Adding or completing a language

A language in BAKED is not one switch. It is a row in about twenty per-language tables plus a dozen detectors and parsers that still keep their own lists. Every earlier language fix that was written against the one run that exposed it left the same defect in the other languages (see Lessons). This skill is the checklist that keeps a language pull request from stepping on the same rakes again.

One pull request per language (or per toolchain of a language), on a feature/<language>-language-support branch from fresh main, closing its issue. Remove the issue's deferred label when you take it.

1. Find what is missing

1. Read the language's issue and the issues it links. Its file:line references may be stale: the tables moved to LanguageRecords in #1717 and #1733, and #1488, #1749 and #1753 filled several rows since. Re-check every claimed gap on current main.
2. List the language's empty rows: grep -rn "^\s*<language>: none(" src. Each none(reason) is either a real "this language has no such thing" (keep it) or a "not done yet" (the reason names an issue or says "yet"/"not read"/"not recognized"): those are the gaps.
3. Walk the layers below that are not LanguageRecords. They do not fail the compiler or the parity test when a language is missing, so check each by hand.
4. Run the language's fixture through tests/language-parity.test.ts (bun run test tests/language-parity.test.ts); a golden that still says null for a row you fill must change in the same branch.

2. The layers

Records (the compiler and tests/language-parity.test.ts ask for them)

Record	File	What a row holds
TEST_CONVENTIONS, TEST_MARKERS, MANIFEST_NAMES	src/core/intelligence/language-traits.ts	test file names; what a test file holds when the name is a plain word (#1848); manifests by file name
GATE_TOOLCHAINS + TOOLCHAIN_GATES	src/core/workflows/gates.ts	which detectors the language gets, in order
FILE_FORMATTERS	src/core/workflows/format-written.ts	the formatter gate's file-scoped check and write forms, range form and probe
TEST_RUNNERS (+ TEST_COUNTERS, FAILED_TESTS)	src/core/verification/output-counts.ts	test count with ran, failed test names, a stable failure signature
TEST_GATES	src/core/workflows/test-gates.ts	build gates a test file declares (features, tags) and the flag that turns them on
DEPENDENCY_MANIFESTS	src/core/workflows/quality/manifests.ts	dependency reader with fields (executable fields, #1793)
ENTRY_POINTS	src/core/verification/entry-points.ts	how a command the project installs is launched
PLACEHOLDERS	src/core/workflows/quality/builtin-rules.ts	unimplemented-body idioms, and not the language's legitimate look-alikes
MASK_SYNTAX	src/core/workflows/quality/code-mask.ts	comments and strings, including raw strings and long brackets
CALL_SYNTAX	src/core/intelligence/call-context.ts	call and import spellings
NUMERIC_LIMITS	src/core/verification/edge-inputs.ts	the integer and float edges worth a probe
MUTATION_TOOLS, PROPERTY_LIBRARIES, PACKAGE_CONTENTS	src/core/workflows/quality/	mutation tool, property libraries, package build
INLINE_HARNESS + TEMPLATES (keyed by test framework)	src/core/workflows/acceptance-rows-inline.ts, acceptance-rows.ts	acceptance rows of a new test framework, both modes (#1763)

Every row you fill gets its golden in tests/fixtures/languages/<language>.json.

Lists outside the records (check each by hand)

