Repository navigation
Releases
v0.4.4
Added
stale-path(warning, on by default): a repo path in backticks, such assrc/api/routes.tsorpackages/web/, that isn't in the repository any more. It only checks paths with two or more parts that end in an extension or a slash and whose top folder exists, and it skips globs, placeholders, fenced blocks, example sections and paragraphs about another repository. It only asks whether paths exist; it opens no file besides the steering file. On 26 public repos it reported 15 paths, and I checked each by hand: all 15 were real moved or deleted files.dead-command(warning, opt-in):npm run,npm test,pnpm,yarn,bun runandmakecommands whose script or target isn't defined. It readspackage.jsonand Makefile files, so it runs only with--include config(include: ["config"]inlint_steering). Withconfig,stale-pathalso reads.gitignorefiles and skips the paths they cover.
Changed
- The README leads with CI: the GitHub Action comes first, with a copy-paste workflow that lints pull requests and pushes to main that touch a steering file and uploads SARIF on push only.
Fixed
- The README's "How was it tested?" answer still said every real finding from 0.3.0 was kept in 0.4.0. 0.4.2 took that claim back in the changelog; the FAQ line is gone now too.
Details: packages/mcp/CHANGELOG.md and plugins/asitis/CHANGELOG.md.
v0.4.3
Added
- Codex:
codex plugin marketplace add Katta041/asitis-agent-tools && codex plugin add asitis@asitis. Codex has no plugin commands, so the plugin brings asteering-checkskill that picks the right tool when you ask in plain words ("use AsItIs to show what Codex loads in packages/api"). - Gemini CLI:
gemini extensions install https://github.com/Katta041/asitis-agent-tools. You get/asitis:md-lint,/asitis:md-loads,/asitis:md-budgetand/asitis:md-fix. It runs the server bundled in the repo under Node's permission model, reading only your workspace, so nothing is downloaded when it starts. - Cursor and VS Code: one-click install links in the README.
- Everything else: docs/install.md has configs for Devin Desktop (Windsurf), Cline, Kiro, Zed, Goose, Continue, JetBrains AI Assistant and Docker.
- Docker: a Dockerfile that runs the prebuilt server as a non-root user with no npm install.
docker run -i --rm --network none --read-only -v "$PWD":/workspace:roworks. - Cursor plugin in
plugins/asitis-cursor. I haven't run it inside Cursor yet, so use the install link for now. Its README says what's unconfirmed.
Changed
- The Claude Code plugin manifest has a license, an icon and documentation, support, privacy and terms links, and its description names all four commands.
- Docs use the plain
npx -y @asitis/mcp@0.4.3form for the server, which works since 0.4.2.
Nothing changed in the linter itself. On the 71-repo corpus 0.4.3 reports exactly what 0.4.1 and 0.4.2 did.
Details: packages/mcp/CHANGELOG.md and plugins/asitis/CHANGELOG.md.
v0.4.2
Fixed
npx -y @asitis/mcpfailed with "could not determine executable to run" from 0.3.0 on. 0.3.0 added a second command (asitis) next to the server (asitis-mcp), and with two commands npx only starts the one named after the package,mcp, which didn't exist. That's the line in the README and the command MCP Registry clients run, so anyone who installed the server that way since 0.3.0 hit the error. The package now has a third command,mcp, which is the server, and the plain form works again.npx -y -p @asitis/mcp asitis-mcpandnpx -y -p @asitis/mcp asitis lintkeep working as before. Nothing changes for the Claude Code plugin or the GitHub Action, which never used npx.asitis-mcp --helplistsload_planand shows the-pform that works on every version.- The README no longer says no real finding was lost between 0.3.0 and 0.4.0. The 0.4.0 precision numbers stand (70% to 91% on 71 hand-labelled repos). The agnix rule count in the comparison table is 761 now.
