v0.4.62026-11-16
@jshookmcp/jshook v0.4.6
Patch release to ship the v0.4.5 P0 fix. 0.4.5 was Blocked by npm's
automated dual-use review and its semver cannot be reused, so the same fix is
re-released under a new version number. This is the same retirement pattern
npm applied to 0.4.0 (see the publish workflow's DUAL-USE notes).
What's in it
Identical product code to the (blocked) 0.4.5 — the P0 registry fix:
- search-tier dynamic activation was broken. A domain loaded on demand
(e.g. any workflow/full domain activated from the default search
profile) updated the registry's local registration map but NOT the shared
read cache. activatetools / activatedomain reported every tool in that
domain as notFound and call_tool answered "exists but is not active" —
tools were discoverable but never invocable. Fixed by keeping the two
registration stores in sync in ensureDomainLoaded.
- Verified end to end: searchtools → activatetools →
calltool(browserattach, {autoConnect:true}) attaches to a running
Chrome; typecheck, registry tests, and the full pnpm check gate pass.
Test-pipeline hardening (also since 0.4.4)
The P0 passed 20,562 tests due to a test-strategy gap. Added:
- no-mock in-process pipeline test (search → activateToolNames),
- real stdio MCP e2e under MCPTOOLPROFILE=search,
- test:e2e:meta CI gate after build,
- registry mock audit (tests/server/registry/MOCK-AUDIT.md).
All three regression tests were verified to fail with the fix disabled.
Install / upgrade:
``jsonc
{
"mcpServers": {
"jshook": {
"command": "npx",
"args": ["-y", "@jshookmcp/jshook@latest"]
}
}
}
``
Docs: https://vmoranv.github.io/jshookmcp/ · npm: https://www.npmjs.com/package/@jshookmcp/jshook
Full changelog: https://github.com/vmoranv/jshookmcp/compare/v0.4.4...v0.4.6
v0.4.52026-11-16
@jshookmcp/jshook v0.4.5
Emergency patch release for a P0 regression in the search tool tier.
P0 fix
- search-tier dynamic activation was broken. A domain loaded on demand
(e.g. any workflow/full domain activated from the default search
profile) updated the registry's local registration map but NOT the shared
read cache. As a result activatetools / activatedomain reported every
tool in that domain as notFound (totalDomainTools: 0) and call_tool
answered "exists but is not active" — tools could be **discovered but never
invoked**. Since search starts with zero loaded domains, this affected all
cross-tier activation, which is the core of the search-first design.
- Fixed in src/server/registry/state.ts + index.ts: lazily loaded
registrations are now appended (idempotently, name-deduplicated) to the
shared cache in ensureDomainLoaded, keeping both stores in sync.
- Verified end to end: searchtools → activatetools →
calltool(browserattach, {autoConnect:true}) now attaches to a running
Chrome; typecheck, 46 registry tests, and the full pnpm check gate
(20,562 tests) all pass.
Also in this release
- docs(publish): clarify npm's validating vs staged lifecycle states
(a "Validating" versions-tab row is automated review, not a pending Approve).
Install / upgrade:
``jsonc
{
"mcpServers": {
"jshook": {
"command": "npx",
"args": ["-y", "@jshookmcp/jshook@latest"]
}
}
}
``
Docs: https://vmoranv.github.io/jshookmcp/ · npm: https://www.npmjs.com/package/@jshookmcp/jshook
Full changelog: https://github.com/vmoranv/jshookmcp/compare/v0.4.4...v0.4.5
v0.4.42026-11-16
@jshookmcp/jshook v0.4.4
405 commits since v0.4.3. The release is dominated by search/retrieval quality,
a rebuilt evaluation pipeline, runtime hardening, and a large test/coverage
campaign — the public surface stays 736 tools across 36 domains.
Highlights
- Retrieval quality — holdout MRR 32.6% → 79.9% (gap 38pp → 5.4pp) (#279);
synonym groups, verb-inflection stemming, scene-keyword phrasing, and
description enrichment so real query wording finds the right tool.
- Search resilience — queries never block on a cold embedding worker / vector
bootstrap; stale vector cache retired; CDP timeouts come from config.
- New evaluation pipeline — E→A→C→B→D evaluation, trace capture,
difficulty-layered gates, floor ratchets, and cross-view reports (#287).
- Observability — tool-failure classification into dim_label error classes,
surviving the soft-fail path and mirrored onto duration metrics.
