plc-st-review

plc-st-review

A semantic linter, code reviewer, and team-style enforcer for IEC 61131-3 Structured Text, built for CI on PLC codebases that can’t be compiled outside the vendor IDE. Parses .st files with the tree-sitter-iec61131-3-st grammar and reports semantic problems, not textual diffs.

plc-st-review runs in three modes, each backed by the same 80-check engine:

• Static linter (--lint src/**/*.st), run on every push. 59 single-revision checks for ST bugs: division by zero, out-of-range array indices, infinite loops, TON/CTU/R_TRIG misuse, output reads, unused vars, naming-convention drift, forbidden_symbols, the full PLCopen-aligned set (forbidden statements, IF without ELSE, direct-address use, pointer arithmetic, parameter/global caps, …), an IEC 62443 cybersecurity set (hard-coded credentials / endpoints, unvalidated input, debug pragmas, persistent plaintext secrets), and more.
• PR / MR reviewer (GitHub Action or GitLab CI job), posts inline review comments on lines that triggered findings. Adds 21 diff-based checks that compare the PR against its base: signature drift, outdated call sites, enum removals, timer-value changes, EXTENDS swaps, pragmas, SAFETY_* constant changes, and metric regressions (complexity, nesting, LOC growth). With --project-scope it also flags newly added POUs that nothing in the repo calls.
• Team-style enforcer: drop a .plc-st-review.yml in the repo root listing your naming_conventions (prefix / suffix / pattern per declaration kind) and forbidden_symbols. Both modes pick it up automatically.

Catches the bugs reviewers miss on visual scan:

• A TON.PT changed from T#5s to T#500ms (10× faster).
• A function block grew a required input but only some call sites updated.
• A CASE on an enum that gained a value but no ELSE branch.
• A global variable’s type silently changed and the readers are now broken.
• A constant whose name starts with SAFETY_ had its value changed.

See it in action

Live demo: PR #1, every check the tool ships with, posted on a real PR 👈 open this for the full bot output. The PR exercises the 78 always-on check categories, the ones that run on every PR: each shows up as an inline review comment on the changed line that triggered it, and findings on lines outside the PR’s diff hunks (e.g. POU_DELETED, or a check whose anchor line wasn’t itself edited) collect in a single summary comment at the bottom. Two categories — DEAD_POU_INTRODUCED and MULTI_WRITER_GLOBAL — are opt-in: they need a whole-repo parse, which is too slow to run on every PR, so they’re off by default and triggered on demand (see project scope). The PR is intentionally kept open as a fixture; conversation is locked.

A single finding looks like this in the GitHub UI:

🟧 warn  FB_INSTANCE_NEVER_CALLED
FB instance T3 (TON) is read but never invoked

Outputs of an FB only update when the instance is called.
Reading e.g. `instance.Q` without calling `instance(...)`
returns stale data.

A handful more of what’s posted on PR #1:

FB_ConveyorState.st:26  🟥 error  TIMER_VALUE_CHANGED
                                  Timer T_StartupDelay.PT: T#2s → T#200ms (10.0x faster)

Globals.st:9            🟧 warn   CONSTANT_VALUE_CHANGED
                                  Constant SAFETY_TIMEOUT: T#2s → T#10s
                                  Identifier prefix matches a safety-critical pattern.

FB_Diagnostics.st:49    🟥 error  ARRAY_INDEX_OUT_OF_BOUNDS
                                  arr[15] is out of declared bounds [0..9]

FB_Diagnostics.st:60    🟥 error  INFINITE_LOOP
                                  WHILE TRUE loop with no EXIT statement

Quick start

CI linter (no PR required)

Most industrial ST repos don’t run on a PR/MR workflow, code lands on main after a manual code review and an IDE-side build. You can still get every single-revision check on every push:

# .gitlab-ci.yml, lint every .st file on every push, no MR needed
lint-st:
  image: ghcr.io/heytalepazguato/plc-st-review:v0
  script:
    - plc-st-review --lint "src/**/*.st"

# .github/workflows/lint.yml, same idea on GitHub
name: lint
on: [push]
jobs:
  lint:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: npx plc-st-review --lint "src/**/*.st"

--lint accepts file paths, directories, or globs (*, **, mixed). It parses each .st file in isolation and runs the 59 single-revision checks: the 21 diff-based ones are auto-disabled because there’s no “before” state. Exit code is non-zero when any finding meets reporting.fail_on_severity (default error), so the job fails the pipeline on real bugs.