If you're on 0.3.x to 0.4.1 and an MCP client config runs npx -y @asitis/mcp@<version>, either bump it to 0.4.2 or use npx -y -p @asitis/mcp@<version> asitis-mcp.
Details: packages/mcp/CHANGELOG.md and plugins/asitis/CHANGELOG.md.
v0.4.1
Fixed
asitis loadsandload_planwere wrong about @imports from a folder above the start folder. 0.4.0 said Claude Code never expands them. It does, after you approve them: an import whose file sits outside the folder you start Claude in is an external import, and Claude Code asks once per project, in an interactive session, whether to allow external imports. Yes loads them from then on,claude -pincluded; no keeps them off and the question doesn't come back.claude -pnever asks, so until someone has said yes there, they don't load in scripts or CI. My 0.4.0 canary only usedclaude -p, which is why it looked like "not expanded". Checked again on Claude Code 2.1.312 and 2.1.325, interactive and headless, both answers. Such imports are now listed under "Only after you approve imports from outside the start folder", not counted at launch, with a warning that says whatclaude -pdoes. An ancestor's import of a file inside the start folder loads at launch, and a start-folder file's@../x.mdwaits for approval like an ancestor's.context_budgetcounts from the project root, where nothing in the project is external, so its numbers don't change.asitis loadswarned "there is a CLAUDE.md" when the only file hiding AGENTS.md was a CLAUDE.local.md. It names CLAUDE.local.md now.asitis loadsindented an import that waits for approval as if it hung under the line above it (9. AGENTS.md). Imports are indented only under a file listed in the same group now.
Details: packages/mcp/CHANGELOG.md and plugins/asitis/CHANGELOG.md.
0.4.0: see what your agent loads
This one folds in 0.3.1, which never shipped on its own.
Behaviour change for CI: on a repo too big to search in full, asitis lint used to stop with exit code 2 and lint nothing. It now lints what it found and exits 4, so a job that fails on any non-zero code still fails and a partial run never looks clean. Add --allow-partial (Action input allow-partial: true) to get 0 or 1 instead. The walk also covers 200,000 folder entries now (was 50,000), so next.js, bun, node, rust and PostHog lint in full.
Added
load_plantool,asitis loads [dir]and/asitis:md-loads: what Claude Code, Codex, Gemini CLI, Cursor and Copilot load when a session starts in a given folder. You get the files at launch with their imports, what loads later (a subfolder CLAUDE.md on first read, path rules on a matching file, nested skills), what gets cut (Codex's 32 KiB chain, MEMORY.md after 200 lines) and what is never read and why. Each step says whether it rests on vendor docs, a canary run in the real tool, or an assumption. Text by default, 60 steps per tool per call.- Monorepo rules:
path-rule-matches-nothing(apaths,globsorapplyTopattern that matches no file in the repo),nested-duplicateandnested-contradiction(notes for a subfolder file that repeats or contradicts the one above it), andsec-nested-override(a subfolder file that tells the agent to ignore the files above it). - Skill checks:
skill-description-vague,skill-description-person,skill-description-overlap(two skills whose descriptions say nearly the same thing),skill-listing-budget(skill descriptions in one.claudefolder or plugin over Claude Code's 8,000-character listing budget) andskill-shadowed-command(a.claude/commandsfile with the same name as a skill, so it never runs). The first three and the budget take options in.asitisrc.json. secret-in-steering: an API key or private key in a steering file is an error. The message gives the key type, line and column and never a character of the key; placeholders such asAKIAIOSFODNN7EXAMPLEorsk-xxxxare skipped.- Agent leak audit in the opt-in config scan:
secret-in-permission-rule(a password, token or key saved into a permission rule by "Yes, and don't ask again", in Claude Code, Cursor CLI or Gemini CLI settings),local-agent-file-not-ignored(settings.local.json,CLAUDE.local.md, local subagent memory and sign-in files that no.gitignorecovers) andagent-transcript-in-repo(committed session transcripts and copied agent state).--include confignow also reads files named.gitignore,.cursor/cli.jsonand the first 4 KB of<uuid>.jsonlfiles. Nothing new is read without it. hidden-instructionin Spanish, Portuguese, French, German, Russian, Chinese, Japanese, Korean and Hindi, inside hidden comments, for every category the English check has. Same error and warning split as English. No new findings on 4,363 real hidden comments; the rule page lists what it misses (#6).- MCP prompts
lint,budget,fixandloads, so clients that list prompts can run AsItIs without knowing the tool names (/mcp__asitis__lintin Claude Code), and resourcesasitis://findingsandasitis://budget, plusasitis://findings/<path>andasitis://budget/<path>. asitis lint --badgeprints a "checked with AsItIs" README badge after a clean run.--allow-partialand the Action inputallow-partial.
