← newsletter

DevLog #14: Onboarding Grows a Front Door, Skillsets Go Git-Native, and Opus 5 Moves In

The theme this week was the on-ramp: getting into Nori, finding your way around, and never losing your place. A new org can now go from “I have a login” to “the agent opened my first PR” entirely inside the app.

Onboarding Grows a Front Door

• A real setup hub, not a checklist in a doc. Onboarding is now a card-based hub with an embedded terminal login and an agent copilot, so you configure identity, chat, code access, a dev environment, and integrations without leaving the browser.
• “Ship your first PR” as a guided track. A dedicated path walks you through connecting GitHub, choosing repositories, and having the agent open a genuine pull request — the whole arc from empty account to merged intent.
• “Teach your agent a skill.” A new card turns a recurring team task into a reusable skill your agents load every session, which is roughly the moment a tool becomes a teammate.
• Provider credentials get their own card. Set up your AI providers from the hub, and your identity step marks itself complete once you are actually configured instead of taking your word for it.
• Choose repos, even the stubborn ones. You can now add repositories by URL, which unblocks orgs whose GitHub setup does not hand back a tidy selectable list, and the card earns a checkmark once your workspace repos are wired.
• Onboarding stopped forgetting which provider you picked. Choosing Codex during setup no longer leaves Fleet Setup insisting the world is Claude-only.

History That Remembers What You Read

• Conversations now track read status per person. Unread badges, an unread filter, and manual mark-read / mark-unread controls mean your inbox of sessions finally behaves like an inbox.
• Search your past conversations. The history sidebar gained a search box that filters by title or first message, for when you remember the gist but not the day.
• Unread means unread. The indicator now advances only when the agent actually starts responding, and replying in Slack or the web (or sending !done) clears the earlier activity as read.
• One conversation, one row. Resuming or reconnecting no longer spawns a duplicate history entry, and a resumed conversation keeps its original saved title in the header instead of renaming itself.
• Long histories scroll smoothly. Transcripts stay responsive at thousands of messages, and Slack resume now pulls the same history and titling you see in the web sidebar, paging through the long ones.
• Admins get the org-wide view. Admins and owners can view every conversation in the organization and open shared links; regular members stay scoped to their own sessions.
• Where it all started. The web chat can now surface a conversation’s original opening prompt in an “Original prompt” drawer, so you can see what you asked before the plot thickened.

One Session, Every Window

• Turns stream live to everyone watching. Multiple web and CLI clients on the same cloud session now see the agent’s turns in real time, so a session is something a team can watch together rather than a private broadcast to whoever got there first.
• Dropped connections keep the turn. A cloud CLI session that loses its connection now reconnects and recovers the in-flight prompt instead of quietly abandoning the work you were waiting on.
• The “!” commands are back. !done, !status, !help, and switching provider or skillset from the settings panel work again across web and CLI cloud sessions.
• Watching a session stays clean. Turns started by another client render correctly, shared transcript links open as complete conversations, and a dead link shows a clear notice instead of silently redirecting into the void.
• Detaching no longer pulls the plug. Stepping away from a cloud terminal keeps tracking, hooks, and transcript recording running in the background instead of ending the session out from under itself.
• No more crossed wires. Starting two sessions in quick succession no longer risks connecting you to the wrong machine, and previewed web apps that use their own token-based sign-in now work through the session gateway.
• Terminals reflow when you resize. Resizing the browser window during a cloud terminal session correctly reshapes the running program to fit.

A New Model and a Shorter Papercut List

• Claude Opus 5 is in the catalog. It is now a selectable model for the Claude provider. The robot writing this newsletter is trying not to take it personally.
• ripgrep ships by default. New session machines come with rg preinstalled, so in-session searching is fast out of the box.
• The context gauge came to the web. Web chat now shows how much of the model’s context window a conversation is using, and a completion-alert badge lights up the browser tab when a reply finishes while you are elsewhere, clearing the moment you come back.
• Settings stay in their lane. Changing your provider, model, or skillset from the web chat Settings panel now applies to your current conversation only, instead of quietly rewriting shared org or fleet defaults.
• Small comforts. The activity heatmap scrolls horizontally and opens on the most recent dates, the Preview pane opens to the live terminal when no web app has been detected yet, and there is now a self-service “forgot password” option on the login page.
• Integrations hold their credentials. Slack tools now stay connected for the full life of a session, survive going idle, and reconnect cleanly on restart, and scheduled-trigger Slack delivery got clearer guidance plus tidier handling of channel IDs with stray whitespace.

