Tandem v0.6.8

See the assets below to download the installer for your platform.

v0.6.8 (2026-07-09)

Tandem 0.6.8 is a hosted-v1 readiness follow-up. It hardens hosted session
isolation, completes the native Linear webhook path and Webhooks hub, expands
memory isolation across tenant/org/user boundaries, stabilizes the
tandem-server test path used by CI and local full-suite runs, settles the
0.6.8 BUSL/open-core licensing posture, removes a production-linked GPL
dependency, and updates the OpenAI Codex catalog to the documented GPT-5.6
preview model ids.

Linear Webhooks And Webhooks Hub

Automation V2 now supports direct Linear-to-Tandem webhook delivery without a
bridge or unsigned dev mode. Linear triggers use the native
linear-signature HMAC-SHA256 scheme over the exact raw request body, validate
the signed webhookTimestamp against the replay window, and fail closed with
provider_secret_not_imported until an operator imports the Linear-owned
signing secret.

Provider-owned secret lifecycle is now explicit: admins can import or re-import
the Linear signing secret through the management API, re-import rotates the
tenant-scoped material and retires the old version, and Tandem-generated
secret rotation is refused for provider-owned triggers. Trigger responses
surface the verification lifecycle from awaiting_secret through active
verified delivery, and the TypeScript client exposes
importWebhookProviderSecret.

The Control Panel now has a dedicated Webhooks hub with Linear setup UX. The
guide and internal docs cover Linear verification, secret import, rejection
reasons, SDK/API usage, and troubleshooting. Rejected Linear events now produce
rejection evidence without entering the runnable inbox, while verified Linear
deliveries dedupe by linear-delivery id.

Memory Isolation And Eval Gate

Memory context-tree endpoints and storage are now tenant-scoped. Context URI
resolution, tree traversal, layer generation, and layer writes all use the
request tenant; foreign-tenant nodes behave like missing nodes, so callers do
not get a cross-tenant existence oracle. The store schema now allows different
tenants to own the same context URI independently.

Ordinary tenant-local memory reads now enforce org-unit ownership. Records can
carry an owner_org_unit_id, read filters receive the caller's verified
org-unit memberships, and memory_put refuses to stamp ownership for an org
unit the writer does not belong to. Source-bound and knowledge-scoped records
continue to flow through the enterprise grant model.

Vector memory chunks now include a per-user subject dimension, and archived
chat exchanges are stamped with the session actor so one user's historical
conversation chunks do not become a tenant-wide recall pool. Prompt injection,
memory list, and memory search paths thread the resolved caller subject through
the governed filter, while imports, consolidation, and model-tool writes remain
shared by design. Unbacked team and curated write/promote paths now fail
closed instead of accepting tier labels with no storage backing.

A new cross-user memory isolation eval dataset, baseline, CI regression gate,
and HTTP/app-state test matrix cover tenant, subject, DM/group, distillation,
forged project, and org-unit boundaries. This gives the memory scoping work a
standing regression signal instead of relying only on per-feature unit tests.

Hosted Runtime Isolation

Hosted sessions now enforce actor scope on command routes, including
/session/{id}/command, and hosted mode fails closed when actor context is
missing. Local desktop and development flows keep compatibility for implicit
sessions, but hosted deployments now have explicit coverage for cross-actor
session access attempts and assertion-verification edge cases.

Licensing And Compliance

The BUSL Additional Use Grant for Tandem BUSL components now allows internal
production use for any organization regardless of revenue. A commercial license
is required to provide the licensed work, or a substantially similar product,
to third parties as a managed, hosted, SaaS, white-label, embedded, or other
commercial offering. The BUSL Change Date policy is now rolling: each release
stamps BUSL components to release date + four years, after which that version
converts to the existing Change License.

tandem-enterprise-server is now licensed under BUSL-1.1, matching the
commercial enterprise layer already used by the governance and plan-compiler
crates. The permissive @frumu/tandem-enterprise installer stays permissive,
but its README now discloses that downloaded enterprise binaries contain
BUSL-licensed components. The repository license map now covers all Rust
workspace members, published npm packages, and Python package metadata, with a
CI guard that fails when the map and manifests drift.

The browser/tooling markdown conversion path no longer ships the GPL-3.0
html2md dependency. It now uses Apache-2.0 htmd, preserves iframe embed
URLs as markdown links, and keeps the remaining auto_generate_cdp GPL
exception documented as build-only code that is not linked into distributed
artifacts.

Provider Catalog And Desktop Healing

The OpenAI Codex provider catalog now exposes the documented GPT-5.6 preview
models as explicit gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna entries
instead of the generic gpt-5.6 placeholder. Persisted bare gpt-5.6
defaults and stale desktop selected-model dispatches now heal to the compiled
gpt-5.5 default, avoiding startup or dispatch failures from a model id that
is no longer valid.

Test And CI Stability

The tandem-server suite now serializes environment-mutating state construction,
cleans up process-wide test overrides, and documents local nextest recipes that
match CI's profile, feature flags, stack size, and isolated TANDEM_HOME.
This addresses the class of tests that passed in isolation but failed under
full parallel execution.

Full Changelog: v0.6.7...v0.6.8