Models on Vertex AI (@gadhs/pi-vertex)

Rotation pickup. Every attempt re-resolves its credential: the anthropic kind’s cached clients are fingerprinted against the ADC file, the google kind’s adapter reads ADC per request, and the maas kind mints its bearer inside the attempt. Re-running gcloud auth application-default login rewrites the file, and the next request uses it — no pi restart.

Turn retry. A request that fails on expired auth is re-issued on a timer instead of killing the turn: re-login in another terminal and the turn resumes by itself - for as long as the turn is alive, unless adcRetry.maxWaitMs caps it. Configure with adcRetry. Auth failures happen at connection, before any content streams, so a retry never duplicates output; non-auth errors propagate immediately. A turn that ends in an error is never announced as a resume: with a cap that closed, the notice says how long it waited and that it gave up; another error stops the retry and says so; after your escape nothing is said. The re-login itself is exercised live (test/adc-relogin.live.test.ts): a copy of the credential with a dead refresh token fails the attempt, is rewritten good while the pump waits, and the same turn completes on the next poll.

One immediate refetch (maas). The bearer is cached per process until two minutes before expiry and single-flight; a 401 on a cached token is refetched once and the request replayed before any event reaches pi, so the consumer sees one stream. A 401 on a fresh token is the timer’s business.

anthropic adaptive (Opus 4.7+/5, Fable 5, Sonnet 4.6): thinking.type: "adaptive" + output_config.effort; pi levels map minimal/low→low, medium→medium, high→high, xhigh/max→xhigh iff "xhigh": true (else clamped to high). The extension injects compat.forceAdaptiveThinking — without it pi-ai silently falls back to a 1024-token budget and drops the effort.

anthropic budget (Haiku, older Sonnet/Opus): thinking.type: "enabled"
budget_tokens, growing max_tokens to keep budget_tokens < max_tokens.

"offSupported": false for anthropic models that reject disabled thinking (Fable 5).

google: pi’s level goes through pi-ai’s Gemini adapter (LOW / MEDIUM / HIGH; the catalogue’s map says what a model lacks — 3.x has no minimal). Thinking off is raised to low before the adapter sees it: pi-ai’s disabled config sends THINKING_LEVEL_MINIMAL for a Gemini 3 Flash and thinkingBudget: 0 for a 2.x, and Vertex refuses both (measured on 3.8 Flash, 3.7 Flash and 2.5 Pro), so every Gemini turn with thinking off was a 400 - including compaction, which runs thinking off. Off is not a thing these models offer; low is the floor pi-ai itself uses for the 3.x Pro line.

openai-compatible: pi’s level goes through as reasoning_effort; whether the vendor honours it is what the probe’s reasoning column measured. The system prompt always travels as role system. pi-ai would send it as developer for a reasoning model on a host its compat probe does not recognise, and the endpoint fronts vendors that drop a developer message unread - MiniMax M2 and Qwen3-Next answered a 113-character system prompt with 13-25 input tokens billed, meaning the mode prompt, the guidance and the agent body were silently absent on those two models (#113). system is the role every vendor behind the proxy honoured when probed. A vendor that opens its answer with a literal <think> tag in the content rather than a reasoning_content field - MiniMax M2 does - has the tag lifted into a thinking block as the stream passes (#112): text_* events become thinking_* up to </think>, the text after it is its own block, every partial and the final message agree, and a block that follows moves one index down. An unclosed tag is thinking to the end - the model ran out of budget mid-thought. Text that does not open with the tag passes untouched, byte for byte. The split is a view: on the way back in, this model’s own earlier turns get the tag put back - <think>…</think>, two newlines, the text, as one content string; the newline count after the tag is normalised, the rest is what the vendor wrote - because pi-ai replays an unsigned thinking block as nothing and MiniMax needs its reasoning in the history (its explore cell fell from 0.81 to 0.52 without it, 0.66 with it and the yield now the answer alone). Another model’s turns are left to pi-ai. Before this, pi’s transcript showed the reasoning as the answer and the cold reviewer’s verdict reader read it (think-tags.test.ts pins the split, the unclosed tag, the pass-through, the index shift and the rejoin; contract.test.ts pins the rejoin reaching the adapter).