Customer Snapshot

Not all data sources are required. The dashboard degrades gracefully -- panels with unavailable data show appropriate empty states.

V2 Architecture

The v2 dashboard uses a deterministic two-stage pipeline:

1. assemble.py -- Deterministic data assembler. Accepts --jira, --bq, --asana, --sentiment JSON file arguments, applies component/parent normalization, theme clustering, trending metrics computation, and Asana task transformation. Outputs complete INTELLIGENCE_DATA JSON.
2. compose.py -- Folder composition. Takes INTELLIGENCE_DATA JSON and assembles template files (shell.html, panel JS, lib/) into a working dashboard folder. Only data.js changes on each refresh -- shell and panel templates are stable.

Output Structure

customers/<name>/dashboard/
  index.html        -- Main shell (loads panels dynamically)
  data.js           -- INTELLIGENCE_DATA constant (refreshable)
  panels/           -- Individual visualization panel JS files
  lib/              -- Shared libraries (echarts.min.js)
  history/          -- Archived previous data.js snapshots

CLI Usage

# Step 1: Assemble data from sources
uv run --project .claude/skills/deep-analytics python \
    .claude/skills/customer-snapshot/templates/assemble.py \
    --customer "GResearch" \
    --jira /path/to/jira.json \
    --bq /path/to/bq.json \
    --asana /path/to/asana.json \
    --sentiment /path/to/sentiment.json \
    --output /path/to/data.json

# Step 2: Compose dashboard folder
uv run --project .claude/skills/customer-snapshot python \
    .claude/skills/customer-snapshot/templates/compose.py \
    --customer "GResearch" --data /path/to/data.json \
    --output customers/g-research/dashboard/

Pipeline

Step 1: Parse customer name

Extract the customer name from the user's input. Common patterns:

• "/customer-tracker GResearch" -> "GResearch"
• "build a tracker for Acme Corp" -> "Acme Corp"
• "customer dashboard for G-Research" -> "G-Research"

Step 2: Load customer registry

Read templates/customers.yaml (project root, NOT inside skill directory) to look up customer configuration:

# Read the file using Claude's Read tool
# Path: templates/customers.yaml

• Find matching customer by name field (case-insensitive, ignore hyphens/spaces)
• If found: extract slack_channels for Step 4, jira_customer for Step 3
• If not found: proceed with Jira-only mode, sentiment will show "Not configured"
• If slack_channels[].id is "PLACEHOLDER": warn and skip Slack fetch
• Optional: if customer has component_normalize map, use it to override built-in normalization

Step 3: Fetch Jira data

Use the Jira skill to pull all issues for the customer with comment metadata:

uv run --project .claude/skills/jira python .claude/skills/jira/scripts/issues.py list \
  --customer "<CustomerName>" --max-results 200 --with-comments

The --with-comments flag includes per-issue comment analysis: comment count, last comment date/author, last eng comment date/author (excluding FE-UPDATE), first comment date, and FE-UPDATE count. This data powers the dashboard's analysis section (staleness, velocity, response cadence). The response also includes resolutiondate for accurate time-to-resolution metrics.

Parse the JSON output. If no issues are returned, generate a dashboard with an empty state message rather than failing.

Step 3.5: Fetch Asana tasks

If the customer has action_tracker_id in customers.yaml (and it is not "PLACEHOLDER", and action_tracker is "asana"):

1. Determine current user GID:

   uv run --project .claude/skills/asana python .claude/skills/asana/scripts/query.py project --gid <action_tracker_id> --pretty
   

   (The user GID comes from the PAT owner, resolved during task filtering)

2. Fetch all incomplete tasks in the customer's Asana project:

   uv run --project .claude/skills/asana python .claude/skills/asana/scripts/query.py tasks \
     --project-gid <action_tracker_id> --limit 100 --pretty
   

3. Filter to incomplete tasks only (completed=false or null)

4. For each task, compute:

   • overdue: true if due_on is before today and task is not completed
   • stale: true if (today - modified_at) > 7 days AND section is "To Do" or "In Progress"
   • stale_days: days since modified_at
   • priority: from custom_fields Priority enum, or parsed from name prefix [P0]/[P1]/[P2]/[P3], or null
   • linked_jira: extracted from task name using regex \(WB-\d+\), or from notes field
   • section: from memberships[0].section.name

5. Default scope: filter to tasks where assignee matches current user ("my tasks")

6. Build the actions object for INTELLIGENCE_DATA (see schema in Step 7)

If action_tracker_id is missing or "PLACEHOLDER": set actions: { available: false, reason: "not_configured" } and proceed. If Asana API fails: set actions: { available: false, reason: "api_error" } and proceed gracefully.

