[Sprint 92] Autonomous cost reduction — q1b-only, weekly cadence, CEE-night doc pipeline

[peter-fusek]

Context

Commit 2c93428 landed the cost-reduction baseline for Sprint 92. This issue captures:

1. What changed in the code.
2. The handoff queries a future infra-cost-optimizer agent should run to verify and tune.
3. The one-liner resume path when e5g / sgu become clinically active again.

Only q1b (Erika Fusekova) is currently clinically active; e5g (Peter) and sgu (Nora) are onboarded but not under treatment review.

What changed (2c93428)

• PatientProfile.paused: bool added (non-destructive pause).
• e5g and sgu marked paused=True in patient_context.py.
• Scheduler gate added at scheduler.py:126 — paused patients contribute zero per-patient jobs.
• 14 agent crons rewritten to once-per-week at CEE night (02:00-03:00 UTC, stacked at 15-min offsets).
• document_pipeline_drain agent added (runs daily 02:00 UTC for q1b only).
• api_document_webhook now:
  • Rejects paused patients with 202 skipped_patient_paused (no Claude spend).
  • Out-of-window uploads enqueued into pending_docs_queue agent_state; drain agent processes nightly.
  • In-window uploads (00:00-04:00 UTC by default) dispatch immediately as before.
• AUTONOMOUS_COST_LIMIT default lowered $5 → $3.
• PAUSED_PATIENTS env var → hot-pause without redeploy.
• DOC_WEBHOOK_CEE_NIGHT_HOURS env var → override the immediate-dispatch window (default 0-4).

Tests: 850 pass, ruff clean.

Railway env vars to set on the oncoteam service

Var	Value	Why
AUTONOMOUS_COST_LIMIT	3.0	Matches the new default; lower than the $10 pressure-test value.
PAUSED_PATIENTS	e5g,sgu	Belt-and-suspenders with code; allows hot-unpause via env.
DOC_WEBHOOK_CEE_NIGHT_HOURS	0-4 (optional, default)	Covers 01:00-05:00 Bratislava year-round.

Handoff — what the infra-cost-optimizer agent should check

1. Verify the spend drop (run 7 days post-deploy)

• mcp__claude_ai_Oncofiles__search_conversations(tags=\"sys:agent-run\", limit=500) → sum cost:* tags per task:* tag. Expect total < $4 / week.
• GET https://api.oncoteam.cloud/api/autonomous/cost — expect mtd_spend on pace for ≤ $15/mo.
• mcp__claude_ai_Oncofiles__search_conversations(tags=\"sys:cost-alert\") — expect zero new alerts.

2. Check the drain queue

• mcp__claude_ai_Oncofiles__get_agent_state(\"pending_docs_queue\") after 02:30 UTC — should be empty or very small (<3 docs).
• If the queue regularly grows >20 docs overnight, split the drain into two cron slots (02:00 + 03:00 UTC) or add per-doc concurrency.

3. Scheduler job census

• GET /health → scheduler.jobs count. With q1b active only, expect exactly 20 jobs (18 per-patient for q1b + 2 system).
• Any :e5g or :sgu job IDs indicate the pause gate regressed.

4. WhatsApp reactive path

• Send a test message to +421800223213. Expect a Haiku reply within ~6s.
• Confirm no background WhatsApp broadcasts to paused-patient phone numbers.

Tuning knobs (for the future infra-cost-optimizer)

• If weekly cadence is too sparse for research on q1b, bump daily_research back to every 3 days (schedule_params={\"day\": \"*/3\", \"hour\": 2}).
• If mtd_spend exceeds $5/mo, downgrade daily_research and trial_monitor to Haiku (set model=\"light\" in agent_registry.py). Keep Sonnet only for pre_cycle_check and weekly_briefing.
• If the drain queue regularly hits >20 entries, split into document_pipeline_drain_a (02:00 UTC) / document_pipeline_drain_b (03:00 UTC).

Resume path (non-destructive)

When e5g or sgu become active again:

1. Via env (hot, no redeploy): set Railway PAUSED_PATIENTS=\"sgu\" (keep whichever stays paused) or remove entirely, then restart the service.
2. Via code: flip paused=False on the profile in src/oncoteam/patient_context.py, push, Railway redeploys.

Tokens, oncofiles data, agent_state, conversation history — all preserved. No migration.

Related

• Sprint 92 meta: Sprint 92: P0 oncopanel reset + enrollment geography + foundations for clinical user onboarding (v0.82.0 → v0.85.0) #389
• Clinical-user architecture (complementary, not blocking): Clinical funnel redesign — agent proposals lane + immutable audit log + human-in-the-loop approval gate #395, Onboard MUDr. Minaroková as oncologist for Erika (q1b) — physician role + research-review-first dashboard #396, Disaster recovery + GCP backup strategy — hourly audit log + daily oncofiles dump + restore drills + business continuity playbook #397, P0: Structured oncopanel as first-class patient profile — versioned genomic dossier + variant-level traceability + physician review #398, P0: /research as physician cockpit — 8 sub-panels + server-persisted audit + discussion + re-surfacing UI #399

[peter-fusek]

T+7d verification window: 2026-04-26 (Saturday)

Sprint 92 cost-reduction shipped 2026-04-19. The verification window is tomorrow-ish. What to check:

1. GET https://api.oncoteam.cloud/api/autonomous/cost → expect mtd_spend on pace for ≤$15/month (was tracking >$50/month pre-Sprint-92)
2. search_conversations(tags='sys:cost-alert') → expect zero new alerts since 2026-04-19
3. GET /health → scheduler.jobs should be 21 (q1b active + system jobs; paused e5g/sgu contribute zero)
4. get_agent_state('pending_docs_queue') after 02:30 UTC → should be empty/small (<3 docs)

If any of these metrics regress, re-open with findings. Otherwise close with the verification numbers as material proof.

Expected call-time: ~5 minutes. Can be automated via a single curl+mcp session.