Repository navigation
Sessions not loading + "sessionController is unavailable" after installing official bundles by name (desktop 0.2.2-rc.1) #13627
Replies: 1 comment
|
Your chain is right, and the part that cost you the most time — "removing the entries from
Deleting the two entries fixes the first and second, leaves the third, and the third is what the loader actually reaches for. So the shadow survives the cleanup. That is also why the app "boots and looks healthy": the composition did run, it just ran a set of rows the guard had already disabled. On your suggestion 4 ("resolve core rows from the app first"), one warning from having implemented the same idea on our side: it has to be done as one change, with one source. If the version comes from the running kernel but the manifest still comes from the older profile copy, you convert "correct denial" into "false denial" — and the reverse change fixes nothing. Version, manifest and patch layer have to be read from the same place, or the guard's answer stops being about one build. Two smaller notes on the diagnosis you posted, because they will matter to the next person:
We keep the three-way split (installed / selected / resolved, with Boundary: this reads manifests and local files, not your machine; declaration-layer admission does not prove runtime health. |
Uh oh!
There was an error while loading. Please reload this page.
Environment
0.2.2-rc.1(commitc59d10a), Electron 44.7.0, Node 24.21.0desktop(Electron-managed), started from the installed appplugin_manager { action: "install_bundle", target: "@deepseek-ai/dsh-experimental-session-search" }and later the same for
@deepseek-ai/dsh-experimental-plugin-market-bundleSymptoms
After a restart following those two installs:
gateway/service-unavailable: typert gateway: session/create: active Service "sessionController" is unavailableplayprofile on the same machine, same binary, works fine.How to see the real error (the app hides it)
The Electron host's startup warnings are not surfaced in the UI or written to a log. They
can be read by booting a copy of the profile as a CLI web host:
stderr then shows:
A fresh profile with only official bundles prints no such warnings and works.
Root cause chain
install_bundlewith a bare package name resolves the newest published npm build.For these experimental packages that was the
0.2.1-alpha.*line, older than theruntime. The official catalog entry for the same bundle is
0.2.2-rc.1.~/.dsh/profiles/desktop/node_modules, wherethey shadow the core packages shipped inside the app (
dsh-session-query-sqlite,dsh-tool-session-query,dsh-plugin-market, ...).session-query-sqlitedisabled, nothing provides thesessionQueryservice.session-controllerand its dependents staypending, sosession/create,session/listand the whole session UI are unavailable — while the app otherwiseboots and looks healthy.
Removing the bundles from
package.jsonalone does not fix it: the shadowing packagefiles remain in the profile's
node_modules. (This is the part that cost the most time.)Why this looks like a product bug, not user error
runtime, although the catalog knows the right pinned version for that profile.
and a service error. The actionable messages (
did not activate,waiting for service)exist only on stderr of a host the user never sees.
app's own "Disable third-party plugins, back up profile patch, and restart" dialog is
not reachable because the app did boot.
Suggested fixes (any of these would have prevented or shortened this)
install_bundleshould prefer the catalog'sinstallTarget.specand pin the versioncompatible with the running DSH; when only a name is given and a catalog entry exists,
use the catalog rather than npm
latest.unavailable, and write the same list to a log the app can show.
core rows from the app first (the shadowing is the structural trigger).
Workaround used
dependenciesanddsh.profile.bundles.node_modules:dsh-experimental-session-search,dsh-experimental-plugin-market-bundle,dsh-session-query-sqlite,dsh-tool-session-query,dsh-plugin-market,dsh-plugin-market-index,dsh-client-ui-plugin-market.~/.dsh/sessions, independent of the profile).Adjacent, not identical: #4309 was also triggered by the profile's
node_moduleslayout.(Not related: #4549 and the dangling
tool/callfamily — that is a session-log integrityfailure, a different mechanism with a different symptom.)
Reported from real use on Windows 11; diagnosis and drafting assisted by an AI agent, verified and posted by the user.
All reactions