Skip to content

build(node-tools)!: adopt OpenClaw 2026.11.5 on Node 24 - #275

Open
jsugg wants to merge 14 commits into
mainfrom
build/node24-openclaw-extended-stable
Open

jsugg wants to merge 14 commits into
mainfrom
build/node24-openclaw-extended-stable

Conversation

@jsugg

@jsugg jsugg commented Nov 18, 2026 •

Copy link
Copy Markdown
Owner

Summary

Moves the node-tools bundle from OpenClaw 2026.8.35 to 2026.11.5 (dist-tag latest) and drops Node 22, by owner decision.

This PR first targeted 2026.9.37 (extended-stable). That release bundles undici 8.10.2, inside high advisories fixed in 8.12.2 (GHSA-4w7w-5xrc-9h26), and no later 2026.9 release exists. 2026.11.5 has neither problem:

  • It no longer bundles undici; it resolves undici 8.12.3.
  • The first-party plugins resolve smol-toml 1.9.1, which clears GHSA-r4xh-jqrq-34v2.

Other details:

  • OpenClaw and the first-party @openclaw/* plugins move in lockstep. The lock was regenerated with tests/scripts/update_node_tools_lock.py (npm ci smoke) and mirrored into src/clawops/assets/.
  • The acpx CLI stays at 0.21.1 (@openclaw/acpx embeds 0.21.0), so the ACP worker adapter ranges are unchanged.
  • Release quarantine. 2026.11.5 was published 2026-11-17, inside the updater's 3-day window, and was taken by owner decision. Every other new or changed package was then re-resolved with the quarantine on. The only packages younger than 3 days are OpenClaw 2026.11.5's own (openclaw, @openclaw/*).

Node support

2026.11.5 engines.node is >=24.16.0 <25 || >=26.1.0. Its node-version.mjs also names 24.22.1 / 26.12.1 as strongly recommended security releases.

Surface Before After
platform_compat.NODE_RELEASE_FLOORS 22.22.3 / 24.15.0 / 25.9.0 24.16.0 / 26.1.0
Setup smoke + memory-plugin matrices 22.22.3, 24.15.0 24.16.0, 26.1.0
Fresh-host lanes + lock toolchain Node 22.23.5 / npm 10.9.9 Node 24.22.3 / npm 11.19.1
Linux bootstrap NodeSource channel setup_22.x (hard-coded) derived from the floor's major (setup_24.x)
  • The floor contract now checks each supported line's floor leg, instead of hard-coding Node 22.
  • Docs (README, SETUP_GUIDE, HOST_PLATFORMS, CI_AND_SECURITY), varlock env templates, and the packaged mirrors are updated.
  • Intel macOS CI lanes (fresh-host + Feature Verification) install the official nodejs.org 24.22.3 package (pinned SHA-256) into /usr/local, through the new hosted_docker.py install-node. Two things required this:
    • The macos-15-intel image ships Node 22.23.2, which the new floor refuses.
    • Homebrew publishes no Intel macOS bottles for any Node formula, so bootstrap's brew install node fallback fails there.

Dependabot

openclaw / @openclaw/* ignore >=2026.12.0, which holds the pinned 2026.11 line. The contract test derives the bound from the pinned CalVer month.

npm overrides

  • Dropped undici: 8.12.4. The tree resolves patched undici 8.12.3 directly.
  • Added npm: 12.2.2. OpenClaw pins npm@12.2.0, which bundles undici 6.28.0. 12.2.2 bundles undici 6.29.1.
  • Added @hono/node-server: 2.1.8. This is the version main already locks.
    • OpenClaw pins @modelcontextprotocol/node@2.1.1, which accepts only ^1.19.9. Every 1.x from 1.19.10 is inside GHSA-rmxm-3fg6-px4f.
    • Upstream @modelcontextprotocol/node@2.1.2 widens that to ^1.19.9 || ^2.1.3. Overriding that package instead would break OpenClaw's exact @modelcontextprotocol/server@2.3.1 pin, since 2.1.2 needs ^2.4.0.

OpenClaw 2026.9+ behaviour adaptations (re-verified on the 2026.11.5 dist)

  • ACP: the session key is still agent:<harness>:acp:<uuid>, and steer replies are still truncated at 800 chars.
  • Fake model: it reads the whole latest user turn, because runtime context is still sent as a separate trailing user message.
  • FeatureHost retries OpenClaw's explicit "retry" refusals during Gateway startup, within 180s. The pattern now covers 2026.11's wording:
    • offline maintenance, now with owner details
    • schema admission
    • "retry after session storage is ready"
    • "retry after the agent store is ready"
  • Timeouts:
    • The daily healthcheck steps get 300s; openclaw doctor takes 20–27s idle.
    • model-auth's read-only OpenClaw queries get 120s; 2026.9+ CLI queries run about 2× as long as on 2026.8.
    • The memory-pro search used by verify-pro-parity gets 120s.
  • Test races fixed:
    • The node-tools retry unit test patched the global time.sleep.
    • The FV _session_reply could read back the previous command's reply.

