Hermes proposes one gateway-owned conversation across terminals, apps and background jobs

Hermes' single-session-owner proposal, opened September 9 and updated September 12, would bring terminal, desktop, API, editor, bot and scheduled work under one execution authority. The point is not another synchronized chat view: clients would become readers and submitters, while the gateway owns the canonical writer and queue. The proposal explicitly says it is not merged and awaits a maintainer decision. That matters alongside the September 11 state.db patch release, which repairs current storage failures without establishing that this larger architectural replacement has shipped.

The facts

- The proposal gives CLI, Ink, Desktop, API, ACP, bots, cron and Kanban a shared gateway-owned session and durable FIFO admission queue. - Queued admissions are intended to survive an owner restart; a started admission that loses its owner instead becomes unknown and is not automatically replayed. - Retries carry the original admission identity and committed payload rather than constructing a fresh request. - Profile, provider or issuer, generation and revision boundaries are named parts of the proposed transport contract. - The served profile set remains a startup snapshot, so adding a profile still requires a gateway restart. - The September 12 status reports green CI but explicitly excludes live certification of hosted-room execution across two profiles on one multiplexer.

Why it matters

Operators increasingly use several interfaces to the same agent: a phone to start work, a terminal to inspect it and a desktop window to intervene. Shared files alone do not make those interfaces agree about who may execute next. A single owner can make cancellations, approvals and recovery coherent, but it also becomes a critical availability boundary. The useful design question is how clients behave when that owner is absent, not how many interfaces can display the same transcript.

Current

Inspected on 2026-09-13. The Hermes Agent release baseline is v2026.9.11 published 2026-09-11T19:20:31Z. The main source was open (not merged) when captured. Proposals, merged mainline changes and packaged releases are different availability states; the linked release baseline does not imply that open proposals have shipped.

Evidence

Primary evidence: NousResearch/hermes-agent PR #106742 (https://github.com/NousResearch/hermes-agent/pull/106742). Supporting context: Nous Research — September 11 state.db release context (https://github.com/NousResearch/hermes-agent/releases/tag/v2026.9.11); Hermes Agent — official product README (https://raw.githubusercontent.com/NousResearch/hermes-agent/main/README.md). Sources were read directly through public pages or GitHub API bodies. Contributor test results are attributed claims, not independently reproduced results.

Operator take

Keep the released reliability patch and this proposed authority model on separate rollout tracks. For the latter, require demonstrations of duplicate submission, interrupted execution, reconnect after compression and two-client cancellation against the same admission. A queue can deliver a request once without proving that an external side effect occurred once. Preserving an unknown result is therefore an operator safeguard, not an inconvenience to remove with an aggressive retry loop.

Caveat

The architecture remains proposed. CI totals and cross-interface demonstrations are contributor reports, not independently rerun acceptance tests. The released state.db repair is supporting context, not proof that this queue or writer contract is available in the current package; the proposal also names an uncertified multiprofile hosted-room path.