Skip to content

Releases

v0.5.10

Choose a tag to compare

@LadsonDavid LadsonDavid released this 21 Nov 17:01

What's new

  • One rule for what you see: plain malveon and the notice when your agent finishes list only problems the code proves, with or without an open session. Everything weaker is counted ("N advisory notes not shown") and listed by malveon --all.
  • A skipped test is reported once, as a weakened test, not a second time as unfinished code. A test that skips only when the machine lacks something (Go t.Skip guards, JavaScript test.skip(condition, reason)) is no longer called a switched-off test.
  • An import through an @/ alias is judged only when your tsconfig.json or jsconfig.json says where the alias points, so a guess can no longer produce a false "import cannot work".
  • The "two routes claim one address" finding is removed: it was right in too few projects to interrupt anyone.
  • Anonymous usage counts can be turned off. --no-telemetry, MALVEON_NO_TELEMETRY=1 or DO_NOT_TRACK=1 stops usage telemetry and the background update lookup, and no anonymous ID is created. Network calls still happen only on an explicit malveon update or malveon account, and for the paid-licence refresh.

The CLI assets are signed and include SHA-256 checksums and build provenance. Source analysis does not execute requests or verify rendered behavior. The GitHub plugin launchers use npm 0.5.10 once it is published.

v0.5.9

Choose a tag to compare

@LadsonDavid LadsonDavid released this 21 Nov 08:17

What's new

  • Plain malveon and the notice when your agent finishes now list only problems the code proves. Everything weaker is counted ("N advisory notes not shown") and listed by malveon --all. When nothing is proven it says "No confirmed problems in changed files".
  • Weakened tests are caught more often: a deleted or newly skipped test, two or more assertions removed, or a test left with no assertion. A single removed assertion in a test that still asserts is listed as likely, not proven.
  • A call to an address nothing serves is no longer reported when a wildcard handler covers that address. A catch-all that serves your app's pages does not hide a missing /api route.
  • TODOs, debug logs, sample data and stub messages are listed only by malveon --all.
  • The editor extension shows your plan and runs left, and the setup prompts say "project" instead of "repo".

The CLI assets are signed and include SHA-256 checksums and build provenance. Source analysis does not execute requests or verify rendered behavior. The GitHub plugin launchers still use npm 0.5.8 until npm 0.5.9 is published.

v0.5.8 — response-field checks and plugin hooks

Choose a tag to compare

@LadsonDavid LadsonDavid released this 20 Nov 12:42

What's new

  • Reports when JavaScript or TypeScript reads a key missing from a matching Next.js GET handler's literal JSON response. Both source locations appear in the finding. Dynamic responses remain unchecked.
  • Omits several findings based on guesses, counts what it could not judge, and limits the default report in a dirty Git checkout to findings involving changed files.
  • Makes desktop notifications opt-in for proven problems. Browser layout remains behind the experimental flag.
  • Bundles prompt and completion hooks for Claude Code, Cursor, and Codex in the GitHub plugin. Codex requires hook trust review. The plugin launchers still use the published npm 0.5.7 engine until npm 0.5.8 is available.

The CLI assets are signed and include SHA-256 checksums and build provenance. Source analysis does not execute requests or verify rendered behavior.

v0.5.7 — desktop notifications and more reliable automatic checks

Choose a tag to compare

@LadsonDavid LadsonDavid released this 19 Nov 10:43
  • Added Windows desktop notifications for automatic checks in Codex, Cursor, and Antigravity. These appear outside agent chat and can be disabled with MALVEON_NOTIFY=0.
  • Fixed Cursor on Windows so automatic checks can read prompt and completion hook input. Malveon status now shows the latest automatic check.
  • Fixed local browser checks from relative project paths and allowed more time for Chrome or Edge to start.
  • List supported multi-part requests separately and report which parts remain unchecked. Added bounded quoted-text checks and same-chat follow-ups.
  • npm installations can update through npm from malveon update.

v0.5.6 — simpler setup across projects

Choose a tag to compare

@LadsonDavid LadsonDavid released this 19 Nov 06:50

Run malveon on, press Enter, then open a new agent session. Setup now applies across projects for the selected agents.

  • Agent detection looks for executables. Project settings folders alone do not select Codex, Antigravity, or another agent.
  • Folders without a Git baseline can use source checks. Before-and-after checks still need Git history.
  • malveon status now handles an indefinite pause.
  • For setup in one folder, use malveon on --root <folder>. Codex still asks you to approve its hooks in /hooks.

The checks are warn-only. The public code-reading core uses the same source as v0.5.5.

v0.5.5 — clearer request checks and fewer false alarms

Choose a tag to compare

@LadsonDavid LadsonDavid released this 17 Nov 05:22

v0.5.5

  • Show the captured request, the agent's recorded interpretation and pre-edit plan, and supported source-effect comparisons. Declarations are not user approval or independent proof of intent.
  • Keep new data-changing actions, package additions, table references and literal outside requests visible for review when a request has no source anchors.
  • Check explicit quoted-text additions/removals directly against source, with remaining behavior unverified.
  • Use one consistent terminal summary and retain detailed evidence under --all without dumping unmatched prompt words.
  • Recognize supported literal GET forms and page-link connections without inventing API handlers.
  • Stop accusing unchanged files as missing behavior, and distinguish deleting content from deleting the containing file.
  • Improve supported existing-change checks, caller locations and optional compiler comparison under --run; preserve baseline limitations.
  • Improve first-notice delivery and supported host detection. Actual native delivery depends on matching configured builds and remains separately verifiable.

