v1.1.0

Added

• The digest follows the request. The first non-trivial turn after init, a reset or a
  compression sends its message to context_bootstrap as query, stripped and clipped to 1,000
  characters, when it runs 12 characters or more. An engine with
  decision 0034
  adds the facts related to it and orders the profile by it. An older engine ignores the argument,
  and so does lumberroom.cloud until it takes 0034. If a server refuses it, the digest goes out
  without query on the next turn. digest_query: false turns it off.

v1.0.4

Changed

• lr-review works undated facts: it fills a start date the fact's own text names, through review_queue source undated and the fill_date verdict. Needs a server with the undated source.

v1.0.3

Added

• Registers the lumberroom:lr-review skill for working the dreaming review queue.

Fixed

• import-builtin no longer points at lumberroom ingest review, which is not a subcommand. The
  hint, the README and the spec now name lumberroom ingest list --state proposed and
  lumberroom ingest approve.

v1.0.2

Fixed

• hermes lumberroom login prints the sign-in URL at once when stdout is a pipe or a file.
  Hermes v2026.9.21 leaves a piped stdout block-buffered, so the URL sat in the buffer while login
  waited for the browser or a paste, and a script driving the login never saw it. A terminal was
  never affected. Every sign-in line now flushes. Later Hermes releases line-buffer a piped stdout
  themselves.
• The end-to-end gate runs against Hermes v2026.9.21 as well as main. The driver passed
  profile_home to set_secret_scope, which v2026.9.21 does not take; it now checks the signature
  first. The missing context_bootstrap row in that failing run came from the same crash: recall
  never ran. The digest reaches the engine on both releases.

v1.0.1

Fixed

• A token refresh answered with a 5xx no longer keeps the refresh token. The engine spends the
  refresh token before the steps that can still fail with a 500, and a proxy can answer 5xx after
  the engine answered, so presenting the token again could trip the engine's replay check and
  revoke the whole sign-in on every machine. The plugin now drops it; the live access token serves
  until it expires, then hermes lumberroom login signs in again. A refresh that never connected
  still keeps the token.

v1.0.0

Added

• The lumberroom memory provider for Hermes Agent. Recall through Hermes's prefetch hook on
  every turn, the digest on the first, and the engine's own memory_search, memory_write,
  registry_get and memory_forget from the live tools/list, over /mcp through the official
  mcp SDK. lumberroom.cloud with browser sign-in is the default; base_url points it at a
  self-hosted engine, which runs the same code with a static token or OAuth.
• Setup turns off Hermes's MEMORY.md and USER.md; hermes lumberroom import-builtin sends
  their entries to the engine's proposal queue.
• Gateway chats admit only listed owners, turn by turn; cron counts as the owner.
• dreaming_review, off by default: Hermes may list and act on lumberroom.cloud dreaming proposals
  through review_queue and review_decide when the host is lumberroom.cloud and the server offers
  them.
• A cross-process refresh fence: two Hermes processes on one profile never present the same
  refresh token twice.
• scripts/hermes-plugin-test.sh, the end-to-end gate, in token and OAuth mode against an engine
  checkout.

Moved from client/hermes/ in the engine repository on 25 September 2026, with its history.