Skip to content

Compat Kit

Jump to bottom
swear01 edited this page Aug 1, 2026 · 30 revisions

Auto Storage Compat Kit

Compat Kit is the development-time workflow for adding deterministic cross-mod support without repeatedly rereading an upstream repository or weakening Auto Storage's transaction boundary. It collects facts, makes unresolved semantics explicit, generates reviewed mechanical code or RED boundaries, and runs declared verification gates. It is never loaded by Minecraft.

This file is authoritative. The GitHub Wiki Compat Kit page mirrors it, but the developer page is intentionally absent from the player-manual Home contents and sidebar.

Safety boundary

Compat Kit automates evidence and boilerplate, not recipe semantics.

  • No player-runtime jar scanning, reflection fallback, serializer-name semantics, machine-name semantics, or EMI/JEI crafting logic.
  • scan structurally follows class-file inheritance for concrete Recipe and RecipeSerializer implementations. Names are only candidate evidence for resource and station surfaces; they never authorize behavior.
  • scan publishes public surfaces, bounded recipe-data structure, and risk evidence only. It also inspects bounded private bytecode to detect hidden randomness or world/entity access, but never writes that private bytecode into the audit.
  • A human or reviewing agent explicitly decides consumed inputs, catalysts/tools, remainders, all outputs, typed units, station rates, costs, deterministic bounds, and rejections.
  • needs_decision blocks scaffold, generate, conformance, and verify.
  • A loaded but incompatible integration fails explicitly. The tool never emits a compiling empty adapter or silently skips a required check.
  • A representative target version and SHA are reproducible audit evidence, not a player-facing exact dependency pin or multi-version support promise.

Use Python 3.11 or newer (tomllib is required) and JDK 21 so javap matches the project toolchain:

export JAVA_HOME=/opt/homebrew/opt/openjdk@21/libexec/openjdk.jdk/Contents/Home
export PATH="$JAVA_HOME/bin:$PATH"
python3 --version

Files and ownership

tools/compat-kit/
  compat-kit                         executable wrapper
  compat_kit.py                     standard-library CLI
  schema/*.schema.json              strict machine-readable contracts
  examples/github-actions/          reusable addon CI example
compat/audits/<mod-id>/<version>.json
compat/contracts/<mod-id>.json
compat/generation/<mod-id>.json
build/compat-kit/cache/<jar-sha>/v*/ ignored versioned scan cache
build/compat-kit/report.json         ignored verification report

The audit and reviewed contract are committed. Caches, downloaded jars, source checkouts, and transient reports remain under ignored build/.

End-to-end workflow

1. Acquire reproducible evidence

Download one official representative jar and, when available, check out its matching source tag or commit. Record the artifact URL, SHA-256, and source revision in the compatibility document. The checkout passed to scan or diff must be clean, including untracked files, so the recorded revision fully describes the inspected source. When the supplied source is a module subdirectory, Compat Kit discovers the enclosing Git worktree, requires that whole worktree to be clean, and records its HEAD while keeping evidence paths relative to the supplied module. A supplied --source must belong to a Git worktree; omit it rather than presenting uncommitted/unversioned contents as source evidence. When the target jar has classified candidates, at least one candidate must map to a tracked, non-ignored, non-symlink Java source file under the supplied module; ignored build output and a clean but unrelated checkout are rejected instead of borrowing HEAD. Matching is segment-aware: a path such as notsamplemod/recipe/CrushingRecipe.java cannot satisfy the candidate samplemod.recipe.CrushingRecipe. Never silently replace a failed source.

2. Scan

tools/compat-kit/compat-kit scan \
  --jar build/compat-kit/artifacts/target.jar \
  --source build/compat-kit/sources/target \
  --classpath build/compat-kit/classpath/dependency.jar \
  --data-root build/compat-kit/datapacks/target-override \
  --output compat/audits/target/1.2.3.json

