Repository navigation
Conversation
A constant is a decision the program has made -- a limit, a retry count, a path, a key, a magic number somebody named. Scattered across the tree those decisions cannot be read as a set, nobody can say which are still true, and the same one gets made twice under two names. const-files names where they live and every declaration outside is an error. It is theme-files for constants. Deliberately not a parser. A rule that has to tell one expression from another needs a tree, which is why test-quality fetches a grammar. This one asks whether a line *declares* a screaming-snake name, and declaration syntax answers that alone -- so it covers all eighteen languages straitjacket calls structured code rather than the nine with a pack, and never reaches the network. It is opt-in only because designating the files is a decision no default can make, and enabling it with no const-files is refused rather than obeyed: every constant would be a finding with nowhere to move it. Three decisions that keep it a rule rather than a nuisance. Declarations, not uses -- referencing a constant is the point of having one, and flagging that would make the rule unsatisfiable. At least two words joined by an underscore, because PI, OK, HTTP, a Go export and a C header guard are all single all-caps words and flagging them would bury the constants among them. And a bare NAME = value counts as a declaration only where the language spells it that way -- in Python, Ruby and Shell only at the left margin, which is what keeps every enum member and every shell local from being a finding. To read code rather than text, rules/comments.rs now yields the comments and a code-only view from one traversal. Masking preserves byte offsets, so a column in the masked line is the column in the real one, and a declaration commented out or quoted inside a string is not one. One lexer answers where the code stops and the prose starts; answering it in two places is how the two answers come to disagree. Run against this repository it reports twenty constants across Rust, TypeScript and shell -- every one a real declaration, with imports, uses and indented shell locals correctly left alone. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WrSzGnURZoupdEfVdwk9pg
Deploying with
|
| Status | Name | Latest Commit | Preview URL | Updated (UTC) |
|---|---|---|---|---|
| ✅ Deployment successful! View logs |
straitjacket | 6da1335 | Commit Preview URL Branch Preview URL |
Aug 30 2026, 08:42 PM |
The first cut matched declaration syntax with a per-language table of keyword patterns. That table is a parser written badly: it cannot tell `MAX_SIZE` in `const MAX_SIZE = 3` from `MAX_SIZE` in `n > MAX_SIZE` without knowing each language's declaration grammar, and every language added means another pattern to get subtly wrong. The question is a tree question, so it moves to beamte, which owns what a construct *is*. `const-declaration` asks the node vocabulary instead: a `_binding` that is neither an import nor a parameter, outside any callable, whose bound name is SCREAMING_SNAKE. That needs no language table at all -- only which pack serves which grammar. What is left here is what beamte refuses to do: fetch a grammar, parse, pick a severity, and name the files constants belong in. The last of those is the policy half, and it stays policy: beamte reports every declaration and `const-files` decides which are licensed. Costs the change accepts: ten languages instead of eighteen, and a grammar download on first scan, which is why the rule is off by default twice over. Gains: no false positive on a use, and shell reassignment in `publish.sh` now reads as the declaration it is. `comments.rs` reverts to main -- the masking it grew existed only to keep the regex off commented-out code, and a parse does that for free. `pack::cached` moves out of `test_quality` so both hosts share one cache rather than each keeping a private one. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WrSzGnURZoupdEfVdwk9pg
A `[patch.crates-io]` entry naming a branch stops resolving the moment that branch is merged and deleted, and the failure lands on whoever pushes next rather than on whoever merged. That is how #63's CI broke: beamte's branch went away under a PR that had not changed, and cargo could not resolve the dependency before a single check ran. A merged commit stays reachable, so the rev survives the merge. The table is temporary either way -- `beamte = "0.3"` above it is already written for the registry, and this whole block goes when 0.3 publishes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WrSzGnURZoupdEfVdwk9pg
Contributor
Author
|
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A constant is a decision the program has made — a limit, a retry count, a path, a key, a magic number somebody named. Scattered across the tree those decisions cannot be read as a set, nobody can say which are still true, and the same one gets made twice under two names.
stray-constreports aSCREAMING_SNAKE_CASEdeclaration anywhere but the files you designate;const-filesistheme-filesfor constants.Or
--stray-const. Enabling it with noconst-filesis refused rather than obeyed: every constant would be a finding with nowhere to move it, which is a configuration nobody means.The analysis is beamte's
The first cut of this rule matched declaration syntax with a per-language table of regexes, here in straitjacket. That was wrong, and the second commit replaces it.
