Carta Compensation App

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):

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

Compare its benchmark_version.id against the printed snapshot_benchmark_version_id:

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:

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 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 call_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 . 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. A benchmark response is ~15k tokens of decimal strings (or numbers, for the export); retyping one is slow and a single wrong digit silently corrupts a salary figure that then looks authoritative in the dashboard. Every response reaches disk through save_benchmark_result.py, one of three ways:

Case 1 — the harness persisted the result to a file (the common case for plan / subscription_status). The tool result says something like "Output has been saved to …". Read that path from the message and pass it straight through — the payload never has to pass through your reply:

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

Don't reconstruct the path — the location is client-dependent. Don't hunt for a sibling *-blob-*.json; the helper unwraps the envelope itself.

Case 2 — a small INLINE result (subscription_status, or any small probe). Write the tool result verbatim to a .raw file with the Write tool — copy, don't paraphrase or reformat — then pass that file:

Write <raw_dir>/<name>.raw          ← the raw tool result, unedited
uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_benchmark_result.py" "<raw_dir>/<name>.raw" "<raw_dir>/<name>.json"

Case 2b — an inline result too large to retype comfortably. Both capture scripts accept - as the source and read the payload from stdin, which is the right route whenever the harness did NOT persist the result to a file and the payload is big (a full export page is ~25k tokens). Pipe it rather than hand-copying it into a heredoc:

pbpaste | uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_benchmark_result.py" - "<raw_dir>/<name>.json"

- composes with --export-page and with save_roster_page.py too — verified, and it produces byte-identical output to the file path.

Or write the result to a .raw file with the Write tool, then pass that path. The Write tool reproduces content exactly, so this is a legal capture route and NOT the hand-copying the contract forbids — the distinction is mechanical reproduction versus retyping. Do NOT use a shell heredoc for this: quoting mangles payloads and it invites transcription. Verify every page after capture — --export-page prints the row and job-area counts it captured, and they must match what the response reported (row_count, jobs_covered). A mismatch means the payload did not survive; re-capture rather than building on it.

If none of these routes is available, stop and say so rather than retyping a benchmark payload by hand: that is the one thing this contract exists to prevent, and a truncated or mistyped page is worse than no dashboard.

Case 3 — a compensation:export:benchmarks page (this is how every benchmark row is now fetched — see Step 2c). Same source rules as Case 1/2/2b (persisted path, verbatim .raw file, or - for stdin), but pass --export-page and give it <raw_dir> itself as the destination, not one .json file — a page covers up to 6 job areas at the recommended job_limit, so it fans out into that many benchmark_<JOB>.json files in one call:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_benchmark_result.py" \
  --export-page <printed_result_path_or_.raw_file> "<raw_dir>"

Every case the helper does the unwrapping (bare JSON, content blocks, base64 resource.blob, string-valued result, or the columnar export envelope) and exits 2 rather than write something the builder would misread.

If you find yourself about to type benchmark numbers into a heredoc or a Python literal, stop — you are in the failure mode this contract exists to prevent. Use Case 1, 2, or 3.

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

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

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

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

2b. Plan. Write the raw response to a temp file, then normalize:

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

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).

Call it with equity_quantity: "FOUR_YEAR_GRANT", benchmark_version_id, exactly one <dimension>_bucket param, job_offset, and job_limit: 6.

PAGING IS THE DEFAULT, not an opt-in. A response that omits jobs/job_limit is the FIRST PAGE — not the whole matrix. There is no way to get all 22 areas in one call.

Keep calling with job_offset = the previous response's next_job_offset until that field comes back null. total_job_areas tells you how many exist overall (22), so you can tell a stalled sweep from a finished one without counting job names yourself. Treating one response as the complete cube would silently publish 6 of 22 job areas as if it were everything — this is the exact failure mode the paging discipline below exists to prevent.

Capture every page the moment it arrives, via the export-specific mode:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_benchmark_result.py" \
  --export-page <printed_result_path_or_.raw_file> "<raw_dir>"

This is a third capture path, in addition to the two in the contract above — pass it the same kind of source (a harness-persisted result path, or a verbatim .raw file you Wrote first), but give it <raw_dir> itself as the destination, not a single .json file. It fans one page out into one benchmark_<JOB>.json per job area the page covered (matching the per-job-area files the old sweep produced, so build_datadir.py is unchanged), and appends the page's paging metadata to <raw_dir>/export_pages.json. It prints whether the sweep is COMPLETE or INCOMPLETE after every page — read that line; it is the authoritative answer to "am I done paging."

Never echo benchmark figures into the conversation — they belong in the data dir. Report progress as "N of 22 job areas covered so far" between pages, not row counts or numbers.

Check coverage between pages the same way as before:

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 — i.e. paging stopped before total_job_areas was covered. 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.

Mid-sweep a non-zero exit is expected — after page 1 of 2, most areas are legitimately still MISSING. Only the final check needs to come back clean (exit 0). Do not treat a non-zero exit between pages as a failure and re-fetch what you already have.

When a page-level call fails (403, 5xx, timeout, malformed response):

1. Retry that page once, same arguments (same job_offset). Transient 5xx and timeouts usually clear.
2. Never fall back to per-job-area compensation:get:benchmark calls to route around a failing export call. That is the ~22-call pattern this command exists to replace — if the export is failing, the fix is the export call's arguments (wrong benchmark_version_id, bad bucket param, auth), not a different endpoint.
3. If a page fails twice, stop before building and surface the error — with only 4 calls in a full sweep, a page that won't succeed after a retry means something systemic, not bad luck on one of 22 job areas. A timeout on a job_limit: 6 page is NOT a reason to retry with a bigger limit; if 6 areas time out, a larger page certainly will.

