v0.2.0 - Onboarding + agents-as-tools UX

Highlights

• New Onboarding sub-agent — registers first-time patients with duplicate detection and input validation. Returns a patient_id the patient can immediately use for booking or coverage.
• Single-chatbot UX (agents-as-tools refactor) — the patient now talks to one ClinicOps Assistant. Sub-agents (Onboarding, Scheduler, Eligibility) are exposed to the master as delegate_to_<name> tools and invoked as stateless subroutines. No more visible routing handoff.
• Interactive REPL — clinicops with no subcommand drops into a prompt-toolkit-powered session: persistent history at ~/.clinicops_history, up/down arrow recall, ctrl+r reverse search, inline autosuggest.
• Streaming responses — agent output streams token-by-token via the OpenAI SDK streaming API.
• Trace propagation via ContextVar — master agent propagates trace_id into nested sub-agent runs so every event lands under one trace in the events store.

Added

• Onboarding agent with lookup_patient and register_patient tools
• Interactive REPL with prompt-toolkit (history, search, autosuggest)
• Streaming responses via the OpenAI SDK streaming API
• Current datetime injection into Scheduler, Onboarding, and master system prompts so agents resolve relative dates
• ContextVar-based trace propagation for nested delegate calls
• register_builtins() helper shared by CLI and eval runner

Changed

• Architecture: moved from visible triage → scheduler handoff to single user-facing assistant with sub-agents as tools
• Config: provider-agnostic LLM_API_KEY / LLM_BASE_URL / LLM_MODEL replace OPENROUTER_* (works with OpenRouter, OpenAI, Ollama, Azure, Groq, any OpenAI-compatible provider)
• REPL conversation memory via prior_messages passed to the master every turn
• Master agent system prompt explicitly suppresses routing disclosure
• CLI chat subcommand simplified — one call to the master, no manual orchestration
• All markdown docs refreshed for the new architecture

Fixed

• register_patient validates inputs before any DB I/O so bad inputs fail fast
• Dashboard Recent decisions panel no longer crashes with ValueError: cannot convert float NaN to integer on events without a latency
• REPL visual polish: httpx logs suppressed, distinct rich.Panel banner and colored you ▶ prompt

See CHANGELOG.md for the full list.

v0.1.0 - Phase 1 complete

What's in v0.1.0

Added

• Scheduler agent — books, reschedules, and cancels appointments; handles double-bookings and provider availability
• Eligibility agent — checks insurance coverage, expired plans, and prior authorization gaps
• Triage agent — classifies patient intents and routes to the right agent; handles Spanish / English-Spanish code-switching
• Agent registry with plugin discovery — drop a .py file into plugins/ to add a new workflow
• clinicops CLI: seed, chat, logs, healthcheck, dashboard, eval
• Streamlit observability dashboard with per-agent metrics and full trace drill-down
• Eval harness with 20 golden test cases (scheduling, eligibility, triage, multilingual, failure modes)
• Provider-agnostic LLM config (LLM_API_KEY, LLM_BASE_URL, LLM_MODEL) — works with OpenRouter, OpenAI, Ollama, and any OpenAI-compatible endpoint
• GitHub Actions CI: lint, format, type check, pytest on Python 3.11 and 3.12

Known Limitations

• Loom demo recording pending
• Install from source only (no PyPI release)

See CHANGELOG.md for the full list.