Amplitude Listener Events Adoption

Amplitude Listener Events Adoption

Produce a markdown adoption report for the last 15 domain events subscribed in the Amplitude listener, grouped by domain, listing organizations that triggered each event and how many times.

Prerequisites

• The Amplitude MCP must be connected (tools prefixed mcp__amplitude__). If not, prompt the user to run /mcp.
• Target project: Packmind V3 [CLOUD], appId = 100032310. Never query the DEV or legacy projects unless the user explicitly asks.

Inputs

Organizations exclusion

Always exclude these internal/test organizations from the report (apply after querying Amplitude, before aggregation):

• AI Configuration orga
• test_package_standards
• packmind
• Joan's orga
• demo-orga
• VIncent test
• cedric.teyton's organization
• cedric-test
• test-skills

Match on the exact organization grp:name value returned by Amplitude. Extend this list if the user provides additional orgs to skip.

Events exclusion

Always exclude these Amplitude event names from the report even if they are among the last N subscribed in the listener:

• user_signed_in
• new_organization_created
• user_signed_up

Rationale: these three events fire once per user/organization creation and produce extreme long tails (hundreds of orgs with a single event each), drowning out the actual feature-adoption signal the report is built for.

Drop excluded events from the event-name list before building the Amplitude query (do not count them toward N — select the next most recent event to fill the quota). Extend this list if the user provides additional events to skip.

Workflow

Step 1 — Identify the last N still-subscribed events

Source file: packages/amplitude/src/application/AmplitudeEventListener.ts.

1. Read the current file and collect the set of event class names that appear in this.subscribe(<Event>, ...) inside registerHandlers(). This is the "still-subscribed" set.

2. For each class, also extract the corresponding Amplitude event name (the string literal passed as the 2nd argument to this.emitAmplitudeEvent(event, '<name>', ...) in the handler). This is what Amplitude stores.

3. Walk the file's git history newest→oldest:

   git log --all --follow --format="%h|%ai|%s" -- packages/amplitude/src/application/AmplitudeEventListener.ts

4. For each commit, extract added this.subscribe( lines:

   git show --format="" <commit> -- packages/amplitude/src/application/AmplitudeEventListener.ts \
     | grep -E "^\+ *this\.subscribe\("

5. Walk commits in order; for every added class name that is also in the still-subscribed set, record (amplitude_event_name, class_name, commit_date, commit_sha) once (deduplicate by class — only keep the most recent addition). Stop once N unique events are collected.

6. Discard entries for events that were later removed or renamed out of the current file (e.g. SpaceRenamedEvent → filtered because no longer in the current registerHandlers).

Multi-line this.subscribe(\n EventName,\n ...) must also be captured — widen the grep context (grep -A2) when a line ends just after this.subscribe(.

Efficient commit walk: rather than running git show once per commit interactively, batch the first ~40 commits returned by git log and loop in a single bash call:

for c in $(git log --all --follow --format="%h" -- packages/amplitude/src/application/AmplitudeEventListener.ts | head -40); do
  echo "=== $c ==="
  git show --format="" $c -- packages/amplitude/src/application/AmplitudeEventListener.ts \
    | grep -E "^\+ *this\.subscribe\(" -A1
done

This surfaces all added this.subscribe( lines with commit boundaries in one tool call, which keeps token cost low and preserves the newest→oldest ordering needed for dedupe.

Step 2 — Query Amplitude

Use mcp__amplitude__query_dataset with:

• projectId: "100032310"
• definition:

  {
    "type": "eventsSegmentation",
    "app": "100032310",
    "params": {
      "range": "Last 30 Days",
      "events": [ /* one entry per amplitude event name */
        { "event_type": "<name>", "filters": [], "group_by": [] }
      ],
      "metric": "totals",
      "countGroup": "organization",
      "groupBy": [{ "type": "group", "group_type": "organization", "value": "grp:name" }],
      "interval": 30,
      "segments": [{ "conditions": [] }]
    }
  }

• groupByLimit: 200
• timeSeriesLimit: 0

Critical gotcha: the CSV returned by query_dataset omits the organization name column when groupBy is set on a group property. Re-query using mcp__amplitude__query_chart with the chartEditId returned by query_dataset; that response includes the real org-name column. Always do this second call to get usable data.

If the response payload is large, rely on timeSeriesLimit: 0 (totals-only) to keep it compact.