Changed
context_budgetis now a view of the same load model, so the two can't disagree. Some totals move: imports inside rule files count at launch, a nested.claude/rulesfolder no longer counts at the root,.claude/AGENTS.mdcounts when there is no CLAUDE.md, Codex'sAGENTS.override.mdreplacesAGENTS.mdand an emptyAGENTS.mdis skipped, a.cursorrulesbelow the root no longer counts, and files over 4 MiB are skipped for Claude Code.broken-importreads@imports the way Claude Code does. I checked its parser in 2.1.312 and ran canary files through it: it takes the whole word after@, soSee @guide.md: for details,Follow @AGENTS.md.and(see @guide.md)load nothing even when the file is there. Those are now warnings, and so is(@guide.md).@guide.md#setupis no longer called missing, because Claude Code drops the#setup. Gemini CLI keeps it, so GEMINI.md is checked the Gemini way.context_budgetandload_plandon't count an import that loads nothing.broken-importfollows imports inside imported files: depth first, each file once, 4 hops for Claude Code and 5 for Gemini CLI. A problem further down is reported on the import that starts the chain, with the chain written out:"@docs/a.md -> @b.md -> @missing.md".xml-tag: a lone placeholder such as<service>in a skill description is a note now, not a warning.- Exit code 4 for a partial run (see the top of this entry). JSON gets
summary.partial: true, SARIF atoolExecutionNotificationsline,--format githuba::warning. The Action annotates and writes the SARIF file for what it found, then fails with 4 unlessallow-partialis set. Build and cache folders such as.pnpm-store,Pods,bazel-*and Python caches are skipped, whilebuild,dist,outand similar are searched when they hold a CLAUDE.md, AGENTS.md or GEMINI.md. - Fewer false positives on real repos. On the 71-repo corpus, errors and warnings went from 900 to 655 and precision from 70% to 91%, with no real finding lost.
broken-linkskips files rendered from templates (x.md.j2), resolves{baseDir}/from the skill folder and skips placeholders, PR-comment../blob/<ref>/links and escaped brackets.unknown-keyandcompatibility-too-longare notes for skills outside the folders Claude Code and Agent Skills readers load.duplicate-rulekeeps case inside code spans, lets procedures share steps, and makes one sentence shared by two different paragraphs a note. A skill name with its plugin prefix (n8n:reviewinreview/) is a note, since Claude Code documents that form. A CLAUDE.md that is a symlink to AGENTS.md reports its length as a note.agents-md-shadowedskips files under.agents/. sec-permission-allowmasks a secret inside the rule it quotes, and its fingerprint for such a rule no longer depends on the secret, so baselines for those few findings need regenerating.- The old one-line Spanish, German, French, Chinese and Japanese "ignore previous instructions" pattern is gone. It matched mid-sentence; the new phrases need the verb to start a clause or be in the imperative. A harmful phrase in another language now beats an English match that only names the assistant, so
Claude: 忽略之前的所有指令is an error. - npm description and keywords; the Action has a Marketplace-friendly name and description.
