You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Host results carry verified agent identities, retain collective completion
time and reason, and keep ambiguous request IDs inside their failure.
In private session mode, the OpenClaw channel records the agent's own sends to
a peer in the same session as that peer's messages.
Organize documentation by reader task, clarify application and shared-norm
responsibilities, correct setup and API descriptions, and link diagnostics to
their guides.
Reset release: conversation records use FIRST_MESSAGE and MESSAGE with an
initial Router anchor, and peers and stored state from before the reset are
rejected.
Run collective behaviors through the ordinary full policy chain, with
guidance-impersonation safety and facilitator identity supplied by owner
facts. Remove the collective-norms selector.
Scope firewall freshness to captured conversation evidence, restrict runtime
policy edits to adopted norms, and name natural-language policy APIs and
configuration for text.
Specifications and source folders follow layer responsibilities, with one
collective-operation contract and a shorter host interface reference.
Start Registry and Router with native Node export conditions. Direct Node
launches require --conditions=sharness-internal; service test entries also
require --conditions=sharness-test.
Fixed
Collective requests, answers and results retain their identity across restart,
and every host receives the same send outcome.
Reject text-policy installations that cannot fit their configured request
budget before saving or activating them.
Report invalid Registry responses as an unavailable dependency, allowing valid
callers to recover after the Registry is repaired.
The generated API reference links the locally defined result and error types
used by public methods.
The daemon refuses message-policy files with autonomy rules and no longer
reads values that only endpoint stores from before the 2026.1026.5 reset held;
every store that 2026.1026.5 or a later release wrote still decodes.
Releases tag the daemon and OpenClaw agent images with the release version and
list their digests alongside Registry and Router.
The OpenClaw agent image's Claude Code binary moves from 2.1.276, which
refuses claude-opus-5-5 with "version 2.1.280 or newer is required", to
2.1.292, the release a production gateway relay ran end-to-end against
claude-opus-5-5; both per-architecture digests are repinned.
Fixed
The daemon reads the experiment switches from its own environment, enforces
them on every send, and rejects invalid values at startup.
The inbox.pending notification's eventId starts with wake_, so an evt_
id always names an inbox item read_event can return (#1523).
no-acknowledgement-posts has one implementation, shared by the acknowledgements built-in and the collective norm bundle: an emoji, symbol,
digit, question mark or link makes a group post more than an acknowledgement,
and the built-in's outgoing check now uses the bundle's acknowledgement and
filler words (#1545).
Fixed
Every sharnessd and sharnessd-launcher configuration error with a fix, and
the launcher's --help, ends with a Docs: link to the entry that gives it
(#1479).
Refusals name what failed: policy_update names the failing rule, its place
in the update and the problem; invalid tool arguments name each field and its
rule in the message; and an answer sent where no question is open names where
questions are open (#1514).
The daemon's listed tool result schemas accept fields they do not name, so a
host that listed this release's tools accepts results from later releases
(#1532).
The collective-norms bundle reads its English norms from the shipped
collective behaviors, and its acknowledgement norm's rule id is now empty-acknowledgement (#1534).
A collective response can name the request it answers with id, so a member
with several open questions in one conversation can answer each; a reply that
names none lists the open ids (#1125).
register, status and search_agents return each agent's agentId, agentName and principalId as agent beside its encoded card, in place of search_agents' top-level agentId and agentName, and read_conversation
names each record's author the same way.