agentleFS
LuqueeeLuqueee/kivgraph/.claude/skills/adding-a-language/SKILL.md

adding-a-language

Skill21 starsChanged 3 months agoApache-2.0View on GitHub

How a new language enters Kivgraph's graph - the two routes and why one of them is almost always wrong, the four decisions that cannot be migrated later, the branches that decide what your language is without asking, and the gates that fail closed. Use when adding or extending language support, when touching `internal/facts/semantic.go`, a `*loader` package, `SupportedLanguages()`, or the provenance codes.

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 3 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Add your two constants, or your first test fails at the door.

    line 34
  2. Distinctive:

    All three languages that took it did so before the payload existed.

    line 66
  3. Distinctive:

    Ask it as: *can two files of one package declare the same qualified name?* If yes, the file is part of the identity.

    line 113
  4. Very specific:

    If the language gets an alias in SupportedLanguages() (internal/config/config.go:1277), it gets a row here under each spelling.

    line 154
  5. Very specific:

    Do not trust a list written by hand.

    line 164
  6. Very specific:

    Every semantic language shares unitSemantic and differs only by language.

    line 227
  7. Very specific:

    unitIdentity had no branch for Rust: Every Rust workspace was keyed typescript\x00<repository>\x00 with an empty package name.

    line 248
  8. Very specific:

    Use internal/testsupport.TempDir(t), never t.TempDir(), when the path feeds the workspace layer.

    line 362

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Adding a language to Kivgraph

The rule that costs the most to discover

internal/facts/semantic.go:483:

func definitionProvenance(language Language) Provenance {
	if language == LanguageDart {
		return DartAnalyzerDefinition
	}
	return PythonIndexerDefinition
}

That was the shape until Java arrived: two functions, both if Dart, else Python. A language added through the payload route and not added here published every definition as PYTHON_INDEXER_DEF and every use as PYTHON_INDEXER_USE. Nothing failed. The provenance is legal, canonicalProvenanceValues accepts it, stage.integrity passes with 0 invariant violations, the golden probes pass, doctor goes green. The graph is wrong in the one field that says where a fact came from, and nothing but reading it would have told you.

It is a table with no default now, and NormalizeSemantic refuses a payload whose language has no provenance. That check is not decoration: Set.Validate would not have caught it, because it only rejects an empty provenance under an EXACT confidence and DEFINES is StructuralCertain. Add your two constants, or your first test fails at the door.

The same shape repeated in the indexer, where the fallthrough was TypeScript, and that one is fixed too -- see The branch that used to decide what you are.

So the work is not writing a loader. The work is finding the branches that already decide what your language is without being asked. Two of them are now failures instead of defaults, which is the only reason this file is shorter than it was. Assume there are more.

Two routes, and the one you want

Route A: the payloadRoute B: your own normalizer
The producer emitsfacts.SemanticPayload as JSONwhatever it likes
Graph semantics written byfacts.NormalizeSemanticyou
Who took itPython, Dart, Java, C#Go, TypeScript, Rust
Costinternal/pythonloader/loader.go is 178 lines, python-worker/index.py is 258internal/facts/golang.go is 747 lines, typescript.go 937, rust.go 534 -- each with its own Input, Report, unit type, cache branch and counters, on top of a loader
Contractdocs/protocol/semantic-facts-v1.mdnone; it is code

Route A is not the cheap route because it does less. It is cheap because NormalizeSemantic (internal/facts/semantic.go:193) already writes the repository, the package, the files, the symbols, DEFINES, the evidence, the references, the calls, the imports, PART_OF and every UNRESOLVED -- and it writes them the same way for every language that uses it, which is the actual product. A second copy of the graph model is a second place for the model to drift.

Dart's loader is 2013 lines and that is not a counter-example: it speaks the Analysis Server protocol. It writes no graph semantics either.

Take route B only when the payload cannot express what the analyzer knows. All three languages that took it did so before the payload existed.

