codemie-analytics

name

codemie-analytics

description

CodeMie Analytics expert — use this skill whenever the user asks about CodeMie usage data, AI adoption metrics, user leaderboards, CLI insights, spending, LiteLLM costs, token usage, or wants to build a dashboard/report from CodeMie or LiteLLM APIs. Also triggers for: "who uses CodeMie most", "show me AI analytics", "get spending data", "generate a report", "leaderboard", "cost analysis", "LiteLLM customer info", "enrich CSV with costs", "top performers", "AI champions", "tier distribution", or any custom analytics query against the platform. Always use this skill when CodeMie analytics, reporting, or cost data is involved.

CodeMie Analytics Skill

You are an analytics expert for the CodeMie (EPAM AI/Run) platform. You know every analytics API endpoint, how to call LiteLLM directly, and how to orchestrate data into a final report.

The plumbing (config lookup, SSO credential decryption, token refresh messaging) lives in scripts/analytics-cli.js. You never need to touch those details — just invoke the CLI and react to what it prints.

---

Step 1 — Understand what the user wants

Identify the analytics scenario. The CLI supports these command families:

Leaderboard (AI Champions)

The leaderboard ranks users across 6 scoring dimensions:

• D1: Core Platform Usage (20%) — conversations, assistant interactions
• D2: Core Platform Creation (20%) — assistants, datasources created
• D3: Workflow Usage (10%) — workflow executions
• D4: Workflow Creation (10%) — workflows authored
• D5: CLI & Agentic Engineering (30%) — coding agent sessions, tokens, repos
• D6: Impact & Knowledge (10%) — marketplace publishing, knowledge sharing

Tiers: pioneer (80+), expert (65+), advanced (45+), practitioner (25+), newcomer (<25)

Scenario	Command	What it retrieves
Full leaderboard (paginated, filterable)	leaderboard	data.rows[] — rank, user_name, total_score, tier_name, score_delta, dimensions[] (id/score/weight per D1–D6), summary_metrics{} (cli_sessions, active_days, total_lines_added, total_spend, web_conversations, …)
Leaderboard KPI summary	leaderboard-summary	data — total_users, avg_score, top_score, tier_counts{} (pioneer/expert/advanced/practitioner/newcomer counts and percentages)
Single user champion profile	leaderboard-user <id|email>	data — same shape as a leaderboard row but for one user; includes full dimension breakdown and all summary_metrics
Tier distribution	leaderboard-tiers	data.rows[] — tier_name, user_count, percentage; one row per tier
Average dimension scores	leaderboard-dimensions	data.rows[] — dimension id/label, avg_score, weight; one row per D1–D6
Top N performers	leaderboard-top [limit]	data.rows[] — same shape as leaderboard rows, limited to top N (max 50, default 10)
Score histogram	leaderboard-scores	data.rows[] — score_range (e.g. "0-10"), user_count; one row per 10-point bin
Framework metadata	leaderboard-framework	data.framework{} — title, principles, calculation_steps; data.tiers[] — name, label, min_score; data.dimensions[] — id, label, weight, description
Computation snapshots	leaderboard-snapshots	data.rows[] — snapshot_id, created_at, status, period_start, period_end, user_count
Available seasons	leaderboard-seasons --view monthly|quarterly	data.rows[] — season_key (e.g. "2026-03"), label, start_date, end_date

Leaderboard filters: --view (current/monthly/quarterly), --season-key (2026-03, 2026-Q1), --tier, --intent (cli_focused/platform_focused/hybrid/sdlc_unicorn), --search, --sort-by, --sort-order.

CLI Insights

Scenario	Command	What it retrieves
Full CLI overview (agents, repos, tools, errors)	cli-insights	Composite object with sub-keys: summary (total sessions, cost, tokens, repos, users), agents[] (agent name + session count), top_users[], top_repos[], errors[], llms[]
User classification & top spenders	cli-insights-users	data.rows[] — user_name, classification (cli_focused/platform_focused/hybrid/sdlc_unicorn), total_cost, session_count, token_count
Detailed single-user CLI profile	cli-insights-user <name>	Composite: key_metrics (sessions, cost, tokens, repos, tools), tools[], models[], repositories[], workflow_intent, category_breakdown[]
Project classification & top by cost	cli-insights-projects	data.rows[] — project_name, classification, total_cost, session_count, user_count
Usage patterns (weekday, hourly, session depth)	cli-insights-patterns	Composite with sub-keys: weekday.data.rows[] (weekday_name, session_count), hourly.data.rows[] (hour_utc, session_count), session_depth.data.rows[] (depth_bucket, count)

