While Pi's ~/.pi/agent/settings.json has a non-empty enabledModels (its Ctrl+P / /scoped-models list), a new session starts on defaultProvider/defaultModel only when that list takes the model in, else on the list's first entry, so a model picked in magpie outside the list was written and never used. Setting Pi's model now adds it to enabledModels as provider/model, as Pi's own picker does, when nothing in the list takes it in already: the same reference with or without a :level after it, or a glob (*, ?, [..], without case) that the reference or its model id fits. The user's entries, their order and comments stay; no list is made where there is none, and an empty one is left as it is. Turning the model back to Pi's own takes magpie/… entries out with magpie's provider, removing the key rather than leaving it empty. The model field shows what a new session starts on: the list's first model when the saved default isn't in the list and that entry names one model outright. The list itself isn't editable in magpie yet; Field is single-valued, and an editor needs a multi-value field kind in the API, GUI (a multi-select, reorderable picker), TUI and CLI.
Pi's models.json now gives each magpie model the API its provider speaks natively, so the gateway passes the request through instead of translating Chat: Responses-native models keep openai-responses; a Messages-only provider's (a custom Anthropic-compatible endpoint, as Cleary's) get api anthropic-messages with baseUrl the gateway's root (Anthropic's SDK adds /v1/messages). A Claude from 4.6 on also gets compat.forceAdaptiveThinking, so Pi asks for adaptive thinking and an effort, not the budget such a Claude refuses (why these stayed on Chat before). Magpie's models stay under providers.magpie; other providers are untouched.
Why Chat→Messages could 400 where Pi's direct Messages didn't: with reasoning on, buildAnthropic raised max_tokens to budget+4096 (28096 for high) past the client's own max_completion_tokens, which Pi sets to the model's output limit, so an upstream capped below that refused it. The budget now fits under a cap the client gives (at most the cap less 1024; no thinking when the cap is 2048 or less); with no cap it still raises max_tokens as before.
Tested: go test -tags nogui ./... (new: TestPiTakes, TestPiModelJoinsEnabledModels, TestPiModelOffLeavesEnabledModels, TestThinkingBudgetFitsTheClientsCap; TestPiModelsAskedOnTheirNativeAPI extended), nogui builds for darwin, linux and windows. In a sandbox HOME: settings.json with enabledModels [xai/grok-4.6, openai-codex/gpt-5.6-sol, deepseek/deepseek-flash], magpie pi magpie/cleary/kimi-k3 appended the model and kept the rest, magpie pi default took it out again; models.json listed cleary's models on anthropic-messages, a Chat relay's on the provider's openai-completions, a Responses vendor's on openai-responses. Pi 0.87.1 loaded that models.json (--list-models) and, against a fake Messages endpoint standing in for the gateway (MAGPIE_ADDR on a spare port), posted to /v1/messages with the gateway key: kimi-k3 with a 15360-token budget under max_tokens 16384, claude-sonnet-5 with adaptive thinking and effort high.
yetone
8060650048