Release summaries
cmp-execution-lanes-delegation
Team workers run under bounded, least-privileged delegation with typed help and two-phase lifecycle work orders, and team readiness and evidence follow the finalization rules
integrator
createTeamRun accepts an optional delegation authority (TeamDelegationAuthorityV1 or a per-run resolver) and workOrderTimeoutMs. Every team worker is launched with a TeamWorkerDelegationV1 projection (derived principal, objective plus claimable tasks only, role tools within the lead's tools, role-required grants only, lease wall-clock bound, no further delegation) in TeamBackendWorkerLaunchV1, and every backend lifecycle call receives its work-order id and deadline. Under an authority, worker-scoped operations must pass the worker's derived principal as actorId. Evidence bundle ids must be saved to the lane store before they are attached, every new run must enable the verification lane, and a worker only claims tasks hinted for its role or unhinted. Typed help uses the canonical agent-protocol help request and response. evaluateLaneReadiness no longer reports ready on a verifier self-signoff.
agent
A team worker acts only as its derived principal, sees only the run objective and the tasks it may claim, holds only its role's tools within the lead's and the grants its role requires, and cannot delegate further. Ask for help with requestHelp and a canonical AgentHelpRequestV1 instead of guessing; the lead, or a responder the host help-response authority verifies, answers with respondToHelp. No worker of the run may answer.
maintainer
Team readiness lives in runtime/team-readiness.ts and shares its approval rules with finalization (lanes/team/finalize.ts); it runs for every snapshot with team state and resolves keyed lane policies through policyRegistry or the default team.coordinated policy. Delegation projections are derived in lanes/team/team-delegation.ts and enforced in team-delegation-authority.ts (including the run's delegation binding); typed help is in team-help.ts and team-help-response.ts; backend lifecycle calls go only through runTeamWorkerLifecycle (team-work-order-runner.ts, team-work-orders.ts, team-work-order-store.ts, team-work-order-apply.ts), never inside withExclusiveTeamMutation. The ledger row delegation.bounded-and-projected is implemented with residual blockers, and completion.evidence-backed names its remaining legacy gap.
cmp-fix-w8-delegation-boundary-composition
Preserve persisted team authority and the exact bounded projection through concrete backends
integrator
Reopening a delegation-bound team run requires its current host authority, even when the new runtime omits the delegation source. Never-bound runs retain legacy behavior. Queue and workflow worker payloads preserve the existing delegation, workOrderId and deadlineAt without forwarding whole launch objects, descriptors, provider handles or unrelated parent context.
maintainer
This patch closes the missing-source and concrete-adapter gaps recorded in the earlier cmp-execution-lanes-delegation capsule. Public optional field types and lifecycle run/worker handle maps remain unchanged. No production host, grant or worker execution is activated.
communication-agent-protocol-owner
Communication projections use the canonical pure help definition
integrator
Existing Communication lifecycle projections and validation remain compatible with canonical durable help request types.
communication-os-ai-gateway-models
CommunicationOS thread summaries and AI reply drafts run only through the Vercel AI Gateway. Users can switch between allowlisted Anthropic, OpenAI and Google models per request, and governance managers set the workspace default.
integrator
comms.thread.summarize and comms.reply.draft 1.1.0 accept an optional allowlisted Gateway
model. The change is additive, with the named errors AI_MODEL_NOT_ALLOWED and AI_MODEL_OVERRIDE_DENIED. COMMUNICATION_OS_AI_MODEL_CATALOG is the single allowlist. communication.ai.model-settings.get|set and the communication.ai.model-setting.changed event define the workspace setting. communication-os-agent adds ./thread-assist-gateway over createVercelGatewayModel. The client adds getAiModelSettings, setAiModelSettings and CommunicationOsApiError.code.maintainer
communicationos-api reads COMMUNICATION_OS_AI_ASSIST_ENABLED=true plus Gateway authentication (AI_GATEWAY_API_KEY, or the per-invocation Vercel OIDC token header), read per request; /v2/health reports aiAssist. COMMUNICATION_OS_AI_DEFAULT_MODEL is optional and must be allowlisted. Direct provider keys (MISTRAL_API_KEY, ANTHROPIC_API_KEY, COMMOS_LLM_*) are ignored. The Gateway requests EU-zone inference, zero data retention and no training. Gateway request processing itself is global.
