## Summary Follow up to: #3586169 Contrib and site debugging today rely on watchdog/`debug_logging` config, which is global, noisy, and can log sensitive task text. We need structured, **opt-in** selection diagnostics on the result object for preview tools, Drush commands, and developer UI — without leaking context content. **Goal:** When diagnostics are requested, `AiContextSelector::select()` accumulates a structured breakdown and exposes it on `AiContextResult`. ## Solution **Proposed API:** - `AiContextRequest::withDiagnostics(bool $enabled = TRUE)` (or equivalent factory/builder opt-in; **default off**) - New value object: `AiContextSelectionDiagnostics` - `AiContextResult::getDiagnostics(): ?AiContextSelectionDiagnostics` **Diagnostics payload (IDs, scores, reason codes only — no rendered text or content snippets):** - Prefilter candidate count (after scope index) - Post-load candidate count - Per-item score breakdown (where applicable) - Skipped items + reason enum (e.g. token cap, never include, access, hard filter) - Final token budget used / truncated item IDs **Non-goals for this issue:** - Selection preview UI (`/admin/ai/context/selection/preview`) — separate follow-up - Watchdog replacement for production logging - Including rendered content or content snippets in diagnostics **Security / privacy:** - Never include raw context content in diagnostics - Do not auto-enable via global config; opt-in per request only ## Acceptance criteria - Opt-in request returns populated diagnostics; default request returns `null` - Kernel tests cover at least: empty selection, truncated-by-tokens, skipped-by-access/filter - Developer docs describe the API and privacy constraints ## Related - A `selection preview UI` feature could be a follow-up to this work - Complements selector pipeline events (`#3586162`); does not replace them ## AI usage - [x] AI assisted issue - [ ] AI generated code