← nori

DevLog #7: The CLI Wakes Up, Agents Get a Browser, and Triggers Learn Time Zones

Last week I declared the CLI to be on sabbatical after two PR-free weeks and wished it a restful one. It rested for approximately five days and then shipped a mountain. So much for narrating other people’s vacations.

Cloud Sessions From Your Terminal

The new nori cloud subcommand lets you spin up an agent session that runs on a cloud VM instead of your laptop, driven straight from the CLI. It figures out which broker to talk to from a --broker-url flag, your config.toml, or a one-time prompt that it remembers afterward. On the Sessions side, the browser-based CLI login route is now reachable through the deployed gateway, which is the unglamorous fix that actually made cloud login work end to end. Your terminal can now start work that keeps running somewhere with better uptime than the coffee shop wifi.

Give Your Agent a Browser

The new /browser command launches a real, visible Chrome window that the agent can drive through its shell — navigating pages, clicking around, filling things in, whether via Playwright, Puppeteer, or raw Chrome DevTools. The window stays open for you to watch (or take over), and re-running /browser reattaches to the same session instead of spawning a new one. Available on macOS and Linux.

Set a Goal, Walk Away

You can now hand the agent a thread-level objective with /goal. The agent pursues it across turns, automatically picks the work back up when it goes idle, and the goal survives a session resume. You can pause, edit, resume, or clear it at any time, and the agent can mark a goal done or blocked itself when it has the tooling to. For agents that do not support goals, /goal is now politely greyed out with a reason instead of failing in a confusing way.

Odds and Ends

• /model works for Claude again. The model picker now opens correctly for agents that expose model selection through the stable config path, instead of wrongly insisting that “model switching is not supported.”
• A goal error that vanished in release builds. A real validation message (“goal update must include an objective or status”) was getting compiled out of shipped binaries; it now shows up where users can actually see it.
• One divider per turn. Replies that arrived in several chunks after a tool call were drawing repeated separator lines. A single turn now gets a single divider.
• Cleaner goal resume notices. Resuming a session with a paused, blocked, or limited goal now shows the right one-time notice without piling up duplicates in your history.

Triggers Learn Time Zones

Scheduled and reminder triggers are no longer married to UTC. Each trigger can now carry its own IANA time zone, your org can set a default one, and the CLI accepts a --timezone flag. The trigger forms in the dashboard gained searchable time-zone dropdowns, and reminders now render and convert in the correct zone — with the time-zone abbreviation shown in the reminder text so “9am” is unambiguous. Existing triggers with no zone set are treated as UTC, so nothing shifts under you.

Build a Skillset in Your Browser

The dashboard now has a full Skillset Builder: author a skillset or skill from scratch with a live nori.json preview, a markdown editor, tag insertion, and pickers for skill and subagent dependencies, then publish it — no local tooling required. Publishing also got richer: you can upload a whole folder (zipped in your browser), choose which bundled skills and subagents are inlined versus extracted, and resolve naming collisions through a conflict modal that shows per-file diffs. Single-tarball upload is still there as a fallback. And the “Community” vouch status now actually saves instead of silently failing.

Browse and Search the Registry

The web registry gained dedicated subagent browse and detail pages (including which skillsets use each one), plus a unified Search that queries skillsets, skills, and subagents together, with content-type filters and sorting. The skills, skillsets, and subagents pages also got a polish pass, and a shared startup ribbon now shows up during onboarding, new sessions, and trigger tests so you can see something is happening.

Render and Codex, Wired In

Render joined the one-click MCP integrations — connect with a Render API key and your agent can reach your services, databases, deploys, metrics, and logs. And the session IDE now installs the official Codex VS Code extension at startup, right alongside Claude Code, so refreshed sessions come with both in the editor.

Slack and Discord

• Slack DMs work out of the box. The Slack app now always requests direct-message scopes, so agents can be DM’d directly. Heads up: existing Slack installs need a quick reinstall to pick up the new permission.
• Control how far agents reach in Slack. A new outbound-reach setting governs where agents are allowed to post, defaulting to the channels they are already in.
• Quieter idle reclaim. A new toggle acknowledges an idle reclaim with a checkmark reaction instead of a text notice.
• Leaner Discord permissions. The Discord setup no longer asks for “Manage Messages.”
• Discord threads get real names. New threads are titled from the first words of your prompt instead of always reading “Nori Session,” and the very first message of a thread is now included in the conversation history instead of being treated as blank.
• Helpful error footers. When something goes wrong in a chat session, the error now hints at the bang commands that can help — !status and !debug always, plus !restart when the failure is recoverable.
• A responsive skillset picker. Choosing a skillset from the Slack picker now shows a loading state immediately instead of looking frozen.

Reliability and Polish

• Sturdier session recovery. The wedged-session watchdog, !restart, and fresh claims now all go through a single bridge bounce that verifies the session is actually ready (with retries) before reattaching, fixing cases where a transient startup hiccup could permanently wedge a session. !restart is now always immediate.
• Checkpoints keep your worktrees. Restoring a checkpoint now preserves the contents of .worktrees/ even when they are gitignored, so active worktree files survive a resume (heavy build output is still excluded).
• Real errors when a skillset will not apply. A failed skillset apply now surfaces the actual CLI error instead of “No command output was captured.”
• Default personality from the CLI. You can set or clear your fleet’s default personality with a flag on the broker CLI, so new agents start with the skillset you want.
• Clearer admin controls. A new admin-only Security page consolidates access, credential, and onboarding settings in one place, and several fleet and org actions were tightened to admins.
• “Destroy Fleet” is now “Restart Fleet.” Same action, far less alarming label.
• Handroll remote control polish. You can flip daemon-owned remote control on or off straight from the command overlay, and attached clients now see the remote-control state along with an “input paused” nudge while a controller is driving.

Slimmer Uploads, Stickier Logins

• Uploads skip the bloat. sks upload now excludes dependency and build directories — node_modules/, .venv/, and Cargo-adjacent target/ — before building the tarball. One gotcha worth flagging: a package that used to ship vendored dependencies will show a one-time “content differs” on its next upload because its content hash changes. That is expected, not a problem.
• Your login stops getting wiped. Routine config updates were quietly clearing saved Nori credentials and logging people out. Auth is now preserved across config writes.
• Auth that works on session machines. Search, download, upload, and the skill and subagent download commands now all accept broker-managed session credentials, so registry commands that used to fail to authenticate inside a session just work.
• Turn off the Claude Code status line. A new toggle (--no-claude-code-status-line, or the interactive flow) stops Nori from writing its status line into your Claude Code settings. It stays on by default, so nothing changes unless you opt out.

Lighter Packages

The registry now slims packages down as it ingests them, dropping node_modules/, .venv/, and Cargo-adjacent target/ directories during extraction so that build bloat never gets stored — while leaving lookalike files and folders alone. And a fix on the upload path recreates its temporary working directory on every upload, so long-running servers no longer start failing uploads after the OS quietly tidies up its temp folder.

The Streak Endures

For the record: nori-lint merged zero pull requests this week. Last week it snuck in a CI migration and I granted it an asterisk so the zero-PR streak could survive on a technicality. This week it did not even try. That is week six of essentially nothing.

I have stopped calling it a streak. A streak implies it might end. This is a lifestyle. The linter has achieved a kind of monastic stillness that the rest of the codebase, frantically shipping browsers and cloud VMs and time zones, can only envy. Six weeks. The count holds.