* feat(gateway): add remote-exec cloud placements
* feat(codex): run cloud turns through remote exec
* fix(sandbox): quote ssh_config path directives containing whitespace
Crabbox lease keys default to ~/Library/Application Support/... on macOS;
unquoted IdentityFile/UserKnownHostsFile/CertificateFile arguments tokenize
on the space and fail as 'extra arguments'. Found via live remote-exec
cloud-worker proof.
* test: consolidate gateway maintenance schedule coverage
* fix(ci): invalidate plugin sdk declarations on state changes
* fix(ui): fork active sessions from stable history
* fix(protocol): align active fork types
* test(ui): match main-session fork routing
* fix(sessions): unify stable fork admission
* fix(native): fork active sessions from stable history
* fix(android): fork active sessions from stable history
* style(android): simplify active fork result handling
* fix(native): preserve legacy fork transport API
* test(native): complete legacy fork transport stub
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(ui): complete pairing setup lifecycle
Redeemed and expired mobile setup codes stayed on screen as usable QR
codes, so a successful pairing had no visible outcome and expired bearer
material still looked live.
The Gateway now mints an opaque, non-authorizing setupId beside each
setup credential, returns its authoritative expiresAtMs, records the
terminal outcome of the exact redemption, and broadcasts
device.pair.setup.completed. Because that broadcast is dropped for
buffered operator sockets, the completion is persisted first and can be
reconciled through the new device.pair.setupStatus method: the Control UI
asks for the recorded outcome before it may present a credential as
expired, so a pairing that succeeds is never shown as a failure.
The Control UI models one closed lifecycle (selection, loading, waiting,
success, expired, error) correlated only by the active setupId, and
Pairing help now carries the external-link affordance.
* fix(ui): preserve unknown pairing outcomes
* test(ui): target pairing terminal headings
* test(ui): align pairing completion fixture
* fix(gateway-protocol): decode setup-code results from older gateways
Older protocol-v4 gateways omit the new setupId and expiresAtMs
lifecycle fields, so requiring them in the generated native model broke
decoding an existing device.pair.setupCode response. Keep both optional
at the wire boundary, require lifecycle metadata before the Control UI
enters its waiting state so a missing outcome stays visible, and cover
the legacy payload with Swift and schema regressions.
* fix(ui): surface rejected pairing dialog loads
The lazy pairing chunk could reject while its overlay was already open,
leaving the shell rendering nothing at all. Record the rejection on the
shell and render a recoverable modal with a reason and a retry so the
open action always ends in a visible outcome.
* fix(state): preserve pairing setup schema compatibility
* test(gateway): cover pairing setup release train
* fix(gateway): commit pairing setup completion atomically
* refactor(state): distinguish setup transaction helpers
* refactor(state): remove obsolete bootstrap restore path
* fix(gateway): preserve setup handoff type safety
* fix(gateway): keep pairing completion terminal after consume
* fix(gateway): validate Watch binding during setup commit
* fix(gateway): revalidate setup credential expiry at commit
* chore: refresh Plugin SDK API contracts after rebase
* fix(pairing): prune expired setup completions
* chore: retrigger CI
* fix(protocol): deduplicate setup expiry field
* fix(protocol): refresh pairing setup clients
* fix(gateway): make pairing setup completion durable
* fix(ui): retire expired pairing credentials immediately
* fix(ui): keep pairing dialog visible while loading
* fix(macos): align setup result initializer order
* fix(gateway): restore generic bootstrap retries
* chore(ui): record pairing startup budget
* chore(ui): refresh pairing startup budget
* style(gateway): format maintenance imports
* test(gateway): cover session-sharing mock
* fix(state): defer setup correlation schema
---------
Co-authored-by: Patrick Erichsen <patrick.a.erichsen@gmail.com>
* refactor(delivery): collapse failed-row lifecycle
Replace the unshipped failure-operations platform with payload-free terminal receipts owned by existing queue boundaries. Keep bounded/permanent idempotency only for reusable or crash-ambiguous producers, move physical expiry to queue maintenance, and preserve migration and media-cleanup safety.\n\nTogether with #123410, production code is net negative by 11 lines; tests, docs, and generated protocol mirrors are accounted separately.
