Babysit Kedro PR

Drive Kedro changes toward a mergeable state — locally before push, or on an open PR with failing CI. Fix mechanical CI failures, resolve clear merge conflicts, ensure DCO sign-off. Use a conservative posture: stop and ask the user before every commit, push, force-push, and merge-conflict resolution.

Use this skill for any local Kedro check or fix — even single-tool runs like "lint my diff", "run ruff on my changes", or "run one pytest". Do not pip install ruff (or mypy, or pytest, or pre-commit) and run them yourself: Kedro's .pre-commit-config.yaml and pyproject.toml already pin those tools with project-specific configuration, and the right venv must be active. The targeted commands in reference.md "Fast lint iteration" reproduce CI's results exactly; ad-hoc installs do not.

This skill complements review-kedro-pr (judgment-based code review). It does not triage PR review comments.

How to invoke this skill

Pick a track first, then run only the steps that track requires. There are two tracks, mapped to the two real situations users hit:

Track	When the user says...	What to do
L. Local dev	"babysit my local changes", "verify before push", "run checks locally", "are my changes ready?", "lint my diff"	Step 2 (env) → Step 5a (full sweep to surface FAILs) → Step 4 (per-failure fix loop with targeted commands) → Step 5a (final re-verify) → optional Step 5b (commit + push). No PR required. Skips GitHub fetch entirely.
C. CI diagnosis & fix	"CI is failing", "fix the lint job", "what's failing on CI?", "fix the X check on this PR", "babysit this PR"	Step 1 (PR detection) → Step 2 (env) → Step 3 (snapshot via scripts/watch_ci.sh) → Step 4 (fix only the failing checks; verify each with the most targeted command — see reference.md section "Fast lint iteration") → Step 5b (commit + push). Skip Step 5a — per-check verify already happened in Step 4.

Picking the right track

• If "babysit this PR" is the only signal, default to Track C (a PR is implied).
• If intent is ambiguous, ask once: "Verify your local changes before pushing (no PR), or diagnose and fix failing CI on an open PR?" Pick a track before invoking any script.
• For very narrow Track L asks ("just lint my diff", "just run ruff on my changes"), skip Step 5a and run the targeted command from Step 4 directly — there's no need for a full-sweep discovery pass.
• Announce the chosen track in one line before invoking any step (e.g. "Picking Track L — no PR signal."), so the user can redirect.

Each step header below is annotated with the tracks that need it (e.g. Tracks: L, C). Skip a step if your track is not listed.

Workflow

1. Identify the PR + preflight Track: C

Detect the PR from the current branch, or accept a PR URL/number from the user:

gh auth status
gh pr view --json number,title,baseRefName,headRefName        # current branch's PR
# or, if the user supplied a PR:
gh pr view <number-or-url> --json number,title,baseRefName,headRefName

If the user provided an explicit PR, thread it through every subsequent command as gh pr view <num> ... and bash scripts/watch_ci.sh --pr <num>. The default (no --pr) only works for the current branch's PR.

Preflight checks:

• gh is installed and authenticated. If not, stop and ask the user to install/authenticate.
• The working tree is clean enough to operate on. Run git status --porcelain; if there are uncommitted changes unrelated to the planned fixes, warn the user and ask whether to stash or proceed.
• A PR exists for the current branch. If no PR exists, ask the user whether to switch to Track L (local-only) or stop and open a PR first (gh pr create); creating PRs is out of scope for this skill.

2. Set up the Python environment Tracks: L, C

Run scripts/bootstrap_env.sh. The script:

• Detects an active isolated env: $VIRTUAL_ENV set, OR ($CONDA_PREFIX set AND $CONDA_DEFAULT_ENV != "base").
• If no isolated env active: ask the user "venv or conda?" + name. Pass to bootstrap_env.sh --type {venv|conda} --name <name>. The script creates the env and prints the activation command. The skill cannot activate the env from a child shell — instruct the user to activate and re-invoke the skill.
• If active but dependencies missing: the script runs make install-test-requirements + make install-pre-commit. It also runs make install-docs-requirements only when invoked with --with-docs — the script does not auto-detect docs files. The agent must determine docs-ness first (see below) and pass --with-docs if needed.
• Probes for system tools vale, lychee, gh. Prints platform-specific install hints if missing; does not install them.

