feat: AI adoption scoring, redesigned adoption card, diffs accepted KPI

- Add composite AI adoption score (accept rate 40%, engagement 40%, consistency 20%)
  with tier labels (AI-Native/High/Moderate/Low/Manual)
- Collect AI code tracking data from /analytics/ai-code/commits endpoint
- Redesign user detail adoption card: tier headline, description, score bar, stat pills with tooltips
- Replace misleading Accept Rate KPI (accepts/(accepts+rejects) ~100%) with
  Diffs Accepted (accepts/applies) showing actual acceptance rate
- Fix consistency calculation: use elapsed days in window instead of full 30
- Add adoption badges to dashboard members table
- Update FEATURES.md and cursor rules with new metrics documentation

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: ofershap

.cursor/rules/cursor-api-data-guide.mdc

@@ -120,16 +120,31 @@ Legacy field from the old fixed-pricing model. It's NOT the current billing mech
 - `acceptedLinesAdded / totalLinesAdded` as "AI %" - WRONG for agent users, only valid for tab/composer-heavy users
 - `totalApplies` - may undercount in YOLO/auto-apply mode
 - `usageBasedReqs` vs `agentRequests` - the gap between these is unclear; likely `usageBasedReqs` counts only requests that consumed usage-based billing
+- `nonAiLinesAdded` from ai-code/commits - NOT "manually typed code". It's "unattributed lines" — lines where Cursor's server couldn't trace back to an AI generation event. Breaks on squash merges (attribution from individual commits is lost), initial commits to new repos (no prior AI tracking context), and code that went through review cycles. Verified against real data: a 100% agent-driven user showed 43% "nonAI" due to squash merges and initial commits. Do NOT use this as "% manual work".
+
+### AI Code Tracking API (`/analytics/ai-code/commits`) Limitations
+
+Investigated Feb 2026 by comparing real user data against git history and studying git-ai's approach:
+
+1. **Squash merges lose attribution**: When a PR is squash-merged, the final commit has no record of the individual AI edits from development. All lines appear as `nonAiLinesAdded`. This is the #1 source of false "manual" attribution for teams using squash merge workflows.
+
+2. **Initial commits to new repos**: The first commit to a new repo has no prior AI tracking context. Even if the agent wrote 100% of the code, it all shows as `nonAiLinesAdded`.
+
+3. **Only Cursor Source Control panel commits are tracked**: Terminal `git commit`, VS Code source control, and external git tools produce no tracking data at all.
+
+4. **git-ai (open source alternative) solves this differently**: Uses real-time IDE checkpoints (pre/post edit hooks) stored in `.git/ai/`, intersected with the actual diff at commit time. Survives rebases, squash merges, cherry-picks by rewriting authorship logs. Requires installing their tool — not replicable from API data.
+
+5. **Industry consensus (DX Framework, Cursor's own research)**: Lines-of-code metrics are unreliable for measuring adoption. The recommended approach combines accept rate + engagement intensity + consistency. See "AI Adoption Score" in Derived Metrics.
 
 ### Now Collected
 - Per-request token/cost data from `/teams/filtered-usage-events` - gives per-model cost breakdown per user. Stored in `usage_events` table. Collected incrementally (since last timestamp).
 - Command adoption from `/analytics/team/commands` - team-level command usage. Stored in `analytics_commands` table.
 - Plan mode adoption from `/analytics/team/plans` - plan mode usage by model. Stored in `analytics_plans` table.
 - Per-user MCP tool usage from `/analytics/by-user/mcp` - which MCP tools each user uses. Stored in `analytics_user_mcp` table.
 - Per-user command usage from `/analytics/by-user/commands` - which commands each user uses. Stored in `analytics_user_commands` table.
+- AI Code Tracking from `/analytics/ai-code/commits` - per-commit AI vs manual line attribution. Stored in `ai_code_commits` table (aggregated per user per day). Provides `tabLinesAdded`, `composerLinesAdded`, `nonAiLinesAdded` per commit. Used for AI Adoption % on user detail page. Only tracks commits made through Cursor's Source Control panel — terminal git commits are not captured.
 
