archforge 0.6.1

Preflight quality linter for AI-generated PowerPoint: catches silent font fallback (Hangul-deep, CJK-aware), tracking damage, text collisions, off-canvas bleed, and AI-generated deck tells in built .pptx files

The preflight linter for AI-generated PowerPoint.

Catches silent font fallback, unreadable sizes, colliding frames, off-canvas text, and AI-tell punctuation in built .pptx files, before a human ever sees a render.

Quickstart · What it catches · CI · Calibration record · Discussions · 한국어 README

AI agents / LLMs: read llms.txt, or pip install archforge then archforge skill --install to teach your agent the build-lint-fix loop.

PowerPoint opens both of these decks without a single warning. One of them is broken:

Code review cannot see any of it, because the defects live in font slots, autofit scales, and coordinates that only materialize at render time. Archforge reads the .pptx itself (XML, font-resolution chain, geometry, image alpha), so it needs no PowerPoint installation and runs anywhere your agent or CI runs.

30 seconds

pip install archforge
archforge demo        # builds broken.pptx + fixed.pptx and lints both, in front of you

Then point it at your own deck:

archforge deck.pptx                 # objective defects only (core profile, the default)
archforge deck.pptx --profile full  # + AI-tell / style rules: machine-made decks want this
archforge deck.pptx --json          # machine-readable JSON (agents / CI)
archforge scan decks/ --profile full   # many files, directories, or globs in one run

The decks in examples/ demonstrate the flagship defects and the profile split, each with expected outputs.

Why

The worst pptx defects are silent. No error is raised when:

• text lands on a font that lacks its glyphs and silently falls back to an OS default (the classic case: CJK text on a Latin-only font)
• positive letter-spacing quietly wrecks CJK character spacing
• autofit shrinks text below readable size
• text frames collide, or glyphs run off the canvas

These are exactly the defects machine-generated decks produce, and exactly the ones an LLM cannot see in its own output. Archforge is the gate between "the build succeeded" and "a human would sign off on the render."

Usage

archforge deck.pptx --fail-incomplete   # incomplete checks (W18) fail: recommended in CI
archforge deck.pptx --fail-on-warning   # WARNs fail too
archforge deck.pptx --e2-no-exemptions  # E2 numeric-range/minus exemptions off (full profile)
archforge deck.pptx --strict            # union of the three flags above
archforge deck.pptx --ghost         # per-page title list (horizontal-logic review)
archforge deck.pptx --render pages/ # add on-image contrast check (W7) from p01.png-style renders
archforge deck.pptx --skip W14,W6   # suppress specific WARNs (recorded in JSON)
archforge deck.pptx --lang en       # report language (default: ARCHFORGE_LANG, then OS locale)
archforge deck.pptx --no-config     # ignore config files (linting untrusted decks)
archforge deck.pptx --sarif out.sarif        # SARIF 2.1.0 (GitHub code scanning)
archforge deck.pptx --write-baseline bl.json # adopt an existing deck as-is (beta)
archforge deck.pptx --baseline bl.json       # report only new findings after that
archforge deck.pptx --hard-min 5 --body-min 9 --small-min 7.5   # size gate thresholds
archforge deck.pptx --w6-sim 0.95 --w6-cluster 5                # W6 repetition thresholds
archforge scan out/**/*.pptx --json          # aggregated JSON; exit 1 if any file fails
                                             # (an input matching nothing exits 2)
archforge rules                              # one-line summary of every rule
archforge explain W15                        # what a rule means and how to fix it
archforge demo --dir tour                    # regenerate the demo pair anywhere

Project defaults live in .archforge.json (or .archforge.yml with pip install archforge[yaml]) next to the deck or in the working directory; CLI flags override the config file, and the applied config path is always visible in the output.