--classpath is repeatable and supplies the target's complete compile-time ancestry without turning dependency classes into target candidates. Every non-JDK superclass and interface reachable from the target jar must resolve; an incomplete classpath fails instead of silently omitting a structurally hidden recipe family. This check runs before name-bucket classification, so a class named like a client viewer, builder, or datagen helper cannot bypass ancestry validation. Before reading platform classes, Compat Kit requires the resolved javap toolchain's release metadata to report major version 21. JDK ancestry is then resolved from that JDK's jmods inventory, not from a package-prefix allowlist, so platform classes such as org.xml.sax.* are recognized without weakening external dependency checks. At most 128 jars and 200,000 classes are read, duplicate/conflicting classes fail, and the sorted artifact SHA/size set is persisted in the audit and cache identity. Every classpath jar is rehashed after inspection just like the target; replacing a dependency during a scan fails instead of mixing two artifacts in one audit. Class-name constants are decoded as JVM modified UTF-8, so unrelated NUL or supplementary string constants cannot abort a valid scan. --data-root is repeatable and follows data-pack precedence: the target jar is the first layer and later supplied roots override earlier recipes by exact ID. Roots and recipe/tag files must be real, non-symlink paths. Every bounded data/*/tags/**/*.json resource and bounded pack.mcmeta enter each root digest without being counted as recipes; the file bound applies globally across ordered roots. Roots declaring top-level data-pack filter or overlays metadata fail because removal and overlay-directory activation are not modeled. The deterministic audit contains NeoForge identity, artifact SHA/size, source revision and exact source paths, structurally classified concrete recipe and serializer classes, recipe types, builders, datagen classes, client/viewer wrappers, block entities, resource/station candidates, per-serializer recipe counts, bounded sample IDs, top-level fields and array cardinalities, NeoForge condition types, override provenance, and explicit risk flags. Risk detection uses bounded javap -c -p output so non-public implementation risks are not missed; only compact flags for randomness/chance, world/entity access, multiblocks, live machine state, generic ingredient surfaces, unbounded outputs, and capability mutations requiring simulation are persisted. Structurally discovered recipes inspect every reachable non-JDK target/classpath superclass and interface implementation, independently from the first path used to classify the class as a Recipe; inherited findings are attributed to each concrete audited recipe. Duplicate ancestry-class definitions fail closed even when their hierarchy metadata matches, so bytecode ownership never depends on classpath order. Each private-bytecode result is reduced to compact flags before the next class is read; platform-neutral concurrent pipe readers retain at most the configured limit plus one byte and terminate javap immediately on overflow, so neither one process nor all classes can bypass the memory bound. RandomSource, ThreadLocalRandom, SplittableRandom, SecureRandom, and RandomGenerator are all randomness evidence. Without --source, scans are cached by jar SHA under build/compat-kit/cache/; repeating the same SHA and scanner format needs no network access. A scanner format change uses a new cache namespace instead of trusting stale evidence. Data-root scans bind the ordered layer digests into their cache identity; ancestry classpaths bind their exact artifact set. Cache metadata also records the selected JDK 21 installation, release version, and module-file identity. JDK validation occurs before every cache return, and an identity change forces a fresh scan rather than reusing evidence from another module inventory. The current scanner format is 12; formats 7, 8, 9, 10, and 11 remain readable only as explicit legacy evidence while committed contracts are migrated. Format 12 retains each candidate's structural classification and a separate sorted top-level structural_hierarchy inventory, requires both copies to agree, and requires hierarchy to win before every name-term bucket. Removing only one indirect path therefore cannot demote that class without conflicting with the other independently persisted record. Validation also cross-checks direct extends/implements parents in the public declaration; generic type bounds are not treated as direct ancestry. The scanner format is also stored in every audit and required by its published schema, so an older structurally valid committed audit cannot authorize current contract review. Multi-release META-INF/versions/ aliases are not treated as binary class names; the root binary name is scanned once. Chance, randomness, generic-ingredient, and capability-mutation detection read both source-like method signatures and the methodName:(descriptor) form emitted by javap -c -p. Named nested classes are audited and mapped back to their top-level Java source, including legal Java identifiers that themselves contain $. The scanner uses SourceFile, InnerClasses, and EnclosingMethod metadata rather than splitting the binary name on every $; a legal top-level Recipe$1Variant therefore maps to Recipe$1Variant.java, while anonymous/local classes and every class carrying the JVM ACC_SYNTHETIC access flag are excluded. Family IDs use normalized simple class names. A legal identifier that normalizes to an empty string uses class_<binary-name-hex> instead, while normalized collisions append the same collision-free binary-name encoding.

