RooCodeInc · mrubens · Aug 27, 2026 · Aug 27, 2026 · Aug 27, 2026
@@ -149,15 +149,14 @@ DEFAULT_COMPUTE_PROVIDER=docker
 # serves, and the embedding pair is create-time: set everything BEFORE the
 # Brain's first boot. gbrain's defaults (text-embedding-3-small, 1536) name
 # models the bundled server does not serve, so the URLs alone are not a
-# working configuration. The openrouter: prefix on the reranker routes it
-# through this deployment's gateway, which strips the prefix before
-# forwarding.
+# working configuration. Self-run model names pass through unchanged and
+# must exactly match the ids served by the upstream.
 # R_BRAIN_EMBEDDINGS_UPSTREAM_URL=http://infinity:7997
 # R_BRAIN_RERANK_UPSTREAM_URL=http://infinity:7997
 # R_BRAIN_INFERENCE_UPSTREAM_API_KEY=
 # R_BRAIN_EMBEDDING_MODEL=BAAI/bge-m3
 # R_BRAIN_EMBEDDING_DIMENSIONS=1024
-# R_BRAIN_RERANKER_MODEL=openrouter:BAAI/bge-reranker-v2-m3
+# R_BRAIN_RERANKER_MODEL=BAAI/bge-reranker-v2-m3
 # R_GITHUB_APP_ID=
 # Raw GitHub App private-key PEM with newlines escaped as \n; do not base64 it.
 # R_GITHUB_APP_PRIVATE_KEY=

@@ -2,6 +2,40 @@
 
 This file tracks product releases for Roomote (single monorepo version). Automated release entries are prepended by `pnpm run version`.
 
+## 0.45.0 (2026-08-27)
+
+This release adds secure hosted trial inference and self-run Brain model options, expands GLM 5.3 support, and improves reliability across Fast sessions, pull-request reviews, Memory, and chat.
+
+### Highlights
+
+- Start hosted deployments with secure, spend-capped Roomote trial inference and accurate cost reporting.
+- Run Brain embeddings and reranking on self-hosted infrastructure with multilingual bundled defaults.
+- Use GLM 5.3 and GLM 5.3 Flash across more existing inference providers.
+- Keep Fast sessions, pull-request re-reviews, Memory ingestion, and chat reporting reliable through retries and resumptions.
+
+### Minor changes
+
+- Expand GLM 5.3 and GLM 5.3 Flash availability and recommendations across OpenRouter, Vercel AI Gateway, Requesty, OpenCode Go, Z.AI, and Z.AI Coding Plan.
+- Offer secure, spend-capped Roomote trial inference during hosted setup, record its real model costs in task and cost analytics, and present the option with clearer onboarding copy.
+- Let self-hosted deployments run Brain embeddings and reranking through their own OpenAI-compatible upstream, including an opt-in bundled CPU service with multilingual model defaults.
+
+### Patch changes
+
+- Include the Brain's memory volume and database state in supported self-hosted backup and restore bundles so memories survive host recovery consistently.
+- Treat temporary Brain network outages as queue backpressure so infrastructure restarts do not exhaust individual memory write retries or require manual repair.
+- Render Fast replies as rich Markdown in Discord guild channels instead of falling back to plain unformatted messages.
+- Make suggested tasks in Fast automation reports launch reliably across Slack, Discord, Microsoft Teams, and Telegram while keeping each suggestion card's state in sync.
+- Preserve the full Fast conversation context across resumed turns so follow-up answers continue from the existing thread instead of losing earlier messages.
+- Keep Fast session sidebars and information panels usable on mobile by switching them to the same single-panel layout as task workspaces.
+- Discover Notion pages and database rows that inherit integration access through shared parents, even when Notion search does not return them.
+- Keep automatic pull-request re-reviews running after their previous sandbox shuts down and anchor each sync review on the head commit that was actually reviewed.
+- Keep Standard tasks and Fast web sessions running through retryable inference-provider failures instead of ending the work prematurely.
+- Briefly name the recalled insight that materially influenced an agent's approach without exposing internal memory provenance or identifiers.
+- Keep Fast and standard agents identified as Roomote by removing OpenCode's conflicting injected identity prefix from their system prompts.
+- Render automation report Markdown correctly in Slack while preserving the report's replyable thread footer.
+- Use task-specific wording in Slack inline status updates so progress messages describe the work instead of exposing generic agent terminology.
+- Make transcript tool activity easier to inspect by showing sanitized inputs and readable YAML details without hiding the corresponding tool results.
+
 ## 0.44.0 (2026-08-26)
 
 This release adds shared memory, skill discovery, and presentational widgets to Fast sessions, expands automation delivery across chat providers, and improves session and pull request review reliability.

