AnchorsFactory

Rule-driven anchor placement for UFO fonts. You describe, in a compact text file, where anchors should sit on your glyphs; AnchorsFactory computes the coordinates from each glyph's own geometry and writes the anchors into the font.

It does the pre-marking step of accent handling: place top/bottom/_top/… anchors consistently across hundreds of glyphs, so a tool like GlyphConstruction can then assemble composite glyphs by snapping mark anchors to base anchors.

Install

pip install anchorsfactory
# or, from a checkout:
pip install -e .

Requires Python 3.10+, fontParts and fontTools.

Quick start

# place anchors using a rule set, save to font_anchored.ufo
# (none ship with the package — start from a copy of examples/rules/default.anchors)
anchorsfactory MyFont.ufo --rules my-rules.anchors

# overwrite in place, with a backup of existing anchors
anchorsfactory MyFont.ufo --rules my-rules.anchors --in-place --backup-dir backups/

# a whole folder of UFOs, addressing the set by name off a rule library
anchorsfactory masters/ --rules-path my-rules --rules default

By default the source UFO is never overwritten — output goes to *_anchored.ufo unless you pass --in-place.

The rule language

Rules are stacked: define reusable labels (@, a list of anchors) and variables (&, one axis's value), then mark glyphs with them, mixing them with one-off anchors. An anchor is a name and a parenthesised X Y placement.

# a variable (reuse one axis value) and a label (reuse a set of anchors)
&mid = capHeight*1/2+xHeight*1/2
@ = top (box.center capHeight), bottom (box.center 0)

# apply it, by name / unicode / range
A = @, ogonek (outline.right 0)
U+0410..U+044F = @            # all Russian Cyrillic
U+0413 += desc (outline.right 0)   # Г also gets a descender anchor

• frame.position works on both axes: width.* (advance), box.* (bounding box) or outline.* (the contour, sampled on a scanline at the other coordinate, with .first/.last to pick a stem). On X the words are left/center/right, on Y bottom/middle/top; box*1/3 is a fractional position (same *n/m as capHeight*2/3) and outline.centroid the area centre of mass.
• Y may also be a number, a font metric (capHeight, xHeight, …), or a reference glyph ($H, $H.bottom, $H*5/6); box.top reads this glyph's own height.
• Selectors: name, U+XXXX, range U+A..U+B, glob *.sc, category {Lu}.
• = replace · += add · -= remove; !extends default inherits a ruleset.
• Terms combine with +/- (on either axis) for a base position plus a bias — outline.centroid-25 nudges a slanted mark off the optical centre.
• &name names a reusable X/Y value; @name a reusable set of anchors. Both are late-bound, so an extending file can override either.
• %name places an anchor relative to another anchor on the same glyph — bottom (%top 0) reuses top's x so the two don't drift apart.
• compN./complast. measure one component's outline (comp1.outline.center@top), for seating marks on ligature parts.
• !propagate = composites makes composite glyphs inherit their components' anchors — write rules for the base letters, precomposed glyphs get theirs free.
• @top-10 / @bottom+8 sample just inside an edge (the fix for a scanline grazing a smooth peak); @xHeight-20 and other metric/variable insets too.

Full reference: docs/anchor-rules.md.

Rule sets: the host carries them

No rule sets ship with the package. Rules are data, not code: bundled, they sat where a user could not edit them, and a change of data went out on a version of code. The library carries the engine and a resolver; you carry the rules.

The sample sets live in this repository under examples/rules/ — copy them and make them yours. Address one by path, or by bare name once a search path is configured:

anchorsfactory MyFont.ufo -r path/to/default.anchors           # by path
anchorsfactory MyFont.ufo --rules-path my-rules -r default     # by name
export ANCHORSFACTORY_RULES_PATH=my-rules                      # or by environment

A name is also looked up next to the file doing the referencing, so a set can !extends default its neighbour with no configuration at all. In Python, load_document(ref, base_dir=…, search_paths=[…]); process-wide, anchorsfactory.set_search_paths([…]) / add_search_path(dir). is_name(ref) tells a bare name from a path without touching the disk.

resolve_ref(ref, base_dir=…, search_paths=…) answers which file a reference loads from — the same decision load_document makes, and the same string that lands in RuleSource.origin, so a host that lets a user edit an inherited set writes back to the file the engine actually read. A name nothing answers raises RuleSetNotFound (a KeyError), kept distinct from the FileNotFoundError of a mistyped path and the DSLError of bad rule text — three different things to tell a user. Nothing is cached: a reload sees an edit made mid-session.

default and default-italics are the samples to start from. The italic one is an overlay: the shear itself is automatic, so it only restates what a slanted drawing changes — eight rules on top of default rather than a second copy of it.

examples/rules/default.anchors covers Latin core — Basic Latin, Latin-1 Supplement and Latin Extended-A — and is generated from the Unicode database: every composite comes from a canonical decomposition, and every base carries exactly the anchors those composites ask of it. It ships with a matching default.glyphsConstruction, so the two halves of the pipeline stay in step — AnchorsFactory places the anchors, GlyphConstruction assembles the composites.

Rules address glyphs by codepoint (U+00C1) wherever the name would be a guess, so they make no assumption about a font's naming scheme; the ASCII letters, which every font names alike, stay spelled out as A–Z / a–z.

Accents get both treatments, because fonts ship them both ways: each mark is ruled on its combining codepoint and its spacing twin (U+0301 and U+00B4), and again under the legacy names a font may use for glyphs that carry no codepoint at all — acutecomb, and the capital-height set Acute / acute.case, which is anchored at cap height rather than the lowercase midpoint.