- Runtime / browser — missing-executable errors made actionable; launch-path
gaps closed; honest process kill/resume reporting; correct Win32 quoting.
- Modules — binary-secrets custom patterns, APK signing-block fixes,
jadx decompile cache, Dart snapshot fingerprinting and package detection.
- Coverage — 736/736 tools measured; full regression suite green at
20,562 passing tests (statements 83%, branches 73.5%, functions 86.9%,
lines 84.5%).
- Deps / toolchain — vitest 5, MCP 2.4, advisory bumps; tsdown pinned to
0.23.0 (TS 7 dts compatibility); pnpm lockfile repairs.
Housekeeping
- Removed machine-local autoresearch/eval-loop scratch files from the repo and
added gitignore rules so they cannot be committed again.
- Hardened the release pipeline: the publish Verify step now satisfies the
CI-strict scorecard gate on a clean checkout.
Install / upgrade
``jsonc
{
"mcpServers": {
"jshook": {
"command": "npx",
"args": ["-y", "@jshookmcp/jshook@latest"]
}
}
}
``
Docs: https://vmoranv.github.io/jshookmcp/ · npm: https://www.npmjs.com/package/@jshookmcp/jshook
Full changelog: https://github.com/vmoranv/jshookmcp/compare/v0.4.3...v0.4.4
v0.4.3 — Runtime hardening2026-10-30
Highlights
- Includes runtime hardening for breakpoint condition evaluation.
- Fixes QuickJS stack configuration for compatibility-layer parsing.
The npm package was built by the GitHub Actions workflow and published with SLSA provenance: https://github.com/vmoranv/jshookmcp/actions/runs/40895558637
v0.4.2 — 735 tools in one MCP server, ~3K tokens to start2026-10-10
735 tools across 36 domains, one MCP server — and it starts at ~3K tokens.
This is the version that made it back onto the public registry after the dual-use
policy change. Two earlier attempts (0.4.0, 0.4.1) were blocked by npm's
automated review and are dead version numbers; 0.4.2 is the one that shipped.
Dual-use compliance
jshook is dual-use by design — it hooks processes, intercepts TLS, and drives
browsers with anti-detection. That capability set is now declared rather than
merely present:
- package.json carries contentPolicy.class = "dual-use".
- DISCLOSURE ships inside the tarball and covers the process/memory injection
surface, which was previously not declared at all.
- The postinstall lifecycle hook is gone. So is inject-dll.ps1 — a stub that
declared OpenProcess / VirtualAllocEx / WriteProcessMemory /
CreateRemoteThread and called none of them.
- CI now fails the build if an install-time script or an injection helper
reappears in the packed tarball.
Since v0.3.5
425 commits. The themes, rather than the changelog:
Search quality. search_tools was rebuilt around a hybrid BM25 + vector
retriever with an incremental embedding cache, a tuner with holdout verification,
and a repaired feedback loop (the tracker had been split across two instances, so
signals were being recorded and never read).
Observability. OTLP export for spans and metrics, an anonymous install.id,
and search telemetry — on by default, with a documented opt-out. Tool-argument
shapes (key names only, never values) are captured on tool.execute spans.
Binary instrumentation. Frida device support, with session-level device
inheritance and remote/device-probe timeout knobs.
Native emulator correctness. A run of ARM semantics fixes: BSL/BIT/BIF
operand roles, ADRP page-base truncation, SQXTN2 upper-half placement,
SM4EKEY lane windows, signed shift amounts.
Process and memory. One MemoryManager shared across startMonitor /
stopMonitor instead of two, plus escaping fixes in the Windows script builders.
Platform. Electron fuse detection migrated to the v2 wire format;
electron-fuse-handler follows.
Context cost
Measured on the built server, using the same 4-bytes-per-token estimate the
runtime's TokenBudgetManager uses:
| profile | tools | tools/list payload | ~tokens |
| --- | --- | --- | --- |
| search | 8 | 7,527 B | ~1,882 |
| workflow | 360 | 209,005 B | ~52,251 |
| full | 743 | 435,820 B | ~108,955 |
The full figure scales with the tool count — re-measure it when the catalog
grows. The search figure does not.
Install
``bash
npx -y @jshookmcp/jshook@latest
``
Node 22.22.2+ or 24.15+. AGPL-3.0.
Docs: https://vmoranv.github.io/jshookmcp/