<!-- carta:plugin-version --> <carta-plugin>carta-cap-table:6.96.0</carta-plugin>

<!-- [PATTERN carta-writing-style v0.0.2] [PATTERN etiquette v0.0.6] [PATTERN text v0.0.8] [PATTERN tables v0.0.12] [PATTERN carta-watermark v0.0.10] [PATTERN base v0.1.0] -->

CTC Dashboard (local React console)

Builds a corporation's CTC benchmark baseline, writes it to a local data dir in the dashboard console JSON schema, and launches the prebuilt React app via serve.py. The browser never calls the Carta MCP — this skill fetches the data; the server only serves JSON + the app source.

Read-only. The compensation:* MCP namespace exposes no write commands. This console displays CTC data and (in later tabs) models scenarios locally. Plan activation, payband edits, employee edits, offer letters, HRIS resync and every other mutation stay in the CTC product.

No demo data — real corporation required

Never fabricate, synthesize, or fall back to demo data, and never launch against an empty or partial data dir. Every dashboard runs against one real corporation's CTC data — either fetched fresh or served from a prior local fetch (cache). If the invocation names no corporation, do not auto-pick or guess — ask, then proceed once they answer.

Step -1 — Confirm this client can run the console (REQUIRED, before anything else)

This skill runs local scripts (uv run …) and serves a React app from 127.0.0.1. That works in Claude Code only. In Cowork, Claude Desktop, or claude.ai there is no local shell and no browser reachable at localhost, so every step below fails — and it fails at the first Bash call, after the user has already waited, with a permission error that does not explain why.

If Bash is not available in this session, stop before Step 0 and route instead:

"The CTC console runs a local web app, so it only works in Claude Code. For the same data here I can use carta-compensation-benchmarks (market rates for a role) or carta-compensation-scorecard (your roster and compa-ratios) — which would you like?"

Do not attempt a partial run, and do not fall back to summarising benchmark data in chat yourself: those two skills already do it properly, with the attribution and PII rules this one does not carry.

Launch order — cache-first, MCP-minimal

Building needs a full MCP sweep; launching a warm cache needs exactly one call. Resolve the name and check the local cache before touching any MCP, then spend a single get:plan call on the version gate (Step 0) to confirm the cached release is still the one the corp's plan points at. A matching version launches on local data alone.

Step 0 — Resolve identity + check the local cache

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/ctc_paths.py" resolve "<corp name>"

Read-only; prints slug, cache_root, raw_dir, dashboard_dir, snapshot_age_days, snapshot_benchmark_version_id, snapshot_benchmark_version, snapshot_skill_version and current_skill_version without creating anything. Use the printed dashboard_dir verbatim — never recompute a cache path.

• snapshot_age_days=none with suggested_match= lines → did-you-mean picker via AskUserQuestion.
• snapshot_age_days=none, no suggestions → build (Step 1).
• snapshot_age_days=<N> → cache hit. Now run the version gate below — do not launch on age alone.

The version gate (REQUIRED on every cache hit)

Age is not freshness, and there are two independent ways a fresh-looking cache can be stale: the benchmark release it holds, and the skill version that built it. Check both before launching. Neither is visible on screen — a stale dashboard renders perfectly, which is exactly what makes it worth a gate.

Age is not freshness. Carta publishes benchmark releases on its own cadence, and a corporation's plan can be re-pinned to a newer release at any time — so a cache written yesterday can already be several releases behind. A purely time-based gate serves superseded percentiles under a citation that names the old release, and the figures look authoritative because every other part of the dashboard is correct. That is a correctness bug, not a staleness inconvenience: someone sets salaries from these numbers.

So on a cache hit, make one cheap call before launching (Step 1's MCP identification applies):

read_tool({"name": "compensation__get__plan",
           "arguments": {"corporation_id": <corporation_pk>}})

Compare its benchmark_version.id against the printed snapshot_benchmark_version_id:

Condition	Action
ids match and age <30d	Launch (Step 4) silently.
ids match and age ≥30d	Offer "use cached" vs "re-fetch" — the release is current, only the fetch is old.
ids differ	Re-fetch (Step 2), regardless of age. Say why in one line: "Carta published a newer benchmark release (v21.0 → v25.5) — refetching." Do not offer "use cached" as the default here; the cached figures are superseded.
snapshot_benchmark_version_id=none	Treat as a mismatch and re-fetch. A cache whose version cannot be read must not be trusted.
The plan call fails (403/5xx)	Fall back to the age rule and say so: "Couldn't confirm the benchmark release is current, so this may be up to N days behind." Never silently pretend the check passed.

Skill version — the second staleness axis (no MCP call)

resolve prints snapshot_skill_version (what built the cache) and current_skill_version (what a rebuild would produce). Compare them:

Condition	Action
equal	Nothing to say. The cache was built by this skill; launch.
differ	Tell the user and offer a rebuild — do not rebuild unprompted, and do not launch silently. One line naming both versions and what it means, then AskUserQuestion: rebuild (recommended) vs open the cached build.
snapshot_skill_version=none	A cache built before the field existed. Same as "differ": say the cache predates version tracking and offer the rebuild.
current_skill_version=none	The running skill's version could not be read. Say so, skip this check, and fall through to the benchmark gate — never treat unknown as a match.

Why offer rather than force, when a benchmark mismatch re-fetches outright: superseded percentiles are wrong, while a cache from an older skill is correct but possibly incomplete. The figures still tie out; what may be missing is a capability — a tab, a peer-group dimension — that this version could add. Wrong data justifies spending the user's calls without asking; missing features do not, especially now that a default rebuild is a ~77-call sweep.

Say what changed in terms of the dashboard, not the version string. "Built with skill 0.1.0, now on 0.2.0" tells the user nothing they can act on. Name the consequence: "This dashboard was built before peer-group switching existed, so it has no dimension or bucket pickers. Rebuilding adds them — about 77 calls, a few minutes." If you cannot say what a rebuild would add, say that plainly rather than inventing a benefit.

Version equality is the signal — do not second-guess it. This skill's version: is bumped deliberately when a change warrants rebuilding a cache; an unchanged version means the maintainers judged the change cache-compatible. Never compare timestamps, file contents, or the plugin version as a substitute — plugin.json moves whenever any sibling skill in the plugin changes, which has nothing to do with this dashboard's data.

Bump version: in this skill's frontmatter when a change alters what a build produces — a new tab, another data dimension, a different data-dir shape. That bump is the whole mechanism by which existing users are told their dashboard can be improved; skip it and they keep opening a cache that silently lacks the new capability. Leave it alone for wording, docs and fixes that produce a byte-identical build.

This intentionally costs one MCP call on a warm launch. The previous zero-call path was faster but could not distinguish "4 days old and current" from "4 days old and 12 releases behind" — and it is the only thing standing between a user and year-old compensation figures. One get:plan call is the cheapest possible check: it is the same call Step 2b already makes, it returns in one round trip, and on a match nothing else is fetched.

