Skip to content

autopipeline.md

Latest commit

 

History

History
206 lines (164 loc) · 30.5 KB

autopipeline.md

File metadata and controls

206 lines (164 loc) · 30.5 KB
Raw

autopipeline: run villa's pipeline from a released scan to a scored ink map, unattended

AutoPipeline (AP) 0.1.9 runs villa's VC3D tools and ink detection on one released scan, from its headers to checked surfaces, ink maps with reversed and shuffled controls, and scores; its code is on public branch submission/progress-2026-10-fix9, not on main.

Field Value
Id autopipeline
Status Branch only
Kind package
Area Pipeline and agents
Run scrolls-kit autopipeline after pip install -e . in the package folder; inside that folder the branch's own kit module runs the same program (alias autopipe); python scripts/demo_synthetic.py; scrolls-kit vc3d mcp serves it as an MCP tool
Code submissions/progress-2026-10-fix9/autopipeline/ on its branch: mainly kit/autopipe.py (cmd_autopipe, build_parser, DEFAULTS), kit/stagereg.py with kit/stagereg_ext_ap.py and kit/data/stages.json (the stage registry), kit/ink3d.py (3D ink readers), kit/autopipe_monitor.py, kit/apconsole.py (the console), kit/qareasons.py, kit/apbackends.py, kit/vc3d_agent.py, kit/vc3d_cache.py, kit/cli.py, scripts/demo_synthetic.py, scripts/get_vc3d.py, scripts/vc3d/ink.py; release record submissions/progress-2026-10-fix9/release.json
Tests submissions/progress-2026-10-fix9/autopipeline/tests/ on its branch: 49 test_*.py files, among them test_autopipe.py, test_stagereg.py, test_ink3d.py, test_autopipe_ink3d.py (the whole chain on a synthetic scroll with stand-in VC3D tools), test_autopipe_monitor.py, test_autopipe_monitor_ui.py (the monitor page in a local Chromium), test_demo_synthetic.py and test_vc3d_agent.py, with stand-in tools in vc3d_fakes.py
Needs Python 3.11 or newer (zarr 3 needs it); numpy, tifffile, zarr 3 and scipy (pyproject.toml, tested 2026-10-12 with numpy 2.4.6, tifffile 2026.3.3, zarr 3.1.6, scipy 1.17.1). For real runs: villa's VC3D tools (vc_gen_umbilicus, vc_grow_seg_from_seed, vc_tifxyz_selfcross, vc_flatten, vc_render_tifxyz), from villa's Linux x86_64 AppImage or VC3D.app on a Mac; for ink, villa's vesuvius package and torch in their own environment (KIT_INK_PYTHON). Optional, for the browser tests only: the playwright Python package and a system chromium or chromium-browser; without them those tests skip and download nothing
Runs on The synthetic demo: 11.5 s as it reports itself in a cloud container, Python 3.13, from the 0.1.9 package's release check (2026-10-25); the package's 1281 tests took 514 s there (27 skipped). With 0.1.8 the demo reported 12.7 s and its 1251 tests took 601 s. With 0.1.7 the demo reported 13.1 s. With 0.1.6 the demo reported 5.5 s. With 0.1.5 it reported 4.7 s on a shared 4-core container (2026-10-18). A real run on PHerc0841 w00 from the released scan and published segment: 593 s, 344 s on 8 CPU cores then 249 s on one L4, on Modal (package README, 9 October 2026); the three-reader rerun took 1163 s on 8 CPU cores and one L4 (package README, 15 October 2026). GPU ink runs on Hugging Face (hard rule 11)
Network The open-data bucket vesuvius-challenge-open-data over HTTPS (scan and prediction headers and chunks, normal grids, published segments and ink labels); api.github.com and github.com for scripts/get_vc3d.py (villa's AppImage, sha256 checked); model downloads by the MCP tool fetch_model. The monitor listens on 127.0.0.1 only and loads nothing from the internet. The synthetic demo needs no network
Writes A run folder work/autopipe/<sample>-<volume>-<digest10>/ under the kit home (the package folder, or KIT_HOME): plan.json, state.json, events.jsonl, traces, renders, ink maps, scores.json, report.json and, when ink is handed off, modal-job.json; work/vc3d/ for the catalog cache and the unpacked AppImage (about 370 MB). The demo writes to a temporary folder under /tmp and prints its path. monitor --snapshot FILE.html writes one HTML file
Go-ahead Needed for a paid batch or any cloud (Modal) launch, which costs money (the go-ahead rule, hard rule 8); not needed for the demo, defaults, plan or a local run
Released Public branch submission/progress-2026-10-fix11 at 22cecae (2026-10-30; README entry dated 2026-10-30 in releases.json, pinned to the top of "Released tools"). Tagged autopipeline-v0.1.11, with a GitHub release not marked Latest (2026-10-30); 0.1.10 is on submission/progress-2026-10-fix10 at eea2df5, tagged autopipeline-v0.1.10 (2026-10-29); 0.1.9 is on submission/progress-2026-10-fix9 at 2144cdf, tagged autopipeline-v0.1.9, with a GitHub release not marked Latest (2026-10-25); 0.1.8 is on submission/progress-2026-10-fix8 at bf807b6, tagged autopipeline-v0.1.8 with a GitHub release not marked Latest (2026-10-25); 0.1.7 is on submission/progress-2026-10-fix7 at 24572e6, tagged autopipeline-v0.1.7 with a GitHub release not marked Latest (2026-10-24); 0.1.6 is on submission/progress-2026-10-fix6 at 8096b46, tagged autopipeline-v0.1.6 with a GitHub release not marked Latest (2026-10-24); 0.1.5 is on submission/progress-2026-10-fix5 at 9e50d61. Earlier versions stay on their own branches, the first, 0.1.0, on submission/progress-2026-10 since 2026-10-14; 0.1.4 is tagged autopipeline-v0.1.4
License Code: Apache License 2.0 (LICENSE on the branch); data and model weights keep their own licenses and are not included
Guide docs/autopipeline/README.md, the full reference (commands, settings, variables, stages, troubleshooting), with rules for agents in docs/autopipeline/AGENTS.md
Branch submission/progress-2026-10-fix9 at 2144cdf
Checked 2026-10-24 against commit 24572e6

