3dprinter — Flashforge AD5X print pipeline

Headless slicing + printing for a Flashforge AD5X (220³ mm, 1 toolhead, 4-channel material station) using OrcaSlicer in CLI mode and the flashforge-python-api library over the printer's LAN API.

Configuration Variables

Set these before running the pipeline, or answer the install prompt and let your agent configure them:

Trigger conditions

0. User wants an object designed (sketch / annotated photo / spoken dimensions) → that is the DESIGN stage: sibling skill openscad-cad (parametric OpenSCAD → STL). Return here once an STL exists.
1. User sends an STL / 3MF / OBJ file → run the slice workflow.
2. User asks "can you print X?" or references a model → same workflow.
3. User asks about printer status, material station, or what's loaded → status/channels.

Setup

1. Install OrcaSlicer if missing — download the Linux AppImage from https://github.com/OrcaSlicer/OrcaSlicer/releases (v2.4.2 asset pattern OrcaSlicer_Linux_AppImage_Ubuntu2404_*.AppImage, x86_64), install deps libopengl0 libglu1-mesa libwebkit2gtk-4.1-0 libjavascriptcoregtk-4.1-0, chmod +x, verify with --help. If {{ORCA_BIN}} is missing or the file does not exist, offer to install it — you may download and set it up yourself rather than just asking for a path. Headless: export APPIMAGE_EXTRACT_AND_RUN=1 or extract once with --appimage-extract and point {{ORCA_BIN}} at squashfs-root/AppRun.
2. Install the Python API: pip install flashforge-python-api.
3. Copy assets/profiles/*.json (SOP-tuned PLA/PETG filament profiles plus quality and fast process profiles) next to your OrcaSlicer profile collection, and point {{ORCA_MACHINE_PROFILE}} at the "Flashforge AD5X 0.4 nozzle" machine profile that ships with OrcaSlicer. The bundled profiles are materialized (self-contained). After any edit, regenerate them with:

   python3 scripts/materialize_profiles.py \
     --profiles-dir <dir with your profile JSONs> \
     --orca-resources <OrcaSlicer resources/profiles dir> \
     --output <extra copy dirs...>

   (See the "CLI ignores inherits" pitfall — without baking, headless slices silently fall back to Orca's generic slow defaults: walls 60 mm/s and a 2 mm³/s volumetric cap instead of 200 mm/s and 25.)

4. Create {{STATE_FILE}} (default ~/3dprinter/state.json; override with the 3DPRINTER_STATE env var). The variables map into it like this:

   • {{PRINTER_IP}}, {{PRINTER_SERIAL}}, {{PRINTER_CHECK_CODE}} → printer.host / printer.serial / printer.check_code
   • {{ORCA_BIN}}, {{ORCA_DATADIR}}, {{ORCA_MACHINE_PROFILE}} → slicer.orca_bin / slicer.datadir / slicer.machine_profile
   • {{STATE_FILE}} → the file itself (or 3DPRINTER_STATE)

   Then chmod 600 {{STATE_FILE}} — it holds printer credentials (serial + check code). The channels.py script enforces 0600 on writes. Full example:

   {
     "printer": {"host": "192.168.10.50", "serial": "SNXXXXXXXXXXXX", "check_code": "a1b2c3d4", "model": "Flashforge AD5X"},
     "channels": {"1": {"material": null, "color": null, "brand": null, "remaining_pct": null}, "2": {"material": null, "color": null, "brand": null, "remaining_pct": null}, "3": {"material": null, "color": null, "brand": null, "remaining_pct": null}, "4": {"material": null, "color": null, "brand": null, "remaining_pct": null}},
     "slicer": {"orca_bin": "~/OrcaSlicer.AppImage", "datadir": "~/orca-data", "machine_profile": "~/OrcaSlicer/profiles/Flashforge AD5X 0.4 nozzle.json", "profiles_dir": "~/profiles", "output_dir": "~/gcode"},
     "defaults": {"layer_height": 0.2, "infill_percent": 15, "supports": "auto", "auto_start_print": false}
   }

Slice workflow

1. Receive model — save the attachment to a local models directory.
2. Ask channels — confirm the material station before slicing: which channel(s), material, and color are loaded (channels.py show first; confirm changes with the user; do not assume the state file is current).
3. Ask print intent — purpose drives parameters (see policy below): decorative / functional / prototype; supports on/off; quantity.
4. Slice:

   python3 scripts/slice.py model.stl --material PLA [--name foo] [--profile "Spock Fast 0.24 @FF AD5X.json"] [--supports none] [--infill 10] [--layer 0.24] [--no-thumbnail]

   → gcode at <output_dir>/<name>_<MATERIAL>.gcode plus est. time, layers and bed_verified (the script fixes the OrcaSlicer bed-temp quirk and verifies). slice.py also embeds a 140x110 LCD thumbnail PNG by default (headless OrcaSlicer never writes the image bytes itself — see pitfall); disable with --no-thumbnail.
5. Report — est time, material, layer count; ask confirm to upload/print (default: upload + confirm before starting; bed must be clear, filament loaded).
6. Upload/start:

   python3 scripts/ff_print.py upload <gcode> [--start]
   python3 scripts/ff_print.py status

Parameter policy (agent-owned)

Pick optimal values for quality/speed balance + easy support removal. Defaults (0.4 mm nozzle, AD5X):

• Supports: tree by default, interface disabled for clean snap-off; switch to none when the model self-supports or geometry is simple.
• Brim: auto; skirt only when a raft is needed (rare on PEI).