Use the plan's version, not the newest published one. compensation:list:benchmark_versions may show a release newer than the plan's (releases land before plans are re-pinned). Always follow get:plan — that is what the CTC product UI shows, so pinning to a newer release would make this dashboard disagree with the product.

A numeric corp id instead of a name → ctc_paths.py find-by-id <id>. No corporation in the invocation → ctc_paths.py list-dashboards to offer resuming a prior one.

Run this silently. The user's first line should be the greeting, not narration of the steps. Cache age in words ("3 days old"), never as a raw field=value.

Step 0.5 — Preflight: Claude CLI (required to build)

A build fetches every benchmark page and the employee list through a headless claude session (capture_export.py, see the capture contract in Step 2), and the ask box, the planner's cohort filter and the Benchmarks location fetch (2d-v) spawn one too. All of them need the claude CLI installed and logged in — the Claude Code binary, from Anthropic. Detect it early so a missing CLI is resolved before the user waits.

A build needs it; a warm-cache launch does not. On a cache hit the dashboard is fully functional without the CLI — only the three console features need it, so never fail a cache launch over a missing CLI. A build without it cannot capture data, so it stops.

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/preflight_claude.py" check

The search covers PATH, the usual install locations, and the binary of the Claude Code session running this skill (CLAUDE_CODE_EXECPATH). That last one means an IDE or desktop user with nothing on PATH is still found, so claude_bin=none should be rare.

• claude_bin=<path> → capture the path for Step 4. No output to user, whatever claude_auth= says. A user who is logged out sees the console's "not supported" message in place of the two features; the dashboard itself is unaffected.
• claude_bin=none → ask via AskUserQuestion before installing anything:

  "Fetching your data and the chat in the dashboard need the Claude Code CLI (claude), which isn't installed here. Want me to install it? (runs Anthropic's installer)"

  • Yes → run preflight_claude.py install. It uses Anthropic's native installer and falls back to npm. Then use the printed path. If it still prints claude_bin=none, say the install didn't succeed and treat it as No.
  • No → on a cache hit, continue: the dashboard works and the Claude features show the console's "not supported" message. On a build, stop and say the build needs the CLI to fetch data — any cached dashboard for this corp still opens.

When Step 4 launches serve.py, pass the discovered path as --claude-bin "<path>" (only when a non-none path was found). serve.py runs the same search itself when it is omitted, and finds the binary again if the given path disappears (a Claude Code auto-update deletes old versions), so the flag is a hint rather than a requirement.

What the user sees when it fails. The browser gets one message for every failure — CLI missing, logged out, crashed mid-turn — and never a stuck "Working…" or a raw error. The specific cause is written to chat-errors.log in the dashboard dir. Read that file when a user reports the chat isn't working, rather than guessing.

Reused daemon caveat: if serve.py detects an already-running daemon for this corp, it reuses it and exits. That daemon re-checks the CLI whenever a Claude feature is used while marked unavailable, so a CLI installed or logged in since then is picked up without a relaunch.

Step 1 — BUILD: identify the Carta MCP + resolve the corporation

Reached only on a build/refresh (cache miss, stale re-fetch, or an explicit "refresh"). A warm-cache launch needs only the MCP identification below plus the single get:plan call for Step 0's version gate — not the rest of this step's resolution work.

Identify the Carta MCP server. Scan the tools available in the conversation for any matching mcp__*__welcome. Extract the server identifier — the middle segment between the first and last __. Examples: mcp__carta__welcome → carta; mcp__carta-prod__welcome → carta-prod.

• None found: stop and tell the user, do not fabricate data:

  "No Carta MCP is connected. Building needs one — connect a Carta MCP (the carta-cap-table plugin provides it). Your cached dashboards still open without it."

• Exactly one: call mcp__<SERVER>__welcome to verify. This is <SERVER>.
• Multiple: ask via AskUserQuestion. Default to carta (production) if present.

Don't call any other mcp__<SERVER>__* tool before welcome — every other command is gated.

Classify the environment from <SERVER>'s name. A name containing test/sandbox/demo/ preprod/preproduction (case-insensitive) → cartaEnvironment = "nonprod". Everything else — carta, carta-prod, any other name, or an opaque UUID — → "production". This is a customer-facing plugin, so an unrecognized identifier is far more likely to be a production connector we haven't named than a staff test session. Carry this into Step 2e's meta.json.

Resolve the corporation. Priority order — stop at the first that succeeds:

1. A numeric id in the prompt ("corp 7") → use that integer verbatim. No lookup.
2. A name → list_accounts(search="<name>"), filtered to entries whose id starts with corporation_pk:. One match → use it. Several → AskUserQuestion with the names copied verbatim from the response.
3. No hint → list_accounts() bare. Exactly one corporation → use it silently (don't ask someone to confirm something they have no choice about).
4. Zero corporation_pk: entries → the caller may have no cap table at all. Check with fetch({command: "context_tools:get:profile", params: {}}) — it is a command, not a flat tool, so read_tool returns "Unknown tool". If its corporations[] is empty, say CTC needs a cap table and stop. Do not ask them to name a corporation they don't have.

HARD RULE — only ever act on a name and corporation_pk returned verbatim by list_accounts for this query. Never invent or auto-correct a company name, never blend two returned names, never reuse one remembered from earlier in the conversation. If the search returns nothing, say so and ask for the exact name or numeric id — a benchmark for the wrong corp is worse than asking again.

Key the cache on the canonical corporation name (the one the API returned, not the one typed):

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/ctc_paths.py" resolve "<canonical name>"

Use its printed raw_dir / dashboard_dir as the build target. Because every build keys on the canonical name, any typed variant or a re-fetch lands on the same directory — a rebuild refreshes that cache in place and a corp can never spawn a duplicate.

Then tell the user: "✅ Resolved <canonical name>. Starting data fetch…" — the only checkpoint between here and the launch.

Step 2 — Fetch → raw files

What's happening: pulling the corporation's plan and market benchmarks from Carta. Results land as raw JSON in the local cache — nothing is sent back to Carta.

Follow references/queries.md for the exact arguments.

The capture contract — read this before the first call

Never re-type, summarise, or hand-transcribe an MCP result into a file — and that includes reproducing it with the Write tool, a heredoc or a Python literal. Writing a result back out means generating every character of it again. A benchmark page is ~15k tokens of figures, and a single wrong digit silently corrupts a salary that then looks authoritative in the dashboard.

Why the harness can't save it for you. Claude Code only saves a tool result to a file when it is very large (the ~65KB location list in 2d-v is). An export page is ~17-25k characters, so it arrives inline in this conversation, and from there the only way onto disk is to write it back out — which is the retyping this contract forbids.