customer
On a CompanyOS thread, users with AI access can choose which model writes the summary or the reply draft. Each result says which model generated it. Workspace administrators choose the default model on the CommunicationOS governance page. Drafts still need a person to approve and send them.
communication-os-ai-inbox-triage
CommunicationOS inbox threads can carry advisory AI triage labels (category, urgency, needs reply, reason) produced through the Vercel AI Gateway. Labels are cached per thread revision, journaled with their model and shown as marked AI chips with filters in the CompanyOS inbox.
integrator
New additive operations comms.thread.triage and comms.thread.triage.batch (1.0.0) plus the communication.ai.thread-triaged event. communication-os.ai.use (2.1.0) provides them. The label taxonomy, strict parser and structured-output JSON schema live in @lssm-tech/lib.communication-spec/contracts/thread-triage. communication-os-agent adds ./thread-triage and a triage() method on the thread assist port. The client adds triageThread and triageThreads. ThreadListItem.aiTriage is optional, so existing hosts are unaffected.
maintainer
No new environment variables. Triage uses the existing COMMUNICATION_OS_AI_ASSIST_ENABLED gate, Gateway authentication and model resolution. Labels are stored as communication.ai.thread-triaged journal events, which carry no roomId, so room and message projections are unchanged. The per-principal rate limit (100 model calls per 10 minutes) is kept in process memory, so on serverless it applies per instance.
customer
In the CompanyOS CommunicationOS inbox, users with AI access see AI-marked chips on each conversation (category, urgency, needs reply). Hovering shows why and which model suggested it. They can filter by AI category or by whether a reply is needed. The labels are only suggestions: nothing is sent, assigned or changed automatically.
communication-os-live-channels-governance
CommunicationOS provider connections, room bindings, governance policies, contact identities and audit export inside CompanyOS
integrator
Additive contracts: communication.audit.export@1.0.0 (query) and communication.audit.exported@1.0.0 (event), plus governance document models for retention, consent, escalation, binding consent and identity review evidence stored in governed-work details. startProviderOAuth accepts providerKey.
customer
CompanyOS gains /companyos/communication-os/connections (provider OAuth return, status refresh, room binding with recorded consent), governance policy publishing, per-network contact identities with merge evidence, and a bounded CSV/JSON journal export that is journaled before download.
communication-os-meeting-threads
Meetings join CommunicationOS in CompanyOS: an OPA meeting gets one CommunicationOS conversation with a shared agenda, participants, shared notes and governed follow-ups. The hCircle meeting host becomes production-qualifiable through a startup configuration verdict, a health field, a proven schema ledger and an exact-SHA canary.
integrator
New additive operations communication.meeting_thread.list, .link and .revise (1.0.0, experimental) with COMMUNICATION_MEETING_THREAD_LIMITS. A meeting thread links one OPA meeting (opa_smart_meeting or opa_meeting_workspace) to one room; follow-ups use the existing communication.governed_work.create with kind action, threadId set to the meeting room and details.meetingThreadId. The runtime adds ./ecosystem-meeting-thread, the client adds listMeetingThreads, linkMeetingThread and reviseMeetingThread, and the module adds the MeetingList and MeetingThreadDetail organisms. Existing contracts are unchanged.