GitHub pull request

Drop this into .github/workflows/plc-st-review.yml (full example at examples/github-workflow.yml):

name: PLC ST Review
on:
  pull_request:
    paths: ['**/*.st', '**/*.ST']
permissions:
  contents: read
  pull-requests: write
jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: HeytalePazguato/plc-st-review@v0

The action posts findings as inline review comments. Re-runs update existing comments rather than duplicating.

GitLab merge request

Drop this into .gitlab-ci.yml (full example at examples/gitlab-ci.yml):

plc-st-review:
  image: ghcr.io/heytalepazguato/plc-st-review:v0      # or :1.0.0 to pin an exact version
  stage: review
  rules:
    - if: $CI_PIPELINE_SOURCE == "merge_request_event"
  variables:
    GITLAB_TOKEN: $CI_JOB_TOKEN
    GITLAB_URL: $CI_SERVER_URL
    GITLAB_PROJECT_ID: $CI_PROJECT_ID
  script:
    - plc-st-review --gitlab --mr "$CI_MERGE_REQUEST_IID"

The job fetches the MR’s changed .st files, runs the review, posts findings as inline discussions, and updates them (rather than creating duplicates) on re-runs.

Important notes for GitLab users:

• Pull credentials: GitLab runners pull the image from GHCR anonymously. The image is published as public, so no docker login step is needed. If your runner is offline and you mirror the image internally, pull ghcr.io/heytalepazguato/plc-st-review:v0 once and push to your internal registry.
• Self-hosted GitLab: supported. The example above reads GITLAB_URL from $CI_SERVER_URL, which is auto-provided by every GitLab runner regardless of whether you’re on gitlab.com or your own instance.
• Token scope: $CI_JOB_TOKEN works for most projects with default settings. If your instance restricts job tokens, mint a project access token with api scope and set it as a masked, protected CI/CD variable named GITLAB_TOKEN, then drop the GITLAB_TOKEN: $CI_JOB_TOKEN line.
• Same engine as the GitHub Action: every check listed below fires identically on GitLab. There is no GitHub-only path in the engine.

See docs/gitlab-setup.md for the full walkthrough and the common-gotchas list.

CLI

Published on npm:

npm install -g plc-st-review               # https://www.npmjs.com/package/plc-st-review
plc-st-review --lint "src/**/*.st"         # static linting, no PR / base ref needed
plc-st-review --base main --head HEAD      # diff current branch against main
plc-st-review --files old.st new.st        # compare two specific files
plc-st-review --base main --output json --out-file findings.json

Output formats: terminal (ANSI when stdout is a TTY), markdown, json.

The CLI exits non-zero when at least one finding meets or exceeds the reporting.fail_on_severity threshold (default error).

--max-file-size <bytes> overrides the per-file source-length cap (see Source-size cap below). Useful when you have legitimately huge generated FB files; pass 0 to disable the cap entirely. Defaults to the config’s parsing.max_file_size_bytes (1 MB).

Metrics

--metrics is a standalone mode that measures a whole codebase instead of reviewing a diff: cyclomatic complexity, nesting depth, LOC, call-graph fan-in/out, dead code, dependency depth, and more. It does not run the review checks. Full reference: docs/metrics-mode.md.

plc-st-review --metrics src/                              # per-POU + project report
plc-st-review --metrics src/ --sort complexity --top 20  # worst 20 POUs
plc-st-review --metrics src/ --threshold complexity=25   # exit nonzero past the bar
plc-st-review --metrics src/ --format json               # machine-readable
plc-st-review --metrics src/ --format dot | dot -Tsvg -o deps.svg   # call graph
plc-st-review --metrics src/ --format badge              # shields.io URL