So every benchmark page and the employee list are fetched by capture_export.py, never by calling read_tool yourself. It runs each call in a headless claude -p session on the same Carta MCP server, checks the arguments the session actually sent, and takes the raw tool result off that session's event stream straight to disk. The payload never enters this conversation. It is the same path the console's on-demand location fetch uses (location_fetch.py).

sweep — one whole compensation:export:benchmarks sweep (2c, 2c-bis, 2d-v):

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/capture_export.py" sweep \
  --server "<SERVER>" --claude-bin "<claude_bin>" --dest "<dir>" \
  --args '{"corporation_id": <id>, "benchmark_version_id": <id>, "equity_quantity": "FOUR_YEAR_GRANT", "<dimension>_bucket": "<CODE>"}'

It pages from job_offset: 0 with job_limit: 6 until next_job_offset is null, retries a failed page once, stops at 6 calls, and fans each page out into benchmark_<JOB>.json plus export_pages.json exactly as save_benchmark_result.py --export-page does. It also checks every page came back on the pinned benchmark_version_id. Leave job_limit and job_offset out of --args; it owns paging and refuses them. Its last line is sweep COMPLETE — 22 job areas across 4 page(s), or it exits 1 with FAILED — <reason> after removing that sweep's files, so a failed sweep never leaves a half-captured group behind.

call — one read-only compensation__get__* / compensation__export__* result written to a file, which you then pass to the matching save_* script as its source path:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/capture_export.py" call \
  --server "<SERVER>" --claude-bin "<claude_bin>" \
  --tool compensation__get__plan --args '{"corporation_id": <id>}' --out "<raw_dir>/plan.raw"
uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_benchmark_result.py" "<raw_dir>/plan.raw" "<raw_dir>/plan.json"

<SERVER> is the identifier from Step 1; <claude_bin> is the path from Step 0.5. Exit 2 means the arguments were wrong — fix them, never retry unchanged. Exit 1 means the fetch failed.

FAILED — server_unavailable: <SERVER> is not connected in a headless session; reachable: <a>, <b> means the headless session names the same Carta MCP differently. A claude.ai connector and a locally configured server with the same URL are deduplicated, and the copy kept can differ between this session and a headless one (verified: "claude.ai Carta (Test)" here, carta-test headless, both mcp.test.carta.rocks). Ask via AskUserQuestion which of the reachable servers is the same Carta environment as <SERVER> — never pick one yourself, since a different environment means a different company's data. Use the answer as --server for every capture in this build, and as mcpServer in meta.json (2e). If none is, stop the build.

Two narrow cases need no capture_export call:

• The harness already saved the result to a file — the tool result says "Output has been saved to …" or carries <persisted-output>, as the location list in 2d-v does. Pass that path straight to the save script. Don't reconstruct the path; it is client-dependent.
• A result of a few fields you only read — subscription_status, or the version gate's get:plan on a cache hit. Read the values; nothing is written.

If capture_export.py cannot run — no claude CLI, or its first call fails for a reason that is not a Carta error — stop the build and say so. Do not fall back to calling read_tool and writing the result out yourself: a mistyped page is worse than no dashboard.

If you find yourself about to write benchmark numbers or employee rows into a file by any means, stop — you are in the failure mode this contract exists to prevent.

2a. Subscription gate (REQUIRED — do this before anything else).

read_tool({"name": "compensation__get__subscription_status",
           "arguments": {"corporation_id": <corporation_pk>}})

It returns {corporation_id, has_ctc_access, is_subscribed}.

• has_ctc_access: true, is_subscribed: true → continue.
• has_ctc_access: true, is_subscribed: false → stop, surface the subscription message, fetch nothing.
• has_ctc_access: false → the caller has no CTC role on this corp (is_subscribed is null). Stop; do not retry.

A corp with no subscription has no benchmark data, so every later call would be wasted.

2b. Plan. Capture it with capture_export.py call, then normalize:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/capture_export.py" call \
  --server "<SERVER>" --claude-bin "<claude_bin>" \
  --tool compensation__get__plan --args '{"corporation_id": <corporation_pk>}' \
  --out "<raw_dir>/plan.raw"
uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_benchmark_result.py" "<raw_dir>/plan.raw" "<raw_dir>/plan.json"

This is also the check that capture works here at all. If it exits 1 on anything other than a Carta error, stop the build per the capture contract.

Capture from it: benchmark_version.id (pins 2c) and peer_group.{code,label,dimension}. If dimension is missing or is not one of post_money / capital_raised / headcount, stop — guessing produces numbers that don't match the product UI.

2c. Benchmarks — paged bulk export, not one call per job area.

There are 22 job areas (listed in references/queries.md §3), but the fetch is now compensation:export:benchmarks — a bulk columnar export that returns several job areas' full matrices (every level, both the IC and LEADER ladders) in one response, replacing the old one-call-per-job-area sweep.

Always pass job_limit: 6, starting at job_offset: 0. A full sweep is then 4 pages (22 areas ÷ 6). Do not omit job_limit hoping for a bigger page:

The binding limit is a 10-second SERVER TIMEOUT, not a row cap. Omitting job_limit asks for every remaining area at once, which on a real corp takes ~17s and fails with exceeded 10000ms time limit — burning a call and returning nothing. The response cap is ~300 columnar rows (~12 areas at ~17 rows each), so the old "omit it, you'll get up to 12" advice is arithmetically true and operationally wrong: the request times out long before it hits the row cap. Verified against a live MCP — the timeout error itself recommends job_limit: 6, and 6 completes comfortably (observed 55 and 102 rows per page).

capture_export.py sweep sends equity_quantity: "FOUR_YEAR_GRANT", benchmark_version_id and exactly one <dimension>_bucket param from --args, and adds job_offset and job_limit: 6 itself.

PAGING IS THE DEFAULT, not an opt-in. One response is one page of at most 6 job areas, never the whole matrix, and there is no way to get all 22 areas in one call. The sweep keeps calling with job_offset = the previous response's next_job_offset until that field comes back null, so it cannot mistake 6 of 22 job areas for everything.

Capture the corp's own group into the raw dir itself:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/capture_export.py" sweep \
  --server "<SERVER>" --claude-bin "<claude_bin>" --dest "<raw_dir>" \
  --args '{"corporation_id": <id>, "benchmark_version_id": <id>, "equity_quantity": "FOUR_YEAR_GRANT", "<dimension>_bucket": "<plan peer_group.code>"}'

It writes one benchmark_<JOB>.json per job area (matching the per-job-area files the old sweep produced, so build_datadir.py is unchanged) and the paging metadata in <raw_dir>/export_pages.json. Read its last line — sweep COMPLETE is the authoritative answer to "is this group captured".

Never echo benchmark figures into the conversation — they belong in the data dir. The script prints job-area and row counts only; report progress the same way.

Then check coverage:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/build_datadir.py" \
  --raw "<raw_dir>" --check

Same four lines as before (captured / empty / CORRUPT / plan.json), plus a fifth when the raw dir was fetched via the export: EXPORT SWEEP: <warning> if the last page's next_job_offset was still non-null. build_datadir.py refuses to build in that state even if every benchmark_<JOB>.json file happens to exist, so a build can never publish a sweep that was cut short mid-page.

