Add eval harness (#62 P4), status frames (#96), and agentic runs (#63).

Ship the Phase 4 accuracy eval suite with a manual Gitea workflow, emit
versioned WS status frames during grounded chat, and introduce opt-in
agent infrastructure (Redis/Celery, AgentRun/Step, tools, orchestrator)
gated by ALLOW_AGENTIC_TASKS so default chat behaviour stays unchanged.

.env.example

@@ -94,6 +94,31 @@ REVENUECAT_WEBHOOK_SECRET=
 # Enforce plan feature + prompt/token quotas on chat turns (default true).
 # ENFORCE_SUBSCRIPTION_GATES=true
 FRONTEND_BASE_URL=http://localhost:3000
+
+# Agentic task execution (#63) — long-running, multi-step, tool-using turns.
+# Default false: chat behaves exactly like the always-on grounded path (#62),
+# no planner/tools/AgentRun rows. Requires a plan with allows_rag (or
+# allows_all_future_features) — see monetization SubscriptionPlan.allows_feature.
+ALLOW_AGENTIC_TASKS=false
+# Redis — optional. Unset = InMemory channel layer (single process, fine for
+# dev/tests) and agent work runs on a daemon thread instead of Celery.
+# REDIS_URL=redis://127.0.0.1:6379/0
+# CELERY_BROKER_URL=redis://127.0.0.1:6379/0
+# Orchestrator plans + synthesises; sub-agents run independent plan steps
+# concurrently on a smaller/cheaper model.
+# OLLAMA_MODEL_ORCHESTRATOR=gpt-oss:20b
+# OLLAMA_MODEL_SUBAGENT=llama3.2
+# AGENT_MAX_PLAN_STEPS=8
+# AGENT_MAX_ITERATIONS=12
+# AGENT_WALL_CLOCK_SECONDS=600
+# AGENT_SUBAGENT_CONCURRENCY=3
+# AGENT_TOOL_TIMEOUT_SECONDS=20
+# AGENT_TOOL_OUTPUT_MAX_CHARS=8000
+# AGENT_MAX_TOOL_CALLS_PER_RUN=40
+# AGENT_FETCH_URL_MAX_BYTES=2097152
+# Run a worker once REDIS_URL/CELERY_BROKER_URL are set:
+#   docker compose --profile agentic up redis worker
+#   uv run celery -A llm_be worker --loglevel=info
 # STRIPE_CHECKOUT_SUCCESS_URL=http://localhost:3000/billing/success?session_id={CHECKOUT_SESSION_ID}
 # STRIPE_CHECKOUT_CANCEL_URL=http://localhost:3000/billing/cancel
 # Customer Portal return URL (plan change / cancel / payment method).