A hard ceiling regardless of outcome: at most 8 export calls per build for the corp's OWN group (4 pages plus retries). 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:

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:

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 — --export-page takes the destination, so point it at peer_<CODE>/ instead of the raw dir:

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_benchmark_result.py" \
  --export-page <printed_result_path_or_.raw_file> "<raw_dir>/peer_<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. Fan the sweep out across subagents instead (see below) and those 68 calls cost roughly one bucket's wall clock, because buckets are independent. Parallelism, not a smaller scope, is the answer to the cost.

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: at most 6 export calls per bucket (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 partial peer_<CODE>/ directory 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. Fan the alternates out across subagents.

Do the alternates in parallel, one subagent per bucket. Buckets are independent — each is a self-contained 4-page sweep writing to its own peer_<CODE>/ directory, with no shared state and no ordering between them. Sweeping them in one context is the slowest possible arrangement and, on a client where MCP results arrive inline rather than as a file path, also the most fragile: every page's payload has to pass through the orchestrator twice (once arriving, once being written), and each of those hand-reproduced pages is an opportunity to corrupt a salary figure. Delegating puts the payloads in the subagent's context instead, so the orchestrator never handles them.

Split one bucket per agent, 17 agents. The harness caps concurrency (~10 at once) and queues the rest, so this lands in about two waves and the wall clock is a couple of buckets deep rather than 17. If the MCP starts rate-limiting, fall back to ~6 agents of ~3 buckets each — do NOT respond to rate limiting by narrowing the scope.

Do the version gate ONCE, in the orchestrator, before spawning anything. Resolve the corporation, run Step 2a/2b, and hand every agent the already-pinned benchmark_version_id. Seventeen agents each calling get:plan is 17 wasted calls, and worse, a release published mid-sweep could leave different buckets pinned to different versions — figures that silently disagree across the picker.

Each agent's brief needs, at minimum:

• Its bucket code(s), and the matching bucket param for that dimension — post_money_bucket / headcount_bucket / capital_raised_bucket, exactly one per call.
• The pinned benchmark_version_id, corporation_id, and equity_quantity: "FOUR_YEAR_GRANT".
• The absolute raw-dir path, and the destination <raw_dir>/peer_<CODE> — the code spelled exactly as in the table above, since the builder reads the dimension and label back out of the directory name.
• The paging discipline: job_limit: 6, page until next_job_offset is null.
• The capture contract, in full. The agent must write the tool result verbatim and pass it to save_benchmark_result.py --export-page; it must never retype or summarise a payload.
• A verification step, done by PARSING — not by eye. The script prints the job areas and row count it captured; the agent parses the response's own jobs_covered and row_count and compares them programmatically, then re-captures on any mismatch rather than proceeding. Say this explicitly in the brief: an agent told only to "confirm the counts match" will skim two numbers that look alike, and a single transposed digit inside a 100-row page changes neither the row count nor the job list. Structural comparison is what catches a payload that arrived intact but was written back wrong — the one failure the capture contract exists to prevent, and one that a real sweep has already produced and caught this way.
• The per-bucket ceiling (6 calls) and the instruction to delete a partial peer_<CODE>/ and report it rather than leaving it half-swept.
• Never echo benchmark figures in its report — the same rule that applies here. Agents report bucket codes, page counts and row counts only.

Sanity-check the figures themselves once the sweep lands. Counts and job lists prove a page arrived; they say nothing about whether its numbers survived being written back out. One cheap pass over the captured rows catches what counting cannot:

uv run - <<'PY'
import json, glob, os
bad = 0; rows = 0
for f in glob.glob(os.path.expanduser("<raw_dir>/peer_*/benchmark_*.json")):
    for r in json.load(open(f)).get("benchmarks") or []:
        rows += 1
        p = (r.get("salary_benchmarks") or {}).get("percentiles") or {}
        v = [float(p[k]) for k in ("p25", "p50", "p75", "p90") if p.get(k) is not None]
        # Percentiles are ordered by construction, and a salary outside this range
        # is not a number the API returns -- either signals a mangled write.
        if v != sorted(v) or any(x < 1000 or x > 50_000_000 for x in v):
            bad += 1
print("rows", rows, "anomalies", bad)
PY

A transposed digit usually breaks the ordering or leaves the plausible range, so this turns a silent corruption into a visible one for the cost of a second. It is a smoke test, not a proof — a wrong digit that happens to preserve both properties still slips through, which is why the per-page parsed comparison above remains the primary defence.

Verify the fan-out centrally when the agents return. Do not trust the reports alone — read each bucket's peer_<CODE>/export_pages.json and confirm sweep_complete: true. A retried page can appear twice in a manifest's pages list; that is harmless, because a re-capture overwrites the same per-job-area files rather than appending, but a manifest whose sweep_complete is false means that bucket really is short and must be re-run or dropped.

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/save_roster_page.py" \
  --export-scorecard <printed_result_path_or_.raw_file> "<raw_dir>"

Same source rules as the capture contract above (harness-persisted path, or a verbatim .raw file you Wrote first), and the same destination convention as --export-page: pass <raw_dir> itself, 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.

Call compensation:get:employee-scorecard with page_size: 10, then page with page: 2, 3, …. Capture every page the moment it arrives (no --export-scorecard flag):

uv run "${CLAUDE_PLUGIN_ROOT}/skills/carta-compensation-app/scripts/save_roster_page.py" \
  <printed_result_path_or_.raw_file> "<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.

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-bis. Equity refresh report — the Refresh

…(truncated)