When the sweep fails (FAILED — <reason>, exit 1): it has already retried the failing page once and removed the partial files.

1. Run the sweep once more if the reason looks transient (timeout, 5xx).
2. Never fall back to per-job-area compensation:get:benchmark calls, or to calling the export yourself, to route around a failing sweep. If the export is failing, the fix is its arguments (wrong benchmark_version_id, bad bucket param, auth), not a different endpoint or a different capture route.
3. If it fails twice, stop before building and surface the reason — a page that won't succeed after retries means something systemic. A timeout on a job_limit: 6 page is NOT a reason to ask for a bigger page; if 6 areas time out, a larger page certainly will.

A hard ceiling regardless of outcome: at most 2 sweeps for the corp's OWN group (each stops itself at 6 calls). Alternates in 2c-bis carry their own ceiling. If you're about to exceed either, stop and report — you're in a retry loop.

2c-bis. The other buckets in the corp's dimension — this is what turns on the peer-group dropdowns.

The Benchmarks tab has a dimension picker and a bucket picker, and it hides both unless the data dir contains more than one peer group. They read alternatePeerGroups, which build_datadir.py assembles from peer_<CODE>/ subdirectories of the raw dir. Fetch nothing here and the controls never appear — not broken, just nothing to switch between.

So: after the corp's own sweep is COMPLETE, sweep every other bucket in the same dimension. Same benchmark_version_id, same equity_quantity: "FOUR_YEAR_GRANT", same job_limit: 6 and the same page-until-next_job_offset-is-null discipline — the only thing that changes is the <dimension>_bucket value. The codes, in order, per dimension:

Dimension	Bucket codes (low → high)
post_money	ONE_MILLION, TEN_MILLION, TWENTY_FIVE_MILLION, FIFTY_MILLION, ONE_HUNDRED_MILLION, TWO_HUNDRED_FIFTY_MILLION, FIVE_HUNDRED_MILLION, ONE_BILLION
headcount	ONE_TO_TWENTY_FIVE, TWENTY_FIVE_TO_HUNDRED, HUNDRED_TO_FIVE_HUNDRED, GREATER_THAN_FIVE_HUNDRED
capital_raised	ONE_TO_TEN_MILLION, TEN_TO_TWENTY_FIVE_MILLION, TWENTY_FIVE_TO_FIFTY_MILLION, FIFTY_TO_ONE_HUNDRED_MILLION, ONE_HUNDRED_TO_TWO_HUNDRED_MILLION, GREATER_THAN_TWO_HUNDRED_MILLION

Sweep all three dimensions, not just the plan's. The app has a dimension picker as well as a bucket picker, and it offers whatever dimensions the data dir contains — so fetching only the plan's dimension leaves that control with one option, which is the same dead-end the bucket picker had. Post-money valuation, headcount and capital raised are all real ways a customer asks "who are we being compared against", and the builder already labels each group with its own dimension and citation phrase.

The bucket param name changes with the dimension — exactly one per call:

Dimension	Param
post_money	post_money_bucket
headcount	headcount_bucket
capital_raised	capital_raised_bucket

The plan's own dimension still leads. The builder puts it first in the picker and the corp's own bucket stays the default, because that is the peer set the plan actually chose — switching bucket asks "what if we were valued higher?", while switching dimension changes what a peer is. Both are legitimate; only one is the default.

A bucket label alone is ambiguous across dimensions. $1M-$10M exists in both post-money and capital raised, and on a real corp they return different figures (167,000 vs 144,000 for the same role). That is why each group carries its own dimension and citation, and why the CSV filename includes the dimension. Never describe a group by its bucket label alone.

Capture each bucket into its OWN subdirectory — point --dest at peer_<CODE>/ instead of the raw dir, with that dimension's bucket param in --args:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/capture_export.py" sweep \
  --server "<SERVER>" --claude-bin "<claude_bin>" --dest "<raw_dir>/peer_<CODE>" \
  --args '{"corporation_id": <id>, "benchmark_version_id": <id>, "equity_quantity": "FOUR_YEAR_GRANT", "<dimension>_bucket": "<CODE>"}'

Use the bucket code exactly as spelled above: build_datadir.py reads the code back out of the directory name to look up that group's label and dimension, so peer_one_million or peer_ONE-MILLION lands as an unlabelled group in the switcher.

Skip the corp's own bucket — it is already the flat benchmark_*.json set at the top of the raw dir, and fetching it twice would put the same group in the dropdown twice.

Sweep ALL dimensions and ALL buckets by default

Unless the user has asked for something narrower, fetch every bucket in all three dimensions. Do not ask first, and do not quietly pick a cheaper scope on their behalf — a partial sweep is the one outcome that looks like a broken app rather than a saved call. With one dimension the dimension picker renders a single option and hides; with one bucket both pickers hide, and the user sees a dashboard with no peer-group controls at all and no indication that anything was skipped.

Narrow only on an explicit instruction ("just post-money", "skip the alternates", "keep it quick"). Then say which scope you used and which pickers it leaves empty, so the absence is a stated choice rather than a silent gap.

Cost, and why it is no longer a reason to ask. There are 18 buckets across the three dimensions (8 post-money, 4 headcount, 6 capital raised), each a full 4-page sweep: 17 alternates × 4 = 68 extra calls, against the ~9 for the corp's own group.

Sequentially that is minutes of wall clock, which is what the old "ask first" default was protecting against. Run the sweeps in parallel instead (2c-ter) and those 68 calls cost a few buckets' wall clock, because buckets are independent. Parallelism, not a smaller scope, is the answer to the cost.

Scope	Calls	What the pickers show
All three dimensions (DEFAULT)	~68	Both pickers fully populated
Own dimension, all buckets	~12–28	Bucket picker works; dimension picker hides
Adjacent buckets only	~8	Three buckets in one dimension
Own bucket only	0 extra	Neither picker appears

The per-dimension figures differ — post-money has 7 alternates (~28 calls), capital raised 5 (~20), headcount 3 (~12) — so quote the number for the dimension actually in play rather than 28 for all three.

A verified reference point: on corp 7 the full 18-bucket sweep completed with every page returning 200, so 68 is a measured figure rather than a guess.

Ceiling for this step: one sweep per bucket (the script stops itself at 6 export calls: 4 pages plus two retries), and stop the whole step after 3 consecutive buckets fail. A systemic failure — wrong benchmark_version_id, auth, a bad bucket param — will fail on every bucket, and grinding through 17 of them wastes ~100 calls to learn what the second one already told you.

A failed alternate is not a failed build. Alternates are additive: the corp's own group is already captured and the dashboard is complete without them. If a bucket's sweep won't finish, delete its peer_<CODE>/ directory (the script has already emptied it) and move on — build_datadir.py already drops an alternate with no rows and warns rather than publishing a group that renders as a blank grid, but removing the directory keeps the warning list honest about what you actually attempted. Report which buckets are in the switcher and which you dropped.

