Flag Release

Record a Flag’s Automated Release

You’re using a skill that takes an existing feature flag (created OFF) that guards a pull request’s change, and records an automated rollout so the change releases safely once the PR merges — no human toggling a flag. It defers per-environment to the team’s release policies and honors any human-stated release intent.

This is the release step of the PR flag workflow, and it’s deliberately atomic:

By the time this skill runs, the flag exists (OFF) and the guarding code is wired and pushed. This skill only records the rollout — it never creates flags or edits code. It can be driven directly by an automation harness, or as the final step of the flag-and-release-change orchestrator.

The deploy is not the release. The merge ships the control path (flag OFF); the release is the flag operation this rollout performs afterward, governed by the environment’s policy.

Honoring a hold is the one thing you must get right. There is no “hold” release type, and policy is NOT a manual gate. On merge, policy automatically performs the environment’s release (immediate, progressive, or guarded) with no human promotion step — a guarded rollout still starts on its own the moment the PR merges. So recording a held environment as policy does not hold it; it releases it on merge, before any notBefore date. The only way to hold an environment is to omit it from the environments array entirely, which leaves the flag OFF there. If the user wants an environment held (or not released until a date), exclude it from the call and report it as held. When in doubt, omit.

Prerequisites

• The remotely hosted LaunchDarkly MCP server.
• The guarding flag already exists in LaunchDarkly, created OFF, with the agreed key/tags.
• A pull-request reference (repoFullName + prNumber, or prUrl) so the rollout binds to the right merge.

MCP tools this skill uses:

• create-automated-rollout-config — record the rollout for the flag against the PR (the deliverable)
• match-release-policies — resolve which release policy governs each environment (call before proposing the plan)
• list-release-policies — see the project’s release policies and the metrics they auto-attach
• get-flag — confirm the flag exists and is OFF before recording

Full release model — simple vs policy, precedence, previewing, prerequisites, metric adequacy: references/auto-release.md (opens in a new tab).

Plan Phase

Record nothing in this phase.

1. Confirm the flag. get-flag to verify the guarding flag exists and is OFF. If it doesn’t exist yet, stop — creation is launchdarkly-flag-create‘s job, and recording a rollout for a missing flag fails confusingly.
2. Pick the target environments. Use the environments named by the user or harness. Don’t hardcode a set — a given change can’t always release to every environment. If none are named, enumerate the project’s real keys and confirm the set rather than assuming.
3. Preview each environment’s policy. Call match-release-policies (by flagKey + environmentKey) to resolve, deterministically, what a policy release will do per environment — winningReleaseMethod (immediate / progressive / guarded / none). Don’t reason about policy scope by hand. For a guarded winner, check the auto-attached metrics can actually compare this change (see the metric-adequacy note in references/auto-release.md (opens in a new tab)).
4. Capture the human’s release intent. Ask (briefly, only if not already stated): release on merge, hold (recorded but not released yet), or wait until a notBefore date? A cohort/segment to target first? A prerequisite parent flag this must not precede? Intent sits above the policy in precedence and is honored or explicitly held — never silently dropped.
5. Present the per-environment plan and stop. For each environment, state either the releaseType it will be recorded with (simple / policy, and what that does on merge) or that it will be held — omitted from the recorded config so the flag stays OFF there — with the reason. Wait for confirmation; revise on feedback.

Implement Phase

Only after confirmation:

1. Sort every target environment into exactly one bucket — RELEASE or HOLD. Do this before you build the call. There is no third bucket, and releaseType does not create one.

   Bucket	Meaning	What goes in the call
   RELEASE	Goes live on merge — now, or per its policy	Add { environmentKey, releaseType } to the environments array
   HOLD	Not yet — waiting on a date, sign-off, segment, or parent flag	Nothing. Leave it out of environments entirely; name it as held in your report

   releaseType (simple vs policy) only chooses how a RELEASE environment goes live — it never holds one. Both release on merge: simple serves true immediately; policy runs that environment’s policy (immediate / progressive / guarded) automatically, with no human promotion step. “policy defers to the policy“ — not to you, and not until a date. Do not reach for policy to park a held environment: it ships that environment on merge, before any notBefore. The only encoding of a hold is absence from the array.

2. Build environments from the RELEASE bucket only — then read the keys back. The array must contain every RELEASE environment and no HOLD environment. Before you send the call, scan the environmentKeys in the array: if any environment you’re holding appears there, delete that entry. The call has no field for a date or a hold — if you catch yourself wanting to add holdUntil, notBefore, or hold to an entry (or to keep it as policy “so it waits”), that is the signal the environment is HOLD: drop the entry, don’t invent a field. The date and reason go in your report, not the call.

   Worked example — “release staging on merge, hold production until 2026-09-01”: staging is RELEASE, production is HOLD.

   jsonCopy

Edge Cases

What NOT to Do

• Don’t create the flag or edit code — that’s launchdarkly-flag-create. This skill only records the release.
• Don’t turn the flag on yourself, or toggle it after recording the config. The rollout owns that; double-toggling causes audit noise and confuses the scheduler.
• Don’t skip match-release-policies. Proposing policy without knowing what it resolves to is guessing.
• Don’t silently release against a stated hold/notBefore. Honor intent or hold — never drop it.
• Don’t handle or print credentials. Access is injected by the environment.

References

• references/auto-release.md (opens in a new tab): the automated-rollout / release-policy model, simple vs policy, precedence (intent → override → policy → default), previewing, prerequisites, metric adequacy. (Core of this skill.)
• launchdarkly-guarded-rollout: for a bespoke rollout no policy expresses — set that env to simple here and drive the guarded rollout by hand after merge.