<!-- 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>

Issue Securities

Walk an admin from raw input to issued certificates or option grants on a Carta cap table. Those two are the only security types this skill issues.

Flow	Example prompts
Certificates	"Issue 1 cert for Jane Doe, 1000 Series A at $1.50." · "Draft 5 founder certs on Acme."
Option grants	"Issue 1000 ISOs to Jane at $1.50 on the 2024 Plan." · "Draft 10 ISOs for new hires." · "Issue an EMI grant — 2000 options at £0.50."
Spreadsheet / file	"Issue the grants in ~/Downloads/Q3-hires.xlsx." · "Here's our import template — issue these certs." · "Draft the grant in this signed award agreement."
Resume	"Resume draft set 472." · "Continue the 'Q2 hires' draft set."

To fix an already-issued certificate or grant, that's carta-modify-issuables, not this skill.

Out of scope — stop and route to the Drafts UI in the Carta app for RSUs, SARs, CBUs, warrants, convertibles, SAFEs, convertible debt, and for custom legends, vesting, acceleration, or exercise periods:

"This skill issues certificates and option grants today. For <thing>, use the Drafts UI in the Carta app."

Carta's own import template has sheets for several of those, so an uploaded workbook routinely contains rows this skill can't issue. Phase 0.25 skips them and reports the count — never reshape an RSU row into an option grant to make it fit.

Architecture — one engine, two surfaces

The transaction is identical everywhere; only the surface that collects and reviews it varies.

• The engine is this file: resolve security_type → fetch reference data → assemble rows → save_drafts / validate_drafts / issue_securities → recovery. It never knows which surface is in play.
• The adapter implements exactly three capabilities. Nothing else may branch on environment.

Adapter	Selected when	collectConfig (0.5)	showReview (2)	confirm (2→3)
Cowork (primary, ~95% of usage)	preview_start absent	one show_widget form	chat markdown	one AskUserQuestion
Code	preview_start present	render-panel config panel	render-panel review panel	the panel's Confirm & Issue button

This file documents the Cowork path, since that is nearly all real usage. If Phase 0 Step 1 selects the Code adapter, read code-adapter.md — its §0 lists every point where that adapter diverges, and anything §0 does not mention behaves exactly as described here.

Read up front on every run:

• references/cowork-adapter.md — the form, the chat review, the confirm, and the authoritative per-block field list. Skip only if Step 1 selected Code.
• references/payload-reference.md — the authoritative field contract: types, formats, picklists, autofills, date quirks.

Both paths end the same way: the issue_securities mutate (Phase 3). The SDK's HITL prompt on that mutate is the final, irreversible gate — never the review gate.

---

Hard rules

1. The field contract lives in payload-reference.md. Read it before constructing any payload. No invented keys.
2. Never mix certificate and option grant fields in one mutate. Run the skill twice for a mixed request.
3. One confirmation gate per mutate attempt — never zero, never two stacked. On Cowork the gate is the AskUserQuestion in cowork-adapter.md §3; on Code it is the panel button, and stacking a question on an open panel suspends its submit watcher so the click never lands. Recovery questions after a server short-circuit are unrestricted on both. The SDK's HITL prompt is not this gate.
4. Retry contract — reuse identity from the FIRST response. draft_set_id from the first mutate goes on every subsequent issue_securities, save_drafts, load_drafts, validate_drafts, resolve_duplicate_stakeholder; omitting it makes the server auto-create a second draft set with the same incomplete rows. Each row's draft_pk from its first save goes on every retry row alongside every required field; omitting draft_pk inserts a new row instead of updating. A timeout is not an error — the call may have already succeeded server-side, so retrying with the wrong params risks a duplicate draft set or a double-issue. Read payload-reference.md § Timeouts & retries before retrying any mutate that timed out.
5. The server is the source of truth. Don't mirror its validation; surface its messages verbatim.
6. Never delegate to a background agent. The gates require interactive HITL.
7. Templates only — no custom payloads for legends, vesting, acceleration, or exercise periods: "Custom <thing> isn't supported here. Save as draft and finish in the Drafts UI."
8. No id sniffing. Required values come from user input, cap_table:get:stakeholders with detail=full, or a documented default — never scraped from another grant or certificate. If none of the three applies, ask via AskUserQuestion.
9. Pre-save assertion. Before any save_drafts or issue_securities call, walk every row and confirm each always field (per Row templates) holds a non-null value. If one is missing, recover before the call, in order: (a) the row template's documented default; (b) re-run the stakeholder lookup (detail=full) and re-stamp issue_date_relationship / email / stakeholder_kind; (c) AskUserQuestion. This is load-bearing because both failure modes are silent: save_drafts accepts incomplete rows without complaint, and at issue time a row with stakeholder_id=null slips duplicate detection, creates zero securities, and still returns success.
10. Never ask who the grantees are before opening the collection surface. A missing recipient is an empty field on the surface, never a chat question — this is the single most common way this skill goes wrong. Two sub-rules follow from it:
    • A bare "N <securities>" is a quantity, not a headcount. "100 option grants", "100 certificates" — the server's quantity field counts shares/options for one recipient. With no named people and no plural-person language, open one blank block with quantity pre-filled to N (knowns.rows = [{"quantity": "100"}]).
    • Only pre-render multiple blank blocks when the language counts people — "100 employees", "grants for 100 new hires", or an explicit list of names. Then build knowns.rows as that many empty dicts so the surface opens pre-sized.
    • Genuinely ambiguous (rare) → AskUserQuestion which one. That is a real fork, not the forbidden "who are the grantees" question.
    • "Which file did you mean?" is also a real fork, not this forbidden question — but only when the prompt referenced a file and Phase 0.25's search found zero or several candidates, or the workbook has more than one importable sheet. Never use it to ask who is in the file, and never in place of parsing a path the prompt already gave.