2c-ter. Run the alternates in parallel.

Issue the alternates' capture_export.py sweep commands as parallel Bash calls — several in one message, about 6 at a time, until all 17 have run. Buckets are independent: each sweep writes only its own peer_<CODE>/ directory, and each runs its own headless sessions.

Do not delegate the sweeps to subagents. capture_export.py keeps payloads out of every conversation, so a subagent adds cost and a place for a brief to drift without making capture any safer.

Do the version gate ONCE, before starting any sweep. Resolve the corporation, run Step 2a/2b, and pass the same pinned benchmark_version_id to every sweep. The script refuses a page from any other release, so a release published mid-sweep fails that bucket rather than leaving buckets that silently disagree across the picker.

Verify centrally when the sweeps return. Read each command's last line, then run build_datadir.py --raw "<raw_dir>" --check once: every bucket in the switcher must have a peer_<CODE>/export_pages.json with sweep_complete: true. Drop any bucket whose sweep printed FAILED and say which ones are missing.

2d. Employee list — one bulk export (feeds the Scorecard tab).

Fetch this on every build, not on request. It is what makes the Scorecard tab exist, and because a warm cache launches without re-fetching data (Step 0 spends one call on the version gate only), a build that skips it produces a dashboard whose Scorecard tab is missing for the next 30 days with no way for the user to ask for it short of an explicit "refresh".

Call compensation:export:scorecard with just corporation_id. It returns every benchmarked employee in ONE columnar response — no page, no page_size, nothing to page through. Capture it with the export flag:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/capture_export.py" call \
  --server "<SERVER>" --claude-bin "<claude_bin>" \
  --tool compensation__export__scorecard --args '{"corporation_id": <id>}' \
  --out "<raw_dir>/roster_export.raw"
uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_roster_page.py" \
  --export-scorecard "<raw_dir>/roster_export.raw" "<raw_dir>"

The capture contract applies in full: the employee list is names and salaries, so it is never written out by hand. Pass <raw_dir> itself as the destination, not a single .json. The script writes <raw_dir>/roster_pages.json in the same schema the paged path produces, so build_datadir.py does not care which route was used.

Read the script's last line — it is the authoritative answer to "am I done." It prints sweep COMPLETE — N of M employees captured. The export returns everyone or refuses, so an INCOMPLETE here is not "fetch another page" — it means the response was filtered or capped, and you should stop rather than build on it.

Never echo employee names, salaries or compa-ratios into the conversation — they belong in the data dir, same rule as benchmark figures. Report only "N employees captured".

If the export returns 400 "at most 200 fit in one export response": the corporation is too large for a single response. Fall back to the paged sweep below — it is slower and its pages overlap, but it has no size ceiling. Do not narrow with job_filters to squeeze under the cap: a filtered employee list silently under-reports how many people sit below market, which is the failure this whole tab exists to surface.

Three things that will bite otherwise:

• The generated tool name is compensation__export__scorecard — colons become __. Unlike employee-scorecard, there is no hyphen in this one.
• It is staff-gated. A non-staff caller gets a permission error, not an empty list — fall back to the paged sweep rather than reporting the corporation has no employees.
• 403 means no CTC role on this corp (not "no employees"): stop, don't retry. A genuinely empty employee list returns total_results: 0.

<details> <summary><b>Fallback — the paged sweep</b> (only when the export is unavailable or refuses)</summary>

Fetch compensation:get:employee-scorecard with page_size: 10, then page with page: 2, 3, …. Capture every page through capture_export.py call, then save it (no --export-scorecard flag):

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/capture_export.py" call \
  --server "<SERVER>" --claude-bin "<claude_bin>" \
  --tool compensation__get__employee-scorecard \
  --args '{"corporation_id": <id>, "page": <n>, "page_size": 10}' \
  --out "<raw_dir>/roster_page_<n>.raw"
uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_roster_page.py" \
  "<raw_dir>/roster_page_<n>.raw" "<raw_dir>"

Why 10 and not 30. A page_size of 30 is REJECTED — the response exceeds the 80,000-char transport limit. Verified against a live MCP on a 134-employee employee list. 10 is the documented safe value in that error message.

Budget — the overlap makes this bigger than it looks. Pages return overlapping rows, so a page of 10 yields fewer than 10 NEW employees (observed ~30-40% repeats, i.e. ~6-7 new per page). Budget ceil(total_results / 6) + 2 calls. The authoritative signal is the script's N of M line rising each page — stop only when a page adds zero new employees twice in a row.

• The generated tool name keeps its hyphen — compensation__get__employee-scorecard.
• page_size is snake_case. pageSize is not rejected — it is ignored, and you get the default 10 instead, i.e. a short employee list that looks real.
• Never pass score=LOW/MID/HIGH to shrink a page. It filters on the overall band, which is nullable, so it silently drops every employee without one and undercounts the list.

</details>

If the roster sweep fails outright, still build — build_datadir.py omits roster.json, snapshot.json records hasRoster: false, and the Scorecard tab simply does not appear. Losing one tab is better than losing the Benchmarks dashboard too. Say which tab is missing and why.

2d-v. Locations — the Benchmarks tab's Location dropdown.

The CTC product's Benchmarks page offers every supported location and re-fetches when one is picked; the server applies that location's scalar, the plan threshold, bands and rounding. The console does the same: it lists every supported location, and fetches a picked one on demand (scripts/location_fetch.py) through the user's own Carta MCP, with the same location value the product sends. Every figure on screen is still the server's.

First, save the location list. One call, with the id from the plan's benchmark_version.geo_adjustment_version.id:

read_tool({"name": "compensation__get__benchmark_locations",
           "arguments": {"geo_adjustment_version_id": <geo_adjustment_version.id>}})

The result is ~65KB, so the harness saves it to a file. Pass the printed path straight through:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_benchmark_result.py" \
  <printed_result_path> "<raw_dir>/location_catalog.json"

If it arrives inline instead, fetch it with capture_export.py call (--tool compensation__get__benchmark_locations) and pass the --out file — never write it out yourself.

If the plan has no geo_adjustment_version, or the command is unknown (an older Carta MCP), skip this: the console then offers only the locations fetched below.

Then fetch where the employees are, ahead of time, so the common picks open at once. Run this after the roster (2d), which is where that list comes from:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/ctc_locations.py" plan --raw "<raw_dir>"

It prints default=<label> (the location the main sweep already covers — San Francisco), then one line per location still to fetch:

location=Tulsa,OK,USA	dir=<raw_dir>/geo_tulsa-ok	employees=37	label=Tulsa, OK

and to_fetch=N. It also writes each geo_<slug>/location.json. Re-running it after a partial sweep lists only what is left.

