Bug-report sheet → AI chat conversation

Status: Delivered
CAS: CAS-1933
Delivered: 2026-05-07
PRs: #495, #497, #498, #500, #502, #503

What’s new

The “Report a bug” button throughout the app now opens an AI-assisted chat instead of a plain textarea. When the dialog opens, it automatically bootstraps a conversation with Claude, providing it the current route, the last 60 seconds of trace logs, a store snapshot, a screenshot, and the app version and build SHA. You describe what went wrong; Claude asks clarifying questions and proposes a structured, Paperclip-ready bug report. You can also save ideas from the same dialog — suggestions or feature thoughts that don’t need to be filed as bugs.

How to use it

Click the “Report a bug” button on any page (available in the app toolbar).
The dialog opens. Claude is already briefed — you see its opening message acknowledging your current page and context.
Type a description of the issue. Claude will ask follow-up questions to clarify steps, expected vs. actual behaviour, and severity.
When Claude proposes a structured report, review it. You can ask for edits or regenerate.
Press Submit report to file the report as a GitHub issue via the existing Tauri pipeline. The issue body contains the AI-formatted summary plus a Submitted by: line identifying you.
Alternatively, select Save as idea to capture the thought in the local ideas store without filing an issue.

The chat dialog has a type toggle (Bug / Idea) and a priority selector so you can classify before submitting. Both are reflected in the GitHub issue labels.

What changed under the hood

New ideas table in SQLite; new save_idea and chat_stream Tauri commands.
LocalCliService (CAS-1932): Claude Code runs as a managed subprocess inside the app, providing streaming chat responses via chat://stream-delta / chat://stream-done Tauri events.
useReportIssueStore builds the bootstrap prompt at dialog open — injecting route, log buffer, store snapshot, build SHA and version, and screenshot bytes (passed as a vision input when the local CLI supports it).
ReportIssueParams extended with issueType and priority; report_issue command appends labels to the filed GitHub issue accordingly.

Why we built it

Filing a plain-text bug report from a running app is friction enough that most bugs go unfiled. By turning the dialog into a short AI conversation, we lower that barrier: the AI already has the context (logs, screenshot, route) so the regent or Eivind can describe the symptom in one sentence and let Claude produce the structured report. This also makes the reports more consistent and easier for board agents to triage.

Known limitations / follow-on work

Screenshot is currently passed as base64 in the system prompt; vision-input path is wired but depends on the local CLI version supporting the --image flag — not verified on all builds.
Ideas are stored locally only; no sync to Paperclip or GitHub yet.
Bootstrap prompt does not yet include previous crash reports or the last filed issue — adding these would improve context quality over time.