* feat(onboarding): recommend plugins and skills from installed apps
Scan installed macOS apps during classic onboarding (TCC-free), gather
candidates from official catalogs + ClawHub search, let the configured
model pick genuine matches, and offer an opt-in multiselect install step.
Adds a device.apps node-host command (default-off sharing, Android-parity
envelope) so remote gateways can request a paired Mac's inventory, and a
wizard.appRecommendations kill switch. Custom setup-inference completions
no longer inherit the 32-token verification-probe output cap.
* feat(onboarding): recommend apps in guided flow
* fix(onboarding): harden app recommendations against ClawHub self-promotion
Third-party ClawHub skills are never pre-selected regardless of model tier
(publisher-controlled listing text reaches the matcher prompt and could
promote itself); their labels now say they install third-party code.
Installed-app scans follow symlinked .app bundles. Matcher output stays
bounded by the resolved model's own maxTokens budget (documented invariant).
* fix(onboarding): key official catalog candidates by resolved plugin id
Real catalog entries are package manifests without a top-level id; keying the
candidate map and channel/provider classification by entry.id collapsed the
whole official catalog into one undefined-keyed entry, so no official plugin
or channel was ever recommended. Regression test runs against the bundled
catalogs.
* fix(onboarding): satisfy lint, types, deadcode, and migration gates
Split the guided-onboarding test into a self-contained custodian suite to stay
under max-lines. Narrow app-recommendation exports (drop dead node-payload
normalizer, unexport internal types/helpers, route candidate tests through the
public API), replace map-spread with a helper, unexport device.apps result
types, add installedAppsSharing to node-host migration expectations, cast the
wizard multiselect mock, and regenerate the docs map.
* test(onboarding): register new live test in the shard classifier
* feat(linux): canvas via CLI-node + Tauri app IPC bridge
* refactor: extract gateway helper modules
* build(linux-canvas): register plugin package in lockfile
* fix(linux-canvas): move canvas advertise test out of core, regen docs/protocol/deadcode
* fix(gateway): break node-catalog/registry import cycle via leaf normalize module; add canvas glossary term
* style: oxfmt invoke.ts and runtime.ts after buildNodeEventParams extraction
* fix(linux): load Canvas WebView via dedicated data_directory context
Wry's Linux/WebKitGTK incognito mode discards Tauri's registered
WebContext (wry webkitgtk/mod.rs), so the Canvas window got a fresh
ephemeral context without the openclaw-canvas:// scheme handler — the
bundled A2UI page never committed (stayed about:blank) and every A2UI
command timed out. Use an isolated cache-backed data_directory instead,
which keeps the protocol handler while still isolating Canvas storage
from the dashboard window.
* fix(linux): keep Canvas WebView ephemeral via incognito + data_directory
Autoreview flagged that a dedicated data_directory alone persists Canvas
browser state (cookies, localStorage, IndexedDB, service workers) across
restarts, so an agent that navigates Canvas to a site could leak an
authenticated session into a later session. iOS uses a non-persistent
store; Linux should match.
Add .incognito(true) alongside .data_directory(): the distinct directory
gives Tauri a fresh WebContext key so it still attaches the
openclaw-canvas:// protocol closure, and incognito makes Wry swap in a
fresh *ephemeral* context carrying those protocols. Live-verified on a
Wayland/WebKitGTK box: the bundled page still loads
(location.href=openclaw-canvas://localhost/index.html, openclawA2UI
present, A2UI renders) and the canvas-webview dir holds no persistent
cookie/storage files.
Catalog session rows (sidebar context menu + click), the built-in viewer
header, and a new "Open Codex/Claude sessions in" preference can launch the
native CLI (codex resume / claude --resume) in the operator terminal on the
machine that owns the session.
- Gateway-local sessions spawn through the existing terminal launch policy
(sandbox/enabled gates preserved) with the resume command in the session cwd.