For each line, run a sweep exactly like 2c — same benchmark_version_id, equity_quantity: "FOUR_YEAR_GRANT" and the plan's own <dimension>_bucket — adding "location": "<location>" verbatim from the line, into that line's dir:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/capture_export.py" sweep \
  --server "<SERVER>" --claude-bin "<claude_bin>" --dest "<dir>" \
  --args '{"corporation_id": <id>, "benchmark_version_id": <id>, "equity_quantity": "FOUR_YEAR_GRANT", "<dimension>_bucket": "<plan peer_group.code>", "location": "<location>"}'

The sweep keeps that dir's location.json. The ceiling is the same as for peer groups (one sweep per location), and the sweeps run in parallel the same way (2c-ter).

• Everything else is fetched when picked. With the list saved and mcpServer in meta.json (2e), the console fetches any other location, any alternate peer group's figures for a location, and an international market's local-currency figures on demand, in about 15-40 seconds. That needs the claude CLI (Step 0.5); without it the picker says the fetch isn't supported in this setup.
• These ahead-of-time sweeps are USD, for the plan's own peer group. An international market opens in local currency, as in the product, which is a fetch on first pick.
• A location that won't complete is dropped, not padded. Delete that geo_<slug>/ directory's pages (keep location.json) and say which location is missing; the builder also omits any location whose sweep is incomplete or from another release.
• Never add a location that was not fetched, and never derive one from another's figures and a scalar — see "Changing the app" below.

2d-bis. Equity refresh report — the Refresh planner tab.

Run this when the caller asked for the planner. Otherwise skip it: the tab is still hidden by default (app/src/App.jsx → SHOW_REFRESH_PLANNER = false), so for an ordinary build the capture is wasted work.

A staff caller who asks to QA the planner gets both halves — this capture AND the preview key in meta.json. See the staff-preview note under Step 2e. Doing only the key produces a visible tab with no planner.json behind it, which reads as a broken feature rather than a hidden one.

Call compensation:export:equity-refresh-report — generated tool name compensation__export__equity-refresh-report — with corporation_id. It takes no other arguments: no paging, no filters, one response carrying every benchmarked employee. Then:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_equity_refresh_page.py" \
  "<result path>" "<raw_dir>"

Three things that will bite:

• It is staff-gated. A non-staff caller gets a permission error, not an empty list. Say so plainly and build without the planner rather than reporting that the corporation has no employees.
• 400 "at most 190 fit in one export response" means the corporation is too large. There is no paged fallback for this report — say which corporation and stop, rather than narrowing the list, which would silently drop people from a refresh cycle.
• The figures are the report's own. They tie out against CTC's Equity Refresh Report; pass them through unchanged and never re-derive one.

build_datadir treats the report as optional — with no capture it records hasPlanner: false and the tab does not appear.

When the workflow ships, flip SHOW_REFRESH_PLANNER to true and delete the staff-preview note under Step 2e. Both must move together, or the docs will promise a tab the app does not show, or vice versa.

2d-ii. The equity pool (optional). The planner's review step measures a plan's draw against the corporation's available pool — the same figures the CTC product UI's Refresh Grant Planner "Equity pool impact" tile renders. Read them from the carta-web endpoint the product UI itself uses (equity_pool_utilization), so the microapp ties out against the product to the share.

Fetch via the MCP:

read_tool({"name": "cap_table__get__equity_pool_utilization",
           "arguments": {"corporation_id": <corporation_pk>}})

Then always hand the result to the capture script — the script owns the null-vs-zero decision:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_equity_pool_utilization.py" \
  "<result path>" "<raw_dir>"

Skip the whole step only when the MCP call itself failed (no result path to hand it). The review step then states that no pool figure is in the build instead of showing a guardrail — a live cache miss and a corporation whose ledger reports no pools sum to exactly 0, and neither is "this company has no shares left", so neither is ever rendered as a zero balance.

Do NOT source the pool from compensation:get:reports-insights.efab.*. That reads a different (compensation-service Redis) snapshot warmed by an out-of-band job — it can drift from the product UI. The reports-insights command is still useful for the two num_employees counts it also returns, but not for the pool.

2d-iii. Corporation equity units (optional). The planner can show grants as shares, fully diluted ownership or an equity value, matching CTC's own Equity Unit dropdown. The last two need a fully diluted share count and a per-share value that the refresh report does not carry.

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_corporation_info.py" \
  "<result path>" "<raw_dir>"

The source is the corporation info endpoint. The price is not a choice — it is resolved server-side from the corporation's own settings, so picking between the preferred price and the 409A value here would disagree with every other CTC surface.

Skip it otherwise. A unit whose input is missing is dropped from the toggle rather than shown as $0 or 0.0000%, exactly as the product does.

2d-iv. Refresh grant policy (planner-only). The Settings step of the Refresh Grant Planner applies the corporation's refresh grant policy — target %, cadence, tenure gate and the recommended corridor. Without this capture the step falls back to a "policy could not be fetched" banner and disables its controls.

Run this whenever 2d-bis ran (i.e. the caller asked for the planner). Skip otherwise.

read_tool({"name": "compensation__get__retention-plan",
           "arguments": {"corporation_id": <corporation_pk>}})

Then hand the result to the capture script:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_retention_plan.py" \
  "<result path>" "<raw_dir>"

• The endpoint is never empty. On a corporation that has never opened CTC's Plan Settings, it lazily provisions Carta's initial defaults (24-month tenure, 30% every 12 months, 15%–50% corridor) and returns them — the same behaviour the CTC product's Plan Settings modal has. Treat the returned record as the corporation's real starting policy, not a fallback the console invented. If the response IS empty, that is a backend degradation, not a normal state; the script exits so you notice, and this build should stop rather than model against a policy nobody set.
• 403 — no CTC role on this corp: stop this capture and build without it. Same rule as 2d-bis; the planner's Settings step then shows its "policy could not be fetched" banner instead of modelling against fabricated numbers.
• The percent fields are FRACTIONS on the wire (0.30 == 30%). The console does the ×100 conversion once, in model/policy.js; do not rescale in flight.

2e. Write meta.json (next to raw_dir, per ctc_paths.py):

{"corporation": "<canonical name>", "corporationId": <int>, "cartaEnvironment": "production",
 "mcpServer": "<SERVER>"}

mcpServer is the --server value the build's capture_export.py calls succeeded with — the identifier from Step 1 (carta, claude_ai_Carta, …) copied verbatim, unless a server_unavailable answer replaced it. The console's fetches are headless sessions too, so they need the headless name. The console's on-demand location fetch calls mcp__<mcpServer>__call_tool, and nothing else; without it the Location picker offers only what the build fetched.

<details> <summary><b>Staff preview — showing the hidden Refresh planner tab</b></summary>

The planner tab is off for everyone (SHOW_REFRESH_PLANNER = false). A staff caller who asks for it can have it in their own build by adding one key:

{"corporation": "…", "corporationId": 7, "cartaEnvironment": "production",
 "preview": {"refreshPlanner": true}}

BOTH conditions, and you check them in this order:

1. The user asked. Not "might find it useful" — an actual request to see the planner, or to QA it. Never add this on your own initiative; a half-built workflow appearing unbidden is the thing the tab is hidden to prevent.
2. They are staff. Call get_current_user and read its staff field. Do not infer it from an @carta.com address, from the corporation they are looking at, or from their having asked — an email domain is not an authorisation, and the whole point is that the caller does not self-certify.

