← newsletter

DevLog #19: Fire One Trigger When Another Finishes, Run OpenCode or pi, and Sign In to Codex from the CLI

Fire a Trigger When Another Trigger Finishes

• Run a trigger off another trigger’s outcome. Pick “Another trigger” as the type in the dashboard trigger editor, name the upstream trigger, and choose which outcome fires it: completed for any outcome, succeeded, or failed. A cleanup or alert trigger that runs only when your nightly job fails is now three fields in a form. Until this week a trigger could only be started by the clock, a webhook, or a reminder.
• Read the upstream result inside the downstream run. The prompt that starts the downstream trigger names the upstream trigger, says whether it succeeded or failed, and carries the upstream run’s response text or error text. That text is fenced as untrusted data with instructions to the agent to treat it as data, because the upstream agent may have pulled it from an external system.
• Wire it from the command line. nori-sessions triggers create --type internal-event --source-trigger-id <id> --source-event failed creates the same thing, and the REST API accepts the same two fields. A trigger cannot name itself as its source. Chains of two or more triggers pointing back at each other are accepted, so read your wiring before you enable it.

Keep a Scheduled Run Through a Broker Restart

• Keep a firing that landed during a restart. A cron tick, a reminder, and a webhook delivery each write a durable queue row at the moment they fire. That row survives a broker restart, a crash, and a moment when every machine is busy. After a crash the broker catches up on firings from the last hour, once each. Repeated ticks for the same trigger collapse into one pending row rather than piling up.
• Get told when a firing gives up. A firing that fails before a machine accepts it retries three times with a 30 second backoff between attempts, then posts a failure notice into the Slack or Discord channel the trigger was created from.
• Watch the queue on the fleet page. The Task Queue card lists waiting trigger firings with the trigger name, attempt count, how many ticks collapsed into the row, and the session that was ended to make room. You can remove a row that is still waiting.
• Choose what happens when the fleet is full. Every trigger defaults to a capacity policy that ends the oldest session idle for at least 10 minutes and takes its machine, at most one per firing, and posts a note in that session’s thread saying it was ended. Set the policy to wait on a trigger that should sit in the queue until a machine frees up on its own.
• Send webhooks without filling the queue. POST /webhook/:triggerId returns the firing id it created. A trigger that already has 10 deliveries waiting returns 429 on the next one.

Run OpenCode or pi, and Pick Claude Fable 5.1

• Select OpenCode or pi as your coding agent. Both are now in the provider list for every organization. Each takes three values on its card in Integrations: an OpenAI-compatible base URL, an API key, and a model id. Both install on the session machine at launch, so the first session start after you select one takes longer than Claude, Codex, or Cursor, which are already on the machine.
• Choose Claude Fable 5.1. claude-fable-5-1 is in the Claude model catalog, which means it appears in the web chat model picker, the Slack !model command, the CLI, trigger settings, custom agents, and fleet defaults. Your organization default is unchanged. One thing to know: Claude Code rewrites a stored claude-fable-5 to a floating alias that now resolves to Fable 5.1, so a fleet, trigger, or Slack agent pinned to Fable 5 runs Fable 5.1 after this update reaches your broker.
• See which agent version your sessions run. The Claude and Codex cards under Token providers now lead with the coding harness version and show the ACP adapter version in parentheses, for example v0.3.257 (ACP: 0.73.0). The card used to show only the adapter number, which is not the version a session runs.

Find Every Integration on One Page

• Set up everything from Integrations. The page is one ordered list: connected integrations, token and model providers, Nori-provided integrations, MCP servers, then your own. Turning a provider on, setting its default model, and choosing which models your organization can pick all moved here from Fleet Setup.
• Recognize an integration by its logo. Slack, GitHub, Google Workspace, Discord, AWS, Luma, Snowflake, Salesforce, and all 67 entries in the MCP catalog now carry artwork, on the setup cards, the configured rows, and the connected summary.
• Get around with three sidebar entries. The sidebar is Chat, Triggers, and Integrations. Organization settings, users, security, and billing are one scrollable Settings page. Agents, Library, and Builder are grouped under Context, and Library is a single search across your skillsets, skills, and subagents. Your old bookmarks redirect to the new locations.
• Open a dashboard page in a new tab. Sidebar entries and session cards are real links now, so command-click, middle-click, and “copy link address” work on all of them.