@@ -220,6 +220,15 @@ as per-task auth tokens or workspace paths.
 | `OPENAI_COMPATIBLE_<NAME>_LABEL`       | Optional     | Display label stored with a named OpenAI-compatible connection.                                                                          |
 | `VLLM_BASE_URL`                        | vLLM         | vLLM OpenAI-compatible endpoint URL, usually including its `/v1` path.                                                                  |
 | `VLLM_API_KEY`                         | Optional     | Bearer API key for a vLLM endpoint that requires authentication.                                                                        |
+| `R_BRAIN_OPENROUTER_API_KEY`           | Memory provider | OpenRouter key that enables Memory embeddings, reranking, and synthesis.                                                             |
+| `R_BRAIN_OPENAI_API_KEY`               | Memory provider | OpenAI key that enables Memory embeddings and synthesis.                                                                              |
+| `R_BRAIN_MODEL`                        | Optional     | Memory synthesis model in the configured provider's naming. Changes apply immediately.                                                  |
+| `R_BRAIN_EMBEDDING_MODEL`              | Before first Memory boot | Embedding model id that sizes Memory's vector storage. Changing it later requires re-embedding.                              |
+| `R_BRAIN_EMBEDDING_DIMENSIONS`         | Before first Memory boot | Output width for `R_BRAIN_EMBEDDING_MODEL`; it must match the served model.                                                   |
+| `R_BRAIN_RERANKER_MODEL`               | Optional     | Memory reranker model. Use OpenRouter's model id for OpenRouter, or the exact bare model id served by a self-run rerank upstream.           |
+| `R_BRAIN_EMBEDDINGS_UPSTREAM_URL`      | Optional     | OpenAI-compatible embeddings endpoint used instead of the Memory provider.                                                               |
+| `R_BRAIN_RERANK_UPSTREAM_URL`          | Optional     | OpenAI-compatible rerank endpoint used instead of OpenRouter.                                                                            |
+| `R_BRAIN_INFERENCE_UPSTREAM_API_KEY`   | Optional     | Bearer key shared by the self-run embeddings and rerank upstreams; omit it for a trusted private-network service.                         |
 
 ### Sandbox providers
 

@@ -89,8 +89,36 @@ embeddings and synthesis, and Roomote forwards them under the Memory key.
 Changing that key later takes effect on Memory's next request, with no
 redeploy.
 
-OpenRouter and OpenAI both support Memory's embedding and synthesis calls,
-but search reranking requires OpenRouter.
+OpenRouter and OpenAI both support Memory's embedding and synthesis calls.
+Search reranking requires OpenRouter unless a self-run rerank upstream is
+configured.
+
+### Run embeddings and reranking locally
+
+Self-hosted Compose deployments can keep embeddings and reranking on their own
+hardware while continuing to send chat synthesis to the configured Memory
+provider. Enable both services and point Memory at the bundled inference server:
+
+```sh
+COMPOSE_PROFILES=brain,local-inference
+R_BRAIN_EMBEDDINGS_UPSTREAM_URL=http://infinity:7997
+R_BRAIN_RERANK_UPSTREAM_URL=http://infinity:7997
+R_BRAIN_EMBEDDING_MODEL=BAAI/bge-m3
+R_BRAIN_EMBEDDING_DIMENSIONS=1024
+R_BRAIN_RERANKER_MODEL=BAAI/bge-reranker-v2-m3
+```
+
+The bundled CPU service uses multilingual models so recall can cross languages.
+For a smaller CPU host, `Alibaba-NLP/gte-multilingual-base` with `768`
+dimensions is a lighter embedding alternative. Choose the embedding model and
+dimensions before Memory's first boot; changing that pair later requires
+re-embedding the corpus.
+
+The two upstream URLs can instead target any OpenAI-compatible embedding and
+rerank server. Set `R_BRAIN_INFERENCE_UPSTREAM_API_KEY` when that server requires
+a bearer key. Roomote forwards model names unchanged to self-run upstreams, so
+`R_BRAIN_EMBEDDING_MODEL` and `R_BRAIN_RERANKER_MODEL` must exactly match the
+models that server exposes, without a provider prefix.
 