Material table (recommended starting points)

Start gcode — PETG-safe purge (recommended)

If you switch between PETG and PLA, add this purge to the machine profile's machine_start_gcode so residual PETG (which needs 235–245°C) doesn't contaminate a PLA first layer (PLA prints at 210°C — too cold to melt leftover PETG). Verified routine — edit the Flashforge AD5X 0.4 nozzle machine profile pointed to by {{ORCA_MACHINE_PROFILE}}:

M190 S[bed_temperature_initial_layer_single]
M109 S[nozzle_temperature_initial_layer]   ; 215 for PLA
G90 / M83 / G1 Z2 / G1 X50 Y220 Z0.25      ; move to purge position
M104 S245 / M109 S245                      ; heat to PETG-safe temp
G1 E10 F600                                ; push 10mm — flushes residual PETG
G1 E-1 F1200                               ; retract to stop ooze while cooling
M104 S[nozzle_temperature_initial_layer]
M109 S[nozzle_temperature_initial_layer]   ; back to print temp
G92 E0 … normal purge line at print temp

• Costs ~1–2 min per print start. Harmless for PLA-only workflows.
• Verify after editing: re-slice and confirm M104 S245 appears before the normal purge line in the output gcode.

Scripts

• scripts/slice.py — headless slice + bed-temp fix + LCD thumbnail embed + verification
• scripts/gcode_thumbnail.py — render/inject the 140x110 LCD preview PNG (used by slice.py)
• scripts/materialize_profiles.py — bake vendor inherits chains into the profile JSONs
• scripts/ff_print.py — status / list / speed [N] (set print-speed override %, e.g. 150, or query) / upload / cancel via flashforge-python-api
• scripts/channels.py — material station inventory (show / set / clear)

All scripts read the state file from {{STATE_FILE}} (or $3DPRINTER_STATE).