 ### Not Currently Collected (but available)
-- AI Code Tracking data from `/analytics/ai-code/commits` - would give us accurate AI vs human line attribution
 - Per-user breakdowns from `/analytics/by-user/*` endpoints for: agent-edits, tabs, models, plans, ask-mode, client-versions, top-file-extensions (we collect mcp and commands per-user, but not these others)
 - Leaderboard from `/analytics/team/leaderboard` - ranks users by tab accepts and agent edits. We chose NOT to collect this because it introduces a third ranking system that conflicts with our own spend_rank and activity_rank, confusing stakeholders.
 - `cmdkUsages`, `subscriptionIncludedReqs`, `apiKeyReqs`, `bugbotUsages` - available in daily usage but not stored
@@ -158,6 +173,22 @@ Low accept rate could mean: picky reviewer (good), bad prompting (fixable), or w
 
 Highly task-dependent. A debugging session produces 0 lines. A scaffolding task produces 500. Not a quality metric.
 
+### AI Adoption Score (0-100%)
+Composite score from three signals in `daily_usage`, weighted:
+- **Accept Rate** (40%) — `total_accepts / total_applies` — trust in AI output
+- **Engagement Intensity** (40%) — `agent_requests / active_days`, normalized against team p90 — how heavily they lean on AI
+- **Consistency** (20%) — `active_days / period_days` — regular usage vs sporadic
+
+Tiers: AI-Native (80%+), High (55%+), Moderate (30%+), Low (10%+), Manual (<10%).
+
+Why this works better than commit-based AI%:
+- Available for ALL users (not just the ~18% who commit through Cursor's Source Control)
+- Not affected by squash merges, initial commits, or external git tools
+- Validated by Cursor's own productivity research (accept rate correlates with developer proficiency) and the DX AI Measurement Framework (combine frequency + acceptance + consistency)
+- Actually differentiating: accept rate ranges 0-100% across real team data, engagement intensity ranges 2-130 reqs/day
+
+The commit-based AI% from `ai_code_commits` is shown as a secondary "AI Code %" metric with a caveat about attribution limitations.
+
 ## Daily Spend Data Sources
 
 `usage_events` (from `/teams/filtered-usage-events`) is the most reliable source for daily spend data. It has per-request cost (`total_cents`) with full billing cycle history and no retention window. `daily_spend` (from `/teams/groups` billing groups API) has only ~2 days retention and systematically underreports compared to `usage_events`.

.cursor/rules/features-file.mdc

@@ -0,0 +1,9 @@
+---
+description: Keep FEATURES.md up to date when adding new features
+globs: "src/app/**/*.{ts,tsx}"
+alwaysApply: false
+---
+
+# Feature Map Maintenance
+
+When adding a new user-facing feature (page, section, chart, badge, metric, setting, or API endpoint), update `FEATURES.md` to reflect the change. Keep entries short — one line per item, grouped under the correct page heading.

.cursor/rules/project-context.mdc

@@ -67,7 +67,7 @@ Single cron endpoint `POST /api/cron` does both: collect → detect → alert in
 
 - `/` — Team overview: stat cards, model cost comparison table ($/request relative multipliers), daily spend trend (sourced from `usage_events` with `daily_spend` fallback, last 2 days marked provisional), spend breakdown by user, members table with search/sort, **group filter dropdown**, time range picker (24h/3d/7d/14d/30d), billing cycle progress
 - `/insights` — Analytics: DAU chart, model adoption, model efficiency rankings, MCP tool usage, file extensions, client versions
-- `/users/[email]` — Per-user detail: KPI cards (cycle spend, $/req, agent reqs, accept rate, team rank), spend trend chart, usage profile radar (activity, intensity, tab usage, precision, on plan, power user), cost breakdown by model, tools & features (MCP tools + commands per user), model preferences, daily activity table, anomaly history
+- `/users/[email]` — Per-user detail: KPI cards (cycle spend, $/req, agent reqs, diffs accepted, team rank), spend trend chart, AI adoption card (tier, score bar, stat pills with tooltips), cost breakdown by model, tools & features (MCP tools + commands per user), model preferences, daily activity table, anomaly history
 - `/anomalies` — MTTD/MTTI/MTTR metrics, open incidents (acknowledge/resolve), anomaly table
 - `/settings` — Detection thresholds, **billing group management** (rename, assign, create), **HiBob CSV import** with change preview
 
