ci(ai-review): onboard the AI review callers

.github/workflows/ai-review-manual.yml

@@ -0,0 +1,73 @@
+name: AI PR Review (manual)
+
+# Manually re-run the AI reviewer and/or the judge on any PR. Use after
+# force-pushing, updating prompts, or to refresh a stale review. Deliberately
+# UNGATED (no gate_on_recommendation): its job names give it distinct check
+# contexts, so a manual run can never satisfy a required check.
+#   what=both    review, then eval of the fresh review (default)
+#   what=review  review only
+#   what=eval    judge the LATEST review comment already on the PR (cheap)
+
+on:
+  workflow_dispatch:
+    inputs:
+      pr_number:
+        description: PR number
+        required: true
+        type: string
+      what:
+        description: What to run
+        required: false
+        type: choice
+        options: [both, review, eval]
+        default: both
+
+permissions:
+  contents: read
+  models: read
+  pull-requests: write
+  issues: write
+
+jobs:
+  review:
+    name: AI PR Review (manual)
+    if: ${{ inputs.what != 'eval' }}
+    permissions:
+      contents: read
+      models: read
+      pull-requests: write
+      issues: write
+    uses: everstake/infra-ai-platform/.github/workflows/ai-pr-review.yml@579156f8f4dabf68087c784cbeb228c3fc58214d # v2.12.1
+    with:
+      pr_number: ${{ inputs.pr_number }}
+      variant: universal
+      # Generated lockfiles carry no review value and inflate the diff.
+      diff_exclude_extra: '^diff --git .*(yarn\.lock|package-lock\.json|pnpm-lock\.yaml|go\.sum)'
+      client_id: ${{ vars.AI_REVIEW_APP_CLIENT_ID }}
+      infra_ai_ref: 579156f8f4dabf68087c784cbeb228c3fc58214d
+    secrets:
+      APP_PRIVATE_KEY: ${{ secrets.AI_REVIEW_APP_PRIVATE_KEY }}
+      ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+
+  eval:
+    name: AI PR Eval (manual)
+    needs: review
+    # Runs after a successful review, or alone when what=eval (review skipped).
+    if: ${{ !cancelled() && inputs.what != 'review' && needs.review.result != 'failure' }}
+    permissions:
+      contents: read
+      models: read
+      pull-requests: write
+      issues: write
+    uses: everstake/infra-ai-platform/.github/workflows/ai-pr-eval.yml@579156f8f4dabf68087c784cbeb228c3fc58214d # v2.12.1
+    with:
+      pr_number: ${{ inputs.pr_number }}
+      variant: universal
+      model: claude-sonnet-5
+      require_codebase: false
+      diff_exclude_extra: '^diff --git .*(yarn\.lock|package-lock\.json|pnpm-lock\.yaml|go\.sum)'
+      client_id: ${{ vars.AI_REVIEW_APP_CLIENT_ID }}
+      infra_ai_ref: 579156f8f4dabf68087c784cbeb228c3fc58214d
+    secrets:
+      APP_PRIVATE_KEY: ${{ secrets.AI_REVIEW_APP_PRIVATE_KEY }}
+      ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

.github/workflows/ai-review.yml

@@ -0,0 +1,72 @@
+name: AI PR Review
+
+# Combined merge-gate caller (review job + eval job). Design: ADR-0005.
+# No paths: filter by design — see the ADR before adding one.
+
+on:
+  pull_request:
+    types: [opened, synchronize, reopened, ready_for_review]
+
+permissions:
+  contents: read
+  models: read
+  pull-requests: write
+  issues: write
+
+jobs:
+  review:
+    # Draft is the only caller-level skip: a draft cannot merge, and
+    # ready_for_review fires a fresh run that reports the real context.
+    # Dependabot/fork skips live INSIDE the reusable (skip_dependabot /
+    # skip_forks below) so the check context is still reported as success —
+    # a caller-level `if:` skip reports a different, shorter context and a
+    # required check then hangs forever. Rationale: ADR-0005.
+    if: ${{ github.event.pull_request.draft == false }}
+    permissions:
+      contents: read
+      models: read
+      pull-requests: write
+      issues: write
+    uses: everstake/infra-ai-platform/.github/workflows/ai-pr-review.yml@579156f8f4dabf68087c784cbeb228c3fc58214d # v2.12.1
+    with:
+      pr_number: ${{ github.event.pull_request.number }}
+      variant: universal
+      # Generated lockfiles carry no review value and inflate the diff.
+      diff_exclude_extra: '^diff --git .*(yarn\.lock|package-lock\.json|pnpm-lock\.yaml|go\.sum)'
+      gate_on_recommendation: true
+      require_codebase: false
+      # Label bypass disabled under a merge gate — the only sanctioned
+      # bypass is the org ruleset's bypass list. Empty matches no label.
+      skip_label: ""
+      skip_dependabot: true
+      skip_forks: true
+      client_id: ${{ vars.AI_REVIEW_APP_CLIENT_ID }}
+      infra_ai_ref: 579156f8f4dabf68087c784cbeb228c3fc58214d
+    secrets:
+      APP_PRIVATE_KEY: ${{ secrets.AI_REVIEW_APP_PRIVATE_KEY }}
+      ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+
+  eval:
+    needs: review
+    # A skipped review concludes SUCCESS (so its check context is reported);
+    # the output is what tells us not to judge a review that never ran.
+    if: ${{ needs.review.outputs.skipped != 'true' }}
+    permissions:
+      contents: read
+      models: read
+      pull-requests: write
+      issues: write
+    uses: everstake/infra-ai-platform/.github/workflows/ai-pr-eval.yml@579156f8f4dabf68087c784cbeb228c3fc58214d # v2.12.1
+    with:
+      pr_number: ${{ github.event.pull_request.number }}
+      variant: universal
+      # Generated lockfiles carry no review value and inflate the diff.
+      diff_exclude_extra: '^diff --git .*(yarn\.lock|package-lock\.json|pnpm-lock\.yaml|go\.sum)'
+      model: claude-sonnet-5
+      gate_min_score: 9
+      require_codebase: false
+      client_id: ${{ vars.AI_REVIEW_APP_CLIENT_ID }}
+      infra_ai_ref: 579156f8f4dabf68087c784cbeb228c3fc58214d
+    secrets:
+      APP_PRIVATE_KEY: ${{ secrets.AI_REVIEW_APP_PRIVATE_KEY }}
+      ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

CODEBASE.md

@@ -0,0 +1,23 @@
+# CODEBASE.md — DRAFT (generated by the AI-review onboarding fanout)
+
+> Refine this file: it is injected into the AI reviewer's system prompt as
+> the repo-specific ground truth. Until it names this repo's real conventions,
+> the reviewer works from the universal contract alone.
+
+## What this repository is
+
+<!-- One paragraph: what the service/library does and for whom. -->
+
+## Tech stack
+
+<!-- Language, framework, build commands (e.g. `make lint`, `make test`). -->
+
+## Forbidden patterns (auto-flag in review)
+
+<!-- Repo-specific hazards the reviewer must flag, one bullet each, with the
+     reason. Examples from sibling repos: secrets in code, raw SQL string
+     concatenation, context.Background() inside request handlers. -->
+
+## Acceptable patterns (DON'T flag as bugs)
+
+<!-- Deliberate conventions that look suspicious to a generic reviewer. -->