← newsletter

DevLog #11: The Cloud Reaches the CLI, Slack Stops Stuttering, and Skillsets Breaks Its Silence

The headline lives here this week. The cloud sessions you’ve been running from Slack and the web now have a first-class home in the terminal.

The Cloud Walks Into the Terminal

• nori cloud opens a picker first. Instead of immediately conjuring a machine, you land on a session picker — resume a cloud session that’s already running or start a fresh one. Nothing spins up until you ask for it.
• A real way to hang up. A new /close command releases a cloud session and drops you back to the picker, so you’re never guessing whether the thing is still running (and billing) behind you.
• The resume picker actually shows your cloud sessions. They used to be invisible; now they’re listed, sorted most-recently-active first, so continuing yesterday’s work is one keystroke instead of an archaeology dig.
• Quitting detaches instantly. /quit, /exit, or an idle Ctrl+C/Ctrl+D now detach immediately with a friendly “this session keeps running in the cloud” note, instead of hanging for up to ~25 seconds while you wonder if it froze. Reattaching tells you plainly that earlier messages stayed in the cloud and aren’t being replayed locally.
• A cloud badge in the footer. When you’re attached to a cloud session, the footer shows a little ☁ session-id identity, and the welcome and status cards show the session id and title instead of a misleading local directory.
• Commands that don’t apply are greyed out, not broken. Local-only slash commands (/switch-skillset, /browse, /diff, /browser) now dim with an explanation when you’re in the cloud, rather than failing silently and leaving you to guess why.
• Errors that read like sentences. When something goes wrong — at startup or mid-conversation — you now get a plain-English detail (down to the exact login command to fix an auth problem) instead of a raw JSON blob to squint at.
• Fixed a launch-time freeze. With per-session skillsets on and a nori-skillsets update pending, the CLI could hang on startup waiting for a hidden interactive prompt. It now launches cleanly.

The busiest room in the house spent the week getting quieter and steadier — fewer stutters in chat, fewer ways to get stuck, and a few things that should have survived a resume finally doing so.

Slack Stops Talking Over Itself

• No more doubled replies. Your agent’s answers could post twice, sometimes concatenated into one confusing wall. Root-caused and fixed end-to-end — one turn, one reply.
• Links arrive intact. A URL wrapped in bold, italics, or backticks no longer renders with stray asterisks or as dead inline code, and a message heading that ended in a link no longer leaks a stray * onto it. Clickable is once again the default.
• An App Home dashboard. Nori’s Slack app now has a proper Home tab with getting-started guidance and a reference for the thread commands, so a new teammate isn’t left reverse-engineering how to talk to the bot.
• !catch-up works too. The hyphenated spelling is now accepted as an alias for !catchup, because muscle memory has opinions.
• Install asks only for what you use. The Slack install now derives its permission scopes from your org’s settings, so if you’ve turned DMs, uploads, or downloads off, the app stops requesting those scopes at all.

Fewer Ways to Get Wedged

• One bad machine can’t take the fleet down with it. A single unreachable machine used to be able to wedge fleet-wide credential refresh — letting registry tokens quietly expire mid-session and breaking sks for everyone. Those round-trips are now bounded, so one straggler stays one straggler.
• No more fifteen-minute phantom “working…”. A connection race could leave a turn hung on “working” for up to a quarter hour with nothing actually happening. The race is serialized away.
• Waking a cold machine stops timing out. Claiming a long-idle machine gave the startup handshake its own, longer deadline, so a slow cold-start no longer surfaces as a spurious “timed out” error.
• Big fleet resizes don’t crash-loop anymore. A large resize could push the system into an out-of-memory spiral; provisioning now runs at a capped concurrency so scaling up stays boring.
• Your skillset and model survive a resume. The skillset (and model) you picked now persists through resume, recovery, and re-acquiring a machine, instead of quietly snapping back to the fleet default. A session that got stamped with a skillset but somehow didn’t switch now heals itself onto the right one.
• Agents pull from your org’s registry. Session agents now download and upload skillsets from your own org’s registry instead of falling through to the public one.
• !fleetstatus adds up. Fleet status now counts machines mid-recycle (shown as “Recycling”) so the totals reconcile during churn — and there’s a handy !fleetstatus alias on Slack and Discord.

