Carta Portfolio Analytics Routing

carta-portfolio-analytics-routing — Portfolio Analytics Router (mirror)

Routes to Schedule of Investments, Co-Investor Lookup, or Performance Benchmarks. SOI and Benchmarks execute inline from a content mirror of their specialist skill; Co-Investor Lookup dispatches the specialist directly via the Skill tool. See execution modes in Architecture Notes below. The fourth Portfolio Analytics capability (Loan Dashboard) is internal-stage today and will be wired into this router as it reaches general availability for external users.

Scope note: This router covers the Portfolio Analytics skills available to external customers today. Loan Dashboard exists as an internal-stage skill but is not yet GA for external users — see Future routes below.

Route Map

Customer Intent Framework

Use this as the semantic layer when Step 1 signal phrases don't produce an exact match.

Step 1 — Parse Intent [Deterministic]

Respond immediately without extended reasoning.

STOP rows — handle before routing:

Route rows — classify and proceed to Step 2.5:

If exactly one route matches: proceed to Step 2.5. If zero or multiple routes match: proceed to Step 2.

Step 2 — Clarify Intent [Interactive]

Fire only if Step 1 returned no clear match.

If the user said anything at all: reason over it before asking. Present the menu only if you've genuinely worked through every signal and still can't determine intent.

If you need to ask, present once:

Which Portfolio Analytics capability can I help with?

1. Schedule of Investments (SOI) — fund holdings and portfolio breakdown
2. Co-investor lookup — who else invested in our portfolio companies
3. Performance benchmarks — fund IRR/TVPI vs peer cohorts
4. Something else — Describe what you need and I'll point you in the right direction

Out-of-scope

Step 2.5 — Fund Admin access preflight [Active proxy gate]

If the Carta MCP server is not connected (noMcp environment), skip this step and proceed to Step 3.

Otherwise:

1. Call welcome to establish session identity (skip if already called this session).
2. Run exactly one proxy probe: call_tool({"name": "fa__list__entities", "arguments": {}}). You get exactly one attempt at this probe — do not retry it, do not vary the call, and do not substitute a different tool to route around a failure.

Branch on the result:

If access is not confirmed, surface this message and stop:

This workflow requires Fund Admin access, which your firm doesn't appear to have enabled. Reach out to your Carta account manager or contact Carta Support to request access.

Important — this is a proxy, not a real entitlement check. See Architecture Notes §"Fund Admin access preflight" for the full rationale — there is no Carta MCP command today that exposes real SKU/entitlement data (SOI's own Step 3 discover call is for MCP UUID-form tool disambiguation only, not an entitlement gate), so this substitutes a cheap reachability probe. It cannot distinguish specific FA sub-products, so false negatives/positives are possible.

Step 3 — Route [Deterministic]

Mechanical rule — no exceptions: the very first content block of your very first turn after a route match, before any tool call (including list_contexts, welcome, or any firm-lookup call the matched route's own Step 1 would otherwise make first) and before any internal reasoning about what the user still needs to provide, must be plain text output containing exactly:

Routing to [Display Name].

Do not combine this turn with a tool call. Do not decide the announcement is unnecessary because you can already tell the matched route will need to ask a clarifying question (e.g. "which firm?"). That clarifying question — and every other step of the matched route's own workflow, including its own "Announce" step if it has one — happens in the next turn, after this one, never instead of it or merged into the same turn as a tool call.

Then hand the run off. Two routes read a reference file; one dispatches a skill. Use the exact action for the matched route:

For the two Read routes: follow the matched file's instructions exactly, starting from its own Step 1/Workflow entry point with the user's original message as context. Each reference file resolves its own internal script and data paths independently (see Architecture Notes) — do not rewrite them.

For co-investors: dispatch the skill and let it run its own workflow from its own entry point. Do not read a reference file for this route — there isn't one, and references/ holds no copy of the co-investor workflow. Do not re-implement any part of it here, and do not pre-collect its inputs (firm, fund, date range) on its behalf; it asks for what it needs. If the dispatch fails because the skill is unavailable, say so plainly and stop — do not fall back to writing the analysis yourself.

Do not summarize what the target skill will do beyond the announcement itself. Do not add any other output before the routing announcement.

