cogwheel0/conduit v4.3.1

on GitHub

Highlights

• LibreChat (Beta) becomes Conduit's fourth backend: sign in, chat with its model specs, agents and models, and keep its history in the Chats tab.
• Hermes 0.23 works over the Desktop Gateway, a gateway's own Webapp listener and the API server, and gets Rooms, chat pinning and hiding, Context & Usage, error cards and background-work controls.
• Open WebUI 0.13.3 is supported, with its terminals' coding agents offered as models and a full-screen Computer viewer for a terminal's desktop and browser.
• iOS 26 uses many more of the system's own controls, and voice and read-aloud float on Liquid Glass islands.

LibreChat (Beta)

• Add a LibreChat server from the backend chooser or the LibreChat tab in the account sheet. Sign in with email and password or LDAP, two-step verification by code or backup code, or the server's single sign-on buttons. (#943, #953)
• Switch between LibreChat and Open WebUI accounts in Accounts. Open WebUI accounts keep their session and data while LibreChat is in use. (#945–#948)
• LibreChat chats fill the Chats tab, stay in sync and searchable, and can be read offline. Rename, pin, archive and delete work offline and reach the server later. (#949, #952)
• The model picker lists the server's model specs, agents and each endpoint's models under foldable sections. Send, stream, stop and regenerate. A reply keeps running while another chat is open. (#950, #954)
• Attach files, photos and camera shots. Turn on web search, code interpreter, file search and MCP servers. Answer tool approvals and questions, and insert prompts from the LibreChat prompt library with /. (#951, #955)
• Share links, Branch into new chat, Duplicate Chat, memories in Personalization, and the server's speech engines. (#956)
• Controls that only feed Open WebUI requests are hidden for LibreChat models.

Hermes

Hermes 0.23

• Desktop Gateway: approvals, clarifications, sudo and secret prompts show again on Hermes 0.22 and later. The Toolsets page and MCP reloads work on 0.23. (#959)
• Webapp listener: chat through a gateway's own Webapp listener: create, send, stop, steer, rename, branch and switch models, plus /model, /branch and /compress. Conduit waits for a gateway that is starting up and lists its installed skills. (#960)
• API server: approvals can be answered on 0.22.1+, and refusals say what went wrong instead of a bare HTTP error. A chat started in another Hermes app offers Continue in a Copy. Runs interrupted or held by a restart recover, and dropped run streams resume. (#961)
• Older servers: the reasoning effort you pick now applies to API server chats. PDFs are no longer offered there, since Hermes rejects them. On the Desktop Gateway, a PDF Hermes can't render is sent as a file the agent can read, instead of failing the message. API server scheduled jobs can be edited. Chats that start with the same words keep their own sessions. (#962)

Chats

• Pin, Archive and Hide Hermes chats, with a Hidden Chats page in Hermes settings. Branch from Here and Undo Last Turn. Editing or regenerating a message in a Desktop Gateway chat now continues the same server session, with its context. (#964)
• All of a turn's questions are answered in one form, also on API server runs. Edited files show as a diff card, with paths relative to the chat's working directory. Paid tools show their price on approvals, and the agent's plan stays above the composer. (#966, #1009)
• Context & Usage in the chat menu shows how full the context window is, what fills it, token and cost totals and your plan's limits, and can compact the chat. (#967)
• A failed turn shows a card that says what broke and offers one fix. (#967)
• A chip above the composer shows running subagents and background processes. Read a subagent's output, steer or stop it, or kill a process. (#968)
• Conduit asks before switching to a model that Hermes holds for confirmation: Switch and Send or Keep Current Model. (#1013)
• The answer to a message Hermes saved during a restart now appears once Hermes replies. (#1014)

Rooms and scheduled jobs

• Rooms: Desktop Gateway connections can open group chats with two to six bots, with @mentions, live bot status and approvals. (#965)
• On the Desktop Gateway, scheduled jobs can be stopped mid-run (Hermes 0.22.6+), can choose what happens to Missed Runs, and can Deliver To messaging apps, Bot Chats and Group Chats. (#963)

Connections

• Conduit asks before a connection sends credentials over plain HTTP to another machine, and such connections are marked Not encrypted. Localhost and HTTPS never ask, and existing connections keep working. (#1011)
• Catalog MCP servers ask for their credentials when added, and OAuth servers go on to sign in. (#1015)
• With two or more connections, switch between them from the name at the top of the sidebar. (#1021)
• The dashboard sign-in page no longer goes blank when the keyboard is up in the account sheet. (#1010)

Open WebUI

• Open WebUI 0.13.2 and 0.13.3 work without a compatibility warning. (#987, #1018)
• Coding agents: with a terminal selected in a saved chat, its coding agents (Codex, Claude Code and others) appear under Coding Agents in the model picker. Agent chats don't offer Regenerate, Edit or Compare, which the server would refuse anyway. Requests from coding agents in chats started on the web can be answered. (#987, #1018)
• Computer: a button in the chat header opens the selected terminal's Desktop or Browser full screen. Watch while a reply runs, or Take Control to type, use the trackpad, zoom and browse. Turn it off under Settings > Chat. (#969)
• Memory recalls, skill lists and other tool calls the server marks hidden no longer show as tool blocks, and tool blocks show the model's label for each call. Notification preferences are saved where the web client keeps them. (#987)

Direct

• Tool calls no longer stop after 8 rounds or 16 calls. A loop now ends after 3 failed calls in a row, then asks the model for a final answer. (#1008)
• New Images per request setting (default 10). Image and document budgets apply per request, so older attachments are left out instead of every later turn failing. (#1008)
• The context limit you pick for a local Ollama model is sent to Ollama as num_ctx. (#1008)
• Reasoning and tool sections survive a reload, and <think> tags are split out of streams. (#995)
• Context fill is measured on a reply's last model call. Streams from servers that send null for unset fields no longer fail. (#984, #999)

iOS

• On iOS 26, the app uses the system's own alerts, sheets, menus, search bars, pop-up buttons, date and time pickers and popovers. A response's usage opens in a popover with context fill, its More menu is grouped, and the note editor has a bottom toolbar. (#970–#979, #992)
• Voice calls and read-aloud float on Liquid Glass islands. Tap anywhere on the track to seek, skip 15 seconds, and follow the highlighted word. (#971, #978, #989, #1012)
• The sliding sidebar shows backends as a tab bar beside a glass New Chat button, and VoiceOver reads each tab's name. (#974, #1012)
• Grouped forms for sign-in and Settings, a calmer chat header, chats and notes listed under date sections, preview tiles for Light, Dark and System, and a greeting by first name. (#999)
• Settings and Workspace pages blur the bar once content scrolls under it. (#1019)

Chats

• Replies are typeset on one rhythm: code blocks wrap long lines, tables size columns to their content, nested lists indent, and follow-up suggestions show as full rows. A reply shows "Thinking…" with the time waited until text arrives. (#999)
• Temporary Chat turns on and off with one tap. One speech speed applies to both cloud and device voices. (#978, #999)
• Long chats scroll without the scrollbar jumping, the Android scrollbar can be dragged, and jumping back to the latest reply leaves it resting on the composer. (#993)
• The chat stays put while you drag the sidebar open, and tapping your profile twice no longer opens Settings twice. (#988, #991)
• Faster streaming Markdown, chat paging, branch previews and message footers. (#985, #986, #997, #998)

Under the hood

• cupertino_native_better 2.5.1, and refreshed dependencies, including openai_dart 12. (#970, #984, #992, #1012)
• LibreChat v0.8.8 and Hermes Agent v0.23.0 vendored as API references, plus a LibreChat dev stack. (#942, #959)