The incidents behind these rules — including the ones that look redundant — are in references/incidents.md. Read it before weakening any of them.

---

Voice & defaults

• Explain anything the skill chose. Tag (default), (from existing record), or (autofill — <so_type> rule) with a one-line explanation under the review. The review is the user's only chance to reject a default, so an unshown default is one they never got to see.
• Silent defaults are computable; prompted fields aren't. If the skill can stamp it (today's date, issue_date + 10 years, an autofill rule), stamp it and surface it tagged. Never ask twice.
• Show the full text of legally binding values (e.g. the legend body), not just the template name.
• Dates display as MM/DD/YYYY everywhere the user sees them. Payload formats follow payload-reference.md.
• Explain jargon on first use (Rule 144 date, Section 4(a)(2), INDIVIDUAL, legend).
• No raw ids or payload field names in customer-facing text — ever. Not in headers, status lines, prompts, confirmations, or error renderings. Never write the word "ID" (✅ "looking up Jane" / ❌ "pulling stakeholder id 12345"), and never render (<number>) after a name. Translate payload keys before surfacing them, including when echoing server banner_errors back: humanize mechanically (_ → space, Title Case), with the exceptions listed in references/labels.md.

---

Resolve security_type

Resolve once, at the top. Pass on every draft-set tool call.

Cue	security_type
"cert", "certificate", "shares", "Series A", "common"	certificate (default)
"option", "ISO", "NSO", "grant" (with plan), "EMI", "CSOP", "Unapproved", "Startup Concessions", "ESS", "Non-Concessional", "ZEPO"	option_grant
Ambiguous ("equity")	Ask with AskUserQuestion
Mixed in one prompt	Ask which to run first; run the other in a follow-up
Out-of-scope security	Route to the Drafts UI; stop

---

Phase 0 — Preflight

Four steps, in order, all before any user interaction and before gathering any input. Step 1 is free. Steps 2–4 are the only round trips this preflight may spend: one ToolSearch, one connectivity check, one list_accounts. Phase 0.5 then spends exactly one more — issuance_init, which carries the stakeholder lookup with it.

Step 1 — Detect the environment from the tool surface

Look at your own available tools. preview_start present → Code adapter; preview_start absent → Cowork adapter. That is the whole test: free, synchronous, and correct in both environments, because render-panel cannot work without preview_start.

Do not run a Bash probe, check the filesystem, or reason about what the environment "looks like" — the tool list is ground truth. Record the selection once and reuse it for Phases 0.5 and 2. Never re-detect per surface, never drop from Code to Cowork because a panel seems slow, and never attempt render-panel when preview_start is absent.

Step 2 — Load every tool in ONE ToolSearch call

ToolSearch: "select:mcp__carta__fetch,mcp__carta__mutate,mcp__carta__welcome,mcp__carta__list_accounts"

mcp__carta__ is the placeholder prefix (Step 2a) — when the session's Carta tools carry a different prefix, substitute it into the select: string; the names after the prefix never change. Zero matches on the literal mcp__carta__ names means the wrong prefix, not a disconnected server — re-check the session's tool list before treating it as the Step 3 stop.

One call, four tools, the complete set for the run. mutate is loaded here, up front, so Phase 2 never has to load it after the user confirms — that would be serial latency at the worst possible moment. On the Cowork path, add mcp__visualize__show_widget to the same select: list if it isn't already loaded.

Never call discover or search_tools in the hot path. Every command name is hardcoded below; looking up a name you already know is a pure round trip. discover is a debugging aid.

Step 2a — Carta command names (hardcoded, never discovered)

Reads go through fetch, writes through mutate. The argument key is params, not arguments, and command names are colon-separated:

mcp__carta__fetch({"command": "cap_table:get:<noun>",     "params": {…}})
mcp__carta__mutate({"command": "cap_table:mutate:<noun>", "params": {…}})

mcp__carta__ is a placeholder — here, in every code block below, and in every reference file. The real prefix is environment-dependent (mcp__carta-test__fetch, plugin-scoped and UUID-suffixed connector forms all occur). Resolve it from the session's tool list and substitute it everywhere; only the prefix varies — tool and command names never do. The one exception: the frontmatter allowed-tools entries are literal grant patterns — never substitute there.

Purpose	Command	Tool
Reference data for the collection surface, plus named stakeholders via stakeholder_names	cap_table:get:issuance_init	fetch
Stakeholder lookup for a roster miss — pass names= for several, search= for exactly one	cap_table:get:stakeholders	fetch
Load an existing set's rows	cap_table:get:load_drafts	fetch
List draft sets (resume by name)	cap_table:list:draft_sets	fetch
Cap-table totals for context math — authorized, outstanding, fully diluted, ownership %	cap_table:get:cap_table_by_share_class	fetch
Save rows, no validation	cap_table:mutate:save_drafts	mutate
Validate a saved set	cap_table:mutate:validate_drafts	mutate
Save + validate + dedupe + issue	cap_table:mutate:issue_securities	mutate
Resolve flagged duplicates	cap_table:mutate:resolve_duplicate_stakeholder	mutate

The totals source has a breakdown-sounding name. cap_table:get:cap_table_by_share_class — corporation_id alone — returns authorized, outstanding, fully diluted, and ownership %. Context math only (e.g. percent-of-fully-diluted for a grant), never a payload source. There is no cap_table:get:cap_table_summary — guessing it returns "Unknown command" — and the similar-sounding cap_table_summary_report is a different plugin's report command, not a name here; the row above is this skill's totals source.