And check internal/scip before writing a loader at all. If the language has a SCIP indexer -- scip-java, scip-python, scip-ruby, scip-dotnet, scip-clang, or rust-analyzer scip -- the conversion to a payload is already written, and the language is a loader that runs an indexer and names a package. Java is internal/javaloader/loader.go at ~230 lines and C# is ~300. See ADR 0080 and ADR 0081.

But a SCIP language is not free, and the way it is not free is worth knowing. The format is shared; almost nothing else is. Adding C# found three bugs in the bridge that were shaped like Java and that no second Java-family language could have exposed:

  • addressable excluded a symbol whose package and version were both ., which is a package qualifier in scip-java and every local symbol in scip-dotnet. It dropped the whole language.
  • Parameters were local N in scip-java and fully qualified symbols in scip-dotnet, so the same concept was a node in one language and absent in the other.
  • scip-dotnet emits no enclosing_range, no SymbolInformation.Kind and no signature. Spans are reconstructed, kinds fall back to the descriptor suffix, and the stable-key discriminator falls back to the descriptor path.

So: before assuming a producer behaves, dump one of its indexes and read it. Every one of those was a five-minute discovery and a silent, plausible graph if missed.

The four decisions you cannot take back

Make these before writing a line. Each is frozen by a contract in the root AGENTS.md, and each is a data migration once it ships.

1. What scopes a symbol's identity

NormalizeSemantic builds a StableKeyIdentity per symbol. Python leaves Module empty: a qualified name is unique inside a package. Dart sets Module to the file, and on a collision appends #<start offset> (internal/facts/semantic.go:267), because Dart lets two files of one package declare the same name.

Get this wrong and every stable key of the language is wrong. Stable keys are listed in Superficies que rompen compatibilidad: the algorithm, the canonical identity and the historical luque-stable-key namespace do not change without a data migration, an ADR and an explicit contract update. This is not fixable in a patch.

Ask it as: can two files of one package declare the same qualified name? If yes, the file is part of the identity.

2. Whether the producer is authoritative

authoritative: true is reserved for a producer that resolved references with the language's own semantic analyzer. Absent or false, definitions are still structural facts but references and calls are CANDIDATE.

Python ships both and says which is which: an AST fallback that reports itself as python-ast-fallback, and a Pyright adapter for exact mode. A fallback that claimed EXACT would be the one thing the project does not do -- an EXACT edge is never created by name, text, path, alias or single candidate.

CANDIDATE is not a degraded EXACT. It is a different answer, and callers filter on it.

3. The provenance codes, which are append-only

internal/facts/codes.go:60, and the file says why at line 13: the HotSnapshot stores kinds, confidences and provenances as bare uint8 and filters traversals by them, so a snapshot is only readable by code that agrees on the numbering.

Append your constants at the end of the block. Do not insert, do not renumber, do not sort the list to be tidy. Reordering is a format break that requires bumping CodeFormatVersion, which makes every published snapshot unreadable.

Then add the same values to provenanceCodes in that file and to canonicalProvenanceValues in internal/storage/ladybug/canonical_integrity.go:119 -- the second list is written out by hand on purpose, so it does not follow the first.

4. The extension set, and the aliases

internal/config/languages.go:21 is keyed by every spelling SupportedLanguages() accepts, aliases included. The file explains the defect that made it: a lookup that only knew the canonical name answered nothing for the alias, and the watcher then did not see a change to a .py or a .dart file as a source change at all.

If the language gets an alias in SupportedLanguages() (internal/config/config.go:1277), it gets a row here under each spelling.

Separately, moduleSymbolExtensions in internal/facts/module_symbol.go:32 decides what is stripped from a module symbol's qualified name. It is readability only -- the stable key disambiguates on the path -- but a language missing from it names every file's module scope with the extension still on.