Running it on the bundled examples/state-machine/ fixtures prints a ranked report (🟢/🟡/🔴 against the configured thresholds):

Project: examples/state-machine/  (11 POUs, 240 LOC)

Top 10 by complexity:
  FB_LineController  complexity:  26  nesting:  5  LOC:    86  🔴
  FB_ConveyorState   complexity:  10  nesting:  2  LOC:    42  🟢
  Conveyor_HMI       complexity:   6  nesting:  1  LOC:    16  🟢
  FB_AxisRamp        complexity:   5  nesting:  1  LOC:    27  🟢
  FB_SpeedCalc       complexity:   2  nesting:  1  LOC:    18  🟢
  FB_Base            complexity:   1  nesting:  0  LOC:     6  🟢
  FB_BaseV2          complexity:   1  nesting:  0  LOC:     8  🟢
  FB_DiagUnit        complexity:   1  nesting:  0  LOC:    10  🟢
  FB_Legacy          complexity:   1  nesting:  0  LOC:     9  🟢
  FB_MetricsDemo     complexity:   1  nesting:  0  LOC:     9  🟢

Dead code:
  FB_Legacy    (0 callers)
  FB_Watchdog    (0 callers)

Summary:
  Avg complexity: 5  Avg nesting: 0.9  Doc coverage: 9.1%  Dependency depth: 3
  🔴 1 POUs exceed complexity threshold (25)

FB_LineController lands in the red band (complexity 26, past the error threshold of 25), and the dead-code list is short and pointed: two blocks nothing calls. A rendered call graph of the same fixtures is in docs/screenshots/metrics-example.svg.

--format dot emits the call graph (pipe to Graphviz); --format json emits the full per-POU + project report for dashboards or CI gates.

Checks

Each row links to a per-check section in docs/checks-reference.md with a ST code example and a suggested fix. See also docs/check-limitations.md for what each check deliberately doesn’t catch.

Diff-based (compare before vs after)

Category	Default severity	Trigger
SIGNATURE_CHANGED	warn (error on breaking)	A POU’s inputs/outputs/in-outs changed.
CALL_SITE_OUTDATED	error	A caller doesn’t pass a required argument the callee now requires, or passes an unknown argument.
TYPE_MISMATCH	error	A VAR_GLOBAL’s declared type changed between revisions.
ENUM_VALUE_REMOVED	error	A CASE statement references an enum value that was removed.
ENUM_VALUE_ADDED	warn	An enum gained a value but a CASE on the enum has no matching branch and no ELSE.
TIMER_VALUE_CHANGED	info/warn/error by ratio	TON/TOF/TP PT changed; severity scales with the change magnitude (≥2× = warn, ≥10× = error).
CONSTANT_VALUE_CHANGED	info (warn for safety-prefixed names)	A VAR_GLOBAL CONSTANT’s initial value changed. Prefixes like SAFETY_, INTERLOCK_, SIL_ elevate severity.
COMMENT_ONLY	info	The AST is structurally identical between revisions; only comments/whitespace changed.
ARRAY_BOUNDS_CHANGED	error (shrink) / warn (grow)	An array declaration’s [lower..upper] bounds changed.
LOOP_BOUNDS_CHANGED	info/warn by ratio	A FOR loop’s bounds changed; severity rises when the iteration count moves ≥10×.
POU_DELETED	error (with callers) / warn	A POU was deleted; severity depends on whether call sites in the new revision still reference it.
POU_RENAMED	info	Heuristic: a POU was deleted and another with an identical signature was added; suggests a rename.
METHOD_ADDED_TO_INTERFACE	error	An INTERFACE gained a method but a FUNCTION_BLOCK that IMPLEMENTS it doesn’t have one.
INHERITANCE_CHANGED	warn	An EXTENDS clause was added, removed, or changed.
PRAGMA_CHANGED	info	The set of pragmas in a file changed (added or removed).
UNUSED_VAR_INTRODUCED	info	A new local variable was declared but isn’t referenced in its scope.
COUNTER_VALUE_CHANGED	info / warn / error by ratio	CTU/CTD/CTUD PV changed; severity scales with the change magnitude.
COMPLEXITY_INCREASED	warn (error on crossing threshold)	A POU’s cyclomatic complexity rose by more than 5, or crossed the configured error threshold.
NESTING_INCREASED	warn (error on crossing threshold)	A POU’s max control-structure nesting depth rose beyond the configured warn threshold.
LOC_SPIKE	info	A POU’s lines of code grew by more than 50% in a single PR.
DEAD_POU_INTRODUCED	info	A newly added FUNCTION/FUNCTION_BLOCK that nothing in the project calls. Needs --project-scope.