• Extension and grammar: EXT_TO_LANGUAGE, BARE_FILENAME_TO_LANGUAGE (src/core/intelligence/types.ts), the tree-sitter queries (backends/tree-sitter.ts), imports in repo-map.ts. A grammar that tree-sitter-wasms lacks goes in vendor/tree-sitter/ with a README row (source, version, how it was built) and its license file; resolveWasm, the bundles and postinstall.mjs already copy that folder (#1499). Vendor the .wasm, never an npm grammar package that runs node-gyp on install.
• Syntax check: CHECKED_GRAMMARS in src/core/intelligence/syntax-check.ts. Add a grammar only after it parses real projects without false errors: run it over the files of a few real packages that the language's own compiler accepts (luac -p, ruby -c), measured on a real corpus (published packages of the language), not on the fixture. A grammar that lags the language stays out, with a comment and the count (Kotlin; Ruby 5 of 2,618 files; Lua 374 of 1,133; tree-sitter-powershell 27 of 98 real scripts, #1498).
• Language server: SERVER_CANDIDATES and SERVER_PROJECT_STATE in backends/lsp/server-registry.ts, the installer (PROJECT_INDICATORS) and the LSP language set in backends/lsp/index.ts. A server shipped as a folder that a runtime starts (PowerShell Editor Services) goes in SERVER_BUNDLES under its Mason package name, with the runtime's arguments built from the install folder; the installer keeps a package without binaries whole in lsp-servers/NAME (#2125). A server that writes project files into the checkout (jdtls .project, #1508) gets its files in SERVER_PROJECT_STATE.
• Agent profile and edit-time format: detectToolchain in src/core/tools/project.ts, read by tools/auto-format.ts; build it from the gate detector (composerGates, luaGates) so the agent runs what the checks run, and give the formatter a stdin probe in formatProbeCommand, or an edited file is never formatted. Read the tool's config from every place the tool reads it (ruff in pyproject.toml, #1537; prettier in package.json, rustfmt's edition, #1681).
• Language reviewer: ProjectLanguage, detectLanguages, LANGUAGE_FILE in src/core/workflows/language.ts, and the language-reviewer profile text (src/core/agents/profiles/builtin.ts) for the language's platform pitfalls.
• Check routing: STATIC_COMMAND, WHOLE_UNIT_COMMAND, fixerCommand and the style-claim lists in src/core/workflows/check-policy.ts (a tool named in FORMAT_CLAIM or LINT_CLAIM also needs FORMAT_COMMAND or LINT_COMMAND, or a passing gate never shows the claim; PHP missed it, #1495); isTestCheck in quality/mutation.ts (a check-mode flag such as Pint's --test is not a test run); the program behind an interpreter or runner in toolchainCommand (src/core/workflows/toolchain.ts); CHECK_COMMAND in src/core/verification/criteria-cases.ts (#1427); noTestsCollected and MISSING_PATH_LINES in src/core/verification/convergence.ts (#1208, #1913); timedOutFailure in check-runner.ts (#1424); REPLAY_TARGETS in quality/baseline-replay.ts (#2020); diagnosticCount in output-counts.ts (#1791, #1326).
• Scoped tests: src/core/workflows/affected-tests.ts: its own isTestFile and isCompiledTestFile (which files a task's written-tests check runs; Ruby's runner branches landed without them, so no written spec ran until #1495), the value-taking flags flagsOnly keeps, the runner's file or filter form in affectedTestCheck, and the per-test timeout in test-timeout.ts (#1655).
• Test integrity and code rules: the skip markers (SKIPS, or a framework's own skips in src/core/intelligence/test-frameworks.ts), CAPABILITY guards and SUPPRESSIONS in src/core/workflows/test-integrity.ts, with their lists in the diff-check paragraph of docs/workflows/runtime.md, the assertion forms of the language's frameworks (#1651), and the lint-suppression rule in quality/builtin-rules.ts with its row in docs/workflows/code-quality.md.
• Workspace and baseline: workspace-facts.ts, sizing manifests (sizing.ts), TOOL_CACHES (src/core/verification/snapshot.ts) and FIRST_RUN_OUTPUTS (tests/helpers/baseline-created-files-scenario.ts) for files a first build writes (Cargo.lock, #1815).
• Manifest approval: requestedManifestEdits in auto-approval.ts (#1552, #1428) and the toolchain-selecting files (wrapper distribution, SDK pin, #1963).
• Qualified names: symbolPath in src/core/workflows/finding-chain.ts splits a reviewer's Container.member spelling, and CONTAINER_MEMBER in tests/workflow-scenario-finding-chain.test.ts holds a row per language (#2137). The tests are not typechecked, so only the test run reports a missing row; R added $ (Reader$read, #1500).
• Repro and examples: reproExecutables in src/core/verification/repro.ts adds the language's runtime for its build tool (#1529); a documented example that needs a build is built first (#1035).

3. Rakes we already stepped on

Each line is a past bug. A new detector, parser or table row answers it before review does.