Read the Real Reason Something Failed

• See why a queued task never started. A task that could not be started used to show one fixed line asking you to mention the agent again. It now names the cause when the cause is known, and expired provider credentials come with the reconnect link. Unrecognized failures keep the general message, so raw internal error text never reaches your thread.
• Get readable agent errors in Slack, Discord, and web chat. Errors the coding agent returns are classified and rendered as plain copy with a link to Integrations where that is the fix. The raw details are preserved for anyone who wants them.
• Find out a Cursor key is bad when you paste it. A Cursor key is now checked against Cursor before it is saved, the same as Claude and Codex. A rejected key fails the upload with an explanation. A Cursor outage never blocks the save. If a working key is revoked later, the session failure says to reconnect at Integrations rather than showing Failed to initialize session services.
• Resume a Codex session whose history moved. Codex reporting a missing session file during history recovery now falls through to loading the session, then to starting a new one. That case used to end the recovery.

Follow a Slack Session from the CLI or the Browser

• Watch a Slack-driven session from your terminal or a browser tab. Attaching to a session that Slack owns now attaches you as a follower. You get the durable history, the turn that is in flight right now, and live updates as they happen, and any number of people can follow at once. Slack stays the only place where prompts are sent, so the two sides cannot both drive the same agent.
• Keep working in the CLI while you follow. The composer stays editable and your draft text and attached images are preserved. Enter does not send. /agent, /new, and /resume still work, so you can leave the followed session for one of your own.

Two Things That Are Gone

• The browser IDE has been removed. The /editor/ route, the !ide command in Slack, Discord, and the CLI, its !help entry, the IDE link in !urls, and the Open IDE button on the fleet dashboard are all gone. Removing it took 122MB out of the software that installs on every session machine. Two more downloads moved out of that bundle, taking it from about 308MB to about 50MB, and a 20MB ceiling per item is now enforced when the bundle is built. Every session machine unpacks that bundle before you can type, so this is time off every session start.
• The Gemini CLI provider has been removed. It is gone from the agent list, the model catalog, and the setup screens, and a fleet still naming it fails loudly at startup rather than trying to run a binary that is no longer there. Antigravity sits in its place on the setup screens as a disabled “Coming soon” row, with nothing behind it yet. Your saved Gemini API key is still stored and still removable; nothing on the machine reads it.

Sign In to Codex from Inside Nori

• Authenticate Codex with /login. The in-app OAuth login wrote its credentials to a directory the Codex agent does not read, so a login that reported success left /agent → Codex failing with “Authentication required for Codex ACP”. The credentials now land in ~/.codex/auth.json, which is where the agent looks. The split has been there since early 2026 and only became visible when the credentials the agent was reading went stale.

Watch a Cloud Session Connect

• See what is happening while a cloud session claims a machine. Starting or resuming a Nori Cloud session now shows live progress in the terminal. Your typed draft is kept while it connects, and the composer refuses to submit until the session is actually running, so a prompt cannot vanish into a session that does not exist yet.
• Read the error when a cloud session cannot start. Failures come back as a message with its detail in your history, and the connecting indicator stops rather than spinning.
• Skip local prompts you cannot answer in the cloud. Local onboarding, directory trust, automatic worktree creation, and per-session skillset selection no longer interrupt a cloud session start. Those are questions about your own machine.

See Your Agent’s Real Model on the Status Card

• Read the model and settings your agent actually resolved. The status card and /status are driven by the configuration the coding agent advertises: provider, model, thought level, then every remaining setting in the order and with the labels the agent gave them. /status lists the agent’s settings in their own block.
• See something on screen the moment Nori opens. The status card and the git branch in the footer are drawn on the first frame. Sessions are created when you send your first prompt, so the card used to stay blank until then, and the footer stayed empty until a background lookup finished. When the session does start, the model and options the agent resolved are announced on their own line.
• Get a quieter footer. The shipped default now shows the approval mode above the composer on the right and the git branch, worktree, and context on the bottom right. The session title, approval label, and token counts are off by default and all still available in /status. Every segment can still be turned back on in your config.
• Read the slash command list. Commands your agent declares get two lines, the command and then its description indented beneath it. One long agent command used to push every built-in command’s description far to the right.