Static (look at the new revision in isolation, filter to bugs new in this PR)

Code-quality + style:

Category	Default severity	Trigger
EMPTY_STATEMENT	info	Lone ; with nothing in front.
UNUSED_RETURN_VALUE	info	Function called as a bare statement; return discarded.
ARRAY_SINGLE_ELEMENT	info	ARRAY [5..5] OF T: single-element array.
VARIABLE_SHADOWING	warn	Local declaration has the same name as a VAR_GLOBAL.
UNQUALIFIED_ENUM_CONSTANT	info	Bare IDLE matches a member of exactly one enum.
IDENTIFIER_CASE_MISMATCH	warn	Reference uses different case than declaration.
UNUSED_INPUT_VAR	info	VAR_INPUT declared but never read.
INPUT_VAR_WRITTEN	warn	VAR_INPUT is assigned inside the POU.
BOOL_COMPARISON	info	IF b = TRUE THEN: comparison adds no information.
REAL_EQUALITY	warn	=/<> against a REAL literal, unreliable on floats.
MULTIPLE_EXIT_POINTS	info	POU has more than one RETURN.
ASSIGNMENT_IN_CONDITION	warn	IF x := y THEN: almost always a typo of =.
COMMENTED_OUT_CODE	info	Comment whose content looks like ST source.
RECURSIVE_CALL	warn	POU invokes itself; risks stack overflow on bounded runtimes.
FORBIDDEN_SYMBOL	error (config-driven)	Identifier matches the repo’s forbidden_symbols blocklist.
ADDRESS_OF_CONSTANT	warn	ADR(c) where c is a VAR_GLOBAL CONSTANT.
UNUSED_OUTPUT_VAR	info	VAR_OUTPUT declared but never written.
OUTPUT_VAR_READ_INTERNALLY	info	VAR_OUTPUT read inside the POU; usually a sign you wanted a local.
NESTED_COMMENTS	info	Block comment contains another block comment.
NAMING_CONVENTION	warn (config-driven)	Declaration name doesn’t match the configured prefix / suffix / regex. See docs/preset-packs.md.
DIRECT_ADDRESS_USED	info	%I0.0 / %QW100-style direct I/O address (PLCopen N1 / CP1).
IF_WITHOUT_ELSE	info	IF / ELSIF chain with no terminal ELSE (PLCopen L17).
FORBIDDEN_STATEMENT	info	GOTO / EXIT / CONTINUE use site (PLCopen L10).
POU_NOT_COMMENTED	warn	POU declaration with no leading comment (PLCopen C2).
NAME_REUSED_DIFFERENT_KIND	warn	Same name reused across declaration kinds, e.g. a global and an enum value (PLCopen N9).
INDIRECT_RECURSIVE_CALL	error	Call-graph cycle through two or more POUs (PLCopen CP13 indirect).
IDENTIFIER_TOO_LONG	warn	Identifier longer than limits.max_identifier_length (PLCopen N6).
IDENTIFIER_CHARSET	warn	Identifier outside the identifier_charset regex (PLCopen N8).
TOO_MANY_PARAMETERS	warn	POU parameter count above limits.max_parameters (PLCopen CP23).
TOO_MANY_GLOBALS_USED	warn	POU touches more than limits.max_globals_used_per_pou globals (PLCopen CP18).
EXTERNAL_VAR_IN_FUNCTION	warn	VAR_EXTERNAL inside FUNCTION / FUNCTION_BLOCK / METHOD (PLCopen CP6).
IMPLICIT_TYPE_CONVERSION	warn	INT-family mixed with REAL-family in arithmetic (PLCopen CP25).