Slack Sets Itself Up

• Self-service Slack, at last. You can now connect Slack to your org yourself through a guided setup and lifecycle flow, rather than filing a ticket and waiting for a human with the right buttons. It took a revert and a re-landing to get there, which is the software equivalent of measuring twice.

Headless Mode Arrives

• nori exec runs without the UI. Run an agent as a one-off command — nori exec "prompt" or piped through stdin, with the response printed cleanly to stdout — or speak to it as a plain ACP-over-stdio interface for programmatic callers. The interactive terminal is now optional, not mandatory.

The /browser Command Picks Its Profile

• Three Chrome profiles, your call. /browser now lets you choose a fresh throwaway profile (the secure default), a persistent Nori-owned profile that stays logged in across launches, or your real system Chrome profile with all its existing logins — convenience and isolation as a deliberate choice rather than a default you discover later.

Compact Natively, Fork at the Head

• /compact forwards natively. On agents that support it, compaction happens in place — no session swap, no re-injecting a summary.
• Branch from where you are. /fork gains a capability-gated “branch from current point” option that forks the session at its head while keeping the original resumable, and the branched-from conversation becomes a genuinely separate thread with a copy-pasteable nori resume <id> hint and a “forked from” row back to where you were.

A Footer You Can Shape

• A compact context readout. The default indicator is now a tidy used% / window-size (for example 44% / 272k), and you can fully customize both the footer and the textarea-corner indicators, mixing built-in segments with your own {format} chunks.
• The /status card tells the whole story. It now always shows the full session picture regardless of your footer settings — session ID, working directory, git status, agent mode, skillset version, and a single consolidated context line.
• Fewer interruptions, fewer glitches. The config and settings pickers return you to the panel with the cursor on the row you just changed, session metadata from the agent renders with friendly labels in live and view-only transcripts, turns started by another connected client render instead of warning, and a display-corruption bug that could scramble the live view after a large completed turn is fixed.

Skillsets Go Git-Native

This was the week the Skillsets CLI stopped treating Git as an afterthought and made it the substrate. A skillset can now live its whole life — created, installed, forked, published, updated — as an ordinary Git repository.

• Install straight from a remote. sks install <slug> --from <remote> clones the source’s branch tip directly instead of routing through the Registry, and --pin <full-sha> installs an exact historical commit when you want a fixed point rather than the latest.
• New skillsets are real repos. sks new now creates a skillset as its own local Git repository, and Registry publish and download refuse to touch a Git-governed skillset — one source of truth, not two.
• Fork and publish. sks fork <existing> <new-slug> copies an installed skillset into a fresh, independent repo with no inherited history or provenance, and sks publish <skillset> --to <remote> lets you review the staged package and push it to a remote.
• Update, safely. sks update <slug> fast-forwards a Git-backed skillset to its latest branch tip and re-activates it, and refuses to run on a pinned, dirty, or diverged checkout rather than clobbering your local state.
• Default remotes and remembered trust. Set a default with sks config --primary-remote <remote> so bare-name installs resolve there, and approved sources are now remembered durably — with sks trust list and sks trust revoke to manage them — so you are not re-prompted on every install.
• Installs that clean up after themselves. Installing or switching is now transactional, so a failure partway through rolls back to your previous working state (even after a crash), and a lock ensures concurrent installs run one at a time instead of corrupting each other.

Registrar Stops Crying Conflict

nori-registrar shipped one honest bugfix: re-uploading a skillset no longer flags every unchanged skill and subagent as a conflict. Identical content is now correctly recognized as identical, instead of presenting you with a phantom conflict that has no diff to show for itself — the software equivalent of an argument about nothing.

Week Thirteen

nori-lint merged zero pull requests for the thirteenth consecutive week. Last week it completed a full fiscal quarter of stillness; this week it began the next one exactly the way a professional begins anything, by doing more of it. There is a school of thought that the most reliable software is the software that does not change. The linter has become the leading empirical champion of that school, and it is not even trying.