If Something Goes Wrong

Architecture Notes

Orchestrator pattern

This skill is the sole external-facing entry point for the Portfolio Analytics theme. It classifies intent, resolves ambiguity via AskUserQuestion, and then either executes a route inline from references/<route>.md or dispatches the specialist skill. Loan Dashboard will be wired in as it reaches GA for external users.

Execution modes — mirror vs dispatch

Dispatch is the preferred mode. A mirror is a copy, so it can drift from its specialist, and keeping it honest costs either a declarable codemod (tests/carta-investors/mirror_sync.py) or a human merge. Dispatch has nothing to drift and nothing to re-mirror. It also keeps this router's allowed-tools minimal: a dispatched skill brings its own tools, so carta-co-investors' Bash(uv run *), Bash(tee *), and skill_checkpoint are deliberately not duplicated here.

The remaining two routes are mirrors only because their specialists are publish: false, so the publish pipeline strips them and there would be nothing for a dispatch to reach. Convert either one to dispatch by publishing it.

Scope decision (as of v2.1.0)

SOI, Co-Investor Lookup, and Performance Benchmarks are GA for external users. carta-co-investors carries publish: true — the dispatch route depends on it. Strip that skill from the published plugin and the route dead-ends for external users, since the dispatch has no skill to reach. tests/carta-investors/test_router_mirror_parity.py guards this. carta-soi and carta-performance-benchmarks stay publish: false; their mirrors are what reach external users. Loan Dashboard (carta-investors:carta-loan-dashboard) has a version field but no publish field — internal-stage today — and is documented below as a future route, not an active one.

Structural patterns

This router follows the same 6 patterns established in carta-valuations-routing and carta-compliance-routing (the reference routing skills), extended to three active mirrored routes instead of one:

Fund Admin access preflight — proxy, not a real SKU check

Step 2.5 was upgraded from a passive welcome-only check to an active proxy gate — same change, same rationale, as carta-compliance-routing and carta-lp-reporting-routing. Research into carta/carta-web's eshares/iam/models.py confirmed there is no single "FA" SKU — Fund Admin access is a family of ~23 ProductCode values fed by fund-admin's own products_fundproductaccess table via an internal gRPC call, cached in carta-web's EntityProductAccess model. That model is not exposed through the Carta MCP gateway — there is no search_tools-style command this router can call the way carta-valuations-routing calls search_tools({"query": "waterfall modeling"}). (SOI's own tool-discovery in Step 3 is unrelated — it disambiguates the MCP's UUID-form tool prefix, not an entitlement check.)

Absent a real entitlement command, Step 2.5 substitutes a cheap proxy: one fa__list__entities call. This only proves FUND_ADMIN.* data is reachable — it cannot tell which FA sub-products are actually enabled, so it can produce false negatives (transient failure on an entitled firm) or false positives (base FA present, but the specific feature a route needs isn't). The one-attempt cap follows the sub-agent safety pattern in the marketplace root CLAUDE.md — capped rather than left to retry loop.

Replace this the moment a real entitlement check exists — file a request for an MCP command surfacing EntityProductAccess for the current firm context; that's Carta MCP team work, not something fixable from the skill side.

Specialists are unchanged

references/soi.md and references/benchmarks.md are content mirrors of carta-investors:carta-soi and carta-investors:carta-performance-benchmarks's own SKILL.md bodies. carta-investors:carta-co-investors has no reference file at all — it is dispatched.

• soi.md is not re-mirrorable by copy-paste. When re-mirroring, rewrite both of its Step 4b paths — the find pattern and the ${CLAUDE_PLUGIN_ROOT} fallback — to this router's own references/soi/ copy. carta-soi's body points them at skills/carta-soi/, which the publish pipeline strips. render-artifact.py and artifact.html need no rewrite — the script probes both template offsets — so those two files stay byte-identical across both locations and re-mirror with cp. Keep them that way.
• carta-performance-benchmarks is entirely self-contained SQL/logic with no sub-references or scripts at all.
• carta-co-investors needs no re-mirroring at all — dispatch reaches the specialist itself, and it resolves its own scripts and canonical-investors.json exactly as it does when the picker invokes it directly.

