mirror of
https://github.com/we-promise/sure.git
synced 2026-09-05 23:01:26 +00:00
* fix(desktop): support a server hosted under a sub-path `normalize_server_url` rebuilt whatever you typed as `scheme://host[:port]` and dropped the path, so a self-hosted Sure behind a reverse proxy that mounts it under a prefix — `https://home.example.com/sure` — could never be added: the health check hit `{origin}/up`, where something else (or nothing) answers, and the app reported "Couldn't reach a Sure server at that address." The same bare origin was then used for the navigation, the SSO hand-off and the IPC capability grant, so even a passing check would have loaded the wrong site. The path was being dropped for a reason — people paste the URL they have in the clipboard, which is usually a deep link like `/sessions/new`. So rather than guess, let the server say where it is: keep the path in the canonical form, and have `check_server` probe `/up` from the address as typed back up to the origin, returning the first base that answers 200. A root-hosted server still resolves to its origin (one extra candidate, tried only after the specific one 404s), and a transport-level failure returns immediately instead of retrying a host that is not there. `check_server` now answers with that resolved base instead of a bool, and both callers save it rather than the raw input. Tests: the sub-path base, trailing slash / query / fragment trimming, the walk-up candidate list, ports preserved on every candidate, and the cap that still keeps the origin. * fix(desktop): carry the mounted base through SSO and deep links Follow-up to the sub-path support in the previous commit: saving a base with a path fixed adding and loading such a server, but left two flows resolving against the bare origin. The injected bridge only intercepted a form action shaped exactly `/auth/{provider}`, so a mounted server's `/sure/auth/{provider}` was never intercepted, and it emitted `location.origin`, which `begin_sso` would then reject as an unknown server. Match the prefix that precedes `/auth/` and emit the base it implies — empty at a domain root, so root-hosted deployments are unaffected. Deep links were gated on `target.server`, which `deep_link::parse` builds from host and port alone. Gate on the destination instead, and let `is_known_server` accept a URL that a saved base covers. `base_covers` matches only at a path boundary, so a saved `https://sure.example.com` does not cover `https://sure.example.com.evil.test`. Reported by automated review on the PR. * fix(desktop): cap candidates from the deep middle, not the shallow end The cap kept the deepest MAX_BASE_CANDIDATES - 1 candidates plus the origin, which discards the shallow ones in between — and a mount point is shallow. A server mounted at /sure, reached by a link pasted three segments into the app (`/sure/transactions/123/edit`), produced `.../edit`, `.../123`, `/sure/transactions`, origin: `/sure`, the only base that answers, was never probed, so the connection failed with "Couldn't reach a Sure server at that address" against a live server. Keep the address as typed plus the shallowest bases instead. Both readings survive the cap: the base typed exactly, and the mount a deep link sits under. Reported by @jjmata in review. * fix(desktop): bound discovery by mount depth, not candidate count Capping the candidate list drops whole mount depths silently: with a budget of four, `https://host/a/b/c/d/e` never probed `https://host/a/b/c`, so a server mounted there was undiscoverable. Bound the supported mount depth instead. The address as typed is always probed first, so a base entered exactly still works at any depth; the rest are every depth up to MAX_MOUNT_DEPTH, which keeps a pasted deep link finite (typed + 4 probes) without skipping a depth in between. Reported by coderabbitai in review.