aporn-tool 0.6.1

Astropornography tool - one command from raw subs to a finished astrophoto

aPornTool

One command from raw sub-exposures to a finished, share-ready astrophoto.

📦 Install: pip install aporn-tool - releases · changelog.

aPornTool takes a folder of raw astrophotography subs and drives them through the full linear-to-nonlinear editing pipeline: stacking, photometric colour calibration, gradient and noise removal, colour-preserving stretch, and star management. Only the deliverables land in your output folder; scratch files get tucked into a hidden working directory.

It's built for the ZWO Seestar workflow (internal calibration → no darks/flats/bias) but works with any OSC .fit subs, and runs natively on Windows, macOS (incl. Apple Silicon), and Linux.

Design goal: processing parity. On the same data, aim to match what a skilled processor would get, so the only remaining variable is capture (aperture, integration hours, sky darkness). The tool can't invent detail that isn't in the data. The final crop, curves, and watermark are yours to do from the 16-bit TIFF (Canva/Photoshop).

---

Contents

• How it works
• Prerequisites
• Installation
• First-time setup
• Usage
• Command reference
• Stages & resume
• Output layout
• Configuration file
• FAQ
• Troubleshooting
• Limitations
• Development

---

How it works

raw .fit subs ──► stage ──► SIRIL: calibrate ─► register ─► stack ─► [SPCC] ─► golden linear stack
                                                                                       │
   deliverables ◄── finish ◄── [StarNet] ◄── [GraXpert BGE + denoise] ◄── [auto-crop] ◄┘
   (.tif/.png/.jpg/.fits)

• Each stage is checkpointed. The golden linear stack (_work/<target>/02_linear/) is the immutable re-processing anchor. Re-finishing never re-stacks.
• The pipeline is parameterised per mode: registration, gradient tool, SPCC placement, and star handling all differ depending on what you're shooting (see the table below).

Mode	Best for	Registration	Gradient	Stars
dso-galaxy	galaxies - M31, M33, M51, M101, M81 (mosaic auto-detected from the subs' pointing spread; override --mosaic/--single)	WCS framing=max+feather (mosaic) or register -2pass+mirrorx (single)	GraXpert BGE	composite dual-layer (StarNet2), blend some back
dso-emission-nebula	Hα HII / SNRs - M8, M20, M42, M16, Veil	star-based 2-pass	SIRIL subsky	keep all (rich field)
dso-reflection-nebula	blue scattered light - VdB106, M78	star-based 2-pass	GraXpert BGE	dual-layer screen blend
dso-star-cluster	globulars & open clusters - M13, M22, M45, M44	2-pass (+FWHM cull)	SIRIL subsky	keep all, stars are the subject
dso-milky-way	wide-field Milky Way from a phone/camera (JPEG/HEIC/PNG/TIFF stills, not FITS)	single-pass global (no cull)	GraXpert BGE (high smoothing)	keep all, stars are the subject

dso-milky-way is the odd one out: it ingests already-processed camera stills instead of raw FITS subs, so it skips calibration, plate-solving, and SPCC. Put your phone on anything steady, take a dozen-plus frames of the same patch of sky, and it stacks them to pull the galactic core out of the noise. See the Milky Way section below.

Planetary (video → AutoStakkert → finish) is planned; it requires a manual GUI stacking step.

---

Prerequisites

Tool	Role	Get it
Python 3.10+	orchestrator	python.org / your package manager
Siril 1.4+ (siril-cli)	stack, register, plate-solve, SPCC, StarNet	https://siril.org/download/
GraXpert 3.x	background extraction + AI denoise (galaxy & reflection)	https://github.com/Steffenhir/GraXpert/releases
StarNet2	star removal	https://www.starnetastro.com/
ffmpeg / ffprobe	planetary + final polish	ffmpeg.org / package manager

Every DSO tool runs natively cross-platform (incl. Apple Silicon). The tool locates each one on PATH, then in the standard install locations per OS, then from aporntool.config.json. You never hard-code paths.

