What changed, in plain language.

This project follows semantic versioning. Each completed release uses one version across package.json, package-lock.json, this file, the Git tag, and npm. Local checks validate the repository metadata; the publishing guide checks the registry separately after publication.

Unreleased

Fixed

• Local Claude install probes now call the companion at 127.0.0.1:14502 with Host relmio-claude:14502. A loopback Host was rejected after Compose reported the container healthy, so the one-time n8n key was never returned.
• A failed local or VPS Claude install after it starts writing removes only that attempt: the companion container, its managed files, and, on a fresh install, the login volume it just created. A failed replace keeps the login volume. The original error is still reported. n8n is not touched.
• A labeled relmio-claude_claude-account volume left behind with no companion folder or container is an absent install that keeps that login. Any other leftover volume is still refused.
• Claude management stays available on a VPS without running n8n, and a failed post-install SSH identity check no longer loses the one-time key.
• Local Claude rollback checks resource ownership before stopping anything. Removal works after image cleanup, stopped companions can renew approval, and strict file-creation permissions no longer block the image's app files.
• Wizard-managed Claude bundles require renewal in the wizard. Approval comparisons ignore JSON key order, VPS checks reject internal networks, and failed native sign-in checks remain errors.

0.24.0 - 2026-11-05

Relmio 0.24.0 adds Claude to the browser wizard (experimental), so you can set up the private Claude companion for n8n without copying a bundle by hand. Claude sign-in stays in your own terminal, and Relmio never handles the paste code or your Claude credentials. It is not affiliated with or endorsed by Anthropic, and Anthropic has not approved it.

Added

• Claude on my server (/claude-vps) for a VPS and Claude for n8n (and Add Claude on a Relmio-created n8n) for local Docker. Each shows the pinned Claude Code version and SHA256, asks four review questions that start unchecked, shows the exact plan and waits for a separate final confirmation before it writes anything. It then shows a one-time n8n key once. The wizard prints the sign-in command for your own terminal: ssh -t root@<host> sh /docker/n8n-openai-oauth/claude/login.sh on a VPS, and the new relmio claude companion login locally (an interactive terminal on macOS or Linux).
• Check sign-in, Check models, Renew approval (7 days minus 1 hour), Sign out and removal for both flows. On a VPS, Replace companion and key takes over a companion made with relmio claude companion prepare and keeps its Claude login. The local wizard refuses a manual companion and changes nothing.
• Hosting options lists Claude as managed on Linux VM presets and local Docker, and as unavailable on container platforms, endpoint products and SDK sandboxes. The local dashboard has a ninth row, Claude for n8n. For an n8n that Relmio created, the local wizard tries to add a Relmio Claude credential.
• A wizard source check and Claude coverage in the Claude, VPS, local, hosting, FAQ, troubleshooting and security guides.

Changed

• The Docker companion bundle now comes from one shared builder used by relmio claude companion prepare, the VPS wizard and the local wizard. The command's behavior is unchanged.

• The Claude guides, README and npm README no longer describe Claude as CLI-only. Native n8n and the advanced Docker path still use relmio claude.

• Owner-run live QA on 2026-11-05 passed local wizard install, sign-in, model check, prompt, removal and reinstall. Hostinger VPS wizard passed SSH, n8n and network discovery only; VPS install, replace, renew, sign-out and removal were not run live.

Fixed

• Local companion probes now send Host: relmio-claude:14502; rejected loopback Host headers had prevented a successful install from returning its one-time key. Failed installs now roll back their own managed files and companion resources.
• A labeled relmio-claude_claude-account login volume with no companion folder or container is reused by a fresh install and kept on failure; other unrecognized leftover volumes remain refused.

Security

• VPS Claude actions need a verified direct-root session. The passwordless sudo -n (model only) session is refused for them, like the other credential-bearing VPS routes.
• The Claude login stays in the companion's Docker volume, which root, Docker administrators and backups can read; Relmio never reads it. Check sign-in runs Claude Code's auth status --json, and the raw output reaches Relmio's local server (over SSH for a VPS). The server forwards only loggedIn, authMethod and subscriptionType to the browser and does not store or log the rest. Relmio keeps only the SHA256 of the one-time key.
• Whether Anthropic permits this use stays unresolved, and the wizard's Claude setup path relates to the third-party login restriction in Anthropic's Agent SDK quickstart even though sign-in stays in the official CLI.

0.23.0 - 2026-11-04

Relmio 0.23.0 adds an experimental, owner-only way to use Claude models from your own Claude Pro/Max subscription in n8n text workflows, through the official Claude Code CLI that you install and log in to yourself. It is not affiliated with or endorsed by Anthropic, and Anthropic has not approved it.

Added

• An experimental relmio claude CLI (macOS and Linux) for local text clients and n8n Basic LLM Chains: private initialization, interactive qualification for local and companion setups, foreground serving and companion-bundle preparation. It runs the unmodified official Claude Code 2.1.317 with its own native login, an independent Relmio bearer, approved exact models (Haiku, Sonnet and Opus 5.5 by default), expiring owner qualification and one active request. It serves a text-only OpenAI-compatible /v1/chat/completions endpoint for n8n's OpenAI Chat Model node. This is not a shared service or Anthropic approval; plan limits and account extra-usage settings still apply.
• A separate private Docker companion recipe for existing n8n networks, with checksum-verified Linux amd64/arm64 binaries and no published inference port. It does not edit or restart n8n. Tools, AI Agent tool loops, the n8n AI Assistant, the Responses API, images and Fable are not supported.
• A hosted Claude subscription (experimental) guide with disclaimers, limitations, tested and untested setups, and the Anthropic API-key alternative for full agent workflows.

Changed

• GitHub and npm guides explain the experimental Claude boundary, required n8n settings and API-key alternative for full agent workflows. The full guide keeps dated owner-run evidence (native macOS with n8n, a Hostinger Docker companion, and a local Colima arm64 Docker companion with n8n and an ngrok webhook) separate from historical research and untested platforms; private host details are omitted. Those records do not establish acceptance on other hosts.

Security

• The hosted website moves from Next.js 16.3.8 to 16.4.2. That fixes GHSA-2464-fhr4-w9xw and GHSA-673q-mjm6-r5xr (critical, unauthenticated remote code execution) and GHSA-p3m5-59mp-jv5r (high, denial of service). The npm package and installers do not include Next.js.
• The development-only zod dependency moves to 4.6.7 for GHSA-mrhx-c7xc-m3w5. It is not shipped in the npm package.

0.22.2 - 2026-10-11

Relmio 0.22.2 makes Help usable on small and short windows and keeps the setup guide off the dashboard's buttons. It also corrects how the docs and progress messages describe ChatGPT sign-in.

Fixed

• Help now opens on short or narrow windows, such as 1280 x 320 or a phone held sideways. When the window is too short for the whole card, the card fits below the top bar and its text scrolls, with its buttons in view.
• On the local dashboard, the setup guide card and the Help chip no longer cover the selected connection's buttons, such as Review removal.
• Sign-in documentation now says ChatGPT opens in a new tab or window in the browser running Relmio when allowed, and otherwise in the default browser. The progress messages say "sign-in page" instead of "window", and a sign-in that cannot start in that tab says "ChatGPT sign-in could not start. Return to Relmio and try again."
• A published source-check record no longer includes local folder paths.

0.22.1 - 2026-10-11

Relmio 0.22.1 fixes Windows sign-in and owned-stack recovery, and keeps Help available without turning the walkthrough back on. Existing consent and installation confirmations still apply.

Changed

• Contextual mascot Help stays available after skipping the walkthrough. It explains the actual prerequisite for Check the server, offers the current feature's tips and versioned bundled FAQ, and highlights controls without activating consent or installation actions.

• A new ChatGPT registration enables eligible text plan use after its first verified plan grant and shows a one-time Got it notice. Existing paused accounts stay paused; n8n background activity still needs separate approval.

Fixed

• Local image sign-in codes are larger and selectable, with a user-clicked Copy code button and manual-copy recovery when clipboard access is unavailable. Unchanged status polls preserve selection; expired or replaced codes clear stale feedback.
• The PowerShell installer works again on Windows computers without Node.js. It now puts its temporary Node.js runtime on PATH while Relmio starts. Before this fix, the wizard stopped with '"node"' is not recognized. The Command Prompt installer already did this.
• ChatGPT sign-in opens in the browser running Relmio when a new tab is allowed, with a system-browser fallback and safe cancellation of an unused attempt. The first-plan notice keeps keyboard focus inside while open.
• Resuming an exact owned local n8n stack works without Compose's newer start --wait flags and verifies health before reporting success.
• Windows QA fixtures handle denied file symlinks explicitly and keep the CMD lock test deterministic. Generated docs retain the first line after the title.
• Escape dismisses Help during a busy operation without unlocking the installation controls. New-n8n Help correctly describes the private default and the optional public ngrok link.

This patch does not establish fresh live VPS-update, image-generation or SuperGrok-workflow acceptance. Those checks remain separate from the Windows and Help evidence. The known braces development-tool advisory GHSA-vfj7-8cjw-p6xm remains unpatched; it is not a production-dependency audit pass for development tools.

0.22.0 - 2026-10-09

Relmio 0.22.0 turns Set up new n8n into a private, one-click n8n that the ChatGPT plan, local model, SuperGrok and Assistant add-ons can join without manual wiring. It also fixes ChatGPT sign-in on Windows.

Added

• The Ready step of Set up new n8n now has Open n8n and buttons to Add ChatGPT plan, Add local model, Add SuperGrok and Add Assistant tools. Each opens with this n8n and its network already selected; Add connection preselects them too. The Assistant tools button is hidden when the stack already includes Code Sandbox.
• For an n8n created by Relmio, an approved ChatGPT plan, local model or SuperGrok install adds its credential to n8n as Relmio ChatGPT plan, Relmio local model or Relmio SuperGrok. OpenAI tokens stay in the sidecar. If Relmio cannot confirm the credential, Ready asks you to check n8n first and shows the key under Manual credential setup. Other n8n installations still use manual entry.

Changed

• New local n8n stacks are private to this computer by default. A public URL through ngrok is now an optional setting, off by default.
• New stacks use n8n 2.42.5. Stack services, and newly installed local models, start again when Docker restarts, like the ChatGPT plan, SuperGrok and Assistant add-ons. Existing local model installs keep working and keep their current setting.
• Removing a stack is refused while add-ons are still attached. Relmio names them so you can remove them first.
• The stack's network is labelled "recommended for this n8n" instead of saying it contains ngrok when it does not.

Fixed

• On Windows, Continue with ChatGPT opens the sign-in page in the default browser again, on this computer and for a VPS. Relmio handed the sign-in link to Explorer, which does not open web links that contain ? and &, so the page stayed on Waiting for ChatGPT until the attempt expired (#115). If the browser cannot be started, Relmio now says so right away instead of waiting.
• Set up new n8n with Code Sandbox or Code Sandbox + SearXNG no longer fails during setup. The certificate step ran the wrong program, and the install check rejected the sandbox runner's declared ports even though none were published on the host.
• The dashboard can read a stack whose local address uses localhost.