Determining whether to pass --with-docs:

• Track L (no PR exists, so origin/main is the default base): git diff --name-only "$(git merge-base HEAD origin/main)"...HEAD | grep -E '^docs/|\.md$'. Pass --with-docs if there's a match.
• Track C (PR's actual base may differ): gh pr view --json files -q '.files[].path' | grep -E '^docs/|\.md$'. Pass --with-docs if there's a match.

Read reference.md section "Environment setup" for manual recipes and cross-platform install hints.

3. Snapshot the PR state Track: C

Single read-only pass:

BASE="origin/$(gh pr view [<num>] --json baseRefName -q .baseRefName)"   # resolves the PR's actual base
gh pr view [<num>] --json mergeable,mergeStateStatus,headRefOid,files
bash scripts/watch_ci.sh [--pr <num>]                                    # snapshot CI + dump failed-job logs
git log "$BASE"..HEAD --format='%H %(trailers:key=Signed-off-by)'        # commits since base, with DCO trailer

Drop the [<num>] / [--pr <num>] brackets when operating on the current branch's PR. When the user supplied an explicit PR in Step 1, keep them — omitting the number will silently resolve the current branch's PR instead.

Don't hardcode origin/main — Kedro PRs may target develop or other branches.

Show the user a one-screen summary:

• Mergeable status (CLEAN / BEHIND / DIRTY / BLOCKED).
• Failing CI checks, each annotated with the local Make target from reference.md section "CI-to-local mapping".
• Commits missing DCO Signed-off-by: trailer.

scripts/watch_ci.sh is snapshot-only: it fetches the current state and exits without blocking on still-running checks. Re-run it later for a refreshed snapshot.

4. Fix in this fixed order Tracks: L, C — sub-bullets gate per track too

Each sub-step ends with stop and ask before staging / committing / pushing / amending.

The Track C sub-bullets (Merge conflicts → CI failures → DCO sign-off) run in that order when multiple apply. Track L only triggers the Local-only failures sub-bullet at the end.

Merge conflicts (Track C). If mergeStateStatus is DIRTY or BEHIND, sync with the base branch. Resolve conflicts only when the intent is unambiguous; otherwise stop and ask the user.

CI failures (Track C). For each failing check:

1. Look up the local invocation in reference.md section "CI-to-local mapping" (and the cookbook entry below it for e2e-tests, which is intentionally out of scope locally).

2. The failed-job logs were already dumped by watch_ci.sh in Step 3. If the snapshot is stale (e.g., new check runs completed since Step 3), re-run bash scripts/watch_ci.sh [--pr <num>] for a fresh snapshot — keep the --pr flag if the user supplied a PR in Step 1.

3. Propose the smallest scoped fix using reference.md section "Common CI failures and fixes".

4. Apply the fix to the working tree only (no git add / git commit).

5. Verify with the most targeted command available — almost never make lint, which runs every pre-commit hook on every file plus mypy on the whole kedro/ package (typically 3–5 min per iteration). Instead:

   • For a single failing pre-commit hook: pre-commit run <hook-id> --files <changed-files> (seconds).
   • For ruff/format: ruff check --fix <file> or ruff format <file> (sub-second).
   • For mypy: mypy <file> --strict --allow-any-generics --no-warn-unused-ignores (5–10s vs. 1–2 min).
   • For Import Linter: lint-imports --config pyproject.toml (it scans the whole project regardless — but skips pre-commit overhead).
   • For unit tests: pytest --no-cov tests/path/to/test_thing.py::TestClass::test_method (sub-second per test) instead of make test (~6 min for the full suite + 100% coverage check). --no-cov is required because pyproject.toml always applies --cov with fail_under = 100, which would otherwise exit non-zero even on a passing single-test run. Never run make test during the fix loop — pick the failing test path from the CI log and target it directly. Sandbox: always request all permissions when running any pytest in this repo (see reference.md section "Sandbox & permissions").