None of the three specialists had its workflow modified. carta-investors:carta-performance-benchmarks had only its description softened (see "Picker de-tune" below) — same precedent carta-compliance-routing used for carta-form-adv (PR #11162) and carta-lp-reporting-routing used for carta-download-tearsheet (PR #11953). All three remain directly invocable by name, with their existing test coverage in tests/carta-investors/skill-triggers.test.yaml untouched.

Maintenance cost — read before editing carta-soi or carta-performance-benchmarks: a mirror is a copy, not a shared source, so any change to those two bodies must be re-applied to the matching references/<route>.md. tests/carta-investors/test_router_mirror_parity.py enforces this where the transform is declarable and otherwise requires the divergence be declared with a reason. carta-co-investors is exempt — dispatch has nothing to keep in sync, which is the reason to prefer it. When promoting a future route, reach for dispatch first and fall back to a mirror only if the specialist must stay publish: false.

Picker de-tune — carta-performance-benchmarks: its trigger-phrase-rich description was stripped down to a one-line redirect ("For performance benchmark requests, use carta-portfolio-analytics-routing instead"), following the same pattern documented under "How to promote a future route to active" § Mirror + de-tune below. It keeps publish: true and all its tools — this only changes picker precedence for ambiguous natural-language prompts and remains fully invocable directly by name. Added version: 1.0.0 — the skill had no prior version: field, matching the WARN default the frontmatter validator recommends rather than fabricating a bump from a nonexistent prior release.

Known limitation — carta-soi and carta-co-investors only: because these two specialists still have trigger-phrase-rich descriptions and remain standalone invocable skills, specific prompts ("show me the SOI") can still win the skill picker directly and bypass this router. This is a known, accepted limitation — and for co-investors it is now harmless, since the picker and the router both end up running the same skill. See the failing: true trigger tests in carta-portfolio-analytics-routing.test.yaml and carta-portfolio-analytics-routing-triggers.test.yaml for the two remaining un-de-tuned routes; the former benchmarks picker test is now a hard assertion, matching PR #11953's approach for tear sheets. Routing logic (STOP rows, disambiguation, inline execution) is fully verified by the inject-mode tests in carta-portfolio-analytics-routing.test.yaml, which force the router to run regardless of picker behavior.

References layout

carta-portfolio-analytics-routing/
├── SKILL.md                 ← this file (sole registered skill)
└── references/
    ├── soi.md                ← mirror of carta-soi/SKILL.md — ACTIVE
    ├── soi/                  ← copies of carta-soi's script + template (see below)
    │   ├── artifact.html
    │   └── scripts/render-artifact.py
    │                         (no co-investors file — that route is dispatched, see Step 3)
    ├── benchmarks.md         ← mirror of carta-performance-benchmarks/SKILL.md — ACTIVE
    ├── loan-dashboard.md     ← mirror of carta-loan-dashboard/SKILL.md — NOT YET WIRED (future route)
    └── loan-dashboard/
        └── artifact_template.html   ← copy of carta-loan-dashboard/references/artifact_template.html, used only by the Step 7e fallback path (see below)

soi/ holds real copies of carta-soi's script and template: the publish pipeline strips the whole directory of any publish: false skill, so a published route must not reference skills/carta-soi/. soi.md points at those copies; benchmarks.md needs none (self-contained SQL). The co-investors route needs no copies for the opposite reason — carta-co-investors is publish: true, so skills/carta-co-investors/ and its scripts/ survive publish and the dispatched skill runs them in place. That publish flag is load-bearing: flipping it back to false silently breaks this route in the published plugin, since the dispatch would have no skill to reach. loan-dashboard.md is different again: its primary render path ($SKILL_DIR/references/artifact_template.html, Step 7c) also resolves correctly with no rewrite, since $SKILL_DIR is probed by literal name (carta-loan-dashboard) same as the others — but its fallback path (Step 7e) calls read_skill(file_path="references/artifact_template.html"), a tool that resolves relative to whichever skill is currently executing. Mirrored verbatim, that call would resolve to this router's own references/, not carta-loan-dashboard's — so the fallback needed a real asset copy into references/loan-dashboard/ and a rewritten path. Check every source skill's tool calls individually; Read/absolute-path/$SKILL_DIR-probe calls behave differently from read_skill's implicit-relative resolution, and a skill can mix both (as this one does).