These are bounded static checks. Rendering, arbitrary business behavior, follow-up intent, long-prompt semantic coverage and measured accuracy across unseen projects remain unverified or unsupported. No general completion or production-readiness guarantee is made.

Validation: published public-core tests and CI passed on Windows, macOS and Linux. CLI package tests were run on Windows with the published core and workspace replacements disabled. One optional-browser profile-cleanup test failed once; isolated and full browser-suite reruns passed, with the initial failure retained. Four CLI targets were cross-built and signed; the exact Windows release bytes passed the quoted-text-removal regression. Runtime behavior on macOS/Linux and current native editor/hook delivery remain separately unverified.

v0.5.4

Choose a tag to compare

@LadsonDavid LadsonDavid released this 14 Nov 06:09

A quieter, clearer first run, and the "done" notice now says something.

  • A shorter malveon on. One plain question of about ten lines (Enter means yes) instead of a wall of hook commands. malveon on --details still shows every file and command first.
  • The real result when your agent finishes. The "done" notice in Claude Code and Codex now shows the short result (the verdict and up to three findings, or "Nothing found in the code I read") instead of one generic sentence. It is off until you say yes in malveon on; a yes writes a setting in Malveon's own folder. It never changes what the agent is told. The text appears in your agent's window, and whether Claude Code or Codex also keep it in the conversation has not been checked.
  • Quieter. The notice speaks only when a project file changed since the last clean check. Your own new prompts no longer count as work. A prompt that names no file, URL or code name is mentioned once, pasted terminal output is ignored, and possible-but-unsure items are counted instead of listed.
  • A new conversation is a new session. A new Claude Code, Codex, Cursor or Antigravity conversation now starts a new Malveon session, so plan documents and earlier work are "already there" and are no longer read as the agent's claims. On one project this took 10 false findings to none.
  • Windows: no more flashing terminal windows from the background watcher.
  • on and off now really switch programs. The program your hooks run is saved under a name made from its content and never overwritten, on replaces a watcher that is running an older build, and off waits for it to stop. Before, an update could leave the hooks on the old program.
  • The "how you will know it's done" request the prompt hook makes can now be answered in local-only mode (it is recorded locally and the reply carries no prompt text).

If you are upgrading, run malveon on once more in each project so its hooks pick up the new program, then start a new agent conversation.

Limits: tested mostly on Windows with Claude Code; how Codex and the desktop Codex app show the notice has not been seen. Two agents or windows prompting in the same project take turns resetting each other's session. The website and its docs pages still describe some earlier behaviour; docs/REFERENCE.md in the source is current.

v0.5.3

Choose a tag to compare

@LadsonDavid LadsonDavid released this 12 Nov 09:43

Much faster on big projects, with the same answers.

  • Faster first look: files are read side by side on all your cores and merged in a fixed order. On a 6,385-file project one run went from about 3 minutes to about 9 seconds (16 cores); across 39 test projects the total went from about 8 minutes to about 30 seconds. The reports are identical to the old ones. On a two-core machine expect roughly 3 to 4 times faster, not 15.
  • Read once, not twice: plain malveon used to read the whole project a second time for the "nothing reaches it" check. It now reuses the first read.
  • A saved copy between runs: if nothing malveon reads has changed (same files, sizes and times, same program build), it loads the saved map instead of reading the project. The same 6,385-file project: about 2.5 seconds. The copy lives in the cache folder of your data directory (not in your project) and holds file paths and names of functions and routes, no file text. Set MALVEON_NO_CACHE=1 to turn it off.
  • One visible difference: with --exclude, the "nothing reaches it" check now skips the excluded folders like the rest of the scan.
  • Fewer false alarms (from running the first-run scan on 100 more public projects): hand-run test-*.mjs scripts, feed/rss routes, types declared in a .d.ts, buttons inside popovers, files in public/, and more are no longer flagged.

Limits: a file changed in the last 3 seconds means that run is not saved, and any real change still rereads the whole project. The first run on files your system has not read recently can be much slower. Findings are still likely, not certain; a quiet result proves nothing.

v0.5.2

Choose a tag to compare

@LadsonDavid LadsonDavid released this 12 Nov 06:49

Fewer false alarms on real projects, found by running the first-run scan on 100 public repos and reading each finding.

  • Fewer wrong "nothing reaches it" notes: files loaded by a framework or tool (Next.js pageExtensions, typed import.meta.glob, Rojo *.server.* scripts, files in public/, hand-run test-*.mjs scripts, files named in a config or an .mdx page) and routes that people or services call from outside (feed, rss, webhooks, health) are no longer listed as unused.
  • Fewer wrong broken imports: ?raw-style imports, types declared in a sibling .d.ts, imports inside strings or comments, and createRequire with another base are no longer called broken.
  • Fewer wrong dead buttons: the chosen option of a toggle group, and a button inside a popover, tooltip or dropdown wrapper.
  • Core (v0.5.2, open source): request helpers that call other request helpers, <a href> links to routes the project serves, and addresses like "/api/log/" + id are now read.

Limits are the same as in v0.5.0: malveon reads code, a quiet result proves nothing, and the findings are likely, not certain. Some "may be unused" notes remain for cron and webhook-shaped routes it cannot see being called.

v0.5.1

Choose a tag to compare

@LadsonDavid LadsonDavid released this 06 Nov 13:55

A wording fix on top of v0.5.0.

  • The report and scan footers no longer say malveon "never runs your app". Since the optional --layout --start check can start your app (with a command you give it and approve once), they now say it "does not click through" the app.
  • The editor extension's results page uses the same wording.

No behavior changes. Everything in the v0.5.0 notes still applies, including the limits listed there.