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
fix(plugin): validate DSH 0.2.2-rc.1 through rc.5 - #2632
The local plugin's npm peers and compatibility matrix omitted DSH 0.2.2 release candidates. This PR explicitly supports the five validated versions 0.2.2-rc.1, rc.2, rc.3, rc.4 and rc.5 across all six optional host-service peers.
The fixture overrides pin Cordis 4.1.0 and select each host's matching session codec. Regression tests verify admission for all five previews and correct fixture dependency selection, while preserving rejection of unvalidated rc.6, stable 0.2.2 and 0.3.x under npm's default semver evaluation. The npm lock importer, changelog and adapter documentation are synchronized.
A fresh source-package install exposed tsx dependency drift. The published tarball's ^4.21.0 range selected tsx 4.23.15 and esbuild 0.28.2 under pnpm, blocking the installer's reviewed build policy. Pin tsx to its existing validated lock version 4.21.0 and synchronize both lock importers. A failing regression covers this packaging mismatch, and the fixed tarball installs successfully through install.sh.
Related Issue (Required) — Refs #2520. Builds on and incorporates #2614. Keep #2520 open for stable local-plugin release tracking and Cloud-plugin validation. Please reconcile #2614 with this PR's overlapping compatibility changes during review.
The initial rc.5 fixture reproduced an ERESOLVE conflict between Cordis ~4.0.5 and the host's ~4.1.0 requirement. New regressions also reproduced missing peers and stale Cordis/codec selection for rc.2, rc.3 and rc.4. Those three versions passed their real API and Web-startup checks before their explicit npm peer declarations were added.
DSH's rc.5 loader evaluates semver with includePrerelease enabled. A real dump-config comparison confirmed that its older >=0.2.1-rc.2 <0.2.2 comparator already admits rc.5 at startup. The explicit preview declarations address npm's default peer-evaluation mismatch.
Bug fix for plugin package compatibility and fixture dependency selection
Documentation update with measured results and verification limits
How Has This Been Tested?
Test environment — macOS arm64, Node.js 25.1.0, npm 11.6.2.
Unit tests and the full plugin suite
Real published DSH API matrix and isolated official Web-profile startup checks
npm test
Test Files 192 passed (192)
Tests 1669 passed | 3 skipped (1672)
npm run test:dsh-compat -- 0.2.2-rc.2 0.2.2-rc.3 0.2.2-rc.4
PASS all 3 DSH versions
npm run test:dsh-compat
PASS all 10 DSH versions
Each version passed type checking, production compilation,
and 71 adapter tests.
pytest apps/memos-local-plugin/tests/python -q
143 passed in 5.64s
make format
All checks passed!
642 files left unchanged
npm run lint
passed
npm run build:package
passed
npm run lock:check
SQLite 13.0.3 lock metadata verified
The ten fixture versions are 0.1.0-rc.6, 0.1.7-rc.2, 0.2.0-rc.2, 0.2.1-rc.2, 0.2.1-rc.3, and all five previews from 0.2.2-rc.1 through rc.5. Each fixture installed its matching published host graph. Existing non-host build tools and unchanged native bindings were reused.
Initial matrix attempts encountered npm/Node registry connection timeouts. Successful runs used an explicit temporary Undici proxy dispatcher for the system HTTPS proxy, with localhost traffic kept direct. The plugin gains no proxy dependency.
Actual Web startup results
Each checked official CLI loaded an npm-packed local plugin runtime into a fresh Web profile with synthetic settings, telemetry disabled and failOnStartupError enabled.
DSH version
Typecheck/build/71 adapter tests
macOS Web startup
0.2.2-rc.1
Passed
Unverified
0.2.2-rc.2
Passed
Passed
0.2.2-rc.3
Passed
Passed
0.2.2-rc.4
Passed
Passed
0.2.2-rc.5
Passed
Passed in initial rc.5 validation
For rc.2, rc.3 and rc.4, profile composition included the MemOS row without a version exemption. Actual startup created SQLite storage and the Viewer health endpoint returned HTTP 200 with agent deepseek-harness. SIGINT stopped each process with exit code 130. The earlier rc.5 smoke produced the same results.
Fresh source installation and real local-model verification
Freshly installed the official DSH 0.2.2-rc.5 CLI with normal npm lifecycle scripts, built this branch with npm pack/prepack, and installed its tarball through the official install.sh DSH target using an isolated pnpm 11.7.0 and a fresh DSH_HOME.