Pitfalls

• OrcaSlicer 2.4.2 AD5X bed-temp quirk: filament bed_temperature* keys can be ignored on this machine (gcode emitted M190/M140 S35 pre-materialization, S55/S60 after baking). slice.py pins EVERY M190/M140 from the material table (regex, not literal S35) and verifies — never ship gcode with the wrong bed temp. Verify: grep -oE "M1(90|40) S[0-9]+" file.gcode.
• Headless CLI does NOT resolve profile inherits — silent speed collapse (2026-09-02): the GUI resolves a profile's inherits chain; the CLI does NOT, and falls back to Orca's generic defaults: outer/inner walls 60 mm/s, infill/top 100, travel 120, accel 500, and — worst of all — filament_max_volumetric_speed = 2 mm³/s which throttled ALL extrusion to ~22 mm/s regardless of profile speeds. Vendor-intended AD5X values: outer 200 / inner 300 / infill 270 / travel 500 mm/s, accel 5000–10000, volumetric cap 25 (PLA) / 12 (PETG). FIX: materialize the profiles (see Setup) and re-run after every profile edit. ALWAYS verify a new slice: grep -m1 '^; outer_wall_speed' x.gcode → 200, grep -m1 '^; filament_max_volumetric_speed' x.gcode → 25.
• Headless CLI never embeds the LCD thumbnail PNG: the gcode carries ; thumbnails = 140x110/PNG but zero image bytes (; thumbnail begin / iVBOR absent) because thumbnail rendering happens in the GUI export path — so the printer LCD shows no object preview. FIX (2026-09-02): slice.py injects a rendered 140x110 iso PNG (via scripts/gcode_thumbnail.py, mesh iso view; PIL fast path, pure-numpy fallback) at the top of the file in the standard Orca/Prusa comment format. Verify: grep -c '^; thumbnail begin' x.gcode ≥ 1. Cosmetic only.
• HTTP API (8898) requires LAN mode: if all HTTP requests get empty replies while TCP 8899 works, the printer is not in LAN mode (touchscreen Settings → Network) — enable it and reboot. LAN mode disables the vendor cloud app's webcam monitoring; use the printer's own MJPEG stream instead: http://<PRINTER_IP>:8080/?action=stream (also returned by POST /detail as cameraStreamUrl; the library exposes it via client.info.get_detail_response() → .camera_stream_url). Open the URL in any browser to watch the print.
• AD5X material-station uploads: for multi-color/material-station prints use upload_file_ad5x with AD5XUploadParams (+ material mappings); the ff_print.py upload command uses the simpler upload_file path (fine for single-material jobs).
• Credentials: serial number + check code are per-printer credentials — keep them out of logs, bug reports, and public repos; chmod 600 the state file (channels.py enforces it on writes).
• Never start a print without confirming the bed is clear and the right channel/material is loaded. defaults.auto_start_print is an agent-level policy (the agent reads it and refuses to auto-start when false); the CLI --start flag is the explicit override after confirmation.
• Station-feed phase is SLOW — don't cancel: once a material-station slot is configured, the printer runs a station feed (state_action == 4) at print start. It pushes filament from the spool through the ~1 m tube and can take 5–10 minutes: nozzle holds ~140°C standby, layer stays 0, run time 00:00. This is normal — wait for the feed to finish, then the nozzle heats to print temp and the purge line runs.
• Runout at layer 1–2 = filament not seated in the extruder, not a spool problem (the station slot sensor can report loaded while the extruder sensor sees nothing). Fix: hand-feed filament into the extruder while purging — API purge works: set_extruder_temp(230, wait_for=False) then extrude(100, 300) (~20 s purge). Note the extrude command acks instantly but executes asynchronously — a 0.0 s return does NOT mean it failed; watch the nozzle for flow.
• Camera stream may serve no frames while the print is paused/feeding — the MJPEG connection opens but returns EOF. Retry during active printing; don't burn cycles diagnosing the network mid-print.
• Direct --profile load ignores the profile's own layer_height key (verified 2026-09-01): slicing with --profile "Spock Fast 0.24 @FF AD5X.json" produced 522 layers — identical to the 0.20 quality slice — even though the profile sets "layer_height": "0.24". Other keys (shells, initial layer, speeds) DID apply. FIX: pass --layer 0.24 (and --infill 10) explicitly — the CLI-override temp profile path works (436 layers, verified). Always confirm the resulting layer count / ; layer_height = line in the gcode.
• slice.py writes enable_support (NOT support_enable) — the invalid key was silently ignored by the OrcaSlicer fork and printed into thin air (2026-09-01 incident: a failed print, no supports in gcode). Fixed in slice.py + a SUPPORT GATE: if --supports auto|tree is requested and the gcode header shows enable_support = 0, slice.py exits non-zero — never print that gcode.
• Verify model orientation BEFORE slicing (same 2026-09-01 incident): slice.py reports base_mm2 (flat contact at Z=0) and warns when it is tiny (<500 mm², e.g. a model standing on a knife edge). Stand/dock models with a flat base often arrive rotated 90° — check min/max extents per axis and rotate so the flat face is down. Render previews with scripts/stl_preview.py model.stl outprefix before confirming the slice.
• MULTI-PART PLATES: raft expansion fuses adjacent parts — space parts for the raft, not the bounding box (2026-09-01 incident: 4 Hook + 4 Assembly plate printed fused; the bonded pieces were scrapped — cuts would look bad). Root cause: raft first_layer_expansion = 2.0 / expansion = 1.5 grows the raft ~2 mm past EVERY part edge, so parts spaced ~4 mm apart (bbox to bbox) get their rafts merged into ONE continuous sheet; the gcode's first raft layer showed a single uninterrupted span across the whole bed. RULE: min bbox-to-bbox spacing = 2 × raft_first_layer_expansion + margin = 8 mm when raft is on. Verify after slicing: the first raft layer (lowest Z) must show GAPS between islands; a continuous raft span = parts will fuse — pack wider. Tree-support bases can bridge tight gaps the same way.

Verification

• Slice test: slice.py test_cube.stl --material PLA --json → bed_verified: true, temps M190 S65 / M140 S62, nozzle 210/215, thumbnail: true.
• Speed/flow sanity on any new gcode: grep -m1 '^; outer_wall_speed' x.gcode → 200, grep -m1 '^; sparse_infill_speed' x.gcode → 270, grep -m1 '^; filament_max_volumetric_speed' x.gcode → 25 (PLA) / 12 (PETG). If any show 60 / 100 / 2 the profile regressed — re-run materialize_profiles.py.
• Thumbnail presence: grep -c '^; thumbnail begin' x.gcode → ≥ 1 (cosmetic).
• Rotating preview GIF (dashboard "Replicator Preview"): python3 lcars-dashboard/preview.py model.stl out.gif → ≥24 frames, loop 0. The dashboard's /api/preview?file=<gcode> resolves a printing gcode name to a local STL and /preview/<stem>.gif serves a cached rotating GIF — see lcars-dashboard/README.md for endpoints and env vars.
• Printer: ff_print.py status → file count (TCP fallback works even when HTTP is down).

Sibling skills (same 3dprinting category)

• openscad-cad — DESIGN stage: turn a sketch / photo / spoken dimensions into a parametrically-coded, verified STL with OpenSCAD, then hand off here.
• multipart-stl-plates — multi-shell STLs (Bambu/MakerWorld plates), plate packing for the 220 mm bed, binary-STL read/write, Orca plate-rejection codes.