Skip to content

v0.4.9

Choose a tag to compare

@LadsonDavid LadsonDavid released this 31 Oct 07:30
· 18 commits to main since this release

Fewer false alarms: a run only reads FAILED for things that are proven.

  • Plan lines that do not say they are done no longer fail the run. A plain bullet, a numbered line or a table row with no status is counted as "not claimed done yet" and listed, not reported as NOT BUILT. A line ticked [x], marked with a check mark, or a table row whose status says Done is still held to the code. On a real 120-line roadmap plan this took the report from 70+ problems to the 3 that were true.
  • Three checks are advice now, not failures: a route registered twice, a sentence in your agent's notes that the code does not show, and a new call to an address no route serves. On 145 open-source projects each finding was checked against the code: they were real in 3 of 9, 0 of 4 and 1 of 3 projects. Broken imports (4 of 4 real), BROKEN and MISMATCH verdicts still fail the run. Across the same 145 projects, malveon review read FAILED for 15 before and 1 now (a real broken import).
  • Fewer wrong "not connected" verdicts. Requests written as BASE + '/path', made through an axios instance with a baseURL, through a request helper (api<T>(url), sendJSON("DELETE", url), or a helper that POSTs when given a body), or with a URL chosen by a ternary, now reach their routes. Express router.route('/p').get().post() chains and Hono app.route() mounts are read. A request field the server does not insist on is no longer a mismatch. A table declared in a migration or schema counts as built. Calls to another company's API are no longer "calls malveon can't read". Measured on the same projects: tables wrongly NOT BUILT 5.8% to 1.2%, components wrongly BROKEN 4.0% to 2.8%, routes reaching a caller 34 to 54 of 177.
  • A - [x] plan line's route in a nested Next.js app, /.well-known/... and GET /swagger.json in plan lines are now read.

Limits: only checks that fired in the 145 projects were measured; called stubs, claims contradicted by verified results and prompts never fired, so their precision is unknown. Unplanned code is advice and was not measured. Large monorepos (35,000 files) still take minutes.

Editor extension packages are not part of this release.