The installer passed SQLite write/read and ONNX CPU backend checks. Installed dependency versions were tsx 4.21.0, esbuild 0.27.7, better-sqlite3 13.0.3 and onnxruntime-node 1.24.3. The source tarball retains the branch package.json version 2.0.16-beta.1 and is an unpublished local build.
Used existing local Ollama models qwen2.5:0.5b and nomic-embed-text for real inference. MemOS used DSH's host-model bridge for auxiliary calls and real 768-dimensional embeddings. Separate headless invocations demonstrated:
Store the synthetic fact that Project Orchid-Z91 uses port 47381. The model replied SAVED.
Restart DSH in a new session and ask for the port without including its value. The model answered Project Orchid-Z91 port 47381.
Ask the model to call memos_search. The tool completed and returned the two earlier memories. The final answer included port 47381.
Compressed V4 session logs contain the producer-owned MemOS recall message with memos_context and the correct port in sessions two and three. SQLite contains three embedded traces and its integrity check returns ok. All three successful headless turns exited 0 and lightweight capture reported warnings=0.
The small model's initial full-tool context exceeded Ollama's 4096-token window and produced off-topic answers. The successful test profile kept MemOS tools and disabled unrelated tool providers. Ollama also logged HTTP 500 responses during auxiliary activity, while the successful turns and capture completed. The temporary Ollama server was stopped afterward. Headless --json stdout also includes MemOS core log lines. This test parsed JSON events separately and preserved all output, and strict NDJSON-only stdout remains an integration limitation.
Remaining verification and release scope
Packaged Desktop applications on macOS, Windows and Linux remain unverified.
Cloud-provider authenticated conversations, live model downloads, and production-profile upgrades remain unverified.
Fresh DSH rc.5/source-plugin installation and native bindings were verified on macOS with Node.js 25.1.0 and pnpm 11.7.0. Other platform/runtime installation combinations remain unverified.
Actual Web startup on rc.1 and full-tool agent answer quality remain unverified.
The three skipped tests remain skipped in the full suite.
Other DSH previews, the Cloud plugin and npm publication are outside this validation.
Remote CI for the updated commit must be reviewed before merge.
Detailed evidence is in apps/memos-local-plugin/adapters/deepseek-harness/COMPATIBILITY.md.
Checklist
I have performed a self-review of my own code
I have commented dependency-selection logic that needs explanation
I have added regression tests proving the compatibility fix
Documentation applicability checked — adapter documentation and validation results are updated in this repository
The cordis assertions are gated on override being truthy rather than on override.cordis. hostOverrides is a per-host table, so if a later entry is added without a cordis pin (e.g. only to change the codec), cordisVersion falls back to the manifest range ~4.0.5; satisfies("~4.0.5", "^4.0.1") returns false, so the script aborts with the misleading Cordis ~4.0.5 does not satisfy plugin peer ^4.0.1. The same applies to the host-peer check (if (override && hostCordis && ...)). Guarding on override?.cordis keeps the checks tied to the pin they validate.
💡 Suggested Change
Before:
if (override && !satisfies(cordisVersion, pluginCordis)) {
After:
if (override?.cordis && !satisfies(cordisVersion, pluginCordis)) {
The keys must be kept manually in sync with the default versions list below. If a host version is added to one list and not the other (or is passed on the CLI), the fixture silently falls back to the checkout's cordis: "~4.0.5" and the shared 0.2.1-rc.3 codec — exactly the regression this change fixes (COMPATIBILITY.md documents that failure mode).
The mere presence of an entry silently flips two behaviors: the Cordis pin here and the codec pin via || override on line 47. Someone adding an entry only to try a different Cordis build would unintentionally repin dsh-session-format-v3-to-v4 to the host version (and vice versa), with nothing in the object shape expressing that intent.
Deriving both from one list and making the codec pin explicit would prevent this. (Note: "4.1.0" is repeated five times, so a single constant also avoids divergence such as one row keeping 4.0.5.)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
The local plugin's npm peers and compatibility matrix omitted DSH 0.2.2 release candidates. This PR explicitly supports the five validated versions 0.2.2-rc.1, rc.2, rc.3, rc.4 and rc.5 across all six optional host-service peers.
