Add new enterprise-weekly-report agent

CLAUDE.md

@@ -11,6 +11,8 @@ When the user makes a request, identify which agent should handle it, read its S
 | If the user asks to... | Activate | Skill file |
 |---|---|---|
 | Create a dashboard, build a dashboard, set up analytics for a feature | **Dashboard Builder** | `agents/dashboard-builder/SKILL.md` |
+| Quick enterprise usage check (SQL-only, no Hex trends or Gong needed) — enterprise usage, account health, how enterprise accounts are doing | **Enterprise Reporter** | `agents/enterprise-reporter/SKILL.md` |
+| Full weekly enterprise report with Hex dashboard trends, comprehensive enterprise analysis with Gong + Hex + BigQuery, enterprise weekly report with 4-week trends | **Enterprise Weekly Report** | `agents/enterprise-weekly-report/SKILL.md` |
 
 If the request doesn't clearly match an agent, ask the user which they need.
 
@@ -42,10 +44,10 @@ Every agent reads from these files. They are the single source of truth — do n
 12. **Validate column names against actual schema** before running queries. Documentation may be outdated.
 13. **Check data quality during query execution.** Inspect results for NULLs, zero/low row counts, date gaps, invalid values (negative counts, retention increases, funnel violations), and schema mismatches. Report issues to users with clear warnings.
 14. **Always use EXACT segmentation CTEs from `shared/bq-schema.md`** without modification. Never simplify or skip steps:
-    - **Enterprise Users:** Must use the two-step pattern (inner CTE + outer SELECT with " Pilot" suffix logic)
+    - **Enterprise Users:** Must use the two-step pattern (`ent_agg` CTE from `ltxstudio_users` + `enterprise` CTE joining `ltxstudio_enterprise_users`). Includes `user_type` (SSO/Code). See bq-schema.md Enterprise Users section.
     - **Heavy Users:** Must include all filters (4+ weeks active, token consumption, etc.)
     - **Full Segmentation:** Must follow hierarchy (Enterprise → Heavy → Paying → Free)
-    - Copy the entire CTE structure from `bq-schema.md` lines 441-516, do not improvise or simplify
+    - Copy the entire CTE structure from `bq-schema.md`, do not improvise or simplify
 
