fix(review): an honest PR comment: real lint-only banner, AI lane status, PR intent, project_dir

RA
11h ago
+419-29

Issue for this PR

Closes #

Stacked on #1719; please review and merge that one first. Takes ideas from the closed #1241, not its code.

Type of change

  • Bug fix
  • New feature
  • Refactor / code improvement
  • Documentation

What does this PR do?

The review comment could misstate what it checked:

  • Lint-only banner: a single undecidable finding set degraded, so every manifest-backed run said "Lint-only run — no dbt manifest". Now it shows only when no changed model resolved against a manifest. Undecidable findings, models without compiled SQL, and files outside the project each get their own line.
  • Silent AI lane: the lane returned [] on every failure, so an expired key, a timeout, or a missing model looked like "no AI findings". It now reports ok | skipped | timeout | error with a reason (credentials masked), the model, and duration. These land in a new signed ai field and in the comment. The timeout now scales with prompt size (60 s + 1 s per 2K chars, capped at 5 min) instead of a fixed 60 s. --ai-model provider/model picks the model explicitly.
  • No PR intent: prTitle/prBody existed but nothing passed them. The CLI now reads them from the GitHub pull_request event, or from GitLab's CI_MERGE_REQUEST_TITLE/_DESCRIPTION.
  • Action: new project_dir input, mapped to --cwd, for monorepos.
  • Docs: a table of what each input unlocks (fetch depth, manifest, compiled SQL, target-base/, catalog.json) and a snippet for compiling the base.
  • Telemetry: review_run gains undecided, unverified_models, ai_state, ai_duration_ms, base_source, out_of_scope_files.

Open question, deliberately not changed: --no-ai still auto-registers the free tier, which is a network call. A release test asserts that on purpose, so the decision is the maintainers'. Changing it is a one-line guard.

How did you verify your code works?

  • New test/altimate/review-ci-honesty.test.ts (13 tests):
    • banner with and without a manifest
    • AI lane: disabled, outside the tier, timeout (signature still verifies), throwing, ok with model and selector, legacy array return
    • timeout scaling
    • PR context from GitHub and GitLab env, and its forwarding to the lane
  • 1,859 tests across the review files and release gates pass. bun typecheck is clean.

Cost side effects

None. The AI lane's timeout ceiling can rise from 60 s to 5 min on very large prompts, but the number of model calls doesn't change (still at most one per review).

Screenshots / recordings

N/A (CLI).

Checklist

  • I have tested my changes locally
  • I have not included unrelated changes in this PR
  • I checked for cost side effects (model calls, queries or cloud resources that grow with user data) and described them, or wrote "none"

🤖 Generated with Claude Code

Summary by cubic

Makes the dbt PR review comment candid about what it checked. Previously a single undecidable finding labeled any manifest-backed run as "Lint-only", and the advisory AI lane returned [] on every failure, so an expired key, timeout, or missing model looked like "no AI findings".

Review honesty

  • The lint-only banner now shows only when no changed model resolved against a manifest; undecidable findings, models without compiled SQL, and out-of-scope files each get their own line.
  • The AI lane now reports ok | skipped | timeout | error with a masked reason, model, and duration, on a new signed ai envelope field and in the comment; it never fails silently.
  • The AI timeout scales with prompt size (60 s + 1 s per 2K chars, capped at 5 min); --ai-model provider/model selects the model explicitly.
  • The CLI passes the PR title and body to the intent check, reading them from GitHub's pull_request event or GitLab's CI_MERGE_REQUEST_TITLE/_DESCRIPTION.

Inputs, docs, and telemetry

  • New project_dir input maps to --cwd for monorepos.
  • review_run telemetry gains undecided, unverified_models, ai_state, ai_duration_ms, base_source, and out_of_scope_files.
  • Docs now list what each input unlocks and how to compile the base into target-base/.

Deliberately unchanged: --no-ai still auto-registers the free tier, which is a network call; a release test asserts that, so the decision is the maintainers'. Model call count is unchanged (still at most one per review); only the timeout ceiling rises, to 5 minutes on very large prompts.