Archive limits, the 1 MiB per-entry NeoForge metadata limit, the 16 MiB per-entry class limit, candidate counts, signature size, source-file counts, malformed metadata, ambiguous multi-mod jars, missing JDK tools, and javap failures all fail closed. Metadata and class sizes are checked before decompression. The target path is rehashed after inspection; a replaced or mutated jar cannot combine one artifact SHA with another artifact's metadata or candidates. Every supplied ancestry jar has the same post-inspection size/SHA check. Current-format cache entries and committed audits are also fully validated down through target, artifact, source, candidate, and risk records; a matching cache path does not authorize malformed or partial evidence. Every candidate must remain in the bucket computed by the current scanner. Both scan and committed-audit validation apply the same current priority for client/viewer, builder, datagen, and structural/name classifications; moving a candidate into a lower-priority bucket cannot hide it from contract review. Every risk-evidence owner must also be one of the audited recipe candidates, so risk cannot be detached from the class under review.

Legacy format-7 and format-8 audits must be regenerated from the exact same reviewed artifact rather than edited in place:

tools/compat-kit/compat-kit migrate-audit \
  compat/audits/target/1.2.3-legacy.json \
  --jar build/compat-kit/artifacts/target.jar \
  --source build/compat-kit/sources/target \
  --output compat/audits/target/1.2.3.json

Migration rejects a current audit, a different target identity, or different artifact bytes. It performs a full current scanner pass and therefore preserves no stale legacy classification.

2a. Generate unresolved machine and requirement proposals

tools/compat-kit/compat-kit propose \
  compat/audits/target/1.2.3.json \
  --output build/compat-kit/target-proposals.json

propose converts only current-format audit evidence into a compact review surface. Public numeric methods and fields that mention time/rate/parallelism are offered as bounded rate-binding candidates; slot counts are deliberately ignored. This includes public tier fields such as Mekanism FactoryTier.processes, but choosing the corresponding tier/item mapping still requires review. Recipe-data fields are classified as transaction-representable, station-descriptor-representable, or unsupported live/world state. Every proposal and binding stays needs_decision; the output cannot authorize a contract, generate runtime reflection, or assert that a candidate machine is a valid station.

2b. Generate a dedicated-server runtime probe

tools/compat-kit/compat-kit probe \
  compat/audits/target/1.2.3.json \
  --plan compat/probes/target.json \
  --game-test-namespace target_auto_storage \
  --output build/compat-kit/target-probe

tools/compat-kit/compat-kit validate-probe \
  build/compat-kit/target-probe/runtime-probe.json \
  --audit compat/audits/target/1.2.3.json \
  --plan compat/probes/target.json

The generated GameTest is server-only and evidence-only. --game-test-namespace is mandatory, is recorded in probe-spec.json, and becomes the class's exact @GameTestHolder; pass the namespace enabled by the Gradle run that will execute it. The probe requires an explicit compatKitProbeOutput system property, sorts and records every loaded RecipeManager recipe ID/type/serializer/concrete class plus common public ingredient/result values, and records target-namespace block, item, and block entity type identities. The probe fails above 50,000 loaded recipes instead of truncating silently. It contains no client imports or reflection. --plan is optional, but when present it must match the exact current audit digest and may add reviewed direct calls for public config values or capability surfaces. The only accessor shapes are a value-wrapper static field, static method, registry block method, or enum-constant numeric field. Target config and capability candidates remain listed as unresolved in probe-spec.json until those bindings are supplied; empty arrays do not authorize those semantics. validate-probe is the mandatory cross-file gate for collected JSON: it loads the committed audit and optional exact plan instead of relying on the output schema alone. It validates the exact canonical probe-plan digest recorded as source_probe_plan_digest, and compat-runtime-probe.schema.json before using it as review evidence. Config records accept only finite JSON numbers and capability records accept only booleans; validation also requires the output IDs, sources, and capability surfaces to match the reviewed plan exactly.

