feat(external-issues): Use LLM generated title/description for ticket creation

[leeandher]

Adds a flag which will support making an LLMGenerateRequest when a user attempts to create an issue link.

A few design decisions:

• It adds a trade-off, a few seconds for an LLM call so that the default response comes with the text, rather than having the browser make a form call, then an LLM call, then replace the content in the browser. This avoids a whole bunch of issues if the user eagerly makes form changes prior to the LLM call.
• This does not apply to all ticket creation. If you set up Jira Tickets or Github Issues to be generated from alerts, those fire through

  sentry/src/sentry/rules/actions/integrations/create_ticket/utils.py

  Lines 146 to 152 in 9644ccb

  	data["title"] = installation.get_group_title(event.group, event)
  	workflow_id = data.get("workflow_id")
  	if workflow_id is not None:
  	data["description"] = build_description_workflow_engine_ui(
  	event, workflow_id, installation, generate_footer
  	)

  , which is a separate code path -- we are not introducing an LLM call on every issue firing, only UI linkages
• We are omitting sentry apps from this change (unfortunately). those titles/descriptions are assembled on the frontend. That said, we could expose this LLMGenerateRequest through an API and service those as well in the future.
• We are retaining the existing title + description in the description, in pseudocode:

# Assuming LLM generates a title + description
Issue Title: LLM Title
Issue Description: f"**{DefaultTitle}**\n\n" + LLM Description + `\n\n---\n\n` + DefaultDescription"

Examples

[linear-code]

ISWF-2529 Allow LLM proxy to generate issue title and description when linking

[Christinarlong]

I'd just watch out for increased latency on the modal loading

[Christinarlong]

Question in general, but will the LLM title/descr. generation be in addition to the integration requests to get things like repo and assignees?

My only callout here would be about latency, since I believe just opening this modal currently can take a while (unless we've already added improvements that idk 'bout). Idk if we have tracing/observability on the getting of issue config process?

[leeandher]

Yeah the latency will definitely increase for organizations opted into this flag. I wanted to test it with production times since I don't have a great gauge for how noticeable it will be when running it locally.

Since it's feature flagged, our SaaS org is probably a good test bed for whether this does need to be retooled once we have it enabled, but either way this is required to power the feature 🤷

[Christinarlong]

Curious Qs, but what do temperature and max tokens do? is temp analogous to effort? What happens if you hit max tokens? Will 750 be enough? is there a way to test how many will be used?

[leeandher]

Honestly, it's not a science and kind of a best guess approximation. these were the values I landed on to consistently get results locally without timeouts or token issues.

if you hit the max tokens the request fails, and we fall back to the default, but since these things aren't deterministic its kinda guess work.

the temperature is a sorta analog for creativity or randomness, closer to 0, is supposedly more deterministic, but we also want to avoid being boring (i.e. every title being "A sentry python issue occurred." or something like that)

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9410c88. Configure here.}

[cursor]

Truthiness check before strip causes inconsistent fallback

Low Severity

The if title and description: truthiness check on line 112 runs on pre-stripped values, but the returned values on line 113 are post-strip(). If the LLM returns a whitespace-only title (truthy pre-strip, falsy post-strip) but a valid description, the function returns {"title": "", "description": "real desc"}. The caller in issues.py then independently checks each field's truthiness, causing the title to fall back to default while the description uses the LLM-formatted template — an inconsistent pairing.

Additional Locations (1)

• src/sentry/integrations/mixins/issues.py#L168-L174

 

^{Reviewed by Cursor Bugbot for commit 9410c88. Configure here.}

[cursor]

Non-dict JSON response causes uncaught AttributeError

Low Severity

If json.loads(content) succeeds but returns a non-dict type (e.g., a list, string, or number), the subsequent content.get("title") call raises an AttributeError. The local except block only catches json.JSONDecodeError, TypeError, and ValueError, so this error escapes to the generic except Exception in maybe_generate_external_issue_details, losing the specific logging context (group_id, viewer_context) and logging a misleading error message.

 

^{Reviewed by Cursor Bugbot for commit 9410c88. Configure here.}