How Mushi Mushi reads a bug report

When a user submits a report, Mushi Mushi turns raw text and screenshots into a plain-English summary — severity, category, and what to try next — before anything hits your editor.

Outputs

category (bug, slow, visual, confusing, other)
severity (low / medium / high / critical). The nightly judge scores severity calibration and feeds the prompt-A/B framework.
component (resolved against your project’s component taxonomy if uploaded).
bug_ontology_tags (tags from the shared bug ontology).
confidence ∈ [0, 1].
summary (one-sentence triage headline).

Vision air-gap

The screenshot is read in a separate, image-only call with its own fixed system prompt and no report text, after the diagnosis is finished. This limits OWASP LLM01 prompt injection via images: a screenshot containing “ignore prior instructions” cannot change the diagnosis, which is already written. The call records any text in the image as data and flags injection attempts. See Prompt-injection defence.

BYOK

Per-project keys (Anthropic and/or OpenAI) override the cluster default. Failover order:

Project’s BYOK Anthropic key (if set)
Cluster Anthropic env var (if set)
Project’s BYOK OpenAI key
Cluster OpenAI env var

Every invocation logs which key source was used so finance can attribute spend without inspecting the prompt.