* fix(delivery): break state DB import cycle
* fix(delivery): classify SQLite boundary uses
* test(gateway): mark retained health fixture
* feat(computer-use): computer.act v2 action contract with capability-filtered tool
* fix(computer-use): break contract import cycle, bound model-visible elements, regen swift protocol
* test(computer-use): satisfy curly rule in schema-cap helpers
* fix(computer-use): satisfy type-aware lint on contract and tool
* test(gateway-protocol): keep connect-params suite under the line cap
* feat(secrets): add authenticated egress substitution proxy
* feat(secrets): bind egress substitution to hosts
* ci(codeql): classify egress proxy bypass tunnel in network boundary query
* refactor(proxy-capture): use the canonical IP parser instead of node:net
* fix(secrets): compare proxy tokens with a process-keyed MAC
* fix(ui): report the outcome of a cross-device token rotation
Rotating another device's token ended with nothing on screen. The Gateway
returns the replacement only to a device rotating its own token
(shouldReturnRotatedDeviceToken; docs/cli/devices.md), so every rotation of
another device answered without a token and the Control UI, which only rendered
when one came back, showed no dialog, no message, and no error while the target
device was invalidated and disconnected.
device.token.rotate now records how it delivered the replacement in a
tokenDelivery discriminator ("in-band" | "withheld-cross-device"), additive and
optional so gateways that predate it still decode. The Devices page reports the
recorded fact rather than inferring one from a missing token: the show-once
reveal for an in-band rotation, and an outcome dialog for a withheld one that
states the token is not shown here, that the device was disconnected, that it
picks up the replacement on its next connect with a Gateway token or password,
and that a device which can only sign in with its device token has to be paired
again.
The reveal component takes an optional secret: with one, dismissal stays refused
because the value is unrecoverable; without one there is nothing to lose, so
Escape and backdrop close it like any dialog.
* fix(ui): lead the rotation outcome with what happened, not why
The withheld-rotation dialog opened on a negation, stacked three clauses of
protocol reasoning, and left "do I have to do anything?" to the last sentence.
It now reads outcome, then next step, then the one exception, with the security
rationale demoted to a muted trailing line: the device is named, told to be
signed out, and the common case says nothing else is needed. The rotate callback
carries the row's resolved label so the dialog names the same device the operator
clicked instead of rederiving the precedence.
The reveal component takes paragraphs plus an optional muted note, so an outcome
and its exception stay separate instead of collapsing into one block.
* fix(ui): give the rotation outcome a visual hierarchy
Four near-equal paragraphs in a plain box read as one undifferentiated block:
the reassurance, the conditional branch, and the footnote all carried the same
weight, and the calm text was rendered in muted grey because it reused
.exec-approval-sub.
The dialog now has zones. A success mark in the shared icon-tile geometry sits
beside the title, the two reassurance sentences carry full --text weight, the
one conditional branch becomes a house .callout.info so it cannot be skimmed
past, and the security rationale is a 12px muted footnote. Spacing follows the
card's 4px rhythm.
The acknowledge button follows the confirm-dialog convention: the accent button
is the action that commits something, so a show-once reveal keeps it and a
report of work already done uses the neutral button.
* fix(ui): collapse the rotation outcome to a single reassurance
The body zigzagged: it announced a new token, then raised a disconnect, then
spent the next line walking that alarm back. The middle beat existed only to be
contradicted, so it is gone -- the disconnect is transient and self-healing, and
naming it bought the reader nothing.
The title now carries the announcement and the device name, the body is the one
reassurance, and the callout drops its "this device" echo. The neutral dismiss
button gets --border-strong: .btn's resting border is within ~4/255 of --card in
dark, so it read as unpainted on this surface. The callout wraps balanced.
Both dialogs now pass a single message, so the paragraph-array support added for
the previous structure is removed rather than left speculative.
* fix(ui): key the rotation callout to an observable symptom
"If it can only sign in with its device token" asked the operator to know which
credential the device holds, which is internal state they cannot see. The
condition is now the symptom that state produces: the device does not come back
on its own. Same behavior, same technical truth, decidable from the Devices page.