maintainer
Meeting threads need no new environment variables or SQL migrations: they are journal events in the existing communication_os journal. The hCircle meeting host stays off unless HCIRCLE_MEETING_RUNTIME_ENABLED=true. When enabled it validates HCIRCLE_GOVERNED_DATA_DATABASE_URL, HCIRCLE_GOVERNED_DATA_MIGRATION_URL (a separate migrator connection), DATABASE_URL, HCIRCLE_MEETING_BYOK_MASTER_KEY and OPA_MEETING_ACCESS_GRANT_PEPPER (each at least 32 characters and different), HCIRCLE_GOVERNED_DATA_SSL_MODE and COMMUNICATION_OS_API_ORIGIN. HCIRCLE_GOVERNED_RECOVERY_TOKEN (at least 32 characters) is also required because governed-data recovery shares HCIRCLE_GOVERNED_DATA_DATABASE_URL and refuses startup without it. /v2/health reports meetingHost and release.sha; a misconfigured host answers 503 on meeting routes instead of aborting startup. The hcircle_meeting_schema_ledger still applies 0039, 0040 and 0041 and now requires the opa_app and opa_migrator roles beforehand and proves forced row-level security on all 21 runtime tables.
customer
In CompanyOS, CommunicationOS > Meetings lists the workspace meetings planned in OPA. Team members with reply access start the meeting conversation, keep a shared agenda, participants and notes, and create follow-ups that are reviewed before anything is sent. Preparation, private notes, transcripts and outcomes stay in the OPA meeting workspace, one click away.
red-gates-dependency-hygiene
Communication spec drops its ProviderSpec dependency and two integrations publish caret workspace ranges; no API changes.
integrator
@lssm-tech/lib.communication-spec no longer installs @lssm-tech/lib.provider-spec. ./contracts/ai-model-catalog keeps the same exports and the same seven allowlisted Gateway models. The entries are now plain data owned by the spec instead of aliases of the provider-spec objects, so compare entries by value, not by reference. provider-linkedin and runtime-local publish caret ranges for their internal dependencies.
sms-consent-stop-start
SMS STOP/START/HELP keywords become CAS preference revisions with sms_opt_out evidence, and outbound SMS fails closed without durable consent
integrator
Inbound STOP, START and HELP (English and French) normalize into keyword signals. STOP and START become compare-and-swap preference revisions in the scoped CommunicationOS journal. The latest revision produces the sms_opt_out evidence that delivery eligibility requires. The delivery dispatcher refuses SMS without an smsConsent gate or an opted-in state. The owner confirmed minor rather than major (release decision B021, 2026-10-26): no in-repo production path dispatches SMS through createCommunicationDeliveryDispatcher (communicationos-worker sends through the Runtime Node delivery port, and createEcosystemDeliveryDispatcher is used only by tests), and sending SMS without durable consent evidence is the compliance defect this change closes.
taketide-public-overview-ownership
Keep public overview rendering in the neutral creator-platform experience and brand content in Taketide composition.
integrator
Use the explicit PublicOverview.web entry with a data-only brand composition; generic and native entries remain unchanged.
ts-01-seq
Tellshift sequences, consent-bound email enrollment, durable due discovery, approved Gmail send and tenant erasure on CommunicationOS
integrator
Run the CommunicationOS migrations to reach 0016_tellshift_sequences_rls. Due discovery calls communication_os.due_sequence_tenants(now) as communication_os_app and receives tenant ids only; all per-tenant work must use the tenant-scoped provider. Wire the Gmail sequence adapter with a grant source and a recipient resolver; uncertain sends are reconciled with the adapter's read-only reconcile.
maintainer
A job sends only its exact approved bytes once: consent, revision approval and policy fingerprint, redaction, content proof, reply stop, reply coverage and credential are rechecked before the claim. A reply stops the sequence; holds never auto-resume; an uncertain send is never resent. Consumers that iterated COMMUNICATION_OS_RLS_TABLES now also see the three sequence tables. The managed migration owner must not be a superuser for RLS to be meaningful; tests prove the path under a NOSUPERUSER owner.