Concept illustration of a continuous papyrus surface traced in copper and flattened into a grid.

Concept illustration: a papyrus surface traced in copper and flattened into a grid. No scan data.

What it does

villa's pipeline turns a CT scan into ink maps in steps: find the scroll's center line (the umbilicus), trace papyrus surfaces (segments, stored as tifxyz grids of 3D points), flatten them, render a thin slab of CT around each surface (a surface volume), and run an ink model on it. Today a person checks the output at almost every step. AutoPipeline replaces each of those choices with a rule fixed before the run, and gives every check a control that can fail. It calls villa's own tools and ink inference; it does not reimplement them.

One run covers one released scan and lives in one folder. plan.json freezes every setting and every resolved input (the zarr headers of every level, the surface prediction, the normal grids, the axis) under a SHA-256 digest, so a changed threshold means a new folder, never an edit to a run whose results were seen. state.json records each of the 11 stages: acquire, umbilicus, seeds, trace, bench, qa, flatten, render, ink, score, report. A finished stage is not repeated, so an interrupted run resumes; each failure is labeled as the tool's own decision or an infrastructure problem. Ink detection always also runs on the layers reversed and shuffled, and scoring (pixel AUC against published ink labels) scores those control maps beside the forward map. AutoPipeline Monitor is a read-only local web page (or a text view) of one run folder. Since 0.1.5 its page has keyboard-accessible controls: inspect a stage and its recorded failure, filter the loaded log tail, pause or refresh the viewer (not the pipeline), list the run's record files, and prepare a status or resume command for review and copying. Copying never runs a command, and a stale or paused view cannot prepare or keep a resume command. Since 0.1.6 the page fits one screen and has a console, on loopback only and off with --read-only: per-run notes kept under the kit home, and one AutoPipeline, kit or VC3D agent command at a time run as a background job.