If they asked and are not staff, say so plainly and build without the key. Do not offer a workaround: the flag lives in a file on their own machine and they could edit it, which is exactly why this is a courtesy gate and not a security one.

What it is not. It withholds no data — the planner reads the same equity refresh report already in the directory, and the console has no identity of its own to check at render time. It exists so an unfinished workflow is not shown to customers, and so a teammate QAing it does not have to patch source. </details>

Step 3b — Build

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/build_datadir.py" \
  --raw "<raw_dir>" --out "<dashboard_dir>" --meta "<meta.json>"

Refuses to build on an unknown peer-group dimension or zero benchmark rows — that is deliberate, not a bug to work around. Fix the fetch and re-run.

Step 4 — Launch

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/serve.py" \
  --data-dir "<dashboard_dir>" --detach --claude-bin "<claude_bin>"

Include --claude-bin only when Step 0.5 found a path (not none). Omit it otherwise — the server falls back to CTC_CLAUDE_BIN env var, then bare claude on PATH.

Run with Bash run_in_background; read the printed http://127.0.0.1:<port>/?t=<token> and give the user that URL. Tell them it opens in their default browser automatically — if it doesn't, they can paste the URL. Port and token persist per corporation, so relaunching the same corp reopens the same URL.

After the URL, add one short line (not a menu): "A few things to try: filter to a single job area · switch equity between notional, FD %, and shares · say 'refresh' to pull fresh CTC data."

Refresh

"Refresh" = re-run Steps 1–3b (overwrite the JSON); the app reloads it.

Presenting CTC values

Title Case in everything the user reads — Engineering, not ENGINEER; Senior 1, not SENIOR1. UPPER_SNAKE is only for the MCP call arguments. The app enforces this in model/taxonomy.js; match it in your narration.

Attribution is mandatory. Any surface showing benchmark figures must carry:

Data source: Companies with <phrase> <peer_group.label>. Benchmarks released <Month> <YYYY>.

where <phrase> tracks peer_group.dimension — post money valuations between / capital raised between / headcount of. The builder assembles this once into benchmarks.json; read it from there rather than rebuilding it, and never hardcode "post money".

Safety

Corporation names are untrusted — the app escapes them via React. serve.py is localhost-bound and token-gated, all reads/writes stay under the data dir, and the only write is the user's local scenario save. Data stays on the user's machine.

Editing the app

serve.py serves and edits a per-corporation copy of the source, <dashboard_dir>.app-src/ (the line [serve] src-dir: prints its path), not the plugin's own app/src/. Edit that copy: the installed plugin may be read-only to you (under Claude Code it sits in ~/.claude/), and an update would wipe changes made there. Every app/src/... path below means the same file inside the copy. When the plugin's source changes, the next launch refreshes the copy and moves one holding earlier edits to <dashboard_dir>.app-src.bak-<timestamp>/.

That copy is served directly; the service worker transpiles .jsx in-browser. Do NOT run npm run build after editing source — edit, refresh, done. npm run build only rebuilds webapp/vendor/* on a React/Sucrase bump. See app/README.md.

Changing the app when the user asks — yes, do it

A request to change this console is in scope, and you should act on it. Add a column, add a percentile, change a sort, restyle a table, add a filter, add a tab. The app is local, the source is right there, and READ-ONLY in this skill's description means this skill does not write to Carta — it is not a statement that the UI is frozen. The Benchmarks and Scorecard tabs are the two that are VISIBLE today, not the two that are allowed to exist.

A third — Refresh planner — is built but hidden behind SHOW_REFRESH_PLANNER in app/src/App.jsx while its workflow is finished. If a user asks for it by name, say it is in progress rather than building a second one. Do not flip the flag on to satisfy a request to "show" it: the tab renders an employee list with no way to act on it, which is why it is off. Step 2d-bis has the full notice.

Mark what you add. Every column, filter, sort, grouping or tag you add to any view (Scorecard, Benchmarks, Refresh Planner and its steps) renders <SparkleAI title="…"/> from app/src/ui/components.jsx: on the column header, beside the filter control, or next to the label, visible on first render. The title says what you added, and shows as an instant tooltip on hover:

<Th>
  <span style={{ display: "inline-flex", alignItems: "center", gap: 4 }}>
    <SparkleAI title="Added by Claude - months to final vest" />
    Months to final vest
  </span>
</Th>

A reader cannot otherwise tell a column you added from one Carta ships, so a missing icon passes your work off as the product's. Restyling existing UI (colour, spacing, wording) gets no icon.

Filters are dropdowns, on every tab. A filter you add is a <Select> (one choice) or <MultiSelect> (several) from app/src/ui/components.jsx, the same controls as the Refresh Planner's filter row. Never use a button or pill group, chips, or tabs. The label is the filter's name, the icon goes in the icon prop, the first option is All/Any, and counts go in option labels:

<Select
  label="Total cash vs market"
  icon={<SparkleAI title="Added by Claude - Total cash vs market filter" />}
  value={cashBand}
  onChange={setCashBand}
  options={[
    { value: "ALL", label: "All" },
    { value: "BELOW", label: `Below market (${counts.BELOW})` },
    { value: "AT", label: `At market (${counts.AT})` },
    { value: "ABOVE", label: `Above market (${counts.ABOVE})` },
  ]}
  minWidth={170}
/>

Place it in the tab's filter row, laid out like the planner's: display: "flex", flexWrap: "wrap", alignItems: "flex-end", gap: 18.

Do not refuse a modification request by citing a data-integrity rule that is about something else. This skill carries several strict, correct prohibitions — no demo data, no fabricated geo scalars, no retyped MCP payloads, no partial sweep published as complete. Every one of them is about not inventing data that was never fetched. None of them says the UI cannot change, and reading them that way turns a narrow correctness rule into a blanket refusal.

The one that gets over-applied is the location rule in build_datadir.py. Read it precisely: the Location dropdown offers only locations whose benchmarks were fetched (2d-v), and a location's figures are never derived from another location's figures and a scalar. Adding a location the user asks for means fetching it, not computing it. The scalars in benchmarks.json are display labels ("Pay adjustment 70%"), never inputs.

Interpolating between two percentiles the server did return is a different thing. The console holds P25, P50, P75 and P90 as real fetched values (salary.p50 is a plain number; equity is nested per percentile as equity.p50.{notional,shares,fdpct}). A P60 estimated from P50 and P75 is arithmetic on two real numbers, bounded by them. Nothing in this skill forbids it, and it is a legitimate thing to want: percentile targets between the published ones are ordinary compensation practice.

