v1.1.0

Added

• The digest follows the session's first request. The plugin sends that request to context_bootstrap as query when it has 12 characters or more after trimming and is no slash command. The Codex app-server harness sends the request alone as currentUserMessage, and the plugin reads that. On the acpx harness the prompt carries OpenClaw's inbound chat context ahead of the request, and the plugin strips the marked context blocks, the active goal line and the message-tool delivery hint before it measures or sends anything. A Telegram reply-quote block stays in. A room event (messages.groupChat.unmentionedInbound: "room_event") sends no query. An engine with decision 0034 (lumberroom/lumberroom PR #159) adds up to four related facts by default and orders the profile by relevance. When the first turn carries no request (a bare /new or /reset, a greeting, a slash command), the plugin serves a digest without query and refetches it once on the first turn that has one; that turn changes the system prompt once and costs one call. Every later turn reuses the digest's bytes. Heartbeat turns and turns with no session key send no query. Any tool_error on a call that carried query gets one retry without it in the same turn, within digestTimeoutMs; an older engine ignores query. digestQuery: false turns it off.

v1.0.2

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.

v1.0.1

Added

• Ships the lr-review skill for working the dreaming review queue.

v1.0.0

The first release, for OpenClaw 2026.9.6 and later on Node 24.16 or 26.1 and later.

Added

• lumberroom takes OpenClaw's memory slot as a plugin of kind memory, against lumberroom.cloud or
  a self-hosted engine.
• Recall on every eligible turn. The context_bootstrap digest enters the system prompt once per
  session and stays byte-identical across its turns. memory_search hits for the user's message
  go ahead of that message for the model only.
• The engine's own tools under their own names: memory_search, memory_write, registry_get and
  memory_forget by default, with the descriptions and schemas the engine serves. tools in the
  config picks from the rest. dreamingReview adds review_queue and review_decide on
  lumberroom.cloud.
• An owner gate for shared chats. Group, channel and thread turns reach memory only when the sender
  is listed in ownerIds, and so does a turn from a room on the main or global session when
  session.groupScope or session.scope routes rooms there. The Control UI, the TUI and the CLI
  keep their digest and recall there. A refused turn gets no digest, no
  recall, no tools and no engine call.
• A write guard. write, edit and apply_patch against MEMORY.md, USER.md or memory/ in
  the workspace are blocked with a reason that points the model at memory_write. The guard reads
  @, ~/ and file:// targets the way OpenClaw's tools do and ignores letter case on macOS and
  Windows. It stays on when the config is invalid.
• Two sign-in modes. token reads a static bearer or an lr_ API token from
  $OPENCLAW_STATE_DIR/.env. oauth signs in through a browser or a pasted redirect URL, and
  refreshes under a file lock, so a gateway and a CLI process sharing one token file never replay a
  refresh token.
• openclaw lumberroom setup, login, logout, status and import. Setup takes the memory
  slot, grants conversation access, turns off both switches that would load memory-core as a
  dreaming sidecar, and adds tools.alsoAllow where the tool profile would hide the plugin's
  tools. status names every problem it finds. import sends existing MEMORY.md, USER.md and
  memory/*.md entries to the engine's review queue, never straight into the store.
• The gateway logs an error at start when either dreaming-sidecar switch is open.
• A bundled snapshot of the engine's tool listing, so the plugin offers its tools before the first
  live tools/list answers, and a cache of the last live listing under the state directory.
• Setup turns off OpenClaw's session-memory hook, which writes memory/*.md through fs where
  the write guard cannot see it; status names it when it is on.
• With owners listed, setup sets the message queue to followup, so a non-owner's message cannot
  steer into an owner's running turn, and its overflow to drop: "old", so dropped messages never
  fold into one turn; status names any steer or collect mode and a summarize drop.