Its default ink model (reader) is Reader V3 since 0.1.8 (domenicor046/reader-v3, MIT): Domenico Russo's Reader v2 continued on 384 px tiles, in ink_9um's checkpoint format, so villa's ink detection runs it unchanged at its own patch size. It is the newest release of the reader family AP already used; no rule fixed in advance chose it, and AP makes no claim that it reads ink better than Reader v2. Its author reports fewer false ink marks on blank papyrus (Community report) and benchmarked PHerc0841 while choosing its recipe, and it trained on w045, as Reader v2 did. "models": ["reader-v2"] in a plan's ink section keeps the previous default. Reader v2, the default from 0.1.3 to 0.1.7, is Domenico Russo's community fine-tune of ink_9um. On PHerc0841 w00, from the released scan and the published segment, it scored AUC 0.8752 forward, 0.5659 reversed and 0.5539 shuffled, against 0.7479, 0.5012 and 0.4851 for ink_9um seed 42 on the same render (package README, Modal, 15 October 2026). PHerc0841 is in the training data of none of ink_9um, v8in, d9v2 and Reader v2, but possibly in Hecate's; w00's labels had been scored before and these numbers chose the default, so they are development data, not a held-out test.

Since 0.1.9 any stage can be swapped by name. python -m kit stages lists every option of every stage (umbilicus, surface, trace, repair, flatten, render, ink, ensemble and others) with who made it, its license, its maturity (default, validated, experimental, pending or failed) and why; autopipeline plan SAMPLE --use STAGE=NAME (or KIT_STAGE_<STAGE>) writes the pick into the plan before its digest is fixed. An experimental pick warns and a pending one is refused. The ink stage also takes 3D readers (kit ink3d): a 3D model's volume is projected onto the 2.5D render canvas, so its map shares pixels, labels and scorer with a 2.5D map, and its control is the sheet moved two pitches out and in, since reversed depth means nothing in 3D. Every 3D reader is experimental: none has beaten the 2.5D default on a held-out test, and the default stays Reader V3. Registry page: stagereg.

What it does not do: traces grown from its own seed points are small patches with no held-out ink score yet, and whole-winding tracing is still open. Version 0.1.4 fixed remote renders for villa builds since 12 October 2026, but no full remote render has run with 0.1.4 or 0.1.5 yet. Version 0.1.5 changes usability and correctness only: the monitor controls, a synthetic demo that fails when a fake model fails or a map is missing, and cache-repair times stored at full precision, so one repair does not trigger an extra render. Its models, scientific settings and recorded measurements are those of 0.1.4, and it adds no ink-accuracy result. Version 0.1.6 adds qa failure reasons (each failed check with its value and threshold), opt-in steps that a plan must name (the Crackle Check gate at report, the published center lines as the axis, other axis backends, a region crop of rejected traces, edge extension, a stack check, trace budgets with partial results, z-scored ink ensembles), backend swap and provenance commands, and two built-in ink models (inkfilter and reader-v3); Reader v2 stays the default. A plan naming none of the new settings runs as before, except that qa.repair without a method now uses the barrier repair. It also adds no ink-accuracy result (CHANGES.md). Version 0.1.7 changes display and data only: ink maps in the monitor and the ink Space in one color ramp, dark umber for a low model score to pale papyrus for a high one, cropped to the surface and enlarged smoothly, with the reversed control on its forward map's scale; a synthetic demo whose CT has planted pen strokes, so its forward map shows them and its control does not; and center lines v2.4, which take the Center Finder's coverage-polished line before the original. A 0.1.6 plan that names umbilicus.centerlines pins the old dataset revision and is refused by 0.1.7, so make it again. Every map and score it makes is model output, not a reading; a triage order of maps is a reading order for a person, not a verdict. Here the release is called autopipeline in code; its short name is AP, while text for villa and the community spells out AutoPipeline.

