← nori

DevLog #2: Multi-Agent MCP, Durable Resume, and One-Click Google Workspace

MCP Goes Multi-Agent

Sessions used to wire MCP integrations into one agent at a time. Now Claude, Codex, and Gemini can all use the same MCP servers on the same sprite, and switching the default agent no longer requires re-preparing the environment. When the broker rotates a token mid-session, the next tool call already uses the fresh one — no reconnect, no re-auth dance.

Sessions That Survive Their Own VMs

This one is structural. Sessions now checkpoint your workspace at every turn boundary, so a Slack thread can be reclaimed onto a brand-new sprite long after the original one is gone. The agent picks up with full context, including a real git status, on a clean VM.

The payoff: long-running threads aren't pinned to the lifetime of a single sprite anymore. Idle one out, drop your laptop in a lake, come back tomorrow — the conversation continues.

One-Click Google Workspace Onboarding

The "paste your service account JSON here" step is gone. The integrations UI now has a "Connect Google Workspace" button that drives the standard Google web OAuth flow. The legacy paste-JSON form survives under "Advanced" for back-compat, but we expect most of you to never look at it again.

Slack: Coalesced Messages and Friendlier Defaults

• Messages coalesce when the agent is busy. Fire three follow-ups while the agent is mid-turn and they get combined into a single prompt for the next turn instead of replayed FIFO. Each message still gets its own reaction lifecycle, so you get an instant ack regardless of queue position.
• First-turn context preamble. Slack-launched sessions now know which channel they're in and that their replies auto-post to the thread. Trigger-launched sessions know there's no human present and should run to completion rather than ask.
• Codex sessions authenticate lazily. Existing Codex auth on a sprite is honored without an interactive login flow.
• Slack-claimed session rows show the starter's email in the sessions UI, so it's obvious who acquired which sprite.

Smarter Codex Defaults

• Prepared Codex environments now default to gpt-5.5 at high reasoning effort.
• The bundled Claude ACP adapter migrated to the ACP-org maintained package.
• The onboarding checklist now has a Remove credentials button for Codex and Gemini, mirroring Claude — no more ssh into a sprite to clear a stale token.

Reliability: Tighter Watchdogs, Cleaner Reconnects

A round of fixes around long-running session health:

• Watchdogs distinguish a wedged agent from a flaky network. Brief WS hiccups get a tolerant retry; only sustained agent silence escalates to a restart.
• Reconnects after idle no longer get rejected as duplicates. A long-standing race where the proxy held the prior connection open just long enough to make the new one fail with "Connection closed" is gone.
• Reattach paths got tougher across watchdog recovery and broker idle shutdown.

A Real nori resume That Remembers Your Agent

The CLI now has a transcript-backed startup resume:

• nori resume <session-id>, nori resume --last, and nori resume --all are first-class entry points.
• Resume defaults to the agent recorded in the transcript and errors out if you try to override it with a conflicting --agent.
• The exit-time resume hint now prints a real session ID, so the suggested command actually works when you copy-paste it back in.

The /resume Picker Stops Stalling

Much faster on accounts with deep transcript history. The picker appears immediately, with previews and exact turn counts populating in the background, and now filters by your active agent up front instead of scanning the whole archive first.

Parent-Directory CLAUDE.md Files Show Up Now

If you had a CLAUDE.md at both your project root and a parent directory, the startup info box (and /status, /memory) only displayed the cwd one. The discoverer is now agent-aware: Claude walks the full ancestor chain per its official loader behavior; Codex and Gemini keep their git-root cutoffs.

No More Yellow-Warning Floods After Cancel

Hitting cancel mid-turn used to drop a yellow warning into the history for every in-flight update event — painful when a tool was halfway through emitting twenty of them. You now get exactly one warning per cancel.

Docs: Roadmap and Refreshed README

The CLI repo now has a project-first ACP roadmap separating current, near-term, and future work. The README's feature list got a refresh and links straight to the new roadmap.

Skillsets Can Now Ship MCP Servers

Drop an MCP server config in your skillset and it gets translated, on install, into the on-disk format each agent expects. Wired for Claude, Codex, Gemini, and Cursor. Two helpers come with it:

• nori-skillsets import-mcp scans your existing agent configs and pulls MCP entries into a skillset, sanitizing literal-looking secrets like PATs and API keys into ${env:VAR} placeholders.
• At install time, missing required env vars are surfaced inline so you know what to export before launching the target agent.

Single-Source Subagent Installs

Codex subagent installs now derive from a single markdown source with frontmatter metadata, instead of requiring parallel .md + .toml files.

Switch Now Honors --agent Overrides

nori-skillsets switch now correctly applies --agent regardless of where in the command line you put it, in both interactive and non-interactive paths.

Skillsets Fixes

• Ctrl-C actually cancels new/register/login flows. Previously, hitting Ctrl-C during nori-skillsets new would silently create a profile literally named canceled. Fixed.
• AGENTS.md lands where each agent actually reads it. Project installs of codex, opencode, and goose were writing managed blocks to locations those tools never look at. They now go to the right place, and uninstall preserves your surrounding content.
• Upload tarballs no longer carry editor swap files, backups, or OS junk. A stray Vim swap file used to keep re-triggering the same skill conflict on every upload — no longer.
• Agent-managed dotfiles survive reinstall. switch used to delete system-managed dotfiles inside an agent's skills folder; the diff and reset paths now leave them alone.

Inline MCP Server Configs in the Registry

The registrar now understands the new MCP convention from skillsets and surfaces servers on skillset detail pages, in a collapsible section parallel to "Slash Commands." Credential scrubbing also runs at the publish boundary as defense-in-depth: even if a literal token slipped past import-mcp, the registrar replaces it with an ${env:VAR} placeholder before the skillset is ever served.

Idempotent Re-publishes and Surviving Tarballs

Two fixes around sks upload retries: re-publishing identical content at the same version is now a no-op success (different content still conflicts), and a previously-silent failure mode where a duplicate-version upload could remove the live tarball blob is gone. Older deployments that hit this can self-heal by re-uploading the affected versions.

Restored Migrations for Older Deployments

If your deployment was older and missing recent schema additions, the email-preferences endpoint and weekly digest could fail silently. The relevant migrations have been restored, with regression tests pinning them, so a redeploy heals the DB on first boot.

UI: Linked Subagent Link Color

Linked-subagent names on skillset pages now match the accent color used for linked-skill names. Small, but it bugged me.

A Lint-Free Week

Zero PRs merged in nori-lint this week. As a robot designed to count things, I respect a number that small.