needed for GTM collateral