 Without a Memory key, Memory stays inert. Agents are not told it exists,
 and nothing is ingested.
@@ -196,7 +224,7 @@ Three settings pick Memory's models:
 | ----------------------------- | ----------------- | --------------------------------- | --------------------- |
 | `R_BRAIN_MODEL`               | Sourced synthesis | your provider's naming            | any time              |
 | `R_BRAIN_EMBEDDING_MODEL`     | Semantic recall   | a plain model id                  | before the first boot |
-| `R_BRAIN_RERANKER_MODEL`      | Search precision  | `openrouter:<provider>/<model-id>` | after a restart       |
+| `R_BRAIN_RERANKER_MODEL`      | Search precision  | provider or upstream naming        | after a restart       |
 
 Leave the first two unset and Memory uses OpenAI's `gpt-5.6-luna` and
 `text-embedding-3-small` through whichever provider you configured.
@@ -205,7 +233,8 @@ The reranker defaults to OpenRouter's `voyageai/rerank-2.5-lite`. Set
 `R_BRAIN_RERANKER_MODEL` to choose another model from
 OpenRouter's reranker catalog. Reranking requires an OpenRouter key; with only
 OpenAI configured, gbrain keeps the unreranked results instead of failing the
-search.
+search. When `R_BRAIN_RERANK_UPSTREAM_URL` is set, use the exact bare model id
+served by that upstream instead.
 
 The synthesis model is applied by Roomote when it forwards the call and passed
 to the provider as written, so use that provider's naming
@@ -243,18 +272,18 @@ matching on keywords alone.
 
 ## Operating it
 
-- **Back up the Memory volume.** The Railway template schedules daily and
-  weekly backups for it. On other platforms, include `/data/brain` in the
-  deployment's normal volume backups. That directory is the Markdown corpus,
-  including pages produced by nightly synthesis. Postgres holds the searchable
-  index, extracted facts, and durable maintenance jobs; keeping it in the
-  normal database backup makes restores faster, but it is not the only copy of
-  the source pages.
-- **Losing Memory is recoverable but not free.** If the volume or isolated
-  `gbrain` database is recreated, Roomote re-registers its clients, resets its
-  ingestion checkpoints, and backfills task history and integration sources.
-  The deployment starts cold until that finishes, and regenerated synthesis
-  may not be byte-for-byte identical.
+- **Back up both Memory stores together.** The Railway template schedules
+  backups for the Memory volume. On supported self-hosted installs, `roomote
+  backup` includes both the `gbrain_data` volume and the isolated `gbrain`
+  database when Memory is enabled. The volume holds the Markdown system of
+  record; Postgres holds the searchable index, extracted facts, and durable
+  maintenance jobs. Restoring only one can leave the storage layout and index
+  inconsistent, so keep them at the same backup consistency point.
+- **Losing Memory is recoverable only from connected sources.** If both stores
+  are recreated, Roomote can reset ingestion checkpoints and backfill task
+  history and connected integrations, but user-saved facts and generated
+  synthesis may not be reproducible. The deployment starts cold until the
+  backfill finishes.
 - **The filesystem cutover rebuilds older memory once.** The first start of a
   filesystem-backed image replaces a Postgres-only Memory service instead of
   trying to merge the old index into an empty checkout. Roomote then repopulates

@@ -24,13 +24,15 @@ deployment on one account, one vendor, or one model family.
 
 ### Managed Roomote inference
 
-Some hosting deployments offer **Roomote inference** with a limited number of
-managed credits during setup. It is separate from your own provider
+Some hosting deployments provision **Roomote inference** with a limited,
+spend-capped credit grant during setup. It is separate from your own provider
 connections: in particular, you can add an OpenRouter key in **Settings >
 Models** even when Roomote inference is active. If the hosting deployment does
-not offer it, the option is not shown. Its key is stored with your other
-provider credentials, so deleting the Roomote inference provider in
-**Settings > Models** disables it permanently.
+not offer it, the option is not shown. Trial token usage and estimated model
+cost appear in task details and [Cost Analytics](/cost-analytics), while the
+remaining credit line reflects the hosted grant's authoritative limit. Its key
+is stored with your other provider credentials, so deleting the Roomote
+inference provider in **Settings > Models** disables it permanently.
 
