Skip to content

0.4.0: see what your agent loads

Choose a tag to compare

@Katta041 Katta041 released this 15 Nov 06:45
· 74 commits to main since this release

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_plan tool, 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 (a paths, globs or applyTo pattern that matches no file in the repo), nested-duplicate and nested-contradiction (notes for a subfolder file that repeats or contradicts the one above it), and sec-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 .claude folder or plugin over Claude Code's 8,000-character listing budget) and skill-shadowed-command (a .claude/commands file 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 as AKIAIOSFODNN7EXAMPLE or sk-xxxx are 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 .gitignore covers) and agent-transcript-in-repo (committed session transcripts and copied agent state). --include config now also reads files named .gitignore, .cursor/cli.json and the first 4 KB of <uuid>.jsonl files. Nothing new is read without it.
  • hidden-instruction in 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, fix and loads, so clients that list prompts can run AsItIs without knowing the tool names (/mcp__asitis__lint in Claude Code), and resources asitis://findings and asitis://budget, plus asitis://findings/<path> and asitis://budget/<path>.
  • asitis lint --badge prints a "checked with AsItIs" README badge after a clean run.
  • --allow-partial and the Action input allow-partial.

Changed

  • context_budget is 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/rules folder no longer counts at the root, .claude/AGENTS.md counts when there is no CLAUDE.md, Codex's AGENTS.override.md replaces AGENTS.md and an empty AGENTS.md is skipped, a .cursorrules below the root no longer counts, and files over 4 MiB are skipped for Claude Code.
  • broken-import reads @ 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 @, so See @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#setup is no longer called missing, because Claude Code drops the #setup. Gemini CLI keeps it, so GEMINI.md is checked the Gemini way. context_budget and load_plan don't count an import that loads nothing.
  • broken-import follows 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 a toolExecutionNotifications line, --format github a ::warning. The Action annotates and writes the SARIF file for what it found, then fails with 4 unless allow-partial is set. Build and cache folders such as .pnpm-store, Pods, bazel-* and Python caches are skipped, while build, dist, out and 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-link skips 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-key and compatibility-too-long are notes for skills outside the folders Claude Code and Agent Skills readers load. duplicate-rule keeps 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:review in review/) 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-shadowed skips files under .agents/.
  • sec-permission-allow masks 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@v0.3.0 keeps 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 a hidden-instruction error (#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_steering on 10 MB of [a lines took about 2.9 s against a 3 s budget; now 1.3 s with the same output.
  • asitis lint on 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.