{ "profile": "full", "skip": ["W14"], "baseline": ".archforge-baseline.json" }

JSON output (single file; scan --json wraps one of these per file plus an aggregate summary):

{
  "schema_version": "1.0",
  "tool": { "name": "archforge", "version": "x.y.z" },
  "target_renderer": "powerpoint-windows",
  "file": "deck.pptx",
  "lang": "en",
  "errors":   [{ "page": 3, "code": "E1", "message": "...", "detail": "...",
                 "location": { "shape_id": 7, "shape_name": "TextBox 6",
                               "bbox": [1.0, 2.4, 5.0, 1.0], "paragraph": 0, "run": 1,
                               "part": "/ppt/slides/slide3.xml" } }],
  "warnings": [{ "page": 5, "code": "W15", "message": "...", "detail": "...",
                 "location": { "shape_id": 4, "bbox": [1.0, 2.4, 3.1, 0.4],
                               "related": { "shape_id": 9, "bbox": [1.2, 2.5, 3.3, 0.4] } } }],
  "ghost":    [{ "page": 1, "title": "..." }],
  "summary":  { "error_count": 1, "warn_count": 2, "pass": false, "incomplete": false,
                "profile": "full", "skipped_codes": [], "baseline_suppressed": 0,
                "config": null }
}

location is the auto-fix target for agents: shape id/name, absolute bbox in inches (group transforms applied), paragraph/run indexes, table cell as [row, col], field: true for auto fields (slide numbers, dates), and for pair findings (W15/W17) a related counterpart. Rule codes are stable, language-independent identifiers; messages follow the report language (--lang > ARCHFORGE_LANG > OS locale), which follows the user, not the deck.

CI

GitHub Action (composite). Pinning the action tag pins the linter: by default it installs the exact source checked out at that ref, not whatever PyPI's latest is. Deck-folder config files are ignored (--no-config) and incomplete checks fail (fail-incomplete: true) unless you opt out, so a PR cannot weaken the gate by committing a config next to its deck. files takes one path, directory, or glob per line; globs are expanded by archforge scan itself, so paths with spaces and ** both behave.

jobs:
  deck-lint:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: Love-Ash/archforge@v0.6.0
        with:
          files: |
            decks/
          profile: full
          sarif: archforge.sarif
      - uses: github/codeql-action/upload-sarif@v3
        if: always()
        with:
          sarif_file: archforge.sarif

pre-commit:

repos:
  - repo: https://github.com/Love-Ash/archforge
    rev: v0.6.0
    hooks:
      - id: archforge
        # args: [--profile, full]

What it catches

ERRORs block shipping (exit 1):

Code	Meaning
E1	The font that will actually render Hangul text is Latin-only: silent Malgun fallback. Resolution follows a measured PowerPoint model (see below)
E2	Dash-family characters used as sentence punctuation (the top AI-generated-deck tell). Numeric ranges (2020 to 2024 with an en dash, Q1 to Q3, 5% to 10%) and minus signs pass by default; --strict blocks everything
E3	Effective size below 5pt after autofit and the full placeholder inheritance chain: unreadable
E4	Positive tracking on consecutive Hangul/Hanja: letter-spacing damage (kana-containing runs are exempt; tracked kana is legitimate Japanese practice)

WARNs are advisory:

Code	Meaning
W1	Body-class frame below 9pt
W5	No font size anywhere in the inheritance chain
W6	Same layout skeleton on 4+ pages (tunable; template systems: tune or skip)
W7	Low text-over-image contrast (needs --render)
W8	Small CJK in narrow frames (device mockups, cards)
W9	Accent vertical bars repeated as list markers
W10	Hand-drawn diagram cloned across pages
W11	AI-tell copy: buzzwords, stock openings
W12	Footer baseline drift
W13	Native PowerPoint shadow/glow/3D effects
W14	Titles are nominal phrases, not claims (Korean heuristic; numeric titles count as claims)
W15	Estimated text-on-text overlap
W16	Text glyphs or picture ink off-canvas
W17	Text straddling an image ink edge
W18	Some spans could not be checked (malformed input): results incomplete. Fails under --strict

Profiles separate objective defects from style policy, and since 0.4.0 the default is core: only the mechanical gates (E1/E3/E4, W1/W5/W7/W8, W15-W18) run unless you opt in. full adds the AI-tell and convention rules (E2 dashes, W6 repetition, W9-W14) and is the right mode for agent build-loops linting machine-generated decks; editorial drops W6/W14 for editorial and portfolio decks. Excluded rules are not merely hidden, they are not executed, and every choice is recorded in the JSON summary, so nothing is silently bypassed.