3. Decide

tools/compat-kit/compat-kit decide \
  compat/audits/target/1.2.3.json \
  --output build/compat-kit/target-contract-draft.json \
  --next-actions build/compat-kit/target-next-actions.md

The draft contains one entry for every recipe-class candidate. Complete each entry as accepted or rejected; do not delete inconvenient candidates. next-actions.md is the compact review surface and lists only unresolved work. The published contract schema applies the same status-dependent boundary: accepted families require recipe type, station, nonempty inputs/outputs with a primary output, and a decision; rejected families require their decision. Once no family remains needs_decision, the schema also requires the target dependency, repositories, and runtime_dependencies scaffold inputs; RED drafts with unresolved families remain valid. The same completed-contract branch requires a non-null fixture, positive GameTest count, authoritative task, nonempty Gradle tasks, all twelve exact checks, and one evidence mapping for every check. source_recipe_inventory_sha256 binds the sorted recipe-class inventory; source_recipe_data_sha256 separately binds the effective recipe-data and data-pack override digest. Every every complete scaffold/verify invocation must also load the committed source audit. Validation compares the contract's exact family-class set, per-family scanner risk set, target identity, artifact SHA, and inventory digest with that separate audit; deleting a family or risk and recomputing a contract field still fails.

An accepted family records:

  1. exact recipe class and RecipeType;
  2. station descriptor, category, variants, and rational rates;
  3. every consumed input, catalyst, tool/durability, and remainder;
  4. every primary and secondary output with exact typed units;
  5. station-work and item/fluid/energy/chemical/addon-resource costs; use an empty costs list only when the reviewed runtime family is genuinely free;
  6. deterministic bounds and source/class/method evidence;
  7. target Maven repositories as explicit HTTPS URLs;
  8. every additional required runtime artifact as an explicit dependency;
  9. fixture, exact GameTest count, authoritative GameTest task, Gradle gates, and every required check mapped to {task, source, marker} evidence.

Process station variants require a positive rational rate. Instant station variants require a zero numerator. These rules match runtime MachineDescriptor validation and fail before scaffolding.

For a bundled module, verification.game_test_task is not arbitrary evidence: it must match the task derived from the fixture's preserved camel case (evilCraftFixture becomes runEvilCraftGameTestServer). This prevents a same-count fixture from supplying another integration's runtime result. The target mod ID must produce a valid, non-reserved Java package segment; identifiers such as class, true, or null fail before files are written.

Commit the reviewed result as compat/contracts/<mod-id>.json.

When a current scan supersedes an older audit of the exact same target and jar, migrate decisions by exact recipe-class identity instead of manually copying the contract:

tools/compat-kit/compat-kit migrate-contract \
  compat/contracts/target.json \
  --old-audit compat/audits/target/1.2.3-old.json \
  --new-audit compat/audits/target/1.2.3.json \
  --output build/compat-kit/target-contract-migrated.json \
  --next-actions build/compat-kit/target-contract-migration.md

The command preserves a decision only when the recipe class, its public signature, its class-owned risk evidence, ancestry artifact SHA/size inventory, and the recipe-data inventory digest are all unchanged. New or changed classes reopen as needs_decision. Removing an accepted family fails; removed rejected legacy false positives are reported. Target or artifact-SHA drift is rejected.

For a one-issue/one-worktree worker, generate the compact handoff package:

tools/compat-kit/compat-kit worker-package \
  compat/contracts/target.json \
  --audit compat/audits/target/1.2.3.json \
  --output build/compat-kit/target-worker

The seven deterministic files contain hashes, compact class/recipe summaries, unresolved decisions, commands, issue/worker context, and a PR checklist. They never embed public-signature bodies or complete upstream source. Existing drift and symlinked output paths fail before any file is written. commands.sh uses the exact safe repository-relative audit path supplied by --audit; it never refers to a package-local audit.json that was not emitted. For a complete contract it runs every declared verification.gradle_tasks entry as its own Gradle invocation, followed by datagen and diff checks; the package cannot replace an authoritative target GameTest with only the compatibility matrix. The script preserves the worker's caller-provided JDK environment instead of hardcoding a macOS/Homebrew JAVA_HOME.

