Never store a credential: redact at every DB write and at boot; keep PDFs when Gemini refuses for budget reasons

Candidate queue: a bulk discovery pass fills agent_candidates, cycles drain it; UDSpace + CORE lanes

discover: one closed-source warn per cycle per lane, shared by the phrase search and the topic walk

Merge the 17 Sep "more papers" lanes onto the hotfix line; parallel ingestion; throughput knobs

Rotate the query frontier on dry cycles; adopt and re-queue never-ingested downloads

crawler: per-host circuit breaker on budget refusals; OpenAlex account key

discovery: NASA NTRS lane (search_ntrs), on by default

agent_runs instrumentation, stall detector, figure linking in the cycle

Review Queue: read the real columns; report an unreadable queue; log failed boot migrations

More papers: budget-aware sources, frontier rotation, schema gate, topic walk, S2 bulk, datasheets, query grid

Keep the agent running: cycle watchdog, Retry-After cap, boot at container start, self keep-alive

Figure & graph mining in the autonomous cycle: figures.py + Postgres port, crops stored in DB, review UI

agent: harvest until a per-cycle paper target, not the first hit

agent: harvest in rounds until a paper is found

fix: run the URL-blacklist reset at boot; log why downloads were skipped

fix: persist crawler retry counts; one-shot reset of poisoned URL blacklist

ui: theme tokens follow the configured theme, not the viewer's OS scheme

ui: Dashboard — health at a glance (header pills, KPI tiles, per-cycle charts, activity feed)

ui: design system + branded shell (agent/ui_theme.py, light theme, brand row, footer)

Fix false 'cycle running' state: boot force-sweeps dead runs and clears leftover lock; UI treats only a fresh lock as running

Anti-hang hardening: TCP keepalives + statement timeout on PG connections, cap Retry-After sleeps, sweep stale 'running' runs at boot and before each cycle

Initial deploy: autonomous ingestion agent (plan→discover→dedupe→download→ingest→report)