OpenClaw 2026.11 breaking changes absorbed

Each item below was found by CI or Feature Verification, then confirmed in the 2026.11.5 dist or its shipped docs.

  • Retired config keys. config validate rejects agents.defaults.heartbeat, channels.defaults.heartbeatVisibility and skills.workshop.approvalPolicy, so these are removed.
    • The baseline only used heartbeat to turn monitoring off.
    • 2026.11 provisions a Proactive check job only from an explicit openclaw setup, which StrongClaw never runs, so no periodic monitor exists.
    • All six render profiles pass config validate on 2026.11.5.
  • Sandbox image. The default moved to openclaw-sandbox:trixie-slim; sandboxed turns fail fast without it. clawops now builds upstream's recipe on a digest-pinned debian:trixie-slim.
  • ACP dispatcher stays off the model.
    • 2026.11 treats tools.deny: ["*"] as "tools disabled" and runs the turn tool-less. The old allow: [session_status] + deny: ["*"] let an ordinary message to the unsandboxed acp-dispatch agent reach the model.
    • The dispatcher now allows only a namespaced sentinel tool that nothing registers, with no * deny. OpenClaw then refuses the turn with "No callable tools remain", verified with openclaw agent --local against the fake provider.
    • The sandbox contract pins this.
  • memory status is offline-only while a Gateway owns the state.
    • The deep clawops doctor and baseline verify now run the offline check and, on that refusal, ask the owning Gateway (gateway call doctor.memory.status {"deep":true}, operator.read).
    • The profile feature tests run memory status / memory index inside a GatewayController.offline() window.
  • CLI exit warning. CLI commands print a Node process warning on exit ("Plugin source receipt was not recorded…"). The vendored memory-plugin host test now drops (node:<pid>) lines like plugin logs.
  • Diagnostics. The deep-doctor feature test now reports the end of each failed check, without plugin log lines.

Remaining advisories in the locked tree (npm audit, root causes)

Advisory Severity Where Fix
GHSA-fc5q-39xc-vjpf high picomatch 4.0.5 bundled inside npm 12.2.2 none released (<= 4.0.7)
GHSA-3g3m-2fjr-6pw4 moderate glob bundled inside npm 12.2.2 npm bundle; not overridable
GHSA-g3hv-75wg-m249 moderate minimatch 10.2.6 (OpenClaw's exact pin, already on main) + npm's bundled 10.2.5 11.x major

Notes on this table:

  • npm audit reports 32 entries; the high-severity ones other than picomatch are npm's bundled internals inheriting GHSA-fc5q.
  • main reports 8 moderate: minimatch, plus smol-toml, which this PR clears.
  • Everything new comes from the npm CLI that OpenClaw has depended on since 2026.9.

Other notes

  • lossless-claw (pinned v0.3.0) logs "runtime is intentionally unavailable during cli-metadata registration" on every 2026.11 CLI command. It is not fatal: CLI commands and Feature Verification pass. Upstream v1.2.0 declares CLI metadata but only claims tests against 2026.9.x, so it is left for a separate change.
  • 2026.11.5 still marks the first-party runtime plugin as unverified when it is loaded via plugins.load.paths; it logs "only bundled plugins can claim reserved command ownership". Installing it through openclaw plugins install (as is already done for acpx) is the likely follow-up.
  • Intel Mac operators whose Node is below 24.16.0 must install Node themselves, because Homebrew has no Intel bottles.

Verification

Local runs, raw:

  • ruff, black, isort, pyright and mypy are clean.
  • pytest -m "unit or contract": 1646 passed.
  • pre-commit (incl. actionlint) is clean.
  • All six render profiles pass openclaw config validate on 2026.11.5.

CI on e16ddac: all required checks pass (Verdict run 44288106963, Attribution Guard run 44288106812). That is 21 pass, 3 skipped by path filters. One macOS fresh-host job hit its 35-minute step cap while setup was still progressing, with no hang signature. The single allowed re-run (attempt 2) passed.

Feature Verification run 44288116824 on e16ddac: 93 PASS / 0 FAIL / 9 UNVERIFIABLE. This matches the main baseline (run 44107296257), with no per-feature verdict changes.