Step 3.7: Fetch BigQuery usage data

If the customer has sfdc_account_id in customers.yaml (and it is not "PLACEHOLDER"):

uv run --project .claude/skills/bigquery python .claude/skills/bigquery/scripts/usage.py \
  --customer "<CustomerName>" --format json

Parse the JSON output. The output matches the INTELLIGENCE_DATA.usage schema (see Step 7).

The usage data powers ECharts time-series and radar charts in the dashboard's Usage panel (replacing the previous CSS horizontal bars). ECharts is loaded from CDN and themed to match the design system. The dashboard coexists: Jira/Slack panels use CSS bars, Usage panel uses ECharts.

If available: false: set usage: { available: false, reason: "<from output>" } and proceed. If sfdc_account_id is missing or "PLACEHOLDER": set usage: { available: false, reason: "not_configured" } and proceed. If BigQuery skill fails: set usage: { available: false, reason: "api_error" } and proceed gracefully.

Step 4: Fetch Slack channel history

For each channel in slack_channels where id is not "PLACEHOLDER":

OLDEST=$(python3 -c "import time; print(time.time() - DAYS*86400)")
uv run --project .claude/skills/slack python .claude/skills/slack/scripts/channels.py history \
  --channel <CHANNEL_ID> --limit 200 --oldest $OLDEST

Default DAYS = 14 (configurable via --days flag on skill invocation). Fetch sequentially, not in parallel (rate limit safety). If Slack API fails or no channels configured: set sentiment to null, proceed gracefully.

Step 5: Analyze sentiment

Claude reads the fetched Slack messages and produces a structured sentiment object:

1. Read messages chronologically
2. Score overall channel tone using 5-point categorical scale:
   • positive (happy, grateful, optimistic)
   • neutral (routine communication)
   • cautiously-negative (hints of frustration)
   • negative (clear complaints)
   • critical (anger, threats to leave, executive escalation)
3. Assign supplementary numeric score (-1.0 to 1.0)
4. Identify hot threads: threads where 3+ messages have negative tone, OR executive expressing frustration, OR mentions of evaluating alternatives
5. For each hot thread, fetch full replies via threads.py replies
6. Produce the sentiment JSON matching INTELLIGENCE_DATA.sentiment schema (see below)
7. For internal section: note risk signals and recommended actions

Output a JSON object -- do not display to user, inject into INTELLIGENCE_DATA.

Step 6: Cluster issues by theme + compute trending metrics

Group issues into product-area themes using Jira field data:

1. Use the issue's first component if available (apply COMPONENT_NORMALIZE map)
2. Otherwise, use the parent epic summary if available (apply PARENT_NORMALIZE map)
3. Otherwise, assign to "Uncategorized"

Labels are skipped entirely -- W&B Jira labels are meta/triage labels (fe-reported, vis-triage), not product areas.

Normalization maps: The template includes COMPONENT_NORMALIZE and PARENT_NORMALIZE JavaScript objects that merge variant names into canonical themes (e.g. "Weave Python SDK" and "weave" both map to "Weave"). These maps are customer/project-specific and should be updated when generating dashboards for new customers. If the customer registry has a component_normalize map, use it to override or extend the built-in maps.

Additionally compute trending metrics:

• Compute opened/closed by month for last 6 months
• Compute raised-to-resolved ratio
• Compute median time-to-resolution using resolutiondate (fall back to updated if null)
• Compute top 5 theme recurrence

Aim for 5-10 themes. Theme names should be short, recognizable product areas. Some Uncategorized issues are expected -- the analysis section flags this as a metric.

Step 7: Build INTELLIGENCE_DATA object

Transform all gathered data into the INTELLIGENCE_DATA constant:

const INTELLIGENCE_DATA = {
  customer: "G-Research",
  generated: "2026-03-17",
  config: {
    sentiment_days: 14,       // --days flag, default 14
    trending_months: 6,        // 6-month lookback
    audience: "internal"       // default view mode
  },
  issues: [
    {
      key: "WB-1234",
      summary: "SDK crash on large artifact upload",
      type: "Bug",           // "Bug" or "Feature Request"
      priority: "P1",        // "P0", "P1", "P2", "P3"
      status: "In Progress", // Raw Jira status value
      assignee: "Jane Doe",  // or null if unassigned
      theme: "SDK & Client Libraries",
      created: "2026-01-15",
      updated: "2026-03-08",
      resolutiondate: null,  // or ISO date string for resolved issues
      url: "https://coreweave.atlassian.net/browse/WB-1234",
      components: ["Weave Python SDK"],
      parent: "WB-900",
      parent_summary: "Weave SDK Improvements",
      comments: {
        comment_count: 5,
        last_comment_date: "2026-03-01T10:30:00.000+0000",
        last_comment_author: "Jane Doe",
        last_eng_comment_date: "2026-02-28T14:00:00.000+0000",
        last_eng_comment_author: "John Smith",
        first_comment_date: "2026-01-16T09:00:00.000+0000",
        fe_update_count: 2
      }
    }
  ],

