With OPENROUTER_API_KEY in a developer's .env the chat agent switched on in dialog tests (real API calls,
spent money, flaky results). An autouse fixture turns the agent off; agent tests pass explicit providers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Live-tested on OpenRouter: Flash-Lite answers in 3–8 s with correct findings at ~$0.10/$0.40 per 1M tokens
(qwen3.7-flash and gpt-oss-120b took 26–32 s). Force a tool call on the first turn, give the model the
current prescription, let check_drug_for_patient default to it and return composition, intolerance check
and general interactions (alcohol); retry once on an empty final message.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Free-text treatment and medication questions go to an LLM agent with read-only tools bound to the current
patient: guideline search (RAG, doses stripped), ЕМИАС record mock (diagnoses, comorbidities, intolerances,
other specialists' drugs, current prescription), drug info (composition, contraindications, interactions)
and a deterministic patient-specific check. Providers: OpenRouter (temporary, synthetic data only) and an
offline mock; GigaChat to follow. Scrubbed input only, forbidden-output filter, any failure falls back to
the existing extractive RAG / instruction answers. Off unless OPENROUTER_API_KEY or CHAT_AGENT=mock.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Patient list sorted red -> yellow -> grey -> green with KPI row, trend arrows,
7-day adherence bars, discontinuation-risk badge with factor tooltip.
- Patient card: scales with severity bands from config/scales (unscheduled marked),
mood/sleep/alcohol small multiples, dose heat-strip, alerts feed with ack,
safety plan, AI summary with disclaimer.
- Live alerts: st.fragment(run_every=2.5s) polling, toast + pinned acute banner on
every page; full rerun on new alert only on doctor pages.
- Connect form -> POST /patients/connect -> code, QR, deep link.
- Thin httpx ApiClient over the frozen schemas with graceful "API недоступен".
- app.py discovers dashboard/pages/*.py automatically (st.navigation); demo pages
of #17 plug in by adding a file.
Refs #16
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- personas degrade: /demo/personas → doctor's patients (hero first) → env token
- scenario file guessed from the transliterated first name (andrey.jsonl), like the backend
- «Подключить демо-пациента»: connects the synthetic hero via /patients/connect when no seed
- connect code prefilled from ?code= of the console link
Refs #17
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- scripts/seed_demo.py: doctor doc-smirnova (on call), Andrey day 0 (connect code,
no consent) or day N with history; 7 patients with seeded-random histories.
Alerts and colours come from the real rules evaluated day by day.
- --reset drops/recreates both DBs and zeroes the demo clock; seed keeps the
persisted clock offset so dates match the running app.
- data/synthetic/andrey.jsonl: console scenario for days 0..13 generated from
config button labels (--write-scenario).
- tests: idempotency, modes, colours, scenario format/sync and a full replay of
the scenario through the real engine, scheduler and monitoring.
Refs #18
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- pages/2_Демо-консоль.py: persona selector, virtual clock bar (+1 h, to 08:00/20:00,
next day, reset with confirm), fallback patient web chat (buttons from Reply.buttons,
kind styling, outbox drained after every action and on a 2 s fragment poll), scenario panel
- pages/3_Сплит-экран.py: patient chat or PATIENT_WEB_URL iframe on the left, compact doctor
panel (patients, live alerts, red banner on a new acute alert) on the right
- console/playback.py: data/synthetic/<persona>.jsonl runner (advance_to/press/text/
expect_color/start), date chaining between days, pause on failure with retry/skip/manual
- console/api.py: own HTTP client over the frozen API contract, ApiError only
Refs #17
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Doctor list/card/alerts/ack/connect, dialog events/outbox/history and demo
controls now read the identity and clinical DBs. Rule-based discontinuation
risk index, trend, adherence; every doctor read is audited.
Refs #12
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- menu button «Вопрос о лечении» -> kb.question -> rag.answer (KR quote + source)
or «уточните у врача»; dose questions get a fixed doctor redirect
- free text in idle that looks like a question (no check-in signals, not a drug
question) -> knowledge base; no answer -> the usual support phrase
- routing order unchanged: consent/safety first, side effects/quit/alcohol -> check-in,
own drug -> med_qa (instruction only, SPEC §7 p.5)
- question cues, small talk and texts in config/texts/knowledge.yaml; scrub placeholders
before "что такое X" no longer hide the definition question
Refs #64
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- converter marks a paragraph glued across a page break with ⟦стр. N⟧; the chunker
strips the marker and extends the chunk's page range (wrong-page citations 24 -> 3,
the rest are scale-form table rows split by a page)
- [NAME]/[PHONE] placeholders from privacy.scrub() no longer count as unknown words
(a scrubbed question used to fall below the coverage gate)
- a question naming one disease boosts its guideline in retrieval and passage choice
("ТМС при депрессии" was answered from the PTSD guideline); "антидепрессант" is not a cue
Refs #62
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- app/scheduler/demo.py: run_until(now) replays due and fresh events in time order: dose rows +
med_reminder, missed after 2 h without an answer → monitoring, daily_checkin, scale_due by the
config/modules.yaml schedules, daily monitoring pass (silence, missed doses) → reach_out/scale_due
- idempotent via scheduler_events (unique token+key); stale events are not sent; no contact without
consent or during an open crisis
- config/scheduler.yaml: times and windows; runs while the questionnaire release is off
- POST /demo/clock/advance replays the scheduler up to the new time
Refs #13
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- app/services/snapshot.py: PatientSnapshot from the clinical DB (field mapping, age, therapy day,
recent C-SSRS risk, weapon access from dialog facts)
- app/services/monitoring.py: RuleMonitoring registered via register_monitoring; persists alert/crisis
actions, crisis_acute audit, enhanced_monitoring flag, cached status_color, notifies the doctor (and
the doctor on duty for crises) after commit; a running crisis is not restarted
- demo clock offset persisted in clock_state, restored on scheduler start
- tests: Andrey's 13 days through the service → G G G Y Y G G G G Y Y Y R
Refs #38
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- No same-day C-SSRS repeat: a risk screen completed within cssrs_cooldown_hours is not queued again
after PHQ-9 item 9 >= 1 (scale crisis path, rule schedule_scale, scale_due); crisis script and alerts stay.
