osr-forge

Convert tabletop adventure module PDFs into playable osrlib Adventure documents: an LLM-assisted extraction pipeline, deterministic map-geometry synthesis, a human correction loop built on an overrides file, and validation against the real osrlib models.

Standalone package + CLI. Consumers need only its artifacts — adventure.json, report.json, overrides.yaml, SVG map previews — or its CLI, regardless of their own tech stack.

Status: released — install from PyPI, docs at https://mmacy.github.io/osr-forge/:

uv pip install osr-forge     # or: pip install osr-forge

The package runs the whole pipeline — preprocessing, the survey (chunked past survey_max_pages) and content extraction stages, monster resolution against the osrlib catalog, deterministic geometry synthesis (doors and secret doors on the stating room's wall, stair and trapdoor transitions from keyed and level-shaped links, guessed landings flagged), and assembly — plus the human correction loop (overrides application, rerun/resume, the playability lint with its smoke delve, cost estimation) and the eval harness that keeps extraction quality a measured number (tools/eval/). The roadmap's phase plans live in docs/ beside the specification.

The library entry points are convert(), assemble(), check(), and estimate(); the osrforge console script wraps them:

osrforge estimate my-module.pdf   # preprocess only; rough token/cost estimate
osrforge convert my-module.pdf    # full pipeline into ./my-module.forge
osrforge assemble                 # stage caches + overrides → artifacts, pure
osrforge check                    # validate_adventure + the playability lint
osrforge report                   # summarize report.json — the review loop's first read
osrforge preview                  # regenerate previews/ (SVG maps + index.html) only
osrforge rerun assemble           # resume any stage through assemble

The workdir-taking commands discover a missing --workdir on their own: the working directory when it is itself a workdir, else the unique *.forge directory within it (here, ./my-module.forge) — several or none is a loud error, and an explicit --workdir always wins.

Recording sessions and live verification runs are driven via tools/extract/run_extraction.py (see tools/extract/README.md); the on-demand eval harness — corpus, scorer, and scoreboard — lives in tools/eval/ (see tools/eval/README.md). The harness also measures your own retail modules: a private corpus directory side-loads with --corpus DIR (integrity via a local hash sidecar, truth authored per tools/eval/AUTHORING.md), and an explicit publish step copies aggregate-only scores — never module text — onto the committed BYOM scoreboard.

The correction loop

Conversion produces a draft — every gap and guess called out in report.json — and corrections live in overrides.yaml, never in hand-edits to generated output (a fresh convert leaves a commented template there). The loop:

1. osrforge report (flags grouped by kind, findings, the monsters summary), then open previews/index.html — each level's synthesized map beside the module's own map-page renders, with a coordinate ruler printing the cell coordinates geometry overrides use.
2. Edit overrides.yaml: monster remaps, per-area field replacement, area adds and removes, geometry (cells, edges, entrance, transitions), town/module metadata. Every entry carries a reason, and every entry must take effect — a typo'd address fails assembly loudly instead of silently doing nothing.
3. osrforge assemble && osrforge check — re-assembly is pure and instant (no model calls), and check exits 0 once validation passes and no error-severity finding remains.
4. Repeat until publishable.

Settings changes on an existing workdir go through rerun --set — for example, osrforge rerun preprocess --set 'blank_page_renders=[21]' blanks a render Azure's content filter rejects, or osrforge rerun assemble --set unresolved_fallback=omit flips the stand-in policy without re-rolling the model.

Development quickstart

Requires Python ≥ 3.14 and uv.

uv sync
uv run pytest

The full check suite, as CI runs it:

uv run ruff format --check && uv run ruff check && uv run pyright && uv run pytest

Tests use no network — model interactions replay from recorded fixtures. The only live-network activity in the repo is manual: the Foundry capability spike (tools/spike/) and the extraction runner (tools/extract/).

Pipeline settings

ConversionSettings holds the deterministic knobs, echoed into each workdir's run.json:

On an existing workdir, change a knob with rerun --set KEY=VALUE: the update lands in the run.json settings echo before the chain runs, and a knob owned by a stage upstream of the rerun stage is rejected with the stage to rerun instead.

Provider configuration

The Azure AI Foundry adapter reads its connection from OSRFORGE_FOUNDRY_* environment variables (an osr-forge-specific prefix, deliberately not AZURE_OPENAI_*, to avoid colliding with other tools' conventions):

Licensing

Package code is MIT (see LICENSE). osr-forge ships no game content: osrlib's OGL data stays in osrlib, and test assets live under tests/assets/ outside the built distribution, each with its provenance and license documented in tests/assets/README.md.

These fences govern what this repository redistributes, not what you convert: bringing your own purchased, non-redistributable module is the primary use case. Conversion runs locally and everything derived from your module stays in your own workdir — nothing is shared unless you share it.