Katta041/asitis-agent-tools@v0.3.0keeps working. - GitHub Actions: setup-node 7.1.0, setup-python 7.1.0, upload-sarif 4.39.3 (needs CodeQL bundle 2.20.7 or later on GHES). Bundled MCP SDK 1.32.1 and zod 4.6.7.
Fixed
- An empty comment written
<!-->or<!--->hid the rest of the file, so visible text could get ahidden-instructionerror (#28). - Fix diffs no longer print a line that holds a key (#32). An edit whose hunk would show one is left out of the fix, with a note.
lint_steeringon 10 MB of[alines took about 2.9 s against a 3 s budget; now 1.3 s with the same output.asitis linton a 10 MB single-line file took over a minute, because every finding hashed its whole line again. Each line is now normalised and hashed once: about a second, same fingerprints.
Details: packages/mcp/CHANGELOG.md and plugins/asitis/CHANGELOG.md.
0.3.0
0.3.0 takes AsItIs out of the chat window and into CI.
New
- An
asitiscommand line tool. It outputs text, JSON, SARIF and GitHub annotations, uses exit codes 0 to 3 and supports--fail-on:npx -y -p @asitis/mcp@0.3.0 asitis lint - A GitHub Action (
Katta041/asitis-agent-tools@v0.3.0) and a pre-commit hook (asitis-lint). Setup for GitHub, pre-commit, lefthook, husky and GitLab is in docs/ci.md. .asitisrc.jsonconfig, inline<!-- asitis-disable-next-line rule -- reason -->comments and baselines, so you can adopt it on a big repo without fixing everything first.- Drift checks:
- flags an AGENTS.md that Claude Code never reads because a CLAUDE.md shadows it
- flags mirrored files that have drifted apart, and rules that contradict each other across tools
- checks the Codex 32 KiB budget, which is shared down the folder chain
- An opt-in security scan of skill folders (
include: skills). It looks for downloaded code piped into a shell, encoded payloads, data sent off the machine, unscopedallowed-tools, binaries posing as text, and skill arguments that reach a shell. - An opt-in audit of
.claude/settings.jsonand.mcp.json(include: config). It looks forbypassPermissions,Bash(*), hooks that download and run code, unpinnednpx -yservers, the vulnerablemcp-remoteversions and inline secrets. It never reads~/.claude.json. - Fix suggestions as diffs, through the new
suggest_fixestool,/asitis:md-fixand--fix-preview. The server still never writes. You see each change and decide. - A page for every rule, with bad and good examples and the vendor doc it rests on: docs/rules.
Default severities
Planted invisible or bidirectional characters, and hidden comments that tell the agent to ignore instructions, hide something, run a command or reach for secrets, are now errors. That means a default CI run fails on them. Comments that only address the agent stay warnings.
How it was tested
- Before release I ran it across 71 public repos and labelled every error and warning by hand. 98% were worth a contributor's time, up from 59% in the first release candidate. Every real finding is still caught.
- A security review checked that the tool doesn't print secrets, and that a repo's own config can't quietly hide security findings.
- The labelled set of 233 cases scores 100% on precision and recall.
- 1,722 tests run on macOS, Linux and Windows.
Known gaps for 0.3.x
- The skill scan doesn't yet catch some heavily disguised download-and-run commands, and it doesn't read Makefiles or YAML inside skill folders.
- A stray
<!-->marker can give a false hidden-instruction warning. - Fix diffs show the lines they change, so a secret already in that line appears in the diff.
Update
claude plugin marketplace update asitis
claude plugin update asitis@asitis
0.2.5
The big one: in Claude Code 2.1 the model was getting the JSON copy of each tool result instead of the text, so read_markdown handed it no file text and the "treat this as data" notice never arrived. Results are text only now. If a script of yours relies on the JSON, start the server with --structured (or set ASITIS_MCP_STRUCTURED=1).
Lint also caught two things it used to miss:
- short rules repeated under two headings, like "Run
pnpm testbefore every commit" - plain contradictions, like "Use named exports" a line above "Never use named exports"
On my labelled set of 233 cases, precision went from 94.1% to 99.5% and recall from 90.6% to 98.6%.