1. Windows wrappers. gradlew.bat, mvnw.cmd, LuaRocks and Composer vendor\bin\*.bat run through cmd /d /c, which launchedCommand unwraps (#1749). Give every command a Windows form and a gate-matrix cell for it.
2. Whole unit or per file. A linter or analyzer that resolves imports cannot pass in a task's isolated worktree: it is whole-unit (#584, #799, #1753). A formatter is per file, and its fixer takes the task's file, never . or the whole crate (#587, #840, #1228, #1269, #1593).
3. Formatter flags and configs. A range or file flag that older releases lack sets probe (#1907). Edit-time format reads the project's own config and never reformats a whole file that was not clean (#1582), within the 30 s limit (#1429).
4. Failure signatures. Strip durations, timings and pass counts so the same red check is recognized again (#1189, #1425, #1678).
5. Failure lines and files. The runner's failed-test and failing-file lines feed correctives, waits and hangs (#1419, #1420, #1424). Read only failing files, not every file the output names.
6. Empty and missing. The runner's "no tests" output and exit code (pytest 5, #1208, #1031, #1864) and its "path does not exist" line (#1913) are read, so a task that runs before its tests exist is not corrected.
7. Diagnostics. Count errors and warnings only, from the tool's default output format (#1791, #1326).
8. Gated tests. A test file gated by a feature, tag or build condition that the command does not turn on runs zero tests: TEST_GATES and requireTests (#1832).
9. Test file names. A plain-word name (PermTest, FeatureSpec) names domain code too, so it needs a TEST_MARKERS match outside a test folder (#1683, #1848).
10. Placeholders and idioms. Recognize the language's "not implemented" forms (#1679, #951) without flagging its legitimate look-alikes (Ruby's abstract NotImplementedError, Java's read-only remove()).
11. Properties and assertions. Recognize the language's property libraries and loops (#1064, #1686) and assertion frameworks (#1651).
12. Manifests. Read every form of the manifest (tables, extras, groups, catalogs, #1682), and register new projects or modules a later task adds (#1684, #1299, #1412).
13. Host-written code must be clean. The acceptance rows the host writes pass the language's formatter and linter as configured, carry no raw control characters or BOM, and keep one type per file where the language expects it (#1927, #1942, #1995, #1804, #1961).
14. Tools that write into the checkout. Language servers and first builds leave files (#1508, #1815): list them, or the run blocks input_changed.
15. Runtimes for examples. The README of a project builds and runs with the language's runtime, not only its build tool (#1529, #1035).
16. Detectors replace the runner fallback. detectGates offers a Makefile, justfile or CI step only when no toolchain detector found anything, so a detector that fires on a config file of a side tool (a .shellcheckrc in a C project) would hide make test. Shell gates are added on top for that reason (#1497); a new detector for a side tool does the same.
17. Shared runners carry their totals. A harness that several languages run through (TAP::Harness behind prove and pg_prove) prints a summary with totals and CPU times; strip them in GROWING_COUNTS, or the same red check never matches again (#1499).
18. Grammar ERROR nodes. A grammar that cannot parse part of the language (PL/pgSQL bodies in tree-sitter-sql) turns that part into ERROR nodes; match the declaration inside the ERROR node too, or its symbols vanish from the call graph (#1499).
19. A linter that passes broken files. A linter run at error severity can still pass a file that does not parse (PSScriptAnalyzer needs ParseError next to Error,Warning); give its gate-matrix cell a file with a syntax error (#1498).
20. Every runner in the replay table. tests/workflow-baseline-replay.test.ts holds a FAILURES and a BUILD_ERRORS row for every runner in BUILD_FAILURES; a new runner adds both (#1498).
21. Tools that survive HOME isolation. Tests run with a temporary HOME, so a tool installed for the current user (PowerShell modules in CurrentUser scope, pip user site) vanishes in them: install CI and device tools for all users (#1498).
22. Output on a CI runner. A runner that detects CI (GITHUB_ACTIONS, TF_BUILD) can change its output: Pester prefixes failed tests with ::error:: (#1498), and lintr's print writes ::warning file=... annotations (#1500). Turn that mode off, or write the lines in the gate's command, and run the real gate-matrix cell with GITHUB_ACTIONS=true before pushing (#1498).
23. Build trees are not sources. A check that writes into the checkout on every run (a CMake build/ folder) makes every later check source_changed unless the project ignores it; sourceManifest skips a folder holding CMakeCache.txt (#1490). A tool that needs a configure or build step first gets that step as its own gate, in order, since a check runs without a shell.
24. Preprocessor lines are code. The integrity and rule scanners treat a line that starts with # as a comment; #if, #ifdef and #elif stay code, so a C or C++ skip under #ifdef _WIN32 reads as guarded (#1490); Swift's #if os(...) guards and #expect/#require assertions are code too (#1492).
25. Bench files the root .gitignore hides. bin/, build/ and the like are ignored at the root, so a bench task's seed/bin/ or solution/bin/ is never committed: run git status --ignored bench/ before the pull request and git add -f them (#1497).
26. A framework's own skip spellings. RSpec's xit, ScalaTest's ignore, package:test's skip: go in the framework's skips column of test-frameworks.ts, not in the shared SKIPS, so they never match another language's code (#1493, #1494, #1496).
27. Output flags are not static checks. A flag that only picks the output format (rspec --format) leaves the check a test run in isStaticCheck (#1493).
28. Load failures in the summary. A runner that counts a test file that did not load as a failure (dart test's loading x_test.dart [E]) needs it subtracted from the test count and read as a build failure, not a failed test (#1494).
29. Shared assertion rows. A shared row (the JavaScript expect) can misread another language's argument order; add a scoped row whose calls come first (#1494).
30. Formatters that exit 0 on a missing path. dart format prints No file or directory found and exits 0: it needs a MISSING_PATH_LINES entry, or a task that formats a file it has not written passes (#1494).
31. Runners that print no totals. Gradle prints no test count; read the reports it writes (build/test-results/**/*.xml) and append them to the check output (#1489).
32. The first run compiles. sbt and SwiftPM print compile and build-step lines only on the run that compiles; strip them from the failure signature, or a repeated failure looks new (#1496, #1492).
33. Build-tool formatters. A formatter that runs through the build tool (sbt scalafmtOnly) needs a fixer that passes the file inside the same build command, not as a separate argument (#1496).
34. Real output only. Parser cases are pasted from a real run of the tool. In the cloud, Maven Central's Google mirror (maven-central.storage-download.googleapis.com/maven2) and Docker images (sbtscala/scala-sbt, swift) give real JVM and Swift tools where the default downloads are blocked (#1496, #1492).
35. Words in a gate's args. isStaticCheck reads the words of a gate's arguments (check, lint, format), so a test command that passes code with such a word (testthat's reporter = 'check') is taken for a linter and the written-tests check skips it; name the thing another way (CheckReporter$new()) and give the gate a check-routing test (#1500).
36. Deterministic runner output. A reporter that prints spinners or running counts (testthat's progress reporter) changes the output of every run; pick the plain reporter, so the failure signature and the failed-test lines stay stable (#1500).
37. One table, all languages. When a fix for this language touches a shared rule, sweep it across every supported language (the file-issue skill, step 3), and prefer a table row over a wider regex.

