Changelog - 2026-09-07

Atlas starts in any repository again

Bug Fix

In one line. A consumer whose repository carries docs/wiki/content and .agents/knowledge no longer looks like the IGNIS checkout, so the server serves its packaged snapshot instead of exiting.

$ cd ~/work/my-monorepo && bunx @venizia/ignis-atlas mcp
ignis-atlas: ENOENT: no such file or directory, open '.../docs/wiki/content/changelogs'   # before

The problem it solves

Repository mode was chosen when two directories existed: the wiki content and the knowledge bundle. That is a layout IGNIS invented and its consumers copy. One consumer had both, but no docs/wiki/content/changelogs, so the corpus loader threw ENOENT before the first JSON-RPC line and the MCP client reported CONNECTION_CLOSED. The packaged snapshot beside the binary was never reached.

What changed

Symbol	Change	Package
isRepositoryCheckout()	Requires all three corpus directories AND a root package.json named @venizia/ignis-workspace	atlas
WORKSPACE_PACKAGE_NAME	New exported constant	atlas
resolveMode()	The explicit --root error names every marker a checkout needs	atlas

• Inside the IGNIS checkout nothing changes: the manifest matches and every directory is there.
• Anywhere else the server falls through to the corpus packaged with the release, which is what a consumer wants.
• An explicit --root that is not a checkout still exits 2 rather than serving the wrong corpus.

Who is affected

• Anyone running bunx @venizia/ignis-atlas mcp outside the IGNIS checkout. Upgrade and the server starts. No configuration changes.
• The IGNIS checkout. No action needed.