0.21.0 - 2026-10-08

Relmio 0.21.0 lets you turn on image generation while you install the ChatGPT plan sidecar, on a VPS or on this computer, so n8n's image model list is ready right after setup.

Changed

• You can turn on image generation while you install the ChatGPT plan sidecar, on a VPS or on this computer. The review step has an optional, closed Image generation (optional) section with the same warnings as the image panel and an unticked box. When you tick it, Relmio starts the Codex sign-in right after a verified install, and the Ready step shows the image status below the one-time Relmio key, with the code one click away. On a VPS, Relmio keeps the SSH connection open only until you finish or cancel. An unticked box changes nothing, and a failed image sign-in never fails the install. Before, images could only be turned on afterwards, so n8n's image Model list stayed empty after a fresh install. Cancel image sign-in in the VPS image panel now also disconnects from the server.

Fixed

• At 320 pixels wide, opening Image generation on the local dashboard no longer makes the page scroll sideways.

0.20.0 - 2026-10-08

Relmio 0.20.0 adds image generation to the local ChatGPT plan sidecar and fixes installing that sidecar for an n8n on this computer, which always failed in 0.19.0 and 0.19.1.

Added

• The local ChatGPT plan sidecar can now generate and edit images through the same opt-in add-on as a VPS sidecar. Open Image generation in the installed sidecar's view on the local dashboard, choose Sign in for images and enter the code at https://auth.openai.com/codex/device. The Codex sign-in is stored in the sidecar's Docker volume on this computer, where administrators and Docker users can read it. Sign out and revoke and Remove bridge sign out of images first and report the revocation result; removal keeps the volume if that step does not finish. A local sidecar built by an older Relmio version needs a replacement or a fresh setup first. The images_off message now names both places; that sidecar file change makes an installed VPS sidecar report a runtime update. The Configure n8n heading for images is now Generate and edit images.

Changed

• The website's docs now include Configure n8n, with the sidecar's request routes, models, tools, image generation and their limits. The FAQ answers whether n8n can generate images and lists the image limits, and the VPS guide's summary mentions image generation.

Fixed

• Installing the ChatGPT plan sidecar for an n8n on this computer works again. In 0.19.0 and 0.19.1 it always failed, first with "Local SIWC private volume initialization failed." and then, past that step, with "Local SIWC installation operation failed." Docker Compose replaced the setup script's $owner with an empty value, and the image copied package.json without making it readable for the user the sidecar runs as. Both failures happen before the sign-in is handed over, so it stays on this computer. Found in local QA on Docker Desktop with Compose 5.5.1.

0.19.1 - 2026-10-07

Relmio 0.19.1 fixes the VPS sidecar update route and adds a ChatGPT sign-in refresh for an installed VPS sidecar. It also fixes n8n's OpenAI node and its AI Agent with Use Responses API off, keeps the wizard's star count, and adds change-aware CI.

Changed

• Release QA uses one n8n workflow, "Relmio release acceptance", on every self-hosted n8n: the Hostinger VPS, a local Docker n8n or another setup. Each n8n keeps one copy and every QA run reuses it. Its Settings node now holds the sidecar Base URL and the workflow version, so the HTTP Request nodes work where n8n reaches the sidecar at another address.
• The 0.16.0 notes and the GPT Image 2.5 experiment record now say that choosing Flare or Sunburst in Relmio 0.16.0 through 0.18.6 never selected those models. The Codex image route ignores the model ID, so every image came from Codex's single image model, which Codex names gpt-image-2.
• Hosting options and Self-hosted model on my server now start with Hostinger KVM VPS selected, the VPS that Relmio is tested on. Pick another provider to change it.
• CI now runs only the checks a change can affect, adds macOS, Linux and an n8n node compatibility job, and reports one required CI gate check. A change to the website or docs no longer reruns the Windows, macOS and Linux wizard suites; pushes to main, release tags and hotfixes still run everything. The website's dependency audit blocks only on production advisories and lists the rest in the job summary.
• The n8n node compatibility job runs n8n 2.40.7 in Docker against the real sidecar and a mock OpenAI upstream, and checks the HTTP Request node, AI Agent (Responses API on and off), Basic LLM Chain, the OpenAI node and the AI Assistant's AI SDK calls. No ChatGPT account is used.

Fixed

• The local wizard's GitHub chip no longer shows ? for the star count after GitHub's limit of 60 unsigned requests an hour is used up. The wizard keeps the count for 15 minutes, shares one request between pages that load together, asks Relmio's website when GitHub refuses, and keeps the last count when both fail. The chip still shows the Relmio version you are running.

• n8n's OpenAI node (Message a Model) no longer fails with "Bad request - please check your parameters" through the ChatGPT plan sidecar. The node asks for store:true unless its Store option is off. The sidecar now accepts that and still sends store:false to OpenAI, as it does for every request; a later request that names a stored response is still refused. Found in the 0.19.1 release check on the VPS. To get this fix on a VPS sidecar, run Review sidecar update.

• The ChatGPT plan sidecar's /v1/chat/completions route no longer refuses n8n's AI Agent on its second turn when the OpenAI Chat Model has Use Responses API off. n8n replays the tool-call message with an empty content list, which the sidecar now treats like no content. To get this fix on a VPS sidecar, run Review sidecar update.

• On Windows, a ChatGPT plan request no longer fails at once with "SIWC session lock is unavailable" when the Relmio process holding the sign-in lock exits while another process is checking it. Windows PowerShell could find the exiting process and then fail to read its start time. Relmio now treats that process as gone once Windows also confirms the process ID no longer exists, as it already did for a process PowerShell could not find.

• relmio stop no longer fails now and then with "Relmio refuses changed local dashboard control state" or "could not safely read its local dashboard control state" after the dashboard has accepted the stop. While stop waits, the dashboard moves its control files away one by one, and a check could list a file just before it moved. Stop now checks again. A control file that stays changed is still refused.

• On a VPS, the main button on Choose your n8n now follows what is installed on the selected n8n. A check of the installed sidecar runs when the step opens. It does not change the sidecar, but a running sidecar may renew its ChatGPT sign-in and ask OpenAI for the account's model list. Where Relmio's ChatGPT plan sidecar already runs, the button reads Review sidecar update: it opens Manage the installed ChatGPT session, checks the installed account again if that check is more than four minutes old, and runs the update review there. It reads Finish sidecar update after an interrupted update, Open recovery for an interrupted install or migration, and Review replacement once the installed sign-in is signed out. Before, choosing Manage ChatGPT plan sidecar changed the button to Review bridge update, which started a new install review that the server refused with vps_sidecar_owned. If the server still refuses a new install, the wizard reads the owner and takes the same route. The update keeps its own review and approval, and n8n is not restarted.

• Manage ChatGPT plan sidecar now opens Manage the installed ChatGPT session and checks the installed account.

• When OpenAI no longer accepts the VPS sidecar's sign-in, or Relmio cannot confirm its last renewal, Check installed account says Needs a fresh sign-in. instead of Signed out., and the main button reads Refresh ChatGPT sign-in. The panel shows the new Refresh ChatGPT sign-in button only in that case. It starts a fresh ChatGPT sign-in on this computer and hands it to the sidecar through the existing reviewed replacement: sign out the old sign-in, reconnect, then review and confirm the replacement, whose button reads Replace the sidecar. The wizard accepts only the registration from that new sign-in and refuses other accounts saved on this computer. It also refuses a new sign-in with a different email, and offers no refresh when the installed account has no email to check against. It cannot tell two workspaces with the same email apart. ChatGPT treats the new sign-in as a new connection, so the old connection's usage settings do not carry over. n8n needs the new Relmio key afterward. No new kind of server write was added.

• The setup guide explains vps_sidecar_owned, vps_sidecar_updating, vps_sidecar_staged and vps_sidecar_partial and points at the panel that resolves each one. It also has tips for the sign-in refresh.

0.19.0 - 2026-10-07

Relmio 0.19.0 adds a skippable setup guide with Relmio the mascot, Plan and usage counts, automatic model discovery and sidecar updates from the wizard, and moves ChatGPT plan sign-in to OpenAI's documented Sign in with ChatGPT flow.

