System health shows critical issues on LLM analytics page

[mottych]

Description

Heal endpoint reports services are down although the AI endpoints work correctly.

To get additional information about the issue, use JAM MCP tools

Screenshot:

Full screen:

---

Website URL:
https://dev.admin.purposepath.app/llm/dashboard

Device and browser info:
Chrome 147.0.7727.101 (1072x761) | Windows (x86) 11

Date and time:
April 19th 2026 | 8:05pm UTC

View developer information (console logs, network requests & timing):
jam.dev/c/25453102-8b42-4e54-8c31-d76f961d8c5e

---

Captured with jam.dev

Priority

high

Attachments

• Attachment
• Attachment

[mottych]

Work on #325 (in progress)

Understanding / acceptance criteria

• Admin LLM dashboard system health must reflect whether LLM configuration/topic data is actually readable the same way coaching uses it, not a stricter control (e.g. DynamoDB DescribeTable) that can fail under least-privilege IAM while Bedrock and topic reads still work.

Approach

• coaching/src/api/routes/admin/health.py: _check_configurations_health now verifies configuration health via TopicRepository.list_all() (optional prefetched_topics to avoid duplicate reads). Removed describe_table on the topics table. Run configurations / templates / models checks with asyncio.gather.
• coaching/src/api/routes/admin/topics.py: /stats passes prefetched_topics=all_topics into the configurations check after the existing list load.

Testing

• Extended coaching/tests/test_admin_endpoints.py with test_configurations_health_prefetch_skips_topic_repo_list.
• Ran: uv run pytest tests/test_admin_endpoints.py -v, ruff, mypy on touched route files.

Feature branch: feature/issue-325-admin-health-configurations-check

[mottych]

Completed (#325)

Cause: The admin health "configurations" probe used dynamodb:DescribeTable on the topics table. That permission is often omitted in least-privilege policies even when the Lambda can still read topic items via the same repository path used for coaching—so the dashboard showed configurations (and overall health) as critical/down while AI calls still worked.

Change (merged to dev):

• Configurations health now succeeds when TopicRepository.list_all() succeeds (same access path as runtime).
• GET /admin/topics/stats reuses the topics list already loaded for metrics via prefetched_topics, avoiding an extra round trip.
• Service probes (configurations, templates, models) run in parallel with asyncio.gather.
• Test added: test_configurations_health_prefetch_skips_topic_repo_list.

If anything still looks wrong after deploy, capture the JSON from GET .../admin/health and we can tune templates/models probes next.