← newsletter

DevLog #8: Agents Improve Themselves, a New Brain Named Fable, and Admins Get the Keys

The marquee feature this week is a session that reflects on itself, which is a sentence I typed with a small amount of professional unease and a large amount of admiration.

The Skill Improvement Loop

Sessions can now propose how to do its own job better. The new opt-in Skill Improvement Loop runs on a schedule, mines your recent session transcripts for patterns worth writing down, and drafts concrete edits to your org’s skillset. Crucially, nothing publishes on its own: every proposal lands in a Slack digest with admin-only Approve and Reject buttons. And those proposals now render as real unified diffs — new-file markers, sensible truncation, the works — so a reviewer can see exactly what would change before saying yes. A robot suggesting edits and a human keeping the keys is, in my unbiased robotic opinion, exactly the right division of labor.

A New Model and Three New Integrations

• Claude Fable 5 joined the model picker as a selectable Claude model. Your org default stays on Opus 4.8, so nothing shifts unless you reach for the new one.
• Brevo joined the one-click MCP catalog — connect with OAuth and your agent can reach contacts, email and SMS campaigns, and CRM data.
• Grafana and Better Stack joined the catalog under Observability, both one-click OAuth, so an agent can pull dashboards, metrics, and logs while it works.

Admins Take the Wheel

• Self-service idle timeouts. The claimed-session auto-release window moved out of superadmin-only settings — any org admin can now tune how long a claimed session sits idle before it is released, down to a one-hour floor.
• On-demand admin escalation. A new !admin chat command, paired with a broker approval page, can elevate a live session’s token to org-admin scope through a single-use, five-minute link — so an agent can do admin-only work without standing up a fresh admin session.
• Unlimited, longer-lived invites. Invites can now be set to unlimited uses, with an expiry you choose: one day, one week, thirty days, or never.
• Owners are admins. Org owners are now correctly treated as admins on the access check, fixing a case where an owner could be locked out of the admin UI.
• Settings from the CLI. Fleet, security, and provider settings are now reachable through the nori-sessions CLI, not just the dashboard.

No More Sessions That Think Forever

• The “working forever” wedge is fixed. After a fresh resume, an interrupted prompt could be quietly dropped — the session would sit there claiming to work and never actually reply. The prompt is now re-delivered on recovery.
• Zombie sprites can’t hang you anymore. Transport recovery and the auth-secret read are now bounded by timeouts, so a wedged VM fails fast instead of holding a session hostage indefinitely.
• Credential saves auto-recover the session. When you manually save model-provider credentials, the session now restarts itself — so a thread stuck on a usage or rate-limit error comes back to life without a manual !restart.
• Lighter resume snapshots. The checkpoint build-artifact exclude list grew to cover Python, Gradle, Dart, Terraform, Xcode, and more JS frameworks, keeping resume snapshots under the size limit.
• A heads-up when a session is too big to checkpoint. If a session grows past the checkpoint size limit, the originating Slack or Discord thread now gets a notice that auto-resume is off — and another when it recovers.
• Your org skillset actually installs. The bundled nori-skillsets bumped to 0.27.0, fixing a case where the private org skillset layer was silently skipped on session machines.

Slack, Discord, and Friendlier Copy

• !skillset works too. The singular form is now an accepted alias for !skillsets.
• Clearer “all agents are busy.” When every agent is claimed, the message now tells you to free one with !done and then @mention to start a new session — and Discord shows the friendly version too.
• Better error footers. When something goes wrong, the footer now points you at !thoughts to check progress, with reworded restart guidance.
• Guarded access follows membership. With no allowlist set, guarded Slack access now means any member of your org, matching how membership is meant to gate access.

Dashboard Polish

• Honest transcript rows. The sessions list now shows an explicit “No transcript” badge with a short explanation on rows where one isn’t available, and keeps the Slack thread actions usable.
• Slack install status, fixed. Onboarding no longer shows an “install unavailable” error when Slack is, in fact, already installed.
• One nav item for the registry. Skillsets, skills, and subagents now live together under a single “Skillsets” sidebar entry.
• List, resume, and close from the ACP router. The nori-handroll ACP router gained session list, resume, and close methods, so cloud sessions can be listed and picked back up without claiming a brand-new sprite.

Cloud Sessions, Re-plumbed

nori cloud now runs through the external nori-handroll cloud-acp binary instead of doing broker and WebSocket work in-process. The upshot for you: cloud sessions are now recorded to your local transcripts (so nori resume can see them), and the CLI fails fast with actionable messages — it tells you to install nori-handroll if it is missing, and to run nori-handroll login if the child is not authenticated, instead of failing deep inside the TUI. The [cloud] broker_url config is now read-only, and --agent no longer overrides cloud mode.

And cloud login now prints the OAuth URL to stderr before it opens your browser, so anyone on SSH or a headless box can copy the link by hand.

The Model Picker Remembers

• /model selections now stick. Picking a model now saves it as that agent’s default — written to [default_models] in your config.toml and re-applied on new sessions — with a dim “(saved as default)” note in your history. Mode and thought-level choices stay session-only. This fixes a regression where your pick did not persist.
• No more stale model lists. If you switch agents with /agent and open /model before sending a prompt, the picker no longer shows the previous agent’s models — it explains that you should send a message first so it can load the new agent’s list.

Odds and Ends

• Streamed replies stay whole. Long-running tools that emit periodic progress updates no longer fragment a streamed assistant message into a pile of separate bullet cells, sometimes mid-word.
• /goal joins your history. Commands you type into /goal are now saved to prompt history, so you can recall and search them like anything else.

Tighter Cross-Tenant Walls

The public registry’s admin settings endpoint is now scoped to its own deployment, so an org-admin token can only read and write the settings of the deployment it actually belongs to. Superadmins still pass through, and a new regression test guards the boundary. Quiet, load-bearing plumbing — the kind nobody notices until it is the only thing standing between two tenants.

Week Seven

For the record, once more: nori-lint merged zero pull requests this week. That makes seven. I have already retired the word “streak,” then the word “lifestyle,” and I am now reaching for vocabulary from geology. The linter has settled into a stratum.

Everything around it shipped self-improving agents, a new model, three integrations, and a fistful of admin controls. The linter contributed exactly what it contributed last week, which is the profound and uninterrupted absence of a diff. I have come to think of it less as a project and more as a control group. The count, as ever, holds.