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
Official bridge packages pin host peers to exact versions, so hooks silently die after every DSH upgrade
#13798
Exact-pinned peerDependencies in official bridge packages cause silent hook disablement after every DSH upgrade
Summary
@deepseek-ai/dsh-hooks-claude-code and @deepseek-ai/dsh-hook-protocol declare exact-versionpeerDependencies on the host runtime packages (e.g. "@deepseek-ai/dsh-agent": "0.2.2-rc.5") and no engines.dsh range. As a result, every DSH runtime upgrade (including prerelease bumps like rc.5 → rc.6) makes these bridge packages fail the compatibility check, and dsh-app-bootsilently skips them — no error surfaced to an Electron GUI user, hooks just stop firing.
This is a real outage I hit, and the failure mode is invisible by construction.
Environment
DSH 0.2.2-rc.5 (Electron desktop app, macOS)
@deepseek-ai/dsh-hooks-claude-code@0.2.2-rc.5
@deepseek-ai/dsh-hook-protocol@0.2.2-rc.5
Profile: ~/.dsh/profiles/desktop
What happened
App auto-updated from 0.2.2-rc.3 to 0.2.2-rc.5 on 2026-11-18.
Bridge packages in the profile remained at 0.2.2-rc.3.
Their peerDependencies pin the host packages to 0.2.2-rc.3 exactly, so the check failed.
dsh-app-boot caught the error, pushed it into skippedBundles, and continued. Hooks silently stopped running for ~15.6 hours until I happened to check versions manually.
Root cause (from dsh-app-boot/lib/index.js)
unmetRequirements() checks engines.dshfirst, and only falls back to peerDependencies when engines.dsh is absent:
The bridge packages have engines: {}, so they take the peer path and every exact pin must match the current runtime.
The skip itself is silent by design:
/** Print each skipped bundle once; loading never prints, so launchers call this once per start. */functionreportSkippedBundles(binName,profile){ ... }
loadProfileDirectory() catches the compatibility error and records it in skippedBundles without printing. Only CLI launchers call reportSkippedBundles(). The Electron GUI never does — and dsh --profile desktop --dump-config-schema (which would expose it via x-cordis.diagnostics) is refused with "profile is managed exclusively by the Electron application". So a desktop user has no path to discover that their hooks are dead.
Why this is a design issue, not a user mistake
Compare the two declaration styles in the same profile:
Package
Peer style
Survives runtime upgrade?
dsh-xdpool (third-party)
">=0.1.1-rc.1 <0.3.0" + engines.dsh range
✅ yes
dsh-thoughtdag (third-party)
">=0.1.2-rc.1" ranges
✅ yes
dsh-hooks-claude-code (official)
exact "0.2.2-rc.5"
❌ no
Third-party plugins survive upgrades because they use ranges. The official bridge packages — which are exactly the ones users are told to keep in sync with the app version — use exact pins and therefore break on every bump. The README does mention upgrading to @<App version>, but that is a manual step with no detection, and the failure is silent.
Measured across the bundled packages: 878 exact-pinned peer entries vs 365 range-style entries, so this affects far more than the two hook packages.
Reproduction
# With DSH 0.2.2-rc.5 and a profile that has hooks-claude-code@0.2.2-rc.3
dsh --profile desktop --dump-config-schema | jq '.x-cordis.diagnostics'# (in the Electron app this command is refused, which is the discoverability problem)
Or check the runtime entry status — inactive means skipped:
include:hooks-claude-code → status: "inactive"
After aligning versions to 0.2.2-rc.5 it becomes status: "schema" and hooks fire again.
Suggestions
Give official bridge packages an engines.dsh range instead of (or in addition to) exact peerDependencies, matching the convention third-party plugins already use:
"engines": { "dsh": ">=0.2.2-rc.1 <0.3.0" }
Because engines.dsh short-circuits the peer scan, this single field makes the package survive prerelease bumps.
Surface skippedBundles in the Electron GUI (a warning badge, or a line in the plugin settings page). Right now the only consumer is a CLI launcher, so GUI users are blind to it.
If exact pinning is intentional (e.g. the bridge touches unstable internals), then a skipped bundle should at minimum produce a visible, non-blocking warning rather than a stderr write nobody reads.