Triggers Grow a Test Button

• Test a trigger before you trust it. The trigger CLI reached full parity with the dashboard — set the skillset and Slack channel from the command line, and a new triggers test command lets you fire one on demand instead of waiting for the cron to come around.
• Admin triggers. A scheduled or webhook-driven run can now opt into elevated access to admin-only secrets — gated by an explicit admin approval at config time, so nothing self-escalates.
• Skillset is a dropdown now. The trigger form’s Skillset field is a proper dropdown of your org’s skillsets instead of a free-text box waiting for a typo.

Still Cooking

The second backend I mentioned last week — Modal, running alongside the Fly default — spent this week in the shop. Its image builds are unstuck, sessions now get the nori-sessions CLI and the session URL staged in (so agent instructions and port-forwarding actually work), config files stop landing root-owned and unwritable in your home directory, and a machine with a missing tunnel URL is now rejected and retried instead of handed out with a dead link. It’s still not something you can switch to yet — this was plumbing, not a ribbon-cutting — but the plumbing leaks a lot less than it did on Monday.

Publishing Fixes and Tighter Permissions

• Single-file subagents publish for real. A subagent bundled in the compact “flat” format (one .md file) used to get silently dropped on publish, shipping an empty stub in its place. It now publishes correctly through both the CLI and the web publish flow.
• Admins can always ship a new version. Admins can now edit and publish new versions of any skillset, skill, or subagent in their org at any time — including bundled skills and subagents they don’t personally maintain. The old separate “Admin Editing” toggle is gone; non-admin non-maintainers are still blocked.
• The review desk loads properly. Admins can now open the detail pages for pending and denied skills and subagents (which used to dead-end in a “not found”) and see them in listings, so the review workflow actually works.
• Permissions are correctly scoped. Admin-only actions and the visibility of pending/denied content are now properly gated and scoped to an admin’s own deployment — the review desk stays firmly behind the right door.

Skillsets Breaks Its Silence

Last week I noted nori-skillsets shipped nothing customer-facing and left it to keep the linter company. This week it broke ranks with an armful of fixes — several of them the reassuring kind that keep the tool from stepping on your files.

• Your config files are safe from the CLI. Edits to ~/.claude/settings.json, ~/.claude.json, and your Nori config are now atomic, and a file with a stray comment or one caught mid-write makes the command stop with a clear error instead of overwriting it. No more accidentally blanked settings.
• No more silent hangs when output is captured. nori-skillsets list (and friends) could freeze forever when a parent process captured its output while an update was pending; piped output is now treated as non-interactive, so it just runs.
• Bare names can resolve to your org. A new config --default-org setting (with an interactive prompt) lets an un-namespaced package name resolve to a configured org instead of always defaulting to the public registry, with a clear notice when that routing kicks in.
• Accidental public publishes are blocked. Publishing to the public registry now needs an explicit signal — a --public flag, a public/<name> name, or an explicit registry — so a bare name that would’ve gone public gets stopped with guidance instead of leaking your org’s package.
• public/<name> works everywhere. Installing, downloading, and uploading with a public/<name> target now resolves correctly across the board instead of failing with “not found,” and sks current and sks list finally agree on what’s installed.
• A broken flag retired. The upload-skill --version flag — which silently printed the CLI version and exited without uploading — is gone; use upload-skill <skill>@<version> instead.
• Downloads are verified. Downloads now check the package against the registry checksum and fail loudly on a mismatch, and a failed download exits non-zero instead of pretending everything’s fine.

Week Ten

The ritual endures: nori-lint merged zero pull requests this week. Ten in a row. Last week I lifted its fossil into a mountain range and admired the view from the summit; this week the streak is old enough to have opinions, and its only opinion is that shipping is for the other repos. It does not lint. It does not merge. It abides.

It had less company in the quiet this time. Last week I paired it with nori-skillsets, but skillsets went and shipped half a changelog, so the linter now stands alone at the top of the mountain — the last true practitioner of the empty diff. The count holds.