  // Sentiment (populated by Step 5, null when Slack unavailable)
  sentiment: {
    available: true,
    channels_analyzed: ["#ext-gresearch"],
    period: { start: "2026-03-03", end: "2026-03-17" },
    overall: {
      score: "cautiously-negative",  // positive | neutral | cautiously-negative | negative | critical
      numeric: -0.3,                 // -1.0 to 1.0
      summary: "Tone shifted negative this week, driven by frustration with SDK stability."
    },
    hot_threads: [
      {
        channel: "#ext-gresearch",
        thread_ts: "1710500000.000000",
        summary: "Frustration about repeated SDK crashes blocking production training",
        sentiment: "negative",
        message_count: 12,
        participants: 4,
        url: "https://coreweave.slack.com/archives/C0XXX/p1710500000000000"
      }
    ],
    internal: {
      raw_analysis: "Detailed sentiment breakdown...",
      risk_signals: ["Repeated mentions of evaluating alternatives"],
      recommended_actions: ["Escalate SDK stability to P0"]
    }
  },

  // Trending (computed client-side from issues data in JS)
  trending: null,

  // Executive summary (computed client-side from issues + sentiment in JS)
  exec_summary: null,

  // SE Actions from Asana (populated by Step 3.5, null/unavailable when Asana not configured)
  actions: {
    available: true,           // false when Asana not configured or fetch failed
    source: "asana",
    current_user: { gid: "12345", name: "Allan Stevenson" },
    scope: "my_tasks",         // or "team"
    project_gid: "98765",
    project_url: "https://app.asana.com/0/98765",
    tasks: [
      {
        gid: "11111",
        name: "Follow up on SDK crash (WB-1234)",
        section: "In Progress",
        due_on: "2026-03-28",
        overdue: false,          // computed: due_on < today
        stale: false,            // computed: 7+ days since modified_at AND section in [To Do, In Progress]
        stale_days: 2,
        priority: "P1",          // from custom field or parsed from name prefix [P1]
        assignee: { gid: "12345", name: "Allan Stevenson" },
        linked_jira: "WB-1234", // extracted from name via regex \(WB-\d+\)
        slack_source: "https://coreweave.slack.com/archives/...",
        url: "https://app.asana.com/0/0/11111",
        modified_at: "2026-03-21T10:00:00Z"
      }
    ],
    summary: {
      total: 8,
      in_progress: 3,
      waiting: 2,
      todo: 2,
      overdue: 1,
      stale: 1
    }
  },

  // Usage data from BigQuery (populated by Step 3.7, null/unavailable when BQ not configured)
  usage: {
    available: true,           // false when BigQuery not configured or fetch failed
    period: { start: "2025-03-24", end: "2026-03-24" },
    seat_utilization: {
      contracted: 50, claimed: 42, active: 35,
      utilization_percent: 70.0,
      history: [{ week: "2025-04-07", contracted: 50, active: 28 }]
    },
    weave: {
      ingestion_gb: 156.3, limit_gb: 500.0, utilization_percent: 31.3,
      unique_users_last_90d: 12,
      history: [{ month: "2025-04", ingestion_gb: 8.2, unique_users: 5 }]
    },
    tracked_hours: {
      last_30d_hours: 1250.0, last_30d_run_count: 342,
      history: [{ week: "2025-04-07", tracked_hours: 180.5 }]
    },
    account_health: {  // internal-only
      renewal_date: "2026-09-15", arr: 250000.0, cs_tier: "Strategic",
      customer_health: "Green", churn_probability_3mo: 0.05,
      churn_probability_5mo: 0.08, subscription_plan: "Enterprise",
      deployment_type: "dedicated-cloud"
    },
    trends: {
      seat_utilization_change: 12.5, weave_ingestion_change: -3.2,
      tracked_hours_change: 8.7, run_count_change: 15.3
    },
    product_areas: [  // NEW - from Plan 01 expansion, powers radar chart
      { area: "Experiments", total_events: 1800, unique_users: 25,
        monthly_events: [{month: "2025-04", count: 150, users: 12}] }
    ],
    power_users: [  // NEW - anonymized by default, real names with --internal
      { username: "alice_ml", total_events: 5000, product_areas: ["Experiments"],
        last_activity: "2026-03-20" }
    ]
  }
};

The comments object is present when data is fetched with --with-comments. The template uses it for:

• Health buckets: Activity-based classification (Needs Triage / Active / Stale / Resolved)
• Attention callouts: Never commented, no eng activity 60+ days, unassigned, recently opened
• Response cadence: Median days to first comment, % within 7 days, zero-comment count
• FE-UPDATE badges: Gold "SE:N" badge in the Last Activity column

FE-UPDATE comments are excluded from eng activity calculations -- SE posting an update doesn't reset the staleness clock.

