/finding-deleted-feature-flags

name: finding-deleted-feature-flags
description: 'Find feature flags that were soft-deleted in the active project within a recent time window. Use when the user asks "what flags were deleted in the last N days", "show me recently deleted feature flags", "who deleted flag X", "audit recent flag deletions", or anything similar. Handles the non-obvious gotcha that system.feature_flags exposes the deleted boolean but does not expose a deletion timestamp — the actual deleted-at time lives in the per-flag activity log and must be cross-referenced.'

Finding recently deleted feature flags

This skill produces a list of feature flags that were soft-deleted in the active project within a user-specified time window, along with who deleted each one and when.

When to use this skill

• The user asks "what flags got deleted last week / in the last N days?"
• The user wants an audit of recent flag deletions (who, when, what was removed)
• The user wants to find when a specific flag was deleted, or by whom
• Any "recently deleted feature flags" framing

Don't use this for **active** stale-flag cleanup — that's `cleaning-up-stale-feature-flags`. This skill is for flags that have already been removed.

The gotcha that makes this non-trivial

`system.feature_flags` exposes `deleted` as a boolean but does **not** expose `deleted_at`, `updated_at`, or `last_modified_at`. There's no way to filter soft-deleted flags by deletion time in a single SQL query — trying to use those columns will return `Unable to resolve field`.

The actual deletion timestamp lives in the per-flag activity log, reachable only via `posthog:feature-flags-activity-retrieve` (one call per flag id). There is no bulk activity endpoint.

So the workflow is two-stage: SQL to enumerate candidates, then parallel activity-log lookups to find each deletion event.

Workflow

1. Clarify the window if ambiguous

"Last week" is ambiguous — it can mean rolling 7 days from now, or the previous calendar week (Mon–Sun). If the user wasn't explicit, ask, or surface both interpretations in the final report.

Always compute the cutoff in UTC and keep the user's local interpretation in your head separately.

2. Enumerate soft-deleted flags via SQL

Query `system.feature_flags` for `deleted = true` in the active project, ordered by `created_at DESC`:

SELECT id, key, created_at
FROM system.feature_flags
WHERE team_id = <team_id> AND deleted = true
ORDER BY created_at DESC
LIMIT 100

Order by `created_at DESC` because deletions empirically cluster near creation — most flags get deleted within a few days of being created — so walking the most-recently-created candidates first finds recent deletions fastest. **But** this is a heuristic, not a guarantee: an older flag deleted recently won't be at the top of this list. Be explicit about that limitation when you report.

`team_id` defaults to the active project, but include it explicitly for clarity.

3. Fan out activity-log lookups in parallel

For each candidate id, call `posthog:feature-flags-activity-retrieve` with `limit: 5, page: 1`. **Issue all calls in one message so they run concurrently** — sequential calls are dramatically slower.

call feature-flags-activity-retrieve {"id": <flag_id>, "limit": 5, "page": 1}

Reasonable batch sizes:

• "last 7 days" → top 20–25 candidates
• "last 30 days" → top 50
• "last 90 days" → walk the full ~100

If you sample fewer than the full set, say so in the report and offer to walk the rest as a follow-up.

4. Extract the deletion event from each response

In each response, find the entry where `activity == "deleted"`. That entry's `created_at` is the actual deletion time, and `user.email` / `user.first_name` identify the deleter. These fields are reliable on every delete path.

For most flags there's exactly one delete event. If a flag has been deleted-and-restored multiple times, take the most recent `activity: deleted` event within the window.

5. Recover the original key and report

Feature flags are renamed to `<original>:deleted:<flag_id>` when soft-deleted while still referenced elsewhere (e.g. a stopped experiment) — the id-based suffix frees the original key for reuse. Don't try to recover the original from the activity log's own fields: `detail.changes` only carries the rename on UI/ORM deletes (and is often empty, or missing the `key` entry, on API/MCP/programmatic deletes), and `detail.name` just mirrors whatever the current key is — tombstoned or not.

Instead, strip the suffix deterministically with [`scripts/strip_deleted_suffix.py`](./scripts/strip_deleted_suffix.py). Pass it the whole step 2 candidate list as JSON in one call — not one invocation per flag:

echo '[{"id": 687432, "key": "high_frequency_alerts:deleted:687432"}]' | python3 scripts/strip_deleted_suffix.py
# prints the same array back (pretty-printed), each object gaining an "original_key" field:
# "original_key": "high_frequency_alerts"

Filter the collected deletion events to those whose `created_at` falls inside the requested window. Present as a table, using each row's recovered original key (not the raw tombstoned form) for the "Key" column:

| Flag ID | Key | Deleted at (UTC) | Deleted by |

State your methodology in the report (how many candidates you walked vs. how many soft-deleted flags exist total), so the user knows what was and wasn't checked.

Watch-outs

• **Borderline cases**: if a deletion is within ~1 hour of the window cutoff, surface it as borderline rather than silently dropping it.
• **Don't trust `created_at` as a proxy for deletion time**: a flag created in 2024 can still have been deleted last week. The activity log is the only authority.
• **Renamed keys are normal**: a flag with key `foo:deleted:12345` was the flag originally keyed `foo` — see step 5 for how to recover it.
• **Walking all candidates is possible but slow**: ~100 parallel activity-log calls is doable. Offer it as a follow-up r