@@ -90,7 +90,7 @@ Single cron endpoint `POST /api/cron` does both: collect → detect → alert in
 
 ## Database Tables
 
-members, daily_usage, spending, usage_events, anomalies, incidents, config, collection_log, metadata, daily_spend, billing_groups, group_members, billing_group_members, analytics_dau, analytics_model_usage, analytics_agent_edits, analytics_tabs, analytics_mcp, analytics_file_extensions, analytics_client_versions, analytics_commands, analytics_plans, analytics_user_mcp, analytics_user_commands
+members, daily_usage, spending, usage_events, anomalies, incidents, config, collection_log, metadata, daily_spend, billing_groups, group_members, billing_group_members, analytics_dau, analytics_model_usage, analytics_agent_edits, analytics_tabs, analytics_mcp, analytics_file_extensions, analytics_client_versions, analytics_commands, analytics_plans, analytics_user_mcp, analytics_user_commands, ai_code_commits
 
 ## Important Caveats
 

FEATURES.md

@@ -0,0 +1,137 @@
+# Feature Map
+
+## Dashboard (`/`)
+
+**Team overview — cost, activity, members at a glance.**
+
+### Controls
+
+- Search by name/email
+- Filter by billing group
+- Time range: 24h / 3d / 7d / 14d / 30d
+
+### Stat Cards
+
+- **Spend** — total team spend, $/day average
+- **Billing Cycle** — day X of Y, days left, reset date
+- **Anomalies** — open count, red border when active
+- **Active** — active members count and % of team
+- **Requests** — total agent requests, /day average
+- **Lines** — total lines added, /day average
+
+### Charts
+
+- **Daily Spend Trend** — area chart with avg line, provisional zone (last 2d), spike detection
+- **Model Cost Comparison** — table with $/request, relative multiplier (1x–8x+), color-coded
+- **Top Spenders** — horizontal bar chart, top 8
+- **Daily Spend by User** — stacked bar, top 6 + Others, clickable legend
+
+### Members Table
+
+- Sortable by: spend, activity, requests, lines, $/req, context, name
+- Filterable by badge type
+- Columns: rank, name, email, spend, requests, lines, $/req, model, profile badges, ranks
+
+### Badges (per user, max 2)
+
+| Category | Badges                                                                                                                                                                |
+| -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| Usage    | Power User, Deep Thinker, Low Usage                                                                                                                                   |
+| Spend    | Cost Efficient, Premium Model, Over Budget                                                                                                                            |
+| Context  | Long Sessions, Short Sessions                                                                                                                                         |
+| Adoption | AI-Native (80%+), High Adoption (55%+), Moderate (30%+), Low Adoption (10%+), Manual Coder (<10%) — based on composite score (accept rate + engagement + consistency) |
+
+### Ranks
+
+- Spend rank ($N) — blue
+- Activity rank (AN) — green
+
+---
+
+## User Detail (`/users/[email]`)
+
+**Per-user deep dive — KPIs, trends, tools, models, anomalies.**
+
+### Header
+
+- Name, email, role, billing group link
+- Profile badges (same as dashboard)
+
+### KPI Cards
+
+- **Cycle Spend** — total $ in billing cycle, $/day
+- **$/Req** — cost per agent request
+- **Agent Reqs** — total requests in time range
+- **Diffs Accepted** — % of agent diffs accepted (accepts/applies), raw counts