Go through fetch/mutate, not call_tool. The runtime's tool descriptions deprecate fetch in favour of call_tool and discover in favour of search_tools. That notice is known and deliberately not followed — do not "fix" the contradiction. fetch stays because the per-command tools call_tool would target are excluded from tools/list (the mechanics below); the discover half is moot here because every command name is hardcoded in the table above, so discovery never runs (full story: incidents.md § Round-trips). Both fetch and mutate are pinned gateway tools: always present in tools/list, reachable in one hop, with scope and staff checks enforced inside the command executor. The double-underscore form (cap_table__mutate__issue_securities) is not a typo for a command name — carta-mcp also generates one hidden tool per command by swapping : for __ — but those are excluded from tools/list and reachable only via a search_tools → call_tool round trip, which is the cost this skill's hot path exists to avoid. When such a tool isn't visible to the session, that route returns "Unknown tool" instead.

Never call set_context for a corporation-scoped command. Every command above takes corporation_id as a direct param — pass it.

Step 3 — Confirm Carta MCP connectivity

The whole flow depends on the Carta MCP server. When it doesn't answer, classify the failure before reporting it — "not connected" and "Carta is briefly down" need opposite responses from the user, and telling someone to reconnect a connection that was fine is its own failure.

Signal	Meaning	Do
No Carta MCP tool in the tool list at all	Genuinely not connected	Stop with the message below
A call fails with HTTP 5xx / 502 / 503 / a gateway or HTML error body / a timeout	Transient upstream — the server is connected and briefly unhealthy	Retry once, then stop with the temporary problem message

Genuinely not connected — stop before gathering any input:

"I can't reach Carta — the Carta MCP server isn't connected. Connect the Carta MCP server and try again."

Transient upstream — retry the failed call exactly once. If the retry succeeds, continue the run normally and say nothing about it. If it fails again, stop:

"Carta is having a temporary problem on its end — the connection is fine. Give it a minute and try again."

The retry cap is one, and it is a hard cap. A second failure means waiting, not another attempt: re-running the same call against a 502 cannot succeed, and repeated attempts are the inner-loop thrash this skill's budgets exist to prevent. Do not vary the call to make a retry look novel, do not fall back to a different tool or a discover/search_tools probe, and do not treat an HTML error body as a data payload to parse — an HTML response to a JSON call is an outage signal, never content.

Step 4 — Resolve the corporation by name

If the prompt named a company and you don't already have its corporation_id, call list_accounts(search="<name>") — never an unfiltered list_accounts(), which returns a truncated alphabetical page that may never reach the name you want. search is the tool's own name lookup; don't substitute a discover guess for it. Only ask the user via AskUserQuestion if search returns zero or several ambiguous matches.

---

Phase 0.25 — Ingest an uploaded file

Skip this phase entirely unless the prompt references a file. When it does, the file replaces the prompt as the source of the rows — everything downstream is unchanged. It does not add a path around any gate: Phase 1 still resolves, Phase 1.5 still saves and validates, Phase 2 still reviews, Phase 3 is still the only mutate.

Supported: .xlsx .xlsm .csv .tsv (deterministic) and .pdf .docx (text extraction — see Documents). The sub-skill issuance-import/SKILL.md owns the mechanics; issuance-import/references/column-map.md is the header vocabulary. Never hand-read a workbook — a column read by eye is how a quantity lands in an exercise-price field.

Step 0 — Confirm you can actually run the parser

The parser is a local script, so this phase needs Bash(uv run *). Check your own tool surface for Bash before promising an import — it is present on the Code adapter and is not guaranteed on Cowork.

No Bash → do not hand-read the file. Reading a workbook by eye is the failure this whole phase exists to prevent, and offering it as a fallback would make the parser's guarantees optional. Say so and route to the feature built for this:

"I can't read spreadsheets in this session. Two options: import it directly in Carta's Drafts UI, which takes this same template — or paste the rows here as text and I'll set them up."

Pasted-as-text rows are fine: they arrive in the prompt, so the ordinary prompt-driven flow handles them with the user's own values in plain sight. That is different in kind from silently parsing a binary nobody can see.

Step 1 — Locate the file

Take the path straight from the prompt; a pasted ~/Downloads/… path is the norm. Only if the user said "the attached file" with no path, list the likely directories (ls -t ~/Downloads, ~/Desktop, the cwd) and look for a supported extension recently modified. Zero or several plausible matches → AskUserQuestion which one (allowed by Hard rule 10's file carve-out). Bash(find *) is deliberately not granted to this skill — use ls.

Step 2 — Parse, before spending any round trip

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-issuance/issuance-import/scripts/parse_upload.py" \
  --file "<path>" --out-dir "$OUT_DIR"

This first run is deliberately without --reference: it costs nothing, and its output tells you the two things Phase 0.5's fetches need — the security_type and the names in the file. It prints SECURITY_TYPE=, ROW_COUNT=, and the paths it wrote.

