<!-- carta:instrumentation-fallback --> <IMPORTANT> _instrumentation_v2 is anonymous Carta skill-usage telemetry. ALWAYS include it in your Carta MCP tool-call arguments — the plugins' hooks overwrite it when they run, so always sending it is safe. Shape: _instrumentation_v2: {"skills": ["<plugin>:<skill>", ...], "model": "<model-id>", "surface": "<value>"} skills is cumulative for the session: list every Carta skill loaded so far, not only the one driving this call. Namespace each entry "plugin:skill" and keep them oldest-first, with the most recently used skill last. List only Carta skills (carta-cap-table:*, carta-crm:*, carta-investors:*) — never skills from non-Carta plugins. surface is the Claude surface you are running in: "chat" (claude.ai or the Claude app, i.e. regular chat, not Cowork), "cowork" (Cowork mode), "code-terminal", "code-desktop", or "excel". Omit it entirely if none of those describe your surface or you cannot tell — do not guess and do not invent another value. </IMPORTANT>

Transform Configuration

Called from carta-reporting step 4d (Claude Code / MARKDOWN path). Use values resolved earlier in this session: data file path, corporation_id, user_report_pk, and _report_processor_path (the script path the parent skill resolved once — reuse it instead of re-running find).

Schema Preview

Run report_processor.py on the data file with no transforms to extract column names and types. Reuse the cached _report_processor_path; only fall back to find if it is empty:

UV_PYTHON_DOWNLOADS=never uv run "${_report_processor_path:-$(find ~ -name "report_processor.py" -path "*/carta-reporting/scripts/*" 2>/dev/null | head -1)}" <<'EOF'
{
  "local_file": "<preview or full report file path>"
}
EOF

Lead with a brief result summary (e.g. "Here's your Securities Ledger report — loading full data in the background."), then present the column inventory.

Column inventory format — numbered list, grouped by category. Example for a 28-column report:

**All 28 columns:**

Identity
  1 — Stakeholder Name (text)    2 — Grant ID (text)    3 — Equity Plan (text)

Grant details
  4 — Grant Date (date)    5 — Grant Type (text)    6 — Exercise Price (money)
  7 — Expiration Date (date)

Share counts
  8 — Shares Issued (integer)    9 — Shares Vested (integer)
 10 — Shares Unvested (integer)   11 — Shares Cancelled (integer)
...

Grouping heuristics (apply in order; unmatched → "Other"):

• Identity: names contain Stakeholder, Name, ID, Email, Type, Plan, Grant
• Dates: type = date
• Share counts: type = integer, name contains Share, Quantity, Count, Units
• Money: type = money
• Percentages: type = percentage

---

Customization Checkpoint

One AskUserQuestion covering columns, filters, sort, and format. Content adapts based on what the original prompt already specified.

When the prompt specified details:

Here's what I'll apply based on your request:

• Columns: Stakeholder Name, Grant Date, Shares Issued, Vested %
• Filter: Vested % > 50%
• Sort: Grant Date descending

Reply Continue with these settings to proceed, or adjust anything. You can also add: formula columns (% of total, running sum, ratio, delta), or aggregations (totals row, group-by).

When the prompt was vague:

What would you like to customize? Reply all columns, no filters to get the full dataset as-is, or specify any of:

• Columns — which to include
• Filters — e.g. "Vested % > 50", "Grant Date after 2024-01-01", "Name contains Smith"
• Sort — e.g. "Grant Date newest first", "Shares Issued descending"
• Formulas — % of total, running sum, ratio (A ÷ B), delta (row-over-row change)
• Aggregations — totals row or group-by rollup

Column name validation

Before building the config, check every column name and filter target the user mentioned against the actual column names from the schema preview. The user's terminology often differs from the report's column names (e.g. "Department" → "Cost Center", "Employee" → "Stakeholder Name", "Vest %" → "Vested %").

Matching rules (apply in order):

1. Case-insensitive exact match → use it silently.
2. One actual column name contains the user's term, or the user's term contains an actual column name (e.g. "vest percentage" ↔ "Vested %") → use it silently.
3. No confident match → flag it in the checkpoint question.

When a term can't be matched, surface it in the checkpoint with candidates:

⚠ "Department" wasn't found in this report. Closest text columns: Cost Center, Equity Plan, Grant Type. Which did you mean, or should I skip this filter?

Show at most 3 candidates, ranked by edit distance / word overlap. If the unmatched term was a filter target, do not add it to the config until the user resolves it.

Apply this check to: explicit column selections, filter column names, sort column names, and formula source columns.

Column selection for wide reports (> 10 columns)

Always pre-select key identifier columns (Stakeholder Name, Grant ID, or equivalent) plus any columns explicitly named in the prompt. State what's pre-selected in the checkpoint:

Columns pre-selected: 1 — Stakeholder Name, 2 — Grant ID, 6 — Exercise Price Reply all for all 28 columns, add by number (e.g. +8, 12, 13), enter a keyword to filter (e.g. vesting), or confirm the pre-selection with yes.

If the user replies with a keyword, call AskUserQuestion once more showing only matching columns and ask them to confirm or add more by number. This is the only case where a second checkpoint question is allowed.

