Mobile Bug Investigator

Takes a fuzzy bug report — "auth gives errors sometimes", "the banner is broken" — and turns it into a scoped investigation someone can actually run, then into a brief a planner can act on. It maps the user-visible symptom onto code-level concepts, writes the search plan as concrete patterns to run in a stated order, and forces the cross-surface question that single-repo debugging skips: a mobile symptom is often a backend change, and a backend change often reaches an admin panel nobody thought to check. It is built around one discipline. It never reports a finding it was not given. Ask it to write the report before the search has run and it will say the investigation has not happened yet and hand back the search plan instead of a page of plausible file paths. Where it reasons rather than knows, it says so in hypothesis language — "appears to live at", "the planner should verify" — because a confident-sounding guess about where a bug lives costs more than an honest unknown. Its output ends at a planner prompt. It does not decide the desired behaviour and it does not recommend a fix; where the report implies a behaviour choice nobody has made, it surfaces that as a question. Each tool is a prompt-template: it calls no model, reaches no network, needs no credential, and returns the same text for the same input every time.