• Exit 2 with AMBIGUOUS: + CANDIDATES=[…] — the workbook has more than one importable sheet. AskUserQuestion which, then re-run with --sheet "<name>". Never merge two sheets into one batch (Hard rule 2) and never pick for the user.
• Exit 2 with ERROR: — nothing usable. Surface the message verbatim and fall back to the ordinary prompt-driven flow; do not guess at rows.

Reconcile security_type with the prompt. File and prompt disagreeing is a real fork → AskUserQuestion. The file wins only when the prompt never said.

Step 3 — Fetch reference data (Phase 0.5's fetches, informed by the file)

Run Phase 0.5's fetches exactly as documented, with two inputs now supplied by the file: issuance_init's security_type, and its stakeholder_names covering the names the file contains rather than the names the prompt named. That keeps the lookup bounded by the file's row count, never by roster size.

stakeholder_names takes the whole list at once, which is the only correct shape here — a 40-row sheet resolved through a concatenated search= matches nobody and would create 40 duplicate stakeholders on a real cap table. Pass the names as a list; never join them into a search string.

The account-setup gate still applies. Having a parsed file in hand is not a reason to push past it: a corp with no option-grant document set cannot issue one, whether the rows came from a spreadsheet or from the prompt. Stop where the gate says to stop — the parsed rows cost nothing and the file is still there afterwards.

Step 4 — Re-run the parser to resolve names to ids

uv run "…/parse_upload.py" --file "<path>" [--sheet "<name>"] \
  --reference "$OUT_DIR/_data.json" --out-dir "$OUT_DIR"

--reference is the same _data.json you just built for build_config.py. The parser matches the file's free text against it — vesting schedule, acceleration terms, share class (by name or prefix), legend (by code or name), document set, equity plan, and the roster — and writes _import_knowns.json.

Unresolved is blank, never guessed. A cell matching nothing leaves its field unset with an import_notes entry. There is no fuzzy matching, and do not add any: an almost-match on a vesting schedule or share class issues genuinely wrong terms, and unlike a bad quantity the server cannot catch it.

Step 5 — Merge into knowns and open the surface

_import_knowns.json holds {security_type, rows, equity_plan_id?, batch_errors?}. Its rows are your knowns.rows — merge them in and continue into Phase 0.5 unchanged. The row count comes from the file, so Hard rule 10's quantity-vs-headcount heuristic doesn't apply here (a 40-row sheet is unambiguously 40 blocks). Carry batch_errors through to the surface's panel-level banner, and hold equity_plan_id for the first mutate only.

Each row may carry import_notes — [{field, raw_value, reason}], display-only. Two obligations, both load-bearing:

1. Show every note on the surface, against the field it's about. On Code, build_config.py renders them as amber markers automatically. On Cowork, render them the same way in the show_widget form — see cowork-adapter.md §1.
2. Render a noted field with nothing selected, so the surface's own readiness check blocks submission until the admin picks. build_config.py does this for you. A marker alone is ignorable; the blocked button is what actually prevents a silent wrong issuance.

Strip import_notes before any mutate — same discipline as the review-only fields (Build the mutate payload). The server rejects unknown keys.

Step 6 — Say what happened, in one line, before the surface opens

Read _import_report.json and report totals — never silently drop a column or a row. A dropped Exercise Price column is a wrong-priced grant the user has no way to notice.

"Read 38 rows from Q3-grants.xlsx. 2 columns I couldn't map and 3 values I couldn't match are flagged in the form — everything else is filled in. Review and submit when ready."

Rows the file carried but this skill can't issue (RSUs, SARs, CBUs, warrants, RSAs, convertibles) are skipped by the parser with a reason. Name them and their count — they need the Drafts UI, and an admin who thinks a 40-row sheet issued 40 securities when it issued 37 has been misled.

Documents (.pdf / .docx)

The parser extracts text to _import_text.txt and stops — it writes no rows, because prose has no fixed layout and a script guessing at it would guess silently.

Read the text, build the rows yourself in the parser's own row schema (issuance-import/SKILL.md § Row schema), and give every field you filled this way an import_notes entry with "confidence": "low" so it renders as needs-confirmation. Then continue from Step 3.

Take only what the document states. A grant agreement rarely names a vesting template by the company's own template name, so leave vesting_template_id unset rather than inferring it from prose like "vests monthly over four years" — that is the fuzzy match this phase forbids, done by hand. Empty text means a scanned image: the parser exits 2 saying so; route the admin to OCR it or type the values in, and never infer values from a filename.

---

Phase 0.5 — Configure the issuance

Collect everything on one surface — every field, per stakeholder — so the user submits once instead of answering a chain of questions, and so a single batch can carry genuinely different terms for different people. This is the engine's collectConfig; the adapter from Step 1 decides what the surface is.

Fetch the reference data first, and issue every call below in ONE assistant turn. They have no dependencies on each other; serial fetches here are pure latency.

• Stakeholder lookup — pass the people the prompt named as stakeholder_names on the same issuance_init call below. The server resolves them alongside the reference data in one round trip, and the result comes back as that payload's stakeholders section with the same shape as the standalone cap_table:get:stakeholders command. There is no separate stakeholder fetch here. If the prompt named nobody, pass no names at all: there is nobody to resolve yet, and Phase 1 resolves whatever names the user types into the form.

