dbt PR review: CI experience hides the engine's proofs (undocumented artifacts, misleading lint-only banner, silent AI lane, guessed base ref)

Start with the end-to-end dbt PR review on jaffle_shop_duckdb, then trace summary.degraded in orchestrate.ts and format.ts and the failure paths in runAiReview. Read github/review/action.yml and github/review/examples/altimate-ingestion.yml alongside the GITHUB_EVENT_PATH path. Done means the listed CI statuses, artifact hints, metadata, telemetry fields, grouped findings, and documentation are represented honestly.

Summary

Running the dbt PR review end to end (altimate-code 0.10.0 on jaffle_shop_duckdb with five injected changes) and mining the 30 most recent reviews in our own dogfood repo surfaced a cluster of CI-experience problems. None change what the engine proves; all change whether users see it.

Problems

1. Artifacts that decide fidelity are undocumented. Equivalence and lineage-breakage proofs need target/catalog.json (dbt docs generate) and compiled SQL for the base ref in target-base/compiled. Neither appears in the usage doc quick-start, github/review/action.yml, or github/review/examples/altimate-ingestion.yml; target-base is mentioned only in prose. 57% of our dogfood reviews ran lint-only.
2. "Lint-only run — no dbt manifest/warehouse was available" is shown when a manifest was present. summary.degraded is runDegraded || findings.some(f => f.degraded), so a single undecidable finding flips the whole-run banner (orchestrate.ts, format.ts).
3. The AI lane fails silently. Every failure path in runAiReview returns [] with only a log; telemetry shows 18% of CI runs call review_ai_prompt and never review_ai_parse. A default run with no credentials attempted a stale local endpoint every time; the docs say the lane is "skipped".
4. Base ref is guessed. defaultBaseRef walks origin/main → origin/master → main → master and ignores the event's pull_request.base.ref. A PR against deployment was diffed against main.
5. One rule is half of all findings. "new model has no uniqueness/grain test" was 427 of 853 findings across 30 dogfood PRs, one bullet per model, never acted on.
6. Headless review telemetry has no project_id. Telemetry.setContext is only called from the interactive session loop, so altimate review events cannot be grouped by repo.
7. PR title/body never reach the AI lane in CI, so its intent check cannot run.
8. Docs describe PII classification as "columns that newly expose sensitive data"; the lane flags any PII column present in a touched model.

Fix (this PR)

Separate lintOnly from undecidableFindings in the envelope and render each honestly; add artifact hints with the exact commands; AI lane returns ok | skipped | timeout | error with a reason, rendered in the summary and recorded in review_run (ai_status, ai_findings, undecidable_findings); scale the AI timeout with prompt size; read base ref and PR title/body from GITHUB_EVENT_PATH; group repeated grain-test findings in the summary while keeping atomic findings; set project_id on the headless path; update the docs, action and example workflow to compile head and base and generate the catalog.

Out of scope, tracked in docs/internal/2026-09-03-dbt-pr-review-deep-dive.md

Verdict provenance (catalog rules can currently block, contrary to the README), loading .altimate/review.yml from the base rather than the PR head, gate lifecycle / override wiring, inline-comment dedupe, positional equivalence false positive (core), feedback loop.