Release v0.98.1
Jul 20, 2026
A rebuilt quote editor with a 30-second undo on send, AI approvals that survive a restart, and a round of agent reliability fixes. A standard upgrade with no breaking changes.
Quotes & Proposals
- A quote editor that looks like the quote. The builder has been reworked so the canvas reads like the proposal your customer will receive — cleaner line entry, seamless fields, and the company identity moved into the header instead of sitting in a form.
- A 30-second undo on send. After you hit Send, the quote shows a short countdown before the email actually leaves. Click Undo and nothing is sent — the quote stays a draft and you can keep editing. Catch the wrong contact or a bad price without an apology email.
- Clearer feedback when a send doesn't go through. If a scheduled send fails at the moment it fires, the quote now carries a persistent banner explaining what happened instead of a toast you might have missed.
AI & Automation
- Pending AI approvals survive a restart. A Tier 3 action waiting for approval is no longer tied to the session that proposed it — close the browser, restart the server, come back tomorrow, and the request is still in Approval History waiting for a decision.
- Control who can approve AI actions. Approving or rejecting a Tier 3 action now requires a dedicated permission, granted by default to Org Admins and not to Technicians or Viewers. Add it to any role you want to give approval authority.
- Approved identity actions run on their own. Once you approve an action against a connected Google Workspace tenant — a password reset, a forced sign-out, suspending a user — it executes against that organization's connection without needing the original chat session to still be open. Microsoft 365 write actions work the same way where the optional connector is configured.
- Fixed a broken CIS compliance lookup in the assistant, and error details returned by AI tools are now scrubbed before they reach the chat window.
Remote Management & Agent Reliability
- Agents no longer get stranded during a credential refresh. Agents periodically rotate their credentials; if the device failed to write the new ones to disk at the wrong moment, it could be left unable to authenticate. The agent now confirms it has safely stored new credentials before the server switches over.
- Full script output, every time. Script results now reliably carry standard output, error output, and the exit code back to Breeze, instead of occasionally arriving with those fields empty.
- Remote desktop links work on Linux. Clicking a remote session link now correctly launches the Linux viewer, which previously wouldn't connect.
- A more reliable admin-elevation prompt on Windows. The deny path now clears only on a confirmed dismissal, so a request can't get stuck in a blocked state.
End-User Experience
- A tidier Breeze Helper. The end-user assistant now presents a single window header with one navigation bar, and on wider windows the chat sits alongside your work as a side panel rather than replacing the whole view.
- The Helper remembers its window. Size and position now persist between launches instead of resetting every time.
Version 0.98.1 is about not having to be perfect the first time — and about work that keeps going when you don’t.
The quote builder has been rebuilt so the editing canvas actually reads like the proposal your customer receives, rather than a form that happens to generate one. The change you’ll feel most is the 30-second undo on send: hit Send and the quote holds for a short countdown before the email goes out. Spot the wrong contact or a stale price in that window and one click stops it — the quote stays a draft, and nobody gets a correction email.
The same idea runs through the AI side. Pending Tier 3 approvals are now durable — a proposed action that needs a human decision is no longer tied to the chat session that raised it, so you can close the browser, restart the server, or pick it up the next morning and it’s still waiting in Approval History. Approving is now its own permission, granted to Org Admins by default and off for Technicians and Viewers, so you decide who holds that authority. And once approved, identity actions against a connected Google Workspace tenant run on their own without the original session open.
Rounding it out is a round of agent reliability work: agents can no longer be stranded unable to authenticate if a credential refresh is interrupted, script results reliably carry their full output and exit code, remote session links now launch the Linux viewer, and the Windows elevation prompt can’t get stuck in a denied state.
Self-hosting note: this is a standard upgrade — 8 idempotent database migrations that apply automatically, with no large data rewrites or blocking backfills, no new required configuration, and no breaking changes. One housekeeping detail: 0.98.0 was withdrawn before it shipped (its web container image failed to build), so 0.98.1 is the version to run — it contains everything 0.98.0 was meant to, plus the fix. As always, review the GitHub release notes before updating.