You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Feature] Updater should align profile @deepseek-ai/* dependencies with the incoming runtime rc before applying an upgrade
#11554
Every in-app runtime upgrade (rc.2 → rc.3 → rc.5, observed twice + once independently) leaves the profile dependency tree on the previous rc. Because first-party plugins pin their @deepseek-ai/dsh-* peers to exact versions, the boot-time compatibility gate (dsh-app-boot → evaluatePluginCompatibility) then rejects every profile-resolved plugin pre-spawn and silently — all mcp__* tools (and plugin-manager rows) disappear after each upgrade until the user manually re-pins packages with pnpm.
Proposal: make the updater align profile @deepseek-ai/* dependencies with the incoming runtime rc as part of the upgrade flow.
even user-pinned direct deps (dsh-mcp-client, dsh-subagent-codex @ rc.3) were left behind
OP
Same signature every time: gate rejects pre-spawn, nothing in logs.db, failOnStartupError never fires, tools vanish with zero user-visible error.
Proposed behavior
The updater already knows the incoming version before it applies it: the installer is downloaded while the old runtime still runs (pending/update-info.json names deepseek-harness-<version>-win-x64.exe). That is the window to act.
Preferred — align before quit/install: after download, before "restart to apply", re-pin every @deepseek-ai/* dependency in each profile's package.json to the incoming rc (exact), run the profile pnpm install, so the first boot of the new runtime sees a matching tree. A working local reference implementation does exactly this (read pending version → pnpm add --save-exact per profile → restart) and has survived the rc.5 upgrade.
Alternative — align at first boot of the new runtime, before plugin gate evaluation: if a profile-resolved plugin manifest's peers mismatch the runtime, auto-heal by preferring the runtime-carried copy of that package (or re-pin + reinstall), then evaluate.
Respect explicit user intent: if a profile deliberately pins an older @deepseek-ai/* version, auto-alignment should be overridable (e.g. an aligned: false marker or a per-profile opt-out), with the incompatibility surfaced instead of silently healed.
Multiple profiles (desktop / web / headless): all need the same treatment.
Non-registry plugins (e.g. stdio commands pointing at local venvs/executables): unaffected — no @deepseek-ai/* manifest involved.
Orthogonal long-term fix: first-party plugins declaring engines.dsh ranges instead of exact peers would end this class entirely (the gate short-circuits on engines.dsh); the updater alignment is still needed for existing pinned plugins.
Success criteria
After any in-app upgrade, all previously-working profile plugins keep working with zero manual steps; if alignment is impossible, the user sees a named, actionable message instead of silent tool loss.
Happy to test release candidates against our profiles (desktop with dsh-mcp-client/telegram-mcp/subagent-codex) and report results.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Summary
Every in-app runtime upgrade (rc.2 → rc.3 → rc.5, observed twice + once independently) leaves the profile dependency tree on the previous rc. Because first-party plugins pin their
@deepseek-ai/dsh-*peers to exact versions, the boot-time compatibility gate (dsh-app-boot→evaluatePluginCompatibility) then rejects every profile-resolved plugin pre-spawn and silently — allmcp__*tools (and plugin-manager rows) disappear after each upgrade until the user manually re-pins packages with pnpm.Proposal: make the updater align profile
@deepseek-ai/*dependencies with the incoming runtime rc as part of the upgrade flow.Evidence (details in #11012)
dsh-mcp-clientstayed rc.2 (lockfile/range)@deepseek-ai/dsh-*deps stayed rc.2dsh-mcp-client,dsh-subagent-codex@ rc.3) were left behindSame signature every time: gate rejects pre-spawn, nothing in
logs.db,failOnStartupErrornever fires, tools vanish with zero user-visible error.Proposed behavior
The updater already knows the incoming version before it applies it: the installer is downloaded while the old runtime still runs (
pending/update-info.jsonnamesdeepseek-harness-<version>-win-x64.exe). That is the window to act.@deepseek-ai/*dependency in each profile'spackage.jsonto the incoming rc (exact), run the profilepnpm install, so the first boot of the new runtime sees a matching tree. A working local reference implementation does exactly this (read pending version →pnpm add --save-exactper profile → restart) and has survived the rc.5 upgrade.logs.dbempty;failOnStartupErrorunrelated; see also Upgrading DSH silently disables version-mismatched profile plugins; the reason lives only on stderr #11080).Edge cases to consider
@deepseek-ai/*version, auto-alignment should be overridable (e.g. analigned: falsemarker or a per-profile opt-out), with the incompatibility surfaced instead of silently healed.@deepseek-ai/*manifest involved.engines.dshranges instead of exact peers would end this class entirely (the gate short-circuits onengines.dsh); the updater alignment is still needed for existing pinned plugins.Success criteria
After any in-app upgrade, all previously-working profile plugins keep working with zero manual steps; if alignment is impossible, the user sees a named, actionable message instead of silent tool loss.
Happy to test release candidates against our profiles (desktop with dsh-mcp-client/telegram-mcp/subagent-codex) and report results.
All reactions