Where the code is

  • Branch submission/progress-2026-10-fix9 of public Scrolls, commit 2144cdfcac4ecfc33150f7e81317e5ab33f248d6 (its parent is main at ff4e9e3; it adds only the package folder), folder submissions/progress-2026-10-fix9/: CHANGES.md (what 0.1.9 changed), SUBMISSION.md, release.json (the package's 165 file hashes and its manifest digest), and autopipeline/ (the installable package vc-autopipeline 0.1.9: kit/, scripts/, docs/vc3d-agents.md, docs/vc3d-headless.md, space/ for the AutoPipeline ink Space, spaces/crackle-check/ with the two Crackle Check engine files reader-v3 runs, tests/, Dockerfile, README.md, AGENTS.md).
  • The package's space/ copy runs Reader V3 with 384 px patches (uploads 384 to 768 px a side; a GPU call asks for 28 s for a 640 px sample). On a CPU its code reproduced earlier Reader V3 scores on the PHerc0841 w00 and ag896 sample crops to four decimals; those crops were chosen with the labels, so they are development numbers. The space/ copy is the same as 0.1.8's. The live AutoPipeline ink Space runs it with its Connect your tool files since revision 137b2f1f (2026-10-25); a live run of the w00 sample gave AUC 0.9414 stored order and 0.6621 reversed, the same as the CPU run.
  • Earlier releases, each on its own branch and never changed: 0.1.0 submission/progress-2026-10 (faada83), 0.1.1 submission/progress-2026-10-fix1 (1534136), 0.1.2 submission/progress-2026-10-fix2 (143c54b), 0.1.3 submission/progress-2026-10-fix3 (bd7d6cf), 0.1.4 submission/progress-2026-10-fix4 (603fe14, tag autopipeline-v0.1.4), 0.1.5 submission/progress-2026-10-fix5 (9e50d61), 0.1.6 submission/progress-2026-10-fix6 (8096b46), 0.1.7 submission/progress-2026-10-fix7 (24572e6), 0.1.8 submission/progress-2026-10-fix8 (bf807b6).
  • The package carries its own kit package, a different and larger one than main's kit/: autopipeline is not a command of main's kit.
git fetch origin submission/progress-2026-10-fix9
git show origin/submission/progress-2026-10-fix9:submissions/progress-2026-10-fix9/autopipeline/README.md
git clone --branch submission/progress-2026-10-fix9 --depth 1 https://github.com/ChaseHendrick/Scrolls.git scrolls-autopipeline

How to run it

From the released entry, then the commands. On 2026-10-25 the demo printed DEMO PASS here with 0.1.9 (the release check's scratch copy, synthetic scroll, stand-in VC3D tools, no network), and autopipeline defaults ran from outside the folder after a fresh pip install. Since 0.1.5 it prints DEMO PASS only when both surfaces are accepted, both fake ink runs exit 0 and all six forward, reversed and shuffled map files are nonempty inside the run folder; otherwise DEMO FAIL with the log to read.

git clone --branch submission/progress-2026-10-fix9 --depth 1 https://github.com/ChaseHendrick/Scrolls.git scrolls-autopipeline
cd scrolls-autopipeline/submissions/progress-2026-10-fix9/autopipeline
python3 -m venv .venv && . .venv/bin/activate      # Python 3.11 or newer
python -m pip install -e .                          # numpy, tifffile, zarr 3, scipy
python scripts/demo_synthetic.py                    # every stage on a synthetic scroll; prints DEMO PASS
(cd /tmp && scrolls-kit autopipeline defaults)    # the install check: works from any folder

Demo output (synthetic inputs, this page's run):

render     done
ink        done
score      skipped
report     done
surface t00: accepted=True flatten=flattened render=yes ink maps={'fake-ink': ['f', 'r', 's']}
surface t01: accepted=True flatten=flattened render=yes ink maps={'fake-ink': ['f', 'r', 's']}
plan digest 553fb8d2387663d0..., run folder /tmp/autopipe-demo-x4yc4frc/run, 4.7 s
DEMO PASS

f, r and s are the forward, reversed and shuffled maps.