Added

• Every wizard page has an optional setup guide. Relmio, the mascot, points at the box or button you need next, says what it is for in plain words and, where it helps, shows an example value. The guide never types into a field. Each quest you finish, such as Check your server, earns a badge. The first visit asks whether to start the guide. Hide folds it into a small Guide button, Skip guide turns it off, and the Setup guide button in the top bar turns it on or off at any time. When a step fails, the guide explains the error, lists the next steps and points at the control that fixes it. With the guide off, a "Need help with this error?" button opens the same help. Only the on or off choice is saved: ui-preferences.json in the SIWC storage root (N8N_OPENAI_OAUTH_HOME or ~/.n8n-openai-oauth), written through the wizard's /api/ui/preferences route.
• Plan and usage shows the requests sent through the ChatGPT plan sidecar in the last 30 days: requests and tokens in total, active days, the peak day, how requests ended and the busiest models. It also shows the account, whether plan use is on, how many models OpenAI lists and how many a completed request verified, and the last plan-usage error with its next step. On a VPS it is under Manage the installed ChatGPT session after Check installed account, and Refresh usage reads the counts over SSH, up to 10 times in 15 minutes. On this computer, the dashboard has a Plan and usage section and the installed sidecar's view has a Plan and usage disclosure. These counts are not your plan's usage. Relmio shows no plan percent, reset time or credits; they stay on ChatGPT's Usage page, which Manage usage opens.
• For Plan and usage, the sidecar now counts each text request it sends to OpenAI, per UTC day and model: how the request ended and, for completed responses, the input, cached, output, reasoning and total tokens. It also keeps the last plan-usage error until a completed response clears it. The counts are stored in activity/<registration ID>.json next to the model-check record, without prompts, outputs, request IDs, IP addresses or headers. Each write keeps the 31 most recent UTC days. Nothing else deletes the file, so after the last request the counts stay until the sidecar's storage is removed. A VPS sidecar starts counting after Review sidecar update.
• The ChatGPT plan sidecar's /v1/chat/completions route now accepts function tools, assistant tool calls and tool results. Function tools go to OpenAI in one additional_tools item, and n8n still runs the tools. Limits are 32 tool calls, 128 KiB of arguments per call and 2 MiB of streamed arguments. A named tool_choice, tool namespaces and custom tools in streamed requests are rejected. A two-turn LangChain tool test passed live on one ChatGPT account on 2026-10-05, streaming over both Chat Completions and Responses.
• An interrupted local or VPS ChatGPT install shows as staged and can be resumed after a review that names the selected account, without deleting data or starting a second refresh writer. Turning plan use on or off, or a token refresh, before resuming does not block it.
• If the destination accepted a session but the acknowledgment was lost, a reviewed reconcile finishes the handoff from the destination's receipt. Without a receipt the sender stays frozen and needs a fresh sign-in.
• An installed VPS ChatGPT sidecar can now be updated from Manage the installed ChatGPT session with Review sidecar update. Relmio rebuilds the image from the current Relmio version and recreates only the sidecar container. The ChatGPT sign-in and the one-time Relmio key stay the same, n8n is not stopped or restarted, and the sidecar is briefly unavailable while it restarts. Plan use must be on. An interrupted update shows as updating; review it again to finish it. The staged-install resume refuses it.
• An installed VPS ChatGPT sidecar can now generate and edit images through an opt-in add-on. Turn it on in Manage the installed ChatGPT session with Sign in for images, which starts a separate Codex device sign-in at https://auth.openai.com/codex/device. In n8n, use the OpenAI node's Image actions with the same Base URL and Relmio key, and pick gpt-image-2 from the list or enter it as the ID. Each request returns one image; masks and URL responses are not supported. A low-quality generation passed live on one Pro account on 2026-10-06. The add-on signs in to Codex the way Hermes Agent does. OpenAI does not document this route for other apps, so it can stop working without notice, and images count against the plan's Codex limits. The Codex refresh token is stored on the VPS. The ChatGPT plan session is never used for images. Signing out of the ChatGPT session also signs out of images, as a best effort. A sidecar built before this release needs Review sidecar update first. The local sidecar has no image sign-in.
• The ChatGPT sidecar now finds new text models without a Relmio release. It asks OpenAI's catalog as the newest stable Codex release on npm, never below the 0.160.0 pin, and checks npm about every 12 hours. That request sends no credentials; registry.npmjs.org is a new recipient and sees the sidecar host's IP address. The catalog is cached for 5 minutes. A model that OpenAI rejects with a model error is hidden from n8n for a day. A VPS sidecar built before this release needs Review sidecar update first.
• On a VPS, Check installed account now shows a Models group. Each model says In n8n or Not in n8n and Ready, Not working or Not checked yet, and has a Copy ID button for n8n's AI Assistant.
• VPS installs can turn on model checks. They are off by default and need confirmation. The sidecar sends one short test request ('Reply with OK') to up to 12 models right away, then to each new model and once a day to a model that failed. Once any model has answered, n8n lists only models that answered. Each test uses a little of your plan. Consent is recorded; turning checks off stops further tests, and Sign out and revoke turns them off first.
• The ChatGPT plan sidecar now fills in earlier output that a client refers to by ID, as n8n's AI Assistant does in tool steps and follow-up messages. From each Responses request that completes, it keeps reasoning items (encrypted content and summary) and assistant text in memory, and replaces each later item_reference with a copy. For this, Responses requests that set reasoning now ask OpenAI for reasoning.encrypted_content. The memory holds at most 4,096 items and 32 MiB, keeps an item for up to 6 hours and is never written to disk. A restart or update empties it; references it no longer knows are dropped, so an open Assistant conversation loses that earlier context. The sidecar adds no logs, so the live Assistant test under Changed doesn't show whether this memory was used.
• The sidecar's /v1/chat/completions route accepts reasoning_effort and sends it to OpenAI as reasoning.effort.

Changed

• The wizard's start cards now say who each route is for and what you need, for example "You need a VPS that runs n8n in Docker, a root login to it and a ChatGPT plan." The step list is now a compact quest track: one row of steps at every width, a check on each finished step and, while the guide is on, Relmio standing in the current step. With the guide on, the Ready screen lists the badges earned and the next steps. The steps and the setup flow are unchanged.

• The dark theme's accent is now pastel green instead of pastel yellow, in the same hue as the light theme's greens. Primary buttons, links, focus rings, the current step and selected items use it in the wizard and on the website. Terminal prompts and the home page's night band, which are dark in both themes, use it too; the night band's moon and stars stay yellow. Success messages keep their cooler green, and contrast still meets WCAG AA.

• The website's Doorway illustration is unframed again, spanning the page against its background instead of sitting inside a rounded, layered card. The current colors, animation and Pause control are preserved.

• The OpenAI Chat Model node and Chat Hub no longer list gpt-image-2. The OpenAI node's text picker still does, because it shares one request with the image picker, and the image picker lists it, so Generate an Image and Edit Image can pick it from the list. The wizard's HTTP recipe no longer offers image IDs as chat models.

• Images stay gpt-image-2 only. In a 2026-10-06 test the Codex image route returned the same token count, size and C2PA provenance for gpt-image-2, gpt-image-2.5-flare and a made-up ID, so Flare and Sunburst need a separate OpenAI credential with your own Platform API key.

• The AI Assistant docs now explain how to point its Self-hosted or OpenAI-compatible endpoint at the sidecar with a pasted model ID, instead of saying the sidecar is not offered. On 2026-10-07, after a sidecar update, n8n 2.40.7's Assistant used the sidecar on one Pro account with gpt-6-astra. The model check passed, "Reply with OK" returned "OK", and "Which workflows do I have?" ran a tool step and listed the instance's workflows in about 35 seconds. That is not a guarantee for other accounts, models or longer conversations, or after a sidecar restart. Whether this use fits OpenAI's Sign in with ChatGPT terms is open. Only do it if nobody else uses that n8n.

• One sidecar request can now list up to 128 tools, up from 32, which leaves n8n's AI Assistant room for its own tools plus those from MCP servers. Tool calls stay at 32 per Chat Completions message or response.

• Document the local SIWC account, permission, and model-access flow, including protected per-registration credentials, private n8n/VPS transfer boundaries, current unsupported capabilities, recovery, and unresolved provider limits.

• A refresh that returns HTTP 503 temporarily_unavailable now keeps the session from before the refresh instead of freezing it. This assumes OpenAI did not rotate the token, which OpenAI does not document.

• If ownership moves but a later finishing step fails, the result still shows the one-time key once, with a warning not to use it until the issue is resolved.

Fixed

• Sign-in no longer fails after you allow access. OpenAI sends the ID token's audience as a one-element list, which Relmio wrongly rejected. Found in a live sign-in test.
• Non-streaming requests and Chat tool streams no longer fail on live OpenAI responses. OpenAI's final response.completed event can have an empty output, so the gateway now rebuilds it from the response.output_item.done events. Chat Completions clients get only final-answer text, not intermediate commentary. The Responses route passes events and phase through unchanged.
• The sidecar no longer refuses n8n's AI Assistant model check with "This Responses parameter is unavailable with ChatGPT plan usage." The check sends max_output_tokens: 16. The sidecar now drops output-token caps instead of refusing them: max_output_tokens on Responses, and max_completion_tokens and max_tokens on Chat Completions. SIWC lists max_output_tokens as unsupported, so no cap can be honored, and the 16-token cap is not enforced. Other unsupported fields are still refused. On one Pro account on 2026-10-07, the check failed this way before a sidecar update and passed after it. A VPS sidecar needs Review sidecar update to get this fix and the related item memory, tool and error changes.
• Failures in a streamed /v1/responses request now reach AI SDK clients, such as n8n's AI Assistant, as errors. The sidecar's own stream errors use OpenAI's error event format, and rewritten response.failed and response.incomplete events keep their sequence_number and any incomplete reason. Before, the AI SDK reported a cut-off stream as a type-validation error and could end a failed response as an empty answer.
• The model list now includes the account's current models. On the tested account that added GPT-6.1-Sol, GPT-6-Sol and GPT-6-Luna. OpenAI filters the catalog by an undocumented client_version parameter; the sidecar now follows the newest Codex release (see Added). The Codex runtime pin moved from 0.147.0 to 0.160.0.
• If SSH connects but the read-only Docker or n8n check fails, the VPS wizard keeps the verified connection and offers Retry discovery without asking for the password again. Changing the server details still needs a fresh identity check.
• Reviewing a ChatGPT sidecar install or bridge update no longer fails with "The selected account or model catalog changed" when the account's access token refreshes while the model list loads. Switching to another account during the review is still refused, and an invalid model list now has its own message.
• VPS installs and resumes no longer stop after the image build with "The VPS sidecar execution configuration changed". Docker Compose reports the sidecar's command and entrypoint as null, which the ownership check treated as a change. A real override is still rejected. Checked against real Compose output, a Docker 29.2.1 container test and a live VPS resume.
• Resuming an interrupted VPS install now rebuilds the sidecar image unless the recorded image was built from the same runtime files. Before, a resume uploaded the current files but could keep running an older image, for example one without the empty-output fix above, so answers came back empty. A resume after the session moved, or of an account replacement, keeps its image; run Review sidecar update once it completes.
• Sign-in no longer fails when OpenAI's callback includes its documented optional scope. An iss value, when present, must match the issuer.
• Declining in ChatGPT now shows a declined message instead of a generic verification failure. Declining plan use keeps the existing registration.
• Function tools without parameters or strict now send both as null.
• The chat route skips reasoning output items, so Chat Completions clients never receive them.
• A callback with the wrong state no longer cancels the sign-in in progress.
• Session locks now record the holder's process namespace and boot. A lock from an earlier boot is reclaimed at once, and one held from another container on the same boot after a 10-minute lease. Holders stop after a 2-minute deadline, and waiters get 503 siwc_lock_unavailable after 2.5 minutes.
• Session lock records are published atomically, so a crash cannot leave a half-written lock, and release waits through short contention. A refresh that has started is no longer cancelled when the caller disconnects, and the rotated token is saved. Keep the SIWC store on a local disk used by one kernel, not on a synced folder or network share. Windows cannot flush a folder to disk, so on Windows the lock skips that step and relies on NTFS.
• On Windows, owner-only file checks now run in one PowerShell process that stays open for the whole Relmio run, so a check no longer starts a new PowerShell. Relmio checks its own process identity once per run, skips rechecks of files it locked down itself until their change time or contents change, and always checks files it did not create.
• On Windows, an existing owner-only SIWC folder is checked but not rewritten, so one sign-in no longer disturbs another process's files. Session locks retry briefly when antivirus or another process holds a lock file for a moment.
• The Codex App Server's live model check now runs inside the running container instead of a one-off helper container.
• Status and sign-out for a completed VPS install keep working after n8n is recreated or the SSH login method changes. Completed installs bind only the SSH host identity and Docker network ID; interrupted installs still bind the full reviewed target.
• The ChatGPT plan notice opens only once its view is visible. Escape closes it without confirming, and Review plan notice opens it again.
• At a usage limit, Manage usage is a readable primary button in both themes. The account card hides Using ChatGPT plan and Pause plan use, and the wizard keeps Check the server off.
• ChatGPT account controls stay disabled in preview after a background check ends. After Sign out, focus moves to the account status.
• The ChatGPT account card fits one screen at laptop sizes in the wizard and on This computer. It no longer repeats the selected account, and the plan badge sits beside the status. Recovery lists show "owned here" and "transfer pending" instead of internal state names.
• When finalization fails after an install, or the runtime cannot be verified, the Ready step says to save the one-time key and not use it yet. It no longer shows Using ChatGPT plan or n8n setup wording, and the warning sits in a callout. The model select, the server's model picker and its HTTP recipe copy stay disabled.
• The Ready step on This computer fits one screen at laptop sizes. Key safety notes sit beside the account card, the installed account's session controls are in a disclosure, and repeated key and URL guidance is gone. Opening the chat tester no longer scrolls the whole page.
• A chat tester error and its Manage usage button scroll into view before focus moves to the error, and again once Send returns. Stop response is now the same height as Send.
• The VPS review shows the SSH host key once, under Host key and build details, so the plan fits one screen. From 1024 to 1279 pixels, step 5 puts the base URL and key under their labels at full width.
• Checkbox and radio labels are at least 24 pixels tall.