4. Tests

• The language's fixture tests/fixtures/languages/<language>.json with goldens for every row you fill (bun run test tests/language-parity.test.ts prints a missing golden ready to paste).
• A describe("gate matrix: <toolchain> (#issue)") block in tests/gate-matrix.test.ts, like the #1749 and #1753 blocks: detection from each manifest (and its absence), the Windows wrapper form, routing (isWholeUnitCheck, fixerCommand on the task's file, coveringChecks), noTestsCollected, the test count and a stable failureSignature, and the "path does not exist" line.
• A real-tool cell under describe.skipIf(!toolchains)("gate matrix: toolchains") with test.skipIf(!available([[tool, "--version"]])), so CI and the laptop run it where the tool is installed.
• Parser cases from real tool output, pasted from a real run, never written from memory.
• A bench-suite task bench/suite/tasks/<language>-<toolchain>-<what>/ (see the bench-suite skill) so the nightly suite and the bench loop run the language.
• The estimate backtest (#1649) when the pull request changes sizing, estimates or forecasts (a language weight, a test-task price).

5. Surfaces and docs

• A row in the gate table of docs/workflows/runtime.md ("Gates, budgets and estimates"), and the test-integrity, baseline and reviewer sections that list languages.
• The feature registry (src/core/features/registry.ts) only when the pull request adds a slash command, /v1 route or desktop screen. Gate detection reaches the TUI gate editor, baked workflow gates and the desktop Gates window through the same detectGates; check that a new gate id renders there and say so in the pull request.
• A Windows, macOS or Linux difference (no shell tools on Windows, bash 3.2 on macOS) is handled in the detector, not left to the user.

6. After the merge

• Tell whoever runs the bench loop that the language is merged, with the toolchain the laptop needs, so the loop adds it to its language rotation.
• A toolchain the cloud containers lack (.NET, PHP, Swift and others) is checked on a device before its bench-suite task graduates.

Lessons

Sources of the rakes above, worth a re-read before a language pull request:

• #1488 and #1717 (one language table, exhaustive per-language records and fixtures) and #1733 (the remaining tables converted).
• The 2026-10-17 sibling audit, #1405-#1435: toolchain coverage is uneven, and one fix leaves the other copies of a computation.
• The 2026-10-20 language sweep, #1678-#1686: fixes written for the language of the run that exposed them missed the other languages.
• Closed Bug issues on toolchains: #587, #1208, #1228, #1326, #1424, #1425, #1508, #1529, #1537, #1582, #1593, #1749, #1753, #1791, #1815, #1832, #1848, #1907, #1913, #1963, #1995.
• The 2026-10-25 language lines (Ruby #2120, Kotlin #2122, Dart #2127, Scala, Swift): rakes 26-34.