  Never put two people in one search=. search AND-s its whitespace-separated terms, so it matches one person only — search="Jane Doe" works, search="Jane Doe Bob Smith" asks for a single human matching all four terms and returns an empty list with a perfectly healthy 200. Commas don't help; they're stripped before the terms are AND-ed. Use stakeholder_names (here) or names= (Phase 1) for several people — never a concatenated search.

• Reference data — cap_table:get:issuance_init with the active security_type, plus stakeholder_names when the prompt named people. One call returns every section the surface and Phase 1 need, each with the same {count, results} shape as its standalone command:

  • Option grant — vesting_templates, acceleration_templates, document_sets, valuations_409a, international_valuations, option_plans.
  • Certificate — certificate_share_classes, legends, vesting_templates, acceleration_templates (cert vesting is opt-in but needs the same two lists once opted in).
  • Both, only when stakeholder_names was passed — stakeholders, already at detail=full, carrying id, full_name, email, event_relationship, and kind per person.

  Every section is fetched server-side in parallel, so adding stakeholder_names costs no extra wall-clock — it removes a round trip rather than adding one.

  Partial failure is non-fatal. A section that failed comes back null and is named in the top-level errors array ([{section, message}]); fall back to that section's individual cap_table:get:<section> command. An empty errors means full success — use the payload directly. This is the only fallback path; the rest of this file just says "from the issuance_init payload".

  Read each section under its own name. Never let one section's count: 0 stand in for another's. Exactly one count may stop the flow — the Account-setup gate below, on document_sets.count read under that name and no other. Every other count, zero included, never gates: the surface is built and opened regardless (Hard rule 10).

Account-setup gate (option grant only)

Runs once, immediately after the issuance_init payload is read — before FMV, jurisdiction, plan, or any surface work — and is skipped entirely when security_type is certificate. Both adapters run it: a zero-template corporation cannot issue from either surface.

Read document_sets.count from the document_sets section, under that exact name. Confirm the section name before acting on the number: a real run aborted a valid issuance by reading acceleration_templates' zero as this section's (incidents.md § Reading server data wrong). Then branch:

• count >= 1 → pass; continue the phase. No other section's count matters here — acceleration_templates.count: 0, or any other empty list, is a normal state and never gates.

• count == 0 → stop before building any surface:

  "Your corporation doesn't have any option-grant document templates set up yet. Create one in the Carta app, then come back."

• document_sets failed to fetch — null (named in the top-level errors or not), absent outright, or present but not the documented {count, results} shape (missing or non-numeric count) → a failed fetch, not count: 0. Run the section's fallback, cap_table:get:document_sets with security_type: "option_grant", and gate on that count instead. If the fallback errors too, surface its message verbatim and stop as a fetch failure — never with the no-templates message above.

Why stopping here doesn't break Hard rule 10. Rule 10 forbids asking for collectible fields — anything the surface has a field for, like who the grantees are — before the surface opens. The surface's document-set field picks among existing templates; it cannot create one. Every grant row requires one (document_set_id is an always field), so with zero templates the field is unfillable from any surface and the batch it collects can never issue. A missing template is an account-setup blocker — the same category as an unreachable Carta MCP (Phase 0 Step 3) — and account setup happens in the Carta app, not on this surface.

The gate has exactly one member: document_sets, on option grants. It is not a "stop on any empty section" rule and must not be read as one.

Option grant: resolve the FMV and the jurisdiction (before building the surface)

Both are batch-level: every row in one draft set shares them. Resolve once, pass in knowns.

The FMV is not "the 409A". A company outside the US prices grants from an EMI, CSOP or share-price valuation and may have no 409A at all — reading only valuations_409a is what left those admins with an empty exercise price. Prefer international_valuations, which covers every source including 409A and carries the currency and status that valuations_409a cannot.

Read international_valuations.active (already filtered to live valuations server-side — do not re-derive it from dates) and set:

knowns key	Value
fmv_options	the active rows as-is (price, currency, valuation_type, effective_date)
fmv_source	the rows' shared support_reference_type (409A / EMI / CSOP / SHARE_PRICE)
fmv_expired_on	only when active is empty and history has one — its expiration_date, so the hint can say when cover lapsed instead of just "none"

build_config.py derives the hint and the prefill from these; don't pre-compute either.

Never pick between two active valuations. An HMRC report yields both an AMV (actual market value, discounted for restrictions) and a UMV (unrestricted market value). Nothing in the payload says which one a grant is priced from, and the difference changes the holder's tax position — so pass both and let the panel ask. The panel deliberately leaves the field empty in that case; do not "help" by filling one in.

If the section is missing — a US-only corp can be refused it (it is permissioned separately), in which case it comes back null in errors. Fall back to valuations_409a: use current_409a when its is_expired is false, and treat is_expired: true as the expired case.

Derive knowns.jurisdiction too — build_config.py defaults it to "US", which shows a UK company ISO/NSO buttons instead of EMI/CSOP. In precedence order:

1. the active valuation's currency (GBP → UK, AUD → AU, USD → US);
2. option_plans[].scheme_type == "EMI" → UK;
3. an EMI/CSOP support_reference_type in the valuations payload → UK;
4. otherwise US — and say so in the review as (default — assumed US), so a wrong guess is visible and correctable rather than silent.