Where the tools usually live per OS

	Windows	macOS	Linux
siril-cli	C:\Program Files\Siril\bin\siril-cli.exe	/Applications/Siril.app/Contents/MacOS/siril-cli	/usr/bin/siril-cli
GraXpert	%LOCALAPPDATA%\Programs\GraXpert\GraXpert.exe	/Applications/GraXpert.app/Contents/MacOS/GraXpert	package / AppImage
StarNet2	(set in SIRIL)	/usr/local/bin/starnet2	/usr/local/bin/starnet2
GraXpert AI models	%LOCALAPPDATA%\GraXpert\GraXpert\{bge,denoise}-ai-models\	~/Library/Application Support/GraXpert/{bge,denoise}-ai-models/	~/.local/share/GraXpert/…
SIRIL config	%LOCALAPPDATA%\siril\	~/Library/Application Support/org.siril.Siril/siril/	~/.config/siril/

---

Installation

Already have Python 3.10+? Install from PyPI:

pip install aporn-tool

That gives you the aporn-tool command. python -m aporntool also works as a module fallback (module names can't contain hyphens, so the Python package keeps the one-word aporntool name), and an aporntool alias command is installed too. Either spelling works.

Or install from source (for development / latest main)

git clone https://github.com/christiantroyandrada/aporn-tool.git
cd aporn-tool
python3 -m venv .venv
.venv/bin/python -m pip install -e .
# Windows: .venv\Scripts\python -m pip install -e .

Installing the Python package does not install Siril/GraXpert/StarNet. Those are separate apps (see Prerequisites). Run aporn-tool config --check to confirm they're found.

---

First-time setup

Run the discovery check. It prints where each tool was found and writes a starter config:

aporn-tool config --check

Then complete the three one-time setup steps (the tool's preflight checks the ones it can):

1. GraXpert AI models (mosaic & reflection) - open GraXpert once and run Background Extraction and Denoise on any image so it downloads the model files. Preflight verifies they exist before stacking, so a missing model fails in seconds, not after a 30-minute stack.
2. StarNet2 (galaxy, emission, reflection) - the composite dual-layer finish calls the StarNet2 CLI directly, so it just needs to be discoverable (set tool_paths.starnet2 in the config, or have it on PATH). No SIRIL-internal StarNet configuration is required. (Star-cluster keeps all its stars and needs no StarNet.)
3. Local Gaia catalogs in SIRIL (SPCC colour calibration) - online VizieR is retired. Install the sky region matching your target: Milky Way for galactic nebulae/low-latitude clusters, Galaxy Season for high-latitude galaxies. Wrong region → SPCC reports "no stars".

Validate everything without processing:

aporn-tool dso-galaxy --in "/path/to/subs" --out /path/to/out --target M31 --preflight-only

---

Usage — from simplest to advanced

1. The simplest run

Point it at a folder of subs and pick the mode, that's all. The object name and coordinates are read from the subs' FITS header, and the results land in a folder named after the target, right beside your subs:

aporn-tool dso-galaxy --in "/path/to/M31 subs"

Produces M31_final.tif (16-bit, the real deliverable), .png, .jpg (quick-look), and .fits next to the input; everything else lives under _work/.

Override either default when you want to: --target for a custom name, --out for a custom location.

⚠️ The output path must not contain spaces (a SIRIL limitation). The default output inherits the subs' parent folder; if that path has spaces, pass a space-free --out. The --in path and sub filenames may contain spaces.

2. Other modes

aporn-tool dso-emission-nebula   --in "/data/M8"
aporn-tool dso-reflection-nebula --in "/data/M78"
aporn-tool dso-star-cluster      --in "/data/M13"

2b. Wide-field Milky Way from a phone

dso-milky-way takes a folder of camera/phone stills (.jpg .jpeg .png .tif .tiff .heic) instead of FITS subs — no --target needed (the run is just named MilkyWay):

aporn-tool dso-milky-way --in "/path/to/phone shots"

How to shoot for it: point at the Milky Way and take a dozen or more frames of roughly the same framing (Night Mode or a long-exposure app gives the most signal). The tool aligns on the stars, so even a hand-held burst stacks fine — star registration is invariant to the shift and rotation between frames. More frames = less noise. Needs SIRIL + GraXpert (no StarNet, no Gaia catalog).

