Production activity feed has not updated since April 22, 2026

## Problem

The production API is reachable, but the activity feed appears stale. As of September 5, 2026, the newest record returned by:

```text
GET https://www.buildcanada.com/tracker/api/v1/feed.json?per_page=20
```

has `occurred_at: 2026-04-22T15:36:59Z`. The feed reports 11,121 total records, but none are newer than April 22.

The MCP backend is also operational through the existing wildcard proxy at:

```text
POST https://www.buildcanada.com/tracker/api/v1/mcp
```

I verified `initialize`, `tools/list`, and `tools/call` with `list_activity`; MCP returns the same April 22 records. This suggests the issue is in data ingestion/evaluation rather than the public API or MCP implementation itself.

## Expected

The configured `agent_weekly_scan` GoodJob cron runs every Sunday and creates current evaluation/activity records.

## Things to inspect

- GoodJob cron and worker history for `AgentWeeklyScanJob` and `AgentEvaluateCommitmentJob`
- failures from the Claude CLI subprocess, including OAuth/token errors
- worker connectivity to the Rails agent API
- production environment values for `RAILS_API_URL` and `AGENT_API_KEY`

One configuration worth checking: `AgentEvaluateCommitmentJob` defaults `RAILS_API_URL` to `http://localhost:3000`, while `docker-compose.prod.yml` runs GoodJob in a separate `worker` container and does not currently pass `RAILS_API_URL` into it. In that topology, localhost points to the worker container rather than the `web` service. This may not match the actual deployment and is therefore only a diagnostic lead, not a confirmed root cause.

The separate missing canonical MCP proxy route is being handled in BuildCanada/OutcomeTracker#169.