# routing.toml — model routing by tier.
# One place to change which model powers each tier. Read by aos-sync and the
# Hermes orchestrator loop. Verify tool-calling quality per model before
# trusting it with tools.

[tiers.T1]
description = "Orchestrator, architect, reviewer, security — strongest available Claude"
model = "claude-sonnet-4-5"
max_tokens_per_call = 200000
allowed_tool_risk = ["low", "medium", "high"]

[tiers.T2]
description = "Department leads, complex workers"
model = "claude-sonnet-4-5"
max_tokens_per_call = 120000
allowed_tool_risk = ["low", "medium"]

[tiers.T3]
description = "Bulk workers: classification, drafts, summaries, reports"
model = "glm-4.5"
fallback_model = "deepseek-chat"
max_tokens_per_call = 60000
allowed_tool_risk = ["low"]
# Rule: no T3 model touches a risk:high tool without a T1/T2 reviewer pass.
requires_reviewer_for_high_risk = true

# Provider configuration — endpoints and which personal-data usage is allowed.
# Personal data (client PII) may only be sent to providers marked
# personal_data_allowed = true AND with an AVV (Art. 28 DSGVO) in place.

[providers.claude]
endpoint = "anthropic"
personal_data_allowed = true
avv_status = "ok"
notes = "EU data residency via AWS eu-west-1 / eu-central-1."

[providers.glm]
endpoint = "zai"
personal_data_allowed = false
avv_status = "pending_legal_review"
notes = "Use for non-PII bulk tasks only (code, public web content, summarization of anonymized data). Do NOT route client personal data here until AVV is signed."

[providers.deepseek]
endpoint = "deepseek"
personal_data_allowed = false
avv_status = "pending_legal_review"
notes = "Same restriction as glm — non-PII bulk only."