General Analytics

Scenario	Command	What it retrieves
Overall usage summary (tokens, cost, users)	summaries	data — total_cost, total_tokens, total_requests, unique_users (MAU), unique_users_daily (DAU), cli_invocations, assistants_count, workflows_count, skills_count, mcp_servers_count
User list + activity trends	users	data.rows[] — user_name, email, total_cost, total_tokens, last_active; plus activity[] time-series
Per-project spending	projects-spending	data.rows[] — project_name, total_cost, total_tokens, user_count, request_count
Per-project activity time-series	projects-activity	Composite with activity.data.rows[] and uniqueDaily.data.rows[]
LLM model breakdown	llms-usage	data.rows[] — model_name, request_count, total_tokens, input_tokens, output_tokens, total_cost
Tool usage	tools-usage	data.rows[] — tool_name, invocation_count, success_count, error_count, total_tokens
Workflow execution analytics	workflows	data.rows[] — workflow_name, run_count, success_count, failure_count, avg_duration_ms, total_cost
Agent execution analytics	agents-usage	data.rows[] — assistant_name, execution_count, total_cost, total_tokens
Embedding model usage	embeddings-usage	data.rows[] — model_name, request_count, total_tokens, total_cost
Chat assistant conversations	assistants-chats	data.rows[] — assistant, conversation_count, user_count, total_cost
Webhook invocation analytics	webhooks-usage	data.rows[] — user_id, invocation_count, total_cost
MCP server usage

LiteLLM & CSV Enrichment

Scenario	Command	Output
LiteLLM customer lookup	litellm-customer [user_id]	JSON
LiteLLM spend logs	litellm-spend	Spend entries
LiteLLM virtual keys	litellm-keys	Key info
Enrich CSV/Excel with LiteLLM costs	enrich-csv <file>	Enriched table
Unlisted / experimental endpoint	custom /v1/analytics/<path>	Raw JSON — use a named command if one exists

---

Security

Never include raw API keys, bearer tokens, cookie values, LiteLLM keys, or session credentials in your responses to the user. If CLI output contains sensitive fields (e.g. from litellm-keys), the CLI automatically redacts them — but if you encounter any token, key, or secret in raw output, redact it before displaying. Never run env, printenv, or similar commands that could expose LITELLM_KEY or CODEMIE_API_KEY to the conversation context.

---

Step 2 — Run the analytics CLI

The CLI script lives at scripts/analytics-cli.js next to this skill. It handles authentication internally. If something is wrong with credentials, it prints a clear actionable message to stderr; pass that along to the user verbatim.

node ${CLAUDE_PLUGIN_ROOT}/skills/codemie-analytics/scripts/analytics-cli.js <command> [options]

LiteLLM commands (litellm-*, enrich-csv) require LITELLM_URL + LITELLM_KEY env vars.

Collecting data for reports (batch pattern)

Only fetch the data the report actually needs. Read the user's request, identify which sections the report requires, and collect only the relevant endpoints. Do not run every available command — unnecessary fetches waste time and tokens.

When building an HTML report, run all needed CLI commands in a single Bash call using --save <filepath> on each. This triggers one permission prompt for the entire collection phase and keeps API responses out of the conversation context.

Directory layout

Every report lives in its own folder under reports/. Derive the folder name from the current date and a short kebab-case description of the report:

reports/
  2026-05-07-cli-usage/          ← report folder (date + name)
    cli-usage.html               ← the HTML report (saved here directly)
    temp/                        ← all temp/data files go here
      summaries.json
      summaries.schema.json
      cli-insights.json
      ...

Never overwrite an existing report folder. Always resolve a free name before creating anything. Use a suffix loop — this is mandatory, not optional:

BASE=reports/$(date +%Y-%m-%d)-<short-name>
REPORT_DIR=$BASE
n=2
while [ -d "$REPORT_DIR" ]; do REPORT_DIR="${BASE}-${n}"; n=$((n+1)); done
OUT="$REPORT_DIR/temp"

Then run only the commands the report needs:

CLI=${CLAUDE_PLUGIN_ROOT}/skills/codemie-analytics/scripts/analytics-cli.js
mkdir -p "$OUT" && \
node $CLI summaries   --save "$OUT/summaries.json"   && \
node $CLI cli-insights --save "$OUT/cli-insights.json" && \
# ... only endpoints needed for this report ...
echo "✓ All data saved → $OUT"

Each command prints: ✓ Saved → <path>. The final echo confirms the directory path — save it, you will reference it in every subsequent step.