 ## MCP Connections
 

agents/enterprise-weekly-report/references/enterprise-context.md

@@ -0,0 +1,162 @@
+# Enterprise Context
+
+## Enterprise Users CTE
+
+The canonical enterprise users CTE lives in `shared/bq-schema.md` (Enterprise Users section). **Always use that exact CTE verbatim** — do not simplify or rewrite it.
+
+The CTE produces: `lt_id`, `org` (with " Pilot" suffix for non-annual orgs), `user_type` (SSO/Code), `email`, `first_active_ts`, `first_active_ts_tokens`.
+
+### How Org Classification Works
+
+- **Annual orgs** (hardcoded list): `Indegene`, `HearWell_BeWell`, `Novig`, `Cylndr Studios`, `Miroma`, `Deriv`, `McCann_Paris` — org name used as-is
+- **Pilot orgs** (all others): org name gets " Pilot" suffix (e.g., "Monster Pilot", "GM Pilot")
+- **McCann split**: `McCann_NY` and `McCann_Paris` are distinct orgs (COALESCE logic)
+- **Excluded**: Lightricks, Popular Pays, None
+- **User type**: SSO (has `organization_name`) vs Code (redeemed code)
+
+### Tables Used
+
+| Table | Purpose |
+|-------|---------|
+| `ltx-dwh-prod-processed.web.ltxstudio_users` | User attributes, enterprise name resolution |
+| `ltx-dwh-prod-processed.web.ltxstudio_enterprise_users` | Enterprise user membership (JOIN filter) |
+| `ltx-dwh-prod-processed.web.ltxstudio_user_all_actions` | Event-level actions (LT team pre-excluded) |
+| `ltx-dwh-prod-processed.web.ltxstudio_enterprise_calls` | Gong call data with AI-analyzed insights per org |
+
+### Enterprise Calls Table (`ltxstudio_enterprise_calls`)
+
+Contains Gong call recordings matched to enterprise orgs, with AI-analyzed fields:
+
+**Identifiers & Metadata:**
+- `conversation_key` — Natural key for the call
+- `call_started_at` — When the call started
+- `organization_names` — REPEATED array of org names on the call (use UNNEST to flatten)
+- `enterprise_participant_emails` / `enterprise_participant_names` — REPEATED arrays
+- `title` — Call title/subject
+- `call_spotlight_brief` — Gong AI summary of the call
+- `clean_text_transcript` — Full transcript (plain text)
+
+**AI-Analyzed Fields (all STRING, values include "Parse Error" for failed parsing):**
+- `customer_sentiment` — "Positive/Excited", "Neutral/Professional", "Negative/Frustrated"
+- `competitors_mentioned` — Comma-separated list or "None" (tracked: Runway, Freepik, Krea, Higgsfield, Weavy, Luma, Kling, Minimax)
+- `primary_objection` — "Price", "Missing Feature", "Timing", "Authority", "Security", "None"
+- `customer_pain_point` — Max 10 words describing the problem
+- `missing_features` — Features the customer wishes existed
+- `mentioned_*_feature` — "Yes"/"No" for: pitch_deck, collaboration, elements, retake
+- `budget_discussion` — Amount/range or "Not Discussed"
+- `is_decision_maker` — "Yes", "No", "Unclear"
+- `next_steps` — Action items summary
+
+**Query pattern:** UNNEST `organization_names` to join with org-level data:
+```sql
+SELECT org_name, ...
+FROM `ltx-dwh-prod-processed.web.ltxstudio_enterprise_calls` c,
+UNNEST(c.organization_names) AS org_name
+WHERE c.call_started_at >= TIMESTAMP(@report_start_date)
+```
+
+**Important — Org Name Mapping:**
+The `organization_names` field contains raw org names (e.g., "Monster", "NBC", "GM") without the " Pilot" suffix that the CTE adds. When correlating Gong call data with usage metrics, map raw names to their CTE equivalents:
+- Raw "Monster" = CTE "Monster Pilot"
+- Raw "Deriv" = CTE "Deriv" (annual, no suffix)
+- Raw "McCann" → depends on McCann_NY / McCann_Paris split
+
+## Hex Enterprise Dashboard
+
+**URL:** https://app.hex.tech/lightricks_prod/app/LTX-Studio---Enterprise-Dashboard-031766Suw1qd6T5K733wDu/latest
+**Project ID:** `01997093-d0d2-700f-9dcc-e624403815fe`
+**MCP Server:** Hex MCP
+
+The dashboard has 5 tabs providing interactive views of enterprise data:
+
+| Tab | Scope | Metrics |
+|-----|-------|---------|
+| Annual Enterprise Data | Annual orgs | WAU, image/video gens, tokens, WoW trends, model mix, feature adoption |
+| Pilot Enterprise Data | Pilot orgs | WAU, image/video gens, tokens, downloads, WoW trends |
+| Image Generations Data | User-level | Per-user image gen counts, model used, tokens consumed |
+| Video Generations Data | User-level | Per-user video gen counts, model used, tokens consumed |
+| Download/Exports Data | User-level | Image/video downloads, project exports, download rates |
+
+**Hex vs BigQuery differences:**
+- Hex uses `griffin_enterprise_name_at_action` for org resolution — this field is NULL for Cylndr Studios; Hex joins through `ltxstudio_enterprise_users` as fallback
+- Hex has PII restrictions — `email` and `full_name` are protected; users identified by `lt_id` + `email_domain`
+- Hex provides 4-week trend windows vs BigQuery queries' 2-week window
+- Hex org names use raw `organization_name` (e.g., "McCann", "General Motors"), not the CTE's suffixed names (e.g., "McCann Pilot", "General Motors Pilot")
+
+## Org Reference
+
+| Org | Type | Notes |
+|-----|------|-------|
+| McCann_Paris | Annual | Image-heavy (~88% Nano Banana 2), creative agency, 34-42 WAU, dominates token consumption (~73% of annual total) |
