<required> **You MUST ask the user the Step 1 mode-choice question before doing anything else.** Do not pick a mode for them. Do not default to a mode. Auto/autonomous mode does NOT authorize you to skip this question - being in auto mode is never a valid reason to bypass the prompt.

There are exactly two narrow exceptions where you MAY skip the Step 1 mode-choice question. Both must be applied conservatively; if you are not certain an exception applies, you MUST ask.

1. No-op since the last check in this conversation. The most recent CI readiness check in the current conversation already ran, and every change since then is obviously/provably outside the scope of any check the skill performs. Examples that qualify: edits only to comments; edits only to files outside any workspace package (e.g. files under .claude/, repo-root README.md / CHANGELOG.md, top-level docs not in a docs/* package). Watch out: docs/api/ and similar are themselves workspace packages — markdown inside a workspace package does NOT automatically qualify. If you cannot quickly verify the changed files are outside every workspace package, ask.
2. Explicit standing instruction. The user has explicitly told you — in this conversation or in a saved memory entry — to auto-skip the CI readiness check without asking. Inferred preference, prior approvals of past runs, or general "be less interruptive" guidance does NOT qualify.

If either exception applies, treat it as equivalent to the user picking Skip in Step 1: do NOT ask the mode-choice question, do NOT create CI-readiness tasks/todos, and do NOT run the CI script or any other build/test/API-report steps. Stop the skill after reporting the skip.

When you apply an exception, you MUST explicitly report — every single time, with no abbreviation across repeated runs — that (a) you skipped asking the mode-choice question, (b) the CI readiness check itself was skipped, and (c) which exception applied. Silent skipping is forbidden.

When neither exception applies, ask the user and wait for their response. Immediately after they respond, create one task/todo item per applicable step using your available task tooling (TaskCreate for Claude, TodoWrite for Copilot) — before doing any other work. Mark each task in_progress when you start it and completed when you finish. This prevents steps from being silently skipped as context grows.

Tasks to create by mode:

• Check: Run CI script → Review output → Report final status
• Build: Run CI script → Review output → Build unbuilt packages → ESLint auto-fix → Regenerate API reports → API changes review → Run build:docs → Regenerate type tests → Report final status
• Test: same as Build, plus Run tests

For Build/Test: if @fluidframework/tree is among the changed packages and its API surface likely changed, add a "Cascade API reports to aggregator packages" task after "Regenerate API reports".

Omit steps that don't apply (e.g. skip "Build unbuilt packages" if everything is already built; skip "Regenerate API reports" and "Regenerate type tests" if the API surface didn't change). </required>

Step 1: Confirm with the user

Before doing anything, ask the user:

I can run a CI readiness check on your branch. Pick a mode (fastest to slowest):

1. Skip — skip the CI readiness check
2. Check (quick) — auto-fix formatting, policy, and syncpack on changed packages
3. Build — Check + build unbuilt packages + ESLint + regenerate API reports and type tests
4. Test (slower) — Build + run the test suite in changed packages

Wait for the user's response. If they say skip (or anything clearly negative), stop here. Otherwise, note their choice and immediately create tasks for all remaining steps as described in the required block above before proceeding.

Step 2: Run the script

Run the bundled script from the repository root:

bash .claude/skills/ci-readiness-check/ci-readiness-check.sh [base-branch]

…