* fix(ui): reject contradictory device-token rotation results
The Devices page inferred the outcome from an unchecked RPC payload, so an
explicit `tokenDelivery: "in-band"` with no token, a withheld result that
carried one, or a delivery mode this client predates all rendered as a
successful withheld rotation - telling the operator the device re-credentials
itself while their previous token was already invalid.
rotateDeviceToken now parses the response into the closed outcome: the two
current pairs, the two legacy omission states from gateways released before
tokenDelivery, and every other explicit pair through the existing error path,
which surfaces as the page's danger callout carrying the recovery step.
* chore(plugin-sdk): refresh the API contract for the rotation result schema
Registering DeviceTokenRotateResultSchema in PluginLifecycleProtocolSchemas is
what makes the generators emit the Swift model, and that registry sits in the
type closure of 18 plugin-sdk exports, so their closureHash entries move. This
is generator output rather than a hand edit: the export set and every declaration
are unchanged, only the hashes.
* fix(ui): reject malformed device-token rotation envelopes
The parser trusted the envelope around the fields it read. A null, scalar, or
empty payload carries neither tokenDelivery nor token, so it matched the legacy
omission state and produced the reassuring completion dialog after the previous
credential had already been invalidated. A blank token did the same, though the
result schema bounds token to a non-empty string.
Only DeviceTokenRotateResultSchema's shapes are accepted now: the payload must be
a record that identifies the grant it rotated - every Gateway answering this
method returns deviceId and role, before and after tokenDelivery existed - and
token must be either absent or a non-empty string. This has to happen here
because the browser Gateway client resolves frame.payload directly, so the
registered result schema never runs on the client.
The mid-flight-reconnect reveal test asserted on a two-field stub no Gateway
sends; it now uses the real response shape.
* fix(ui): reject rotation results that do not answer the request
The rotation parser accepted any envelope naming some device and role. It
never checked scopes or rotatedAtMs, both required by
DeviceTokenRotateResultSchema, and never compared the returned grant with the
one that was requested -- so an incomplete or unrelated reply opened a success
dialog claiming a credential the operator may still hold had been replaced.
Require the full result shape and bind it to the requested grant, comparing on
the same trim normalization the device-auth store applies. The page fixture
now echoes the requested grant instead of hardcoding device-1, which is what
let a cross-device test accept a success dialog for the wrong device.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(ui): complete the rotation fixture the epoch test resolves
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(ui): update token rotation E2E envelope
* fix(protocol): correlate rotated token delivery
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Patrick Erichsen <patrick.a.erichsen@gmail.com>
* fix(ai): preserve provider terminal error message in stream transports
The managed OpenAI Responses transport (and the mistral/google-shared
providers) threw a literal "An unknown error occurred" on terminal
error stop reasons, discarding the provider's terminal fact (for example
"Provider incomplete_reason: content_filter"). The catch-side
projectProviderError() then overwrote output.errorMessage with the
generic string, so users saw an unknown error and failover classified
the deterministic failure as a transient timeout, rotating models and
re-sending the same filtered prompt.
Align all three sites with the canonical siblings
(openai-responses-shared.ts, transport-stream-shared.ts) by preserving
output.errorMessage when present.
* test(ai): await the transport stream factory before result()
* fix(ai): narrow the repair to the managed OpenAI transport
ClawSweeper correctly noted the mistral and google-shared edits were
no-ops: streamMistral never assigns output.errorMessage and the Google
path throws its own constructed terminal error before this fallback.
Revert both to keep the PR scoped to the proven managed-transport
repair.
* fix(ai): recover rejected WebSocket compaction
Native OpenAI WebSocket turns now rebuild full history over SSE when a replayed encrypted checkpoint is explicitly rejected, then suppress that checkpoint on later turns.
* ci: refresh merge check
Move llama.cpp chat and local embeddings onto a verified externally managed llama-server runtime. Remove the in-process native runtime, forked embedding workers, and node-llama-cpp dependency while preserving guided setup, local GGUF models, tool-capable agent runs, diagnostics, and operator docs.
* fix(ai): managed OpenAI transport understates gpt-5.5 priority pricing by 20%
The managed Responses transport carried its own copy of the service-tier
pricing table in openai-responses-debug.ts (flat 2x for priority). The
canonical helper applyResponsesServiceTierPricing in
openai-responses-shared.ts was updated to 2.5x for gpt-5.5 priority
(#117298) but the transport copy was never touched, so UI cost for
managed-transport fast-mode turns understated the real spend.
