Use when adding support for a NEW LANGUAGE (Go, C, C++, PHP, ...) or a NEW FRAMEWORK / RUNTIME PRESET (like spring-boot or the JS runtimes) to safelint itself. Walks the complete development checklist so nothing is missed - registry, rule audit, tests, docs, config examples, bundled skill files, drift tests, counts. Invoke before writing any code for such a change.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Use when adding support for a NEW LANGUAGE (Go, C, C++, PHP, ...) or a NEW FRAMEWORK / RUNTIME PRESET (like spring-boot or the JS runtimes) to safelint itself. Walks the complete development checklist so nothing is missed - registry, rule audit, tests, docs, config examples, bundled skill files, drift tests, counts. Invoke before writing any code for such a change.
Adding a language or framework preset to safelint
This is the development checklist for extending safelint itself. The
tracked human walkthrough is docs/contributing/adding-a-language.md (read it
first; it has a worked TypeScript example). This skill is the enforcement
layer: every item below must be done or explicitly recorded as deliberately
skipped with a rationale. History shows the misses are never the engine code -
they are the docs fan-out, the counts, the examples, and the bundled skill
files.
Two workflows: Part A (new language), Part B (framework / runtime
preset). Part C lists the conventions and the validation gate shared by both.
Part A: Adding a new language
A1. Grammar packaging
pyproject.toml: add tree-sitter-<lang> as its own optional extra
(every language is opt-in; the base install ships no grammars), fold it
into the [all] extra (single source of truth - dev pulls
safelint[all]), and extend the extras comment block.
uv lock regenerates; commit uv.lock.
A2. Language module: src/safelint/languages/<lang>.py
Optional grammar import with _GRAMMAR_AVAILABLE flag (silent
except ImportError fallback carries # nosafe: SAFE203, matching the
five existing modules).
EXTRA_NAME, GRAMMAR_INSTALL_HINT, parser factory raising
ImportError with the hint when the grammar is absent.
LanguageDefinition (extensions, comment_node_type,
comment_prefix). Line-comment directives only; block-comment nosafe
is a documented non-feature.
Node-type string constants (rules must import these, never hardcode
strings) and a FUNCTION_TYPES aggregate containing body-bearing
function nodes only (see the Rust module's note on excluding signatures).
A3. Registry: src/safelint/languages/__init__.py
Add the parallel registration block (grammar present -> _REGISTRY;
absent -> _UNAVAILABLE_EXTENSIONS + _UNAVAILABLE_EXTRA_NAMES so the
CLI can hint pip install 'safelint[<lang>]'). Keep the blocks
grep-parallel with the existing five.
Export the definition in __all__. Do NOT reach into _REGISTRY from
elsewhere; supported_extensions() is the public surface.
A4. Per-rule portability audit (the core design work)
For every rule in ALL_RULES, decide: ports cleanly (widen its
language tuple, add a per-language node-type table entry), needs a
language-specific replacement (new rule, category band 1xx-8xx), or is
deliberately skipped (record the rationale - it will be needed in A7).
Add _<lang>-suffixed config defaults in DEFAULTS["rules"] for every
configurable list the ported rules consult (io_functions_<lang>,
sinks_<lang>, assertion_calls_<lang>, ...). Python uses bare keys;
every other language uses the suffix. TS inherits JS via the fallback in
core/_validators.py - a new language gets no fallback unless designed.
Dataflow rules (8xx) need a per-language tracker:
analysis/dataflow_<lang>.py modelled on the four existing ones
(iterative worklists, NOT recursion - SAFE105 polices the codebase).
tests/core/test_engine.py: the _RULES_* allow-list buckets assert
every rule's language tuple EXACTLY. Every widened tuple must move its
rule to the matching bucket or the suite fails.
A5. Plumbing
.pre-commit-config.yaml AND .pre-commit-hooks.yaml: add the
language's pre-commit tag to types_or (both files).
safelint list-rules and the at-a-glance docs table derive from the
registry automatically - verify the new language shows up, no manual
edit expected.
A6. Tests
Per-rule, per-language test files (tests/rules/test_<rule>_<lang>.py):
violation case AND clean case for every ported rule.
Engine-level file: discovery picks up the extensions, // nosafe
(or the language's comment form) suppression works, one known-bad file
fires.
tests/core/test_optional_grammars.py: the missing-grammar hint path.
Coverage gate is fail_under = 97 (pyproject), not the 80 some older
notes claim.
A7. Documentation fan-out (where additions historically go missing)
docs/languages/<lang>.md: full language page - extensions, quick
start, rules table with per-language notes, "rules not registered"
list with rationale per skipped rule, per-language config keys.
mkdocs.yml: nav entry under Languages.
docs/configuration/rules.md: "Currently supported" bullet, the
"Rule scope (current languages)" table, and REMOVE the language from
the "Planned" list.