Since villa commit e1aa625a6 (29 October 2026 UTC), villa builds no longer include vc_flatten or vc_tifxyz_selfcross, which the flatten stage and qa's self-intersection check call; villa's latest AppImage is now built after that commit, and its stable AppImage predates vc_tifxyz_selfcross. AutoPipeline 0.1.11 runs on those builds: its self-intersection check compiles the same census code from villa's source once (scrolls-kit selfcross build), and its flatten stage uses villa's flatboi. (package)

A real run on a released scan:

APPIMAGE="$(python3 scripts/get_vc3d.py)" && export VC_APPIMAGE="$APPIMAGE"   # villa's current Linux AppImage, sha256 checked
scrolls-kit vc3d selftest                                                    # renders and traces synthetic volumes; PASS
scrolls-kit autopipeline plan SAMPLE --settings '{"purpose": "benchmark", "segment": "0841-w00"}'
scrolls-kit autopipeline run work/autopipe/FOLDER [--until STAGE] [--monitor]
scrolls-kit autopipeline status work/autopipe/FOLDER
  • SAMPLE: a scroll id as the open-data catalog names it, such as PHerc0841.
  • FOLDER: the run folder plan printed, <sample>-<volume>-<digest10>; a relative folder is read from the kit home.
  • STAGE: the last stage to run, one of the 11 stages above.

Usage blocks, as --help prints them on the branch, shown with scrolls-kit (the program itself prints python -m kit, the branch's own module):

usage: scrolls-kit autopipeline [-h] {plan,run,status,defaults,monitor} ...
usage: scrolls-kit autopipeline plan [-h] [--settings SETTINGS] [--out OUT] sample
usage: scrolls-kit autopipeline run [-h]
                                    [--until {acquire,umbilicus,seeds,trace,bench,qa,flatten,render,ink,score,report}]
                                    [--monitor [PORT]]
                                    out
usage: scrolls-kit autopipeline status [-h] out
usage: scrolls-kit autopipeline monitor [-h] [--port PORT] [--host HOST]
                                        [--snapshot FILE.html] [--open]
                                        [--thumbs {auto,on,off}]
                                        [--stale-minutes STALE_MINUTES]
                                        [--text] [--width WIDTH]
                                        [--color {auto,on,off}] [--dither]
                                        [--watch SECONDS]
                                        out
Action and option Default Meaning
defaults Print the default settings as JSON (DEFAULTS) and exit 0
plan sample required Freeze a plan for one scan: resolves the scan, surface prediction, normal grids and axis
plan --settings SETTINGS none A JSON object, or @file, of settings over the defaults; an unknown key is refused
plan --out OUT work/autopipe/<sample>-<volume>-<digest> Run folder
run out required Run or resume the planned run in this folder
run --until STAGE none (all stages) Stop after this stage
run --monitor [PORT] off; port 8765 when given bare Serve the live monitor page on 127.0.0.1 while the run lasts
status out required Each stage's status and each surface's result
monitor out required Run folder (a relative one is read from the kit home)
monitor --port PORT 8765 Port; 0 takes any free one
monitor --host HOST 127.0.0.1 Address to listen on (this machine only)
monitor --snapshot FILE.html none Write one self-contained page and exit, no server
monitor --open off Open the page in the default browser
monitor --thumbs {auto,on,off} auto Ink map thumbnails; auto shows them for smoke runs and published labeled segments only
monitor --stale-minutes STALE_MINUTES 30 With no update for this long a running stage looks stopped
monitor --text off Print the run as text, ink maps drawn in characters, and exit
monitor --width WIDTH 40 With --text, a map's width in characters
monitor --color {auto,on,off} auto With --text, gray half blocks; auto is on in a terminal unless NO_COLOR is set
monitor --dither off With --text and no color, ordered (Bayer) dithering
monitor --watch SECONDS none With --text, print again every SECONDS (at least 1) until Ctrl-C

The settings themselves (umbilicus, seeds, trace, qa, flatten, render, ink, score, strict and others) are in section 6 of the reference; print them with defaults. The main environment variables:

Variable Meaning
KIT_HOME Where runs, caches and experiments go; default the package folder
KIT_INK_PYTHON A Python with villa's vesuvius and torch; without it a run stops at ink and writes modal-job.json (stage state handoff)
KIT_AUTOPIPE_DEVICE cpu (default; hides CUDA from the ink engine), cuda, cuda:N, or mps with a TorchScript export
KIT_AUTOPIPE_RUNNER modal marks a process as the remote runner where a handed-off stage runs
KIT_SURFBENCH_REFS Winding reference folder for the bench stage
KIT_ALLOW_MODAL 1, set by the owner, lets kit run anything on Modal; without it nothing runs there
VC_APPIMAGE, VC_BIN, VC_ROOTFS Where VC3D's tools are; the full list, with the relay, cache and thread variables, is in section 7 of the reference

Inputs and outputs

Inputs: OME-Zarr scans and surface predictions, normal grids, published tifxyz segments, and ink label zarrs for benchmarks, all from the open-data bucket; settings as JSON. Outputs: tifxyz surfaces, zarr surface volumes, TIFF ink maps and JSON records, all in the run folder (section 8 of the reference lists every file).

File in the run folder What it holds
plan.json version, created, settings, resolved, digest; never edited
state.json stages, surfaces, running, planned_traces, trace_failures; each surface's qa result with reasons and unmeasured
events.jsonl Progress and stage events, rotated at 256 KiB to events.1.jsonl
ink/<sid>/<model> Forward, reversed and shuffled maps per model and surface
scores.json, report.json AUCs of forward and control maps on published labels; the report with its digest
triage.json With triage, a reading order of maps; on a scroll without published labels it stays in the git-ignored run folder
modal-job.json The ink job, when ink is handed off

plan, run and status print a JSON object on stdout and progress lines on stderr.

Exit code Meaning
0 Done, or defaults printed
1 The run's result holds failed: a stage failed (read state.json)
2 A tool error, printed on stderr, such as Error: FOLDER has no plan.json; make one with action plan (seen on 2026-10-16 with a folder that held no plan), or a usage error

Common failures and their fixes (section 13 of the reference has the rest): the plan's digest does not match its content (plan.json was edited: make a new plan); <out> holds a run of another plan (use a new folder); unknown plan setting <key> (compare with defaults); ink stays handoff (set KIT_INK_PYTHON); no surface passed the acceptance rule (read qa.reasons and qa.unmeasured; the stable AppImage lacks vc_tifxyz_selfcross and vc_gen_umbilicus); a VC3D tool exits 139 (start it through the AppImage loader: set VC_APPIMAGE); remote renders fail on 0.1.3 (upgrade to 0.1.7, the current release; 0.1.4 fixed it); the demo prints DEMO FAIL (a dependency is missing or a fake ink command failed: read the log it names); the demo prints DEMO FAIL with got 0 surfaces and state.json says a folder is inside the repository but not ignored by git (--out named a folder inside the clone that git would track: use --out work/demo-run, or no --out for a temporary folder); the monitor's Prepare resume command is unavailable (the run is active, finished, unreadable or refused by its plan's rules, or the view is paused or older than 30 seconds: press Refresh now).

Why it is not on the default branch

It is a reviewed contribution package. 0.1.0 to 0.1.4 were released by the Progress Prize submission workflow: the owner approves an exact list of files, a manifest binds their hashes, and the release step puts them on a new public branch and checks the commit before its URL goes in the prize form (SUBMISSION.md on the branch). 0.1.5, 0.1.6 and 0.1.7 are ordinary software releases, not new prize entries: the same kind of reviewed, hash-bound file list, on a branch whose one commit has no parent, with release.json recording each file's hash. A released branch never changes; each fix goes out as a new reviewed package under a new submission id (progress-2026-10-fixN), which is why 0.1.0 to 0.1.7 each have a branch. The package also carries its own kit package, which would clash with main's kit/. main holds only its reference (docs/autopipeline/README.md), its README entry, and the every-scan tools (python -m kit run all, run watch, run catalog), which plan and price runs and launch nothing.