Hand-held? Add --no-tripod. Because the tool aligns on the stars, a fixed foreground (rooftops, trees, wires) smears into a ghost as the framing drifts between hand-held frames — that is the one thing star-aligned stacking can't fix. --no-tripod recovers a sharp foreground: it keeps the deep stacked sky but paints the foreground back in from a single frame.

aporn-tool dso-milky-way --in "/path/to/phone shots" --out ~/Pictures/mw_out --no-tripod

It finds the foreground automatically from how much each pixel moves between the registered frames (the sky is aligned = still; the foreground drifts = moving) — no horizon line to draw, and it works whether your foreground is along the bottom, the side, or scattered rooftops. Hand-held drift also leaves a wider ragged, colour-fringed border around the stack. For a clean, fringe-free framing, pass an explicit --crop "X Y W H" box that crops past that border (the default auto-crop keeps the border as a mild sky fringe rather than risk clipping into it). On a real tripod, leave --no-tripod off (there's no ghost to fix). Tuning knobs live under pipeline.no_tripod in the config.

Two practical notes:

• Output path with spaces: phone shots often live under a spaced path (e.g. ~/Pictures/Milky Way), and the default --out lands beside them, which SIRIL can't handle. Pass an explicit space-free --out, e.g. --out ~/Pictures/mw_out.
• HEIC (iPhone default): supported only if your SIRIL was built with HEIF support (most current builds are). The tool prints a heads-up when it sees .heic; if registration finds no frames, export the frames to JPEG or TIFF and re-run.

2c. DSLR / mirrorless DSO (FITS-less)

The four DSO modes also take DSLR/mirrorless lights — raw (.cr2 .cr3 .nef .arw .dng .raf …), TIFF, or JPEG — no FITS required. Point --in at your lights and pass calibration frames as folders:

aporn-tool dso-galaxy \
  --in    "/data/M31/lights" \
  --darks "/data/M31/darks" \
  --flats "/data/M31/flats" \
  --bias  "/data/M31/bias" \
  --target M31 --focal 530 --pixel 3.76 --out /data/M31_out

Notes:

• A folder is read as one format (priority FITS > raw > TIFF > JPEG), so raw + in-camera JPEG pairs just work — the JPEGs are ignored.
• Master darks/flats/bias are optional and applied when given (a masters stage builds them).
• --target is required (raw has no OBJECT header); pass --coords RA,DEC for a target not in the built-in catalog. Raw/FITS get debayered; TIFF/JPEG are treated as already-RGB.
• --focal/--pixel set the plate-solve geometry for the SPCC in galaxy/reflection (SPCC runs in preprocess there). Emission/star-cluster solve blind in the finish phase, so those two don't use --focal/--pixel yet; SPCC may not calibrate DSLR emission/cluster colour without a catalog target.
• Master darks/flats/bias also work for cooled astro-camera FITS lights, not just DSLR raw.
• For accurate SPCC colour, set your camera under pipeline.spcc.sensor in the config (e.g. "Canon EOS 2000D" for a T7). SPCC uses the local Gaia catalog if installed, otherwise it reverts to the online Gaia catalogue automatically.
• Emission targets need the right camera. Nebulae like the Rosette or Orion are mostly hydrogen-alpha (656 nm), which a stock (unmodified) DSLR's IR-cut filter largely blocks — you can't process in red that wasn't captured. For strong Halpha, use an astro-modified camera or a dual-band/Ha filter. Broadband targets (star clusters, reflection nebulae like the Pleiades, galaxies, the Milky Way) are unaffected.

Already stacked elsewhere? Add --stacked to feed a finished linear integration (from DSS, Siril, etc.) straight into the finish — no re-stacking:

aporn-tool dso-reflection-nebula --in "/data/PLEIADES_STACKED.tif" --target M45 --stacked --out /data/m45_out

--stacked works on any mode and imports the single image as the golden anchor, then runs that mode's BGE / StarNet / stretch. (Galaxy and reflection do their SPCC in preprocess, which --stacked skips, so they use the image's existing colour; emission and star-cluster still colour-calibrate in finish.)

