Repository navigation
Memory MCP/CLI abort on Node 24.19–24.21: agentdb's better-sqlite3 11.x (RemoveEnvironmentCleanupHook … (env) != nullptr) #4055
Description
Environment: ruflo 3.56.1 and 3.56.3 (global npm), agentdb 3.0.0-alpha.20 (nested better-sqlite3 11.10.0), Node v24.21.0, Linux x64 (WSL2).
Symptom: the first memory call in a process (MCP memory_stats/memory_search, CLI memory search, hooks session-end) intermittently aborts:
node::RemoveEnvironmentCleanupHook(v8::Isolate*, CleanupHook, void*) at ../src/api/hooks.cc:142
Assertion failed: (env) != nullptr
4: Database::~Database() [.../agentdb/node_modules/better-sqlite3/build/Release/better_sqlite3.node]
Through the ruflo-core plugin launcher the MCP server dies and Claude Code reports Connection closed, dropping all ruflo tools for the rest of the session. One real session: server idle 41 min, then its first memory_stats died, so this is not a startup race.
Cause: Node regression nodejs/node#65446: legacy node::ObjectWrap addons BUILT against ≥ 24.19.0 headers abort in GC finalizers. The fix landed as abe1b673e9b (PR #65943, 2026-09-21) but is not in any release through v24.21.0.
Evidence (A/B on the same store, Node 24.21.0): original binaries aborted 8/22 runs; the same binaries rebuilt against 24.18.1 headers aborted 0/17 (and 0/10 more through the plugin launcher after a 3.56.3 reinstall).
Ask:
ruflo doctor: flag native addons whosebuild/config.gypinodediris ≥ 24.19 while the runtime is a 24.x without the fix.- Document the workaround:
node-gyp rebuild --release --nodedir=<24.18.1 headers>in eachbetter-sqlite3dir (agentdb-nested, top-level, agentic-flow). - Longer term: agentdb on better-sqlite3 13.x (node-addon-api) removes the legacy ObjectWrap path.
Activity
Merge pull request #4069 from ruvnet/fix/doctor-node24-sqlite-4055
ruvnet commentedon Oct 11, 2026
Thanks @ronmikailov, this was a precise report. What is done and what is not:
Shipped in #4069 (next
@claude-flow/clirelease):ruflo doctor(anddoctor --component native-addon) now readsbuild/config.gypiof eachbetter-sqlite3copy in the resolved tree and warns when a locally compiled copy used Node 24.19+ headers while the runtime is Node 24.19-24.21, listing each path and version with the workarounds (Node 22 / 24.18.1 / a release with the fix, rebuild against 24.18.1 headers, or the 12.x prebuilt). It never loads an addon. The workaround is also indocs/USERGUIDE.md.Already shipped earlier (#3955, in 3.56.4+): ruflo opens AgentDB with its own prebuilt 12.x
better-sqlite3, which keeps the source-built 11.x addon out of that code path.Verified: a clean install of
[email protected]still contains nestedagentdb/node_modules/better-sqlite3andagentic-flow/node_modules/better-sqlite3at 11.10.0 (agentdb 3.0.0-alpha.20 declares it as an optional dependency^11.8.1), and theoverridesin theruflowrapper do not apply to an installed package.Not verified / not done: I did not reproduce the abort (this machine runs Node 22; I did not install Node 24), so I am relying on your A/B (8/22 aborts with the original binaries, 0/17 rebuilt against 24.18.1 headers). The nested 11.x copies cannot be removed from inside the published package; the real fix is an
agentdbrelease depending onbetter-sqlite3@^12, which is a separate repo and a maintainer decision. The upper bound of the check (24.21.x) should be revisited when a Node release ships fix (nodejs/node PR #65943).feat(doctor): flag better-sqlite3 copies built against Node 24.19+ he…
Merge pull request #4074 from ruvnet/fix/doctor-4055-pr-number
fix(doctor): cite the right Node PR (#65943) in the native-addon work…
ronmikailov commentedon Oct 15, 2026
Follow-up: Node v24.22.0 (2026-10-14, LTS) ships the runtime fix,
abe1b673e9"src: keep global list of addon-provided cleanup hooks" (doc/changelogs/CHANGELOG_V24.md, credited to nodejs/node#63985; #65943, whichdoctor-native-addon.jscites, is its v24.x backport).A/B on ruflo 3.56.10 (WSL2, x64), 30 serial
ruflo memory store --path <scratch>runs per arm.LD_DEBUGshows the command loads onlyruflo/node_modules/better-sqlite3(12.11.1):(env) != nullptr, exit 134)npm i -g [email protected]on 24.22.0 (12.11.1 prebuilt)The prebuilt never aborted, so rows 3-4 show the stock install is healthy, not the fix; rows 1-2 are the A/B. So
isAffectedNodeVersion()'s upper bound (v[1] <= 21) is right, and the fix is runtime-only: on 24.22.0 a binary built against 24.21.0 headers no longer aborts. Two suggestions for the FIX text:nodejs/node#65446 is still open upstream.