- Paired-node sessions run through a new seq-ordered node PTY relay: a
duplex node-host command streams PTY output via node.invoke.progress and
receives keystrokes/resize via a new node.invoke.input event, behind the
unchanged terminal.* client protocol (TerminalSessionManager gains a backend
abstraction; node relay reuses the streaming-invoke controller).
- Owner boundary: each plugin owns its resume command and builds argv from a
validated thread id; the gateway routes node opens through the node command
allowlist and plugin invoke policy (no advertisement-only trust), and nodes
re-verify session eligibility before spawning.
- UI setting catalogOpenTarget + canOpenTerminal capability gate every entry
point; capability requires the owning host to actually have the CLI.
Node PATH is normalized before command-availability probes, Windows .cmd/.bat
shims spawn via ComSpec, and catalog terminal opens reattach persisted tabs
before opening the new tab.
* refactor(deadcode): trim auto-reply and CLI exports
* refactor(deadcode): trim cron and task exports
* refactor(deadcode): trim fleet and process exports
* test(deadcode): exercise live task and process seams
* test(fleet): cover stream redaction through owner module
* refactor(security): trim dead internal exports
* refactor(secrets): trim dead internal exports
* refactor(deadcode): trim remaining src exports
* refactor(deadcode): remove test-only runtime exports
* refactor(deadcode): trim pairing test exports
* refactor(deadcode): reconcile refreshed baseline
* test(auto-reply): deduplicate queue state imports
* feat: node-hosted plugins — dynamic tools, MCP servers, and skills
Nodes become declarative plugin hosts:
- node.pluginTools.update: node hosts publish plugin-registered agent tool
descriptors; gateway materializes them as agent tools executing via
node.invoke under the node command allowlist, with tools.effective
invalidation and node online/offline removal.
- Trusted paired-node descriptors: no gateway-side plugin registration
required; gateway.nodes.pluginTools.enabled off-switch (default on);
description/count caps; deterministic node-prefixed collision names.
- Declarative node-hosted MCP: nodeHost.mcp.servers (McpServerConfig shape)
starts MCP clients on the node host, publishes tools as pluginId node-mcp,
executes via built-in mcp.tools.call.v1 with per-layer timeouts, failure
isolation, and orphan-safe shutdown. No re-pairing when servers change.
- Node-hosted skills: node.skills.update publishes ~/.openclaw/skills
content (64 skills/64KB/512KB caps both sides); gateway merges them into
the skills snapshot while connected and exec host=node is available, with
node:// locators, node-prefixed collisions, disabled command dispatch,
and gateway.nodes.skills.enabled + nodeHost.skills.enabled switches.
- Security: node-supplied pluginIds cannot satisfy pluginId-scoped tool
allowlists unless gateway-registered; reserved node-mcp id requires the
core MCP descriptor shape; protocol registry kept out of public
plugin-sdk dts.
- E2E: pond harness proves publication, MCP round-trip, skills locator, and
disconnect/reconnect for all three surfaces.
* style: format node-plugin-tools test
* fix(skills): keep status loader unfiltered when eligibility is passed
skills.status started passing eligibility for the node-skill merge, which
flipped loadWorkspaceSkillEntries into filtered mode and dropped disabled
skills from status reports (QA plugin-lifecycle-hot-reload timeout). Status
now merges node skills explicitly around an unfiltered load. Also: regen
docs_map for new node docs sections; add the intentional node-host MCP
onclose suppression to the lint-suppression allowlist.
Restart stale local node-host processes when they reconnect to a newer gateway with a released-version mismatch, so launchd/systemd can restart them with updated code instead of leaving old dynamic imports alive.
Adds gateway mismatch detail propagation, node-host terminal pause handling, and regression coverage for the GatewayClient reconnect-pause path.