3. Combine multiple nights (more integration = the #1 quality lever)

--in is repeatable; all .fit from every source are staged and stacked together:

aporn-tool dso-galaxy \
  --in "/data/M31/2026-07-04" \
  --in "/data/M31/2026-07-05" \
  --out /data/out --target M31

4. A target not in the built-in catalog

Nothing special needed. The coordinates come from your subs' FITS header, so unlisted targets (VdB 106, Sh2-155, …) just work. Pass --target only when you want a custom output name:

aporn-tool dso-reflection-nebula --in "/data/vdb106"
aporn-tool dso-emission-nebula   --in "/data/sh2-155" --target Sh2-155

5. Control the crop

Auto-crop (default) trims empty registration/mosaic borders. Override it:

# keep the full frame (no crop)
aporn-tool dso-galaxy --in "/data/M31" --out /data/out --target M31 --no-crop

# explicit SIRIL crop box: X Y W H (x from left, y from top)
aporn-tool dso-galaxy --in "/data/M31" --out /data/out --target M31 --crop "162 108 2310 4378"

6. Check status and resume

Re-running the same command auto-resumes at the first unfinished stage, nothing done is repeated. Inspect the ledger any time:

aporn-tool status --out /data/out --target M31

dso-galaxy / M31  (fingerprint 8cd20ff7d100564d)
  calibrate  done
  register   done
  stack      done
  spcc       done
  bge        done
  denoise    done
  finish     failed
Resume at: finish

7. Re-run specific stages (cheap, from the golden anchor)

# redo colour calibration and everything after it (e.g. once the right Gaia region is installed)
aporn-tool dso-galaxy --in "/data/M31" --out /data/out --target M31 --redo spcc

# restart at a named stage
aporn-tool dso-galaxy --in "/data/M31" --out /data/out --target M31 --from finish

# ignore all checkpoints and re-run everything
aporn-tool dso-galaxy --in "/data/M31" --out /data/out --target M31 --force

8. The advanced, fully-specified run

aporn-tool dso-galaxy \
  --in "/data/M31/2026-07-04" --in "/data/M31/2026-07-05" \
  --out /data/out \
  --target M31 \
  --crop "162 108 2310 4378" \
  --star-reduce 0.35 \
  --config /data/my-aporntool.config.json \
  --redo spcc

• --star-reduce 0.35 - after StarNet removes stars (mosaic), blend 35% of them back (lower = fewer/dimmer stars; default 0.5).
• --config - use a specific config file (tool paths, catalogs, Seestar defaults).
• --redo spcc - recompute from the spcc stage onward, reusing the stack.

---

Command reference

aporn-tool <command> [options]

Command	Purpose
dso-galaxy, dso-emission-nebula, dso-reflection-nebula, dso-star-cluster	process a deep-sky target (FITS subs)
dso-milky-way	stack a wide-field Milky Way from phone/camera stills (JPEG/HEIC/PNG/TIFF)
config --check [--config PATH]	show tool discovery; write a starter config
status --out PATH --target NAME	print the resume ledger
--version	print the version

Options for the processing modes:

Option	Description
--in PATH	subs folder (required, repeatable for multi-night)
--out PATH	output root (optional; defaults to a <TARGET> folder beside the subs; must be space-free)
--target NAME	object name (optional; auto-detected from the subs' FITS OBJECT header)
--crop "X Y W H"	explicit SIRIL crop box (default: auto-crop)
--no-crop	disable auto-crop; keep the full frame
--star-reduce F	mosaic star blend-back fraction (default 0.5)
--darks PATH	folder of dark frames (DSLR calibration; DSO modes)
--flats PATH	folder of flat frames (DSLR calibration; DSO modes)
--bias PATH	folder of bias/offset frames (DSLR calibration; DSO modes)
--focal MM	focal length (mm) for the SPCC plate solve (DSLR; default: Seestar)
--pixel UM	pixel size (microns) for the SPCC plate solve (DSLR; default: Seestar)
--no-tripod	hand-held Milky Way: recover a sharp foreground from one frame, de-ghosting the house/trees/wires (dso-milky-way only)
--clean	on success, delete working files except the golden anchor (reclaims disk)
--from STAGE	restart at this stage
--redo STAGE	re-run this stage + everything downstream
--force	re-run all stages, ignore checkpoints
--preflight-only	validate the environment, then stop
--config PATH	config file (default aporntool.config.json)

Catalogued targets get canonical coordinates: M3 M4 M5 M6 M7 M8 M11 M13 M15 M16 M20 M22 M31 M33 M42 M44 M45 M51 M81 M92 M101 NGC869 NGC6960 NGC7000. Anything else uses the RA/DEC from your subs' FITS header automatically.