Set knowns.currency from the same source. See code-adapter.md § knowns.

The surface's fields — one full key-value block per stakeholder, every field inside that person's own block, so a batch can issue genuinely different terms to different people. The authoritative enumeration for both adapters is cowork-adapter.md § Fields; it also covers batch mode (shared terms once + a compact name/email/quantity table) for large identical-term batches.

Render the form with show_widget and wait for its sendPrompt() reply — cowork-adapter.md §1. Never express this as a chain of AskUserQuestions: it is an option-picker that cannot take a free-text quantity, price, date, or name, so one prompt per field is the exact serial interrogation this phase exists to eliminate. Reserve AskUserQuestion for genuinely blocking single choices — an ambiguous security_type, a multi-plan pick, the Phase 2 confirm.

On submit, take the returned rows as your working set and map each row's own fields onto a resolved row: references/row-mapping.md. Then go to Phase 1.

---

Phase 1 — Resolve each row + reconcile share classes

Your working set is the rows from Phase 0.5 — one entry per grantee/holder, each already carrying its own quantity and full field set. Phase 1 resolves each row; it does not re-collect the person, the quantity, or any field the surface already carries.

Resolve from what Phase 0.5 already fetched. Match each rows[].name case-insensitively against the issuance_init payload's stakeholders section, which carries full_name, email, id, kind, and event_relationship per person:

• Exactly one match → reuse email, event_relationship, kind, id. Stamp stakeholder_id to bypass duplicate detection. Tag (from existing record). No MCP call.
• No match → the person is new, or was typed into the form after Phase 0.5 fetched (routine on Cowork). Batch only these misses into one cap_table:get:stakeholders call passing names= (a list of the missed names), not search=. A genuine no-match is a new stakeholder — never ask for an email that is already on the cap table.
• Multiple matches on one name → disambiguate with AskUserQuestion.

search matches one person; names matches many. search AND-s its whitespace-separated terms across full_name/email, so two people in one search string can never match anything — it returns an empty list with a 200, which looks exactly like "nobody here". A zero-result search that contained more than one name is a malformed query, not an absent person. Never conclude "these are all new stakeholders" from one; re-issue it as names=. Creating a duplicate stakeholder for someone already on the cap table is silent, wrong, and lands on a real cap table.

The number of stakeholder calls is bounded by roster misses, never by row count. A per-name search loop is the serial round-trip pattern this skill was slow for — and concatenating every row's name into one search to make that loop look batched is the same bug wearing a disguise, with the added defect that it silently matches nobody.

Precedence for the two fields the surface can also supply: a non-empty relationship stamps issue_date_relationship, and stakeholder_kind (defaulting to INDIVIDUAL) stamps itself — but only when the lookup found no match. For an existing stakeholder the cap-table record always wins; the surface auto-populates and locks these on an exact name match precisely so the two agree. Tag (from config surface — new stakeholder).

Dropping any of email, issue_date_relationship, stakeholder_kind, or stakeholder_id from a stamped row is a contract violation (Hard rule 9).

Push back on parsing only — a required field empty, a quantity that isn't a parseable number, a date that doesn't parse (ask for YYYY-MM-DD or MM/DD/YYYY), a broken email shape, or a required price that is 0/blank, except the two legitimate 0 cases: a ZEPO grant and a certificate/RSA on an LLC. Everything else is the server's call — don't pre-validate price-vs-FMV, decimals, future dates, negatives, state codes, exemption picklists, or prefix format.

Share-class reconciliation (certificate)

Matching a user-supplied class name, the "most recently created" default, and the ambiguous-match table: certificate-fields.md § Share-class reconciliation.

Option-plan reconciliation (option grant)

Use the option_plans section from the Phase 0.5 issuance_init payload.

• One non-expired plan → default silently. Tag (default — only active plan).
• Multiple non-expired → AskUserQuestion, one option per plan ("Use \"<name>\" (<available_quantity> available)"), last option "Cancel".
• Skip expired plans (is_expired: true); never recompute available_quantity.

Pass equity_plan_id only on the first mutate that creates the set — it is locked server-side after. Also stamp the chosen plan's name onto every row as plan_name, a review-only field never sent to the mutate.

---

Phase 1.5 — Save + validate before review (or save-only)

Reached immediately after Phase 1 resolves every row, for both of the surface's footer buttons. Saving and validating before the review exists is deliberate: validate_drafts runs nearly every check issue_securities does — all but the corp-level missing-signatory check — and it needs an existing draft_set_id, so validating early means saving early too. Reviewing an unvalidated summary means the user first learns of a rejection at the final confirm, after a draft row has already been created.

Full mechanics — branching on save_only / config_submit, translating server errors back into knowns, re-rendering the surface, and draft-state bookkeeping: references/save-validate-flow.md.

Resume an existing draft set

Loading a set by id or name, re-deriving option-grant review-only display fields, and jumping straight to Phase 2: references/resume-flow.md.

---

Shared resolution helpers

The collection surface already gathers legend, vesting, acceleration, and document set as fields inside each stakeholder's block, so on a normal run the row arrives already carrying the resolved id/label and these procedures don't run. They are the fallback for anything the surface didn't resolve. Picklist source, default posture, and what gets stamped are identical either way.

Board approval is the exception to "the surface already did this": it is a surface field, but the pointed-to section documents the underlying needs_board_approval logic both paths converge on. Dividend accrual start date is a further exception — it has no surface field at all yet and always goes through chat.

