New skill proposal: dfam-check — printability rules for meshes

[knewnothing-git]

Hi maintainers,

I'd like to propose adding a dfam-check skill that fills a gap between $cad (generates meshes) and $gcode (slices them): checking whether the mesh is actually printable for the target AM process before it gets sliced.

What it does

Measures mesh files (.stl, .obj, .ply, .3mf) against DfAM rules from ISO/ASTM 52910 and published process design guides:

Overhang angle histogram (build-plate contact excluded)
Wall thickness field via ray-cast sampling
Support-volume upper-bound estimate
Six axis-aligned orientation candidates ranked by support area and build height

Reports facts only — the workflow does the ✅/❌/❓ comparisons against a limits table per process (FDM, SLS, SLA/DLP, metal PBF, MJF), same pattern as $sendcutsend.

Why it fits

Today an agent can go $cad → $gcode → print with no printability gate. $sendcutsend is the sheet-metal preflight; there's no equivalent for FDM/SLS/SLA/metal AM before slicing. dfam-check would be that gate, and would hand off to $cad for regeneration and $cad-viewer for visual inspection.

Status

Working prototype, tested on three fixtures (thin-walled box, cantilevered bracket, solid block) — measurements match the planted defects exactly (0.6mm walls flagged, 1200mm² of 0° overhangs located, 226% support ratio on the bracket).

Structure matches the $sendcutsend / $gcode convention: SKILL.md + references/process-limits.md + scripts/dfam_tool.py + agents/openai.yaml + MIT LICENSE.

Would this fit the roadmap? Happy to open a PR against develop if there's interest — CONTRIBUTING.md is clear on the flow.

Thanks!

[earthtojake]

Hey, yes I've been wanting to add a skill for ISO/ASTM 52910! Please proceed. How do you plan on storing and reading the PDF since it's very large?

[knewnothing-git]

No PDF storage needed — the skill doesn't parse the standard at all.

The limits are encoded as deterministic rules, and each finding cites the clause as a reference string next to an independently measured value. So there's nothing large to read or ship.

That also avoids a licensing problem: ISO/ASTM 52910 is copyrighted and sold by ASTM, so bundling the document would be redistribution. Citing clause numbers alongside our own measurements isn't.

What it does:

Runs entirely locally — no API key, no network call, no service dependency. Only trimesh, numpy, rtree.
scripts/dfam_tool.py is fact-only: it measures and never emits pass/fail. Verdicts belong to the workflow, so the thresholds stay inspectable rather than buried in the tool.
Measures ray-cast wall thickness (shape-diameter function, 9-ray cone per sample), face-angle overhang maps, support volume estimate, and a 6-orientation ranking.
Per-process limits for FDM, SLA/DLP, SLS, metal PBF, and MJF in references/process-limits.md, so they're readable and editable without touching code.
Accepts .stl, .obj, .ply, .3mf. Z-up build convention.
Sits naturally before $gcode — check printability, then slice.

On validation: the engine is tested against a 100-part parametric rig built with build123d, covering 33 features and 8 surface types, where each part declares its expected verdict. Currently 100/100, 0 mismatches. It's caught real bugs — a single-ray thickness probe was producing false failures on high-genus meshes, which is why it now uses the cone method.

Happy to open a PR. Would you prefer the test rig in the same PR, or kept separate to keep the diff reviewable?

[knewnothing-git]

No PDF storage needed — the skill doesn't parse the standard at all. The limits are encoded as deterministic rules, and each finding cites the clause as a reference string next to an independently measured value. So there's nothing large to read or ship. That also avoids a licensing problem: ISO/ASTM 52910 is copyrighted and sold by ASTM, so bundling the document would be redistribution. Citing clause numbers alongside our own measurements isn't. What it does: Runs entirely locally — no API key, no network call, no service dependency. Only trimesh, numpy, rtree. scripts/dfam_tool.py is fact-only: it measures and never emits pass/fail. Verdicts belong to the workflow, so the thresholds stay inspectable rather than buried in the tool. Measures ray-cast wall thickness (shape-diameter function, 9-ray cone per sample), face-angle overhang maps, support volume estimate, and a 6-orientation ranking. Per-process limits for FDM, SLA/DLP, SLS, metal PBF, and MJF in references/process-limits.md, so they're readable and editable without touching code. Accepts .stl, .obj, .ply, .3mf. Z-up build convention. Sits naturally before $gcode — check printability, then slice. On validation: the engine is tested against a 100-part parametric rig built with build123d, covering 33 features and 8 surface types, where each part declares its expected verdict. Currently 100/100, 0 mismatches. It's caught real bugs — a single-ray thickness probe was producing false failures on high-genus meshes, which is why it now uses the cone method. Happy to open a PR. Would you prefer the test rig in the same PR, or kept separate to keep the diff reviewable?

…

On Wed, Aug 12, 2026 at 6:50 AM earthtojake ***@***.***> wrote: Hey, yes I've been wanting to add a skill for ISO/ASTM 52910! Please proceed. How do you plan on storing and reading the PDF since it's very large? — Reply to this email directly, view it on GitHub <#188?email_source=notifications&email_token=BSLL7SQLWZVSD7VI4ABWOTD5JPA7LA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZZHAZDOMRYUZZGKYLTN5XKM3LBNZ2WC3FFMV3GK3TUVRTG633UMVZF6Y3MNFRWW#discussioncomment-17982728>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BSLL7SWWWNW2XG6IW5XFZKL5JPA7LAVCNFSNUABJKJSXA33TNF2G64TZHMYTEMJXGUZDEMBWG45UI2LTMN2XG43JN5XDWMJQGUZTSNRSHCQXMAQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.*** .com>

[earthtojake]

Yep put it all in a single PR

How many rules are there? Curious to see how it looks!

[knewnothing-git]

26 numeric limits — 6 categories across 5 processes (FDM/FFF, SLS, SLA/DLP, PBF-LB metal, MJF), with 4 n/a cells where powder processes make the rule meaningless.

The categories: min supported wall, min unsupported wall, self-supporting angle, min hole diameter, min positive feature, max unsupported bridge.

They live as a markdown table in references/process-limits.md, not in code. dfam_tool.py only measures and never compares — it emits facts, and SKILL.md does the comparison. That way a user with a machine- or material-specific datasheet can override a limit by editing one cell, and the workflow cites whichever source was actually used rather than a hardcoded default.

The tool has 5 measurement families: mesh integrity (watertight, euler, body count), overhang maps by face angle, ray-cast wall thickness, support volume estimate, and a 6-orientation ranking. Two subcommands, measure and orientations.

Limits are conservative defaults from published process design guides — Hubs, Formlabs, HP MJF, EOS — with the category structure following ISO/ASTM 52910 §6 (feature limits §6.5, supports §6.7).

Will put the test rig in the same PR as you asked.