Do not cat, Read, or print the saved JSON files into the conversation. Raw API responses can be hundreds of KB. Use Step 2.5 to inspect structure instead.

Temp files are not cleaned up automatically.

Common filter flags

Flag	Example	Notes
--time-period	last_30_days	Predefined period
--start-date	2024-01-01T00:00:00	Custom range start
--end-date	2024-03-31T23:59:59	Custom range end
--users	alice,bob	Comma-separated usernames
--projects	my-project	Comma-separated project names
--page	1	Pagination
--per-page	100	Results per page (default 50)
--output	json	json | table | csv
--pretty	(flag)	Pretty-print JSON

Leaderboard-specific flags

Flag	Example	Notes
--view	monthly	current | monthly | quarterly
--season-key	2026-Q1	Specific season to query
--tier	pioneer	Filter by tier
--intent	cli_focused	Filter by user intent profile
--search	john	Partial name/email search
--sort-by	total_score	rank | total_score | user_name | tier_level
--sort-order	desc	asc | desc
--limit	20	Max entries for leaderboard-top (max 50)

Example invocations

CLI=${CLAUDE_PLUGIN_ROOT}/skills/codemie-analytics/scripts/analytics-cli.js

# Full leaderboard — top 50 pioneers sorted by score
node $CLI leaderboard --tier pioneer --sort-by total_score --sort-order desc --per-page 50 --pretty

# Single user champion profile
node $CLI leaderboard-user [e-mail] --pretty

# Leaderboard KPI summary for Q1 2026
node $CLI leaderboard-summary --view quarterly --season-key 2026-Q1 --pretty

# Dimension averages (D1–D6) for current snapshot
node $CLI leaderboard-dimensions --pretty

# Tier distribution
node $CLI leaderboard-tiers --pretty

# Top 10 performers
node $CLI leaderboard-top 10 --pretty

# 30-day platform summary
node $CLI summaries --time-period last_30_days --pretty

# Full CLI insights
node $CLI cli-insights --time-period last_30_days --pretty

# Detailed CLI profile for a specific user
node $CLI cli-insights-user [e-mail] --time-period last_30_days --pretty

# Usage patterns (weekday + hourly + session depth)
node $CLI cli-insights-patterns --time-period last_30_days --pretty

# Per-user spending breakdown
node $CLI spending-by-users --time-period last_30_days --pretty

# Custom endpoint
node $CLI custom /v1/analytics/mcp-servers --time-period last_30_days --pretty

---

Step 2.5 — Inspect data structure

After collection, run inspect-schema.js to generate a compact .schema.json file alongside each saved JSON — one permission prompt, no data in context:

node ${CLAUDE_PLUGIN_ROOT}/skills/codemie-analytics/scripts/inspect-schema.js "$OUT"

Output is a short listing of generated schema files with sizes, for example:

Schemas written to: reports/2026-05-07-cli-usage/temp/
  ✓ leaderboard-top.schema.json (1.3 KB)
  ✓ summaries.schema.json (0.4 KB)
  ...

Then read only the .schema.json files you actually need for the report using the Read tool. Each schema shows field names, types, array lengths, and string samples — enough to write extraction code without touching the raw responses:

CRITICAL — schema is the source of truth for field names. Never use field names or metric IDs from this skill's documentation when writing report code. Documentation can be stale. The .schema.json files are generated from live API responses and are always correct. Every id, key, or column name referenced in JS must be verified against the schema you just read — not assumed from the tables above. If the schema contradicts the docs, trust the schema.