---

Stages & resume

Stage names are per-mode. Run aporn-tool status to see yours. Current orders:

Mode	Stages (what --from / --redo accept)
dso-galaxy	calibrate → register → stack → spcc → bge → denoise → finish
dso-reflection-nebula	calibrate → register → stack → spcc → bge → denoise → finish
dso-emission-nebula	calibrate → register → stack → mirrorx → finish
dso-star-cluster	calibrate → register → stack → mirrorx → finish
dso-milky-way	register → stack → anchor → bge → denoise → finish (convert is merged into register; no calibrate/mirrorx/SPCC); --no-tripod appends → deghost

A stage is only marked done after its output is verified (exists, non-empty, right type). A crash or a raising stage is marked failed and reported with the log tail. Re-run to resume. Changing a parameter re-runs that stage and everything downstream, nothing upstream.

---

Output layout

<OUT>/                       ← DELIVERABLES ONLY
├─ M31_final.tif             ← 16-bit, the real deliverable
├─ M31_final.png
├─ M31_final.jpg             ← quick-look
├─ M31_final.fits
└─ _work/                    ← everything scratch, hidden here
   └─ M31/
      ├─ 00_lights/          hardlinked .fit subs
      ├─ 01_process/         SIRIL sequences + frames
      ├─ 02_linear/          M31_Linear.fit  ← GOLDEN ANCHOR (never deleted)
      ├─ 03_graxpert/        BGE / denoise intermediates
      ├─ 05_finish/          finish scratch
      ├─ logs/               generated .ssf + per-stage stdout
      └─ aporntool.json      run manifest / resume state

Multiple targets can share one --out; each gets its own _work/<target>/ and deliverables at the root.

---

Configuration file

Every tunable parameter lives in one file, aporntool.config.json — tool paths, Seestar optics, local Gaia catalogs, and a pipeline block with every processing knob grouped by stage/mode. The file is optional and appears out of the box: the first run writes it (all defaults) next to you, editing a value overrides just that knob, and deleting the file restores every default. You can also write it on demand:

aporn-tool config --init      # write the file pre-filled with ALL defaults, ready to edit
aporn-tool config --check     # same, and verify Siril/GraXpert/StarNet are discoverable

{
  "tool_paths": { "siril": "…", "graxpert": "…", "starnet2": "…", "ffmpeg": "…" },
  "seestar_focal_mm": 150.0,
  "seestar_pixel_um": 2.9,
  "catalog_astro": null,
  "catalog_photo": null,
  "pipeline": {
    "stack":    { "sigma_low": 3.0, "sigma_high": 3.0, "feather_mosaic": 100, "filter_round": "2.5k", "filter_wfwhm": "2.5k" },
    "graxpert": { "bge_smoothing": 0.0, "bge_correction": "Subtraction", "denoise_strength": 0.8 },
    "crop":     { "bg_frac": 0.25, "margin_frac": 0.02, "target_blocks": 700 },
    "spcc":     { "sensor": "Sony IMX662", "osc_filter": "UV/IR Block", "whiteref": "Average Spiral Galaxy", "catalog": "localgaia" },
    "mosaic_finish":     { "autostretch_clip": -2.8, "autostretch_bg": 0.15, "ght_d": 0.8, "ght_b": 3.0, "ght_sp": 0.15, "ght_hp": 0.85, "rmgreen": 1.0, "satu": 0.7, "star_reduce": 0.5 },
    "emission_finish":   { "subsky_degree": 1, "satu": 0.7, "satu_bg": 0.1 },
    "cluster_finish":    { "subsky_degree": 1, "denoise_mod": 0.5, "ght_d": 0.7, "ght_b": 3.0, "ght_hp": 0.9, "satu": 0.6, "satu_bg": 0.1 },
    "reflection_finish": { "target_bg": 0.35, "shadows_clip": -2.8, "sat_r": 0.30, "sat_g": 1.3, "sat_b": 4.5, "midboost": 0.55, "lc": 1.3, "bgpull": 0.08, "gamma": 0.85, "bg_desat": 0.14, "bg_desat_soft": 0.14, "st_bright": 1.5, "st_sat": 1.2 },
    "milkyway_finish":   { "bge_smoothing": 1.0, "bge_correction": "Subtraction", "denoise_strength": 0.8, "autostretch_clip": -2.9, "autostretch_bg": 0.10, "rmgreen": 1.0, "satu": 0.85, "satu_bg": 0.1 },
    "no_tripod":         { "barrier_pct": 85.0, "barrier_dilate": 4, "feather": 12.0, "min_island_frac": 0.002, "fg_target_bg": 0.10, "fg_shadows_clip": -1.8, "fg_gain": 0.6 },
    "jpeg_quality": 95
  }
}