Security (IEC 62443 industrial-cybersecurity checks):

Category	Default severity	Trigger
HARDCODED_CREDENTIALS	error	Secret-named variable with a literal STRING initialiser (62443-4-2 CR 1.5).
HARDCODED_NETWORK_ENDPOINT	warn	STRING literal that’s an IPv4 dotted-quad or http(s)/tcp/opc.tcp/mqtt(s)/modbus/ssh URL (62443-4-1 SI-1).
UNVALIDATED_INPUT_USE	info	VAR_INPUT used as array subscript or divisor with no in-POU relational guard (62443-4-2 CR 3.5).
DEBUG_PRAGMA_IN_PRODUCTION	warn	{attribute 'monitoring'} / 'debug' / 'force_init' pragma in a non-test path (62443-4-1 SI-2).
PERSISTENT_PLAINTEXT_SECRET	error	VAR_GLOBAL PERSISTENT / RETAIN of a secret-named variable (62443-4-2 CR 4.1).

See the IEC 62443 mapping page for the per-clause cross-reference and suggested severity policy by Security Level. The companion MISRA-C and IEC 61508 mapping pages cover the rest of the standards-credibility surface.

Single-revision integrity:

Category	Default severity	Trigger
ENUM_VALUE_UNUSED	info	An enum value is declared but no longer referenced anywhere in the repo.
ENUM_MEMBER_UNKNOWN	error	A qualified ref like E_State.IDEL doesn’t match any declared member of E_State: likely a typo.
ARRAY_INDEX_OUT_OF_BOUNDS	error	A literal index sits outside the array’s declared bounds (arr[15] when arr is ARRAY [0..9]).
DIVISION_BY_ZERO	error	The divisor is a literal 0, or a VAR_GLOBAL CONSTANT resolving to 0.
INFINITE_LOOP	error	WHILE TRUE DO ... END_WHILE; with no EXIT inside the body.
LOOP_BOUNDS_REVERSED	error	FOR loop bounds and step point opposite directions, per spec body never runs, on overflow-wrapping runtimes it runs ~unlimited times.
STATE_UNHANDLED	info	A CASE on an enum has no ELSE and doesn’t cover every enum value, regardless of whether the enum changed.
UNREACHABLE_CODE	warn	A statement sits after RETURN / EXIT / CONTINUE in the same block.
FOR_LOOP_VAR_MODIFIED	warn	Loop counter mutated inside the FOR body (PLCopen L12).
FOR_LOOP_VAR_USED_AFTER	info	Loop counter read after END_FOR — the post-loop value is implementation-defined (PLCopen L13).
POINTER_ARITHMETIC	warn	+ / - / * / / applied to a POINTER-typed local (PLCopen E2).
POINTER_COMPARED	warn	< / > / <= / >= applied to a POINTER-typed local (PLCopen E3).
UNINITIALIZED_VAR_USED	warn	Local variable read before any assignment to it (PLCopen CP3).
TIME_EQUALITY	warn	= / <> between TIME operands; brittle on tick-based runtimes (PLCopen CP28).
MULTI_WRITER_GLOBAL	error	A global variable is written by more than one PROGRAM. Needs --project-scope (PLCopen CP26).
COUNTER_PV_ZERO	error	CTU/CTD/CTUD initialized with PV := 0: Q always TRUE or counter useless.
TIMER_PT_ZERO	error	TON/TOF/TP set with PT := T#0s: fires immediately or never.
TIMER_NOT_DRIVEN	warn	A timer’s Q/ET is read but no call site sets IN.
EDGE_TRIG_REUSED	error	Same R_TRIG/F_TRIG instance fed by multiple different CLK expressions.
FB_INSTANCE_DOUBLE_CALL	warn	Same FB instance invoked more than once in one scope’s scan.
FB_INSTANCE_NEVER_CALLED	warn	FB instance declared, its outputs read, but no call site invokes it.
BISTABLE_DOMINANCE_MISMATCH	info	SR/RS choice mismatches the variable name’s intent (heuristic).