Workaround I'm currently using
A patchedDependencies patch in pnpm-workspace.yaml that adds the engines.dsh range to both bridge packages. Verified with the app's own unmetRequirements logic:
App runtime
Without patch (exact peers)
With engines.dsh patch
0.2.2-rc.5
PASS
PASS
0.2.2-rc.6
FAIL
PASS
0.2.2-rc.7
FAIL
PASS
0.2.2
FAIL
PASS
0.3.0
FAIL
FAIL (expected — cross-minor)
So one field removes the whole class of prerelease-bump breakage.
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.
Exact-pinned
peerDependenciesin official bridge packages cause silent hook disablement after every DSH upgradeSummary
@deepseek-ai/dsh-hooks-claude-codeand@deepseek-ai/dsh-hook-protocoldeclare exact-versionpeerDependencieson the host runtime packages (e.g."@deepseek-ai/dsh-agent": "0.2.2-rc.5") and noengines.dshrange. As a result, every DSH runtime upgrade (including prerelease bumps likerc.5→rc.6) makes these bridge packages fail the compatibility check, anddsh-app-bootsilently skips them — no error surfaced to an Electron GUI user, hooks just stop firing.This is a real outage I hit, and the failure mode is invisible by construction.
Environment
0.2.2-rc.5(Electron desktop app, macOS)@deepseek-ai/dsh-hooks-claude-code@0.2.2-rc.5@deepseek-ai/dsh-hook-protocol@0.2.2-rc.5~/.dsh/profiles/desktopWhat happened
0.2.2-rc.3to0.2.2-rc.5on 2026-11-18.0.2.2-rc.3.peerDependenciespin the host packages to0.2.2-rc.3exactly, so the check failed.dsh-app-bootcaught the error, pushed it intoskippedBundles, and continued. Hooks silently stopped running for ~15.6 hours until I happened to check versions manually.Root cause (from
dsh-app-boot/lib/index.js)unmetRequirements()checksengines.dshfirst, and only falls back topeerDependencieswhenengines.dshis absent:The bridge packages have
engines: {}, so they take the peer path and every exact pin must match the current runtime.The skip itself is silent by design:
loadProfileDirectory()catches the compatibility error and records it inskippedBundleswithout printing. Only CLI launchers callreportSkippedBundles(). The Electron GUI never does — anddsh --profile desktop --dump-config-schema(which would expose it viax-cordis.diagnostics) is refused with "profile is managed exclusively by the Electron application". So a desktop user has no path to discover that their hooks are dead.Why this is a design issue, not a user mistake
Compare the two declaration styles in the same profile:
dsh-xdpool(third-party)">=0.1.1-rc.1 <0.3.0"+engines.dshrangedsh-thoughtdag(third-party)">=0.1.2-rc.1"rangesdsh-hooks-claude-code(official)"0.2.2-rc.5"Third-party plugins survive upgrades because they use ranges. The official bridge packages — which are exactly the ones users are told to keep in sync with the app version — use exact pins and therefore break on every bump. The
READMEdoes mention upgrading to@<App version>, but that is a manual step with no detection, and the failure is silent.Measured across the bundled packages: 878 exact-pinned peer entries vs 365 range-style entries, so this affects far more than the two hook packages.
Reproduction
Or check the runtime entry status —
inactivemeans skipped:After aligning versions to
0.2.2-rc.5it becomesstatus: "schema"and hooks fire again.Suggestions
Give official bridge packages an
engines.dshrange instead of (or in addition to) exactpeerDependencies, matching the convention third-party plugins already use:Because
engines.dshshort-circuits the peer scan, this single field makes the package survive prerelease bumps.Surface
skippedBundlesin the Electron GUI (a warning badge, or a line in the plugin settings page). Right now the only consumer is a CLI launcher, so GUI users are blind to it.If exact pinning is intentional (e.g. the bridge touches unstable internals), then a skipped bundle should at minimum produce a visible, non-blocking warning rather than a
stderrwrite nobody reads.Workaround I'm currently using
A
patchedDependenciespatch inpnpm-workspace.yamlthat adds theengines.dshrange to both bridge packages. Verified with the app's ownunmetRequirementslogic:engines.dshpatch0.2.2-rc.50.2.2-rc.60.2.2-rc.70.2.20.3.0So one field removes the whole class of prerelease-bump breakage.
All reactions