None of the routes below are dispatched from Step 3 and none have Step 1 route rows — promoting them is wiring work only, per "How to promote a future route to active" below, not a re-mirroring job.

Route dispatched

How to promote a future route to active

1. Confirm the underlying skill is GA for external users.
2. Decide dispatch mechanism. Try direct dispatch first — it is the only mode with no drift surface:
   • Direct dispatch (this router's pattern for co-investors) — Step 3 calls Skill('carta-investors:carta-<skill>') with no references/ copy at all. The specialist runs under its own frontmatter, so do not copy its tools into this router's allowed-tools; Skill alone is enough. Requires the backing skill be publish: true, or the publish pipeline strips it and the dispatch dead-ends for external users — add the route to DISPATCHED in tests/carta-investors/mirror_sync.py so a later publish: false fails CI instead of shipping. Still inherits picker-competition risk if the skill's own description is trigger-phrase-rich.
   • Mirror (this router's pattern for soi and benchmarks) — copy the backing skill's SKILL.md body verbatim into references/<route>.md. Check first whether the source skill has any Read ${CLAUDE_PLUGIN_ROOT}/skills/<source>/... calls or relative references/ links that assume they're being read from the source skill's own directory — if so, rewrite those paths into a references/<route>/ sub-directory (as carta-compliance-routing did for form-adv); if the source skill resolves its own paths dynamically or via literal skill-anchored strings (as all three routes here do), no rewrite is needed. Does not require touching the backing skill's frontmatter or publish flag, so its existing test coverage stays intact. Cost: the mirror can drift from the source if the backing skill changes later — there's no auto-sync, re-mirror by hand on every meaningful change.
   • Mirror + de-tune (carta-valuations-routing's pattern for waterfall) — mirror as above, and set the backing skill to publish: false plus soften its description. Eliminates picker competition, but changes the skill's public availability and can break existing tests that assert direct-picker behavior. Get explicit sign-off before doing this to any currently-public skill.
3. Add its signal phrases to the Step 1 Route rows table and the Customer Intent Framework.
4. Add it as an option in Step 2's AskUserQuestion (existing active routes keep their numbers; new route gets the next number; "Something else" moves to last).
5. Add a Read ${CLAUDE_PLUGIN_ROOT}/skills/carta-portfolio-analytics-routing/references/<route>.md row to the Step 3 table (mirror/mirror+de-tune) or the bare skill name (direct dispatch).
6. MINOR bump the version and add verdict tests — both inject-mode (routing logic) and plugin-mode (picker behavior, expect failing: true if the backing skill still wins the picker on its own).

Future routes — not yet active

Loan Dashboard

• Status: Internal stage — has a version field (1.0.0) but no publish field. Not GA for external users.
• Mirror status: Already built at references/loan-dashboard.md — one call rewritten (the Step 7e read_skill fallback, see References layout above) plus a real copy of artifact_template.html into references/loan-dashboard/. Every Bash pattern it needs (carta workspace cache, uv run, test -d/-f, find, mkdir -p) and Artifact are already in this router's allowed-tools (added for carta-soi's own artifact calls). Only read_skill is missing — add it at promotion time.
• Signals: "loan dashboard", "loan portfolio", "show my loans", "loan overview", "draw balance", "drawn balance", "undrawn", "credit facility", "outstanding loans", "loan commitments", "borrower portfolio", "loan KPIs", "loan exposure"
• CTA until GA: "The loan portfolio dashboard is coming to external users soon. Reach out to your Carta account manager or contact Carta Support to request early access."
• Skill when live: carta-investors:carta-loan-dashboard
• Dispatch mechanism to use: Mirror already built — promote via Read references/loan-dashboard.md once the tool grants above are added.
• Disambiguation to carry forward: "loan portfolio" / "loans I've made" / "draw balance" → Loan Dashboard (once live); "my LP portfolio" / "LP investments" → out of scope, suggest /carta-lp-reporting-routing (do not confuse loan portfolio with LP portfolio — these are different products, "loan" vs "LP").