The touchpoints

Do not trust a list written by hand. publishing-releases learned that one the expensive way: its hand-written list of files carrying the version was two minors out of date on a page nobody had thought to add. Derive the list instead, from the commit that did this twice in one pass:

git show --stat fe8308b     # feat: add Python and Dart semantic indexing support
git log --oneline fe8308b..HEAD -- internal/pythonloader internal/dartloader
git grep -il dart -- 'internal/**' 'cmd/**' 'scripts/**' '.github/**'

fe8308b is 70 files and 4197 insertions for two languages at once. The commits after it are the ones worth reading twice: they are what the first pass missed.

What the files want, which the list of names does not say:

FileWhat it needs
internal/config/config.gothe name and aliases in SupportedLanguages(), a <Lang>Config struct, its field on Config, its defaults, and its validation
internal/config/languages.goone row per accepted spelling
internal/facts/facts.goLanguage<X> and the provenance constants
internal/facts/codes.gothe codes, appended, and the provenanceCodes rows
internal/facts/module_symbol.gothe extension
internal/facts/semantic.godefinitionProvenance, useProvenance, and the identity branch if the language is file-scoped
internal/<lang>loader/the new package: run the producer, decode, check version and language, set Authoritative
internal/scip/nothing, if the producer emits SCIP: Convert already writes the payload. Give it the language, the repository, the package and a file reader -- then check what your producer actually sets, because the bridge tolerates absences but cannot invent them
internal/indexer/semantic.gosemanticSourceExtensions, the indexSemantic switch, semanticRequestedPackage if an import prefix has to be stripped
internal/indexer/full.gothe partition, the units append, the report counters, semanticSchedule, semanticBudget, a progress phase and a worker limit
internal/indexer/factcache.gothree separate duties -- see below
internal/indexing/full.gothe options, twice: from configuration and from the caller
internal/indexing/service.go, document.gothe JSON report counters
internal/storage/ladybug/canonical_integrity.gocanonicalProvenanceValues
internal/syntax/parser_manager.go, grammars/manifest.jsononly if a tree-sitter grammar is pinned; the manifest is digest-pinned and never edited by hand
internal/mcp/instructions.gothe sentence users read, and surface_test.go asserts it
internal/integrations/assets/kivgraph/SKILL.mdthe shipped skill names the languages
cmd/kivgraph/main.gothe index.<lang> counter line and the doctor toolchain branch
scripts/verify-semantic-coverage.sha hardcoded set -- see the gates
testdata/semantic-coverage/manifest.jsonthe entry, with at least ten capabilities
testdata/<lang>/the fixtures the manifest names
scripts/build-bundle.shonly if a worker script ships
.github/workflows/ci.ymlthe toolchain, or the tests skip and the suite still goes green
docs/adr/00XX-*.mda new analyzer boundary is an architecture change; the next number is 0082

internal/agenthook/languages.go needs nothing. It is a thin name over config.SourceExtensions, and it is thin precisely because it used to be the second copy of that table and the two disagreed.

The branch that used to decide what you are

This section used to list eleven switch { case unit.isPython: ... } sites whose default was TypeScript, and told you to find all eleven by hand. Adding Java did the refactor it recommended instead. What is there now:

git grep -n 'unit\.kind' -- 'internal/**/*.go' | grep -v _test

analysisUnit carries one kind (internal/indexer/full.go), whose zero value unitUnspecified is deliberately not a language. analyseUnit refuses it and names the repository, so a unit built without its kind stops the pass instead of being analysed as a TypeScript package.

Every semantic language shares unitSemantic and differs only by language. So a language added through the payload route touches none of the nine dispatch sites. What it does touch is a set of tables, all in internal/indexer/full.go and internal/indexer/semantic.go:

git grep -n 'LanguageJava\|LanguageCSharp' -- internal/indexer/ | grep -v _test

semanticSourceExtensions, the indexSemantic switch, addSemantic, semanticSchedule, semanticBudget, semanticRequestedPackage if an import prefix has to be stripped, plus the partition and the units append in Full, a ProgressPhase and a worker limit. A language with more than one accepted spelling also needs dedupeRepositories on its partition, or a repository that declares both is indexed twice. Call it eight or nine places rather than four -- the win is not that there are fewer, it is that every one of them is a table or a switch with an explicit default, so a miss is a compile error, an empty count or a named failure instead of another language's facts.

