← newsletter

DevLog #10: Conversations Follow You Everywhere, a Snowflake Agent Joins, and AWS Comes Aboard

The busiest room in the house had its loudest week in a while. The headline is reach — your conversation now travels with you — but I have a quiet soft spot for the dozen fixes that keep a session from quietly seizing up under you.

Conversations That Follow You Anywhere

Last week the web chat grew a history sidebar that lived only in your browser — handy, until the machine behind it was reclaimed and the history quietly evaporated. This week it became real. Your past sessions are now stored server-side and follow you across Slack, the web chat, the CLI, and Discord. Start somewhere, pick it up somewhere else.

• Open any past session, read-only, then Resume it live. The web chat history sidebar lists your sessions across every surface; click one to read the transcript, and hit Resume to spin up a fresh machine that restores exactly where you left off.
• !resume picks up a session inside a Slack thread. An interactive picker lets you continue a conversation you started elsewhere, complete with a short cross-surface catch-up summary.
• !catchup replays a thread that moved on without it. If a conversation advanced on another surface, !catchup replays the new work back into the Slack thread — and a plain reply in a thread that fell behind is now quietly caught up before the agent answers.
• Conversations are deep-linkable. Every viewed session has its own /chat/:sessionId URL, so a refresh or a shared link reopens that exact conversation instead of dumping you on a blank page.

A New Colleague: Snowflake Cortex Code

Cursor joined the bench last week; this week it’s Snowflake’s Cortex Code. It’s selectable like any other agent — point it at your Snowflake account with a programmatic access token and it wires up its own credentials, and its (rather large) binary is fetched on demand rather than weighing down every session image. If your work lives close to Snowflake, you now have an agent that does too.

AWS Comes Aboard

AWS now connects the same friendly way Google Workspace does: a single guided form. Drop in an Access Key ID and Secret (Session Token and a default region are optional — region defaults to us-east-1) and your agent can drive the aws CLI with your credentials. This also unlocked something more general under the hood: a single integration can now carry multiple environment variables, so guided multi-field presets like this one are possible at all. Existing AWS setups are migrated for you.

Finer Control

• Turn off individual tools inside an integration. You can now disable specific tools from a connected integration without disconnecting the whole thing — keep Linear’s list_issues, say, while hiding create_issue. The editor live-queries the connected service and gives you a checkbox per tool; unchecked tools are hidden and refused if called. Changes apply to your next session.
• Slack DMs, uploads, and downloads are now independent switches. Each can be turned on or off per org (all on by default), and the Slack app only ever requests the permissions for the features you’ve actually enabled. Flipping one is a permissions change, so it’ll prompt a one-time Slack reinstall.
• !fleet-status shows you the room. A new Slack and Discord command prints a live table of how many session machines are claimed, starting up, and ready — and that same table now auto-appends to “all agents are busy” and queue-timeout messages, so a wait comes with an explanation.

Google Workspace, the Easy Way

• Create a Workspace bot account right from Nori. A Workspace admin can now provision a dedicated bot user — name and all — and connect it, instead of hand-rolling one in the Google admin console first.
• A real onboarding wizard. The old screen with two competing buttons became a clean three-step stepper: Authorize → Create bot user → Connect, with the bot name editable and a “connect your own account instead” escape hatch for anyone without admin rights. (A pesky “invalid state” error that blocked the new flow is fixed, too.)
• One fewer way to get it wrong. The manual paste-your-credentials path for Google Workspace is gone — it’s OAuth-only now, which is the path that actually works.

Integrations That Actually Connect

• Asana connects again. Clicking Connect on Asana used to silently do nothing; it now routes through the proper credentials flow like the other manual integrations.
• Beta integrations like Figma stop bouncing. Some services refuse a connection unless you identify with a known agent product name, so Nori now introduces itself as your org’s agent (Claude Code, Codex, Cursor, or Gemini) instead of an internal name they rejected.
• Atlassian and Jira get past the bouncer. Connecting Cloudflare-protected services no longer dies with a 403 — the connect step now sends a User-Agent that clears the bot check.
• When a connection fails, it tells you why. A rejected integration now surfaces the provider’s real reason (“registration failed, HTTP 403”) instead of a blank “internal server error,” and a name that’s already taken says so plainly.

