[shyoldman]

Problem

Qoder's live catalog gives several models a max thinking effort, usually marked as the default. GLM-5.3-Flash comes back with:

{ "high": {}, "max": { "is_default": true } }

PI_THINKING_LEVELS enumerates ["minimal", "low", "medium", "high", "xhigh"], so max never makes it into the map. pi's model picker then offers only high for that model, even though upstream defaults to max. I hit this on GLM-5.3 and GLM-5.3-Flash; Kimi-K3 has the same shape. Qwen Ultimate also advertises max — you just don't notice because it has xhigh too.

The ThinkingLevel type in pi-ai already contains "max", and pi core treats it as a first-class level. The array lags behind the SDK it targets.

Fix

Add "max" to PI_THINKING_LEVELS, so it maps max → max like the other identity mappings.

Verification

• npm run lint, npm run check, npm test (138 tests) all pass.
• On a locally patched build I ran /thinking max on GLM-5.3-Flash and captured the outgoing request body (before qoderEncodeBody, via a local echo endpoint):

{"model":"gfmodel","parameters":{"max_tokens":131072,"enable_thinking":true,"reasoning_effort":"max"}}