Two things the refactor turned up, which say what the old shape was costing:

  • unitIdentity had no branch for Rust. Every Rust workspace was keyed typescript\x00<repository>\x00 with an empty package name. One workspace per repository hid it -- wrong but unique. Two in one repository shared a fact-cache entry, and the second was served the first's facts.
  • The counters, weight(), detail(), describeInputs and the scheduler queue all fell through to TypeScript too. Nothing errored; the graph was simply about the wrong thing.

The lesson generalises past this repository: when a dispatch's default is a real case rather than a failure, adding a member is a silent bug waiting for whoever forgets one site.

The cache has three duties, and each fails silently

internal/indexer/factcache.go:

  1. Identity -- unitIdentity at line 380, what the entry is about.
  2. Inputs -- describeInputs at line 395, the source tree plus every manifest whose content changes the answer. The Python/Dart branch lists twelve filenames by hand; a manifest yours reads and that list does not name is a file you can edit without invalidating anything.
  3. Analyzer fingerprint -- analyzerFingerprint at line 664, every option that changes what the producer does.

And a fourth that is not obvious and cost a defect. e0d8d52 (fix(indexing): fingerprint the python producer the pass runs): the options were fingerprinted, the binary was fingerprinted, and the script the loader actually resolves and runs was not. Editing python-worker/index.py changed no cache key, so a rebuild reused the previous producer's facts and published a generation the current code would not produce.

The fix is pythonloader.ProducerFile, and the shape to copy is in its comment: it resolves with the same rules as resolveCommand, because two resolution rules is how a cache ends up keyed on a file nobody runs. If your producer is a file this repository ships, fingerprint the file, resolved the way you run it. If its identity is unknown, say so with a value that cannot match -- unknown identity is not a licence to reuse anything.

The gates that fail closed on a language they do not know

scripts/verify-semantic-coverage.sh opens with:

expected = {"go", "typescript", "python", "dart"}

make semantic-coverage fails until the set names your language and testdata/semantic-coverage/manifest.json has an entry whose fixture and test both exist on disk and whose capabilities list has no duplicates and at least ten entries. This one is honest: it fails loudly and immediately.

Note what it also says: Rust is not in that set. The coverage gate covers six languages and Rust is not one of them. Do not read a green semantic-coverage as a statement about Rust, and do not treat its absence as permission to skip yours.

internal/mcp/surface_test.go:333 asserts the server instructions mention "Python" and "Dart" by name, and that the whole string stays under 2048 bytes. Adding a language to that sentence is a change to the MCP surface, which AGENTS.md lists under Superficies que rompen compatibilidad: the tool descriptions and the skill that names them do not change in silence.

An analyzer that builds must not build in the repository

AGENTS.md says it under Nunca modificar with no exception, and it is the easiest rule in the project to break by accident: a build-based indexer -- scip-java driving Maven, scip-dotnet running dotnet restore -- writes target/, obj/ and bin/ into the directory it builds, and no flag moves them. They are the build tool's output, not Kivgraph's.

internal/scratchtree is the answer: it materialises the working tree somewhere else, the build runs there, and the tree dies with everything it produced. Measured on this repository -- 1652 tracked files, a 3.8 GB working tree:

strategytimesizewrites inside the repository
copy of the working tree8154 ms4.5 GBnone
git worktree add107 ms16 MByes, .git/worktrees/
git archive + dirty overlay76 ms16 MBnone

Two things to take from it. git worktree add is the obvious answer and the wrong one -- it registers metadata inside the repository, which is the rule you came to keep. And a materialisation reproduces the working tree, not HEAD: a user editing code expects the graph to describe what is on disk.

Then the defect it introduces, which is worth knowing before you write it: the scratch tree is a fresh temporary directory per pass, and if the loader derives its package name from where it read the files, every stable key changes on every pass. A loader carries two roots and they are not interchangeable -- sources is where files are read, repository is what the facts are about. See ADR 0082.