Moves OpenClaw and the first-party plugins in lockstep from 2026.8.35 to
2026.9.37, the extended-stable line. The ACP backend plugin now resolves
smol-toml 1.9.0, clearing GHSA-r4xh-jqrq-34v2. The acpx CLI stays 0.21.1,
so the ACP worker adapter ranges (already aligned with its built-ins) are
unchanged.

2026.9.37 requires `>=24.22.1 <25 || >=26.12.1`: Node 22 and 25 are no
longer supported.

- platform_compat mirrors the new per-line floors; bootstrap provisions
  the NodeSource channel of the floor's release line instead of 22.x
- CI smoke and memory-plugin matrices on 24.22.1/26.12.1; fresh-host and
  the lock toolchain on Node 24.22.3 / npm 11.19.1
- floor contract checks every supported line's floor, not just Node 22
- dependabot holds the 2026.9 line (ignore >=2026.10.0)
- overrides: undici 8.12.4 dropped (2026.9.37 bundles its own undici,
  which overrides cannot reach); npm pinned to 11.21.2 so the npm CLI
  OpenClaw now depends on carries undici 6.29.1 instead of 6.28.0

Known residual: OpenClaw 2026.9.37 bundles undici 8.10.2, inside
advisories fixed in 8.12.2 (e.g. GHSA-4w7w-5xrc-9h26). No later 2026.9
release exists to pick up the fix.

BREAKING CHANGE: Node 22.x and 25.x hosts are refused; Node 24.22.1+ or
26.12.1+ is required.
The macos-15-intel image ships Node.js 22.23.2 in /usr/local/bin, which
the fresh-host scenario resolves first. Under the new 24.22.1 floor that
Node is refused, bootstrap falls back to `brew install node`, and
Homebrew publishes no Intel macOS bottles for any Node formula, so the
fresh-host macOS jobs and the Feature Verification macOS lane fail before
setup.

Both lanes now install the official nodejs.org package (pinned version
and SHA-256) into /usr/local first, the same path an Intel operator takes.
The new `hosted_docker.py install-node` helper refuses versions outside
the supported range and checks the installed binary's version.
The healthcheck steps ran with the workflow runner's 30s shell default.
On OpenClaw 2026.9.37 `openclaw doctor` alone takes 20-27s on an idle
host with a bare config, so the step timed out on a provisioned host
(Feature Verification run 44223925453), and the deep security audit was
close to the limit too. Setup already gives doctor 300s.
…ehaviour

- fake model: read the latest user *turn* (consecutive user messages).
  2026.9 appends per-request runtime context as its own trailing user
  message, so the prompt's echo token was lost and agent/ACP turns
  replied "acknowledged"
- FeatureHost: retry openclaw and gateway admin calls while OpenClaw
  answers with its explicit "retry" refusals (state store in offline
  maintenance, session storage not ready) during gateway startup,
  bounded to 180s
- node-tools unit test: record only the module's own retry sleeps;
  patching the global time.sleep also captured subprocess's wait-poll
  backoff, which made the test (and `clawops baseline verify`) flaky
`model-auth ensure` gave `openclaw agents list` 30s, and the other
read-only queries 10-60s. In isolation, 2026.9.37 runs these about twice
as long as 2026.8.35 (`agents list --json`: 2.0-3.9s -> 4.8-6.4s on a
bare config). On the Intel macOS host, right after setup, the queries
also share the CPU with a Gateway that took 64s to start. `agents list`
hit its 30s limit there, ensure exited 1 before applying the model
chain, and the agent stayed on the profile default model (Feature
Verification run 44231180562).

The read-only queries now share one 120s budget, the same budget as the
inference probe.
`memory verify-pro-parity --mode openclaw` gave `openclaw memory-pro
search` 30s. On OpenClaw 2026.9.37 the CLI was still loading the
memory-lancedb-pro plugin when that limit hit (Feature Verification run
44235709488, linux-lancedb), so the live side came back empty and the
parity check failed. The search now gets the same 120s as the module's
memory-pro import.
`_session_reply` accepted the session's latest assistant message when
it was no more than 1s older than the send. In Feature Verification run
44240797608, `/acp spawn` was answered less than 1s before `/acp steer`
was sent, so the spawn confirmation was read back as the steer reply. A
reply must now also be newer than the session's latest message from
before the send.
2026.9.37 bundles undici 8.10.2, inside high advisories fixed in 8.12.2
(GHSA-4w7w-5xrc-9h26), and no later 2026.9 release exists. 2026.11.5
no longer bundles undici: it resolves undici 8.12.3, and the
first-party plugins resolve smol-toml 1.9.1.