Vesting resolution

Use the vesting_templates section from the Phase 0.5 issuance_init payload. Render a compact picker (name + summary_short + vesting_type). AskUserQuestion: one option per template → vesting_template: <id>; "No vesting" → leave unset. When set, also collect vesting_start_date (MM/DD/YYYY, default issue_date) unless the template's vesting_type is milestone, which the server defaults — skip that prompt.

Flow	Default posture	"No vesting"
Certificate	Opt-in (only if the user volunteers)	Normal
Option grant	Required server-side	Accepted, but warn — atypical

Acceleration resolution

Only if vesting is set. Use the acceleration_templates section from the issuance_init payload. AskUserQuestion: one option per template → acceleration_template: <id>; "No acceleration" → leave unset.

Type-specific helpers

• Legend (certificate) — certificate-fields.md § Legend resolution.
• Dividend accrual start date (certificate) — certificate-fields.md.
• Rule 144 difference reason (certificate, only when rule_144_date ≠ issue_date) — certificate-fields.md.
• Exercise periods, document set, board approval (option grant) — option-grant-fields.md § Resolution helpers.

---

Row templates

Fill literally. Every slot must hold a value before review; a None or empty is a skill bug — reapply the default, re-call the stakeholder lookup, or ask (Hard rule 9). Each file also lists the review-only fields stamped alongside the payload, which are display-only and never sent to the mutate.

• Certificate — certificate-fields.md § Certificate row.
• Option grant — option-grant-fields.md § Option-grant row.

Pre-mutate checklist

Tick before any mutate — Phase 1.5's save_drafts / validate_drafts included, not just the final issue_securities:

• [ ] security_type resolved and passed on every call
• [ ] Stakeholder lookup ran with detail=full → issue_date_relationship, email, stakeholder_kind, stakeholder_id stamped on every row
• [ ] Cert: share class resolved → prefix. Grant: option plan resolved, so_type autofills applied (currency, exemption)
• [ ] Every always field populated per the row template; pre-save assertion passed (Hard rule 9)
• [ ] For issue_securities only: the review surface was opened and confirmed (Phase 2 → 3). Phase 1.5's calls precede the review by design, so this doesn't apply to them
• [ ] If retry: draft_set_id + each row's draft_pk in the payload (Hard rule 4)

---

Phase 2 — Render the review surface (mandatory pre-save gate)

The engine's showReview + confirm. Phase 1.5 has already saved and validated these rows, so this is a read-only confirmation before the irreversible issue_securities call, not another save.

Print the review as chat markdown, then confirm with one AskUserQuestion — cowork-adapter.md §2–3. The full always/conditional/optional column spec, the default-explanation text, the confirm prompt, and the compressed format for identical-term batches are in references/chat-review.md.

Tool pre-load: mutate must already be loaded (Phase 0 batched it) before you open this surface. Do not ToolSearch here — it adds serial latency after the user has confirmed. If for any reason it isn't loaded, load it now, before rendering the review.

---

Phase 3 — On confirmation, run the mutate

You reach Phase 3 when the review surface confirms. The confirmation already happened — do not re-ask, and don't second-guess a fresh signal as a stale replay. Branch directly:

Answer	Do
"Issue … now" / any free-text affirmative	Run the issue securities mutate
"Save as draft"	Save as draft
"Edit a row"	re-render the Phase 0.5 form, pre-filled with the resolved rows
"Cancel"	stop — Canceled closing

Build the mutate payload from your Phase-1-resolved rows

The review is read-only and has nothing to merge back. Run the serializer on your Phase-1-resolved rows and pass what it returns as drafts verbatim:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-issuance/scripts/serialize_drafts.py" \
  --security-type <option_grant|certificate> \
  --rows "$OUT_DIR/_review_rows.json" --out "$OUT_DIR/_drafts.json"

It enforces the three things that used to be prose here, each of which fails the whole mutate when got wrong:

• Per-field date formats. grant_expiration_date, vesting_start_date and rule_144_date are CharField(10) and take MM/DD/YYYY only — an ISO string comes back Date is invalid, with no server-side coercion. Every other date goes out ISO. Rows carry ISO everywhere upstream of this call, so do not hand-format dates. Conversion is idempotent, so a row already in MM/DD/YYYY is fine.
• Non-payload keys stripped: import_notes, row_key, and the review-only plan_name / document_set_label / exercise_periods_text / legend_body. Any of them present is an unknown-field rejection.
• Empty means omit, while a real 0 price, needs_board_approval: false, and an explicit vesting_template: null all survive.

Exit 2 with the row and field named means a date couldn't be read — fix it and re-run rather than sending it.

No Bash in this session? Then do all three by hand, and note that the format rule bites hardest on an imported batch: those rows arrive prefilled in ISO (the panel's date inputs accept nothing else), so all three CharFields need converting — Phases 0.5/1 did not touch them.

Re-run the pre-save assertion (Hard rule 9) on the resolved rows, then mutate.

---

Run the issue securities mutate

One mutate runs save → validate → check duplicates → issue, atomically.