3a. Generate reviewed mechanical code

After the contract is complete, write a separate generation plan bound to the exact canonical contract digest and run:

tools/compat-kit/compat-kit generate \
  compat/contracts/target.json \
  --audit compat/audits/target/1.2.3.json \
  --plan compat/generation/target.json \
  --output build/compat-kit/target-generated

The generator emits direct typed registry calls for reviewed safe shapes: single-item deterministic recipes or a reviewed provider returning an exact TypedRecipePlan and RecipeFamilyCost. It can mechanically emit fixed, config-tick-ratio, public numeric getter, tier multiplier, parallel-lane, and speed-times-parallel station variants. Config wrapper fields are read with .get(); enum tier fields are read directly, for example FactoryTier.BASIC.processes. Every dynamic value must be positive, multiplication is checked, and every reviewed dynamic accessor is declared integral and converted through BigDecimal.longValueExact() rather than a lossy cast. Process rates are positive and Instant rates are fixed zero. Generated code uses MachineVariant.of/derived, public Auto Storage APIs, and exact target types; it never uses reflection. Families sharing one station descriptor must provide byte-identical station and rate definitions, and the generated register validates the descriptor's reviewed namespace rather than assuming auto_storage.

The plan is a reviewed binding, not another inference layer. Unsupported recipe shapes stay as explicit red_boundary entries. A typed provider remains the small handwritten semantic boundary for N-input/N-output selection, catalysts/tools/remainders, and mixed resource costs; code generation only owns registration boilerplate. Every generated family therefore declares an exact registration_id; the audit family ID identifies review evidence and must not be guessed to be the runtime registry path. The registration namespace must match the reviewed station descriptor namespace. Commit plans and generated source, then keep a byte-for-byte regeneration test so drift is visible.

The built-in single_item_to_item template accepts only its canonical contract shape: one consumed Item selected by recipe.input with amount 1, one primary Item selected by recipe.output with amount recipe.output.count, and either no cost or the single auto_storage:station_work cost selected by recipe.processing_time. Any changed amount, selector, extra input/output, or different cost remains a handwritten/RED boundary rather than being silently compiled with different semantics.

Contract recipe type IDs, station descriptor IDs, and station variant item IDs must use normal lowercase resource-location grammar. A descriptor cannot list the same station variant item twice. Generation plans must bind each station item exactly once; duplicate rate-item bindings fail before any Java is emitted instead of being collapsed by set comparison. A registry_block_method accessor also declares its exact block_id; the station's representative item ID is not assumed to be the owning block registry ID.

3b. Generate conformance and resource scaffolds

tools/compat-kit/compat-kit conformance \
  compat/contracts/target.json \
  --audit compat/audits/target/1.2.3.json \
  --plan compat/conformance/target.json \
  --output build/compat-kit/target-conformance

tools/compat-kit/compat-kit resource-scaffold \
  compat/contracts/target.json \
  --audit compat/audits/target/1.2.3.json \
  --plan compat/resources/target.json \
  --output build/compat-kit/target-resource

conformance emits a reusable assertion harness plus target scenarios for success/batching, shortage, capacity, overflow, stale holders, catalyst/tool/remainder preservation, multi-output merging, mixed-resource rollback, dedicated-server isolation, and all-mod coexistence. The generated harness performs snapshot/delta/unchanged and success/failure assertions; the integration supplies only the reviewed scenario setup and operation. Happy, catalyst/tool/remainder, and multi-output paths carry separate expected deltas, and zero-valued resource keys are normalized before comparison. Dedicated- server isolation checks the physical NeoForge distribution instead of trusting an addon-supplied boolean. Every conformance plan declares the exact game_test_namespace; the generated class uses it in @GameTestHolder. Every family batch must be at least 2, so happy_path_and_batching always proves a repeated operation rather than duplicating a single-craft assertion.

resource-scaffold emits API-only kind/container/block/renderer boundaries and real persistence, transfer, rollback, and dedicated-server test scenarios for an optional custom resource kind. It rejects Item, Fluid, and NeoForge Energy, which must reuse Auto Storage's built-in support. Generated common source may not import Core internals or client classes; the renderer bridge remains generic and client registration stays isolated. Each resource plan binds a sample amount, unique snapshot key, and exact game_test_namespace; the generated test class uses that namespace in @GameTestHolder. The tests own the before/after snapshot, first assert that reset/seed produced that exact key and amount, then own the delta, save/load round-trip, rollback, and physical-side assertions; an addon provider exposes operations and bytes, not self-attested persistence or client-isolation booleans. Resource IDs may begin with digits; every constant, registration method, and generated GameTest identifier is derived by one collision-checked Java-identifier normalizer that prefixes _ when needed.