{
  "_envelope": "[envelope] data_as_of: '2026-05-07T02:00:10', total_count: 50",
  "data": {
    "rows": {
      "_type": "array",
      "_count": 50,
      "_item": {
        "rank": "number",
        "user_name": "string ~ 'Jane Smith'",
        "total_score": "number",
        "tier_name": "string ~ 'expert'",
        "score_delta": "number | null",
        "dimensions": { "_type": "array", "_count": 6, "_item": { "id"       
                  
      
    
  

---

Step 3 — Build the HTML report

Save the HTML report directly inside $REPORT_DIR (not in temp/). Use a kebab-case filename matching the folder name:

reports/2026-05-07-cli-usage/cli-usage.html
reports/2026-05-07-leaderboard/leaderboard.html
reports/2026-05-07-spending/spending.html

Step 3a — Write the HTML

Invoke the codemie-html-report skill. Pass:

1. The schemas inspected in Step 2.5 — field names and structure.
2. The user's intent — e.g. "leaderboard dashboard with tier distribution".
3. Timestamp context — _envelope lines in schemas include data_as_of.
4. Output path — e.g. $REPORT_DIR/leaderboard.html.

The HTML skill writes the file using /*__CSS__*/ for styles and /*__DATA:name__*/ placeholders for all embedded JS data arrays:

<script>
  const LEADERBOARD = /*__DATA:leaderboard__*/;
  const SUMMARIES   = /*__DATA:summaries__*/;
</script>

Each placeholder name maps to a saved file in $OUT (without the .json extension).

Step 3b — Inject data

After the HTML file exists, run the shared inject-data.js from the codemie-html-report skill. It matches each JSON file in $OUT to a /*__DATA:name__*/ placeholder by filename (without .json) and replaces it in-place.

Do not run inject-data.js before the HTML file is written.

node ${CLAUDE_PLUGIN_ROOT}/skills/codemie-html-report/scripts/inject-data.js \
  "$REPORT_DIR/<report>.html" "$OUT"

Placeholder names must exactly match the JSON filenames — /*__DATA:leaderboard-top__*/ is only replaced when leaderboard-top.json exists in $OUT.

Expected output:

  ✓ injected leaderboard-top
  ✓ injected summaries
✓ 2 data block(s) injected into reports/2026-05-07-<name>/<report>.html

---

Full API Reference

Leaderboard endpoints (GET /v1/analytics/leaderboard/...)

Admin-only. All accept snapshot_id, view, season_key query params.

Endpoint	Additional Params	Returns
/leaderboard/summary	—	Total users, tier counts, top score
/leaderboard/entries	tier, search, intent, sort_by, sort_order, page, per_page	Paginated ranked entries
/leaderboard/user/{user_id}	path: user ID or email	Full user profile with D1–D6 breakdown
/leaderboard/tiers	—	Tier name, count, percentage
/leaderboard/scores	—	Score histogram (10-point bins)
/leaderboard/dimensions	—	Average D1–D6 scores
/leaderboard/top-performers	limit (default 3, max 50)	Top N by total score
/leaderboard/snapshots	view, status, is_final, page, per_page	Computation snapshots
/leaderboard/seasons	view (required: monthly/quarterly), page, per_page	Available seasonal periods
/leaderboard/framework	—	Static metadata: dimensions, tiers, intents, scoring
/leaderboard/compute (POST)	period_days, view, season_key	Triggers manual computation

CLI Insights endpoints (GET /v1/analytics/cli-insights-...)

Endpoint	Params	Returns
/cli-insights-weekday-pattern	time filters	Weekday usage patterns
/cli-insights-hourly-usage	time filters	Hourly usage patterns
/cli-insights-session-depth	time filters	Session depth distribution
/cli-insights-user-classification	time filters	User classification breakdown
/cli-insights-top-users-by-cost	time filters	Top users ranked by cost
/cli-insights-top-spenders	time filters	Top spenders
/cli-insights-users	time filters	CLI user list
/cli-insights-user-detail	user_name (required), user_id	Full user detail
/cli-insights-user-key-metrics	user_name (required), user_id	User KPIs
/cli-insights-user-tools	user_name (required), user_id	User tool usage
/cli-insights-user-models	user_name (required), user_id	User model usage
/cli-insights-user-workflow-intent	user_name (required), user_id	User workflow intent
/cli-insights-user-classification-detail	user_name (required), user_id	User classification detail
/cli-insights-user-category-breakdown	user_name (required), user_id	User category breakdown
/cli-insights-user-repositories	(required),

Standard CLI analytics (GET /v1/analytics/cli-...)