Security

• VPS installs bind the reviewed n8n container and network IDs and check them again before writing and before the session transfer.
• VPS SIWC files are published atomically under /docker/n8n-openai-oauth, and unsafe parents or linked targets are rejected. Every SSH command has a finite deadline: 45 minutes by default, 30 minutes for a build, 2 minutes for handoff acceptance and 5 minutes for file publication. A timeout reports an unknown outcome and is not retried.
• Recovery never restarts an old writer automatically. Switching to a fresh account after a "not accepted" result is refused if the original receipt appears, and on a VPS while a one-off helper container is still present.
• The docs now state plainly that the Codex App Server target is high trust. Its relay forwards every client call except a short deny-list to an App Server running as the same user as the SIWC store, so connect only trusted local clients.

0.18.6 - 2026-10-04

Relmio 0.18.6 makes the selected theme and keyboard focus easier to distinguish.

Fixed

• The selected theme in the local wizard and website has a high-contrast inset ring in both themes, so it stays visible after focus moves away. In forced colors, selection and keyboard focus have separate indicators. Theme behavior and control sizes are unchanged.

0.18.5 - 2026-10-03

Relmio 0.18.5 changes the light theme from pastel yellow to pastel green.

Changed

• The light theme of the local wizard and the website is now pastel green instead of pastel yellow. Pages, panels, lines, selected items and the home page's day scene use the new greens, and success messages use a cooler green so they stay distinct from the page. Text contrast still meets WCAG AA. The dark theme is unchanged.

0.18.4 - 2026-10-03

Relmio 0.18.4 restores browser format checks on the /hosting plan details.

Fixed

• Hosting plan detail fields on /hosting check their format in the browser again. Most field patterns used characters that browsers reject, so the browser skipped the check and a malformed value only failed later with a general error. Accepted values are unchanged.

0.18.3 - 2026-10-03

Relmio 0.18.3 names SSH fields by their visible labels in setup errors, and records that the hosted chat on the website is turned off while Relmio applies to OpenAI for access.

Changed

• The hosted chat and its ChatGPT sign-in are turned off while Relmio applies to OpenAI for access. The home page's chat section now says so, links to the local wizard and offers Remove saved sign-in from this browser for people who signed in before. The top bar and footer no longer link to the chat, and /chat still redirects to the notice at /#chat. ChatGPT sign-in now works only in the local wizard.

Fixed

• SSH setup errors name the field by its visible label, for example "Server address is invalid." instead of "Hostname is invalid.", on the home server route, /assistant, /supergrok-vps and /local-model-vps.

Security

• /api/chat no longer reads credentials or forwards anything to OpenAI. Every request gets 410 Gone with Cache-Control: no-store. The site no longer ships the hosted sign-in packages, and its Content Security Policy no longer allows connections to auth.openai.com or the Firefox extension frame on localhost:1455. Removing the saved sign-in clears it from the browser only; it does not revoke access at OpenAI.

0.18.2 - 2026-10-03

Relmio 0.18.2 fits the remaining wizard steps on one screen at laptop and desktop sizes, including SSH sign-in, review, install and failure states.

Fixed

• On /local and /assistant, a long error or an opened rail disclosure no longer makes the whole window scroll. The rail notes scroll on their own and the step panel keeps its actions in view. The rejected-port error with "This computer only" open and every local install failure now fit at 1024 x 768, 1280 x 720 and 1366 x 768.
• The /local rail shows the ChatGPT credential copy only for ChatGPT for n8n, and /assistant shows the verified server once on the review step.
• Wizard steps that overflowed at 1024 px now fit on one screen: the home Assistant result and install-failure error, /assistant while installing, the /local review, the /supergrok-vps server check, the /local-model-vps removal review and long /hosting detail forms. Fit panels use slightly tighter padding from 1024 to 1279 px, the home error takes a full-width row under the step title there, and the home Assistant result says its settings note once.

0.18.1 - 2026-10-02

Relmio 0.18.1 returns the hosted site to its scrolling layout with a new footer, and fixes accessibility, keyboard and security-policy findings from a Front-End Checklist audit of the wizard and the site.

Added

• Add a footer to every page of the hosted site with product links, the creator's GitHub, X, LinkedIn, YouTube, Facebook and Ko-fi links, and the copyright and license line.
• The hosted home page describes Relmio and its creator to search engines with structured data.

Changed

• The hosted site scrolls as a normal page again. The home page returns to its hero, How it works, hosted chat and safety sections in the shared theme, and /chat redirects to the chat section at /#chat. The one-screen layout stays with the local wizard and dashboard.
• The hosted site now enforces a Content Security Policy with a fresh nonce on every page. Only the site's own scripts run. Network connections stay on the site or at auth.openai.com, with one frame allowed for Firefox's extension detection endpoint. HSTS now covers subdomains.
• The local dashboard's Connections view has one primary button, Add connection. Setup, resume and sign-in actions in rows and the detail card keep the 40 px size in the default style.

Fixed

• The hosted ChatGPT extension dialog takes focus, keeps keyboard navigation inside, closes on Escape and returns focus to sign-in. Its title and focus ring remain readable in dark theme, and its links announce a new tab.
• The hosted sign-in flow can check Firefox's extension through its exact loopback detection URL without relaxing the script nonce policy.
• The decorative footer wordmark no longer blocks clicks on the license link.
• Local management views keep one primary action and hide the Docker status when no Docker check runs.
• SuperGrok and Local model VPS errors sit directly after the footer actions on narrow screens and remain in the rail on desktop.
• The wizard's initial SSH status check returns a normal not-connected response instead of logging an HTTP error. Connection-required operations still reject requests until the VPS is connected.
• New-tab links in the hosted footer, extension installation callout, sign-in dialog, sign-in credit and local ngrok checklist show an external-link indicator and announce the new tab.
• Docs, install, changelog and guide descriptions are unique, complete summaries of 120 to 160 characters that end at a full sentence, and the FAQ opens with a short intro.
• The GitHub chip's version follows the newest stable npm release within minutes.
• Standalone links on the hosted site, including guide links, reference lists in the docs and the sign-in attribution, have at least a 24 px target. Quiet links keep their underline inside sentences.
• Primary and removal buttons on the local dashboard use the standard 40 px size.
• In Windows high contrast and other forced-color modes, the stepper shows the current step as a filled, ringed disc and finished steps as a check.
• Local dashboard rows show each service's full name, state and boundary instead of cutting text off. The selected connection card holds the longer description.
• A rejected port, or another rejected first-step field, on the local setup page is marked invalid and linked to the visible error until you correct it.
• On narrow screens, a blocking error on the local setup page sits right after the step's footer buttons instead of below the Docker status.
• Standalone wizard links, including Compare hosting routes, the server setup links, hosting sources and the local troubleshooting link, have a 24 px target.
• Docs code blocks, wide docs tables and the wizard's hosting comparison table can be scrolled with the keyboard and show the focus ring.
• The stars on the home page's night band stop twinkling within 5 seconds.
• Browsers that ask for /favicon.ico or an Apple touch icon get the Relmio logo instead of a missing page.

0.18.0 - 2026-10-01

Relmio 0.18.0 brings private local models for n8n, wider VPS support and hosting plans to the stable channel, with a redesigned wizard and website. Features that are not fully tested carry an Experimental label.

Added

• Install one CPU Ollama model from an allowlisted catalog beside an existing n8n container, on local Docker or a VPS. Relmio measures memory, CPU and disk, reviews the exact download and resources before you confirm, publishes no host port, and needs no provider sign-in or API key. Model-only checks ran on local Docker and a Hostinger VPS. Linux full-stack acceptance is tracked in issue #86, and a reported 120-second local inference timeout on cold model loads remains open.
• Sign in to a VPS with a local SSH agent and the image's real username. A verified passwordless sudo -n login can manage the model only; the OAuth bridge, AI Assistant and SuperGrok still need direct root.
• Compare hosting products and generate manual plans for provider profiles, model and search artifacts, Daytona handoffs and restricted relays. Plans do not contact providers, provision resources or change n8n.
• Add a maintainer harness for model acceptance on disposable Linux n8n hosts. No live Linux result is claimed yet.
• Publish per-page canonical URLs and social previews, plus robots.txt and a sitemap, for the hosted site, and return a branded page for unknown wizard and website addresses.

Changed

• The stable channel (relmio@latest and the hosted installers) now includes local models and hosting plans. Relmio is tested on Hostinger KVM VPS and local Docker. Other VPS hosts, hosting-plan platforms, the SuperGrok adapter, Codex App Server and Codex Chat Adapter are labelled Experimental.
• Set up a VPS in five visible steps: choose a route before ChatGPT sign-in, confirm the SSH fingerprint before entering credentials, and keep the required plan review in view. The credential-upload destination comes from the verified SSH identity; going back or changing the identity clears the reviewed plan and approval.
• Share one UI kit and top bar across the local wizard and website, with a pastel-yellow light theme and black dark theme. Fit wizard and dashboard views to one screen on laptop and desktop displays, use a smaller top-bar logo, move hosted chat to /chat, and clarify its disclosures.
• Clarify callback recovery, the system-browser Codex login boundary, Windows launcher path limits and API-key scope: managed OAuth flows do not configure upstream API keys, and API-key connections and hosting artifacts are configured separately. Record the 2026-09-27, 2026-09-29 and 2026-10-01 OpenAI source reviews.

Fixed

• Wizard: keep focused controls and view headings clear of the top bar and footer on phones, show each step's primary action on the first phone screen, keep long steps and the local model removal review within one screen, and stop progress cards covering warnings and footer actions.
• Hosted chat on narrow phones: starter prompts stay usable, and Jump to latest appears only when a conversation exists.
• Read progress and status updates aloud while an operation runs.
• Accessibility: copy buttons keep stable names, announce results politely and copy the text shown; rejected fields are linked to their errors; Tab keeps working while an operation runs; focus returns to the composer or transcript after a chat response or sign-out.
• Hosted chat: name the sign-in flow as the unofficial third-party Codex flow, describe who receives the prompt and tokens, label sign-out as local, and return every response with Cache-Control: no-store.
• Model readiness: accept Docker's canonical Buildx current record and enabled no-new-privileges forms without weakening those checks, keep the owned Ollama cache across restarts and retries, and verify the manifest digest and quantization before readiness.
• Track the matching Codex login attempt and require confirmed helper shutdown before another sign-in. A delayed SSH-status response no longer replaces the identity adopted for review. Managed Windows paths decode as strict UTF-8, and opening hosting options keeps the private session.
• Ignore forwarded hosts when building crawler and metadata URLs, and send Permissions-Policy with the other hosted security headers.

