Items 1 & 2 of the #351 remainder. Item 1 — the flat/nested generative property tests now read SEED and NUM_RUNS from PROPERTY_SEED / PROPERTY_NUM_RUNS (via an envInt helper that honors an explicit 0 and falls back to the current defaults — 20250705/300 flat, /100 nested — on unset/empty/non-numeric). New .github/workflows/nightly-property.yml runs the generative suite daily (and on workflow_dispatch) with a random seed and NUM_RUNS≈5000; on failure it files a Gitea issue containing fast-check's shrunk counterexample (jq-escaped, dedup'd by title). No build step — the suite imports the converter from src/, so a tsc error can't masquerade as a property failure. Item 2 — new packages/prosemirror-markdown/README.md documents the counterexample process (surface -> shrink -> permanent fixture in test/fixtures/counterexamples/ + counterexamples.test.ts -> fix the converter, never weaken a property; maintainer-approved ACCEPTED/allowlist entries carry a reason) plus the two golden layers, the coverage allowlist, and how to run with the env knobs. Linked from AGENTS.md. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
5.0 KiB
@docmost/prosemirror-markdown
The single, canonical ProseMirror ↔ Markdown converter plus the Docmost schema mirror (#293/#345). Headless and framework-free: no React, no browser runtime. There is exactly ONE copy of this converter in the repo, consumed by:
packages/mcp(the MCP server),packages/git-sync(two-way Git sync),apps/server(server-side markdown import/export, #345).
src/lib/docmost-schema.ts mirrors the upstream Tiptap schema that lives in
packages/editor-ext. The mirror is not free-floating: serializer-contract.test.ts
guards the boundary — every schema node must have a converter case, so a drift
between editor-ext and this package surfaces as a failing test rather than a
silent divergence.
Why byte-stability matters
Git sync exports a page to markdown, and re-imports it on the next pull. If
export → import → export is not byte-stable, every pull rewrites files
that nobody edited, and the user's history churns with phantom diffs. So the
converter is held to more than "it roughly round-trips": the second export pass
must be a byte-for-byte fixpoint. That is what the property suite below proves.
Two golden layers (do not mix them)
-
Corpus fixtures —
test/fixtures/corpus/. A fixed, hand-curated set of representative documents (headings, marks, lists, tables, diagrams, columns, details, mentions, …). These are the readable, deterministic "known-good" snapshots. Edit them deliberately. -
Generative property suite —
test/generative/. fast-check draws random documents and asserts invariants over them. Two entry points:flat-roundtrip.property.test.ts— flat documents, attribute-level fuzzing.nested-roundtrip.property.test.ts— deeply nested structures.
The invariants (P1–P4):
- P1 — semantic round-trip:
mdToPm(pmToMd(doc))is canonically equal todoc(no data loss for the round-trip-supported space). - P2 — byte fixpoint:
pmToMd(mdToPm(pmToMd(doc))) === pmToMd(doc)(the first pass may normalize once; the second pass must be a fixpoint). - P3 — totality: neither converter throws; output is bounded.
- P4 — parser fuzz totality: for ANY input string,
markdownToProseMirrordoes not throw and returns a schema-valid document.
These invariants are kept STRICT — no
it.fails, skip, or weakening. A failure means the generator found a REAL converter bug.
The counterexample process (the DoD)
This is the point of the generative layer. When a property run diverges:
- A property run surfaces a divergence (locally, in CI, or in the nightly cron — see below).
- fast-check shrinks it to a minimal, human-readable counterexample and prints the reproducing seed.
- Commit the shrunk doc as a permanent fixture under
test/fixtures/counterexamples/, with a matching case incounterexamples.test.ts. The fixture stays forever, as a regression pin. - FIX the converter so the counterexample round-trips. Never weaken a property to hide the bug.
- If — and only if — a maintainer decides a particular markdown-representable loss is genuinely acceptable, it is recorded as an explicit ACCEPTED / allowlist entry with a written reason, not by silently relaxing an invariant.
Attribute-coverage allowlist
flat-roundtrip.property.test.ts maintains ATTR_VALUE_FUZZ_ALLOWLIST. The
suite asserts that every attribute in the live schema is EITHER value-fuzzed by
the generator OR explicitly listed in this allowlist. This forces any newly
added node attribute to be consciously classified — you cannot add an attr and
leave it silently un-exercised; the coverage test fails until you either fuzz it
or record why it is held out.
Running
# The full package suite (corpus + generative + contract tests):
pnpm --filter @docmost/prosemirror-markdown test
The generative suite honours two env knobs (invalid/empty → falls back to the default):
| Env var | Default (flat) | Default (nested) | Meaning |
|---|---|---|---|
PROPERTY_SEED |
20250705 |
20250705 |
fast-check seed (reproducibility) |
PROPERTY_NUM_RUNS |
300 |
100 |
runs per property |
# Reproduce a specific counterexample seed with a bigger budget:
PROPERTY_SEED=12345 PROPERTY_NUM_RUNS=5000 \
pnpm --filter @docmost/prosemirror-markdown exec vitest run test/generative/
Nightly cron
.github/workflows/nightly-property.yml runs the generative suite every night
with PROPERTY_NUM_RUNS cranked to ~5000 and a random seed, to reach deeper
counterexamples than a fixed-seed PR run can. On failure it files a Gitea issue
containing the reproducing seed, the run count, and the tail of the output (the
shrunk counterexample), which kicks off the counterexample → fixture process
above. It can also be triggered manually (workflow_dispatch) with custom
num_runs / seed.