launchdarkly

Use the official ldcli only for LaunchDarkly operations. Read CLI preflight; reuse authentication without printing tokens or config. Audit requests are read-only: "can we kill it?" asks for a recommendation, not a targeting change, deprecation, archive, deletion, or code removal.

Scope

Resolve the project from the request or repository configuration. Otherwise list projects with ldcli projects list --help and ask which project; never assume default. Paginate before claiming an account-wide count.

Use Integration (integration) and Production (production) only. Discover these exact keys with ldcli environments list; a missing/inaccessible required environment is unknown, not empty or disabled. Ignore testing, staging, preproduction, development, and other unused environments in the cleanup heuristic. If a project uses different keys, resolve that mapping before adapting the command; do not silently substitute another environment.

Inspect a flag

Read ldcli flags get --help, then fetch full configuration for each selected environment. Report existence, creation date, archive/deprecation state, on/off, and the values actually served. Distinguish missing flags from authentication, permission, project, and environment failures. Missing does not authorize creation; follow CLI changes only when requested.

For "who can see it?", trace default variation, individual/context targets, ordered rules, segments, percentage/experiment rollouts, and prerequisites. on: true alone does not mean everyone receives the feature. Resolve indexes against variations[].value; never assume index zero means enabled. Describe audiences without exposing context identifiers unnecessarily. Separate feature recipients from dashboard permissions and maintainers. For multivariate flags, establish which value means enabled; otherwise report unknown.

For detailed dependency, code-reference, or archive analysis, read lifecycle and evidence.

Count long-enabled cleanup candidates

With Bun available, run from this skill's installation directory:

bun run scripts/audit.ts --project <project-key>

The command paginates inventories, fetches full targeting with --summary 0, and checks scoped change history through ldcli. It never writes to LaunchDarkly. Keep raw output local and out of commits or public PRs.

Count each non-archived flag once when all hold at UTC auditedAt:

• Created strictly more than 30 × 24 hours ago.
• Targeting is on in both Integration and Production, and all possible default, rule, target, and positive-weight rollout outcomes serve boolean true.
• No selected configuration or flag history change in the last 14 × 24 hours. A change exactly at the cutoff excludes the flag. Use selected-environment resource specifiers, not all-environment history that disqualifies unused edits. Shared metadata changes affecting selected environments also count as changes.

Unresolved prerequisites, non-boolean enabled values, malformed/missing targeting, and inaccessible history are unknown, not verified candidates. Resolve them before claiming an exact total. Matching means candidate for removing the flag check while preserving the released feature, not "turn the feature off" or "safe to archive now".

Return evidence

Lead with candidateCount, project, both environment keys, UTC audit time, and scannedCount. List candidate keys with createdAt and lastTargetingChangeAt; the latter is not a global last-change timestamp. Name ignored environments and unknowns. exactCount: null / exit 2 means a lower bound; exit 1 means inventory failed and no count is proven. Aggregate only complete project reports; restricted token access cannot prove account-wide scope.

State archive blockers separately. Snapshot reads are not atomic; rerun before acting. Stop after the report unless the user explicitly requests a change.