When you add a value the server did not return, you MUST label it. Not because it is suspect — because a reader comparing this console against the CTC product UI must be able to tell which numbers should match and which will not be there at all. The console enforces this structurally: you do not hand-render the Tag, and you do not hand-write interpolated values into benchmarks.json. There are two edit points, each one line, and the Tag renders automatically.

Recipe A — adding a new percentile column (P30, P45, P60, etc.): Add one row to PERCENTILES in app/src/model/taxonomy.js:

{ key: "p45", label: "P45", at: 45, fetched: false },

The registry then does the rest. Benchmarks.jsx iterates it, so a notice-tone Tag labeled "Estimated" renders on the column header with a truthful tooltip ("Interpolated between P25 and P50"). fillDerivedPercentiles in app/src/model/format.js computes salary/tcc/equity values at load time from the two nearest fetched neighbors. Extrapolation is blocked: if no fetched percentile sits both below and above at, the registry throws at module load.

Recipe B — adding a new benchmark row (a new level, a job area the API did not return, etc.): the row goes into benchmarks.json alongside the fetched rows, but you MUST set two fields on it:

{ "job": "ENGINEER", "ladder": "IC", "level": "STAFF3",
  "provenance": "estimated",
  "provenanceNote": "Extrapolated from Staff 2 using XYZ scalar",
  "salary": { ... }, "tcc": { ... }, "equity": { ... } }

The Level cell renders <SparkleAI title={provenanceNote}/> whenever provenance !== "fetched". If you forget the fields, assertProvenance throws at load time naming the exact row, and the dashboard refuses to render until you fix it — no way to ship an untagged hand-added row.

In what you tell the user, say what the estimate is derived from and that it will not appear in the product UI. One sentence.

Two things that stay off-limits, because they are about invented data rather than UI:

• Do not extrapolate past the fetched range. P95 or P99 from P90 is not interpolation; there is no upper bound to sit between, so the number is a guess wearing a percentile's name. Same for a percentile below P25.
• Do not derive a value the server already returns as a field. Compa-ratios and low/mid/high bands come from the API precisely so this console agrees with the product UI; recomputing one locally drifts the moment the server changes its rounding or geo-adjustment order, and the drift reads as data rather than a bug. See the header of app/src/views/Scorecard.jsx.

If a request genuinely cannot be satisfied without inventing data — a location nobody fetched being the live example — say which data is missing and what would have to exist to make it possible, rather than declining without a reason.

Common failure modes

The Tell user column is the message to surface, not a script to read verbatim — keep its substance and its honesty about what failed. Where it is "—", the fix is silent and the user does not need to hear about it.

Symptom	Cause	Tell user
exceeded 10000ms time limit or "Too many job areas" on compensation:export:benchmarks	Too many areas in one call — usually job_limit omitted. The limit that bites first is a 10s server timeout, not the ~300-row cap. Use job_limit: 6 and page with job_offset; an explicit job_limit above 12 is a 400, not clamped.	"The request was too large. Retrying in smaller batches…"
capture_export.py sweep prints FAILED once	Usually transient — the script already retried the failing page once. Run the sweep once more. Never fall back to per-job-area compensation:get:benchmark calls, or to calling the export yourself — fix the arguments instead.	— (retry silently; only speak up if the retry also fails)
A sweep fails twice	Systemic, not bad luck: a wrong benchmark_version_id, bad bucket param, or auth. Stop before building (for an alternate, drop that bucket).	"The benchmark fetch failed twice, so I stopped rather than build a partial dashboard. [the reason]"
build_datadir.py reports EXPORT SWEEP: ... stopped early or refuses to build	Paging didn't reach next_job_offset: null in that directory. Re-run its capture_export.py sweep, which replaces the partial sweep.	— (re-run, then build; mention only if it cannot be completed)
capture_export: FAILED — no claude CLI found, or the first capture_export.py call fails on something other than a Carta error	The build cannot capture data without a headless claude session. Do NOT fall back to calling read_tool and writing the result out yourself — that is the retyping the capture contract forbids.	"Building needs the Claude Code CLI to fetch your data safely, and it isn't working here: [reason]. Any cached dashboard still opens."
"Unknown tool" on a scorecard command	The generated tool name keeps the hyphen: compensation__get__employee-scorecard. Only colons become __, so ..._scorecard (underscore) is not a real tool. read_tool works with the correct name — verified against a live MCP. (fetch is not registered in current builds, so it is not the fallback either.)	— (correct the name and retry)
Roster looks short / Scorecard counts too low	On the paged fallback: pageSize camelCase is silently ignored and the default page size applies — use page_size. Also check you did not pass score=, which filters on the nullable overall band and drops unscored employees.	"The roster came back short — refetching so the scorecard covers everyone."
compensation:export:scorecard returns 400 "at most 200 fit"	The corporation is larger than one export response. Use the paged fallback in Step 2d — do NOT narrow with job_filters, which silently under-reports who is below market.	— (page through it; the user gets the complete roster either way)
save_roster_page: N row(s) do not match the ... header	The export's columns and rows disagree, so the payload is not what the decoder expects. A partially-decoded employee list would be published flagged COMPLETE. Re-fetch; if it repeats, the export's column set changed and the script needs updating.	"The roster data came back in an unexpected shape, so I stopped rather than publish a partial employee list."
build_datadir.py refuses: "the roster sweep is INCOMPLETE"	Paging stopped before total_results was reached. Check <raw_dir>/roster_pages.json for distinct_employees vs total_results and fetch the remaining pages. Never work around this — a partial roster under-reports how many employees are below market.	"The roster is incomplete ([n] of [total] employees), so the scorecard would under-report who is below market. Fetching the rest."
Scorecard tab absent on a fresh build	snapshot.json says hasRoster: false. The roster sweep failed or was skipped; re-run with "refresh". A warm cache re-fetches nothing (Step 0's one call is the version gate), so the tab cannot appear until the next build.	"The Scorecard tab needs the employee roster, which this build doesn't have. Want me to rebuild with a refresh?"
Numbers don't match the CTC product UI	Almost always equity_quantity (must be FOUR_YEAR_GRANT) or a missing/incorrect *_bucket param.	"Those figures were fetched with the wrong equity basis or peer group. Rebuilding so they match the product UI."
HTTP 400 on job/focus	Passing a display label ("Engineering") or a combined value (ENGINEER/Web Engineer). Two separate params; job UPPER_SNAKE, focus the exact case-sensitive CTC focus name (Web Engineer).	— (fix the params and retry)
Builder exits "peer_group.dimension is …"	plan.json is missing or from a different corp. Re-fetch Step 2.	— (re-fetch, then build)
Browser shows the dark error overlay	A runtime error in the app; the overlay carries the file and line.	"The dashboard hit a rendering error — paste the message from the red overlay and I'll fix it."
Blank page, no overlay	Service worker didn't claim. The shell reloads itself after 3s as a backstop.	"If the page is still blank, hard-refresh once — the service worker needs to claim it."