plat 711

← coding-agents / pi

PLAT-711: Pi: deployment-supplied providers/models staged into each session

Field	Value
State	deployed
Priority	P2
Product	coding-agents
Area	pi
Summary	A deployment can supply Pi custom providers/models (PI_CLI_AGENT_TEMPLATE_DIR: models.json + thinking settings, no keys) that are staged into every session agent dir; first user Citymall's gateway.

What happened

Every AgentWorks Pi session runs with its own private PI_CODING_AGENT_DIR, so a Pi models.json placed in the service HOME was never seen: a deployment could not run Pi on its own OpenAI-compatible gateway (Citymall).

Fix

• multi-llm-provider-go af24edf (pi-cli adapter): PI_CLI_AGENT_TEMPLATE_DIR names a folder with models.json and an optional settings.json (only thinking keys). Both are validated and staged (0600) into each session agent dir before Pi starts; a bad template fails the launch. The template cannot hold a key: apiKey and credential headers must be $<PROVIDER>_API_KEY, the variable the adapter injects from the scoped provider credential (server env or a Pi account's underlying-provider key); !command values are refused. Launch-arg redaction covers any <PROVIDER>_API_KEY. Test: TestPiAgentTemplateStagedIntoSessionDirAndRefusesLiteralKeys.
• multi-llm-provider-go a7fcb3c: the template's models are in the Pi model catalog, so the app accepts them as a chat's model (without it a Code/Brain chat on citymall/gpt-6-luna was refused: "model is not offered for engine pi-cli").
• agent_go: the template's models lead the Pi model picker (first is the default) and a template provider key in the server environment makes the Pi server account configured (Pi's own --list-models cannot see it).
• Citymall's template (deploy/rootless-linux/products/citymall/pi-agent): gpt-6-luna first, gpt-5.6-luna second, thinking off mapped to "none" and off by default, because the gateway refuses tool calls with any other reasoning setting.

Left

• Pricing and context limits of the Citymall models are unknown; the picker shows none.