--- title: "Issues" description: "Recurring call problems, grouped into patterns you can triage — and fix with AI help." --- The **Issues** page (`/dashboard/issues`, under Quality & Testing) collects problems detected across your calls and groups repeat occurrences into **patterns**, so fifty instances of "agent quoted the wrong hours" show up as one row with a trend line — not fifty separate reports. ## Where issues come from - **System** — raised automatically by ThunderPhone's call analysis when it detects something going wrong on a call. - **User** — raised by people: reported from a call's detail page or filed via the [issue reports API](/api-reference/issue-reports). Each issue has a **severity** — Critical, Warning, or Info. ## Triage Every pattern row shows its severity, an occurrence **trend** sparkline, the number of affected **calls**, its **status**, and when it was last seen. Statuses: | Status | Meaning | | --- | --- | | **Open** | Needs attention — the default view shows only these. | | **Dismissed** | Acknowledged; not being acted on. | | **Resolved** | Fixed. New occurrences push a pattern back to the top. | Change a pattern's status inline from its status control, or select several rows (including "all matching") and update them in bulk. Filter pills mirror the columns: **Status** (defaults to Open), **Severity**, **Source**, free-text **Search** (title, trace, or call ID), and a **Date** range — plus **Reset filters**. ## Dig into a pattern Click a row to open the detail panel: the pattern's description, its occurrence history, and the affected calls, each linking to the full transcript in call history so you can see the failure in context. ### Fix with AI Fix with AI needs recorded audio from at least one affected call. Calls made while recording was disabled cannot be replayed. The investigation replays stored turns; it does not place live calls or create billable simulation calls. From the detail panel, **Fix with AI** checks whether the problem still reproduces, then tests focused prompt changes against the affected turns and clean comparison calls. The result shows the before-and-after replay rate, detected side effects, and whether the issue was **Fixed** or **Improved**. You review the winning redline before applying it. Leave the page at any time; **Resume fix** returns to the saved investigation. Nothing is applied automatically: - **Apply fix** stages the prompt change in the agent's draft. - **Apply & deploy** stages the change and deploys it immediately. Both actions preserve unrelated draft edits. Until you deploy, the current deployed prompt remains live. If the audio cannot be reconstructed, the investigation ends as **Not replayable**. If the issue no longer appears, it ends as **Not reproducible**. It also stops without changing the agent when its replay budget ends before confirmation. ThunderPhone applies the change only when it can still find exactly one unchanged match for the prompt section that was tested. Unrelated draft edits are preserved. If that section changed or matches more than one place, apply reports a conflict. Rerun the investigation against the current prompt before applying it. See the [Investigations API reference](/api-reference/investigations) for the complete status and replay objects. ### Report to ThunderPhone When the investigation finds no safe prompt fix, select **Report to ThunderPhone**. ThunderPhone also routes some unresolved investigations automatically. If you use organization-wide webhooks, the `issue.escalated` event tells you that an issue pattern was reported and whether the report was automatic. [Configure issue webhooks](/webhooks/events#issue-escalated). ## Next steps Get notified when quality metrics slip, before issues pile up. Turn a fixed issue into a regression test scenario.