README.md + docs/index.md: language tables, the "Rule coverage"
paragraph, and every count ("N rules", "N languages", "the
all-five-languages set" wording will need rewording at six languages).
docs/power-of-ten.md: per-language fidelity notes if the language
changes any Holzmann-rule mapping (e.g. its analogue of rule 8/9).
Config examples in BOTH forms for every new _<lang> key:
[tool.safelint.rules.<rule>] (pyproject.toml) and [rules.<rule>]
(standalone safelint.toml). Check docs/configuration/toml.md too if
the language gains a preset axis (see Part B).
Scattered enumerations (the docs that list every language's
extension / grammar / --language value OUTSIDE the headline language
tables - the v2.5.0 Go addition missed all of these; they are not
caught by the language-table edits above):
- docs/configuration/cli.md: the --all-files supported-extension
list AND the --language <LANG> accepted-value list.
- SECURITY.md (repo root - docs/project/security.md is a
build-time copy, edit the root): the supported-versions table
("current" row), the bundled-grammar supply-chain list
(tree-sitter-<lang>), and the "files read from the working tree"
extension list.
- docs/configuration/toml.md: the "enable opt-in rules" walkthrough
(add the language's new opt-in / language-only rules).
src/safelint/skill_files/README.md: counts and layout tree.
The drift tests in tests/test_skill_install.py
(test_skill_documents_every_supported_extension and
..._every_active_rule) fail per client until every extension and
every rule code+name appears - land skill-file updates in the SAME
commit as the registry change or the suite is red.
A9. Bookkeeping
CHANGELOG.md under [Unreleased] -> Added. New language = MINOR,
never MAJOR (project semver policy; JS itself shipped as 1.13.0).
pyproject.toml package metadata: project.description and
project.keywords enumerate the supported languages - add the new one
to both (the v2.6.0 PHP release missed this; PyPI showed a stale
language list). These are part of the scattered-enumeration family.
Stale-count + enumeration sweep: grep -rn for the old counts
("N rules", "five languages", "N cross-language") AND for the
previous language's tokens - its extension (e.g. .rs), its name as a
whole word (grep -w rust, since plain grep does not treat \b as
a word boundary), and its grammar (tree-sitter-rust) - across
docs/, README.md, SECURITY.md, pyproject.toml, plan/README.md
(the roadmap title / table / counts go stale too),
src/safelint/skill_files/.
Every extension list, grammar list, --language value list, and
rule-set enumeration that names the prior languages must gain the new
one; these scattered lists (NOT the headline language tables) are the
most common miss. Tip - substitute the previous language's extension
for <prev-ext> and the new one for <new-ext> (e.g. .rs and .go
when Go follows Rust): for f in $(grep -rl '<prev-ext>' docs README.md SECURITY.md pyproject.toml plan/README.md src/safelint/skill_files); do grep -L '<new-ext>' "$f"; done
surfaces files that mention the old extension but not the new one.
Bump the version - this is the most-missed step (missed twice). A
new language is additive = next X.Y.0. The work lands via the release
branch flow (see CLAUDE.md "Release workflow"): the feature/* -> development PR carries the RC bump (project.version = "X.Y.0rcN",
e.g. 2.6.0rc1); the later development -> main PR flips it to the
production "X.Y.0". Do NOT leave project.version at the previous
release. Keep the CHANGELOG.md heading at ## [Unreleased] - it is
dated only at the production tag. The owner controls release timing and
tagging; the number itself is derived by the additive=MINOR convention,
so apply it as part of finishing the work rather than waiting to be
reminded.
SECURITY.md supported-versions table: add the new release line as
"current" (e.g. **X.Y.x** (current; <Lang> support)) and demote the
prior current line into the maintained band. This rides in the
production (development -> main) bump, and is part of the same
scattered-enumeration family as the sweep above.
A10. Common pitfalls (review-caught in the Go port - design for these upfront)
Real bugs the bot reviewers caught after the Go PR opened. They generalise;
build them in from the start rather than waiting for review:
Validate config lists. A new language-only rule that reads a config
list (error_names_<lang>, panic_calls_<lang>, ...) must pass it
through _validated_string_list(...). A mistyped scalar
(error_names_go = "err") otherwise becomes a set of single characters
and silently disables matching.
Test samples must be VALID in the target language. Tree-sitter
parses leniently, so a type-invalid sample (a value returned from a
void function, a single-target bind of a multi-value call) does NOT
raise a parse error - the test then passes for the wrong reason. Write
samples that would actually compile.