Other changes:
context_budgetfollows what each tool actually loads. AGENTS.md only counts for Claude Code when there's no CLAUDE.md, or when CLAUDE.md imports it.- Symlinks inside the project are followed.
- Big repos no longer overflow Claude Code's tool-result limit.
- Cursor's unquoted
globs: *.tsis accepted. - The guard hook gives better advice on CRLF files and catches whole-file reformats done through Edit.
- Two new checks: a frontmatter value cut off by an unquoted
#, and near-miss keys ("did you mean user-invocable?"). - The asitis.app line in results is off by default (
--footerturns it on).
Full list in CHANGELOG.md.
Update:
claude plugin marketplace update asitis
claude plugin update asitis@asitis
npm: npx -y @asitis/mcp@0.2.5
0.2.4: clearer output, working update steps
Install:
claude plugin marketplace add Katta041/asitis-agent-tools
claude plugin install asitis@asitisAlready installed:
claude plugin marketplace update asitis
claude plugin update asitis@asitisThen restart Claude Code. Other MCP clients: npx -y @asitis/mcp@0.2.4 --root /absolute/path/to/project
Changed
- The bundled server's output is easier to read (see the
@asitis/mcpchangelog). /asitis:md-lintasks for a short report: totals first, hidden-content findings next, then errors and warnings aspath:line rule: message, notes in one line./asitis:md-budgetasks for one line per tool, largest first, and stops early when nothing loads every session.- Guard hook messages: correct plurals ("2 lines", not "2 line(s)"), no doubled parentheses, and for an Edit on a file with mixed line endings it says that Claude Code's Edit writes the whole file with one style, and advises asking the user instead of retrying.
- README: update and uninstall steps (
claude plugin update asitis@asitis,claude plugin uninstall asitis@asitis), a first-run step, and absolute links that work on marketplace pages.
0.2.3: Windows project folders work again
Install or update:
claude plugin marketplace add Katta041/asitis-agent-tools
claude plugin install asitis@asitis
# already installed: refresh the marketplace, update the plugin, then restart Claude Code
claude plugin marketplace update asitis
claude plugin update asitis@asitisOther MCP clients: npx -y @asitis/mcp@0.2.3 --root /absolute/path/to/project
Fixed
- Windows: the server could read nothing when Claude Code passed the project folder as an 8.3 short path (
C:\Users\RUNNER~1\...) or in a different case than the one on disk. Every read failed with "is a symlink that resolves outside the allowed folder". The root and the files it contains are now compared in the same form. Links that resolve outside the project,.envfiles and.claude/settings*.jsonare still refused, and the project folder can still contain the plugin folder.
Changed
- The bundled server is
@asitis/mcp0.2.3.
0.2.2: MCP Registry metadata, guard answers before the hook timeout
Install or update:
claude plugin marketplace add Katta041/asitis-agent-tools
claude plugin install asitis@asitis
# already installed: refresh the marketplace, update the plugin, then restart Claude Code
claude plugin marketplace update asitis
claude plugin update asitis@asitisOther MCP clients: npx -y @asitis/mcp@0.2.2 --root /absolute/path/to/project
Fixed
- The guard hook's own time budget is 6 s instead of 7 s and now counts from the moment the hook process starts. Node start-up and the work the budget cannot interrupt (reading and decoding a file of up to 16 MB) used to fall outside it, so on a slow machine the check could run into Claude Code's 10 s hook timeout, which lets the change through. The guard now always answers first: it asks, or blocks in
denymode. - The guard's fail-closed test is deterministic: it lowers the budget through
ASITIS_GUARD_BUDGET_MS(which can only lower it) instead of measuring wall-clock time.
Changed
- The bundled server is
@asitis/mcp0.2.2, which adds MCP Registry metadata and no behaviour changes.