DCO sign-off (Track C). List commits without Signed-off-by:. Apply one of the recipes from reference.md section "DCO sign-off recipes" (typically git rebase <base> --signoff for older commits). Ask the user before amending; explicitly warn that amend requires a force-push (git push --force-with-lease).

Local-only failures (Track L). Track L drives this loop from the output of Step 5a's full sweep — apply the same targeted-command flow as above for each FAIL.

5. Verify locally → commit + push

Two sub-steps. Run the ones your track requires.

5a. Verify locally (full sweep) Track: L

bash scripts/run_local_checks.sh

The script auto-detects scope from the changed files (code, docs, or code+docs). It refuses to run if no isolated env is active.

Local scope is narrower than CI's path filter. Only docs/** changes trigger local docs checks. Generic *.md files outside docs/ (e.g., RELEASE.md, root README.md) are skipped locally because they don't affect the docs build — but CI's docs-only-checks workflow will still fire on them. The script prints a one-line note when this divergence applies, with the recommended local pre-check command (make linkcheck / make language-lint dir=docs). If those checks fail on CI later, Track C is the safety net.

Before invoking the script, warn the user about the wait time. Total wall time depends on scope:

• code only: ~10–12 min (lint ~3-5 + test ~6 + detect-secrets ~30s)
• docs only: ~5–10 min (lint ~3-5 + linkcheck ~2-5 + language-lint ~30s)
• code+docs (mixed PR): ~12–17 min (everything above)

Request all permissions upfront when invoking via a sandboxed shell tool — the script's plan typically includes make test and/or make linkcheck, which need all and network respectively (see reference.md section "Sandbox & permissions").

The script prints the same plan + per-check duration estimates upfront before running, so the user can confirm they're happy to wait or background it. Tell the user to leave the terminal open — each check streams output as it runs (make test prints per-file dots; make lint prints >>>>>>>> <hook name> headers).

Show the user the per-check pass/fail/skip summary at the end. Any FAILs feed into Step 4's targeted fix loop — re-run run_local_checks.sh only as the final belt-and-suspenders check once all targeted recipes pass.

Track C skips 5a entirely — per-check verification already happened in Step 4 for the specific failing checks; running the full sweep would be redundant and slow.

5b. Commit + push Tracks: L (optional), C

On user confirmation:

• git add only the touched files.
• git commit with a message you propose (or the user provides). Use -s for DCO sign-off.
• git push (or git push --force-with-lease for the DCO rebase case — warn explicitly).

On Track L, commit + push is optional: the user may just want a clean local working tree before continuing to iterate. Always confirm intent.

After push on Track C, do not block waiting for CI. If the user wants to see the new CI state, they can re-run bash scripts/watch_ci.sh [--pr <num>] later for a fresh snapshot — keep the --pr flag if a PR was supplied in Step 1.

Out of scope

• Watching CI to completion (blocking) — watch_ci.sh is snapshot-only by design. The skill surfaces current state and exits; the user re-runs it later if they want a refreshed view. This avoids long blocking sessions where the agent is idle waiting on CI.
• Running e2e tests locally — slow, requires packaging the wheel, and rarely needed during the babysit loop. If e2e-tests is the failing check, the skill surfaces the CI logs and stops — no local reproduction.
• Comments triage — PR review comments, reviewer asks, and bot-generated review feedback (such as Bugbot, the Cursor-style review bot that posts inline comments). The author handles reviewer feedback directly.
• Judgment-based code review — defers to review-kedro-pr.
• RELEASE.md compliance — adding or editing RELEASE.md entries is a reviewer convention, not a CI check (gh pr checks does not surface it). Defers to review-kedro-pr, which flags missing entries as Critical.
• Security review — defers to a future kedro-security-review skill.
• Authoring features / large refactors — out of scope; this skill only applies the smallest scoped fix to make a check pass.
• Performance regressions — pipeline-performance-test (label-triggered on PRs with the performance label) reports timing, not pass/fail. The skill surfaces the timing delta and asks the user to investigate manually.
• Installing system tools — the skill prints install hints for vale, lychee, gh but never invokes brew/apt/cargo on the user's behalf.