Repository declarations are build inputs, not discovery hints. A target published through Modrinth Maven, Curse Maven, or its own Maven must list the required repository in target.repositories; an empty list means Maven Central is sufficient. Generated addons put these reviewed repositories first in declared order and restrict them to target and explicit runtime dependency groups. Explicit runtime groups cannot fall back to Maven Central even when target.repositories is empty; the target may fall back only under its exact SHA gate. Fixed Maven Central, BlameJared, and release-Ivy filters reserve Patchouli and Auto Storage for their owning repositories. The generator never guesses, sorts, or otherwise changes target repository precedence from the dependency coordinate. Required target-side runtime artifacts belong in target.runtime_dependencies; bundled descriptors and external addon builds copy the exact reviewed list instead of depending on transitive metadata or hand edits. Target and explicit runtime dependencies are non-transitive; every required companion must therefore appear in the contract. Repository URLs, dependency coordinates, and derived group filters are serialized as escaped Groovy literals, so $, quotes, and backslashes cannot alter the reviewed build input. Target and runtime dependencies must use exact group:name:version Maven coordinates; malformed or control-character-bearing values are rejected before scaffolding.

4. Scaffold a RED integration

Bundled module:

tools/compat-kit/compat-kit scaffold \
  --bundled compat/contracts/target.json \
  --audit compat/audits/target/1.2.3.json

Independent addon:

tools/compat-kit/compat-kit scaffold \
  --addon compat/contracts/target.json \
  --audit compat/audits/target/1.2.3.json \
  --output ../target-auto-storage

Bundled output owns its src/compat/<mod-id>/compat-module.json, isolated source set, one-call AutoStorageCompatModule, present-mod fixture, and GameTest structure. The fixture must be a Java-safe camel-case identifier ending in Fixture; path-like values fail before any file is created. Gradle discovers the source set, reviewed HTTPS repositories, target dependencies, fixture mod, run task, expected test gate, and audited target artifact from that descriptor; expectedTests must be a positive JSON integer representable by Gradle's Integer parser; fractions and wider overflow values are rejected rather than truncated. Before writing a bundled module, Compat Kit compares its derived module ID, entrypoint, source-set name, and fixture name with every existing descriptor; different mod IDs that normalize to the same Java/Gradle identifiers fail before materialization. The target is on both compile and fixture runtime classpaths, with both declarations non-transitive. build and the module GameTest resolve exactly one target jar and reject a SHA different from the reviewed audit. For an audited descriptor, that same coordinate must be the primary compile dependency and must also appear unchanged in the explicit runtime dependencies. No central compatibility switch is added.

External output is an ordinary NeoForge project with the Gradle wrapper, compile-only Auto Storage API dependency, reviewed target repositories and target dependency, one-call AutoStorageAddon.register(...), dependency metadata, a RED present-target GameTest, required Patchouli runtime, and reusable GitHub Actions workflow. It cannot compile against Auto Storage implementation classes. The separate source audit is copied to compat/audit.json; its build and runGameTestServer gates also resolve exactly one target jar and verify source_audit_sha256. Every target-derived metadata value, including addon display names and bundled fixture descriptions, is serialized as a TOML basic string, so quotes, backslashes, control characters, newlines, and multiline-literal delimiters cannot break generated metadata. DEL (U+007F) is escaped explicitly because TOML forbids it even though JSON serializers commonly leave it literal.

The scaffold adapter and fixture are deliberately RED. Rerunning scaffold is byte-deterministic and refuses to overwrite drift. The manifest binds the scaffold to the reviewed contract; implementation edits are expected after the initial RED generation. Verification regenerates the security-sensitive bundled descriptor or external Gradle execution chain (build.gradle, settings.gradle, gradle.properties, both launchers, and both wrapper artifacts) from the reviewed contract and generator, then checks every byte and manifest entry. Editing a launcher or gate and self-attesting a new manifest hash therefore cannot bypass artifact or task gates. Before writing anything, scaffolding preflights every destination for type/content drift and every required parent as a real directory. Symlinked roots, ancestors, and generated targets are rejected before the first write; a late-sorting conflict, file occupying src, or src symlink cannot leave a partial project or write outside the requested output root. The lexical output path is normalized before checking existing ancestors, so missing/../link/output cannot hide a symlink. Existing ancestors above a not-yet-created output root are checked too. Rerunning an unchanged scaffold restores 0755 on gradlew and tools/compat-kit/compat-kit.

