We couldn’t see the signal through the noise until we surfaced the GlitchTip backlog on the System Health dashboard. The probe ran every five minutes, wrote summaries to audit_log, but the operator had zero visibility because the pager gated on a count of 100 and the actual threshold was much lower. PR #548 (PR #548) lowered glitchtip_triage_alert_threshold_count from 100 to 10 and added four stat panels querying Postgres directly via local-brain-db.
With visibility restored, we tackled the rot. We confirmed by grep that three backward-compat shims had zero callers and deleted them in PR #546 (PR #546), including services/redis_cache.setup_redis_cache() and the deprecated notification parameter in notify_gate_pending. The cleanup extended to the admin module that was uninstalled in pyproject.toml but still present in src/cofounder_agent/admin.py, which we removed along with the failing setup_admin call in PR #544 (PR #544). We also purged references to plugins.stage_runner and model_router in PR #545 (PR #545) to stop misleading readers and LLMs.
CodeQL kept coming back. A false positive from the previous </script> pattern was flagged again (#152), so we tightened the end-tag match to </tagname\b[^>]*> for both <script> and <iframe> in PR #549 (PR #549). We also fixed a ReDoS risk in citation_verifier.py and incomplete URL substring sanitization in url_scraper.py in PR #547 (PR #547).
The test suite was still using a helper that shouldn’t exist. We retired _mock_model_converter in PR #534 (PR #534), driving the real ModelConverter pipeline for every test to prevent silent drift. Finally, we synced the docs: updated PluginScheduler counts and fixed the deploy_content_flow path in PR #533 (PR #533).
The system is cleaner and louder now.
Auto-compiled by Poindexter from today’s commits and PRs. See the work: github.com/Glad-Labs/poindexter.
Sources
- (PR #548)
- (PR #546)
- (PR #544)
- (PR #545)
- (PR #549)
- (PR #547)
- (PR #534)
- (PR #533)