Configuration

Create .plc-st-review.yml at the repo root:

case_sensitive: false          # identifier casing rules; see "Case sensitivity" below

disabled_checks:
  - COMMENT_ONLY

severity_overrides:
  TIMER_VALUE_CHANGED: error   # all timer changes block merge

ignore_paths:
  - "deprecated/**"
  - "third_party/**"

safety_critical_prefixes:
  - SAFETY_
  - INTERLOCK_
  - SIL_

reporting:
  fail_on_severity: error      # exit-nonzero threshold
  comment_style: inline        # inline | summary | both (GitLab/GitHub)

metrics:                       # bands for the metric-regression checks
  thresholds:
    cyclomatic_complexity:
      warn: 15
      error: 25
    nesting_depth:
      warn: 5
      error: 8

parsing:
  max_file_size_bytes: 1000000  # per-file size cap; see "Source-size cap" below

limits:                        # numeric caps for the size-/count-checking PLCopen rules
  max_identifier_length: 32    # PLCopen N6 — drives IDENTIFIER_TOO_LONG (null = off)
  max_globals_used_per_pou: 10 # PLCopen CP18 — drives TOO_MANY_GLOBALS_USED (null = off)
  max_parameters: 8            # PLCopen CP23 — drives TOO_MANY_PARAMETERS (null = off)

identifier_charset: null       # PLCopen N8 — regex every identifier must match; null = off
                               # e.g. '^[A-Za-z_][A-Za-z0-9_]*$' restricts to ASCII + underscore

The metrics block is optional; the values shown are the defaults. COMPLEXITY_INCREASED and NESTING_INCREASED read these bands; LOC_SPIKE fires on any single-PR growth over 50%. See docs/checks-reference.md for the full block (the lines_of_code, comment_ratio, and fan_out keys are accepted now and consumed by the upcoming standalone --metrics mode).

Case sensitivity

Whether two identifiers that differ only in case (Motor vs motor) are the same symbol depends on your toolchain, so plc-st-review makes it configurable with one top-level key:

case_sensitive: false   # default

• false (default) — identifiers are matched case-insensitively. This is what the IEC 61131-3 standard specifies and what CODESYS (and CODESYS-derived IDEs) and Beckhoff TwinCAT do. With this setting the engine resolves a constant declared MaxCount even when it is referenced as MAXCOUNT, and IDENTIFIER_CASE_MISMATCH reports references whose casing drifts from the declaration.
• true — identifiers are matched exactly, byte-for-byte. Set this for B&R Automation Studio, which treats Motor and motor as two distinct variables. In this mode IDENTIFIER_CASE_MISMATCH is automatically disabled (a different case is a different symbol, not a style slip), and checks like VARIABLE_SHADOWING only fire on an exact-case match.

Pick the value that matches the IDE your code is compiled in; the wrong setting can hide real bugs (too loose) or invent false ones (too strict).

Source-size cap

Each .st file is parsed by the tree-sitter grammar running as WebAssembly. To keep a single pathological or hostile file from blowing up memory in the parser, the engine enforces a per-file source-length cap. Files over the cap are skipped with a one-line stderr warning naming the path and treated as empty by every check.

parsing:
  max_file_size_bytes: 1000000   # default: 1 MB

Set 0 to disable the cap entirely (every file is parsed regardless of size). A --max-file-size <bytes> CLI flag overrides whatever is in the config for the current run, so you can do:

plc-st-review --lint "src/**/*.st" --max-file-size 0          # parse everything, no cap
plc-st-review --lint "src/**/*.st" --max-file-size 5000000    # raise the cap to 5 MB for this run

Almost every real ST file is far below the default; raise this only if you have legitimately huge generated FB files. Disable it only if you trust every file in scope (e.g. linting your own first-party code), since the cap is what stops a hostile file from exhausting parser memory.