+- **Team Rank** — spend and activity rank of N
+
+### Charts
+
+- **Spend Trend** — same area chart as dashboard, per-user
+- **AI Adoption** — tier label (AI-Native/High/Moderate/Low/Manual) with one-line description. Score bar (0-100). Three stat pills: diffs accepted %, requests/day, active days. Composite score from Accept Rate (40%), Engagement Intensity (40%), Consistency (20%). Tooltips on hover with raw numbers.
+
+### Sections
+
+- **Cost Breakdown** — per-model table: requests, $/req, total $, included vs overage bar, errors
+- **Tools & Features** — top 10 MCP tools + top 10 commands used
+- **Context Efficiency** — avg cache read/req, org median, vs org ratio, rank, color-coded band
+- **Model Preferences** — model, days used, requests
+- **Daily Activity Table** — date, model, requests, spend, lines +/-, accepts, tabs, version (spike rows highlighted)
+- **Anomaly History** — detected date, type, severity, message, status
+
+---
+
+## Insights (`/insights`)
+
+**Team analytics — adoption, efficiency, trends.**
+
+### Stat Cards
+
+- Avg DAU, Commands total, Agent lines accepted, Tab lines accepted, MCP tools count
+
+### Sections
+
+- **Plan Exhaustion** — users who exceeded plan, days to exhaust, buckets (1-3d, 4-7d, 8-14d, 15+d)
+- **Model Rankings** — biggest spenders, most/least cost efficient, full scorecard
+- **DAU Chart** — daily active users by type (DAU, Cloud Agent, CLI)
+- **Model Adoption Share** — stacked area, top 5 models over time
+- **Model Usage Breakdown** — table: model, messages, users, % of total
+- **Top File Extensions** — horizontal bar by AI lines accepted
+- **Commands Adoption** — top 20 commands, usage counts
+- **MCP Tool Adoption** — top 20 tools by server, call counts
+- **Client Versions** — pie chart + table with "latest" / "needs update" badges
+
+---
+
+## Anomalies (`/anomalies`)
+
+**Incident monitoring and response tracking.**
+
+### Stat Cards
+
+- Open Anomalies, Resolved, Open Incidents, Avg MTTD, Avg MTTI, Avg MTTR
+
+### Sections
+
+- **Open Incidents** — table with acknowledge/resolve actions
+- **All Anomalies** — table: date, user, type, severity, metric, message, status
+
+---
+
+## Settings (`/settings`)
+
+**Detection thresholds, budget, billing groups.**
+
+### Detection Config
+
+- Static thresholds (max spend/cycle, max requests/day)
+- Z-score detection (std dev multiplier, lookback window)
+- Spend trend detection (spike multiplier, lookback, cycle outlier multiplier)
+- Collection schedule (cron interval)
+- Team budget alert threshold
+
+### Billing Groups
+
+- Group management: create, rename, assign members
+- HiBob CSV import with change preview
+- Backup export/import

src/app/dashboard-client.tsx

@@ -139,10 +139,13 @@ export function DashboardClient({ initialData }: DashboardClientProps) {
     "premium-model": "cpr",
     "cost-efficient": "cpr",
     "power-user": "reqs",
-    "tab-completer": "reqs",
     "deep-thinker": "spend",
     "light-user": "reqs",
-    balanced: "reqs",
+    "manual-coder": "lines",
+    "low-adoption": "lines",
+    "moderate-adoption": "lines",
+    "high-adoption": "lines",
+    "ai-native": "lines",
   };
 
   const handleBadgeFilter = useCallback(
@@ -178,7 +181,8 @@ export function DashboardClient({ initialData }: DashboardClientProps) {
         (u) =>
           u.usage_badge === badgeFilter ||
           u.spend_badge === badgeFilter ||
-          u.context_badge === badgeFilter,
+          u.context_badge === badgeFilter ||
+          u.adoption_badge === badgeFilter,
       );
     }
     const sorted = [...users].sort((a, b) => {