Lintlang Audit

Audit a config or prompt with LintLang

LintLang is a static linter for the natural-language instructions that control AI agents: system prompts, tool descriptions, and agent configs. It is zero-LLM — deterministic parsing and structural checks only, no model call, no telemetry, no network access during a scan (https://github.com/hermes-labs-ai/lintlang).

Run this skill on request for the file the user names. It reports a scan verdict and does not rewrite the file or block a tool call.

What to do

1. Resolve the target. Audit the file or files the user named. If no file was named, ask which one — do not guess, and do not sweep every candidate in the repository.

   LintLang reads .yaml, .yml, .json, .md, .txt, .prompt, and .py. A .py file is scanned by AST extraction for embedded prompts and uncalibrated thresholds (P1/P2); it is not general Python linting, so do not offer this skill as one.

2. Resolve a runner, in this order. Stop at the first that works.

   • lintlang --version prints lintlang 0.8.2 → use lintlang for both the version check and scan.

   • Otherwise, if uvx is available and the pinned release runs, use it with no persistent install and no PATH change:

     uvx --from lintlang==0.8.2 lintlang --version

     Use uvx --from lintlang==0.8.2 lintlang for the scan too. Keep the ==0.8.2 pin so an unreviewed newer release is never fetched. This downloads the package into uv's cache once; the scan itself still makes no network call.

   • Otherwise, if lintlang --version succeeded with another version, use that installed lintlang command and report its version with the result; available checks and findings may differ from 0.8.2.

   • If neither runner works, stop and relay the install line: python -m pip install lintlang==0.8.2. Do not install anything persistently on the user's machine yourself.

3. Scan, once, with JSON output. Run one of these commands, matching the runner that worked in step 2:

   file='./prompt.md' # replace with the exact selected path, shell-quoted
   lintlang scan --format json -- "$file"

   file='./prompt.md' # replace with the exact selected path, shell-quoted
   uvx --from lintlang==0.8.2 lintlang scan --format json -- "$file"

   Set file before running the chosen command; ./prompt.md is only an example. To scan more files, append each additional quoted path argument after "$file", for example "$next_file" after assigning next_file. Treat every named path as data: pass it as one argv element. If using a shell, put each path in a variable and quote the expansion as shown; never paste a raw path into a shell command. The -- keeps a path that begins with - from being read as a flag. JSON is an array with one object per input file, each with file, verdict, input_error, skipped, and structural_findings.

   Add --fail-on fail (blocks on CRITICAL/HIGH) or --fail-on review (blocks on MEDIUM and above) only when the user asked for a gate or a CI exit status. See the exit codes below before you do.

4. Read input_error and verdict before anything else.

   • input_error is non-null → the scan never ran on that file (missing file, unreadable, unsupported). verdict is ERROR. Report what the message says. This is not a clean result.
   • verdict is SKIPPED → no covered agent-facing content was inspected. Report the skipped reason. Do not call this a pass.
   • verdict is FAIL (CRITICAL or HIGH present), REVIEW (MEDIUM present), or PASS (nothing above LOW).

5. Report. Summarise; do not paste the whole payload back. Lead with the verdict and the counts by severity, then the specific findings that matter, naming each by its code (H1.1, H1.6, P2, …) and location. Say which file each finding belongs to when more than one was scanned.