Guards (a hand-edited file can't break a run): no file / deleted / empty / corrupt → all built-in defaults (a corrupt file just warns and falls back). A partial file overrides only the keys it sets; missing keys, unknown keys, and wrong-typed or non-finite values are ignored and keep their default. Change a value and re-run: only the affected stage recomputes (a stretch tweak → just finish; a stack/crop change → from that stage), reusing the golden linear stack.

Precedence: an explicit CLI flag beats the config — e.g. --star-reduce 0.3 overrides pipeline.mosaic_finish.star_reduce. Tool-path resolution: explicit config path → PATH → known install locations.

Scope: pipeline holds the tunable look dials. Rig optics (seestar_focal_mm / pixel_um) stay top-level (they describe the scope, not the processing), and a few deep reflection-algorithm internals (white-point percentile, bloom kernels) remain structural constants — ask if you want any of those exposed too.

---

FAQ

Do I need darks/flats/bias? No. The Seestar calibrates internally. For other rigs, calibrate your subs beforehand (the tool ingests already-calibrated .fit).

How many subs do I need? More integration is the biggest quality lever. Combine nights with repeated --in. There's no hard minimum, but a handful of subs won't stack meaningfully.

Will it delete my raw data? No. Subs are hardlinked into _work/00_lights/; the golden linear stack and your originals are never overwritten.

Can I re-process without re-stacking? Yes, that's the point of the golden anchor. Tweak a finishing parameter and re-run; only finish (or the changed stage onward) recomputes.

It didn't crop the mosaic tightly. Auto-crop trims the black/near-empty borders, but it thresholds on brightness rather than per-pixel coverage, so a noisy low-coverage band can survive on mosaics (those edge pixels are normalized to near the sky level). Pass an explicit --crop "X Y W H", or do the final crop by hand in Canva/Photoshop. Tighter mosaic cropping is on the roadmap.

Which mode for the Pleiades (M45)? It's an open cluster in blue reflection nebulosity, so use dso-reflection-nebula, not the plain cluster finish.

Does it need a GPU? No. GraXpert uses the GPU when available and falls back to CPU.

---

Troubleshooting

Symptom	Fix
StarNet2 not found (galaxy/emission/reflection)	The composite finish calls the StarNet2 CLI directly - set tool_paths.starnet2 in the config or put it on PATH.
Preflight: GraXpert models missing (but installed)	Open GraXpert, run BGE + Denoise once so it downloads the model files.
SPCC "no stars" / imprecise	Install the Gaia region matching the target (Milky Way vs Galaxy Season).
--out path must not contain spaces	Choose a space-free output folder (SIRIL limitation).
golden anchor not found on --from/--redo	Run the full pipeline first (drop --from/--redo) so preprocess produces the anchor.
Plate-solving fails on a mosaic	Ensure SIRIL's astrometry catalog is installed; the tool verifies solves by frame count, not exit code.
Noisy/ragged band left after auto-crop	Known mosaic limitation (brightness-based crop misses low-coverage edges). Pass --crop "X Y W H", or crop in post.

Every stage saves its SIRIL script and stdout under _work/<target>/logs/. Check <stage>.log for the exact tool error.

---

Limitations

• Deconvolution / wavelets are skipped for undersampled Seestar data (they amplify noise without adding real detail).
• The last 5% (final crop, curves, watermark) is done by hand from the 16-bit TIFF.
• Planetary needs a manual AutoStakkert stacking step (no usable CLI) and is not yet shipped.
• Auto-crop trims black borders but can leave a noisy low-coverage band on mosaics (it crops by brightness, not coverage). Tighten with --crop or in post; a coverage-aware crop is planned.

---

Development

.venv/bin/python -m pip install -e . pytest pytest-timeout
.venv/bin/python -m pytest -q          # unit suite; external tools are mocked

Tests mock Siril/GraXpert/StarNet, so they run anywhere. Integration runs need the real tools and real subs. See REQUIREMENTS.md for the full design, the per-mode pipeline contract, and the load-bearing SIRIL/GraXpert/StarNet gotchas.