feat: route platform LLM calls through CommonStack first

[FlyM1ss]

DO NOT MERGE — work in progress: the CommonStack-first routing flip still lands on this branch.

Goal: route every platform-credits LLM call through CommonStack first, and fail over to OpenRouter when CommonStack runs out.

Done so far: Claude Haiku 4.5 (anthropic/claude-haiku-4-5) is added to COMMONSTACK_MODEL_ALLOWLIST. CommonStack's /v1/models confirms it serves that exact id.

• The provider seed is ON CONFLICT DO NOTHING, so changing the constant alone never reaches prod's existing commonstack row.
• A new migration marker, commonstack-allowlist-v1, fixes that in both the SQLite and Postgres stores. It adds only missing ids and keeps any an admin added.
• It leaves an unreadable row untouched, and it records the marker even when nothing changed, so an admin's later removal sticks.

To do:

• Add a platform provider-order helper (ATL_PLATFORM_PROVIDER_ORDER, default commonstack,openrouter). Apply it at model_providers/service.py resolve_platform_execution_candidates and list_execution_options, and at the legacy fallback in execution/service.py.
• Add a test pinning that a CommonStack exhaustion response triggers failover. The real error CommonStack returns when its balance runs out is unverified: failover keys on HTTP 402 or the balance/quota strings in adapters/base.py.
• Add a CLAUDE.md env entry.
• Out of scope: the AI Hedge Fund runtime stays on OpenRouter Nemotron, which CommonStack does not serve.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
agentic-trading-lab	Ready	Preview	Sep 25, 2026 9:47pm UTC