- Scale intro registry: register_scale_intro(scale_id, text_or_fn); start_scale(intro=None) uses it.
veteran_daily migrated from per-scale on_enter hooks.
- Prefix triggers "button:<prefix>:*" with trigger_payload(event); medication buttons carry the dose
(med:taken:<dose_id>, ...), med_pending dropped; payload-less legacy buttons answer as stale.
- Abandoned-step timeout: settings.abandon_after_hours (default 6 h) returns abandonable states
(daily.*, meds.*, scale.* not in onboarding) to idle before the next event; never in crisis or with intake on.
Refs #56
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Start screening of the veteran module: peer-tone intro, trauma exposure gate for
PC-PTSD-5, PCL-5/PHQ-9/AUDIT-C via scale sessions (sex asked if unknown), weapon
access question stored as facts for lane B with a safe-storage block on risk,
plain-language result, modules enabled in profile, FAQ, resume after a crisis.
Refs #36
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Deep-link /start <code> binds the chat via IdentityProvider, text and
button presses go to get_engine().handle(), replies render as inline
keyboards and are sent verbatim (no parse_mode). TelegramOutbox delivers
scheduler pushes to bound chats and falls back to the previous outbox.
No-op without TG_BOT_TOKEN; logs carry no text or chat ids.
Refs #11
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DfbiLXFJswz2EK3qtWF7S7
build_structured() aggregates a PatientCard into scores, flags and
adherence (no patient text, no names); build_summary() sends it through
the LLM gateway, falls back to a config-driven template, keeps an open
acute alert first and caches per (token, provider, data hash).
Refs #19
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Topic from engine or keyword classifier, section via integrations.drug_info,
sentence retrieval by keyword overlap (dosing sentences dropped), LLM only
rephrases the excerpt through gateway.phrase("med_answer") with forbidden-output
and added-content guards; dose questions and missing answers -> None + fallback text.
Refs #15
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Freeze interfaces between lanes A (bot), B (backend), C (dashboard) so three developers work in
parallel without touching each other's files: dialog Event/Reply + engine registry with echo stub,
integration protocols with in-memory defaults, MonitoringService with noop stub, /api/v1 schemas
with fixture-backed routes, frozen main.py wiring, config/modules.yaml + due_scales(), DAR-5,
tzdata, Andrey trajectory, lanes in WORKFLOW, ADR-010.
Refs #34
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Single entry point app/llm/gateway.py: extract_item, extract_checkin,
risk_classify, phrase, summarize_for_doctor. Two-level design:
LLMProvider (transport, GigaChat) and LLMTasks (PromptedTasks over a
provider, or rule-based MockTasks). Shared policy in the gateway:
assert_scrubbed backstop, confidence threshold 0.5, risk failure -> NONE,
forbidden-output filter with config/texts/phrases.yaml fallbacks.
Refs #7
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rule-based PII scrubber: phones, e-mails, SNILS (checksum-aware), passport,
INN, cards, dates of birth, addresses, URLs/handles and person names
(dictionary in config/privacy/names_ru.yaml + declension, patronymic and
surname heuristics). Idempotent via fixed-point iteration; false-positive
guards for ordinary clinical text. pseudo_id normalises SNILS to 11 digits
and returns HMAC-SHA256 truncated to 32 hex chars.
Refs #2
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two DeclarativeBases on two engines (ADR-003): identity holds direct
identifiers, clinical holds token-keyed pseudonymised data plus audit_log.
Portable types (JSONB variant, tz-aware UTC timestamps rejecting naive
values) so tests run on SQLite. init_db() runs create_all at startup.
Guardrail #5 checked by AST and import-graph tests.
Refs #6
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pydantic schema for config/scales/*.yaml, cached loader, score() with
severity bands (sex-specific for AUDIT-C), rule-based C-SSRS triage,
flags (PHQ-9 item 9, PCL-5 >= 33, PHQ-4 subscales) and delta().
Starting literature thresholds for phq9, gad7, pcl5, auditc, pcptsd5,
cssrs_screen, phq4; final values are approved by the psychiatrist (#8).
Refs #3
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One Postgres with isolated identity/clinical databases and roles (ADR-003),
FastAPI /health, Streamlit placeholder, typed settings with offline defaults.
Refs #1
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ElebobaandClaude Opus 5.5