fix(ai-cost): stop the Claude Team path losing data it already has

[Gregory91G]

Closes #2606. Part of #2479. Depends on #2528 (merged) — the Bronze key it added is what makes the overage half of this safe.

Six findings from reading the Claude Team path end to end. Each is its own commit.

The seat status was a filter; it is now a column

WHERE status = 'active' in Claude Team staging looked like a seat-state gate. It is retroactive: the endpoint restates the status for every day it re-reads, so the value describes the seat as of the read and not as of the metric date. The day a person's seat is deactivated, the filter deletes their entire history — and a rebuild removes it from silver altogether. Nothing about a seat's present state licenses deleting the work it recorded.

The value now rides in the class-contract column seat_status, appended last so the contract grows without a rebuild, and emission is gated on the activity counters instead — the same gate every other contributor to the class already uses. Sources with no seat lifecycle concept emit NULL and read as unknown at gold, where the value becomes a dimension of the ai_usage source.

Audit decision D1 said the opposite (keep the filter, add a check counting non-active rows); the spec now records why that reading was wrong.

One incremental boundary for several producers

Silver classes carried WHERE _version > (SELECT max(_version) FROM {{ this }}). A class is written by every connector that feeds it, and union_by_tag rebuilds it from all feeders on whichever connector's run fires. _version is a timestamp, but its meaning differs per producer — extraction time for the AI connectors, write time elsewhere — so a single maximum over the whole table lets whichever producer commits first raise the boundary above another producer's rows. Those rows sit below it forever.

The boundary moves to a maximum per (tenant, source instance) — the shape already merged for the crm classes and class_person_attribute_claims. _version, its values and its role in ReplacingMergeTree are untouched; only the group the maximum is taken over. coalesce(max_version, 0) is load-bearing: a source absent from the target has no boundary row, and comparing against NULL would drop every row of every new source.

This fixes the future only. Rows already below a boundary stay below it, so recovering them is a separate, deliberate operation.

assert_ai_staging_rows_reach_silver is new and reports what staging can prove a class is missing, per source — there was no check for this at all. check-dbt-conventions prescribed the table-wide form and would have kept reintroducing it.

Counts that stopped in silver

prs_with_cc_count and prs_total_count have been in the class unread since it gained them. They are served as ai.prs_with_assistant and ai.prs_total, emitted only where the vendor supplies a value, so an organisation without the vendor's code-host connection returns no value rather than a zero asserting "no pull requests involved the assistant". ai.prs_total is context for the other measure and not a goal of its own: it is the vendor's count over the vendor's own window, which need not be the day it is reported against.

Seat spend joins identity coverage

identity_resolution_coverage listed every evidence relation except the seat one — the only relation whose unresolved rows are money rather than activity.

Vendor fields that reached nothing

The code-metrics response carries org-level pull-request attribution and top-user rankings beside the per-user array; they were read and discarded on every sync, and the vendor keeps no history. claude_team_code_metrics_org keeps them as their own stream: one request and one row per day, per-user array stripped so the roster is not stored twice. Copying them onto every user row would multiply one organisation fact by the roster and make any sum wrong.

disabled_reason, disabled_until and account_name join the overage extras blob, where a zero spend can now be told apart from a seat the vendor blocked.

A rebuild that deletes months

full_refresh=false guards claude_team__ai_overage. That model is the only place a past month's closing spend exists — the endpoint keeps no history — so a rebuild from Bronze deletes those months rather than reproducing them, and reconcile-connectors dispatches exactly that rebuild on a MAJOR descriptor bump. Silver still rebuilds freely: it reads the staging model, not Bronze.

Also

conversation_count is documented for what it holds. For the dev-agent tools no vendor publishes a separate conversation counter, so the connectors map their session or thread count into it and the value equals session_count on those rows. That is the intended mapping rather than a duplication to collapse, and it does not hold on the assistant class, where the two differ per surface. ai.dev_conversations gains the explanation its number needed.

The connector README described the pre-#2528 shape of the overage key.

Descriptor versions

claude-team 2.1.0 → 2.2.0, cursor / chatgpt-team 2.0.0 → 2.1.0, claude-enterprise 3.0.0 → 3.1.0. All MINOR, deliberately not MAJOR: a MAJOR dispatches the one-shot dbt --full-refresh, and the guard that makes it safe arrives in this same change — it has to be on main before any MAJOR bump reads it. Rows written before this version keep an empty seat_status and read as unknown, which is what they are.

Test plan

Automated, run on this branch:

• dbt parse — clean
• dbt compile of the touched models — the compiled SQL carries seat_status, both new measures and the ai_cost_metric_evidence branch
• cargo test -p analytics --bins — 420 passed, 0 failed
• cargo fmt --check and cargo clippy -- -D warnings — clean
• source.sh validate and validate-strict for ai/claude-team — manifest valid; discover returns five streams
• both DDL snapshot edits reproduced against a throwaway ClickHouse running the real destination connector and the real promotion, rather than hand-written
• ruff at the pinned version, plus the remaining pre-commit hooks

Needed the CI rig, now run on this PR:

• e2e-bronze-to-api, ai shard — 4m20s, green, including the new ai_pr_attribution.test.yaml. The other three shards are green too, and Metric coverage gate passes, so both new metric keys carry the period / peer / timeseries / breakdown coverage that gate requires
• connectors-ddl snapshot and field-parity gate — 6m31s, green: a warehouse built from scratch by the real pipeline reproduces both edited snapshots, and every staging contributor still lines up with its silver union target
• tests/stand — the api-smoke job of the Stand E2E lane, green

Build runner image was cancelled on the first attempt at its 30-minute ceiling while the merge queue was saturated; re-running the same commit passed, so the timeout was contention rather than a defect. Every other check is green.

[coderabbitai]

Important

Review available on request

• 🔍 Trigger review

Reviews should be triggered manually for repositories with fewer than 10 stars. Select Trigger review above or comment @coderabbitai review to review the latest changes. For a full review, comment @coderabbitai full review.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: e2261fff-8eb9-4dc3-aa37-20dc34bba3a3

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}