Note: trending and exec_summary are computed client-side in the template JS, not server-side. The data they need (issues with dates, sentiment object) is already in INTELLIGENCE_DATA.

For priority mapping: use P0/P1/P2/P3 directly from Jira. If priority uses names like "Critical"/"High"/"Medium"/"Low", map to P0/P1/P2/P3 respectively.

Step 8: Generate the intelligence dashboard

Use the v2 two-stage pipeline:

1. Save data source JSON files to temp directory
2. Run assemble.py with --jira, --bq, --asana, --sentiment arguments to produce INTELLIGENCE_DATA JSON
3. Run compose.py with --customer, --data, --output to assemble the dashboard folder
4. Output folder: customers/<kebab-case-name>/dashboard/
   • Example: customers/g-research/dashboard/
5. Open the dashboard: open customers/g-research/dashboard/index.html

The shell template and panel JS files handle all rendering -- charts, filters, theme sections, sentiment panel, trending, exec summary, animations. The data.js file contains the INTELLIGENCE_DATA constant and is the only file that changes on each refresh. Previous data.js snapshots are archived in history/.

Step 9: Present to the user

Tell the user:

• Dashboard folder path and how to open it (open .../dashboard/index.html)
• Brief summary: total issues, sentiment score (or "not configured"), backlog trajectory
• Usage stats if available (seat utilization %, Weave ingestion %, trend direction)
• Asana action counts if available (e.g., "8 open actions, 1 overdue")
• Hot issues or threads needing attention
• Flag any P0/P1 stale issues

Design Rules

Read references/design-system.md for the complete visual specification. Key principles:

• Custom HTML/CSS horizontal bars for Jira/Slack panels -- no external charting libraries for those
• Instrument Serif + Outfit + JetBrains Mono typography
• Gold accent (#d4a853), warm cream light mode (#f6f4f0), deep navy dark mode (#0c0f14)
• Inline header stats, not KPI cards
• Pill toggle filters with "All" default
• Collapsible theme sections with priority mini-badges
• Resolved issues hidden by default with toggle to show
• Both light and dark mode via prefers-color-scheme
• Internal/External audience toggle in header area

Visualization

• Usage panel: Apache ECharts (CDN, custom 'wandb' theme) -- seat utilization (line), product adoption (radar), weave ingestion (bar), tracked hours (bar)
• Jira/Slack panels: Custom CSS bars (unchanged from Phase 3/6)
• Both coexist in the same template
• ECharts loaded from https://cdn.jsdelivr.net/npm/echarts@5/dist/echarts.min.js
• Custom theme registered as 'wandb' matching design system colors (dark/light mode aware)

Status Mapping

The template maps Jira statuses to display categories for filter pills and issue badges:

Health Buckets (Analysis Section)

The analysis section uses activity-based health classification layered on top of raw statuses:

FE-UPDATE comments are excluded from eng activity -- only non-FE-UPDATE comments count. This prevents SE updates from gaming the staleness metric.

Analysis Section

The dashboard's primary value is the analysis section above the issue list:

1. Health summary: Horizontal bar showing bucket distribution (Needs Triage / Active / Stale / Resolved)
2. Attention callouts: Clickable cards that filter the issue list:
   • Never Commented (zero comments)
   • No Eng Activity 60+ days (VERY_STALE_DAYS)
   • Unassigned (no assignee)
   • Recently Opened (last 14 days)
3. Velocity chart: Opened vs resolved by month (last 6 months), custom CSS bars
4. Response cadence: Median days to first comment, % responded within 7 days, zero-comment count

Callouts are interactive -- clicking one filters the issue list, auto-expands matching themes, and scrolls to the issue section.

Anti-Patterns

When generating or modifying dashboards, never introduce these patterns:

• Chart.js or any external JS charting library
• KPI cards (big number + small label in identical card grid)
• Gradient text, glassmorphism, neon accents
• Inter/Roboto/Arial fonts
• Sidebar navigation
• Copying from the prototype at .claude/skills/customer-tracker/ in older commits
• LLM-based theme clustering (use Jira components/labels only)

Future View Types

The skill is structured for multiple view modes of the same data. The Intelligence Dashboard is the primary view:

1. Intelligence Dashboard (current) -- filterable dashboard with theme groupings, sentiment, trending, exec summary
2. Timeline (planned) -- issues plotted against target dates
3. Slide Deck (planned) -- presentation mode for QBRs
4. Confluence Page (planned) -- XHTML for publishing

Troubleshooting