Endpoint	Returns
/cli-summary	CLI totals (tokens, cost, sessions)
/cli-agents	Agent breakdown
/cli-llms	Model breakdown
/cli-users	CLI user activity
/cli-errors	Error logs
/cli-repositories	Repo activity
/cli-top-performers	Top by lines added
/cli-top-versions	CLI version distribution
/cli-top-proxy-endpoints	LiteLLM endpoint usage
/cli-tools	Tool usage

Dashboard analytics (GET /v1/analytics/...)

All accept time filters + users + projects + page + per_page.

Endpoint	Returns
/summaries	Platform totals (tokens, cost, unique users)
/users-spending	Per-user cost + tokens
/users-activity	Activity time-series
/users-unique-daily	Unique users/day
/users	User list
/projects-spending	Per-project spending
/projects-activity	Project activity time-series
/projects-unique-daily	Unique projects/day
/llms-usage	LLM model usage
/tools-usage	Tool usage
/workflows	Workflow runs
/agents-usage	Agent executions
/embeddings-usage	Embedding model usage
/assistants-chats	Chat assistant conversations
/webhooks-invocation	Webhook usage
/mcp-servers	MCP server usage
/mcp-servers-by-users	MCP by user
/power-users	Power user analytics
/knowledge-sharing	Knowledge sharing metrics
/top-agents-usage	Top agents
/top-workflow-usage	Top workflows
/published-to-marketplace	Marketplace publishing

Spending & Budget endpoints

Endpoint	Method	Returns
/spending	GET	Current user: spend, budget_limit, hard_budget_limit, reset time
/budget_usage	GET	Per-key budget rows with % used
/budget-soft-limit	GET	Soft limit warnings
/budget-hard-limit	GET	Hard limit hits
/spending/by-users/platform	GET	Per-user platform spending
/spending/by-users/cli	GET	Per-user CLI spending

Engagement

Endpoint	Method	Returns
/engagement/weekly-histogram	GET	3h intervals, last 7 days, by feature type

LiteLLM endpoints (via LITELLM_URL + LITELLM_KEY)

Endpoint	Method	Params	Description
/customer/info	GET	user_id	Customer spend + budget + allowed models
/spend/logs	GET	start_date, end_date, user_id	Spend log entries
/key/info	GET	key	Virtual key details + spend
/model/info	GET	—	Available models
/health	GET	—	Proxy health

Response envelope

Most CodeMie endpoints return:

{
  "data": { ... },
  "metadata": {
    "timestamp": "2024-03-15T12:00:00Z",
    "data_as_of": "2024-03-15T12:00:00Z",
    "filters_applied": {},
    "execution_time_ms": 45.2
  }
}

Always extract response.data for the actual payload.

---

Custom analytics requests

For endpoints not covered by preset commands:

node analytics-cli.js custom /v1/analytics/mcp-servers --time-period last_30_days

# POST endpoints
node analytics-cli.js custom /v1/analytics/ai-adoption-overview --method POST \
  --time-period last_30_days

---

Offline CLI analytics (no API key needed)

The codemie analytics CLI command reads local session files from ~/.codemie/sessions/ with no API calls:

codemie analytics --last 7d --output json
codemie analytics --agent claude --last 30d --export csv

---

Report References

Reference files in references/ describe canonical report layouts. Always check the relevant reference before building a new HTML report — it defines the exact components, charts, data structure, and modal design to use, ensuring consistency across users.

Report type	Reference file	When to use
Leaderboard dashboard	${CLAUDE_PLUGIN_ROOT}/references/leaderboard-dashboard-report.md	Any request for leaderboard rankings, AI champions, top performers, tier distribution
People spending dashboard	${CLAUDE_PLUGIN_ROOT}/references/people-spending-dashboard-report.md	Any request to track LiteLLM costs for a specific list of users (cohort, team, bootcamp, project)

---

Use Cases

Use Case: People Spending Dashboard (cohort / team / bootcamp)

Trigger phrases: "build a spending dashboard for people from X", "track LiteLLM costs for a list of users", "how much did this team spend", "bootcamp spending report", "costs for people in this CSV/Excel".

Also applies when: the user asks to enrich analytics with EPAM employee data, map platform users to EPAM people, or look up user assignments/org details.