Verification:
- node scripts/run-vitest.mjs run src/gateway/client.test.ts -t 'CLIENT_VERSION_MISMATCH' --reporter=verbose
- node scripts/run-vitest.mjs run src/gateway/server.node-version-mismatch.test.ts src/node-host/runner.credentials.test.ts src/gateway/client.test.ts --reporter=verbose
- /Users/steipete/Projects/agent-skills/skills/autoreview/scripts/autoreview --mode local
- Crabbox AWS run_292dcbfd78d9: focused GatewayClient mismatch regression plus server/node-host mismatch tests passed
Co-authored-by: scotthuang <scotthuang@tencent.com>
* fix(browser): prevent permanent timeout after stuck evaluate
Thread AbortSignal from client-fetch through dispatcher to Playwright
operations. When a timeout fires, force-disconnect the Playwright CDP
connection to unblock the serialized command queue, allowing the next
call to reconnect transparently.
Key changes:
- client-fetch.ts: proper AbortController with signal propagation
- pw-session.ts: new forceDisconnectPlaywrightForTarget()
- pw-tools-core.interactions.ts: accept signal, align inner timeout
to outer-500ms, inject in-browser Promise.race for async evaluates
- routes/dispatcher.ts + types.ts: propagate signal through dispatch
- server.ts + bridge-server.ts: Express middleware creates AbortSignal
from request lifecycle
- client-actions-core.ts: add timeoutMs to evaluate type
Fixes#10994
* fix(browser): v2 - force-disconnect via Connection.close() instead of browser.close()
When page.evaluate() is stuck on a hung CDP transport, browser.close() also
hangs because it tries to send a close command through the same stuck pipe.
v2 fix: forceDisconnectPlaywrightForTarget now directly calls Playwright's
internal Connection.close() which locally rejects all pending callbacks and
emits 'disconnected' without touching the network. This instantly unblocks
all stuck Playwright operations.
closePlaywrightBrowserConnection (clean shutdown) now also has a 3s timeout
fallback that drops to forceDropConnection if browser.close() hangs.
Fixes permanent browser timeout after stuck evaluate.
* fix(browser): v3 - fire-and-forget browser.close() instead of Connection.close()
v2's forceDropConnection called browser._connection.close() which corrupts
the entire Playwright instance because Connection is shared across all
objects (BrowserType, Browser, Page, etc.). This prevented reconnection
with cascading 'connectOverCDP: Force-disconnected' errors.
v3 fix: forceDisconnectPlaywrightForTarget now:
1. Nulls cached connection immediately
2. Fire-and-forgets browser.close() (doesn't await — it may hang)
3. Next connectBrowser() creates a fresh connectOverCDP WebSocket
Each connectOverCDP creates an independent WebSocket to the CDP endpoint,
so the new connection is unaffected by the old one's pending close.
The old browser.close() eventually resolves when the in-browser evaluate
timeout fires, or the old connection gets GC'd.
* fix(browser): v4 - clear connecting state and remove stale disconnect listeners
The reconnect was failing because:
1. forceDisconnectPlaywrightForTarget nulled cached but not connecting,
so subsequent calls could await a stale promise
2. The old browser's 'disconnected' event handler raced with new
connections, nulling the fresh cached reference
Fix: null both cached and connecting, and removeAllListeners on the
old browser before fire-and-forget close.
* fix(browser): v5 - use raw CDP Runtime.terminateExecution to kill stuck evaluate
When forceDisconnectPlaywrightForTarget fires, open a raw WebSocket
to the stuck page's CDP endpoint and send Runtime.terminateExecution.
This kills running JS without navigating away or crashing the page.
Also clear connecting state and remove stale disconnect listeners.
* fix(browser): abort cancels stuck evaluate
* Browser: always cleanup evaluate abort listener
* Chore: remove Playwright debug scripts
* Docs: add CDP evaluate refactor plan
* Browser: refactor Playwright force-disconnect
* Browser: abort stops evaluate promptly
* Node host: extract withTimeout helper
* Browser: remove disconnected listener safely
* Changelog: note act:evaluate hang fix
---------
Co-authored-by: Bob <bob@dutifulbob.com>
Peter Steinberger