Security

• Keep model-only passwordless sudo separate from OAuth bridge, Assistant and SuperGrok VPS access. Temporary Buildx state stays inside reviewed operation locks; an unknown SSH or SFTP outcome keeps the lock and state for inspection.
• Require the VPS model ownership marker to stay root-owned, mode 0600 and single-link. Generated relay runtimes process request bodies and send the operator's upstream bearer to the fixed upstream; their logging and retention remain unknown.
• Credential contents and recipients, OAuth behavior and provider permissions are unchanged. The ChatGPT bridge remains unofficial, private and policy-uncertain; provider approval and account entitlement are not established.
• Update the hosted web framework to Next.js 16.3.8 and pin patched undici, brace-expansion and fast-uri releases to address newly reported advisories.

0.18.0-experimental.4 - 2026-09-29

Fixed

• Remove stale pre-publication wording from experimental installation docs. Wizard and runtime behavior are unchanged.

0.18.0-experimental.3 - 2026-09-29

Changed

• Put the VPS route chooser before ChatGPT sign-in and present setup in five visible steps, with SSH fingerprint confirmation before credentials and the required plan review kept in view.
• Identify the credential-upload destination from the currently verified SSH identity. Back navigation or an identity change clears the reviewed plan and approval; separate human confirmation remains required before remote writes.

Fixed

• Prevent a delayed startup SSH-status response from replacing the identity adopted for review or leaving stale consent text after a newer connection decision.

Security

• Credential contents and recipients, OAuth behavior, provider permissions, and bridge capabilities remain unchanged. The bridge remains unofficial, private, and policy-uncertain; provider approval and account entitlement have not been established.

0.18.0-experimental.2 - 2026-09-28

Fixed

• Accept a bounded canonical Docker Buildx current record with an empty Name without interpreting its Key. A named default still requires the exact reviewed Docker host, and fallback, shadowed-default, and other saved selector checks remain fail-closed.

• Accept Docker's bare, :true, and =true enabled no-new-privileges forms during local and VPS model readiness checks. Conflicting or malformed values still fail closed; this fixes false readiness failures without weakening the security requirement.

0.18.0-experimental.1 - 2026-09-27

Added

• Add provider-free local and dedicated VPS Ollama model setup for n8n Chat Models, with reviewed network attachment, measured Docker capacity, an allowlisted catalog, and bounded inference-readiness checks.
• Add local SSH-agent and editable image-qualified Linux usernames, plus a verified noninteractive sudo mode limited to model management.
• Add the authenticated hosting catalog and manual-plan generator for provider profiles, model/search artifacts, Daytona handoffs, and restricted relays. Plans do not contact providers, provision resources, modify n8n, or prove runtime compatibility.
• Add an interactive Linux model-acceptance harness for disposable n8n scenarios; live Linux/provider acceptance remains separate and unclaimed.

Changed

• Keep 0.17.5 as the stable npm latest and hosted-installer default. The experimental wizard is opt-in through relmio@experimental or its exact prerelease version.
• Clarify callback recovery, the system-browser Codex login boundary, Windows launcher path limits, and the distinction between planner generation and operator-deployed runtime behavior.
• Record the 2026-09-27 OpenAI source review. It distinguishes partner identity sign-in from the bridge's complete host-side Codex credential copy and records unresolved scopes, permissions, model entitlement, runtime data recipients, logs, and retention.
• Document SSH host-key, privilege, temporary Buildx state, and recovery requirements without enabling root/password SSH or changing n8n.
• Shorten the experimental server wizard to five steps, with optional setup guidance in expandable details and the required plan review kept in view.

Fixed

• Track the matching Codex login attempt and require confirmed helper shutdown before another sign-in; inconclusive callback-port probes are not treated as proof that the port is free.
• Preserve the owned Ollama cache across restarts and retries, verify catalog manifest digest and quantization before readiness, and retain bounded pull streaming, parsing, and inference deadlines.
• Decode managed Windows paths as strict UTF-8 before applying existing owner-only ACL checks.
• Preserve the private setup session when opening hosting options directly.
• Pin model-specific setup commands to this prerelease so they do not open the stable wizard, which does not include the new model setup.

Security

• Keep model-only passwordless sudo separate from OAuth bridge, Assistant, and SuperGrok VPS access. Constrain temporary Buildx state to reviewed operation locks and attest the default builder without changing the user's saved config.
• Require the VPS model ownership marker to remain root-owned, mode 0600, and single-link through pre/post-upload checks. Across OAuth/model/Grok VPS operations, unknown SSH/SFTP outcomes retain the operation lock and Buildx state for inspection; verified failures use normal cleanup.
• Document that generated relay runtimes process request bodies and send the separately configured upstream bearer to the fixed upstream. Cloudflare observability is enabled in generated artifacts; actual field-level logging and retention remain unknown.

New provider deployments, local-model inference, throughput, and provider entitlements were not live-tested for this prerelease. Offline planner/profile checks are not runtime certification.

0.17.5 - 2026-09-26

Added

• Add a dated official-source review confirming that the callback-port check does not alter Relmio's authentication, permissions, data handling, logging, or recipients.

Fixed

• Probe the default callback port on both IPv4 and IPv6 loopback before local and VPS OpenAI OAuth sign-in, and report that an existing listener could interfere. The page reported by the original sign-in failure did not establish which application displayed it or received a callback.

Changed

• Explain the OpenAI OAuth "Signed in to ChatGPT / This sign-in request expired" page, add Windows PowerShell commands that inspect port 1455 without stopping anything, and state that an MCP Transport closed error comes from the AI client's MCP server, not from Relmio.

0.17.4 - 2026-09-21

Added

• Add a dated official-source review confirming that the canonical-domain change does not alter Relmio's authentication, permissions, model request, data handling, logging, or recipients.

Changed

• Make relmio.jpfusin.tech the canonical website, documentation, package homepage, and hosted-installer origin.
• Permanently redirect requests from relmio.vercel.app to the matching path on the new canonical domain so existing links keep working.

0.17.3 - 2026-09-18

Changed

• Lead bare relmio, NPX, and hosted installer launches with ChatGPT on my server while keeping relmio local as the explicit computer-only route.

0.17.2 - 2026-09-18

Changed

• Open bare relmio, NPX, and hosted installer launches as a foreground browser wizard without creating persistent Relmio state. A clean first-run machine may have no .relmio directory and no local n8n stack; prerequisite guidance now remains inside the wizard.
• Run the complete release gate plus focused PowerShell, CMD, browser handoff, ACL, and Codex login checks on native Windows CI.

Fixed

• Keep a successfully dispatched Windows browser handoff alive even when explorer.exe later returns a nonzero status, preventing the temporary page from disappearing with ERR_FILE_NOT_FOUND.
• Retry native Command Prompt temporary-runtime cleanup when Windows briefly holds an executable open, and report an actionable nonzero failure if cleanup remains blocked.
• Give the hosted Git Bash launcher a real native child terminal and normalized Windows runtime path through its bundled winpty and cygpath tools, while replacing GUI-dependent Windows CI coverage with a headless native-console test.
• Use native Windows Node/npm process boundaries, owner-only temporary browser handoffs, and isolated official Codex login attempts with validated atomic credential promotion and actionable cancellation errors.

0.17.1 - 2026-09-14

Relmio 0.17.1 makes streamed replies calmer and clearer in both the hosted chat and the browser wizard's Test AI Chat console.

Changed

• Distinguish sending, connecting, waiting for first text, active streaming, completion, interruption, and failure without announcing every text chunk to assistive technology. Reduced-motion mode keeps equivalent static cues.
• Keep partial text primary and visible when a response is stopped or fails, with stable message sizing and explicit retry-ready terminal states.

Fixed

• Ignore empty stream deltas as non-visible output and keep pre-token progress monotonic, preventing blank replies or backward status changes.

The Codex Chat Adapter POST /chat protocol and external clients such as n8n are unchanged. Node.js 24+ and the hosted chat's exact gpt-5.6-luna request remain in place.

0.17.0 - 2026-09-14

Relmio 0.17.0 is a pre-1.0 compatibility milestone that moves supported setup paths to Node.js 24 and updates the hosted web chat to GPT-5.6 Luna.

Added

• Add a dated official-source review for the Node.js 24 runtime and hosted Luna request, including authentication, data-flow, policy, entitlement, and native-platform limits.

Changed

• Require Node.js 24 or newer across the browser wizard, hosted launchers, package metadata, CI, release validation, and current setup guidance. Portable launchers now select a checksum-verified Node.js 24 runtime; the native CMD launcher pins the reviewed official Node.js 24.21.0 archives.
• Use the exact gpt-5.6-luna model ID for the hosted web chat request and its visible model label. This does not change n8n, image, audio, console, provider, or local endpoint defaults.

Hosted Luna access still depends on the signed-in account and compatibility transport. This release does not establish broader model entitlement or Terms approval.

0.16.0 - 2026-09-12

Relmio 0.16.0 brings GPT Image 2.5 discovery and selection guidance to the browser installer for new and existing OpenAI OAuth bridges.

Added

• Discover the exact Flare and Sunburst image model IDs alongside existing models, including GPT Image 2. Both 2.5 variants were tested with n8n's native image generation and editing nodes on the current account. Correction, 2026-10-07: these requests went to the Codex image route, which ignores the model ID. A test on 2026-10-06 got identical results for gpt-image-2, Flare and a made-up ID, so choosing Flare or Sunburst in 0.16.0 through 0.18.6 produced Codex's own image model, not GPT Image 2.5 Flare or Sunburst. Relmio 0.19.0 offers only gpt-image-2 on this route.
• Show discovered image choices and copyable IDs after bridge installation or update, with instructions for n8n's image model selector.

Fixed

• Recognize the unchanged runtime from Relmio 0.15.0 during a local bridge update instead of rejecting it as file drift. Modified or unowned files still fail closed; updates retain the saved sign-in and leave n8n untouched.

Changed

• Simplify local and VPS setup with clearer visual choices and plain-language labels. Keep specialist local tools and technical explanations behind expandable details while retaining every setup option and approval step.
• Keep GPT-Live and Realtime models out of this bridge's discovery response and return explicit unsupported-operation guidance for their session routes. Audio remains a separate unsupported capability through this OAuth bridge.

Security

• Update the hosted web framework, image-processing dependency, and YAML parser to versions that address the newly reported Next.js, sharp, and js-yaml advisories.

Existing bridges need a reviewed runtime update to receive the new catalog. Exact output dimensions, all image options, and universal account access remain unverified. The bridge remains unofficial and policy-uncertain; this release adds no Platform API-key fallback or new credential permissions.

[0.15.0] - 2026-09-08

Relmio 0.15.0 adds browser-driven updates for existing OpenAI OAuth bridges and repairs n8n OpenAI node compatibility on local Docker and VPS deployments.

Added

• Update an existing local or VPS OpenAI OAuth bridge from the browser wizard. Local updates preserve the saved sign-in; VPS updates upload the current local sign-in. Both rebuild only the owned sidecar, retain its selected network, leave n8n running, and keep port 10531 private.