The fixture overrides pin Cordis 4.1.0 and select each host's matching session codec. Regression tests verify admission for all five previews and correct fixture dependency selection, while preserving rejection of unvalidated rc.6, stable 0.2.2 and 0.3.x under npm's default semver evaluation. The npm lock importer, changelog and adapter documentation are synchronized.
A fresh source-package install exposed tsx dependency drift. The published tarball's ^4.21.0 range selected tsx 4.23.15 and esbuild 0.28.2 under pnpm, blocking the installer's reviewed build policy. Pin tsx to its existing validated lock version 4.21.0 and synchronize both lock importers. A failing regression covers this packaging mismatch, and the fixed tarball installs successfully through install.sh.
Related Issue (Required) — Refs #2520. Builds on and incorporates #2614. Keep #2520 open for stable local-plugin release tracking and Cloud-plugin validation. Please reconcile #2614 with this PR's overlapping compatibility changes during review.
The initial rc.5 fixture reproduced an ERESOLVE conflict between Cordis ~4.0.5 and the host's ~4.1.0 requirement. New regressions also reproduced missing peers and stale Cordis/codec selection for rc.2, rc.3 and rc.4. Those three versions passed their real API and Web-startup checks before their explicit npm peer declarations were added.
DSH's rc.5 loader evaluates semver with includePrerelease enabled. A real dump-config comparison confirmed that its older >=0.2.1-rc.2 <0.2.2 comparator already admits rc.5 at startup. The explicit preview declarations address npm's default peer-evaluation mismatch.
Suggested reviewers @hijzy and @whipser030, matching #2614.
Type of change
How Has This Been Tested?
Test environment — macOS arm64, Node.js 25.1.0, npm 11.6.2.
The ten fixture versions are 0.1.0-rc.6, 0.1.7-rc.2, 0.2.0-rc.2, 0.2.1-rc.2, 0.2.1-rc.3, and all five previews from 0.2.2-rc.1 through rc.5. Each fixture installed its matching published host graph. Existing non-host build tools and unchanged native bindings were reused.
Initial matrix attempts encountered npm/Node registry connection timeouts. Successful runs used an explicit temporary Undici proxy dispatcher for the system HTTPS proxy, with localhost traffic kept direct. The plugin gains no proxy dependency.
Actual Web startup results
Each checked official CLI loaded an npm-packed local plugin runtime into a fresh Web profile with synthetic settings, telemetry disabled and failOnStartupError enabled.
For rc.2, rc.3 and rc.4, profile composition included the MemOS row without a version exemption. Actual startup created SQLite storage and the Viewer health endpoint returned HTTP 200 with agent deepseek-harness. SIGINT stopped each process with exit code 130. The earlier rc.5 smoke produced the same results.
Fresh source installation and real local-model verification
Freshly installed the official DSH 0.2.2-rc.5 CLI with normal npm lifecycle scripts, built this branch with npm pack/prepack, and installed its tarball through the official install.sh DSH target using an isolated pnpm 11.7.0 and a fresh DSH_HOME.
The installer passed SQLite write/read and ONNX CPU backend checks. Installed dependency versions were tsx 4.21.0, esbuild 0.27.7, better-sqlite3 13.0.3 and onnxruntime-node 1.24.3. The source tarball retains the branch package.json version 2.0.16-beta.1 and is an unpublished local build.
Used existing local Ollama models qwen2.5:0.5b and nomic-embed-text for real inference. MemOS used DSH's host-model bridge for auxiliary calls and real 768-dimensional embeddings. Separate headless invocations demonstrated:
Compressed V4 session logs contain the producer-owned MemOS recall message with memos_context and the correct port in sessions two and three. SQLite contains three embedded traces and its integrity check returns ok. All three successful headless turns exited 0 and lightweight capture reported warnings=0.
The small model's initial full-tool context exceeded Ollama's 4096-token window and produced off-topic answers. The successful test profile kept MemOS tools and disabled unrelated tool providers. Ollama also logged HTTP 500 responses during auxiliary activity, while the successful turns and capture completed. The temporary Ollama server was stopped afterward. Headless --json stdout also includes MemOS core log lines. This test parsed JSON events separately and preserved all output, and strict NDJSON-only stdout remains an integration limitation.
Remaining verification and release scope
Detailed evidence is in apps/memos-local-plugin/adapters/deepseek-harness/COMPATIBILITY.md.
Checklist
Reviewer Checklist