Added two cohort metrics — "Issues resolved by cohort" and "Defect originated by cohort" — to the AI usage cohort stats partial (app/views/shared/resource_actions/ai_cohorts_stats.html.haml). This fixes missing cohort-level tracking of resolutions and defect origins, improving visibility into where defects and fixes are coming from for AI cohorts. Deployed to production.

Add a new transcript-inline-chart.js and integrate it into transcript-event.js (renderToolBlock/renderToolActions) to enable rendering inline charts in the Data Deep Dive popup. Implements chart rendering logic and updates tool hooks so the popup can display inline visualizations, improving data inspection and exploration.

Update app/javascript/apps/data-interrogator.js to add a resume bar and follow-up question handling: new UI and logic were implemented inside renderDrawer and runQuestion (including renderResumeBar) to surface resume state and run subsequent prompts. Deployed to production on June 1 — enables users to resume Data Deep Dive sessions and continue exploratory follow-up queries.

Implement recording of per-provider token-type counts in LlmUsageMetric for the past year by adding provider_usage JSONB builders in ExternalUsageClient::AnthropicConcerns (platform and teams) and updating PopulateLlmMetrics. Removed several non-essential columns and deleted an obsolete fetch_cursor_ai_usage_stats helper. Addresses missing token breakdown in usage metrics for reporting/billing and improves historical accuracy.

Introduce a migration to persist cached LLM token counts (db/migrate/20260525152531_add_llm_usage_cached_token_count.rb), update LLM metrics collection (ExternalCursorAiUsageClient) to drop a few non-essential columns, and adjust telemetry tests (ApiAiTelemetryTest). This captures all provider token-count types added over the past year to improve billing, analytics, and historical reporting.

Extended CommitCodeFileExtra to record richer line-level metrics (total_lines, safety/added_safety, added_rescue, duplicate/added_duplicate, churn, LLM-authored, connective invokes, consequential changes). Added a capped backfill_code_line_data rake task to limit daily backfills (oneoff.rake). Tuned RepoCodeCategory multipliers (docs=1.0, CSS=0.3, tests=0.8) and simplified PR sidebar summary rendering (review-comment-item.js). Improves code-line risk scoring and reduces backfill load.

Fixes a bug in app/javascript/bundles/area-chart/area-chart.js where the render/mouseMove callback skipped the zeroth data point and its tooltip didn't appear. Adjusted the mouseMove/render logic to include index 0 so the first point's tooltip displays correctly. Deployed to production 2026-05-15; improves chart usability and data visibility.

Add a migration (create_line_safety_em) and update app/models/code_line to detect and persist "safety words" markers and duplicate content blocks. Also adjust vendor/gems (duplicate_content_block) and commit_crunch_engine scala_line to copy/save body_safe_em data. Released to production to enable collection of duplicate-content and AI-signature research data.

Add server-side validation for v1 AI API tokens triggered after token entry. Changes update llm_metrics external usage clients (external_usage_client.rb, external_augment_code_usage_client.rb, external_cursor_ai_usage_client.rb) including stat_dates_between and get_with_retry adjustments. Deployed May 8 — prevents invalid tokens from being accepted, reducing downstream errors and improving usage metrics accuracy.

Add defect_rate calculation and persistence to LLM metrics so each LLM resource stat now records a defect_rate. Changes touch the llm_metrics gem (stat creation/calculation concern and LlmResourceStat) and were released to production on May 5, 2026. This fills a gap in per-LLM quality telemetry, enabling dashboards and alerts to track regressions and support SLA/quality monitoring.

Add a dedicated _ai_roi_stats.scss, refine app/javascript/apps/compare-stat.scss and make small updates to the ai_roi_stats HAML partial; also tidy a minor ordering in report_tabs.rb. These changes polish layout and visual consistency of AI ROI comparison stats across reports, improving readability and presentation.

Replaces the legacy imperative commit-value editor with a new JS component (moved code into app/javascript/apps/commit-diff-delta-editor.js and app/assets/javascripts/commits/_show.js) that lets users set the percent of a commit attributed to humans vs LLMs. Also fixes assorted prop warnings. Deployed to production; improves UX and reduces console noise.

Remove legacy CommitValue states pending_review and rejected and replace them with an AI-assignment marker. Changes (notably in CommitConcerns::ValueEvaluationConcerns) add deduce_ai_assignment_state and update state-derivation logic; tests adjusted for conflicts. Result: simpler commit-state model that treats anomalous/high-velocity commits as LLM/CI authored to reduce false human attribution.

Update app/models/commit_value.rb to remove pending_review and rejected states and introduce pending_ai_assignment and llm_complete. Rename and add callbacks (detect_zero_pending_ai_assignment, sync_pending_ai_assignment_values, handle_pending_ai_assignment_complete), expand included_in_stats, and adjust approved_value, impacting? and derive_number_authority_em to treat exceptional-velocity commits as LLM/CI-authored. Result: LLM-generated CI runs are identified and correctly included in stat...

Surface the Replace Markdown action directly in the selection menu (app/javascript/lib/ample-editor/components/selection-menu.js): moved and consolidated the ConvertFromMarkdownButton logic and adjusted related ExtractContentButton code so replacing markdown no longer requires a second click when applicable. Deployed to production 2026-03-31 — reduces user clicks and streamlines editor workflow.

Add a callback on UserUsageProfile to enqueue SuggestStarredReportsJob when a user supplies a mission. Update app/jobs/suggest_starred_reports_job.rb to build and send an email containing clickable report suggestions that can be added to the user's dashboard. Impact: proactively surfaces relevant reports after mission setup, improving report discoverability and user engagement.

Add new client-side app (app/javascript/apps/starred-report-picker.js) implementing filterSuggestions, autosuggest rendering, textarea auto-expand, keyboard handlers, search input change handling and requestAiSuggestions/renderAiSuggestionCard logic. Enables interactive selection and saving of starred reports with AI-assisted suggestion cards and improved UX for typing and navigation.