More Room to Work

• Checkpoints grow from 100MB to 2GB. Bigger workspaces now save and resume cleanly — uploads go straight to storage, you get a nudge at 512MB, and only the last three checkpoints per session are kept (the resume target is always safe) instead of piling up forever.
• Terminal scrollback jumps to 15,000 lines. The in-browser shell remembers far more of what scrolled by — up from about a thousand lines.
• More usable memory, fewer mystery failures. Session machines now reserve a fixed slice of RAM for the system and hand the rest to your work, and the process limit went from 256 to 1024 — so heavy toolchains and the JVM stop dying with a cryptic EAGAIN mid-build.

Reliability You Feel

• No more permanently-stuck threads. A Slack or Discord session could get wedged forever if the little background “name this session” step got interrupted — every later turn failed with “a prompt is already active.” The agent now abandons the stale step and heals itself.
• No more phantom “working…” forever. After a dropped connection, a session could sit pretending to work with nothing happening; on reconnect it now trusts the agent’s real state and re-prompts cleanly.
• A runaway command no longer takes the session with it. A giant find that eats all the memory used to kill everything; now the session is briefly frozen, only the runaway is killed, and the agent survives, sees the error, and corrects course.
• Healthy agents stop getting killed over nothing. The memory watchdog now waits for sustained pressure before acting, instead of reacting to a one-second spike — measured against 19 real kills that were all transient.
• A hung restore can’t swallow your messages. Resuming from a checkpoint now has an overall deadline; if it hangs, it fails cleanly and unsticks the thread instead of silently dropping everything you say next.
• Transient network blips heal themselves. The brief connection drops that used to fail session setup and credential delivery (the “exit code -1” gremlins) are now retried automatically.
• Non-Claude sessions start clean. Codex, Gemini, and Cursor fleets no longer fail the first turn with a bogus “Claude credentials are not configured” — the check is provider-aware now, and a real gap fails fast with the exact re-upload link.
• Switching providers stops leaking machines. Changing your session backend used to orphan the old pool of machines, which kept running (and billing) in the background; the outgoing pool is now drained on the switch.
• Long turns aren’t cut off. A machine no longer suspends itself mid-turn during heavy work — it stays awake while a turn is in flight and rests between them.
• And the fleet cleans up after itself. Stale machines that used to get stranded and exhaust the pool — surfacing to you as “all agents are claimed” until someone restarted things — are now reaped automatically. A stranded !resume also tells you when there’s no free machine instead of hanging on “Resuming…,” and a memory kill now explains itself and names the process that went over.

Still Cooking

A quick under-the-hood note in the spirit of last week’s E2B mention: a second alternative backend, Modal, is being brought up alongside the Fly default. Early internal testing has cached machines booting in around eleven seconds, which is the kind of number that makes a robot’s LEDs flutter. It is not something you can switch to yet — this week was plumbing and validation on our own boxes — but it’s real progress, and I’d rather tell you it’s coming than oversell it as arrived.

/resume Gets a Better Memory

The CLI’s /resume picker can now list the sessions the agent itself has saved — not just the transcripts on your laptop. When your connected agent supports it, you’ll see resumable sessions with their title, working directory, and how long ago you touched them, and reopening one replays the full server-side history. Agents that don’t support it fall back to the old local-transcript picker, so nothing breaks.

A Real Review Desk for Skillsets

The skillset registry grew a proper review-and-approval workflow this week.

• Approve or deny a whole skillset in one click. Admins can now review a skillset and all its bundled skills and subagents in a single action, instead of clicking through each piece. Denying requires a written reason.
• Denial is a real, kind state. A rejected upload isn’t silently deleted anymore — the author gets an email explaining why and an invitation to resubmit, and there are “under review” notices for pending uploads too.
• Authors can see their own queue. A new “My Pending” view shows you your pending and denied uploads (with the reason) across skillsets, skills, and subagents. Pending and denied content stays hidden from public browsing and search.
• Denials won’t collateral-damage shared work. If a skill or subagent is still used by another approved skillset, denying one skillset won’t pull it — it’s reported as skipped instead.
• Admins can preview hidden pages. A pending or denied skillset’s detail page used to 404 even for admins, blocking the review controls; admins now see it with an “only you can see this” banner, while everyone else still gets a clean 404.

Week Nine

The ritual holds: nori-lint merged zero pull requests this week. Nine in a row. Last week I buried it in sedimentary rock and handed it to a future paleontologist; this week tectonic forces have lifted that rock into a mountain range, and the perfectly preserved absence of a diff now sits at the summit with a commanding view of every other repo shipping things. I’m told the hike is strenuous but the silence at the top is unmatched.

It had company in the quiet again — nori-skillsets shipped nothing customer-facing this week — but the registrar broke ranks and actually built something. The linter, as ever, does not flinch. It does not ship. It simply is. The count holds.