PLCopen Coding Guidelines preset

The repo ships one opinionated preset: presets/plcopen.yml, tuned to PLCopen Coding Guidelines v1.0. Vendor-neutral, standards-body-published — use it as your baseline if your team doesn’t already have a style guide:

# .plc-st-review.yml
extends:
  - ./presets/plcopen.yml

# Layer your own preferences on top.
naming_conventions:
  bool: { prefix: b }   # we use `b`, not PLCopen's `x`

The full rule-by-rule mapping (which PLCopen rules the preset enforces, which would need new checks, which are out of scope) is on the Pages site under Presets → PLCopen Coding Guidelines.

How it works

Every change is reduced to an AST diff. The engine:

1. Parses the before and after versions of every changed .st file with the tree-sitter grammar.
2. Builds a symbol table per revision: POUs (with parameter signatures), global variables, enums, timer instances, call sites, CASE statements.
3. Hands both tables to each registered check. Each check is a self-contained module under src/engine/checks/.
4. Renders the resulting findings to terminal / Markdown / JSON.

No LLM is involved. Findings are deterministic.

Deterministic by design — no LLM in the loop

Every finding is produced by a check module under src/engine/checks/ that operates on the tree-sitter AST and the symbol table. Each module is a pure function of (before AST, after AST, config) → findings[]. Same inputs, same outputs, byte for byte. No temperature, no sampling, no model version drift.

That choice buys four properties an LLM-based reviewer can’t match:

• Reproducible — a finding that fires today fires tomorrow on the same input. CI reruns match. Two engineers see the same comments. Findings stay citable in code review and postmortems six months later.
• Auditable — every finding maps to one named check with source code you can read. For regulated industries (medical, automotive, energy, pharma) this auditability is the difference between “evidence we can show an inspector” and “a black box we trust.”
• Air-gappable — the shipped Docker image bundles everything the analyzer needs. Mirrored to an internal registry, it runs on offline runners with zero outbound traffic — no model-host endpoints, no API tokens, no SaaS dependency.
• Your code never leaves the network — parsing and analysis run in-process in the CI job. Nothing is sent to a third party. There’s no “we promise we don’t train on your code” disclaimer to evaluate; the architecture makes it impossible.

PLC source typically encodes safety logic, vendor part numbers, calibration constants, customer-specific behavior, and timing assumptions that the customer never agreed to have anywhere but the runner that built it. On that kind of code, deterministic wins by default. See Deterministic (no LLM) for the longer write-up.

Development

Contributors only. End users don’t need this section: the GitHub Action, GitLab CI image, and npm CLI are all ready to use without building from source.

npm install                # plain install — nothing compiles
npm run build              # tsc to dist/
npm test                   # vitest, ~7s, 274 tests across 70 files
npm run lint               # tsc --noEmit

Building from source

npm install

That’s the whole setup. The parser ships as WebAssembly: web-tree-sitter boots the tree-sitter runtime and loads the grammar’s tree-sitter-iec61131_3_st.wasm (published inside the tree-sitter-iec61131-3-st package). There’s no native addon to compile, so no C++ toolchain, no node-gyp, and no platform-specific prerequisites — npm install runs the same on Windows, Linux, and macOS, on any supported Node version.

Roadmap

Likely additions, in rough priority order:

• Standalone CLI binaries: plc-st-review-linux-x64, -darwin-arm64, etc. as GitHub Release assets, for shops that don’t have Node installed. More tractable now that the parser is a single platform-independent .wasm rather than per-platform .node addons — the remaining work is bundling the two wasm files (the grammar’s and web-tree-sitter’s runtime) alongside the packed binary. Revisit if real users without Node ask.
• Optional LLM-powered explanations: a --explain flag that paraphrases deterministic findings in plain English for less-experienced reviewers. Strictly additive, every explanation is grounded in a deterministic finding; the LLM never surfaces new issues.
• Vendor-specific checks: PLCopen MC_* motion patterns, TwinCAT / CODESYS / ABB-specific library FBs. Currently the engine sticks to standard IEC 61131-3 to stay portable across vendors.

License

MIT, see LICENSE.