Changed

• Document all 16 n8n OpenAI actions and their current bridge limits. Audio, file management, stored conversations, moderation, video, background jobs, and stored responses return specific unsupported-operation guidance.
• Make the completion-page Responses API notice dismissible while preserving the required setting in the permanent instructions.

Fixed

• Allow n8n Message a Model requests with Background Mode off by removing the disabled background parameter before forwarding to the OAuth transport. Both sidecar installers bundle the same compatibility adapter.
• Explain failed VPS model checks and route rejected ChatGPT credentials back to fresh sign-in. Clear the previous plan and approval before another update.
• Separate the completion notice from the credential heading and report Docker network refresh failures without claiming the install or SSH session stopped.

Existing bridges need a runtime update from the wizard to receive these fixes. Relmio's ChatGPT connection remains unofficial and does not provide every OpenAI Platform API action.

0.14.0 - 2026-09-06

Relmio 0.14.0 adds experimental SuperGrok OAuth for local apps and existing local or VPS n8n deployments without requiring a ChatGPT sign-in.

Added

• Add SuperGrok Chat Completions for local apps and private n8n clients using a fresh official Grok device sign-in, isolated provider session, and separate local Relmio bearer. n8n executes its own tool calls.
• Keep the simple /chat interface through the same direct HTTP transport.
• Add a dedicated SuperGrok VPS wizard for verified SSH identity, read-only n8n discovery, reviewed installation, device sign-in, model checks, status, sign-out, cancellation, and owned-only removal. The companion joins one selected Docker network and publishes no host port.

Changed

• Use one guided browser flow for provider, local or VPS destination, discovery, review, sign-in, and completion. Provider changes invalidate stale plans and keep the OpenAI OAuth and SuperGrok instructions distinct.
• Redesign the public homepage with the original Relmio mascot, a full-width day/night landscape, timed destination labels, and accessible motion controls. Keep the install wizard focused on large controls and plain instructions.
• Limit provider setup to OAuth. Remove upstream API-key gateways, profiles, registration and selection routes, and the API-key n8n xAI sidecar.
• Show seven dashboard services and four provider-owned OAuth entries. Keep runtime health, provider readiness, and inventory freshness independent.
• Preserve the last observed state while inventory is stale and disable maintenance actions until a successful refresh.
• Document the provider-specific n8n setting prominently: the Relmio OpenAI OAuth recipe uses Responses API on, while SuperGrok requires it off for workflow model nodes and Chat Hub.
• Keep existing API-key gateways and credential data running but outside the 0.14.0 dashboard. Refuse to adopt an earlier SuperGrok development install that lacks the fresh-session marker.

Fixed

• Keep endpoint URLs readable on narrow dashboard layouts.
• Use supported Docker Compose run options for Grok login and version probes.
• Keep Windows identity and unsafe-path tests portable by forwarding filesystem metadata options and using a directory junction that does not require Developer Mode or administrator-created symbolic links.
• Document that Git Bash 2.38.1 needs per-process MSYS=enable_pcon for the portable launcher TTY check, with native PowerShell and Command Prompt as the supported alternatives. No global Git setting is required.

Security

• Pin Grok's executable inside the image and disable automatic update checks.
• Read only the current runtime's marked private OAuth session. The official CLI remains the credential writer; the HTTP handler does not execute CLI tools, import other applications' credentials, or consume refresh tokens.
• Preserve existing API installations and credential data without adopting, migrating, or deleting them through the OAuth-only runtime.
• Require SSH host-key confirmation, an exact reviewed plan, and final human confirmation before every VPS write. Keep the companion under /docker/n8n-openai-oauth, leave n8n unchanged, and remove only resources that pass exact ownership checks.

0.13.0 - 2026-09-04

Added

• Add owner-scoped local dashboard lifecycle commands: relmio start, relmio status, relmio open, and relmio stop.
• Reopen Relmio's local launcher to rediscover a fixed inventory of six local services: the OpenAI API endpoint, both Codex endpoints, the owned n8n stack, the OAuth bridge, and the AI Assistant tools.
• Copy verified service URLs and use only actions supported by the latest inventory state, including Codex sign-in, client-credential rotation, OAuth bridge refresh, owned-stack resume, removal, and a return to setup.

Changed

• Keep hosted curl, PowerShell, and Command Prompt launches foreground-only when they use a verified temporary runtime, while persistent package installs can manage the dashboard between terminal sessions.
• Make the persistent dashboard the default local launcher's home while keeping the existing four-step Add connection setup flow intact and the separate VPS flow available through relmio vps.
• Present Docker availability, service boundaries, component state, and recovery choices in a responsive dashboard that marks stale inventory and refreshes it read-only.
• Pair every dashboard rail label with an accessible inline SVG icon, using a neutral workflow mark for n8n at the compact navigation size.

Fixed

• Re-attest exact generated Docker ownership and health before reporting a service as healthy, including each companion's selected external n8n container and Docker network on Windows.
• Keep owned n8n-stack recovery guidance accurate across Compose validation, image pulls, and startup waits. Safe single-line validation details can now explain the problem without exposing paths, credentials, or raw Docker output, extending the Compose diagnostics introduced in v0.12.2.

Security

• Open the local dashboard through an owner-only, single-use browser handoff whose route-bound capability expires after 30 seconds and never appears in process arguments, the visible URL, redirects, or cookies.
• Leave ChatGPT OAuth ownership with the official Codex App Server, keep one active account per Codex target, reject automatic account or key changes after authentication, rate-limit, or quota failures, and deny undocumented provider authentication methods by default.
• Return only sanitized states, components, and allowlisted endpoints from persistent inventory. Relmio never reveals stored secrets, and one-time values are consumed before returning to the dashboard.
• Recheck owner-only Windows ACLs at mutation time and fail closed before any Docker or Compose change when a managed path or file no longer has its attested permissions.
• Let operators disconnect the VPS explicitly and close idle authenticated SSH sessions after 15 minutes without interrupting an active remote operation.

0.12.2 - 2026-09-03

Fixed

• Let Code Sandbox + SearXNG local n8n setup work on current Docker Desktop (Compose v5) instead of failing before images are pulled, and show the actual Compose validation error when that step still fails.

0.12.1 - 2026-09-02

Fixed

• Give the first local n8n + ngrok start enough time to pull images and pass health checks, and explain image-download versus startup-wait failures without exposing Docker output or credentials.
• Explain Windows Docker Desktop WSL 0x800705aa engine-start failures in the local n8n wizard instead of a generic Compose error.

0.12.0 - 2026-09-02

Added

• Support local Docker endpoint, n8n bridge, Assistant companion, and owned n8n stack workflows on native Windows with Docker Desktop's attested desktop-linux engine.
• Add manage/edit actions for detected local and VPS n8n integrations, including a fresh ChatGPT sign-in path and Assistant/SearXNG configuration.
• Add a Ko-fi support link to the GitHub and npm README guides.

Changed

• Guide beginners through ngrok agent-token and user-created Basic Auth fields with inline validation, missing-field warnings, locked install controls, and visible progress while setup is running.
• Keep the same reviewed browser wizard behavior across PowerShell, Command Prompt, npm/npx, curl, and Homebrew launch paths instead of relying on shell-specific setup logic.

Fixed

• Recover interrupted owned n8n-stack operations without moving a newer live lock, including crashed processes, PID reuse, incomplete lock publication, and a preserved partial-stack recovery result when lock cleanup also fails.
• Wait for generated n8n, ngrok, and sandbox health checks consistently, accept the sandbox API's private 8080/9090 metadata, and continue rejecting UDP or host-published Assistant ports.
• Preserve existing n8n enabled modules when adding AI Assistant settings and leave existing n8n containers, Compose files, images, and restarts untouched.

Security

• Protect every Windows managed credential directory and file with a read-back-verified NTFS DACL limited to the current account before writing secrets or invoking a Docker mutation. Remote Docker selectors, Unix sockets, and unrecognized named pipes remain rejected on Windows.
• Bind lifecycle ownership to the process creation identity, fail closed when liveness is ambiguous, and arbitrate stale recovery before changing the canonical local n8n operation lock.

0.11.0 - 2026-09-01

Added

• Add a hosted changelog generated from the repository release history.
• Add a wordless animated sketch banner that keeps Relmio's original two-eyed green mascot and cream doorway.

Changed

• Rewrite the GitHub and npm guides, hosted site, and browser wizard in shorter, plainer language while keeping credential and safety boundaries explicit.
• Add clearer step-by-step guidance for local ngrok and n8n companion setup.
• Use icon-only copy controls with accessible labels, larger touch targets, and responsive layouts across the hosted site and browser wizard.

Fixed

• Keep a completed local wizard from reopening a consumed setup plan; starting another option now prepares a fresh plan.

[0.10.0] - 2026-08-31

Added

• Add local n8n companion choices to the browser wizard: an unofficial private openai-oauth sidecar for an existing n8n container and n8n AI Assistant Code Sandbox support with optional SearXNG JSON web search.
• Add a one-click, separately owned disposable n8n + ngrok stack with mandatory Traffic Policy Basic Auth, loopback-only local ports, and optional Code Sandbox plus SearXNG.
• Add an animated Gateway Android README banner, a two-lane credential-boundary diagram, package and project badges, and a concise product introduction.

Changed

• Bind each companion plan to the exact running n8n container and selected existing Docker network; neither path publishes a host port or edits, restarts, stops, recreates, rebuilds, or executes inside n8n. Assistant settings remain operator-applied.
• Let every completed local setup return to the token-preserving start screen so another endpoint can be configured without restarting the wizard.
• Use Homebrew's formula-scoped trust command for the public Relmio tap instead of asking users to trust an entire third-party tap.

Fixed

• Accept only Docker Compose's strict unpublished publisher placeholder while retaining fail-closed rejection for real or malformed host publication.
• Give disposable n8n's cache an owner-writable temporary filesystem so n8n AI Assistant workspace initialization can complete.
• Keep bind-mounted ngrok and SearXNG configuration readable by their non-root containers while parent managed directories remain owner-only.

Security

• Keep credentials separate: the sidecar uses a private local ChatGPT OAuth volume, while the Assistant model-provider credential is configured directly in n8n and is not handled by Relmio. SearXNG remains optional and off by default.
• Scope the new public exception to the owned n8n route behind mandatory Basic Auth; port 10531, Code Sandbox, and SearXNG remain unpublished, and removal requires exact project-wide ownership attestation.

[0.9.1] - 2026-08-31

Added

• Add a prominent hosted installer launch option for the separate n8n AI Assistant wizard, while retaining every existing general installer method.

Changed

• Redesign the hosted app as an Editorial Console with an interactive four-route Signal Plotter and an improved multi-turn chat console that can stop in-flight responses.
• Describe the AI Assistant's SSH-connected self-hosted target without provider-specific Hostinger or VPS labels, and document that direct local Docker-socket discovery is not supported.

Fixed

• Stabilize reduced-motion hydration by giving the Signal Plotter a deterministic initial signal state before animation begins.
• Keep the hosted repository control's offline metadata fallback synchronized with the prepared release version.

