- run_agent_streaming() calls POST /complete/stream and yields token strings
- Reads NDJSON lines from the streaming response, yields on 'chunk' events
- On 'error' or connection failure, yields the fallback response string
- Tool calls are not supported in the streaming path
- Existing run_agent() (non-streaming, tool-call loop) is unchanged
- Add ChannelType.WEB = 'web' to shared/models/message.py
- Add webchat_response_key() to shared/redis_keys.py
- Create WebConversation and WebConversationMessage ORM models (SQLAlchemy 2.0)
- Create migration 008_web_chat.py with RLS, indexes, and channel_type CHECK update
- Pop conversation_id/portal_user_id extras in handle_message before model_validate
- Add web case to _build_response_extras and _send_response (Redis pub-sub publish)
- Import webchat_response_key in orchestrator/tasks.py
- Write 19 unit tests covering CHAT-01 through CHAT-05 (all pass)
- Move key imports to module level in tasks.py for testability and clarity
- Pop WhatsApp extras (phone_number_id, bot_token) in handle_message before model_validate
- Build unified extras dict and extract wa_id from sender.user_id
- Change _process_message signature to accept extras dict
- Add _build_response_extras() helper for channel-aware extras assembly
- Replace all _update_slack_placeholder calls in _process_message with _send_response()
- Add escalation pre-check: skip LLM when Redis escalation_status_key == 'escalated'
- Add escalation post-check: check_escalation_rules after run_agent; call escalate_to_human
when rule matches and agent.escalation_assignee is set
- Add _build_conversation_metadata() helper (billing keyword v1 detection)
- Add channel parameter to build_system_prompt(), build_messages_with_memory(),
build_messages_with_media() for WhatsApp tier-2 business-function scoping
- WhatsApp scoping appends 'You only handle: {topics}' when tool_assignments non-empty
- Pass msg.channel to build_messages_with_memory() in _process_message
- All 26 new tests pass; all existing escalation/WhatsApp tests pass (no regressions)
- Add supports_vision(model_name) to builder.py — detects vision-capable models
(claude-3*, gpt-4o*, gpt-4-vision*, gemini-pro-vision*, gemini-1.5*, gemini-2*)
with provider prefix stripping support
- Add generate_presigned_url(storage_key, expiry=3600) to builder.py — generates
1-hour MinIO presigned URLs via boto3 S3 client
- Add build_messages_with_media() to builder.py — extends build_messages_with_memory()
with media injection: IMAGE -> image_url blocks for vision models / text fallback for
non-vision models, DOCUMENT -> text reference with presigned URL
- image_url blocks use 'detail: auto' per OpenAI/LiteLLM multipart format
- Add 27 unit tests in test_multimodal_messages.py (TDD)
- Add escalation pre-check in _process_message: assistant mode for escalated threads
- Add escalation post-check after LLM response: calls escalate_to_human on rule match
- Load Slack bot token unconditionally (needed for escalation DM, not just placeholders)
- Add keyword-based conversation metadata detector (billing keywords, attempt counter)
- Add no-op audit logger stub (replaced by real AuditLogger from Plan 02 when available)
- Add escalation_assignee and natural_language_escalation fields to Agent model
- Add Alembic migration 003 for new Agent columns
- AuditEvent ORM model with tenant_id, action_type, latency_ms, metadata
- KnowledgeBaseDocument and KBChunk ORM models for vector KB
- Migration 003: audit_events (immutable via REVOKE), kb_documents, kb_chunks
with HNSW index and RLS on all tables
- AuditLogger with log_llm_call, log_tool_call, log_escalation methods
- audit_events immutability enforced at DB level (UPDATE/DELETE rejected)
- [Rule 1 - Bug] Fixed CAST(:metadata AS jsonb) for asyncpg compatibility