Wider Latin sits alongside it: examples/rules/latin-ext-b.* and latin-ext-additional.* (Vietnamese and the dot-below accents), plus the pre-generation legacy-* sets and the script sets devanagari / hebrew / thai.

Old .txt rule files (see examples/) convert to the new syntax — verified lossless:

anchorsfactory-convert examples/default-anchors-list.txt -o my-rules.anchors

Library API

from anchorsfactory import process_ufo, load_document, apply_document

process_ufo("MyFont.ufo", "my-rules.anchors")  # high-level: open, apply, save

from fontParts.world import OpenFont
font = OpenFont("MyFont.ufo")
apply_document(font, load_document("my-rules.anchors"))
font.save()

Composites (GlyphConstruction)

anchorsfactory.composites assembles composite glyphs from the anchors just placed, using a vendored, unmodified GlyphConstruction.

from anchorsfactory import load_document, apply_document
from anchorsfactory.composites import build_composites
from anchorsfactory.presets import construction_text

RULES = ["examples/rules"]                              # your rule library
apply_document(font, load_document("default", search_paths=RULES))   # anchors first
built = build_composites(font, construction_text("default", search_paths=RULES))
built["Aacute"].glyph        # a ConstructionGlyph: .draw(pen) / .width / .components
built["Aacute"].source_line  # the construction's line, for click-to-rule
built["Aacute"].problems     # missing base/mark glyph or anchor, if any

Constructions use stock GlyphConstruction syntax with one AnchorsFactory extension: a glyph may be addressed as U+XXXX, resolved to the font's own glyph name before the engine sees it (the font's cmap first, then a combining mark's spacing twin, then the AGL name, then uniXXXX). Plain names still work and are the better choice where a name is unambiguous.

A mark may take a .case suffix, asking for the font's capital-height accent set — a legacy cut that exists under a name (Acute, acute.case) but never a codepoint. Fonts without one fall back to the plain mark, so the same file builds either way:

U+00C1 = A + U+0301.case@top   | 00C1  # Á · LATIN CAPITAL LETTER A WITH ACUTE
U+00E1 = a + U+0301@top        | 00E1  # á · LATIN SMALL LETTER A WITH ACUTE

That keeps a rule set portable across naming schemes, at the cost of stock GlyphConstruction not reading such a file — call anchorsfactory.composites.resolve_unicode_refs(text, font) to get plain-name text back. Files with no U+ token are untouched, so name-based constructions keep working unchanged.

Compute without mutating (preview)

compute_document is the functional core: it returns the anchors a rule document would place, keyed by glyph, without touching the font. It owns the same orchestration as apply_document (suffix expansion, shift_x, rounding, same-name dedup), so a preview never drifts from what gets written.

from anchorsfactory import compute_document, load_document, accumulate, resolve

doc = load_document("my-rules.anchors")
placed = compute_document(font, doc)          # {glyph_name: [(anchor, x, y), ...]}

# resolve() is the pure, single-anchor primitive behind it:
specs = accumulate(doc, "A", font["A"].unicodes)   # the anchors due on glyph A
x, y = resolve(font, font["A"], specs[0])

For an interactive editor, on_error="collect" never raises — it places what it can and reports the rest as structured diagnostics:

result = compute_document(font, doc, on_error="collect")
for d in result.diagnostics:
    # d.severity == "error"   -> anchor skipped (geometry raised)
    # d.severity == "warning" -> anchor placed via a fallback, but suspect
    print(d.severity, d.glyph, d.anchor, d.reason)

result is a plain dict subclass (so it works anywhere a dict does) carrying a .diagnostics list of ComputeDiagnostic(glyph, anchor, reason, severity, rule). The default on_error="raise" is unchanged and keeps .diagnostics empty.

Editor support (anchorsfactory.vocabulary)

Everything an editor needs to highlight and complete the rule language, taken from the same tables the parser is built from — so a client never hard-codes a word the parser might later reject.

from anchorsfactory.vocabulary import (
    FRAMES, X_ALIGNS, Y_EDGES, RUNS, METRICS, CENTROID,
    DIRECTIVES, PROPAGATE_VALUES, SUFFIX_KEYWORDS, OPERATORS, SIGILS,
    completions_after_dot, completions_for_slot, as_dict,
)

completions_after_dot("box", axis="y")     # ("bottom", "middle", "top")
completions_after_dot("width", axis="y")   # () — the advance frame is X-only
completions_for_slot("x")                  # what may open the X slot (no metrics)
as_dict()                                  # the whole thing, JSON-serialisable

Both axes share one frame.position grammar but not one alignment table: left/center/right is X, bottom/middle/top is Y, and a font metric only opens a Y slot. Passing the slot the caret sits in therefore narrows a menu from six words to three; pass axis=None when the slot is unknown and both come back.

Note it is deliberately not the IR: Frame.ADVANCE is a node name, the word a user types is width. A client completing straight off anchorsfactory.model gets the former and offers a token the parser rejects.

Development

make venv      # create .venv, install the package (editable) + dev deps, via uv
make test      # run the test suite
make build     # build sdist + wheel into dist/
make release   # bump minor, update CHANGELOG, build, upload to PyPI, tag + push

Releases are cut with make release: it bumps the minor version, fills in a CHANGELOG section from the commit log, publishes to PyPI, and tags vX.Y.Z. Set UV_PUBLISH_TOKEN first; use make release-test to rehearse against TestPyPI.

License

MIT — see LICENSE.