Resolve callee names with call_name, not a hand-rolled
identifier-only check. It covers BOTH bare calls (panic(...)) and
qualified / selector / method calls (log.Fatal(...), pkg.Fn(...)).
Matching only the bare-identifier shape silently drops every configured
qualified-call name (this hit SAFE211's Fatal / Exit).
Resource-cleanup detection (SAFE401-family). The cleanup must occur
after the acquisition and on all exit paths - reject cleanups that
precede the acquire or sit inside a conditional branch; map a
multi-acquirer statement positionally to each handle; a directly
returned acquirer transfers ownership (not a local leak). Tailor the
message to the shape (no-handle / package-scope cases can't use the
normal defer/with/finally fix).
Declaration-site walks (global-state-family). Iterate the
declaration's direct children; do NOT walk() into initializer
expressions (a nested local declaration is not module/package-level
state). Skip blank / discard identifiers (_).
Compound assignments preserve taint in the dataflow tracker:
x += clean is read-modify-write, so it must OR with x's prior taint,
not overwrite it.
Keep the Python that implements all this passing safelint's own
rules - new helpers must obey function_length / nesting_depth /
complexity / no_recursion; factor into small helpers rather than
annotating (the Go port needed several such splits).
Part B: Adding a framework / runtime preset
Precedents: [tool.safelint.javascript] runtime (node / browser / deno /
cloudflare-workers / bun) and [tool.safelint.java] framework (vanilla /
spring-boot). A preset changes rule DEFAULTS for an existing language; it
never changes parsing or rule logic.
B1. Config machinery (core/config.py)
Preset dict whose nested shape mirrors DEFAULTS["rules"], containing
ONLY the keys it overrides (the baseline preset, e.g. node /
vanilla, is the empty dict).
A frozenset of valid preset names.
_resolve_<lang>_<axis>(): reads cfg["<lang>"]["<axis>"], validates,
and on unknown / non-string / non-table values surfaces a
safelint: warning: via core/_diagnostics and falls back to the
default preset. Never raise for a bad preset name.
_apply_* merges the preset into the DEFAULTS copy before the
user's TOML is overlaid, so explicit user keys always win.
B2. Framework-specific structural rules (if any)
New structural rules tied to the framework go in the 9xx band
(9xx is reserved for framework-specific rules - Spring today; never
open a per-language band). Disabled by default; the preset enables
them.
Each follows the full "Adding a new rule" checklist in CLAUDE.md
(registry, defaults, order list, tests, rules.md, language page, all
14 skill files - the rule drift test enforces the last).
B3. Tests
Preset-resolution tests modelled on
tests/core/test_javascript_runtime_presets.py /
tests/core/test_java_framework_presets.py: each preset's overrides land,
user TOML beats the preset, unknown names warn + fall back.
An e2e fixture if structural rules ship (precedent:
tests/fixtures/spring_boot/ + tests/integration/test_spring_boot_e2e.py).
B4. Documentation
Language page (docs/languages/<lang>.md): preset table (when to pick
it, what changes).
docs/configuration/toml.md: a preset section with both TOML forms
([tool.safelint.<lang>] and bare [<lang>] for standalone
safelint.toml) - the Java framework section was missed for two releases;
do not repeat that.
docs/configuration/rules.md: sections for any new 9xx rules.
README.md / docs/index.md: mention on the language row.
Shared language addendum (skill_files/languages/<lang>.md): preset
table; client rule tables gain rows for any new 9xx codes.
CHANGELOG.md under [Unreleased].
Part C: Conventions and the validation gate (both workflows)
Non-negotiables (full detail in CLAUDE.md):
safelint must pass itself: uv run safelint check src/ --all-files with zero
blocking violations (the --all-files flag matches CI; without it the check
only scans git-modified files and can read clean falsely) - including the new
rules on the new code (no recursion, nesting <= 2, complexity <= 10, function
length <= 60).
Never rename or repurpose existing rule names / codes.
No auto-fix, ever; fixes are advisory Suggestions only.
Rule numbering: slot by CATEGORY into 1xx-8xx; 9xx is framework-specific
only.
Indian English ("behaviour", "-ise"); NO em-dashes anywhere.
Additive scope = MINOR semver, always.
Final gate, run all in order; every one must pass:
uv run pytest # coverage gate fail_under = 97
uv run ruff check src/ tests/
uv run ruff format --check src/ tests/
uv run ty check src/
uv run safelint check src/ --all-files # zero blocking violations
uv run mkdocs build --strict # docs integrity, broken anchors fail
After merging: refresh installed skills with
safelint skill install --force (bundled docs changed), and remind users in
the changelog when bundled skill content drifts.