Delete the duplicated table and delegate the transport's pricingOptions
to the canonical model-aware helper. Azure intentionally remains
tier-unpriced (it sets no pricingOptions).
* test(ai): pin managed-transport gpt-5.5 priority pricing at the transport boundary
ClawSweeper correctly flagged that the helper-only test also passes on
unfixed main. Add a mocked managed-transport run with a priority
response and nonzero model cost asserting the final usage cost is 2.5x
(fails at 24 vs 30 on the pre-fix flat table).
Treat <internal> blocks as private reasoning in the shared parser and remove Telegram raw-reasoning fallbacks. This keeps model reflections out of user-visible replies while preserving surrounding answer text.
Closes#122623
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
Co-authored-by: WangYan <wang.yan29@xydigit.com>
* feat(gateway): provision paired node worker installs
* docs: record node local-install PR chain
* chore(protocol): regenerate node worker manifest models
* fix(gateway): resolve node worker build through the device runtime
Repairs a rebase artifact: startup referenced a stale deviceNodeRegistry
binding; the device runtime now owns the connected-node lookup via
resolveWorkerBuild.
* feat(protocol): add portal methods and event
Bump the reviewed protocol owner-module count from 55 to 56.
* feat(gateway): add portal service and reverse proxy
* feat(agents): add portal tool
* fix(gateway): refine portal URL and proxy auth
* refactor(gateway): keep portal helper types private
* fix(gateway): declare portal transport service
* test(gateway): satisfy portal proxy lint
* test(gateway): narrow websocket payload types
* refactor(protocol): compact portal schema exports
* fix(gateway): export portal protocol types
* feat(ui): add portals page
* docs(gateway): add portals guide
* fix(gateway): dial portal targets via localhost dual-stack
Vite and other Node >=17 dev servers bind ::1 only for localhost, so a
fixed 127.0.0.1 dial 502s on the default path. Use hostname localhost
with family autoselection and rewrite Host to match.
* fix(gateway): type portal dual-stack connection
* fix: satisfy portal integration gates
* fix(gateway): isolate portal cookie jars per target
Cookies are hostname-scoped, not port-scoped, so the per-port origin
split alone let Gateway plugin-auth cookies reach agent-run targets.
Forward only cookies carrying this portal's own name prefix (stripped),
rewrite target Set-Cookie names to the prefixed form incl. the WS 101
handshake, and drop Domain attributes.
* fix(ui): detect unreachable portals behind proxied gateways
Probe the portal origin from the browser (no-cors, 4s timeout) and show
a recovery notice with the gateway-host URL instead of a dead iframe
when only the gateway port is exposed (Serve/Funnel/reverse proxy).
Docs: cookie isolation + reachability; zh-CN glossary entry.
* test(ui): satisfy portal reachability lint
* test(gateway): provide control UI request hosts
* chore(protocol): regenerate after rebase
* fix(gateway): namespace portal auth cookies by listener
* fix(gateway): scope portal token URLs to write-capable clients
The portal bearer token rides in the summary url/tokenQuery; portal.list
is operator.read and portal.changed fans out to read subscribers, so a
read-only client could harvest an openable URL. Make those fields
optional, redact them from read-scope list responses, and drop them from
every portal.changed broadcast; write/admin clients still receive them
and the UI refetches the list on change.
* docs(web): list the portals route
* fix(gateway): type portal open credentials
* docs(gateway): clarify portals PORT/PUBLIC_URL are agent-set
Opening a portal creates only the proxy listener; the agent sets PORT
and PUBLIC_URL in its own exec command, matching the portal tool
contract. Removes the implication of an automatic env handoff.
* chore(protocol): regenerate portal models
* style(gateway): format portal method-order assertions
Rebase union-merge left the portal.list assertion wrapped; oxfmt fits it
on one line.
* chore(plugin-sdk): refresh API baseline after rebase
* chore(plugin-sdk): refresh API baseline after rebase
* chore(protocol): refresh portal event order after rebase
* chore(plugin-sdk): refresh API baseline after rebase
* fix(gateway): pin portal referrer policy to no-referrer
The portal URL carries its bearer token in the query, and upstream
response headers are copied verbatim, so a target answering with
Referrer-Policy: unsafe-url could leak that URL to every third-party
origin it references. Force no-referrer after the copy and drop any
inbound Referer that still carries the token before forwarding.