0.9.0 - 2026-08-28

Added

• Add relmio assistant, a dedicated local wizard and isolated companion Compose plan for n8n AI Assistant's self-hosted sandbox and optional SearXNG web search.

Changed

• Add sourced OpenAI policy context to the GitHub README, npm README, and canonical security guide, including the maintainer's Codex for Open Source acceptance, while preserving the distinction between supported Codex/ChatGPT sign-in patterns and unsupported subscription-to-API conversion, resale, account sharing, or safeguard bypass.

Security

• Keep the privileged Docker-in-Docker runner separate from the selected n8n network, publish no companion host ports, generate and redact independent sandbox secrets, attest ownership-bound random Compose identities and network aliases, serialize VPS-sidecar and assistant mutations under one single-use plan lock, and retain the strict no-n8n-mutation boundary.
• Pin every generated AI Assistant companion production image to its reviewed immutable tag and OCI index digest, with regression coverage that rejects floating or digestless references, including the nested sandbox image.

[0.8.1] - 2026-08-26

Changed

• Adopt the Gateway Android logo across the GitHub and npm READMEs, hosted site, and local wizard, and stop publishing the retired Harbor Gate mark.
• Refresh the Open Graph and social-preview card with the Gateway Android while preserving the Relmio relay message and visual flow.
• Add a prominent Legal warning against bypassing rate limits, restrictions, or safeguards.

[0.8.0] - 2026-08-26

Added

• Let trusted local backends and development servers receive Chat Adapter turns as opt-in Server-Sent Events, with progress and text deltas followed by one explicit terminal outcome so a completed response is distinguishable from a redacted failure.
• Let the setup-token-protected local wizard tester show that incremental response flow after a short-lived encrypted credential handoff, without a direct browser-to-adapter request.
• Refresh the hosted chat experience and local installer presentation, and adopt the Harbor Gate abstract mark across Relmio surfaces.

Changed

• Document the streaming contract, local tester behavior, and its limits in the canonical endpoint/reference guides and generated hosted documentation.

Security

• Keep the Chat Adapter experimental, loopback-only, and limited to trusted local backends or development servers; it rejects browser origins and is not an OpenAI /v1 endpoint or a substitute for an OpenAI Platform API key.

[0.7.0] - 2026-08-16

Added

• Add an encrypted in-wizard tester for the experimental Chat Adapter, plus safe sample Chat Adapter and Codex App Server commands for local testing.
• Add generated hosted guides for getting started, local endpoints, VPS and n8n, troubleshooting, FAQ, security, and reference information.

Changed

• Synchronize concise root and npm READMEs around product, installation, security, and common-problem overviews that link to hosted guides.
• Fact-check ChatGPT/Codex token-refresh guidance across documentation: tokens refresh during active use, the official documentation specifies no fixed 10-day lifetime, and the provider credential remains distinct from Relmio's rotatable client capability.

Security

• Limit tester destinations to literal loopback HTTP addresses, retain private keys only in memory for a bounded lifetime, encrypt entered credentials before they cross the browser boundary, require POST after a completed Chat Adapter install, keep sample bearer values out of process arguments, and erase or abort sessions when forgotten, rotated, or shut down.

[0.6.0] - 2026-08-15

Added

• Add an experimental loopback-only Codex Chat Adapter for trusted local backends and development servers, with bearer authentication, multi-turn conversation IDs, strict resource bounds, and a small Relmio-specific POST /chat contract.

Changed

• Make Codex device sign-in target-aware so the experimental Relmio /chat adapter and native App Server retain isolated, persistent ChatGPT credentials; a Platform API key powers neither target and remains reserved for the generic OpenAI-compatible /v1 endpoint.

Security

• Reject browser-origin adapter requests, keep the adapter separate from Platform-key-backed generic OpenAI-compatible /v1 semantics, explicitly deny model turns access to the private Codex credential store, run chat turns read-only without network access, and preserve loopback-only publication plus credential rotation.

[0.5.0] - 2026-08-15

Added

• Add a Rotate client credential action for installed local endpoints that shows the replacement capability before activation and preserves the upstream Platform key or Codex credential/workspace volumes.

Changed

• Keep System, Light, and Dark appearance controls plus Ko-fi, GitHub stars, and the current package version available throughout the local install wizard.

Fixed

• Install operating-system CA certificates in the isolated Codex image so the official ChatGPT device-code sign-in can establish its trusted TLS connection.
• Wrap local safety and error notifications instead of clipping longer text.

Security

• Verify the generated Codex capability with a strict authenticated WebSocket upgrade before reporting installation or rotation success.
• Serialize installation, sign-in, restart, and credential rotation across Relmio processes, with attested stale-lock recovery and fail-closed rollback that restores the prior verifier and re-attests endpoint readiness.

[0.4.1] - 2026-08-14

Changed

• Add a manual Stop sign-in action while a fresh ChatGPT login is pending, then detect and reject results from superseded wizard attempts.

Fixed

• Terminate the OAuth helper process tree when sign-in is stopped or Relmio exits, preventing a rejected or abandoned attempt from continuing to hold the localhost:1455 callback port.

Security

• Fail closed when OAuth process cleanup or credential promotion cannot be confirmed, blocking another login until Relmio restarts instead of risking an ambiguous helper or credential state.

[0.4.0] - 2026-08-13

Added

• Add a local Docker wizard for private compatible clients through a Platform-key-backed OpenAI-compatible /v1 endpoint, plus a separate official experimental Codex App Server target for trusted ChatGPT-sign-in clients.
• Add compact Ko-fi support links to the hosted navigation and public package guides.

Changed

• Make the local credential boundary explicit across the product: a Platform API key powers compatible /v1 requests, while ChatGPT sign-in powers only the experimental Codex App Server protocol.

Fixed

• Keep the controlling terminal attached when the macOS/Linux installer is piped through sh, so the Relmio wizard can open its interactive browser setup flow.
• Install Homebrew dependencies in their required order during release-candidate validation.

Security

• Bind local endpoints exclusively to loopback, require one-time Relmio capabilities, pin every managed operation to an attested local Docker socket, and isolate provider credentials in target-specific containers.
• Restrict the managed Codex endpoint to the ChatGPT login method.

[0.3.1] - 2026-08-10

Changed

• Make the browser wizard beginner-friendly with a modern fixed-viewport layout: all five active steps fit without document scrolling on common 1280x720 laptops, while progress and safety context stay persistent beside the active task.
• Keep narrow-phone documents fixed to the viewport and contain unavoidable long-form overflow within the active task panel instead of the page.
• Expand the GitHub and npm walkthroughs with a hosted-install selector and packaged, sanitized screenshots that document the current n8n workflow.

Security

• Restore a clean hosted-web dependency audit by pinning patched js-yaml and nanoid releases and using the compatible vinext release that does not include the currently vulnerable image-size parser.

[0.3.0] - 2026-08-05

Added

• Add a copy-ready n8n HTTP Request recipe to the local wizard, including the private Chat Completions URL, Generic Credential Type → Bearer Auth fields, the harmless local-only bearer placeholder, JSON headers, the structured response-format body, and a full recipe copy action.
• Add the same structured gpt-5.6-sol example and importable cURL recipe to the GitHub README, npm README, and n8n configuration guide.
• Add a repository-local changelog skill that standardizes Relmio's patch, pre-1.0 feature, and stable major release numbering and metadata checks.

Changed

• Treat 0.3.0 as Relmio's major feature release within the pre-1.0 series; it consolidates the key improvements shipped from v0.2.10 through v0.2.14: resilient Windows OAuth/bootstrap flows, native Command Prompt installation, compact accessible wizard recipes, verified Homebrew/package-manager preparation, and release-time package checks.
• Keep the HTTP Request body aligned with n8n's messages format by targeting /v1/chat/completions; the separate OpenAI Chat Model guidance continues to support the Responses API where that node exposes the switch.

[0.2.15] - 2026-08-05

Fixed

• Exit cleanly when WinGet or another non-interactive validator probes the portable command without arguments, while keeping the browser wizard for interactive Command Prompt and PowerShell sessions.
• Add a redirected-stdio portable smoke test so future WinGet candidates cannot regress into a never-ending default launch.
• Exclude local npm cache directories from portable release archives.

[0.2.14] - 2026-08-04

Added

• Add staged Homebrew formula and WinGet portable-package generation with x64/ARM64 manifests, installed-command smoke tests, and review-only CI artifacts for package-manager publication.
• Add relmio --version and relmio -v for noninteractive installer and package-manager verification.

Changed

• Report Homebrew and WinGet publication status in the hosted install page and documentation, while keeping unapproved commands out of the primary picker.

Fixed

• Replace the hosted Command Prompt installer route's PowerShell launch with a PowerShell-free, non-admin native batch bootstrap that reuses Node.js 22+ when available or verifies a pinned official Windows runtime before use.
• Keep downloaded checksum-manifest text out of CMD evaluation and use reviewed Node.js 22.23.2 x64/ARM64 digests embedded in the release.
• Download the CMD bootstrap to a collision-resistant temporary name without overwriting an existing install.cmd, then clean it after execution.
• Show deterministic download, checksum-verification, and extraction stages in every bootstrap so temporary Node.js runtime setup does not appear stalled.

[0.2.13] - 2026-08-04

Changed

• Compact the local setup wizard so its active step stays near the top, move safety and status notices into dismissible toasts, and collapse the optional AI Agent and HTTP Request recipes until users choose to open them.
• Complete the HTTP Request recipe with authorization, content type, and a copyable sample Responses API JSON payload.

Fixed

• Copy credential values reliably in Opera GX on Windows by preserving the synchronous user gesture for the selection-based clipboard path before falling back to the modern Clipboard API.

[0.2.12] - 2026-08-04

Fixed

• Install the pinned openai-oauth@2.0.0 helper with its exact compatible zod@4.1.8 peer even when inherited npm settings omit peer dependencies, preventing the Windows sign-in helper from exiting before it prints a URL.
• Accept the same strictly validated authorization line from either helper output stream, including Windows terminal framing and final drained output.
• Run npm package builds through the current Node.js runtime on Windows instead of executing npm.cmd directly with shell: false.

[0.2.11] - 2026-08-04

Fixed

• Open the private Windows wizard URL through the documented default-browser association instead of asking Explorer to treat the URL as a folder, while retaining the printed URL and Enter-to-retry fallback.
• Make the Command Prompt installer copy call the system Windows PowerShell executable directly, avoiding ambiguous powershell command resolution.
• Parse the supported openai-oauth@2.0.0 login line across Windows terminal control sequences and chunk boundaries, and report a sanitized, actionable callback-port conflict instead of a generic missing-link error.

Security

• Pin the hosted web tooling to brace-expansion 5.0.9, which includes the upstream denial-of-service fix required by the release audit.