How it works

The E1 font-resolution model is measured, not guessed from the OOXML spec. Probe decks rendered through PowerPoint COM (pixel-compared, serif vs sans) established the actual priority: run a:ea > paragraph pPr/defRPr > lstStyle inheritance chain (shape > layout placeholder > master placeholder > master txStyles > defaultTextStyle) > theme ea (majorFont for title placeholders, minorFont otherwise; a non-empty theme ea beats run a:latin) > run a:latin only when the theme ea slot is empty > OS fallback. The record lives in docs/CALIBRATION.md. Themes resolve per slide master, so multi-master decks are judged against the theme they actually use.

Effective sizes walk the same inheritance chain down to defaultTextStyle, so size gates work on template/placeholder decks where nothing sets explicit sizes. Text in auto fields (a:fld: slide numbers, dates) passes the same gates as regular runs, and line breaks (a:br) count as line breaks in punctuation context.

Geometry checks (W15-W17) approximate effective glyph and image-ink areas: per-run sizes, real line spacing, autofit (including percent-string form), wrap mode, group transforms, alignment, image alpha trim, crop and flip. Intentional composition (drop caps, echo typography, decorative bleed, captions on cards) is excluded.

Thresholds are calibrated against rendered output of a ~50-deck real corpus and hardened with adversarial reproduction fixtures. Methods and per-gate rationale: docs/CALIBRATION.md.

Malformed input never dies silently: guards are per-run and per-slide, and anything a guard skips surfaces as W18 in the JSON output plus a machine-readable summary.incomplete flag. summary.pass reflects the active failure policy (recorded in summary.policy): with the defaults it covers ERRORs only, so gate CI with --fail-incomplete (the GitHub Action does this for you) and key on summary.pass.

Scope notes. Font-coverage knowledge (E1/E4) is currently Hangul-deep: runs written in other scripts are never falsely flagged (script detection is per text run, via Unicode code points), and per-script coverage tables are the extension path for JP/CN depth. Geometry estimation skips vertical text and RTL/complex scripts honestly (W18) instead of guessing. Rotated text frames are out of scope for geometry by design (rotation is overwhelmingly decorative practice) and do not mark the result incomplete. The render model targets PowerPoint for Windows; other renderers (Mac/web/LibreOffice) may resolve fonts differently.

Agent integration

Designed for LLM-agent build-lint-fix loops:

build deck.pptx
loop:
    result = archforge deck.pptx --profile full --json   # machine-made decks: AI-tell rules on
    if result.summary.error_count == 0 and not result.summary.incomplete: break
    fix listed defects (location payloads point at the exact shape/run), rebuild
review WARNs against renders

The Agent Skills pack (standard SKILL.md + YAML frontmatter) teaches this loop and per-code fixes to any supporting agent (Claude Code, Codex, ...). It ships inside the wheel: archforge skill --install. If you cloned the repo, skills/archforge-pptx-lint/ is the same file.

A passing lint is not a finished deck: the linter owns the mechanical defect class; composition and narrative still need eyes on renders.

Community and contributing

• Found a false positive? Report it with the FP template: a repro deck makes it a permanent regression fixture, the most valuable contribution this project takes.
• Questions, ideas, decks you are unsure about: GitHub Discussions.
• Want to contribute code? CONTRIBUTING.md explains the evidence bar (gates are calibrated against renders, not taste); issues tagged good first issue are scoped for a first PR.
• Security: SECURITY.md.

Name

archforge = arch (structure) + forge. A forge where a deck's structure and typography get hammered straight before shipping.

Author

Built and calibrated by Minjae Kwon (Ash) (@Love-Ash · LinkedIn). If archforge caught something before your audience did, a star helps the next person find it. I write up the measurement work behind the gates (how PowerPoint actually resolves fonts, and what AI-built decks silently break); say hi on LinkedIn.

License

MIT © Minjae Kwon (Ash)