Generated independent addons require the current compatible Auto Storage minor line. A 0.3.0 kit emits [0.3.0,0.4): patches remain compatible, while the next pre-1.0 minor requires an explicit addon review.

5. Implement with strict TDD

Run the generated failing fixture before editing the adapter. Implement only the reviewed contract through the public SDK. Keep target imports under the isolated compatibility source set. Do not add target classes to main/common, the API jar, or dedicated-server paths when the mod is absent.

Behavior evidence covers target-absent classloading, target-present registration, exact success, shortage, capacity, long overflow, stale holder, catalyst/tool/remainder behavior, mixed-resource rollback, dedicated-server client isolation, API-only compilation, and all-mod coexistence.

6. Verify

tools/compat-kit/compat-kit verify \
  compat/contracts/target.json \
  --audit compat/audits/target/1.2.3.json \
  --bundled . \
  --output build/compat-kit/target-report.json

For an external addon, use --addon <directory>. Addon contracts must declare exactly build and runGameTestServer, use main as their fixture, and map evidence only to those actual tasks; the verifier never rewrites bundled task names into addon task names. It runs both gates and gives runGameTestServer a fresh world. Verification rejects an unresolved contract, contract/source-audit mismatch, contract/manifest mismatch, remaining RED marker, forbidden implementation links, missing evidence source/marker, a source annotation count different from expected_game_tests, GameTest output that does not report one and only one exact passing summary, an undeclared evidence task, or any non-zero command. Multiple success summaries are rejected even when one matches, because editable fixture output cannot outrank the framework result. expected_game_tests is restricted to 1..2147483647, matching the positive JSON Integer required by the generated Gradle descriptor. A marker assigned to a run*GameTestServer task must occur inside an annotated @GameTest method body; a detached constant, helper, or comment does not prove that the behavior ran. Comments inside a real GameTest are removed before matching evidence, while markers in executable string arguments remain valid. Eligible Java Unicode escapes are rejected in evidence sources before raw marker matching because Java translates them before comment/token parsing; escaped literal \\u text remains allowed. The declaring source's @GameTestHolder must resolve to the exact namespace enabled by that Gradle run; evidence in a different namespace cannot borrow another test's successful task result. Literal holders and bounded static final String references are accepted, while missing, unresolved, ambiguous, or mismatched holders fail closed. A check is marked passed only when all of its declared source markers exist in the correct execution boundary and the associated Gradle task succeeds. The report records those per-check evidence links, exact commands, exit codes, output hashes, tool version, target, and manifest hash. Bundled verification runs every declared Gradle task as a separate process and removes only run/world before each task so stale GameTest state cannot leak between fixtures. Cleanup rejects a symlinked run/world path or any resolved path outside the verification root before deletion. RED and forbidden-link scans are scoped to src/; ignored build/ outputs and previously extracted scaffolds cannot poison a later verification. The expected hashes for the bundled descriptor or complete external Gradle execution chain are recomputed from the reviewed contract and current generator rather than trusted from the manifest. The target-SHA gate or launcher therefore cannot be replaced after scaffolding while retaining a passing report, even if the manifest is edited too. The published report schema requires a strict target identity, all twelve exact check IDs, and at least one successful command; external-addon reports additionally require exactly the authoritative build and runGameTestServer command records. GameTest counting and body extraction share one comment/string-aware Java mask, so @GameTest text inside a comment or literal cannot lend a helper method's markers to a passing runtime count. The extractor tracks annotation and method parentheses until the actual method body, so a brace-bearing intermediate annotation cannot become the evidence body. Escaped triple quotes inside Java text blocks remain literal content and cannot extend evidence into a detached helper. Each GameTest evidence file must also be under the fixture source set executed by its declared run*GameTestServer task; a marker in base or another fixture source set cannot prove behavior for the target run.