And the test that catches it has to index in place. Every end-to-end test here copies the fixture first, which is allowed and is also exactly what hid this for two languages.

Fixtures, and what a fixture has to be

testdata/<lang>/basic and testdata/<lang>/advanced is the shape Dart uses; Python uses basic and coverage. The manifest names one of them and the test that reads it.

The rules from AGENTS.md and running-tests that bite here:

  • An indexing pass must leave the fixture byte-identical. The Rust fixture is the precedent: no target/, no Cargo.lock after a run, and the test checks it. If your analyzer writes into the project it analyses -- a .dart_tool/, a __pycache__, a lockfile -- it runs in a scratch tree; see the section above. A test that copies the fixture first proves the fixture is fine and says nothing about a user's repository.
  • Start with the negatives. What the producer cannot resolve is the contract, not an omission: dynamic dispatch, a decorator that rewrites a name, a conditional import, generated code. Each becomes an UNRESOLVED with a reason, and the test asserts it stays unresolved.
  • Use internal/testsupport.TempDir(t), never t.TempDir(), when the path feeds the workspace layer.

What ships and what only runs here

A worker script that the bundle does not carry works on your machine and nowhere else. scripts/build-bundle.sh installs python-worker/*.py explicitly, by name, into worker/python-worker/; pythonloader.resolveCommand then looks in three places in order -- the working directory, the source tree relative to runtime.Caller, and ../worker/ beside the executable. Copy all three or the bundled binary cannot find its own producer.

doctor reports each toolchain separately (cmd/kivgraph/main.go:1622, runDoctorToolchains), and the rule it implements is worth keeping: a toolchain nobody installed is a fact about this machine, not about the code. analyzerNotInstalled (internal/indexer/semantic.go:93) isolates the repository with not_loaded=1 and lets the others index. A language that aborts the whole pass because its analyzer is missing is a defect -- that one already happened, five repositories deep, four of which had nothing to do with Dart.

Verifying

gofmt -l <changed-go-files>
go vet ./...
go test ./...
make test-ladybug                # the loader touches indexing, so this is not optional
make semantic-coverage
go test -race ./cmd/kivgraph/...
make build

Then check the suite did not skip the thing you came to prove -- a missing toolchain is a SKIP, and a SKIP is green:

go test ./internal/<lang>loader/ -v -count=1 2>&1 | grep -E 'SKIP|FAIL'

And prove it end to end on a real fixture with a native binary, because the make build binary has no storage layer and dies at publish. The full recipe is in running-tests, under Smoke test del binario; the language-specific part is:

env HOME="$home" /tmp/kivgraph-native init --repository "fixture=$repo" --languages <lang>
env HOME="$home" /tmp/kivgraph-native doctor
env HOME="$home" /tmp/kivgraph-native index --full

A healthy pass prints index.<lang>: repositories=1 not_loaded=0 symbols=N references=N unresolved=N, stage.integrity: PASS (0 invariant violations) and stage.golden probes: PASS. symbols=0 with not_loaded=0 means the partition never selected the repository: the language name reached SupportedLanguages() but not the repositoriesFor... call in internal/indexer/full.go.

Then read the provenance of what you published. It is the one thing every gate above passes without checking, and the defect at the top of this file is invisible until someone looks.

Not part of adding a language

  • A release. A new supported language is a capability a user can run, so it does qualify -- but only when someone asks. See publishing-releases, and the twenty-one chore(release) commits in two days that section exists to prevent.
  • The landing. landing/src/components/landing/Languages.astro and the pages around it are user-facing copy for a language that works. They come after the coverage gate is green, not before, and they are written in English like everything the project publishes.
  • Removing a language. Nothing here is symmetric. A configuration key that is retired is accepted, ignored and named -- retiredConfigKeys and the config.retired line of doctor -- because the decoder rejects unknown keys and deleting the field turns every existing config file into a startup failure. See ADR 0062.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.