adbc-drivers · eric-wang-1990 · Jul 31, 2026 · Jul 31, 2026 · Jul 31, 2026 · Jul 31, 2026
diff --git a/.bot/config.yaml b/.bot/config.yaml
@@ -23,16 +23,14 @@ bot_login_prefix: peco-engineer-bot       # <-- your engineer-bot App's login pr
 
 # safe_path denylist (relative to repo root). .git + the secret-scanner allowlist
 # at minimum. NOTE (security): csharp/src is NOT denied — the agent must edit the
-# code it fixes. But .github + .bot ARE denied as defense-in-depth: a bug fix
-# never needs to write the bot's own CI/workflows or config/prompts, and both the
-# author and follow-up system prompts already forbid it. Enforcing it in-sandbox
-# means a prompt-injected issue/review body cannot induce such edits at all,
-# rather than relying solely on a reviewer spotting them. The human PR-review gate
-# remains the backstop — reviewers should still scrutinize any .bot/.github diffs.
+# code it fixes. `.bot` stays denied as defense-in-depth (the agent should never
+# rewrite its own config/prompts). `.github` is INTENTIONALLY allowed (maintainer
+# decision): the engineer-bot may need to fix its own CI/workflow wiring (e.g. a
+# missing engine flag), and the human PR-review gate is the backstop — reviewers
+# must scrutinize any `.github` diffs, since workflows carry secrets/tokens.
 denied_subpaths:
   - .git
   - .gitleaksignore
-  - .github
   - .bot
   # csharp/hiveserver2 is a SEPARATE repo (git submodule -> adbc-drivers/hiveserver2).
   # Edits there cannot be committed to a parent-repo PR, so deny them. The author

@@ -241,4 +241,9 @@ jobs:
           MODEL_ENDPOINT:   https://${{ secrets.DATABRICKS_HOST }}/serving-endpoints/databricks-claude-opus-4-8/invocations
           REPO_RULES_FILES: csharp/CLAUDE.md,rust/CLAUDE.md   # adbc keeps conventions per-language (no root CLAUDE.md)
           DATABRICKS_TOKEN: ${{ secrets.DATABRICKS_TOKEN }}
-        run: python -m databricks_bot_engine.engineer_bot.run --phase followup --bot .bot
+        # --system-prompt is REQUIRED for the followup phase (engine #89 removed the
+        # .bot/ prompt-name fallback). Point at the engineer-followup system prompt
+        # that already exists in .bot/prompts/ (mirrors databricks-jdbc's onboarding).
+        run: >-
+          python -m databricks_bot_engine.engineer_bot.run --phase followup --bot .bot
+          --system-prompt .bot/prompts/engineer-followup/system.md
@@ -246,20 +246,10 @@ jobs:
           OUT_OF_SCOPE:    ${{ steps.author.outputs.out_of_scope }}
         run: python -m databricks_bot_engine.engineer_bot.publish
 
-      - name: Label the fix PR for review
-        # publish applies `engineer-bot` (engineer opt-in) but the reviewer gates
-        # strictly on `review-bot`, so add that label here too — otherwise the fix
-        # PR would never be reviewed and the loop wouldn't close. Keeps the labels
-        # single-purpose (engineer-bot ⇒ engineer, review-bot ⇒ reviewer); the fix
-        # PR simply opts into both. The `labeled` event fires reviewer-bot.yml.
-        if: steps.author.outputs.outcome == 'success' && steps.publish.outputs.coverage_pr_url != ''
-        env:
-          GH_TOKEN: ${{ steps.token.outputs.token }}
-          # Bind the engine-produced URL to an env var rather than interpolating
-          # ${{ }} into the run-script, mirroring the "Comment outcome" step — the
-          # runner shell never parses it as command syntax (defense-in-depth).
-          COVERAGE_PR_URL: ${{ steps.publish.outputs.coverage_pr_url }}
-        run: gh pr edit "$COVERAGE_PR_URL" --add-label review-bot
+      # (Removed "Label the fix PR for review": the reviewer now auto-runs on every
+      # non-draft, non-fork PR — no `review-bot` label needed — so the fix PR is
+      # reviewed without an explicit label-add. publish still applies `engineer-bot`
+      # for the engineer follow-up opt-in.)
 
       - name: Comment outcome on the issue
         if: always()

@@ -55,12 +55,14 @@ concurrency:
 jobs:
   followup:
     # SECURITY: skip fork PRs — keep model/LLM secrets out of untrusted code's reach.
-    # Same `review-bot` label gate as reviewer-bot.yml — the follow-up only acts on
-    # PRs the reviewer was allowed to review.
+    # Like reviewer-bot.yml, there is NO label gate (dropped for consistency with the
+    # other consumer repos). Unlike the initial reviewer, the follow-up deliberately
+    # does NOT require draft == false: it must keep responding to existing review
+    # threads even after a PR is converted to draft. So it runs on any OPEN, non-fork
+    # PR (draft or not).
     if: >-
       github.event.pull_request.head.repo.fork == false
       && github.event.pull_request.state == 'open'
-      && contains(github.event.pull_request.labels.*.name, 'review-bot')
     runs-on: [self-hosted, Linux, X64, peco-driver]
     timeout-minutes: 15
     steps:

@@ -32,12 +32,7 @@ name: Reviewer Bot
 
 on:
   pull_request:
-    # `labeled` is included so adding `review-bot` to an already-open PR triggers
-    # a review immediately, not just on the next push. (The `if:` gate below only
-    # honors `review-bot`; applying `engineer-bot` alone fires this event but the
-    # job is skipped. Engineer-bot fix PRs are reviewed because engineer-bot.yaml
-    # separately adds the `review-bot` label after publish.)
-    types: [opened, synchronize, reopened, ready_for_review, labeled]
+    types: [opened, synchronize, reopened, ready_for_review]
 
 permissions:
   contents: read
@@ -51,18 +46,14 @@ concurrency:
 
 jobs:
   review:
-    # Non-fork, non-draft, `review-bot`-labeled PRs only:
-    #  - fork PRs must never see the LLM/engine secrets; drafts aren't ready.
-    #  - label gate: review ONLY PRs carrying `review-bot`. `review-bot` guards the
-    #    reviewer; `engineer-bot` guards the engineer — strictly separate, never
-    #    mixed. A maintainer opts a PR in with `review-bot`; the bug-fix bot's own
-    #    fix PRs get `review-bot` applied by engineer-bot.yaml after publish, so
-    #    they're reviewed and the loop still closes. Without the label the reviewer
-    #    stays silent, so it doesn't comment on every PR in the repo.
+    # Non-fork, non-draft PRs — auto-review, matching the other consumer repos
+    # (databricks-sql-python/go/nodejs/kernel), which review every non-draft,
+    # non-fork PR without an opt-in label. fork == false keeps the LLM/engine
+    # secrets off untrusted code; drafts aren't ready. (Previously also required a
+    # `review-bot` label; dropped for fleet consistency.)
     if: >-
       github.event.pull_request.draft == false
       && github.event.pull_request.head.repo.fork == false
-      && contains(github.event.pull_request.labels.*.name, 'review-bot')
     # ACTIVATION BLOCKER: before taking this workflow out of DRAFT, switch this to
     # the protected `peco-driver` self-hosted runner (or adopt OIDC + a scoped,
     # short-lived model credential distinct from the CI DATABRICKS_TOKEN). As long