Settings

AI models, background schedules, A/B benchmarking, and access control — all in one place.

The same controls live on the triage page. Model IDs must carry a publisher prefix (anthropic/ or google/) — bare names silently reroute on the gateway and are normalized on save.

Dry-run (proposals only) SAFEenv default

When enabled, triage drafts internal-note proposals but writes nothing to Zendesk.

Preview mode — hold suggestions for review

The AI drafts as normal but writes nothing to the Zendesk ticket. Each suggestion waits on Waiting for Review until a human approves it — where it can also be edited first. Different from dry-run: dry-run discards the draft, preview keeps it so a good one can still be posted.

Auto-execute on new tickets Off

Server poller fires /api/triage/run every 5 min for eligible tickets.

Config grounding for unlinked customers env default

Orgs with a real linked repo always use their own config. This controls what UNLINKED orgs get.

AI model configuration

Triage (main agent) — triage_llm_model

Every /api/triage/run call. Sonnet 5 by default.

Enrichment (backfill) — enrichment_llm_model

Bulk resolution-enrichment pipeline. Haiku is 3-5x cheaper for the same quality here.

Attachment analysis — attachment_analysis_model

Config/log/zip inspection on ticket attachments.

Image vision — vision_model

Screenshot description via Gemini multimodal.

Text embedding (similarity) — text_embedding_model

Powers find_similar_tickets + corpus backfill. Changing it invalidates the existing 3072-dim vectors.

Multimodal embedding (reserved) — multimodal_embedding_model

Not consumed yet — reserved for future image+text similarity.