- openclaw and the first-party plugins move to 2026.11.5 in lockstep;
  the acpx CLI stays 0.21.1
- Node floors follow its engines, `>=24.16.0 <25 || >=26.1.0`; CI floor
  legs are 24.16.0 / 26.1.0, and docs name the 24.22.1 / 26.12.1
  security releases OpenClaw recommends
- dependabot holds the 2026.11 line (ignore >=2026.12.0); its contract
  derives the bound from the pinned CalVer month
- overrides: npm 12.2.2 (OpenClaw pins npm 12.2.0, which bundles undici
  6.28.0; 12.2.2 bundles 6.29.1), and @hono/node-server 2.1.8, main's
  version. The exact @modelcontextprotocol/node 2.1.1 pin otherwise
  forces a 1.x copy inside GHSA-rmxm-3fg6-px4f; upstream 2.1.2 accepts
  ^2.1.3
- the FV retry seam matches 2026.11's wording: offline maintenance with
  owner details, schema admission, and "agent store is ready"

2026.11.5 was published 2026-11-17, inside the updater's 3-day release
quarantine. The pin was taken by owner decision. Every other new or
changed package was then re-resolved with the quarantine on, so the only
packages under 3 days old are OpenClaw 2026.11.5's own.

BREAKING CHANGE: Node 22.x and 25.x hosts are refused; Node 24.16.0+ or
26.1.0+ is required.
@jsugg jsugg changed the title build(node-tools)!: adopt OpenClaw 2026.9.37 extended-stable on Node 24 build(node-tools)!: adopt OpenClaw 2026.11.5 on Node 24 Nov 18, 2026
OpenClaw 2026.11 CLI commands print a Node process warning on stderr as
they exit ("Plugin source receipt was not recorded: Gateway is closing;
plugin shared-state admission is closed"). The vendored host test merges
both streams for plain output, so `memory-pro version` no longer matched
the package version exactly on either Node line. Its log filter now drops
`(node:<pid>)` warning lines along with plugin logs.
`openclaw config validate` on 2026.11.5 rejects three keys the baseline
and channels overlays render, which failed `clawops doctor-host` on
fresh hosts:

- agents.defaults.heartbeat: monitors are now ordinary automation jobs.
  The baseline only used the key to switch heartbeat off. 2026.11
  provisions a Proactive check job only from an explicit
  `openclaw setup`, which StrongClaw never runs, so no periodic check
  exists without it.
- channels.defaults.heartbeatVisibility: it moved into each migrated
  automation's delivery policy.
- skills.workshop.approvalPolicy: Skill Workshop proposals were removed.

All six render profiles now pass 2026.11.5 `config validate`. The
feature test for overlay merging now changes a nested `subagents` key
instead of the retired `heartbeat` one.
OpenClaw 2026.11 moved its default agent sandbox image from
`openclaw-sandbox:bookworm-slim` to `openclaw-sandbox:trixie-slim`, and
fails sandboxed turns when it is missing ("Sandbox image not found"). In
Feature Verification run 44275092875 that failed every agent turn.
StrongClaw now builds the new tag from upstream's documented recipe (the
same packages, user and command) on a digest-pinned `debian:trixie-slim`
index.
The unsandboxed `acp-dispatch` agent relied on OpenClaw stopping any
turn whose explicit tool allowlist leaves no callable tool. It used
`allow: [session_status]` plus `deny: ["*"]`. OpenClaw 2026.11 treats a
`*` deny as "tools disabled", enforces only runtime allowlists in that
case, and runs the turn tool-less. An ordinary message to the dispatcher
reached the model (Feature Verification run 44275092875).

Probed in isolation with `openclaw agent --local` on 2026.11.5:
- allow session_status + deny "*": the provider replied
- allow session_status only: the provider replied
- allow only a tool nothing registers, no deny: "No callable tools
  remain", and the provider was never contacted

The dispatcher now allows only `strongclaw_dispatcher_has_no_tools`,
with no `*` deny. The sandbox contract pins that sentinel and rejects a
`*` deny.
OpenClaw 2026.11 runs `openclaw memory status|index|...` only while no
Gateway owns the state directory ("This command is offline-only"). The
deep `clawops doctor` and `clawops baseline verify` ran `memory status
--deep` against a live Gateway.

`run_openclaw_memory_status` runs the documented offline check, and on
that refusal asks the Gateway for the same deep check through
`gateway call doctor.memory.status {"deep":true}` (operator.read). Any
other failure is reported as before. The runner is injected, so callers
keep their existing test seams.

The profile feature tests run their offline-only `memory status` and
`memory index` inside a new `GatewayController.offline()` window, which
stops and restarts the systemd, launchd or harness gateway as an
operator would.
The deep doctor test reported the first 600 chars of each failed check.
On OpenClaw 2026.11 that space fills with plugin registration log lines,
hiding the error itself. It now drops `[plugins]` log lines and keeps
the last 600 chars.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant