Fluid Framework Client Release

Release workflow for the client release group. Supports two modes: interactive (default) and autonomous.

Pipeline behavior can change: verify the actual run's candidate, metadata, and stages before recommending a publishing action. If the run differs from these instructions, report the discrepancy and investigate before proceeding. If you cannot access the run state, report that limitation rather than guessing or queueing a duplicate run.

Environment Detection

Check the CI environment variable at the start of every session:

• CI=true: Running in a GitHub Actions workflow. Use only origin (no upstream). Use CI-safe commands (see CI-safe alternatives). Never prompt for input. Log blockers and phase completions to workflow output for human review.
• CI unset or false: Running locally. Use upstream/origin detection as described in Key Context.

Select the remote once for repository-state checks:

release_remote=origin
if [ "$CI" != "true" ] && git remote get-url upstream >/dev/null 2>&1; then
    release_remote=upstream
fi

Mode Selection

At the start of every release session, ask the user:

Would you like to run in interactive or autonomous mode?

In CI (CI=true), always use autonomous mode — do not prompt.

Interactive Mode (default)

Run commands autonomously but pause before creating PRs, pushing branches, or triggering builds. Ask for version confirmation at key points. This is the current behavior.

Autonomous Mode

Run the entire selected phase end-to-end without pausing. Auto-detect the release state from the repo (see below). If the user provides version info upfront, skip the version questions entirely.

Requirements:

• Version detection: Infer version numbers from repo state by default (see below). If the user provides versions, use them unless they conflict with repo state.
• No confirmation pauses: Create PRs, push branches, and run flub release without asking. Include clear commit messages and PR descriptions.
• Phase-scoped execution: Each phase runs to completion, then reports what the user needs to do next (e.g., "queue the ADO release build", "monitor the automatically triggered publish run", or "prepare type test updates after publication").
• Fallback to issues: If any step fails (permission errors, CI-safe command failures, git push rejected, or any other error that prevents progress), stop and open a GitHub issue in microsoft/FluidFramework describing what was completed, what failed, and what remains. Use the title format Release <VERSION>: <brief description> and label it with release-blocking. Include the exact commands remaining so a human can finish using the skill in interactive mode.

Auto-detecting release state (autonomous mode and CI)

Auto-detect the release state from the repo using the detection steps below. In interactive mode, this detection also runs when the user gives a generic request like "do the release" without specifying a version or phase.

Step 1: Identify the most recent release.

# Get the latest client release tag
git tag -l 'client_v*' --sort=-version:refname | head -1

Step 2: Identify the release version from repo state.

Inspect release-prep branches/PRs, release branches, and tags for a release already in progress. Use the version on main for a new minor release and increment its minor version by 1 for the next development version. If the version bump has already merged, use the release-prep PR or pre-bump commit to recover the release version.

# Inspect the development version and existing release branches
git fetch "$release_remote" main
git show "${release_remote}/main:package.json"
git ls-remote --heads "$release_remote" 'release-prep/*' 'release/client/*'

Ask for clarification only when repo state is ambiguous or conflicts with user-provided versions or phase. In CI, report the ambiguity or conflict and stop instead of prompting.

Step 3: Check if a release is in progress.

# Check for release-prep branches for the next version
git ls-remote --heads "$release_remote" 'release-prep/<NEXT_VERSION>/*'
# Check for the release branch
git ls-remote --heads "$release_remote" 'release/client/<NEXT_MAJOR>.<NEXT_MINOR>'
# Check for a release tag
git tag -l 'client_v<NEXT_VERSION>'
# Check for open PRs
gh pr list --repo microsoft/FluidFramework --search "release-prep/<NEXT_VERSION>" --state all

For a release branch with no completed release, also inspect ADO pipeline state:

1. Find the Build - client packages (definition 12) run for the release branch tip with releaseBuildOverride: release.
2. If that run succeeded, find the automatically triggered Publish - client packages (definition 219) run whose candidate resource matches that build. Verify the candidate's release branch and commit, and confirm the Publish to public npmjs.org through ESRP Release API stage succeeds, rather than being skipped. Publication destinations come from the candidate's recorded metadata.
3. If ADO pipeline state cannot be queried, do not infer which action is next. Report that a human must identify whether the release build or publish pipeline has completed.

Step 4: Determine the phase and act.

Present the detected state and chosen action to the user (or in the issue body). Example:

Detected state: main is at 2.91.0. No release-prep branches found. The most recent release is 2.90.0. Action: Minor release prep needed for 2.91.0 (next version on main: 2.92.0).

Workflow Selection

Use the requested phase, or auto-detect it from repo state as described above. Ask only if the phase is ambiguous or conflicts with the request; in CI, report the blocker and stop.

For patch releases, skip directly to release execution on an existing release branch.

Human checkpoints

These steps require human action and should be clearly reported in CI workflow logs:

1. Merge release-prep PRs in the correct order (version bump last) after CI creates them
2. Create the release branch (Step 5) — requires elevated permissions on the release/ branch prefix
3. Queue the ADO release build (Step 6) — choose the "release" option in ADO
4. Approve manual publishing recovery, if needed (Step 6) — normal publishing auto-triggers; a manually queued recovery run requires the publication approval gate
5. Announce the release in the Fluid Framework "General" Teams channel

Key Context

• The repo uses pnpm as the package manager
• flub is the Fluid build CLI (pnpm flub ... or pnpm exec flub ...)
• Version scheme: The version numbering is not a simple incrementing pattern (it is NOT always multiples of 10). When suggesting a next version, default to incrementing the minor version by 1 (e.g., 2.90.0 -> 2.91.0). Trust the version the user provides unless it is more than 7-8 minor versions away from the current version (which likely indicates an error).
• Release branch naming: release/client/<major>.<minor> (e.g., release/client/2.90)
• The release branch is created from the commit before the version bump on main
• There is no lerna.json in this repo
• Git remote preference: When pushing branches, prefer pushing to upstream if one is configured for the repo. Check with git remote -v if unsure. Only fall back to origin if no upstream remote exists. Exception: In CI (CI=true), always use origin — there is no upstream.
• Working branch naming: Do NOT use the release/ prefix for working branches because release/ is protected on upstream. Use the standard naming convention below — these branches double as progress markers.

CI-safe Command Alternatives

Some flub commands require interactive TTY input. In CI, use these alternatives:

After using npm pkg set to bump versions, also run pnpm -r run build:genver to update packageVersion.ts files, and pnpm install --no-frozen-lockfile to update the lockfile.

Non-interactive flub commands that are safe in CI (no TTY required): flub generate releaseNotes, flub generate changelog, flub typetests, flub release prepare.

Working Branch Convention

Working branches follow a numbered naming scheme under release-prep/<VERSION>/:

Example for releasing 2.90.0 with next version 2.91.0: release-prep/2.90.0/1-tag-asserts, release-prep/2.90.0/4-bump-2.91.0

Detecting Prior Progress

Before starting any phase, check for existing progress by looking at branches and PRs:

# Check for existing release-prep branches on the selected remote
git ls-remote --heads "$release_remote" 'release-prep/<VERSION>/*'
# Check for open release-prep PRs
gh pr list --repo microsoft/FluidFramework --search "release-prep/<VERSION>" --state all

If branches or PRs already exist, skip completed steps and resume from where the process left off.

Before Starting

Run pnpm flub release prepare client to check readiness. Then check for release blockers:

gh issue list --repo microsoft/FluidFramework --label release-blocking --state open
gh pr list --repo microsoft/FluidFramework --label release-blocking --state open

If either command returns results, stop and report the blockers to the user. In autonomous mode, do not proceed past this check if blockers exist. Also remind the user to check ADO for release-blocking issues (cannot be queried via CLI).

Blocker handling (autonomous and CI): If the agent is blocked at any point (release blockers, missing release tag, npm packages not available, permission errors, or any other issue that prevents progress), open a GitHub issue in microsoft/FluidFramework describing what was completed, what failed, and what human action is needed. Use the title format Release <VERSION>: <brief description> and label it with release-blocking. Include the exact commands remaining so a human can finish using the skill in interactive mode. Then exit gracefully. For confirmed post-release feed propagation delays, follow type test update recovery instead of immediately filing a release blocker.

Behavior by Mode

Commands (both modes)

Run these autonomously in both modes: policy-check:asserts, layerGeneration:gen, flub generate releaseNotes, flub generate changelog, build:genver, flub typetests, flub release prepare

For version bumps, use flub bump locally or CI-safe alternatives in CI (see CI-safe alternatives).

PR Conventions

Use the build: conventional commit prefix for all release PR titles (e.g., build: tag untagged asserts for 2.90.0 release).

Checkpoints

Autonomous Mode: Phase Completion Reports

At the end of each autonomous phase, provide a summary:

1. What was done — list all PRs created, branches pushed, commands run
2. What to do next — specific manual steps needed (e.g., queue ADO build, merge PRs in order)
3. When to continue — timing guidance for the next phase (e.g., "after PRs merge" or "after publication is verified; prepare type test updates without waiting for feed propagation")

Autonomous Mode: Fallback to Issues

If any step in autonomous mode fails (permission errors, command failures, git push rejected, etc.), stop and open a GitHub issue in microsoft/FluidFramework with:

1. What was completed — PRs created, branches pushed, commands run successfully
2. What failed — the specific error and which step it occurred at
3. What remains — the exact commands for remaining steps, ready to copy-paste
4. How to finish — remind the human to invoke the fluid-release skill in an interactive GitHub Copilot session

Use the title format Release <VERSION>: <brief description of failure> and label it with release-blocking. Confirmed post-release feed propagation delays use the retry guidance in type test updates instead.

Read the appropriate reference file for the phase the user selects, then guide them through it step by step.