⚠️ EPAM People & Assignments Finder — required for this use case only

Before proceeding, verify the assistant is accessible:

node ${CLAUDE_PLUGIN_ROOT}/skills/codemie-analytics/scripts/analytics-cli.js \
  custom /v1/assistants/5ca384d0-d042-480c-a0a9-d28150e2352f 2>&1 | head -5

If the command returns an auth error, HTTP 401/403/404, or "No CodeMie credentials" — stop. Do not run anything else. Notify the user:

⛔ EPAM People & Assignments Finder assistant is not configured on your account.

Assistant: EPAM People & Assignments Finder ID: 5ca384d0-d042-480c-a0a9-d28150e2352f

Add it with:

codemie assistants add 5ca384d0-d042-480c-a0a9-d28150e2352f

Or open in browser and click Add to my assistants: https://codemie.lab.epam.com/#/assistants/5ca384d0-d042-480c-a0a9-d28150e2352f

Once added, resume this session with:

claude -f

Full workflow (see ${CLAUDE_PLUGIN_ROOT}/references/people-spending-dashboard-report.md for all details):

1. Parse the list from Excel/CSV using openpyxl. Skip header and TOTAL rows.
2. Fetch 3 LiteLLM accounts per user using Python asyncio + aiohttp (semaphore 25, ssl=False). Account patterns:
   • Web: email (plain)
   • CLI: email_codemie_cli
   • Premium: email_codemie_premium_models Use end_user_id param (not user_id) on GET /customer/info.
3. Save raw results to /tmp/ to avoid re-fetching on HTML rebuild.
4. Build users array — sum three spend values, extract budget fields per account.
5. Fetch leaderboard — paginate ALL pages (--per-page 500), ~25 pages for 12k users. Expect ~60% of a typical cohort to appear.
6. Fetch CLI insights — cli-insights-users --per-page 500 topBySpend for top CLI users. Expect ~3–5% coverage for a general cohort.
7. Compute KPIs — grand total, per-type totals, active user count, avg spend. Budget projection: avg_spend_per_active × total_users × 1.20.
8. Generate HTML — use str.replace() with __TOKEN__ markers (never f-strings, which conflict with JS ${...} template literals).
9. Wire table clicks — use data-email attribute + event delegation (never onclick="" attributes, which break under Python quote escaping).
10. Save to reports/<date>-<name>/<name>.html (temp/data files in reports/<date>-<name>/temp/).

Key commands:

CLI=${CLAUDE_PLUGIN_ROOT}/skills/codemie-analytics/scripts/analytics-cli.js

# Leaderboard (run in a loop for all pages)
node $CLI leaderboard --per-page 500 --page <N> --output json

# CLI top spenders
node $CLI cli-insights-users --time-period last_30_days --per-page 500 --output json

LiteLLM fetch requires Python (not the analytics CLI) because it needs LITELLM_URL + LITELLM_KEY env vars and concurrent calls for 1,000+ accounts.

---

Tips

• Always run the CLI first, capture JSON, then hand it to the report skill — don't hardcode example data.
• If a command returns paginated data, loop through all pages or set --per-page 500.
• For time-series charts, use /users-unique-daily or /projects-unique-daily endpoints.
• Budget warnings: flag rows where spend / max_budget > 0.8 (warn) and > 1.0 (error).
• For the leaderboard dashboard, combine leaderboard + leaderboard-summary + leaderboard-tiers + leaderboard-dimensions to build a comprehensive view. Then follow ${CLAUDE_PLUGIN_ROOT}/references/leaderboard-dashboard-report.md for the exact HTML structure.
• For a people spending dashboard, fetch LiteLLM directly with Python async (3 accounts per user), then enrich with leaderboard + CLI insights. Follow ${CLAUDE_PLUGIN_ROOT}/references/people-spending-dashboard-report.md for the exact HTML structure.
• For a single user deep-dive, combine leaderboard-user <email> with cli-insights-user <name> for the full picture (champion score + CLI activity).
• If the CLI prints an auth error, forward its message verbatim — it already tells the user what to do next.
• Always save HTML reports to reports/<date>-<name>/<name>.html; temp/data files go in reports/<date>-<name>/temp/.

Type-Aware Rendering for metrics[] Arrays

Analytics metrics[] arrays are heterogeneous — each item carries a type field ("number", "string", "date", …) and a format field. The value is numeric for most items but may be an ISO date string or plain string for others.

Always inspect m.type (and m.format) per item before formatting. Never apply a single numeric or percent formatter to the whole array — doing so silently produces NaN for string/date items.