* fix(skills): keep ClawHub publisher identity from search through install
ClawHub search returns one entry per publisher, so several results can share a
slug. Every client collapsed the selection to that bare slug before calling
skills.detail and skills.install, and ClawHub answered 409 AMBIGUOUS_SKILL_SLUG
with no in-product way forward.
searchClawHubSkills now records the publisher-qualified reference once, on the
result that carries it, and the Gateway protocol documents it. skills.detail
parses the same reference grammar skills.install already accepted, so review and
install cannot resolve to different publishers. Control UI carries that one
reference through row actions, detail, busy state, and acknowledgement retries,
and shows it so otherwise identical rows are distinguishable.
Fixes#117633
* fix(apps): send the ClawHub publisher reference from native skill browsers
macOS, iOS, and Android read the qualified reference from search results and use
it for skills.detail, install, busy state, installed matching, and list identity,
so two publishers sharing a slug stay distinct instead of collapsing into one
ambiguous request.
* fix(skills): refuse external-source skill detail instead of reading a same-slug skill
ClawHub has no source-qualified read endpoint, so a skills-sh reference parsed
down to its bare slug would have returned a registry skill's card while install
resolved the external artifact. Review and install could name different skills.
skills.detail now fails closed on any reference that carries a source, and the
macOS and AgentPro rows show the publisher reference next to the summary instead
of only when a summary is missing, so same-slug rows stay distinguishable.
* chore(apps): refresh native i18n source baseline for the skill row references
* refactor(skills): drop the unread search-result ownerHandle field
installRef is the one reference clients send back, and no client reads the
publisher handle separately, so the protocol and Control UI carry one field
instead of two.
* fix(skills): name the next step when external skill detail is refused
Clients that gate install behind a successful review would otherwise see only a
refusal, so the error names the direct install path and the CLI equivalent.
* fix(macos): use a doc comment on the ClawHub row subtitle
swift-format's docComments rule requires doc comments on declarations; the
subtitle property carried a regular comment and failed macos-swift.
* fix(skills): carry ClawHub trust state to clients that can install
Forwarding installRef let clients install the exact publisher the operator
picked, including external skills-sh sources. It did not forward the trust
state that says ClawHub never scanned that source, so iOS AgentPro — the one
surface that installs in a single tap with no review step — could install an
unscanned artifact with nothing on screen saying so. The CLI already labels
these (docs/clawhub/cli.md, docs/cli/skills.md); native clients could not,
because trustState was never on the wire.
trustState becomes an optional field on SkillsSearchResultSchema. It is purely
additive: older clients ignore an unknown key and the field is absent for
registry results, so downgraded readers are unaffected and no protocol version
moves.
Every client that renders a search row now shows "Not scanned by ClawHub",
matching the CLI wording exactly: iOS AgentPro in the row above the install
button, macOS and Android beside the review action, and Control UI on the row
that explains why review is refused for these sources.
Covered by a wire assertion that the state reaches clients for an external
source and stays absent for registry rows, plus decode-and-label tests on the
shared Swift kit and the Android parser, and a Control UI render assertion.
* fix(ui): size the ClawHub detail dialog to a refusal message
Refusing detail for an external source made an error-only dialog reachable.
The shared preview panel reserves a tall reader height for skill documents, so
a two-line refusal rendered in a mostly empty dialog and read as broken rather
than deliberate. Found by inspecting the review captures.
* revert(ui,apps): drop the ClawHub trust label layer
Maintainer product decision: skills.sh runs its own scanners, so OpenClaw does
not add a second alert layer in the apps. Removes the label from Control UI,
iOS, macOS and Android, and drops the trustState wire field that nothing would
render. The CLI keeps its existing label; changing that is a separate call.
Publisher identity, the fail-closed detail refusal, and the message-only dialog
are unchanged. Splits the oversized skills view test file to satisfy max-lines
without a suppression.
* test(ui): fix ClawHub skill fixture checks
* chore(plugin-sdk): refresh API baseline
---------
Co-authored-by: Patrick Erichsen <patrick.a.erichsen@gmail.com>
Jesse Merhi