+| Deriv | Annual | Financial services, 10-18 WAU, declining tokens (818K→410K), needs portrait video support |
+| HearWell_BeWell | Annual | Single power user, 23% Veo 3 usage (highest of any org), volatile activity |
+| Novig | Annual | Low activity (4 WAU), 86% Nano Banana 2, minimal video |
+| Cylndr Studios | Annual | Small team (2-7 WAU), volatile usage, `griffin_enterprise_name_at_action` is NULL |
+| Indegene | Annual | Healthcare, most balanced feature adoption (Elements 63%, Storyboard 38%, Retake 38%), ramping image generation |
+| Miroma | Annual | Video surge (638% WoW spike), high LTX-2 adoption (14%), Script-to-Storyboard adopter (32%) |
+| McCann Pilot | Pilot | Image-heavy (10K+ images), 6-11 WAU, highest download rate, separate from McCann_Paris |
+| McCann_NY Pilot | Pilot | Low activity (1 WAU), sporadic usage |
+| General Motors Pilot | Pilot | Video-heavy (~6.4M tokens total), VEO-3 dominant, enterprise security key differentiator |
+| Meta Pilot | Pilot | High-volume evaluation (~6M tokens), 9-16 WAU, declining trend, split between VEO-3 and LTX-2 |
+| Fanatics Pilot | Pilot | Sports collectibles, image-focused (6K images), 1-6 WAU, growing user base |
+| Jazz Side Pilot | Pilot | New org, rapid ramp-up (600 video gens last week), high growth trajectory |
+| Plarium Pilot | Pilot | Gaming, massive onboarding spike (1→30 WAU), recently activated |
+| Disney Pilot | Pilot | Broadest pilot user base (7-13 WAU), moderate volumes, exploratory usage |
+| EōS Fitness Pilot | Pilot | Print + digital imagery, 3-5 WAU, declining trend |
+| Bent Image Lab Pilot | Pilot | Animation studio, 2 WAU, low volume but consistent |
+| Bosch Pilot | Pilot | Minimal activity, intermittent usage |
+| Comcast Advertising Pilot | Pilot | In training phase, near-zero generation activity |
+| Telemundo Pilot | Pilot | Broadcasting, near-zero activity |
+| Monster Pilot | Pilot | Energy drinks brand, pilot ended — zero recent activity |
+| NBC Universal Pilot | Pilot | Broadcasting, pilot ended — zero recent activity |
+| Craft Pilot | Pilot | New, single user, no generations yet |
+| Eset Pilot | Pilot | Cybersecurity, pilot stopped |
+
+## Key Business Context
+
+- LTX shifted strategy in ~October 2025 to prioritize Enterprise over self-serve
+- "Active" for enterprise = generated content (not just page views)
+- McCann_Paris is distinctively image-focused (~70 images/user/week vs 5-10 videos)
+- Deriv showed strong LTX-2 adoption but needs portrait video support
+- Token costs vary significantly by model (Nano Banana 2 >> Flux >> LTX-2)
+- Pilot orgs have time-limited evaluations — declining usage is a churn risk
+- Gong call data provides qualitative context — sentiment, feature requests, pain points
+- Real-time qualitative context comes from `#ltx-studio-enterprise` Slack channel (read last 7 days each report run)
+
+## Model Distribution Benchmarks (from Hex Dashboard)
+
+Based on dashboard data, these are the typical model usage patterns:
+
+| Model Group | Type | Typical Share | Notes |
+|-------------|------|---------------|-------|
+| Nano Banana 2 | Image | 47-88% of gens | Dominant image model across all orgs |
+| Flux (Flux 2 Pro) | Image | 5-36% of gens | #2 image model, used for quality-critical work |
+| Z-Image | Image | 1-6% of gens | Newer model, lighter usage |
+| LTX-2 (Pro/Fast) | Video | 1-16% of gens | In-house video model, strongest at Miroma/Indegene |
+| Veo 3 / Veo 3.1 | Video | 1-23% of gens | External video model, highest at HearWell_BeWell |
+| Veo 2 | Video | <1% of gens | Legacy, declining |
+| LTXV (13b) | Video | <1% of gens | Legacy, declining |
+
+## Feature Adoption Benchmarks (from Hex Dashboard)
+
+| Feature | Adoption Range | Benchmark |
+|---------|---------------|-----------|
+| Gen Space | 63-100% | Core workflow — nearly universal |
+| Elements | 20-63% | Moderate — highest at Indegene (63%), lowest at Novig (20%) |
+| Script-to-Storyboard | 0-38% | Growing — zero at McCann_Paris/Cylndr, highest at Indegene (38%) |
+| Retake | 7-38% | Lowest adoption — highest at Indegene/Cylndr (38%), lowest at Deriv (7%) |
+
+## Qualitative Data Sources
+
+1. **Gong calls** (`ltxstudio_enterprise_calls`) — Real-time, queried per reporting week. Shows customer sentiment, competitor mentions, feature interest, objections, and next steps.
+2. **Hex Enterprise Dashboard** — Interactive dashboard accessible via Hex MCP. Provides 4-week trend windows, per-user drill-downs, model distribution charts, and feature adoption visuals. Use for trend depth beyond the 2-week BigQuery window and for cross-validation of metrics.
+
+## Important Column Mappings
+
+- `action_name_detailed` — Use for filtering event types (NOT `action_name`)
+- `source_interaction_id` — Video press count (distinct count)
+- `native_action_id` — Image press count (distinct count)
+- `tokens_charged` — Tokens consumed per action
+- `model_name` — Model user intended to use
+- `fetch_result` — Success/failure of generation (NOT `result`)
+- `page_workspace_name` — Workspace context (gen_space, storyboard)