Problem reports
When one of your application’s tools fails on the server side, the assistant can offer to report it. The user sees a draft in the consensus dialog, edits it, switches off any attachment they do not want to share, and approves or rejects it. Only an approved report is stored. The team that runs the assistant reads reports on the Problems page, and can forward them to its own tracker through webhook sinks.
You do not integrate anything in your app for this. It works through the same consensus dialog and the same credential-blind rules as every other gated action.
When the assistant offers a report
Section titled “When the assistant offers a report”Every failed application tool result carries a label the platform itself assigns from the failure category, before the model reads it:
| Label | Failure categories | What the assistant does |
|---|---|---|
host_fault | server error (5xx), network error or timeout, circuit open | Explains briefly and offers a report in the same turn. |
denied | 401, 403 | Explains the permission problem. No offer. |
invalid_input | other 4xx (bad arguments) | Asks the user to fix the input. No offer. |
unsure | tool error, in-doubt, unknown | Says what failed and asks how to proceed. No offer. |
The offer is the consensus dialog: the assistant calls its report_problem tool, which is always gated, and
the dialog opens. It does not ask “shall I report this?” in plain text first. It also does not offer for a
feature request, or for its own mistake (a wrong argument, a misread result).
Once per problem per session. After the user rejects a report, or one is filed, the assistant does not offer the same problem again in that conversation. A problem here is the failing server, tool and error category. The marker is kept with the conversation, not in server memory, so it holds across instances and after a reload.
The consent view
Section titled “The consent view”The dialog is a built-in consensus view for the
report_problem tool. It shows:
| Part | What the user can do |
|---|---|
| Title, Summary | Edit. Drafted by the assistant. |
| Steps to reproduce, Expected, Actual | Edit (optional). |
| Conversation excerpt | Switch off. Shown exactly as it would be stored. |
| Page | Switch off. The page path the user is on. |
| Failed tool calls | Switch off. Server, tool, category, status, error text and time of each failed call. |
| Related tool calls, Assistant assessment | Read only. The assistant’s own facts about the report. |
What the user sees is what is stored: the preview is built once, before the dialog opens, and stored as shown, minus whatever they switched off. It is never assembled a second time at approval. An item the hub could not redact is shown as “Not attached: could not be redacted” and is left out, and an item with nothing to attach reads “Nothing to attach”.
Only the five text fields are editable. The hub refuses an edit to anything else, including which failed calls are attached and the assistant’s assessment. Rejecting the dialog stores nothing. Approving stores the report; the assistant is told only the report id, never what the report holds.
Consent for reports cannot be switched off: no approval policy covers report_problem, and a user always sees
the dialog. Turning reports off is a per-assistant setting (see Switching it on and off).
The approval record of a report holds the draft’s text but never the preview.
What a report holds, and never holds
Section titled “What a report holds, and never holds”A stored report holds:
- the five text fields as the user left them;
- optionally, an excerpt of the conversation: the last 10 user and assistant messages that have text, each cut to 1,000 characters and the whole to 8,000, oldest first. Never tool results, file contents, images, system messages or tool arguments;
- optionally, the page path the user was on (the path the hub holds, up to 2,048 characters, not the full URL);
- optionally, up to 20 failed calls, each with server, tool, error category, status code, the error text (cut to 2,000 characters) and time. Only failures from this conversation; ids from another conversation, or of calls that did not fail, are dropped. The tool arguments of the call are never copied;
- the source (
assistant_offeroruser_command) and the assistant’s assessment (defect,not_a_defectorunsure); - who filed it: the signed-in user’s id, or “anonymous”, and the conversation it came from.
A report never holds credentials or auth headers. The preview is built only from named hub records (this conversation’s failed-call log rows, its messages and its page path). It never reads the session’s bound auth artifacts, so BYOA credentials cannot reach the report, the preview or the model, and the model never sees the preview at all. Error text and conversation text pass through the hub’s secret redaction before they are shown.
The /problem command
Section titled “The /problem command”When the license and the assistant’s setting allow reports, the composer’s / menu lists problem. Typing
/problem (optionally followed by a few words about what went wrong) starts a report on purpose. The match is exact
and case-sensitive: /Problem, a full-width slash or /problems are ordinary text. There is no other alias.
The command goes through the assistant, not around it:
-
After a server-side failure it drafts the report from what happened and opens the dialog with the assessment
defect. -
After a denial or invalid input, or for a feature request, it first explains why this does not look like a fault of the application, and does not open the dialog yet.
-
If the user insists (“file it anyway”), it opens the dialog with the assessment
not_a_defect. It never refuses a report the user asks for.
A report started this way has the source user_command. The command lifts only the once-per-problem rule, for
that one dialog, and only until the second user message after the command: /problem after the assistant already
offered or the user already rejected is honoured. It does not lift the limits below.
There is no button; the menu is the only entry. Commands registered by your own app are not supported, and the
@ mentions in the composer work as before.
Limits
Section titled “Limits”| Who | Limit |
|---|---|
| Signed-in user | 5 reports per conversation |
| Anonymous visitor | 1 report per conversation, and 20 anonymous reports per assistant per rolling hour |
When a limit is reached no dialog opens and the assistant is told “The problem report limit for this conversation has been reached.” The counts come from stored reports, so they hold across instances. The limits are fixed.
Switching it on and off
Section titled “Switching it on and off”Reports are available when the license has problemReports and the assistant has reporting on. The setting is
per assistant, defaults to on for every assistant on Pro (existing ones included), and has a deployment-wide
default for assistants that have not chosen. See Assistants & Roles,
System settings and the one-time rollout notice there.
Do not name an MCP tool report_problem
Section titled “Do not name an MCP tool report_problem”The built-in tool is named report_problem. If any MCP tool in the assistant’s reach resolves to the same name
(for example an instance named report with a tool named problem), DioscHub withholds the built-in for that
assistant until the tool is renamed. The customer’s tool then runs as usual with the generic approval form,
no problem report can be filed, and the playbook is left out of the prompt. The / menu no longer lists
/problem for that assistant, and typed /problem is plain text. A warning is logged on every turn that finds the clash:
An MCP tool is named report_problem, like a problem-reports built-in: problem reports are off for this graph until it is renamedRename the MCP tool (or its instance) to fix it. The built-in’s consent view is locked for the same reason: a
consensusView('report_problem', …) registered by your app is ignored.
Where this connects
Section titled “Where this connects”- Read, filter and forward reports: Problems.
- The dialog and its views: Consensus views and MCP consensus.
- What is stored, kept and erased: Data handling.