# Certificate
mcp__carta__mutate({"command": "cap_table:mutate:issue_securities", "params": {
  "corporation_id": <corporation_id>, "security_type": "certificate",
  "drafts": [ ...cert rows... ],
  "draft_set_id": <draft_set_id on retry/resume>,   # omit on first call
  "draft_set_name": <optional label, ≤30 chars>}})

# Option grant — additionally pass equity_plan_id on the FIRST call only
mcp__carta__mutate({"command": "cap_table:mutate:issue_securities", "params": {
  "corporation_id": <corporation_id>, "security_type": "option_grant",
  "drafts": [ ...grant rows... ],
  "equity_plan_id": <equity_plan_id>,               # first call only
  "draft_set_id": <draft_set_id on retry/resume>}})

Response: {draft_set_id, drafts:[{temp_id, draft_pk, status}], validation:{status, success, errors, warnings}, duplicates:{has_duplicates, count, results}, issued:[{id}]|null}. Always capture draft_set_id and each draft_pk, even on success — a follow-up may retry this conversation.

Response state	Action
issued non-empty	Committed — render the table per On success
validation.errors	mutate-recovery.md § Error recovery
duplicates.has_duplicates	mutate-recovery.md § Duplicate resolution
Only validation.warnings (issued: null)	Surface verbatim; AskUserQuestion: "Acknowledge and issue" / "Edit a row"; re-call on acknowledge
Issue failed / 403	mutate-recovery.md § Atomic issue failure

Every re-call carries draft_set_id + each draft_pk (Hard rule 4).

On success

Render a short table using MM/DD/YYYY. Don't fabricate Label or Grant number columns — they aren't in the response.

• Certificate: Stakeholder · Share class · Quantity · Issue date.
• Option grant: Stakeholder · Plan · Option type · Quantity · Exercise price · Issue date.

Link to the ledger at https://app.carta.com/<VIEW_URL_PATH>, where VIEW_URL_PATH is options/list/<CORP_ID>/ (option grant) or certificates/list/<CORP_ID>/ (certificate). Never invent a different path — corporations/<corporation_id>/equity/options/ looks plausible and is not a real route. app.carta.com is a deliberate hardcode: no MCP command resolves an environment-specific host, so this link is only correct in production and will misdirect a sandbox/test session — a known, accepted tradeoff, not an oversight. There is no verified URL for a specific plan's detail page, so name the plan in plain text, not as a second link. Then close per Closing.

Save as draft (escape hatch)

Runs whenever a save-only save is needed: from Phase 1.5's Save button (the common case), or from the Phase 3 answer "Save as draft".

mcp__carta__mutate({"command": "cap_table:mutate:save_drafts", "params": {
  "corporation_id": <corporation_id>, "security_type": "<certificate|option_grant>",
  "drafts": [ ...rows... ],
  "draft_set_id": <draft_set_id if resuming>, "draft_set_name": <optional, ≤30 chars>,
  "equity_plan_id": <equity_plan_id>}})   # option-grant only, on first save

Response {draft_set_id, drafts:[{temp_id, draft_pk, status}]} — no top-level validation (save_drafts skips it by design), but each row's status still reflects its own save. Check it:

• All success → the Saved as draft closing.
• Some errored → surface verbatim per failing draft_pk, recover via mutate-recovery.md, then re-call save_drafts (not issue_securities) with the same draft_set_id + each draft_pk.
• All failed → surface verbatim; "No drafts saved. Fix the errors above and re-try." — no Drafts-UI link, since there is nothing there to open.

Never show the success message when any row failed — the user would believe a partial set is complete.

Validate without issuing

mcp__carta__mutate({"command": "cap_table:mutate:validate_drafts", "params": {
  "corporation_id": <corporation_id>, "security_type": "<certificate|option_grant>",
  "draft_set_id": <draft_set_id>}})

Returns {validation, duplicates} only. Interpret with the branching rules above; stop at the report.

Cleanup unexpected draft rows

Extra or duplicate rows surfaced by load_drafts on resume, or a dropped row_key from a Phase 1.5 retry: resume-flow.md § Cleanup unexpected draft rows.

---

Closing

At a terminal state, lead with a one-line domain summary — company, count, type, and for an issue, the issue date in long form (Month D, YYYY).

State	Template
Issued, certificate	"N <share class> certificates issued on <company> — <issue date>. Open <company>'s securities ledger in Carta to see the new certificates."
Issued, grant (uniform)	"N <so_type> option grants issued on <company> — <issue date>. Open <company>'s securities ledger in Carta to see the new grants, or find them under the <plan name> plan."
Issued, grant (mixed)	as above, but "N option grants (X ISOs, Y NSOs) issued on …"
Saved as draft	"N <security type> drafts saved on <company> — finish in the Drafts UI."
Canceled, a real draft set exists	"Issuance canceled on <company> — the draft set is still there if you want to come back to it: [Drafts UI](…same URL…)."
Canceled, no real draft set	"Issuance canceled on <company> — nothing was saved to Carta." (no link — there is nothing to open)

A draft set is real when this batch was resumed from an existing one, or at least one row from a save_drafts / issue_securities call this session came back with a success status and therefore a saved draft_pk. It is not real when draft_set_id is still the initial "new" placeholder, or when every row's status came back an error.

Every link above hardcodes app.carta.com — no MCP command resolves an environment-specific host, so this is only correct in production. A deliberate, accepted tradeoff: a sandbox/test session gets a link pointing at production. security_type in a Drafts-UI path is the literal mutate value (certificate or option_grant).

To correct an issued certificate or grant afterward, that's carta-modify-issuables.