That table is a parser written badly. The entire content of the rule is the difference between declaring a name and using one, and text cannot tell those apart without knowing each language's declaration grammar — so every language added meant another pattern to get subtly wrong, and each one had to re-answer where a string literal ends and where a comment starts.
So the analysis moved to beamte, which owns what a construct is: PowderworksCode/beamte#23 adds
const-declaration, which asks the node vocabulary instead — a_bindingthat is neither a_directivenor a_parameter, outside any_callable. One form, every grammar, no table to drift.What remains in this PR is everything beamte refuses to do: fetch a grammar, parse, choose a severity, and name the files constants belong in. That last one is the policy half and stays policy — beamte reports every declaration it finds,
const-filesdecides which are licensed.What the change costs and buys
n > MAX_SIZETOKEN_VAR=…inpublish.shThe lost eight languages are the honest price: a language going quiet here is now a grammar problem rather than a missing table entry, which is at least a problem with one fix. Both costs are why the rule is off by default twice over — a scan that reaches the network because the tool was upgraded is not a surprise anyone should get for free, and where constants belong is a decision no default can make.
rules/comments.rsreverts tomain. The code/comment masking it grew existed only to keep the regex off commented-out code; a parse does that for nothing.pack::cachedmoves out oftest_quality.rsintopack.rs, so both beamte hosts share one grammar cache rather than each keeping a private copy.Two decisions that keep it a rule rather than a nuisance
MAX_SIZEis the entire point of having it; flagging uses would make the rule unsatisfiable. This is the one the regex could not actually deliver.PI,OK,HTTP, a Go export and a C header guard are all single all-caps words — flagging them would bury the constants among them.One caveat worth knowing: an
enummember spelled in screaming snake is a binding outside a callable, so it reads as a declaration. That is the rule working as specified rather than a bug, andstraitjacket-allowcovers the cases where it isn't wanted.Dogfooding
Run against this repository with
const-files = ["src/config.rs"]: 24 findings across 43 files — 13 Rust, 6 shell, 5 TypeScript.MINIMUM_PACK_ABI,WASI_BADF,HEADER_LINE_LIMIT,LINE_MARKER,FILE_MARKER,STRUCTURED_CODEand the rest. Every one is a real declaration.The four above the regex version's twenty are shell reassignments in
publish.sh, which no tree can distinguish from declarations — shell has no separate declaration form. Spot-checked as correctly not flagged:use crate::language::{…, STRUCTURED_CODE}(a_directive), andhas_facet(&STRUCTURED_CODE)(a use, carrying no binding role at all).Running the tree version over
site/is also what caught a real bug in beamte before it shipped: TypeScript nestsvariable_declaratorinsidelexical_declarationand both carry_binding, somdxFiles(DOCS_DIR)came back as a declaration. Fixed upstream by reporting only the innermost binding.Verification
cargo fmt --check,cargo clippy --workspace --all-targets, 93 workspace tests (5 unit + 12 integration for this rule, covering 8 language cases), the 96 site docs-vs-manifest tests, the musl release build, vale on the prose, and straitjacket clean on itself.The integration tests fetch real packs, for the reason
tests/pack_host.rsgives at length: a test that passes because it found no grammar is worse than no test.Blocked on the same thing as #63
gatewill fail atscripts/publish.sh --dry-rununtil beamte is on crates.io.cargo publishignores[patch.crates-io]by design, so the dry run resolvesbeamte = "0.3"against the registry for real, where only 0.1.0 exists. The detail is on #63: beamte'sv0.2.0release run hit a 403 from an under-scoped crates.io token, which is a beamte settings fix rather than anything either PR can carry.The patch table here pins a commit, not a branch — a branch name stops resolving the moment it is merged and deleted, which is exactly how #63's CI broke under a PR that had not changed.
Merge order: beamte#23 → beamte 0.3 published → drop the patch table here.
Independent of #63 in code — the two share no source and can merge in either order. They are complementary:
env-varsgoverns where the environment may be read, this governs where constants may be declared, and a repository's env module is naturally named in both lists.🤖 Generated with Claude Code
https://claude.ai/code/session_01WrSzGnURZoupdEfVdwk9pg