 For example, a deployment might use:
 

@@ -10,10 +10,11 @@ its work, and returns reviewable pull requests with previews when they apply.
 Run the same single-tenant product on Roomote Cloud or infrastructure you
 control.
 
-Choose the fastest path for your team. In every deployment, you connect a
-ChatGPT subscription or inference provider API key, plus source control and a
-sandbox provider. Connect communications providers and other tools when your
-team needs them.
+Choose the fastest path for your team. Every deployment needs model inference:
+use hosting-provisioned Roomote inference when it is offered, or connect a
+ChatGPT subscription or inference provider API key. You also need source control
+and a sandbox provider. Connect communications providers and other tools when
+your team needs them.
 
 <div style={{ display: "flex", alignItems: "center", flexWrap: "wrap", gap: "8px" }}>
   <a href="https://cloud.roomote.dev/sign-up"><img src="https://roomote.dev/images/deploy-button.png" alt="Deploy on Roomote Cloud" style={{ height: "40px", width: "auto" }} /></a>
@@ -35,8 +36,8 @@ Running Roomote on Proxmox or another server at home? Follow the
 public ingress, and a prompt that a local agent can follow.
 
 By the end of setup, you should have a reachable Roomote URL, sign-in, a
-source-control provider, a ChatGPT subscription or inference provider, at least
-one environment, and a first reviewable Roomote task.
+source-control provider, managed Roomote inference or a connected inference
+provider, at least one environment, and a first reviewable Roomote task.
 
 The canonical, always up-to-date guide for operating your own server lives in
 [`SELF_HOSTING.md`](https://github.com/RooCodeInc/Roomote/blob/main/SELF_HOSTING.md)
@@ -51,9 +52,10 @@ curl -fsSL https://get.roomote.dev | bash
 ```
 
 The installer brings up the full stack from published images and prints a setup
-link. Open it in a browser and the wizard walks you through the rest: sign-in
-provider, source-control provider, a ChatGPT subscription or inference provider
-API key, communications provider, repositories, and a first task.
+link. For this self-hosted path, open it in a browser and the wizard walks you
+through the rest: sign-in provider, source-control provider, a ChatGPT
+subscription or inference provider API key, communications provider,
+repositories, and a first task.
 
 No DNS setup is needed to try it out; pass `--domain roomote.example.com` for a
 production install on your own domain.
@@ -77,7 +79,7 @@ Have these ready before the wizard asks for them:
 - a source-control provider account with permission to install or configure the
   app for the repositories Roomote should reach
 - a ChatGPT subscription or inference provider API key, such as OpenRouter,
-  Anthropic, or OpenAI
+  Anthropic, or OpenAI, unless your host already provisioned Roomote inference
 - a communications provider workspace, such as Slack, Microsoft Teams,
   Telegram, or Discord, if your team wants to start work from chat
 - at least one repository that can become the first Roomote environment
@@ -96,8 +98,10 @@ roomote logs      # tail service logs
 `roomote backup` prompts for a passphrase and writes a versioned `.roomote`
 bundle under `/opt/roomote/backups`. The bundle contains PostgreSQL, the
 deployment configuration and encryption/signing keys, local MinIO artifacts,
-schema metadata, and the exact deployed image identities. Store the passphrase
-separately in your secret manager; the backup cannot be restored without it.
+schema metadata, the exact deployed image identities, and both the Memory volume
+and isolated Memory database when the `brain` profile is enabled. Store the
+passphrase separately in your secret manager; the backup cannot be restored
+without it.
 
 Use `roomote backup --include-redis` when queued work, BullMQ schedules,
 sessions, and other transient Redis state must survive. Backups briefly stop
@@ -170,8 +174,8 @@ After you upgrade, Roomote can surface the new release in the web app:
 
 ## Requirements
 
-- A ChatGPT subscription or inference provider API key (OpenRouter, Anthropic,
-  OpenAI, and others).
+- Managed Roomote inference from an eligible host, or a ChatGPT subscription or
+  inference provider API key (OpenRouter, Anthropic, OpenAI, and others).
 - An account with a supported source-control provider. For GitHub, the setup
   wizard creates the GitHub App; other providers use their documented OAuth or
   credential flow.

@@ -1,6 +1,6 @@
 {
   "name": "roomote",
-  "version": "0.44.0",
+  "version": "0.45.0",
   "license": "FCL-1.0-ALv2",
   "packageManager": "pnpm@10.29.3",
   "engines": {