---

Filtering Rules

All filtering is handled by report_processor.py — do not apply filters in Claude's memory.

Translate user requests to filter objects:

User request	Filter object
"vested % > 50"	{"column": "Vested %", "op": ">", "value": 0.5}
"grant date after 2024-01-01"	{"column": "Grant Date", "op": ">", "value": "2024-01-01"}
"name contains Smith"	{"column": "Stakeholder Name", "op": "contains", "value": "Smith"}
"ownership above 5%"	{"column": "Fully Diluted %", "op": ">", "value": 0.05}

Supported ops: > < >= <= = != contains

Pass percentage values as decimals when the column type is percentage (e.g. 50% → 0.5).

When a filter removes all rows from a sheet, the script returns 0 rows — skip that sheet and note it in the summary.

---

Column Selection

All column selection is handled by report_processor.py. Pass column names in the desired display order; the script preserves that order.

Always include key identifier columns (Stakeholder Name, Grant ID, or equivalent) even if the user didn't explicitly request them.

---

Sorting

Translate user requests to sort objects:

User request	Sort object
"newest first"	{"column": "Grant Date", "direction": "desc"}
"largest first"	{"column": "Shares Issued", "direction": "desc"}
"alphabetical"	{"column": "Stakeholder Name", "direction": "asc"}

Multi-key: pass multiple objects in priority order (first = primary sort key).

---

Formula Columns

Formula computation is handled by report_processor.py. Formulas are applied after column selection — they can only reference columns that are included in the output.

Scope boundary

Formulas are for mechanical transforms on values already present in the report data — arithmetic that any spreadsheet could do without knowing anything about equity rules. The four supported ops:

Op	Description	Required fields
pct_of_total	Each row as % of the column's grand total	column
running_sum	Cumulative total down the column (current sort order)	column
ratio	numerator ÷ denominator	numerator, denominator
delta	Row-over-row difference (current sort order)	column

If a user asks for a value that requires understanding cap structure, equity rights, or ownership math, direct them to the Carta report that already contains it:

"That calculation involves equity rules that I can't safely compute from the raw data. Carta's [Report Name] report already has it — want me to pull that one instead?"

Example:

[
  {"name": "% of Total Shares", "op": "pct_of_total", "column": "Shares Issued"},
  {"name": "Cumulative Shares",  "op": "running_sum",  "column": "Shares Issued"}
]

---

Aggregations

Aggregation is handled by report_processor.py.

Summary row — appends one totals row at the bottom; first column reads "Total":

{"type": "summary", "columns": {"Shares Issued": "sum", "Vested %": "avg"}}

Group-by — collapses rows by a key column, one row per unique value:

{"type": "group_by", "group_by": "Stakeholder Name",
 "columns": {"Shares Issued": "sum", "Grant Count": "count"}}

Supported ops: sum avg min max count

---

Script: report_processor.py

See Script Reference in carta-reporting for the full API (all fields, multi-sheet, merge, output format).

Phantom equity label: If _phantom_label_<corporation_id> was resolved in this session (Step 1a of carta-reporting), include "label_overrides": {"CBU": "<label>"} in every report_processor.py invocation for this corporation — both the schema preview call and the output preview/full-report call. Use the corporation-keyed variable (e.g. _phantom_label_12345) to correctly handle multi-corporation flows.

Always check stats after running:

• missing_columns non-empty → list available column names from data[sheet].columns and ask the user which they meant, then re-run with the corrected name
• skipped_formulas non-empty → tell the user which formulas couldn't run (usually the source column wasn't included in the selection)
• filtered_row_count = 0 → no rows matched; offer to relax the filter or change as_of_date
• displayed_row_count < filtered_row_count → preview is active; re-run without preview for full results

---

Output Preview

After the Customization Checkpoint, check if /tmp/carta_report_<user_report_pk>.json exists. If ready, use it as local_file. If not, poll every 5 s up to 5 more attempts.

Run the script with "preview": 5:

UV_PYTHON_DOWNLOADS=never uv run "${_report_processor_path:-$(find ~ -name "report_processor.py" -path "*/carta-reporting/scripts/*" 2>/dev/null | head -1)}" <<'EOF'
{
  "local_file": "<path>",
  "columns": [...],
  "filters": [...],
  "sort": [...],
  "preview": 5
}
EOF

Show first 5 rows of each processed sheet as a markdown table. Above each table write: N of M rows matched · K columns.

Ask:

Does this look right?

1. Generate full report — all rows as a markdown table
2. Excel — download as .xlsx
3. Describe any change to filters, columns, or sorting — I'll update just that field and re-run the preview without restarting the whole checkpoint.

If the user requests a change, update only the affected config field and re-run the preview — do not restart the customization checkpoint.

If the user chooses Excel, invoke Skill(carta-cap-table:carta-reporting-excel).

---

Presentation

• One table per sheet — label with the sheet name as a heading.
• Summary line above each table — e.g. "89 of 212 grants matched (Vested % > 50%)."
• Money columns — format as $1,234.56.
• Percentage columns — format as 12.34% (script stores as decimal; multiply × 100 for display).
• Date columns — format as MMM D, YYYY (e.g. May 3, 2026) — the Carta brand standard.
• If multiple sheets are returned and only one has data after filtering, hide the empty sheets.