Agentrix v1.3.26-beta.8

Windows test build: This prerelease is for Windows only and may be unstable. Choose Beta in Settings > About and accept the test agreement to receive test updates.

Changes in 1.3.26-beta.8

• Step lines In native mode a tool step shows what it did and the file or command, without the output; a failure adds one line saying why.
• Codex commands Codex in native mode takes /model, /effort, /fast, /new, /clear, /compact, /review and /status, and ! runs a command.
• Token rate In native mode the activity pane shows the main model's token rate.
• One picture once A picture mentioned more than once in a reply is shown once.
• Quote from the menu Right-click a message to quote or copy it; with text selected, only the selection is used.
• No click frame Clicking a window or split no longer draws a frame.
• Compaction errors When native mode fails to compact the context, it says why.
• Command titles Commands that write a file are no longer labelled as reading one.
• Native mode Codex always uses native mode in Agentrix, including when you type codex in a terminal.
• Message senders Sub-agent splits identify each message's actual sender, distinguishing your interventions from the main model's instructions.
• Higher turn allowance Claude native mode no longer has a $5 per-turn cap and continues to follow your monthly budget.
• Voice input Click the microphone beside the message box to dictate text in Claude and Codex native mode.
• Big pictures keep running A detailed picture or a very long output no longer stops the task; a result too large to show is skipped alone, and the app says so.
• Codex runs natively Codex always uses native mode in the app. Startup failures explain why, and long conversations can be resumed.
• Native mode commands Claude native mode supports /clear, /new, /model and /effort, and ! to run commands; the output goes into the conversation.
• Compaction shown Compacting the context shows progress, and a line says whether it worked.
• Workers read settings Claude sub-agents read CLAUDE.md, skills and project settings, as they do in the terminal.
• Finished calls leave Finished calls leave the list by themselves, failed ones after 30 seconds, and stay while their view is open.
• API session fixes Fixed API sessions that stopped at their first tool or showed no reply, and project settings that could send the saved key elsewhere.
• Run mode applies at once Changing the run mode, or saving or clearing the API key, also moves running sessions, and the conversation carries on without reopening: idle ones switch at once, working ones when the turn ends, and ones with background tasks ask you first.
• Agent tab The models called moved to the left sidebar beside Files and Git, and show by default when a project opens.
• Activity pane no longer scrolls as a whole When space runs out only the lists scroll. Context length and limits sit together: an API connection shows spend and budget, a subscription shows the 5-hour and weekly limits.
• Worker fixes Fixed API workers that could hang at the end, Codex workers that could not write after a roundtable, ended workers that kept showing as running, and model aliases wrongly reported as a mismatch.
• Billing check The connection test confirms Claude Code reports an API key as its billing source; if not, new sessions move to the terminal and say why, instead of failing.
• Native path fixes Fixed Codex's native connection and a startup failure when Claude runs through the API.
• Native mode Claude and Codex both support native mode. Commands, changes and plans are shown directly. Questions and policy approvals are handled here too. Claude can switch back to the terminal in Settings.
• Run mode Settings stores the API key, tests the connection and sets a monthly budget; Claude uses the API by default and moves to the terminal on your subscription when the quota runs out, continuing the same conversation.
• Connection and quota The activity pane shows whether Claude and Codex are connected through the API or a subscription, plus today's and this month's estimated API spend, the budget and rate limits.
• Direct interventions A note to a sub-agent goes straight to it (Claude takes it mid-run, Codex after the current step), not through the main session, and its split shows the delivery state.
• Real-time state The terminal's state reaches the reading view and the activity pane at once, so you can continue the moment it finishes; while a terminal confirmation is open, your message waits and goes out afterwards.
• Update channel Settings › About lets you choose the stable or beta channel; beta needs the beta agreement first, and you can switch back to stable at any time without a downgrade.
• Confirm in terminal In the terminal view, whenever you need to confirm or choose, and whenever you type a slash command, the terminal pops up at once; a terminal that pops up while you type never takes the focus.
• Dashboard Colour charts now compare each model’s time, cost and pass rate at a glance, refreshed every minute.
• Sub-agent split Click a subagent in the Agent tab to see what it is doing beside the conversation. One shows at a time, and clicking another replaces it. Dispatched workers can take your notes right there.
• Quote a message In the reading view you can quote a message and continue the conversation from it.
• Built-in skills The built-in version always applies: a project copy with the same name is moved to a recovery folder automatically; to extend one, name your skill “-”.
• Skills update live After a skill changes, a running Claude Code session receives the new version at its next compact or clear.
• Virtual Lab Choose any of the three task sources (this project's history, well-known open-source projects and the AA benchmark); each runs on its own and the results are compared side by side. Models in an experiment cannot reach Docker on your computer or change how they are graded.
• Type while it works Keep typing and sending while Claude Code or Codex works; the message waits until the current step is done. Stop and Esc work at any time.
• Light terminals In the light theme, text drawn for a dark terminal turns dark, so body and secondary text are both easy to read.
• Model routing Back-end design and debugging go mainly to GPT-6.1-Sol, which now defaults to high + fast.

Windows: install Agentrix-Setup-1.3.26-beta.8-x64.exe to enable automatic update checks and background downloads.

Install Codex CLI on the target computer.

Commit: 64f2619
Build: https://github.com/noeigenstate/Agentrix/actions/runs/42497223865
SHA-256 checksums are attached in SHA256SUMS.txt.