[0.2.10] - 2026-08-03

Fixed

• Probe installed Windows Node.js runtimes with the literal node --version output instead of a node -p expression, avoiding the PowerShell [eval]:1 quoting failure while still reusing Node.js 22 or newer.
• Let interactive wizard terminals reopen the local browser page when the user presses Enter, while retaining the printed private URL as the fallback for noninteractive launches.
• Navigate the preopened local OAuth tab before severing its opener access, show an immediate preparing state, and close an unnavigated waiting tab if sign-in setup fails.

[0.2.9] - 2026-08-02

Fixed

• Launch the local ChatGPT OAuth helper through the current Windows Node.js runtime and npm's JavaScript CLI instead of executing npx.cmd directly, preventing spawn EINVAL while preserving the native npx path on macOS, Linux, WSL, and Git Bash.
• Explain how to recover when a refreshed wizard page no longer has its private session URL.

[0.2.8] - 2026-08-02

Changed

• Replace the local browser wizard's text appearance selector with compact, accessible System/Light/Dark icons, while keeping the original horizontal Signal Spine flow and touch-friendly behavior.
• Serve the bundled Lucide SVG assets from the wizard and include them in the npm package so offline and Node-free browser launches render consistently.

[0.2.7] - 2026-08-02

Changed

• Restore the original Relmio hosted layout and local browser wizard flow, with the horizontal five-step Signal Spine and GitHub star/version control kept visible.
• Add Astryx's built theme and accessible segmented appearance control to the hosted app, plus lightweight System/Light/Dark preference support to the local wizard.
• Add responsive dark-mode logo treatment, phone-sized controls, and matching browser-wizard guidance to the GitHub and npm README variants.

Fixed

• Keep the checksum-verified temporary Node.js 22 runtime on the child process path so Git Bash and other Node-free systems can launch the Relmio package shim without falling back to a missing or outdated system node command.

[0.2.6] - 2026-08-02

Added

• Add a curl-based wizard bootstrap for macOS, Linux, WSL, and Git Bash that reuses Node.js 22+ or downloads and checksum-verifies a temporary official runtime when Node.js is not installed.
• Add a native Windows PowerShell bootstrap for PowerShell and Command Prompt that works without Git Bash or a preinstalled Node.js runtime and verifies the temporary official Windows archive before execution.
• Expand the hosted installer into an accessible macOS/Linux, PowerShell, Command Prompt, and NPX terminal switcher.

Changed

• Revamp the hosted Vercel experience and local browser wizard around the Signal Spine composition and Patchbay Ledger design language.
• Integrate the Astryx component system, neutral theme, CLI, and AI-readable setup guidance in the hosted React application while keeping the published wizard dependency-light.
• Add persistent route context, sanitized preview status, responsive layouts, reduced-motion behavior, and synchronized GitHub/npm setup documentation.

[0.2.5] - 2026-07-31

Added

• Add a sanitized successful hosted-chat screenshot and a visible guide to the required Sign in with ChatGPT browser extension across the website, GitHub README, npm README, and troubleshooting documentation.
• Explain that the hosted extension requirement is separate from the local npm wizard callback, where a callback-capturing extension may need to be disabled temporarily during sign-in.

[0.2.4] - 2026-07-31

Changed

• Add explicit foundation and attribution language for Evan Zhou Dev's openai-oauth project to the GitHub and npm README explanations.

[0.2.3] - 2026-07-31

Added

• Add sanitized GPT-5.6 Sol and Luna AI Agent examples, a model-selector compatibility preview, and the completed Docker sidecar state to both the GitHub and npm README experiences.

[0.2.2] - 2026-07-30

Added

• Add aligned npm keywords and GitHub repository topics for Relmio, GPT model variants, n8n, AI agents, and API-key discovery.

[0.2.1] - 2026-07-30

Added

• Add the hosted ChatGPT site link to the package metadata and public guides.
• Document the upstream Codex relay model, known limitations, and legal responsibilities in both the GitHub and npm README variants.
• Add a command-first install page for the current n8n and Hostinger VPS wizard, plus a GitHub control with live package and repository metadata.
• Credit Evan Zhou Dev's openai-oauth method on the hosted Relmio page.

Changed

• Refresh the local setup wizard and hosted chat presentation with the Relmio redesign, including clearer progress, copy feedback, responsive layouts, and request-state affordances.
• Move the hosted chat and package homepage to relmio.vercel.app with Node.js 22 and repository-driven preview and production deployments.
• Return ChatGPT OAuth callbacks to the deployment that started sign-in so Vercel preview URLs and the production domain both work.
• Run hosted web linting, type checks, builds, tests, and dependency auditing in GitHub Actions alongside repository-driven deployments.
• Point package and documentation metadata at the canonical relmio repository.
• License Relmio under Apache 2.0 and preserve the upstream openai-oauth attribution in the distributed notice.

Fixed

• Stream hosted chat responses incrementally through deployment proxies and surface safe request errors instead of leaving an empty assistant message.
• Distinguish a ChatGPT hosting-network challenge from an expired OAuth session without exposing upstream response bodies or credentials.

[0.2.0] - 2026-07-29

Added

• Add a provider-neutral product roadmap with a gated SuperGrok/xAI OAuth feasibility track, entitlement checks, and explicit security boundaries.
• Add a trusted-publisher GitHub Actions workflow for short-lived npm authentication after the first package publication.

Changed

• Rename the public product and npm package to Relmio and relmio so the project can grow beyond its initial n8n setup path.
• Replace the generic plus icon with an original two-lane relay mark and add a small brand guide with reusable SVG and source concept assets.
• Publish a concise npm-specific README with absolute image and documentation URLs while preserving the full GitHub README and its Mermaid diagrams.
• Build and inspect a deterministic npm tarball so the registry receives the npm-specific README instead of the repository README.
• Keep the legacy n8n-openai-oauth-setup executable alias and every deployed n8n-openai-oauth compatibility and safety identifier unchanged.

[0.1.8] - 2026-07-29

Changed

• Restore the complete manual sidecar installation path to the README for wizard failures, debugging, and contributor reproduction.
• Add plain-English Mermaid diagrams that explain the private sidecar and help readers choose between the browser wizard and manual setup.
• Keep the README and standalone manual Docker templates synchronized with automated documentation checks.

[0.1.7] - 2026-07-28

Changed

• Add prominent workflow-backup reminders to the README, manual guide, troubleshooting guide, and browser wizard before VPS access.

[0.1.6] - 2026-07-28

Added

• Add a public npm quick-start guide with five sanitized setup screenshots, Mermaid architecture diagrams, and a YouTube walkthrough outline.
• Add individual Base URL/API-key copy controls and n8n recipes for OpenAI Chat Model, AI Agent, Basic LLM Chain, and HTTP Request nodes.
• Add a release metadata validator that keeps the package, lockfile, changelog, and release tag on one version.
• Add GitHub Actions checks with immutable action pins, no persisted checkout credential, and the repository's pinned npm 10.9.8 runtime.

Changed

• Expand troubleshooting for stale wizard sessions, npm versions, local OAuth callbacks, SSH failures, Docker networks, real port mappings, and manual sidecar collisions.
• Use the wizard-only ~/.n8n-openai-oauth/auth.json path consistently in the manual and maintenance guides.
• Replace pre-publication wording and add the local context file to the shared ignore policy.
• Disable npm lifecycle scripts explicitly in every documented and nested npx invocation.
• Add a prominent workflow-backup reminder before local setup and inside the wizard because VPS access remains a real write boundary even with sidecar-only commands.
• Separate OpenAI credential fields from OpenAI Chat Model settings and explain the Responses API compatibility behavior for Chat Model node version 1.3.
• Polish the public README with a collapsible contents list, clickable project links, experimental-use disclaimers, and a contributor guide.
• Document Graphify as an optional local maintainer map while keeping raw graph exports out of Git and npm.

Fixed

• Prevent sanitized preview mode from generating or opening a live OpenAI authorization URL.
• Refuse to show the ready screen when the sidecar returns no usable model ID.
• Tag the exact commit that passed CI and was published instead of relying on the shell's current HEAD.
• Make the sanitized preview follow the production OAuth service contract and show the correct private n8n Base URL.
• Fall back to a temporary selected text field when a browser denies the modern Clipboard API, so the final credential copy buttons still work.
• Always remove the fallback copy field and restore focus when legacy browser clipboard access throws.
• Stop and remove only the named wizard-managed sidecar service when its final safety check detects an unexpected host-port publication; report an explicit manual cleanup path if that removal cannot be confirmed.
• Clean up the sidecar when publication inspection fails or returns malformed metadata, rate-limit install attempts, close the VPS connection after every install outcome, and show actionable browser recovery messages.

[0.1.5] - 2026-07-28

Fixed

• Detect a newly approved ChatGPT credential as soon as its complete file is available instead of waiting for the OAuth helper process to close.
• Poll the local sign-in state more frequently during the first ten seconds so the wizard responds quickly after browser approval.
• Show the local credential's update time and announce when a fresh sign-in has been saved.

[0.1.4] - 2026-07-28

Fixed

• Run ChatGPT login against a new wizard-only credential file so the bridge CLI never needs an interactive terminal to confirm replacement.
• Validate the completed credential before storing it at ~/.n8n-openai-oauth/auth.json with owner-only permissions.
• Stop reusing or overwriting the Codex app credential at ~/.codex/auth.json.
• Open the exact fresh authorization URL returned by the pinned bridge CLI and report its completion separately, avoiding stale browser sign-in tabs.
• Verify Docker Compose publisher metadata so an internal-only 10531/tcp declaration is not mistaken for a published VPS host port.
• Explain that browser extensions which intercept the localhost OAuth callback must be disabled temporarily during a fresh sign-in.

[0.1.3] - 2026-07-27

Fixed

• Let the explicit Refresh ChatGPT sign-in action confirm replacement of an existing local OAuth credential. Previously, the bridge CLI prompt had no input stream, defaulted to “No,” and the wizard reported that sign-in did not finish.
• Clarify when the wizard will reuse an existing credential and when to refresh it.

[0.1.2] - 2026-07-27

Fixed

• Provide a writable /home/node/.local tmpfs so the non-root bridge can start while the container root filesystem remains read-only.
• Use the collision-resistant Docker hostname n8n-openai-oauth so an existing manual openai-oauth sidecar cannot capture n8n requests.
• Treat a wizard-managed deployment as an update, allowing a fresh local ChatGPT sign-in to refresh its OAuth credential safely.

[0.1.1] - 2026-07-27

Fixed

• Quote the generated Compose healthcheck command so Docker Compose validates it as a string.

[0.1.0] - 2026-07-27

Added

• Local browser wizard for installing the OpenAI OAuth sidecar beside a self-hosted n8n Docker deployment.
• Read-only n8n and Docker-network discovery.
• Explicit review and confirmation before remote sidecar writes.
• Safety checks that prevent changes to the existing n8n Compose project, image, container, or host port mappings.
• npx-friendly CLI entry point and beginner documentation.

Security

• OAuth credentials stay on the local computer until the user approves an SFTP upload to the installer-managed sidecar directory.
• SSH host-key confirmation is required before password authentication.
• The sidecar uses an internal-only Docker network endpoint and no published VPS port.