The repository compatibility-matrix task also depends directly on every descriptor-generated audited-artifact verifier and the separately pinned Botania verifier. Running either the normal 10k matrix or 30k stress command directly therefore verifies the exact audited target bytes first.

7. Review an upstream update

tools/compat-kit/compat-kit diff \
  compat/audits/target/1.2.3.json \
  build/compat-kit/artifacts/target-1.2.4.jar \
  --source build/compat-kit/sources/target-1.2.4 \
  --output build/compat-kit/target-1.2.3-to-1.2.4.json

Read the compact delta first. It reports only added, removed, or changed recipe/resource/station signatures and risk evidence. Any target jar SHA change sets contract_affected=true, including private implementation-only changes whose public signatures happen to be stable. The compact surface delta narrows the review, but it never waives contract review for different artifact bytes.

Distribution

./gradlew compatKitArchive

This produces build/distributions/auto-storage-compat-kit-<mod_version>.zip with the CLI, schemas, Gradle wrapper, RED templates, complete public-SDK addon example, CI example, README, and license. Only the explicit repository-owned addon example template is packaged; local build/ or untracked files under examples/addon are never traversed. The ten published JSON schemas are also an explicit allowlist, not a directory glob: audit, contract, conformance plan, delta, generation plan, proposals, report, resource plan, runtime-probe plan, and runtime-probe output. Gradle passes mod_version to publish, and publication fails unless it exactly matches the tool version embedded in generated projects. assemble/build includes the archive, CI uploads it, and tagged releases add it beside API, sources, and Javadocs artifacts.

Archive member order, timestamps, permissions, and bytes are reproducible. The extracted archive is also self-contained: scaffold --addon resolves its Gradle wrapper template beside the extracted CLI instead of requiring an Auto Storage source checkout.

PYTHONDONTWRITEBYTECODE=1 python3 -m unittest scripts.test_compat_kit
./gradlew compileCompatKitGeneratedFixtureJava

src/compatKitGeneratedFixture/ is a committed compile-only contract for the generated conformance and custom-resource outputs. Python regenerates those sources byte-for-byte from the committed example plans, while Gradle compiles them against only the public API jar plus NeoForge/Minecraft. The providers deliberately throw if executed: runtime semantics are supplied and asserted by each real integration's GameTests, while this fixture permanently guards the generator's Java/API surface.

First dogfood: AE2 Inscriber

The committed AE2 19.2.17 scanner-format-12 audit, migrated contract, generation plan, and generated registration prove this workflow against a real target. Structural classification plus the reviewed NeoForge/Minecraft ancestry jar finds twelve actual Recipe implementations while inventorying 556 effective recipe JSONs across 18 serializer groups. The accepted slice remains only InscriberRecipe:

  • middle is consumed;
  • optional top/bottom are retained for INSCRIBE and consumed for PRESS;
  • output is one exact stack;
  • cost is 200 station-work steps and 2,000 configured AE converted through AE2's public PowerUnit API to a finite positive exact integer FE amount;
  • the plain Inscriber contributes 2 work per tick; speed-card state is not inferred from its item;
  • Charger, Entropy, six custom crafting families, Quartz Cutting, Matter Cannon Ammo, and Transform remain rejected.

Ae2GeneratedCompat.java owns deterministic descriptor/family registration; the handwritten Ae2Compat methods own exact eligibility, typed transaction planning, and AE-to-FE cost semantics. A golden test regenerates the source and requires byte identity before the isolated AE2 source set is compiled.

The bounded rate templates are additionally compiled against source-shaped fixtures for Iron Furnaces' public config wrapper Config.ironFurnaceSpeed.get() and Mekanism's public factory parallel field FactoryTier.BASIC.processes. Their real isolated GameTest modules remain the runtime evidence for configured speed and factory parallel throughput; these representative versions are CI fixtures, not player dependency pins.

Its required-check evidence uses markers inside annotated method braces, not method names or declarations. The ingredient-shortage check omits the PRESS middle ingredient and asserts that both remaining items, FE, station work, and output are unchanged.

See AE2 compatibility.

Auto Storage 0.3.0

Clone this wiki locally