For agents

  • Revive: Do not merge the branch into main: its kit/ would overwrite main's. Use it from a clone of the branch, as above. A change goes out as a new reviewed release under a new submission id, with its README entry replaced in place by python -m kit releases add (the releases rule, hard rule 15); never edit the released branch.
  • Check first: That python scripts/demo_synthetic.py prints DEMO PASS (since 0.1.5 it fails when a fake model fails or a map is missing) and scrolls-kit vc3d selftest passes with the AppImage you have; that the villa build takes the URL form of --volume (scrolls-kit vc3d call status shows volume_url); qa.unmeasured before trusting a rejection; that a score's reversed and shuffled controls sit near 0.5 and below the forward map; whether the scored segment is held out from the reader (PHerc0139 w045 is not held out from ink_9um, d9v2 or Reader v2).
  • Never: Edit plan.json or state.json (a changed rule is a new plan in a new folder); call an ink map or a triage order a reading or a finding; paste a command the monitor prepared without reading it first (it prepares commands for review and never runs them); copy a run folder, map, triage list or snapshot of a scroll without published ink into a public repository; launch a paid batch or a Modal job without the owner's yes; run GPU ink on a laptop CPU for hours.

History

  • 2026-10-25: 0.1.9 (progress-2026-10-fix9), an ordinary software release: the stage registry (kit stages, --use STAGE=NAME in plans) and 3D ink readers in the ink stage (kit ink3d, opt-in and experimental, with a moved-sheet control); the default reader stays Reader V3 and the ink Space is unchanged (2144cdf, tagged autopipeline-v0.1.9).
  • 2026-10-25: 0.1.8 (progress-2026-10-fix8), an ordinary software release: Reader V3 the default reader (reader-v3 in villa's fetchable models; the koine route renamed reader-v3-koine), the ink Space on Reader V3 with 384 px patches (bf807b6, tagged autopipeline-v0.1.8).
  • 2026-10-24: 0.1.7 (progress-2026-10-fix7), an ordinary software release: ink maps in color in the monitor and the ink Space (one lightness-even ramp from dark umber to pale papyrus, cropped to the surface, enlarged smoothly, control on the forward map's scale), a synthetic demo with planted pen strokes, center lines v2.4 with the polished Center Finder line first (24572e6, tagged autopipeline-v0.1.7).
  • 2026-10-24: 0.1.6 (progress-2026-10-fix6), an ordinary software release: the monitor on one screen with a console (notes and background commands, loopback only), qa failure reasons, opt-in steps (Crackle Check gate, published center lines, axis backends, region crop, edge extension, stack check, trace budgets, z-scored ink ensembles), backend swap and provenance commands, built-in inkfilter and reader-v3 models; qa.repair without a method now uses the barrier repair (8096b46, tagged autopipeline-v0.1.6).
  • 2026-10-18: 0.1.5 (progress-2026-10-fix5), an ordinary software release: keyboard-accessible monitor controls (stage inspection, log filters, viewer pause and refresh, run-record list, commands prepared for review and copying); a synthetic demo that fails when a fake model fails or a map is missing; cache-repair times at full precision; portable test fixtures for hosts with an ancestor Git checkout and for MCP SDK 1.x (9e50d61).
  • 2026-10-15: 0.1.4 (progress-2026-10-fix4): remote renders pass the URL as --volume on villa builds since ba2bf75; monitor snapshots keep no piece of the run folder's path; real paths keep private lists out of tracked folders on macOS; four macOS-only test failures fixed (603fe14).
  • 2026-10-15: 0.1.3 (progress-2026-10-fix3): Reader v2 the default reader; ink.orient fallback; faster remote reads; Docker build and AppImage download fixed (bd7d6cf).
  • 2026-10-14: 0.1.2 (143c54b) and 0.1.1 (1534136), fix releases.
  • 2026-10-14: 0.1.0, the first public release (faada83).

Related

  • docs/autopipeline/README.md: the full reference, sections 1 to 19, including every setting and the every-scan catalog, batch and watcher.
  • docs/autopipeline/AGENTS.md: rules for agents that describe or run it.
  • batch: python -m kit run all, run watch and run catalog on main, which plan AutoPipeline over every bucket volume and launch nothing.
  • auc: the AUC score with the reversed control, as main's kit computes it.
  • releases: how the README's "Released tools" entry is written and replaced.
  • unroll: another released tool, listed under AutoPipeline in the README.